close

Вход

Забыли?

вход по аккаунту

?

Mangrove Species Mapping and Leaf Area Index Modeling Using Optical and Microwave Remote Sensing Technologies in Hong Kong

код для вставкиСкачать
Mangrove Species Mapping and Leaf Area Index Modeling Using
Optical and Microwave Remote Sensing Technologies in Hong Kong
WONG, Kwan Kit
A Thesis Submitted in Partial Fulfilment
of the Requirements for the Degree of
Doctor of Philosophy
in
Geography and Resource Management
The Chinese University of Hong Kong
May 2012
UMI Number: 3534634
All rights reserved
INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.
In the unlikely event that the author did not send a complete manuscript
and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.
UMI 3534634
Published by ProQuest LLC (2012). Copyright in the Dissertation held by the Author.
Microform Edition © ProQuest LLC.
All rights reserved. This work is protected against
unauthorized copying under Title 17, United States Code
ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106 - 1346
ACKNOWLEDGEMENTS
I would like to express my sincere gratitude to my supervisor, Prof. Fung Tung who works
along with me and providing me with guidance and support in this five-year research. He
has provided me with freedom and opportunities contributing to the success of this study.
He is my best teacher and friend in my life. Besides, I would like to thank my committee
members, Prof. Lin Hui, Prof. Lawal M. Marafa, Prof. Zhou Qiming and Prof. Wang Jinfei for
their time to go through such a long thesis and I treasure their valuable comments and
criticisms. I am also deeply thankful to the Institute of Space and Earth Information Science
(ISEIS) providing me with ENVISAT dataset which constitutes an important data source for
this study. In particular, I would like to thank Dr. Matthew Pang and Dr. Kin Yeung for
arranging the dataset for me.
Another critical component of this study was to collect field data deep in the mangrove
forest. I would like to thank Byford Tsang, Lowa Lo, Louis Lee and Nesta Poon who have
accompanied me to carry out the tedious and harsh survey work. Their hard work and
endurance in such an unfavorable environment are highly appreciated. I am greatly
indebted Byford Tsang because of the injury in the field.
I would like to show my sincere thanks to Honda Lee, Kai-hang Choi, Kelvin Lee, Fung-wai
Lui and Oscar for their technical and logistic supports; Anna Ng, Jane Wan, Sophie Yuen and
To-ze for their administrative supports. At the same time, I would also like to show my
sincere thanks to all professors in the Department of Geography and Resource
Management in CUHK who have kindly provided supports and comments for my studies
and works.
Whole-hearted thanks are given to my colleagues and friends in the Planning Department:
Ida Cheung, Terry Chan, Jonathan Tse, Patrick Lau, Wing Leung, Erin Yeung, Vincent Cheung,
T.S. Wong, T.W. Ng, Anna Chan and Salina Mou for their friendship, encouragement and
sharing my ups and downs in the past three years.
I would also like to take this opportunity to thank Professor Petr Somol for providing me
guidance on pattern recognition, especially his enthusiastic support on FST3. I won’t be able
to learn and compile the scripts in such a short time without his sincere advices.
ii
Thank you to my parents, my aunt, my little cousin, my sister. Sincere thanks are given to
my grandparents who have been taking caring of me since I was born. Your altruistic care is
engraved deeply in my heart. I am proud to let you know that I have finally completed my
doctoral program. I am deeply grateful to have Carmen Ng and my two little dogs, Jai-Jai
and Charcoal as companionship. They have been supporting me, sharing my burdens and
giving me fun while I was working hard with my research.
iii
ABSTRACT
Abstract of thesis entitled:
Mangrove Species Mapping and Leaf Area Index Modeling Using Optical and
Microwave Remote Sensing Technologies in Hong Kong
Submitted by WONG Kwan-kit
for the degree of Doctor of Philosophy
at The Chinese University of Hong Kong in (October 2011)
Mangrove is one of the most productive ecosystems flourished in the intertidal
zone of tropical and subtropical regions. Hong Kong has ten true mangrove species
covering an approximate area of 350 hectares. Mai Po locating in the northwestern
part of Hong Kong nourishes the largest mangrove stand and it was listed as a
Wetland of Importance under the Ramsar Convention in 1995. Over the years, areas
of mangrove have been shrinking globally due to development, pollution, and other
unsustainable exploitation and Hong Kong was no exception. In Hong Kong,
mangroves are usually sacrificed for urban development and infrastructure
construction. Therefore, it is crucial to monitor their growth conditions, change of
extent and possible unsustainable practices threatening their existence. Remote
sensing being a cost-effective and timely tool for vegetation conservation is most
suitable for such purpose.
Taking Mai Po as study area, this study acquired satellite-borne hyperspectral and
radar data supplemented with in situ field survey to achieve three purposes. First,
features from the remotely-sensed data that are significant to species
discrimination were identified through pattern recognition. Second, selected
features grouped into different subsets were used to delineate the boundary of
iv
mangrove species through supervised classification. In the meantime, classifiers
including maximum likelihood (ML), decision tree C5.0 (DT), artificial neural network
(ANN) and support vector machines (SVM) were tested for their accuracy
performance. The third purpose is to understand the current biophysical condition
of mangrove through leaf area index (LAI) modeling by regressing field-measured
LAI against vegetation indices, backscatter and textural measures.
Results from feature selection revealed that hyperspectral narrowbands locating in
green at 570nm, 580nm, 591nm, 601nm; red at 702nm; red-edge at 713nm; near
infrared at 764nm and 774nm and shortwave infrared at 1276nm, 1316nm and
1629nm as well as the multi-temporal filtered backscatter captured in different
seasons have high sensitivity to species difference.
Species-based classification using multi-temporal backscatter features alone do not
provide a satisfactory accuracy. Comparatively, results from pure spectral bands
have better overall accuracy than that from combining spectral and radar features.
However, radar backscatter does improve accuracy of some species. Besides, all
classifiers had similar variations of training accuracy under the same feature subset.
However, the testing accuracy is much lower with the exception of ANN.
Performance of ANN was more stable and robust than other classifiers while serious
overtraining occurs for the DT classifier. Moreover, most species were mapped
accurately as revealed by the producer’s and user’s accuracy with the exception of
A. corniculatum and Sonneratia spp. due to deficiency of training samples.
Simple linear regression model with VIs revealed that triangular vegetation index
(TVI) and modified chlorophyll absorption ratio index 1 (MCARI1) had the best
relationship with LAI. However, weak relationship was found between fieldmeasured LAI and radar parameters suggesting that radar parameters cannot be
used as single predictor for LAI. Results from stepwise multiple regression
suggested that TVI combined with GLCM-derived angular second moment (ASM)
can reduce the estimation error of LAI. To conclude, the study has demonstrated
spectral and radar data are complementarity for accurate species discrimination
and LAI mapping.
v
論文摘要
生長於潮間帶的紅樹林是熱帶和亞熱帶地區最具生產力的生態系統之一。香港擁有十
個紅樹品種,其覆蓋面積約共三百五十公頃。位於香港西北面的米埔是現時香港最大
的紅樹林區。這片紅樹林及其鄰近濕地於一九九五年被列為拉姆薩爾重要的濕地。隨
著經濟的迅速發展、污染及一些不可持續的開發,全球紅樹林的面積不斷地萎縮。而
香港的紅樹也正面對城市發展及基建的直接威脅。因此,了解及監測紅樹林的生長狀
況、覆蓋面積的轉變是紅樹林保育的基礎。遙感是具有成本效益和能提供及時數據的
技術,在紅樹林的生態保育及監測上發揮著重要功能。
是次研究選擇位於米埔的紅樹林區。通過結合高光譜和雷達數據以及實地磡測,以達
到三個目的。第一,利用模式辨認分析找出可提高品種辨識度的光譜帶及雷達數據。
第二,把挑選出來的光譜帶及雷達數據組合,利用不同的分類法包括最大概似法、决
策樹 C5.0 演算法、類神經網路及支持向量機進行紅樹林的品種分類,並籍此測試各分
類法的精度。第三,利用植被指數及雷達數據中取得的參數為獨立變量,而在野外點
測的葉面積指數 (LAI) 為因變量,通過迴歸分析以估算整片紅樹林的葉面積指數,籍此
了解紅樹林現時的生物物理狀況。
根 據 特 徵 選 擇 的 結 果 , 位 於 高 光 譜 數 據 中 的 綠 波 段 (570nm, 580nm, 591nm 及
601nm)、紅波段 (702nm)、紅邊位 (713nm)、近紅外波段 (764nm 及 774nm)、 短波紅
外波段 (1276nm, 1316nm 及 1629nm) 以及在不同季節取得的過濾後向散射數據是最能
辨識品種差異。
據品種分類的結果顯示,單用多時後向散射特徵數據存在很大誤差。而在大多的情況
下,單用光譜數據比起混合光譜及後向散射數據的分類表現為佳。但對於某些品種來
說,後向散射數據能給予比較準確的預測。另外,在同數據組合下,分類法在訓練精
度上沒有多大的分別。除了類神經網路分類法以外,其他分類法的測試精度總比其訓
練精度低。這說明類神經網路模型比起其他分類法的模型要為穩定,而决策樹模型則
被過度訓練。根據生產者及使用者精度分析,因為缺乏足夠的訓練樣本,桐花樹及海
桑屬的精度較其他品種為低。
據不同植被指數的簡單線性迴歸模型顯示,利用三角植被指數 (TVI)及修正葉綠素吸納
比例指數一 (MCARI 1) 對於葉面積指數的估算是最準確。相反地,葉面積指數與從雷
vi
達數據中取得的參數關係則比較弱。這表示單用雷達參數不能對葉面積指數進行準確
的估算。在結合植被指數及雷達參數的多元逐步迴歸分析下,三角植被指數及在灰度
共生矩陣下得出的角二階矩參數能減低葉面積指數估算的誤差。總結以上兩項分析,
光譜及雷達數據在紅樹林的品種分類及葉面積指數估算上有互補的作用。
vii
TABLE OF CONTENTS
ACKNOWLEDGEMENTS ........................................................................................ II
ABSTRACT ……… .................................................................................................. IV
論文摘要 …… ..................................................................................................... VI
TABLE OF CONTENTS......................................................................................... VIII
LIST OF ABBREVIATIONS ................................................................................... XIII
LIST OF TABLES .................................................................................................. XV
LIST OF FIGURES ............................................................................................. XVIII
CHAPTER 1
1.1.
INTRODUCTION ............................................................................. 1
BACKGROUND TO THE STUDY ..................................................................................................... 1
1.1.1.
Mangrove Mapping and Monitoring .............................................................................. 1
1.1.2.
Mangrove Mapping and Monitoring .............................................................................. 3
1.1.3.
Role of Remote Sensing in Mangrove Study ................................................................... 4
1.2.
OBJECTIVES OF THE STUDY ......................................................................................................... 6
1.3.
SIGNIFICANCE OF THE STUDY ...................................................................................................... 7
1.4.
ORGANIZATION OF THE THESIS.................................................................................................... 8
CHAPTER 2
LITERATURE REVIEW .................................................................... 10
2.1.
INTRODUCTION ...................................................................................................................... 10
2.2.
FACTORS AFFECTING VEGETATION REFLECTANCE.......................................................................... 11
2.2.1.
Foliar structure and principal constituents ................................................................... 12
2.2.2.
Foliar optical properties................................................................................................ 14
2.2.2.1.
The visible region (400-700nm) ........................................................................................ 14
2.2.2.2.
The red edge (690-740nm) ............................................................................................... 15
2.2.2.3.
The near-infrared region (700-1300nm) ........................................................................... 16
2.2.2.4.
The short-wave infrared region (1300-2500nm) .............................................................. 17
2.2.3.
Canopy architecture ..................................................................................................... 18
2.2.4.
Background reflectance ................................................................................................ 19
2.2.5.
Atmospheric perturbation ............................................................................................ 20
2.2.6.
Sun-sensor relationship ................................................................................................ 22
2.3.
HYPERSPECTRAL IMAGING AND VEGETATION CLASSIFICATION ......................................................... 23
2.4.
RADAR IMAGING AND VEGETATION CLASSIFICATION ..................................................................... 31
viii
2.5.
PATTERN RECOGNITION FOR VEGETATION CLASSIFICATION............................................................. 39
2.5.1.
The Hughes Phenomenon and Dimensionality Reduction ............................................ 39
2.5.2.
Statistical Pattern Recognition and Feature Selection ................................................. 44
2.5.2.1.
Search Method ................................................................................................................. 47
2.5.2.1.1.
Exhaustive search ........................................................................................................ 48
2.5.2.1.2.
Branch and bound ....................................................................................................... 49
2.5.2.1.3.
Sequential forward/ backward selection ..................................................................... 55
2.5.2.1.4.
Sequential Floating search........................................................................................... 57
2.5.2.1.5.
Oscillating Search ........................................................................................................ 61
2.5.2.1.6.
Genetic algorithm ........................................................................................................ 64
2.5.2.2.
Evaluation criteria ............................................................................................................. 66
2.5.2.2.1.
Distance measure ........................................................................................................ 67
2.5.2.2.2.
Information measure ................................................................................................... 68
2.5.2.2.3.
Classification error ....................................................................................................... 71
2.5.2.3.
2.5.3.
2.6.
Feature Selection Stability ................................................................................................ 72
Feature extraction ........................................................................................................ 75
BIOPHYSICAL PARAMETERS MEASUREMENT AND ESTIMATION ........................................................ 77
2.6.1.
Leaf Area Index (LAI)..................................................................................................... 78
2.6.2.
Fraction of Absorbed Photosynthetically Active Radiation (fAPAR) ............................. 79
2.6.3.
In-situ Leaf Area Index Measurement .......................................................................... 81
2.6.3.1.
Direct and Indirect Methods ............................................................................................. 81
2.6.3.2.
LAI Estimation through Gap Fraction Inversion ................................................................ 85
2.6.3.3.
Gap Fraction Ground Measurement ................................................................................. 89
2.6.3.3.1.
LAI-2000 Plant Canopy Analyzer .................................................................................. 89
2.6.3.3.2.
Hemispherical Photography ........................................................................................ 92
2.6.3.4.
2.6.3.4.1.
Clumping.................................................................................................................... 100
2.6.3.4.2.
Mixture of Green and Non-green Elements .............................................................. 101
2.6.4.
Empirical Relationship with Spectral Vegetation Indices............................................ 102
2.6.4.1.
Traditional Vegetation Indices ........................................................................................ 103
2.6.4.2.
Leaf Area Index Estimation from Hyperspectral and Radar Images ................................ 106
2.6.5.
2.7.
Correction of Indirect LAI Measurement .......................................................................... 99
Physically-based Canopy Reflectance Model Inversion .............................................. 111
2.6.5.1.
Canopy Reflectance Model ............................................................................................. 111
2.6.5.2.
Model Inversion and Biophysical Parameters Extraction ............................................... 115
SUMMARY .......................................................................................................................... 118
CHAPTER 3
METHODOLOGY ..........................................................................120
3.1.
INTRODUCTION .................................................................................................................... 120
3.2.
STUDY AREA DESCRIPTION ..................................................................................................... 120
3.3.
METHODOLOGICAL FLOW ...................................................................................................... 124
ix
3.4.
REMOTE SENSING DATA ACQUISITION AND PROCESSING ............................................................. 127
3.4.1.
Hyperion - EO-1 .......................................................................................................... 127
3.4.1.1.
Radiometric correction ................................................................................................... 127
3.4.1.1.1.
Vertical strips removal ............................................................................................... 128
3.4.1.1.2.
Atmospheric correction ............................................................................................. 129
3.4.1.1.3.
Wavelength recalibration .......................................................................................... 135
3.4.1.1.4.
SNR enhancement through MNF ............................................................................... 137
3.4.1.2.
Geometric correction...................................................................................................... 139
3.4.1.3.
Atmospheric correction algorithms comparison ............................................................ 140
3.4.2.
ASAR - ENVISAT .......................................................................................................... 141
3.4.2.1.
Data Acquisition .............................................................................................................. 141
3.4.2.2.
Data Processing .............................................................................................................. 143
3.5.
3.4.2.2.1.
Radiometric and Geometric Correction ..................................................................... 145
3.4.2.2.2.
Speckle Filtering ........................................................................................................ 146
FIELD MEASUREMENTS AND DATA PROCESSING ......................................................................... 149
3.5.1.
Species Distribution .................................................................................................... 149
3.5.2.
Leaf Spectra Measurement ........................................................................................ 151
3.5.2.1.
Leaf Collection and Handling .......................................................................................... 152
3.5.2.2.
ASD FieldSpec 3 Setup .................................................................................................... 154
3.5.2.3.
Laboratory setup............................................................................................................. 156
3.5.2.4.
Spectra Measurement .................................................................................................... 158
3.5.2.5.
Spectral similarity and variability .................................................................................... 159
3.5.3.
In situ Leaf Area Index Measurement ......................................................................... 161
3.5.3.1.
The optical instrument.................................................................................................... 161
3.5.3.2.
The LAI survey campaign ................................................................................................ 163
3.5.3.3.
Data processing and canopy analysis .............................................................................. 166
3.5.3.4.
Canopy parameter computation – gap fraction, LAI, clumping index, mean inclination
angle ............................................................................................................................... 170
3.5.3.5.
Field LAI and Their Correlation with Reflectance and Backscattering Coefficient Data
Exploration ..................................................................................................................... 175
3.6.
FEATURE SELECTION ............................................................................................................. 175
3.6.1.
Data Preprocessing and Preparation .......................................................................... 178
3.6.2.
Data Format and Split ................................................................................................ 183
3.6.3.
Wrapper-based Approach .......................................................................................... 185
3.6.4.
Search Algorithm ........................................................................................................ 187
3.6.5.
Stability Evaluation ..................................................................................................... 187
3.6.6.
Feature Frequency analysis ........................................................................................ 188
3.7.
MANGROVE SPECIES CLASSIFICATION ....................................................................................... 189
3.7.1.
Species Separability .................................................................................................... 193
3.7.2.
Gaussian Maximum Likelihood Classifier ................................................................... 193
x
3.7.3.
Decision Tree Classifier ............................................................................................... 194
3.7.4.
Artificial Neural Network Classifier ............................................................................ 197
3.7.5.
Support Vector Machines Classifier ............................................................................ 199
3.7.6.
Accuracy Assessment .................................................................................................. 204
3.8.
LEAF AREA INDEX MODELING ................................................................................................. 206
3.8.1.
Preliminary Exploration of Relationship between Hyperspectral bands and LAI ........ 206
3.8.2.
Vegetation Index Derived from Hyperspectral Data................................................... 206
3.8.3.
Radar Backscatter and Derived Textural Parameters ................................................ 208
3.8.4.
Regression Analysis .................................................................................................... 211
3.8.5.
Error Estimation.......................................................................................................... 217
3.9.
SUMMARY .......................................................................................................................... 218
CHAPTER 4
RESULTS AND DISCUSSION (I) – FEATURE SELECTION AND
MANGROVE SPECIES CLASSIFICATION .........................................221
4.1.
INTRODUCTION .................................................................................................................... 221
4.2.
DATA PROCESSING AND EXPLORATION ..................................................................................... 221
4.2.1.
Atmospheric correction algorithms comparison ........................................................ 222
4.2.2.
Radar Data Speckle Reduction ................................................................................... 227
4.2.3.
Statistical Discrimination of Mangrove Spectral Class ............................................... 230
4.3.
FEATURE SELECTION ............................................................................................................. 249
4.3.1.
Sequential Forward Selection (SFS) ............................................................................ 250
4.3.2.
Sequential Floating Forward Selection (SFFS)............................................................. 256
4.3.3.
Oscillating Search (OS) ............................................................................................... 262
4.3.4.
Search Algorithms comparison ................................................................................... 268
4.3.5.
Final Subset Selection ................................................................................................. 270
4.3.6.
Correlation Analysis .................................................................................................... 280
4.4.
IMAGE CLASSIFICATION ......................................................................................................... 283
4.4.1.
Mangrove Spectral Class Separability ........................................................................ 284
4.4.2.
Gaussian Maximum Likelihood (ML) .......................................................................... 288
4.4.3.
Decision Tree (DT) ....................................................................................................... 297
4.4.4.
Artificial Neural Network (ANN) ................................................................................. 304
4.4.5.
Support Vector Machines (SVM) ................................................................................ 312
4.4.6.
Algorithm Comparison................................................................................................ 321
4.5.
DISCUSSION AND IMPLICATION ............................................................................................... 325
4.5.1.
Feature Selection ........................................................................................................ 325
4.5.2.
Mangrove Classification ............................................................................................. 342
4.6.
SUMMARY .......................................................................................................................... 351
xi
CHAPTER 5
RESULTS AND DISCUSSION (II) - LEAF AREA INDEX MODELING .....353
5.1.
INTRODUCTION .................................................................................................................... 353
5.2.
DATA EXPLORATION ............................................................................................................. 353
5.2.1.
Dependent Variable: Field measured LAI.................................................................... 353
5.2.2.
Independent Variables: Vegetation Index and texture measure ................................ 355
5.2.3.
Hyperspectral bands and LAI ...................................................................................... 356
5.2.4.
Normality testing ........................................................................................................ 359
5.2.5.
Linearity testing .......................................................................................................... 363
5.2.6.
Outliner detection....................................................................................................... 365
5.3.
SIMPLE LINEAR REGRESSION ANALYSIS ..................................................................................... 366
5.3.1.
5.4.
LAI2000 Generalized method ..................................................................................... 369
STEPWISE MULTIPLE REGRESSION ANALYSIS .............................................................................. 381
5.4.1.
5.5.
LAI2000 Generalized method ..................................................................................... 384
DISCUSSION AND IMPLICATION ............................................................................................... 391
5.5.1.
LAI model comparison ................................................................................................ 391
5.5.2.
Species composition and LAI....................................................................................... 393
5.5.3.
Hyperspectral Bands, Vegetation Indices and LAI ...................................................... 397
5.5.4.
Backscatter, texture measures and LAI ...................................................................... 407
5.5.5.
Complementarity of Vegetation Index and Radar Parameters .................................. 414
5.6.
SUMMARY .......................................................................................................................... 421
CHAPTER 6
CONCLUSION ..............................................................................423
6.1.
SUMMARY OF THE STUDY ...................................................................................................... 423
6.2.
LIMITATION OF THE STUDY ..................................................................................................... 427
6.3.
RECOMMENDATION.............................................................................................................. 431
REFERENCE …....................................................................................................434
APPENDIX A GEOMETRIC CORRECTION OF HYPERSPECTRAL DATA ....................473
APPENDIX B SCRIPTS DERIVED FROM FEATURE SELECTION TOOLBOX (FST) FOR
FEATURE SELECTION ...................................................................475
APPENDIX C PREDICTED LAI(BON) AND LAI (2000) FROM SIMPLE LINEAR
REGRESSION MODELS .................................................................513
APPENDIX D PREDICTED LAI(BON) AND LAI(2000) FROM MULTIPLE STEPWISE
REGRESSION MODELS .................................................................524
xii
LIST OF ABBREVIATIONS
AFCD
: Agriculture, Fisheries and Conservation Department
ANN
: Artificial Neural Network
ASD
: Analytical Spectral Devices
ASM
: GLCM-derived angular second moment texture
ATCOR-2
: Atmospheric and Topographic Correction model 2
BS
: Backscatter
CORR
: GLCM-derived correlation texture variable
CT
: GLCM-derived contrast texture variable
C-VV
: Band C, Vertically transmitted and vertically received
DN
: Digital Number
DSM
: GLCM-derived dissimilarity texture
DT
: Decision Tree
ENT
: GLCM-derived entropy texture
FLAASH
: Fast Line-of-Site Atmospheric Analysis
FS3
: FieldSpec® 3 Portable Spectroradiometer
FST
: Feature Selection Toolbox
GCPs
: Ground Control Points
GLCM
: Grey-Level Co-occurrence Matrix
GPS
: Global Positioning System
HOMO
: GLCM-derived homogeneity texture
IO
: Influential outliners
Knn
: k-nearest neighbor
LAI
: Leaf Area Index
LAIBon
: LAI derived by Bonhomme and Chartier’s method
LAI2000
: LAI derived by LAI2000 plant canopy analyzer method
LAI2000G
: LAI derived by LAI2000 plant canopy analyzer generalized method
xiii
MCARI
: Modified Chlorophyll Absorption Ratio Index
MN
: GLCM-derived mean texture
MNF
: Minimum Noise Fraction
ML
: Maximum Likelihood
MO
: Multivariate outliners
MODTRAN
: Moderate Resolution Transmittance
MSAVI
: Modified Soil-Adjusted Vegetation Index
NDVI
: Normalized Difference Vegetation Index
OS
: Oscillating Search
RBF
: Radial Basis Function
RDVI
: Renormalized Difference Vegetation Index
RMSE
: Root Mean Square Error
SAR
: Synthetic Aperture Radar
SAVI
: Soil-Adjusted Vegetation Index
SD
: GLCM-derived standard deviation texture
SFS
: Sequential Forward Selection
SFFS
: Sequential Floating Forward Selection
SNR
: Signal-to-noise ratio
SVM
: Support Vector Machine
SWIR
: Short-wave Infrared
TD
: Transformed Divergence
TVI
: Triangular Vegetation Index
UO
: Univariate outliners
VI
: Vegetation Index
VNIR
: Visible-Near Infrared
xiv
LIST OF TABLES
TABLE 2.1. SUMMARY OF STUDY USING FEATURE SELECTION TECHNIQUE FOR HYPERSPECTRAL FEATURE SELECTION ...... 27
TABLE 2.2. MICROWAVE BANDS AND SAR SENSORS ........................................................................................ 32
TABLE 2.3. THE DESCRIPTION OF INPUT PARAMETERS IN CR MODELS ................................................................ 113
TABLE 3.1. THE MAGNITUDE OF STRIPPING IN HYPERION DATA ........................................................................ 128
TABLE 3.2. THE PARAMETRIC SETTING FOR ATCOR-2 ATMOSPHERE CORRECTION ............................................... 131
TABLE 3.3. THE FLAASH PARAMETERS SETTING FOR ATMOSPHERIC CORRECTION................................................ 133
TABLE 3.4. THE SAR DATA LIST FOR ASAR-ENVISAT ................................................................................... 141
TABLE 3.5. THE GENERAL WEATHER CONDITIONS AND TIDE INFORMATION FOR MULTI-TEMPORAL ASAR DATA ......... 143
TABLE 3.6. COMBINATION OF IN-SITU AND MANUALLY COLLECTED SAMPLE POINT OF DIFFERENT MANGROVE SPECIES . 150
TABLE 3.7. THE CANOPY MEASUREMENTS EXTRACTED FROM HEMISPHERICAL PHOTOS .......................................... 171
TABLE 3.8. SUMMARY OF HYPERSPECTRAL FEATURES (0-152) ......................................................................... 179
TABLE 3.9. SUMMARY OF MULTI-TEMPORAL SAR FEATURES (153-184) ........................................................... 182
TABLE 3.10. NUMBER OF TRAINING AND TESTING OBSERVATIONS FOR THE MANGROVE SPECTRAL CLASSES ............... 190
TABLE 3.11. THE RANGE AND STEP OF CHANGE FOR SEARCH OF OPTIMAL C AND

.......................................... 204
TABLE 3.12. THE SPECTRAL VEGETATION INDICES (VIS) FOR LAI MODELING ....................................................... 207
TABLE 3.13. GLCM-DERIVED TEXTURAL VARIABLES AND THEIR CORRESPONDING EQUATIONS ................................ 210
TABLE 4.1. ACCURACY STATISTICS AND CONSISTENCY INDICES AFTER SFS FEATURE SELECTION ................................ 251
TABLE 4.2. GLOBAL MEAN AND STANDARD DEVIATION OF CLASSIFICATION ACCURACY CALCULATED BASED ON SUBSET SIZE
........................................................................................................................................ 252
TABLE 4.3. NUMBER OF TRIAL WITH ACCURACY BELOW 1 STANDARD DEVIATION FROM MEAN (SFS ALGORITHM) ...... 252
TABLE 4.4. ACCURACY STATISTICS AND CONSISTENCY INDICES AFTER SFFS FEATURE SELECTION .............................. 258
TABLE 4.5. NUMBER OF TRIAL WITH ACCURACY BELOW ONE STANDARD DEVIATION FROM MEAN (SFFS ALGORITHM) . 259
TABLE 4.6. ACCURACY STATISTICS AND CONSISTENCY INDICES AFTER OS FEATURE SELECTION ................................. 264
TABLE 4.7. NUMBER OF TRIAL WITH ACCURACY BELOW ONE STANDARD DEVIATION FROM MEAN (SFFS ALGORITHM) . 265
TABLE 4.8. BEST 45 FEATURES ARRANGED IN DESCENDING CRITERION SCORE...................................................... 279
TABLE 4.9. LIST OF SIGNIFICANTLY HIGHLY CORRELATED FEATURES IN THE FINAL FEATURE SUBSET ........................... 282
TABLE 4.10. THE FINALIZED FEATURES HAVING THE HIGHEST CRITERION SCORE AFTER FEATURE SELECTION ANALYSIS .. 283
TABLE 4.11. FEATURE SUBSETS FORMULATED BY THE COMBINING BEST FEATURES IN THE FINAL FEATURE SET ............ 283
TABLE 4.12. THE CONFUSION MATRIX AFTER MAXIMUM LIKELIHOOD (ML) CLASSIFICATION USING FEATURE COMBINATION
OF SUBSET 1 ....................................................................................................................... 293
xv
TABLE 4.13. THE CONFUSION MATRIX AFTER MAXIMUM LIKELIHOOD (ML) CLASSIFICATION USING FEATURE COMBINATION
OF SUBSET 2 ....................................................................................................................... 294
TABLE 4.14. THE CONFUSION MATRIX AFTER MAXIMUM LIKELIHOOD (ML) CLASSIFICATION USING FEATURE COMBINATION
OF SUBSET 3 ....................................................................................................................... 295
TABLE 4.15. THE CONFUSION MATRIX AFTER MAXIMUM LIKELIHOOD (ML) CLASSIFICATION USING FEATURE COMBINATION
OF SUBSET 4 ....................................................................................................................... 295
TABLE 4.16. THE CONFUSION MATRIX AFTER DECISION TREE (DT) CLASSIFICATION USING FEATURE COMBINATION OF
SUBSET 1 ........................................................................................................................... 301
TABLE 4.17. THE CONFUSION MATRIX AFTER DECISION TREE (DT) CLASSIFICATION USING FEATURE COMBINATION OF
SUBSET 2 ........................................................................................................................... 302
TABLE 4.18. THE CONFUSION MATRIX AFTER DECISION TREE (DT) CLASSIFICATION USING FEATURE COMBINATION OF
SUBSET 3 ........................................................................................................................... 302
TABLE 4.19. THE CONFUSION MATRIX AFTER DECISION TREE (DT) CLASSIFICATION USING FEATURE COMBINATION OF
SUBSET 4 ........................................................................................................................... 303
TABLE 4.20. THE CONFUSION MATRIX AFTER ARTIFICIAL NEURAL NETWORK (ANN) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 1 ................................................................................................... 309
TABLE 4.21. THE CONFUSION MATRIX AFTER ARTIFICIAL NEURAL NETWORK (ANN) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 2 ................................................................................................... 310
TABLE 4.22. THE CONFUSION MATRIX AFTER ARTIFICIAL NEURAL NETWORK (ANN) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 3 ................................................................................................... 310
TABLE 4.23. THE CONFUSION MATRIX AFTER ARTIFICIAL NEURAL NETWORK (ANN) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 4 ................................................................................................... 311
TABLE 4.24. THE OPTIMIZED HYPERPARAMETERS COMBINATIONS AND THEIR CORRESPONDING ACCURACY IN DIFFERENT
SUBSETS............................................................................................................................. 312
TABLE 4.25. THE CONFUSION MATRIX AFTER SUPPORT VECTOR MACHINES (SVM) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 1 ................................................................................................... 317
TABLE 4.26. THE CONFUSION MATRIX AFTER SUPPORT VECTOR MACHINES (SVM) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 2 ................................................................................................... 318
TABLE 4.27. THE CONFUSION MATRIX AFTER SUPPORT VECTOR MACHINES (SVM) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 3 ................................................................................................... 319
TABLE 4.28. THE CONFUSION MATRIX AFTER SUPPORT VECTOR MACHINES (SVM) CLASSIFICATION USING FEATURE
COMBINATION OF SUBSET 4 ................................................................................................... 319
TABLE 4.29. COMPARISON OF CLASSIFICATION ACCURACY OF DT MODELS WITH AND WITHOUT BOOSTING OPTION .... 342
TABLE 5.1. THE DESCRIPTIVE STATISTICS OF LEAF AREA INDEX (LAI) COMPUTED USING BONHOMME AND CHARTIER’S,
LAI2000 AND LAI2000 GENERALIZED METHODS ...................................................................... 355
TABLE 5.2. THE DESCRIPTIVE STATISTICS OF VI AND RADAR PARAMETERS ........................................................... 355
xvi
TABLE 5.3. THE LEVEL OF SIGNIFICANCE AFTER KOLMOGOROV-SMIRNOV TEST .................................................... 362
TABLE 5.4. CORRELATION BETWEEN DEPENDENT VARIABLES (LAI) AND INDEPENDENT VARIABLES (VEGETATION INDEX AND
RADAR PARAMETER) ............................................................................................................. 364
TABLE 5.5. THE TYPES AND NUMBERS OF OUTLINERS DETECTED........................................................................ 366
TABLE 5.6. THE PEARSON CORRELATION COEFFICIENT BETWEEN LAI AND VIS AFTER OUTLINER REMOVAL ................ 366
TABLE 5.7. SIMPLE LINEAR REGRESSION MODELS COMPUTED FROM DIFFERENT VIS .............................................. 368
TABLE 5.8. DESCRIPTIVE STATISTICS FOR LAI (2000G) PREDICTED OVER THE STUDY AREA ..................................... 378
TABLE 5.9. STEPWISE REGRESSION MODELS COMPUTED FROM DIFFERENT VIS AND RADAR TEXTURAL VARIABLES........ 383
TABLE 5.10. DESCRIPTIVE STATISTICS FOR LAI(2000G) PREDICTED OVER THE STUDY AREA ................................... 390
TABLE 5.11. GROUP LAI CLASSES OF THE STUDY AREA ................................................................................... 398
2
TABLE 5.12. COEFFICIENTS OF DETERMINATION (R ) AND RMSE DIFFERENCES BETWEEN SIMPLE LINEAR AND STEPWISE
MODELS ............................................................................................................................. 414
TABLE 5.13. CORRELATION BETWEEN GLCM-DERIVED ANGULAR SECOND MOMENT (ASM) FROM BACKSCATTER (BS)
DATA AND DIFFERENT VEGETATION INDICES (VIS) ....................................................................... 420
xvii
LIST OF FIGURES
FIGURE 1.1 CHANGES IN WORLD MANGROVE AREAS (1980 – 2005) .................................................................... 2
FIGURE 2.1. A TYPICAL FOLIAR STRUCTURE..................................................................................................... 12
FIGURE 2.2. A TYPICAL SPECTRAL SIGNATURE OF VEGETATION IN 350 – 2500NM ................................................. 14
FIGURE 2.3. THE ABSORPTION REGIONS OF DIFFERENT ATMOSPHERIC CONSTITUENTS ............................................. 22
FIGURE 2.4. THE BACKSCATTERING MODEL COMPONENTS OF (A) WOODY AND (B) HERBACEOUS MANGROVE .............. 35
FIGURE 2.5. THE CLASSIFICATION ACCURACY AS A FUNCTION OF TRAINING SAMPLES AND DIMENSIONALITY ................. 41
FIGURE 2.6. THE PROCEDURE OF FEATURE SELECTION, A) FILTER APPROACH AND B) WRAPPER APPROACH. ................. 45
FIGURE 2.7. THE COMMONLY USED SEARCH ALGORITHM FOR FEATURE SELECTION ................................................. 48
FIGURE 2.8. EXAMPLE OF BRANCH AND BOUND SOLUTION, WHERE 2 FEATURES ARE TO BE SELECTED FROM THE ORIGINAL
FEATURE WITH 5 FEATURES (D =5 AND D =2). THE AIM IS TO MAXIMIZE THE ILLUSTRATIVE CRITERION
FUNCTION ......................................................................................................................... 51
FIGURE 2.9. THE CONCEPT OF MONOTONICITY FOR CLASSIFICATION PROBLEMS ..................................................... 52
FIGURE 2.10. THE SEQUENTIAL FORWARD FLOATING SELECTION (SFFS) ALGORITHM ............................................. 58
FIGURE 2.11. THE PROCESS OF OSCILLATING SEARCH ....................................................................................... 62
FIGURE 2.12. GRAPHICAL COMPARISON OF A) SEQUENTIAL FORWARD SELECTION, B) SEQUENTIAL FORWARD FLOATING
SELECTION, AND C) OSCILLATING SEARCH................................................................................. 63
FIGURE 2.13. FLOWCHART SHOWING THE GENERALIZED PROCEDURE USED IN A GENETIC ALGORITHM ........................ 65
FIGURE 2.14. A) DIURNAL VARIATION OF FAPAR WITHIN A SPATIAL UNIFORM CANOPY; B) PHOTOSYNTHETICALLY ACTIVE
RADIATION (PAR) REFLECTION, TRANSMITTANCE AND ABSORPTION CHARACTERISTICS VARIES ACROSS LAI
....................................................................................................................................... 81
FIGURE 2.15. THE OPTICAL SENSOR UNIT OF LAI-2000 PLANT CANOPY ANALYZER ............................................... 90
FIGURE 2.16. THE EQUILINEAR FISHEYE LENS PROJECTION ................................................................................. 93
FIGURE 3.1. THE LOCATION OF MAI PO NATURE RESERVE AND THE DISTRIBUTION OF MANGROVE STAND IN HONG KONG
..................................................................................................................................... 121
FIGURE 3.2. LAND USE CHANGE IN DEEP BAY, 1924-1985 ............................................................................ 122
FIGURE 3.3. THE METHODOLOGICAL OVERVIEW OF THE STUDY ......................................................................... 126
FIGURE 3.4. THE SAMPLED BANDS WITH DIFFERENT DEGREE OF STRIPPING ......................................................... 129
FIGURE 3.5. THE SCHEMATIC FLOW OF FLAASH ALGORITHM .......................................................................... 132
FIGURE 3.6. THE CENTER WAVELENGTH SHIFT OF SELECTED BANDS WITH RESPECT TO THE WAVELENGTH AT THE CENTER
PIXEL COLUMN AT 128 OF THE ACQUIRED HYPERION IMAGE ...................................................... 136
xviii
FIGURE 3.7. THE EIGENVALUES OF 158 COMPONENTS AFTER MNF TRANSFORMATION. THE EMBEDDED ZOOM PLOT
SHOWS EIGENVALUES FROM 30 COMPONENTS ONWARDS. ........................................................ 138
FIGURE 3.8. THE EFFECT OF MNF NOISE REMOVAL OF BAND 8 (428.4NM) ....................................................... 139
FIGURE 3.9. THE PROCESSING PROCEDURE OF ASAR-ENVISAT DATA .............................................................. 144
FIGURE 3.10. THE SPATIAL DISTRIBUTION OF SAMPLE POINTS OF DIFFERENT MANGROVE SPECIES ............................ 150
FIGURE 3.11. THE MAP SHOWING THE LOCATIONS OF LEAF SAMPLE COLLECTION ................................................. 153
FIGURE 3.12. COLLECTED LEAVES COOLED WITH ICE IN A THERMAL INSULATING BOX ............................................ 154
FIGURE 3.13. SPECTRAL STEPS AT THE WAVELENGTH OVERLAP REGIONS (I.E. 1000NM AND 1800NM) ................... 155
FIGURE 3.14. THE 100% LINE DURING WHITE REFLECTANCE CALIBRATION ......................................................... 156
FIGURE 3.15. THE LABORATORY SETTING FOR SPECTRA MEASUREMENT ............................................................. 157
FIGURE 3.16. THE ILLUSTRATIVE DIAGRAM SHOWING 25° PROJECTION CONE ..................................................... 158
FIGURE 3.17. THE SAMPLING PLATE DURING SPECTRA MEASUREMENT............................................................... 159
FIGURE 3.18. THE WINSCANOPY IMAGING SYSTEM.................................................................................... 162
FIGURE 3.19. THE LAI SAMPLING TRANSECTS AND LOCATIONS IN MAI PO ......................................................... 164
FIGURE 3.20. THE ILLUSTRATIVE DIAGRAM SHOWING THE IDEAL MEASUREMENT LOCATIONS WITHIN A SAMPLING
QUADRAT ........................................................................................................................ 165
FIGURE 3.21. FIELD PHOTOS DURING THE SURVEY ......................................................................................... 166
FIGURE 3.22. THE SAMPLED IMAGES ACQUIRED UNDER DIFFERENT ILLUMINATION CONDITIONS.............................. 167
FIGURE 3.23. THE HEMISPHERIC PHOTO OVERLAID WITH SUNTRACKS (RAINBOW IN THE MIDDLE OF IMAGE) SKY GRIDS
(YELLOW LINES) DIVIDED ACCORDING TO SELF-DEFINED AZIMUTHAL AND ZENITHAL DIVISIONS .......... 169
FIGURE 3.24. THE TRANSFORMATION FROM RAW TO BINARY IMAGE THROUGH INTERACTIVE THRESHOLD SELECTION .. 170
FIGURE 3.25. AN ILLUSTRATIVE DIAGRAM SHOWING THE LOG-AVERAGE METHOD ................................................ 174
FIGURE 3.26. THE PROCEDURAL FLOW OF FEATURE SELECTION IN FST3............................................................. 177
FIGURE 3.27. THE .TRN FILE FORMAT REQUIRED BY FST3 .............................................................................. 184
FIGURE 3.28. DATA SPLITTING SCHEME: OUTER AND INNER SPLIT FOR MODEL TRAINING, VALIDATION AND TESTING. .. 185
FIGURE 3.29. THE NODE, STREAM AND WORKFLOW IN VISUAL CLASSIFIER AND CLEMENTINE 12.0 ......................... 191
FIGURE 3.30. A SIMPLE BINARY-SPLIT DECISION TREE CLASSIFIER ...................................................................... 195
FIGURE 3.31. A TYPICAL THREE-LAYER MULTILAYER PERCEPTRON ARTIFICIAL NEURAL NETWORK ............................. 197
FIGURE 3.32. THE GRAPHICAL DESCRIPTION OF SUPPORT VECTOR MACHINE (SVM) FOR A HYPOTHETICAL TWO-CLASS CASE
..................................................................................................................................... 201
FIGURE 3.33. THE TWELVE SITES OF VISUAL INSPECTION FOR ACCURACY ASSESSMENT WITH DESCRIPTION OF EACH SITE
ADHERED. ....................................................................................................................... 205
FIGURE 3.34. THE WORKFLOW OF REGRESSION ANALYSIS ............................................................................... 213
FIGURE 4.1. THE CORRELATIONS OF BETWEEN ATMOSPHERICALLY CORRECTED IMAGES (BY ATCOR-2 AND FLAASH) AND
LABORATORY-MEASURED SPECTRUM FOR FULL RANGE AND REGIONS OF THE SPECTRUM FOR VARIOUS
xix
SPECIES INCLUDING (A) AEGICERAS CORNICULATUM; (B) ACANTHUS ILICIFOLIUS; (C) AVICENNIA MARINA;
(D) KANDELIA OBOVATA GROUP 1; (E) KANDELIA OBOVATA GROUP 2; AND (F) SONNERATIA SPP. .... 223
FIGURE 4.2. THE SCATTERPLOTS SHOWING FIT VALUE FROM SPECTRAL FEATURE FITTING ANALYSIS OF IMAGE PIXELS FOR (A)
AEGICERAS CORNICULATUM; (B) ACANTHUS ILICIFOLIUS; (C) AVICENNIA MARINA; (D) KANDELIA
OBOVATA GROUP 1; (E) KANDELIA OBOVATA GROUP 2; AND (F) SONNERATIA SPP. ....................... 225
FIGURE 4.3. THE PERCENTAGE OF CHANGE OF MEAN BACKSCATTER VALUE OF VARIOUS HOMOGENEOUS OBJECTS WITH
DIFFERENT COMBINATIONS OF FILTER AND WINDOW SIZE .......................................................... 229
FIGURE 4.4. THE DECREASE OF STANDARD DEVIATION OF VARIOUS HOMOGENEOUS OBJECTS WITH DIFFERENT
COMBINATIONS OF FILTER AND WINDOW SIZE ......................................................................... 229
FIGURE 4.5. THE AVERAGE TRANSFORMED DIVERGENCE OF SPECIES USING DIFFERENT COMBINATIONS OF FILTER AND
WINDOW SIZE .................................................................................................................. 230
FIGURE 4.6. THE MEAN REFLECTANCE AND BACKSCATTER SIGNATURE OF THE MANGROVE SPECTRAL CLASSES – (A) MEAN
SPECTRAL REFLECTANCE EXTRACTED FROM LABORATORY-MEASURED LEAVE SAMPLES; (B) MEAN SPECTRAL
REFLECTANCE EXTRACTED FROM HYPERION DATA; (C) MEAN RAW BACKSCATTER COEFFICIENT IN DB SCALE;
AND (D) MEAN FILTERED BACKSCATTER COEFFICIENT IN DB SCALE. .............................................. 232
FIGURE 4.7. SUMMARY OF THE WAVELENGTH AT WHICH DIFFERENT SPECIES PAIRS SHOW STATISTICALLY SIGNIFICANT
DIFFERENCES IN MEDIAN REFLECTANCE FOR (A) LEAVE SAMPLES (B) HYPERION. THE MEAN REFLECTANCE
CURVE OF A. CORNICULATUM IS OVERLAID AS A REFERENCE TO VEGETATION REFLECTANCE FEATURES. 236
FIGURE 4.8. SUMMARY OF MULTI-TEMPORAL RADAR DATA IN WHICH SPECIES PAIRS SHOW STATISTICALLY SIGNIFICANT
DIFFERENCES IN MEDIAN BACKSCATTER COEFFICIENT (1=RAW Σ IN LINEAR UNIT; 2=LEE FILTERED Σ IN
LINEAR UNIT; 3=RAW Σ IN DB UNIT; 4=LEE FILTERED Σ IN DB UNIT). ............................................ 237
FIGURE 4.9. FREQUENCY PLOT OF SIGNIFICANTLY DIFFERENT SPECIES PAIRS AFTER MANN-WHITNEY U-TEST AT
  0.01 FOR (A) LEAVE SAMPLES (B) HYPERION. THE MEAN REFLECTANCE CURVE OF A.
CORNICULATUM IS OVERLAID AS A TYPICAL VEGETATION CURVE. ................................................. 241
FIGURE 4.10. FREQUENCY PLOT OF SIGNIFICANTLY DIFFERENT SPECIES PAIRS AFTER MANN-WHITNEY U-TEST AT
  0.01 FOR MULTI-TEMPORAL RADAR BACKSCATTER DATA (1=RAW Σ IN LINEAR UNIT; 2=LEE
FILTERED Σ IN LINEAR UNIT; 3=RAW Σ IN DB UNIT; 4=LEE FILTERED Σ IN DB UNIT). THE TIME SERIES
ATTAINING THE HIGHEST FREQUENCY OF 17 ARE HIGHLIGHTED. .................................................. 242
FIGURE 4.11. THE WAVELENGTH AT WHICH INDIVIDUAL MANGROVE SPECIES HAVING STATISTICALLY SIGNIFICANT
DIFFERENT MEDIAN REFLECTANCE FROM ALL OTHER SPECIES UNDER MANN-WHITNEY U-TEST AT
  0.01 FOR (A) LEAVE SAMPLES (B) HYPERION. THE MEAN REFLECTANCE CURVE OF A.
CORNICULATUM IS OVERLAID AS A REFERENCE TO VEGETATION REFLECTANCE FEATURES. ................. 247
FIGURE 4.12. THE MULTI-TEMPORAL RADAR BACKSCATTER COEFFICIENT OF SPECIES WITH STATISTICALLY SIGNIFICANT
DIFFERENT MEDIAN REFLECTANCE TO ALL OTHER 6 MANGROVE SPECIES UNDER MANN-WHITNEY U-TEST
AT
  0.01 (1=RAW Σ IN LINEAR UNIT; 2=LEE FILTERED Σ IN LINEAR UNIT; 3=RAW Σ IN DB UNIT;
4=LEE FILTERED Σ IN DB UNIT). ............................................................................................ 248
FIGURE 4.13. FREQUENCY PLOTS OF SELECTED FEATURES USING SFS WRAPPED BY KNN IN DIFFERENT SUBSET SIZES .. 254
xx
FIGURE 4.14. FREQUENCY PLOTS OF SELECTED FEATURES USING SFS WRAPPED BY SVM IN DIFFERENT SUBSET SIZES .. 255
FIGURE 4.15. FREQUENCY PLOTS OF SELECTED FEATURES USING SFFS WRAPPED BY KNN IN DIFFERENT SUBSET SIZES 260
FIGURE 4.16. FREQUENCY PLOTS OF SELECTED FEATURES USING SFFS WRAPPED BY SVM IN DIFFERENT SUBSET SIZES 261
FIGURE 4.17. FREQUENCY PLOTS OF SELECTED FEATURES USING OS WRAPPED BY KNN IN DIFFERENT SUBSET SIZES ... 266
FIGURE 4.18. FREQUENCY PLOTS OF SELECTED FEATURES USING OS WRAPPED BY SVM IN DIFFERENT SUBSET SIZES ... 267
FIGURE 4.19. COMPARISON OF MEAN TRAINING ACCURACY OF DIFFERENT SEARCH ALGORITHMS ............................ 269
FIGURE 4.20. COMPARISON OF MEAN TESTING ACCURACY OF DIFFERENT SEARCH ALGORITHMS .............................. 269
FIGURE 4.21. CRITERION I. FREQUENCY COUNT OF FEATURES WITH SUBSET SIZE OF 5 AFTER REMOVAL OF FEATURE
SUBSETS HAVING ONE STANDARD DEVIATION BELOW MEAN TRAINING AND TESTING ACCURACY ........ 271
FIGURE 4.22. CRITERION II. FREQUENCY COUNT OF FEATURES WITH SUBSET SIZE OF 10 AFTER REMOVAL OF FEATURE
SUBSETS HAVING ONE STANDARD DEVIATION BELOW MEAN TRAINING AND TESTING ACCURACY ........ 272
FIGURE 4.23. CRITERION III. FREQUENCY COUNT OF FEATURES WITH SUBSET SIZE OF 15 AFTER REMOVAL OF FEATURE
SUBSETS HAVING ONE STANDARD DEVIATION BELOW MEAN TRAINING AND TESTING ACCURACY ........ 273
FIGURE 4.24. CRITERION IV. FREQUENCY COUNT OF FEATURES WITH SUBSET SIZE OF 20 AFTER REMOVAL OF FEATURE
SUBSETS HAVING ONE STANDARD DEVIATION BELOW MEAN TRAINING AND TESTING ACCURACY ........ 274
FIGURE 4.25. CRITERION V. FREQUENCY COUNT OF FEATURES WITH THE BEST TRAINING AND TESTING ACCURACY IN
DIFFERENT SUBSET SIZES UNDER DIFFERENT SEARCH ALGORITHM-CLASSIFIER COMBINATION ............. 275
FIGURE 4.26. CRITERION VI. FREQUENCY COUNT OF FEATURES WITH THE BEST TRAINING AND TESTING ACCURACY SIZE IN
DIFFERENT SUBSET SIZES REGARDLESS OF SEARCH ALGORITHM-CLASSIFIER .................................... 276
FIGURE 4.27. THE CRITERION SCORE COMPUTED FROM FREQUENCY COUNT OF FEATURES SATISFYING THE SIX CRITERIA278
FIGURE 4.28. CORRELATION MATRIX OF THE SUBSET OF 15 FEATURES ............................................................... 281
FIGURE 4.29. TRANSFORMED DIVERGENCE OF 7 MANGROVE SPECTRAL CLASSES UNDER INCREMENTAL INCREASE OF
SUBSET SIZE (FEATURE INPUT IN DESCENDING ORDER OF CRITERION SCORE) .................................. 286
FIGURE 4.30. TRANSFORMED DIVERGENCE OF 7 MANGROVE SPECTRAL CLASSES UNDER DIFFERENT FEATURE SUBSETS
SELECTED FROM THE FINAL FEATURE SET ................................................................................ 287
FIGURE 4.31. ML CLASSIFICATION RESULT USING FOUR FILTERED RADAR FEATURES (161, 162, 163 & 166) – SUBSET 1
..................................................................................................................................... 290
FIGURE 4.32. ML CLASSIFICATION RESULT USING FIVE NARROW SPECTRAL FEATURES (9, 23, 29, 72 & 94) – SUBSET 2
..................................................................................................................................... 290
FIGURE 4.33. ML CLASSIFICATION RESULT USING THE FIRST FIVE FEATURES WITH THE BEST CRITERION SCORE (9, 23, 161,
162 & 163) – SUBSET 3 .................................................................................................... 291
FIGURE 4.34. ML CLASSIFICATION RESULT USING ALL NINE FEATURES WITH THE BEST CRITERION SCORE (9, 23, 29, 72,
94, 161, 162, 163 & 166) – SUBSET 4............................................................................... 291
FIGURE 4.35. DT CLASSIFICATION RESULT USING FOUR FILTERED RADAR FEATURES (161, 162, 163 & 166) – SUBSET 1
..................................................................................................................................... 298
FIGURE 4.36. DT CLASSIFICATION RESULT USING FIVE NARROW SPECTRAL FEATURES (9, 23, 29, 72 & 94) – SUBSET 2
..................................................................................................................................... 298
xxi
FIGURE 4.37. DT CLASSIFICATION RESULT USING THE FIRST FIVE FEATURES WITH THE BEST CRITERION SCORE (9, 23, 161,
162 & 163) – SUBSET 3 .................................................................................................... 299
FIGURE 4.38. DT CLASSIFICATION RESULT USING ALL NINE FEATURES WITH THE BEST CRITERION SCORE (9, 23, 29, 72, 94,
161, 162, 163 & 166) – SUBSET 4 .................................................................................... 299
FIGURE 4.39. ANN CLASSIFICATION RESULT USING FOUR FILTERED RADAR FEATURES (161, 162, 163 & 166) – SUBSET 1
..................................................................................................................................... 306
FIGURE 4.40. ANN CLASSIFICATION RESULT USING FIVE NARROW SPECTRAL FEATURES (9, 23, 29, 72 & 94) – SUBSET 2
..................................................................................................................................... 306
FIGURE 4.41. ANN CLASSIFICATION RESULT USING THE FIRST FIVE FEATURES WITH THE BEST CRITERION SCORE (9, 23,
161, 162 & 163) – SUBSET 3 ............................................................................................ 307
FIGURE 4.42. ANN CLASSIFICATION RESULT USING ALL NINE FEATURES WITH THE BEST CRITERION SCORE (9, 23, 29, 72,
94, 161, 162, 163 & 166) – SUBSET 4............................................................................... 307
FIGURE 4.43. SVM CLASSIFICATION RESULT USING FOUR FILTERED RADAR FEATURES (161, 162, 163 & 166) – SUBSET 1
..................................................................................................................................... 314
FIGURE 4.44. SVM CLASSIFICATION RESULT USING FIVE NARROW SPECTRAL FEATURES (9, 23, 29, 72 & 94) – SUBSET 2
..................................................................................................................................... 314
FIGURE 4.45. SVM CLASSIFICATION RESULT USING THE FIRST FIVE FEATURES WITH THE BEST CRITERION SCORE (9, 23,
161, 162 & 163) – SUBSET 3 ............................................................................................ 315
FIGURE 4.46. SVM CLASSIFICATION RESULT USING ALL NINE FEATURES WITH THE BEST CRITERION SCORE (9, 23, 29, 72,
94, 161, 162, 163 & 166) – SUBSET 4............................................................................... 315
FIGURE 4.47. COMPARISON OF (A) CLASSIFICATION ACCURACY AND (B) KAPPA COEFFICIENT USING DIFFERENT
CLASSIFICATION ALGORITHMS UNDER DIFFERENT SUBSETS ......................................................... 324
FIGURE 4.48. COMPARISON OF SPECTRAL BANDS SELECTED BY THIS STUDY AND OTHER STUDIES ............................. 328
FIGURE 4.49. RELATIVE FEATURE IMPORTANCE IN DT AND ANN MODELS.......................................................... 344
FIGURE 4.50. COMPARISON OF ANN CLASSIFICATION RESULTS USING PURE SPECTRAL FEATURES AND COMBINATION OF
SPECTRAL AND RADAR FEATURES – A. ILICIFOLIUS G1 ............................................................... 349
FIGURE 4.51. COMPARISON OF ANN CLASSIFICATION RESULTS USING PURE SPECTRAL FEATURES AND COMBINATION OF
SPECTRAL AND RADAR FEATURES – A. ILICIFOLIUS G2 ............................................................... 349
FIGURE 4.52. COMPARISON OF ANN CLASSIFICATION RESULTS USING PURE SPECTRAL FEATURES AND COMBINATION OF
SPECTRAL AND RADAR FEATURES – A. CORNICULATUM AND SONNERATIA SPP. .............................. 350
FIGURE 5.1. FREQUENCY PLOTS OF FIELD MEASURED LAI COMPUTED USING DIFFERENT METHODS .......................... 354
2
FIGURE 5.2. COEFFICIENTS OF DETERMINATION (R ) BETWEEN 158 HYPERSPECTRAL BANDS AND LAI COMPUTED USING
DIFFERENT METHODS; SOLID AND DASHED LINES REPRESENT LAI COMPUTED USING LOG-AVERAGE AND
LINEAR METHOD RESPECTIVELY ............................................................................................ 358
FIGURE 5.3. THE NORMALITY Q-Q PLOT FOR THE INDEPENDENT (LAI) AND DEPENDENT (VIS) VARIABLES ................ 361
xxii
FIGURE 5.4. THE NORMALITY Q-Q PLOT FOR THE TWO RADAR PARAMETERS – CONTRAST AND DISSIMILARITY AFTER
LOGARITHMIC TRANSFORMATION ......................................................................................... 363
FIGURE 5.5. THE LINEAR REGRESSION MODEL BETWEEN FIELD-MEASURED LAI AND VIS ........................................ 372
FIGURE 5.6. SCATTERPLOT OF MEASURED AND PREDICTED LAI (LAI2000G) USING DIFFERENT VIS AFTER LINEAR
REGRESSION ANALYSIS........................................................................................................ 373
FIGURE 5.7 PREDICTED LAI(2000G) FROM REGRESSION MODEL USING NDVI AS PREDICTOR ................................ 375
FIGURE 5.8. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING RDVI AS PREDICTOR ............................... 375
FIGURE 5.9. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING SAVI AS PREDICTOR ................................ 376
FIGURE 5.10. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING MSAVI AS PREDICTOR .......................... 376
FIGURE 5.11. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING TVI AS PREDICTOR ................................ 377
FIGURE 5.12. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING MCARI 1 AS PREDICTOR........................ 377
FIGURE 5.13. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING MCARI 2 AS PREDICTOR........................ 378
FIGURE 5.14. THE LINEAR REGRESSION MODEL BETWEEN FIELD-MEASURED LAI AND RADAR PARAMETERS ............... 380
FIGURE 5.15. SCATTERPLOT OF MEASURED AND PREDICTED LAI (LAI2000G) COMBINING VIS AND ASM AFTER STEPWISE
REGRESSION MODEL .......................................................................................................... 386
FIGURE 5.16. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING RDVI AND ASM AS PREDICTORS ............. 388
FIGURE 5.17. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING SAVI AND ASM AS PREDICTORS.............. 388
FIGURE 5.18. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING TVI AND ASM AS PREDICTORS ................ 389
FIGURE 5.19. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING MCARI 1 AND ASM AS PREDICTORS ....... 389
FIGURE 5.20. PREDICTED LAI(2000G) FROM REGRESSION MODEL USING MCARI 2 AND ASM AS PREDICTORS ....... 390
FIGURE 5.21. COMPARISON OF RESULTS FOR LAI COMPUTED USING THREE DIFFERENT METHODS BASED ON SIMPLE LINEAR
TVI REGRESSION MODEL .................................................................................................... 392
FIGURE 5.22. LOCATION OF MANGROVE SPECIES CLASS WHERE LAI WERE EXTRACTED FOR COMPARISON ................. 394
FIGURE 5.23. MEAN LAI (LAI2000G) PLOT OF MANGROVE SPECIES CLASSES UNDER SIMPLE LINEAR VI MODELS ...... 395
FIGURE 5.24. MEAN LAI (LAI2000G) PLOT OF MANGROVE SPECIES CLASSES UNDER MULTILPLE VI-ASM MODELS ... 396
FIGURE 5.25. CHANGE OF MEAN LAI OF SPECIES AFTER ADDITION OF RADAR PARAMETER ASM IN DIFFERENT VI MODELS
..................................................................................................................................... 397
FIGURE 5.26. MEAN SPECTRAL REFLECTANCE OF LAI CLASSES GROUPED ACCORDING TO THE PREDICTED LAI (LAI2000G)
FROM SIMPLE VI REGRESSION MODEL ................................................................................... 400
FIGURE 5.27. MEAN SPECTRAL REFLECTANCE OF LAI CLASSES GROUPED ACCORDING TO THE PREDICTED LAI (LAI2000G)
FROM MULTIPLE REGRESSION MODEL .................................................................................... 400
FIGURE 5.28. MEAN VALUE OF VARIOUS VIS FOR FOUR LAI CLASSES GROUPED ACCORDING TO THE PREDICTED LAI
(LAI2000G) FROM SIMPLE VI REGRESSION MODEL (*TVI IS SCALED) ......................................... 402
FIGURE 5.29. MEAN VALUE OF VARIOUS VIS FOR FIVE LAI CLASSES GROUPED ACCORDING TO THE PREDICTED LAI
(LAI2000G) FROM MULTIPLE REGRESSION MODEL (*TVI IS SCALED) ......................................... 402
FIGURE 5.30. THE PROFILES CUT ACROSS THE CENTRAL PART OF THE MANGROVE AREA ......................................... 403
FIGURE 5.31. RELATIONSHIP BETWEEN LAI AND VIS ACROSS DEFINED PROFILES A AND B ..................................... 406
xxiii
FIGURE 5.32. BOXPLOTS OF RADAR PARAMETERS FOR FIVE LAI CLASSES GROUPED BASED ON THE PREDICTED LAI
(LAI2000G) FROM MULTIPLE REGRESSION MODEL ................................................................. 410
FIGURE 5.33. RELATIONSHIP BETWEEN LAI AND SELECTED RADAR PARAMETERS ACROSS DEFINED PROFILES A AND B . 413
FIGURE 5.34. LAI PROFILES EXTRACTED FROM TVI LINEAR AND MULTI-REGRESSION MODEL .................................. 416
xxiv
CHAPTER 1
INTRODUCTION
1.1. Background to the Study
1.1.1. Mangrove Mapping and Monitoring
The mangrove communities or, mangals are one of the most important types of
vegetation found in the intertidal zone of tropical and subtropical regions (Tam et
al., 1997). They are highly specialized plants forming one of the most productive
ecosystems in the world with above-ground biomass ranging from 5.4 to 18.4
kilogram/m2 ((Hutchings and Saenger, 1987). The high productivity makes them
habitats, nurseries and feeding grounds for marine and terrestrial wildlife
(Nagelkerken et al., 2008). Apart from being naturally significant, they are
recognized for its community and economic values to human beings. Besides, they
can stabilize the shoreline from erosion and minimize the impact of natural disaster
such as tsunami (Danielsen et al., 2005). Located at the subtropical region, Hong
Kong has nourished forty-four mangrove stands of varying sizes from 0.2 – 172
hectares occupying a total of 350 hectares. They are mostly distributed in the
eastern part of Hong Kong but with the largest stand located at the northwestern
part close to the border area. Compared with the tropical regions, Hong Kong has
relatively lower diversity of mangroves. The humid and hot climate supports 8
native species of mangrove including Acanthus ilicifolius, Aegiceras corniculatum,
Avicennia marina, Bruguiera gymnorrhiza, Excoecaria agallocha, Heritiera littoralis,
Kandelia obovata, and Lumnitzera racemosa and two exotic species Sonneratia
caseolaris and Sonneratia apetala.
Although mangroves are recognized as important habitats, they are globally
threatened. According to the thematic study by Food and Agriculture Organization
of United Nations, the total area of mangrove worldwide has been reduced from
18.8 million hectares in 1980 to 15.2 million hectares in 2005, which means a total
of 3.6 million hectares have been lost in the past 25 years. Although the rate of
annual loss has been reduced from 1.04% (187, 000 ha) in 1980 to 0.66% (102,000
1
ha) in period between 2000 and 2005, mangrove decline is still significant (FAO,
2007). Figure 1.1 shows the estimated worldwide loss of mangrove forest areas.
Asia has the most extensive area of mangrove and also experienced the largest loss
of mangrove area in the past 25 years.
Figure 1.1 Changes in world mangrove areas (1980 – 2005)
(Adopted from FAO, 2007)
Although natural phenomenon contributes to the natural loss of mangrove, human
activities have aggravated pressures and menaces to mangrove worldwide.
Conversion of mangroves to fishponds, reclamation of mangrove areas for
development, pollution, unsustainable exploitation and practices and other related
disturbance are all direct human threats to mangroves (Melana et al., 2000, Valiela
et al., 2001). Hong Kong has experienced the reclamation of mangrove for
agricultural land uses, but that happened in 1920s - 1950s. Nowadays, mangrove
was sacrificed for development due to high demand on land resources and
infrastructure. For instances, about 42% of mangrove in Tolo Habour were
2
sacrificed for construction of new highway, new town and race course development
(Holmes, 1988). In northern Lautau, about 50 hectares of mangrove were cleared
for new airport and associated facilities development (Maunsell, 1991). Apart from
development, effluents discharged from various sources such as domestic wastes,
organic agricultural wastes, inorganic and heavy metal pollution from industries and
runoff from road or highway has been intensified due to rapid economic expansion
in Pearl River Delta. Serious water pollution poses severe threats to wetland
ecosystem, which affects the habitats of marine life and birds. For instance, the
concentration of heavy metal measured in Shenzhen’s mangrove forest has
exceeded the fifth categories of seawater quality standard (“Mangrove”, 2006,
September 16). However, its impacts on the largest mangrove stand in Deep Bay are
still unknown and unpredictable.
1.1.2. Mangrove Mapping and Monitoring
For mangrove conservation to be effectively carried out, information such as
species distribution, leaf area index (LAI), canopy closure, tree height, density, and
other related biophysical parameters are important baselines to understand the
current status of mangrove stands. Based on the understanding to current situation,
appropriate conservation management plans can then be drafted and formulated to
protect the mangroves from further deterioration.
The species map provides an important inventory/ database to understand the
current distribution of mangrove and is useful to formulate conservation measures.
In traditional broad land use and land cover map, mangrove is identified as a single
class. A detail map describing the spatial distribution of mangrove species covering
the stands in Mai Po was not available only until 2008. However, it has not been
updated afterwards. As mangroves are extending towards Deep Bay every year, it is
essential to have an updated species inventory and to monitor the growth and
spread of mangroves. In recent years, the two exotic species have colonized rapidly
in the Mai Po stands. Their effects to the native species are still unknown. The
3
inventory update can locate and monitor the distribution of Sonneratia and proper
management measures can be taken if necessary.
Apart from species distribution, biophysical parameters are important indicators of
an ecosystem. Among the parameters, leaf area index is the commonly adopted
parameter measuring biophysical characteristics of vegetation. It is defined as the
total one-sided area of all leaves in the canopy within a defined region (Gong et al.,
2003). The popularity of using LAI as a quantitative measurement of an ecosystem
found on its significant influence on the energy and mass exchange between
terrestrial ecosystem and the atmosphere (Bonan, 1993). Besides, it can be used as
an input attribute to many ecosystem process models. Consequently, the success of
LAI estimation can model the ecological process and predict the response of
ecosystem (Green et al., 1997, Green and Clark, 2000).
1.1.3. Role of Remote Sensing in Mangrove Study
The acquisition of such information for the entire area is tedious and sometimes
infeasible without the aid of proper tools. Besides, limited by the environment and
accessibility to mangrove areas, intensive field surveys would involve significant
labor resource and cost. Moreover, mangroves are sensitive to disturbances and in
many cases; they are protected by local conservation policy. For instances,
mangrove stand in Mai Po has been listed under Ramsar Convention as wetlands of
international importance. The critical issue is how species and biophysical
information can be acquired efficiently and effectively for the area without causing
much disturbance to the mangrove ecosystem. With the capability to acquire land
surface characteristics in broad scale, remote sensing has long been a cost- and
time-effective tool for mangrove conservation study. With minimum field survey
efforts, essential information can be estimated from the remotely-sensed data
especially when field efforts in mangrove areas sometimes involve enormous labor
resources and costs. Besides, sun-synchronous satellites allow repetitive data
capture, which enables acquisition of temporal series of data for seasonal
comparison. Although the applications of remote sensing for mangrove mapping
4
and monitoring were well developed, they were essentially based on traditional
optical sensors such as SPOT and Landsat ETM+ (Gao, 1999, Sulong et al., 2002). In
Hong Kong, detail map of mangrove species distribution was produced based on
high resolution Quickbird image in 2006 (Fung et al., Under review).
With the advancement of sensor technology, satellite sensors can now capture high
spectral resolution and radar images that offer new prospects for mangrove studies.
Hyperspectral imaging is able to capture a number of narrow contiguous spectral
bands mostly ranging from 350 to 2500 nm. Compared with broadband optical
sensors, hyperspectral imaging offers much richer spectral details. The variation of
spectral reflectance across wavelengths is mainly the function of leaf optical
property and canopy characteristics. As species and maturity variations cause
differences in chemical concentration (that affect leaf optical property) and canopy
structure, their reflectance across the wavelength would be different. Based on
such difference, the narrow bands are comparatively more capable to improve the
accuracy of mangrove mapping and biophysical parameter estimation.
A major problem facing the subtropical region is the presence of clouds especially
during the rainy seasons in summer that always hinders the acquisition of quality
image by the optical sensors. Operating in a much longer wavelength (from 1mm to
1m), radar signals are mostly free from the impact of atmospheric conditions, i.e.
weather-independent (Pasqualini et al., 1999, Kushwaha et al., 2000, Filho and
Paradella, 2002, Miles et al., 2003). The most popular active-sensing radar
technology called synthetic aperture radar (SAR) offers an additional data
dimension for mangrove studies. Different from the spectral bands which react
essentially with the chemical properties of vegetative components (Held et al., 2003,
Le Toan et al., 2004), radar signal can penetrate the vegetation canopy and acquire
canopy structure (Patel et al., 2006) and dielectric properties (Baghdadi et al., 2001,
Lu and Meyer, 2002) beneath the canopy layer. Besides, backscatter returns from
the vegetative layers can be polarized and various modes of polarization are
sensitive to different vegetative structures. As mangrove species have different
canopy structure, their responses to radar signals tend to be varied. Such
information can be useful to distinguish species. Besides, as the intensity of radar
5
backscatter or attenuation is related to canopy structure such as leaf density,
branch density and ground surface roughness (dependent on the penetration
capability), it can also be used to estimate biophysical parameters. Moreover, the
weather-independent characteristics allow continuous data acquisition which
makes multi-temporal comparison more feasible.
Although remote sensing has been extensively applied to many vegetation studies,
some major aspects are still not well-investigated.

Discriminating mangrove from other vegetation types are well researched,
but study on species-level mangrove classification is relatively limited;

Estimation of LAI for different vegetation types such as forest ecosystem and
agricultural crops were reported, however, very limited researches focus on
LAI estimation in the mangrove ecosystem;

Vegetation indices are mostly derived from broadband satellite sensors for
LAI and biomass estimation, only a few used satellite hyperspectral imaging
for LAI and biophysical parameter mapping; and

Radar remote sensing to study species distribution and biophysical
parameters estimation is still under-explored.
1.2. Objectives of the Study
It is worthwhile to explore the potential and assess the feasibility in using the
hyperspectral or radar data alone or in combining the two types of data in
improving mangrove classification and biophysical parameter estimation especially
when extensive ground measurements are constrained by accessibility and other
environmental factors. The study takes advantage of hyperspectral and multitemporal SAR data coupled with in situ field survey to achieve the following
objectives:
6
I. To select spectral and radar features that are important in discriminating
mangrove species;
II. To map the species distribution of mangrove;
III. To compare classification accuracy under different combinations of spectral
and SAR features;
IV. To compare the performance of various classification algorithms;
V. To estimate the leaf area index of mangrove;
VI. To explore the relationship between vegetation indices derived from
narrowband of hyperspectral data and field-measured LAI;
VII. To explore the relationship between radar backscatter and its derivatives
and field-measured LAI; and
VIII. To explore the compatibility of optical and radar in species mapping and
monitoring the conditions of mangrove stands based on LAI retrieval.
1.3. Significance of the Study
There are five purposes for this study. First, through species mapping, it establishes
a spatially-continuous inventory or database for species-level distribution of
mangrove. Focus research on species-level mangrove mapping is relatively limited.
Practically, the results provide an important source of information for species
inventory update.
Second, through feature selection of hyperspectral data, a set of spectral bands and
radar features are identified to effectively differentiate mangrove species. As
spectral bands react primarily with the chemical properties of vegetative
components, selected spectral regions reflect indirectly the main constituents that
cause species difference. Canopy structural and moisture content differences are
reflected in the selected radar features.
7
Third, as radar is weather independent, it is an ideal monitoring tool as data can be
consistently and regularly acquired. If relationship between radar backscattering
and biophysical parameters can be established successfully, it will become a fast
and reliable tool for long-term monitoring of status of mangrove habitats.
Fourth, optical remote sensing has been extensively applied in vegetation mapping
and biophysical parameter estimation. Numerous significant findings have been
obtained relating vegetation and spectral reflectance. Comparatively, application of
radar remote sensing on vegetation studies, especially on mangrove study is not as
widespread as the optical one. As the interpretation of radar signal from vegetation
is not straightforward, many uncertainties still exist. Findings from this study can be
contributory to existing researches to understand the relationship between radar
backscatter and mangrove.
Fifth, using optical remote sensing as benchmarking, this study explore the
complementarity of microwave remote sensing to the optical one for species
classification and leaf area index (LAI) estimation in terms of accuracy improvement.
The roles of optical and microwave data is identified.
1.4. Organization of the Thesis
This thesis consists of six chapters. Chapter one outlines the research rationale,
objectives, significance and contribution of the research.
Chapter two provides a comprehensive review of literatures on a few aspects by
firstly identifying factors affecting spectral reflectance of vegetation in both micro
and macro-scale. This is followed by discussion on how hyperspectral and radar
data can contribute to vegetation classification. After that, problem of high
dimensionality is pinpointed. Statistical pattern recognition and feature selection
strategies in reducing data dimensionality are discussed. Finally, biophysical
parameter estimation with emphasis on Leaf Area Index (LAI) is reviewed. The
discussion begins by focusing on in situ field measurement techniques with
emphasis on indirect gap fraction inversion method. LAI estimation through
8
empirical model and physically-based canopy reflectance model inversion based on
remotely-sensed data are then discussed. In the review, special attention will be
given to studies related particularly to mangrove.
Chapter three describes the methodological framework of the study. The chapter
begins with a description of the study area and provides an overview of the
methodological flow of the study. This is followed by identifying the types of
remotely-sensed data acquired and their related processing procedures. The
followed session focuses on field data collection and measurement. Data analyses
and the corresponding workflows related to statistical feature selection, speciesbased classification and LAI modeling are then revealed.
Chapter four provides results and discussion on feature selection and species-based
classification. The chapter firstly shows the data processing and exploratory results.
This is followed by revealing the results of feature selection under various search
algorithms. After that, mangrove species-based classification results using various
classifiers are presented. Based on the results, implications are drawn and discussed
in the last session of this chapter.
Chapter five analyzes and discusses the outcome of LAI modeling. The chapter
begins with data exploration of the dependent and independent variables. After
that, LAI predictions under different models are presented. The chapter ends by
drawing implications from the results and discusses the complementarity of spectral
and radar data.
This thesis ends with a concluding chapter, Chapter 6 in which the findings of the
study is summarized. A discussion on limitations of the study and recommendations
are provided.
9
CHAPTER 2
LITERATURE REVIEW
2.1. Introduction
Remote sensing has been widely used for vegetation-related applications. The rapid
development of sensor technologies has opened up new opportunities for efficient
and
accurate
vegetation
classification
and
monitoring.
Specifically,
the
hyperspectral and microwave satellite imaging reveal great research potential in the
field as they provide data that are relatively different from the traditional
broadband sensors. Hyperspectral imaging captures a large number of narrow
contiguous spectral bands covering a wide range of spectrum (Schmidt and
Skidmore, 2003) while microwave can reveal beneath-canopy information
(Bourgeau-Chavez et al., 2001). Not only will the combination of data sources
provides more description of the scene, it also presents an opportunity to realize
the inherent ecological process in the vegetation. However, the high dimensional
data pose opportunity as well as challenges. Bands/ features which influence the
vegetation discrimination process should be identified. This is especially prominent
when sample observations are hard to gather and indisputably, it is always the
problem faced for many vegetation studies.
Ecological monitoring of vegetative habitat is another major application. Among the
biophysical parameters, Leaf area index (LAI) is one of the popular indicators of
vegetation biomass. The estimation of LAI using remote sensing techniques requires
a priori field measurement in order to understand the existing condition. Based on
the field-measured samples/ observations, remotely-sensed data or their
derivatives can be used to predict LAI over the area of interest.
This chapter provides a comprehensive review of literatures on a few aspects by
firstly understand factors that affect vegetation reflectance in optical sense. In
micro-level, foliar structure and constitutes response differently across the spectral
ranges. In a much large scale, factors such as canopy architecture, background
reflectance, atmosphere and sun-sensor geometry have influence on signal received
10
by the sensor. This is followed by discussion on how hyperspectral and radar data
can contribute to vegetation classification. After that, problem of high
dimensionality is pinpointed and statistical pattern recognition that can solve the
dimensionality problem and the feature selection strategies are discussed. Finally,
biophysical parameter estimation with emphasis on Leaf Area Index (LAI) is revealed.
The discussion begins by focusing on in-situ field measurement. It is then followed
by highlighting two major methods, i.e. empirical model and physically-based
canopy reflectance model inversion for estimation of LAI using remotely-sensed
data. Since mangrove is the vegetation of interest for this study, special attention
will be on reviewing studies that are related particularly to mangrove.
2.2. Factors Affecting Vegetation Reflectance
The spectral property of vegetation is closely related to the foliar constituents,
organization of vegetative components and background environment. Spectral or
optical properties refer to ‘the material properties which specify the response of
the material to sinusoidal component waves at every frequency of wavelength
(Kumar et al., 2001 p.113)’. The radiance recorded by remote sensor is the
combination of atmospheric effect and spectral properties of surface. The unique
reflectance spectrum of vegetation is determined by factors including biochemical
composition, leaf structure and liquid water content of leaves in micro-scale (Curran
et al., 1992, Fourty et al., 1996). Biochemical composition refers to the pigment and
chemical compounds while leaf structure refers to the arrangement of cells, liquid
and air inside a leaf. Although the general shape of vegetation spectra is alike, the
variability of these factors owing to species difference or health status will affect
the response of absorption and reflectance in different wavelengths (Asner, 1998,
Martin et al., 1998, Schmidt and Skidmore, 2003). At landscape scale, reflectance is
also affected by canopy structure and soil background reflectance (Schmidt and
Skidmore, 2003). In order to understand the spectral properties of vegetation, the
fundamental factors affecting vegetation reflectance will be examined at foliar and
11
canopy level. The relationship between foliar tissues and optical properties in
different wavelength regions will be discussed in this section.
2.2.1. Foliar structure and principal constituents
According to Verdebout et al. (1994), the structure of a leaf can be coarsely
separated into three layers including epidermis, mesophyll and central vascular
cylinder (xylem and phloem). Figure 2.1 shows the internal structure of a leaf. The
epidermis is the first and outmost layer with which the solar radiation interacts and
both diffuse and specular reflectance takes place.
Figure 2.1. A typical foliar structure
(Adopted from Missouri Department of Conservation, USA)
As light penetrates into the leaf, it encounters two kinds of photosynthetic cells –
palisade parenchyma and spongy parenchyma in the mesophyll layer. For most
leaves, the palisade parenchyma cells are elongated in shape and located in the
upper part of the mesophyll while the spongy parenchyma cells are irregular in
shape, loosely organized and occupied in the lower part of the mesophyll (Jensen,
12
2007). There are different kinds of pigments in the palisade parenchyma layer,
which control primarily the absorption of light as a function of wavelength. Within
the foliar structure and mainly in the mesophyll, the principal biochemical
constituents include pigments such as chlorophylls a and b, carotenoids,
xanthophylls, polyphenols and brown pigments and chemical compounds including
water, cellulose, lignin and starch. These molecules show characteristic absorption
and reflection of radiation as a function of wavelength (Verdebout et al., 1994,
Jensen, 2007). The absorption of radiation by leaf is due to two processes –
electronic transitions and vibration and rotation of polyatomic molecules. The
former process is related to the foliar pigments, which affect the visible spectral
domain while the latter process is concerned with water which influences the near
and middle infrared region of the light spectrum (Guyot et al., 1992, Verdebout et
al., 1994). Among those, chlorophyll a and liquid water are the key features
affecting the spectral reflectance of most vegetation. Chlorophyll, found in the
chloroplasts of green leaves in the mesophyll is one of the most prominent
biochemical compounds for photosynthesis. As for liquid water, it accounts for
about 40% to 90% of leaf fresh weight (Verdebout et al., 1994). The two features
tend to mask the subtle spectral signature related to other biochemical constituents
of vegetation (Vane and Goetz, 1993).
The spongy mesophyll layer is the combination of many cells and intercellular air
spaces, within which gaseous exchanges (carbon dioxide and oxygen) for
photosynthesis and respiration to take place (Jensen, 2007). This layer controls the
scattering and diffuse character of reflectance of a leaf (Guyot et al., 1992). As for
the central vascular system, they are mainly responsible for water (by xylem) and
sugar (by phloem) transportation within the leaf (Verdebout et al., 1994). Since the
basic elements within a leaf are the same, it is the assembly or organization of these
elements leads to leaf optical properties differences in different leaves (Verdebout
et al., 1994).
13
2.2.2. Foliar optical properties
Many studies have explored the relationship of main foliar constituents/ structure
with spectral absorption, transmittance and reflectance with a view to understand
the optical properties of foliar and vegetation. With the capability of measurement
of wider spectral range by both handheld and onboard spectrometer, wavelength
ranging from 400 to 2500 nanometers (nm) becomes the prominent interest. A
typical spectral signature of vegetation is shown in Figure 2.2. With the distinctive
optical response of vegetation over the spectrum, spectral domain has been broadly
divided into four regions of interest including the visible region (400-700 nm), the
red edge (690-740 nm), the near-infrared region (700-1300 nm) and middle-infrared
region (1300-2500 nm). The foliar molecules and internal structure works together
to shape the spectral character in these regions.
Leaf
pigments
Cell
structure
Water
content
Water
absorption
Major
absorption
bands
Chlorophyll
absorption
Visible
Primary
factors
governing
leaf
reflectance
Near-Infrared
Shortwave Infrared
Figure 2.2. A typical spectral signature of vegetation in 350 – 2500nm
2.2.2.1. The visible region (400-700nm)
The visible region is portrayed by low reflectance and transmittance resultant from
the strong absorption by pigments such as chlorophyll and carotenoids for
photosynthetic purpose (Clevers, 1988, Wessman, 1994a, Kumar et al., 2001). For
14
vegetation studies, this region is also termed as photosynthetically active radiation
(PAR). Chlorophyll is the dominant pigment within the mesophyll in a green leaf
responsible for photosynthesis. Verdebout et al. (1994) pointed out that since the
concentration of chlorophylls a and b are on average ten times more than
carotenoids, the absorption effect of carotenoids is masked by that of chlorophylls.
Chlorophyll pigments absorb violet-blue and red light for photosynthesis (Kumar et
al., 2001) which results in a main deepening and major widening of two chlorophyll
absorption features (Curran, 1994). The absorption peaks locate at 430/660 nm and
455/640 nm for chlorophyll a and b respectively (Verdebout et al., 1994, Wessman,
1994a, Kumar et al., 2001). Since green light is not absorbed, a reflectance
maximum occurs around 550nm, which is called green peak and that makes most
vegetation appears green.
The change in chlorophyll content of green leaves is reflected from the change in
the magnitude of spectral absorption in this region. During senescence, as
chlorophyll diminishes, the red reflectance increases because the absorption in
chlorophyll decreases (Jensen, 2007). Besides, chlorophyll degraded faster than
carotenoids, and carotenoids become the main pigment, which appear yellow
(Boyer et al., 1988). These two processes give the senesced leaf a yellow colour.
Lichtenthaler et al. (1996) identified regions including wavelength between 530 nm
and 630 nm and around 700 nm are sensitive to the variations in chlorophyll
content. Generally, the increase in leaf reflectance in the visible spectrum is related
to plant stress due to seasonal senescence or environmental stress, which can be
monitor effectively through high spectral resolution imaging spectrometer (Jensen,
2007).
2.2.2.2. The red edge (690-740nm)
The red edge is the visible-near-infrared transition region characterized by the
strong chlorophyll absorption in red wavelength followed by an abrupt swell in near
infrared reflectance at 700-740 nm due to leaf internal scattering (Horler et al.,
1983, Dawson and Curran, 1998, Broge and Leblanc, 2001, Kumar et al., 2001,
15
Mutanga and Skidmore, 2004). The red edge is unique to vegetations because they
contain chlorophylls that is absent in other materials such as soils, rocks, plant litter
(Vane and Goetz, 1993, Elvidge and Chen, 1995). The region shows strong
reflectance and transmittance, which means light rays can penetrate deeper inside
leaf tissue and thus provide more information on leaf structure (Verdebout et al.,
1994). Compared with other spectral regions, the red edge is the most studied
reflectance region because it contains rich information and shows high sensitivity
and correlation with vegetation biophysical and biochemical parameters such as
chlorophyll content (Horler et al., 1983, Demetriades-Shah et al., 1990, Curran et al.,
1991, Filella and Penuelas, 1994, Wessman, 1994a, Lichtenthaler et al., 1996, Kumar
et al., 2001); indicator of vigor, stress and senescence (Boochs et al., 1990,
Vogelmann et al., 1993, Dawson and Curran, 1998, Treitz and Howarth, 1999); and
leaf area index (Kumar et al., 2001). Through laboratory experiments, Horler et al.
(1983) and Curran et al. (1991) have shown positive relationship between the rededge inflection point and the chlorophyll concentration of leaf samples.
Environmental stresses such as nutrient deficiency, soil salinity, heavy metal stress
or exposure to atmospheric pollutants lower the chlorophyll concentration, which
in turn narrow the absorption band in red region (Wessman, 1994a). Therefore, the
spectral shift of the red edge to the shorter wavelengths indicates stress on
vegetation (Vogelmann et al., 1993, Wessman, 1994a, Treitz and Howarth, 1999).
However, it does not specify any particular types of environmental stress
(Verdebout et al., 1994). Besides, Demetriades-Shah et al. (1990) argued that rededge may be good in estimating chlorophyll content in foliar level, but not at
canopy level.
2.2.2.3. The near-infrared region (700-1300nm)
Vegetation is characterized by high reflectance and transmittance and low
absorption in the near infrared region. The leaf pigments such as chlorophyll show
negligible absorption in this region because the energy levels in the near infrared
region is not strong enough (Kumar et al., 2001). Besides, the high reflectance and
16
transmittance is the self protection mechanism of vegetation in order to get rid of
excessive heat energy which will irreversibly denature the proteins (Jensen, 2007).
The spectral response of this region is dominantly controlled by the cellular
structure within the leaf (Wessman, 1994a, Kumar et al., 2001). To be specific, it is
the spongy mesophyll layer which is the composition of many cells and intercellular
air spaces controls the intensity of reflectance in the NIR region (Jensen, 2007).
Studies showed that factors such as the distribution of air spaces, size, shape and
arrangement of cells within a leaf have an effect on light passage, which in turn
influences the light diffusion and scattering character (Kumar et al., 2001). Early
studies showed that the volume of intercellular air spaces in the spongy mesophyll
layer of leaves affects the near-infrared reflectance (Gausman et al., 1970).
However, (Buschmann and Nagel, 1993) found that it is the total surface area of airwall interfaces that matters. On the near-infrared plateau, there are two minor
water absorptions located around 960 and 1200 nm, which is mainly caused by the
cellular arrangement and hydration state within a leaf (Wessman, 1994a, De Jong
and Epema, 2001). A healthy green leaf demonstrates a high reflectance in this
region (Jensen, 2007).
2.2.2.4. The short-wave infrared region (1300-2500nm)
Compared with the near infrared region, the reflectance is much lower in the midinfrared region. Absorption of foliar biochemicals including lignin, cellulose, starch,
proteins, nitrogen and water characterize this region (Kumar et al., 2001). Since
water is a good absorber of mid-infrared energy (Jensen, 2007), leaf water
absorption is the dominant absorption features and determines the overall shape of
the mid-infrared spectrum. The major water absorption bands centered at 1450,
1940 and 2500 nm (Verdebout et al., 1994, De Jong and Epema, 2001) and they
tend to masks the minor absorption features produced by other biochemicals
(Elvidge, 1990, Verdebout et al., 1994, Kumar et al., 2001). After leaves are dried,
the reflectance and transmittance increase due to the reduction in absorption by
water and increase in air-wall interfaces (Verdebout et al., 1994, Jensen, 2007).
17
Besides, when leaves are dehydrated, other biochemical absorption features
become more observable and distinguishable (Kumar et al., 2001). Therefore, the
mid-infrared region combined with the near-infrared bands is a good indicator of
vegetation water status and examination of water stress (Wessman, 1994a).
2.2.3. Canopy architecture
One of the merits of remote sensing is its capability to acquire spectral information
of different land covers over a large area in landscape or global scale. The
interaction of electromagnetic radiation with vegetation canopies is very complex
process (Schlerf and Atzberger, 2006). Apart from the intrinsic optical and structural
properties, the radiation regime of vegetation canopies is complicated by a number
of spatially- and temporally- dependent factors related to canopy architecture,
background reflectance, atmospheric effect and sun-sensor relationship (Gao et al.,
2000, Kumar et al., 2001, Boegh et al., 2002).
A canopy is a composition of different vegetation elements including leaves,
branches, stems, stalks, flowers, bark, etc. Canopy architecture, structure or
geometry refers to the spatial distribution, density, and orientation of the
vegetation elements within the vegetated areas (Goel, 1988). Canopy structural
variables like leaf area index (LAI) and leaf inclination angle distribution (LAD) are
used to describe the density and orientation of leaves within a canopy (Baret and
Jacquemoud, 1994, De Jong and Epema, 2001). Baret and Jacquemoud (1994)
tested the sensitivity of canopy reflectance to various canopy parameters and found
that LAI is one of the primary factors affecting canopy spectral reflectance over the
whole spectrum, primarily in the near-infrared region while LAD has similar effects.
Characterized by the strong absorption in the visible region, LAI increase results in
decrease in reflectance. The reflectance approaches a value of zero when the LAI is
at about 2 – 3 (Goel, 1988). With low absorption in near infrared domain, the
topmost layer of vegetation canopy reflects about 40 – 60% of incident near
infrared radiation and transmitted the remaining (45 – 50%) to the lower layers. The
lower layer reflects and transmits the penetrated infrared energy once again
18
(Jensen, 2007). The additive reflectance effect of canopy layers increase the level of
infrared reflectance when compared with single leaves (Belward, 1991, Jensen,
2007). The contribution of near infrared reflectance reduces with increasing depth
in the canopy and is insignificant from LAI 6 – 8 onwards (Clevers, 1988). LAD works
together with the sun-sensor geometry to determine its effects on reflectance. The
LAD influences canopy reflectance through changing the gap fraction as a function
of the solar-viewing angle. It determines whether the solar incident will be
intercepted by the vegetation (Goel, 1988). (Kupiec and Curran, 1995) found that
the influences of canopy parameters on leaf reflectance rise dramatically at
wavelengths 1400 nm and beyond.
2.2.4. Background reflectance
Background reflectance is regarded as confounding spectral signal contributions
from undergrowths, and non-photosynthetically-active components such as litters
or senescent leaves and soil (Kumar et al., 2001, Boegh et al., 2002). The amount of
background reflectance is determined by canopy closure which is defined as ‘the
sum of the vertically projected crown envelopes which include within-crown canopy
gaps, divided by ground surface area (Rautiainen, 2005) p.298’, which in turn has
close relationship with LAI, i.e. the lower the value LAI, the more prominent the
effect of background reflectance will be, vice versa (Huemmrich and Goward, 1997,
Zarco-Tejada et al., 2001, Eriksson et al., 2006). The contribution of background
signal tend to undermine the ability for further analyses when they mixed with the
canopy signal (Gao et al., 2000). With a view to come up with a more reliable and
consistent results, efforts have been placed on investigation of the different kinds of
background contamination such as soil (Huemmrich and Goward, 1997, Gao et al.,
2000), litter, and understory (Spanner et al., 1990, Eriksson et al., 2006). Variability
of soil reflectance affects mainly the reflectance in the near infrared and shortwave
infrared domain while the effects on visible region is relatively small (Asner et al.,
2000). However, it will generally lower the canopy reflectance because the multiple
scattering in the soil layers induce more absorption (Goel, 1988). Mangrove grows
19
along the coastal areas and the soil is affected by periodic wet-and-dry process due
to tidal effect (Blasco et al., 1998). The wet soil has an overall lower reflectance
than dry soil. When soil reflectance is negligible, canopy reflectance data is strongly
correlated with LAI and Leaf Angle Distribution (LAD) in dense canopies (Asner,
1998). Eriksson et al. (2006) examined the effect of understory vegetation on
canopy reflectance and estimation of LAI. They reported that the compensation of
the effect of understory vegetation yields better estimation of canopy LAI.
2.2.5. Atmospheric perturbation
The main source of energy for imaging spectrometry is coming from the Sun. When
the solar photons cross the atmosphere of the Earth and reach the earth surface,
they interact with the surface and travel through the atmosphere again. Finally,
they reach and record by the sensor through deposition of certain amount energy
on the detectors (Verstraete, 1994b). This Sun- surface-sensor radiation trajectory
modifies the original surface information and renders the contamination of satellite
measured data (Rahman, 2001). During the travel of solar photons through the
atmosphere, the atmospheric constituents will absorb and scatter/ diffuse the
radiation before and after their interaction with the surface before they reach the
sensor (Lillesand et al., 2004, Jensen, 2007). The gases and particles such as oxygen,
nitrogen, carbon dioxide, water vapor and aerosols determine the scattering
properties of the atmosphere as well as the position and width of the absorption
regions in the spectrum (Verstraete, 1994b). Atmospheric scattering is a critical
issue in remote sensing study. The turbidity of atmosphere because of high
concentration of dust or water vapor particles lessens the information content of
remotely sensed data and image contrast (Jensen, 2007), which in turn render
unreliable measurements of vegetation (Gao et al., 2000). Therefore, the
atmospheric conditions govern the fraction of scattering radiation and render
different effects across wavelength. The effect is more influential in the visible
region than in the near infrared region when dust and water vapor particles are
considered (Goel, 1988). The scattering by molecules tends to have effect up to
20
750nm, which can be fully explained by the Rayleigh scattering theorem while the
aerosol scattering effect poses significant impact up to the SWIR region (1300nm)
(Griffin and Burke, 2003).
The atmospheric absorption determines the spectral intervals in which the solar
photons can pass through and the surface can effectively be surveyed. Figure 2.3
shows the absorption regions of different atmospheric constituents as a function of
wavelength adopted from Jensen (2007). The primary atmospheric water
absorption bands locate at 970, 1190, 1450, 1940 and 2700 nm provide limited
spectral information (Jensen, 2007).
The atmospheric perturbation can be largely mitigated through atmospheric
correction algorithms using radiative transfer codes such as ATREM, CAM5S,
MODTRAN (Gao et al., 2000, Rahman, 2001). These codes are used to remove
atmospheric scattering and gas particle effects under different solar illumination
and sensor viewing conditions and to retrieve correct reflectance of surface
materials (Staenz et al., 2002). Rahman (2001) studied the atmospheric effects,
mainly atmospheric aerosol and water vapor, on accuracy of crop biophysical
parameters retrieval. The results showed that atmospheric aerosol content has
relatively significant impacts on biophysical parameters retrieval than water vapor.
21
Figure 2.3. The absorption regions of different atmospheric constituents
(Adopted from Jensen, 2007)
2.2.6. Sun-sensor relationship
The sun-sensor relationship is usually described by Bidirectional Reflectance
Distribution Function (BRDF), which suggested the reflectance of a target is a
function of sun illumination geometry and sensor viewing geometry. Since the
scattering of radiation by targets at Earth’s surface is non-Lambertian or anisotropic
(Goel, 1988, Epiphanio and Huete, 1995), the magnitude of reflectance taken from
the same vegetation canopy is different because of variations in illuminationviewing geometry. The non-Lambertian property of vegetation canopies influenced
canopy
bi-directional
reflectance
function
through
specular
reflection,
comparatively strong backscattering and hot spot effect (Goel and Qin, 1994). Hot
spot effect occurs when canopy reflectance reaches the peak when the sun is
directly located behind the sensor and the shadowing is equal to zero (Goel, 1988,
22
Roberts, 2001). Generally, the increase of view zenith angle has the tendency of
raising reflectance from both VIS and NIR region for most of the azimuth angle (Goel,
1988). Specifically, the BRDF effect is more pronounced in high spectral absorption
wavelength ranges such as blue and red and less apparent in high reflectance
regions such as green and NIR bands (Jensen, 2007). Apart from sun-surface-sensor
geometry, canopy structural orientation and arrangement such as leaf area index,
leaf angle distribution and clumping which in turn affects the exposure of soil and
undergrowth to sensor also affects the bi-directional reflectance from a vegetation
canopy (Roberts, 2001). There are evidences showing that BRDF is strongly
dependent on LAD to determine the variability of reflectance. For examples,
canopies with more horizontally tilted leaves tend to have least influences by the
sun-view geometry and have the highest reflectance among different kinds of
inclinations (Kimes, 1984). Besides, the intensity ‘hot spot’ effect is also affects by
LAD (Goel, 1988).
Since the bi-directional effects have comparatively signification impacts on
atmospheric perturbations (Roberts, 2001), correction of this effect through
geometric normalization of hyperspectral images to common geometry is
sometimes needed to attain more reliable reflectance data for different purposes
(Chen et al., 1999). BRDF models based on radiative transfer theory are always
applied prior to any analyses in order to minimize this effect (Gao et al., 2000). Chen
et al. (1999) compared the results of LAI retrieval from CASI reflectance data before
and after BRDF correction and reported that the sensitivity of LAI is higher after
correction.
2.3. Hyperspectral Imaging and Vegetation Classification
The development of hyperspectral sensor enables the capture of narrow and
contiguous bands across the spectrum. Hyperspectral imaging, or its synonyms
imagining spectrometry, imaging spectroscopy is defined as the transformation of
record of energy deposited by photons in many narrow contiguous spectral bands
(covering the visible and the infrared region) by photographic or electronic detector
23
into analog or digital count by converter (Verstraete, 1994b). Curran (1994)
identified four major differences between broadband sensors and narrowband
hyperspectral sensor including spectral coverage, number of bands available to
users, spectral resolution and dynamic range of the signal. The hyperspectral
sensors perform better than the broadband sensors in the four areas, i.e. wider
spectral coverage (include mid-infrared region); more wavebands available to users
(like airborne AVIRIS with 224 bands and space-borne Hyperion sensor with 242
bands); higher spectral resolution (between 1nm to 10nm); higher dynamic range of
the signal (8-bit to 16-bit with the majority of 12-bit). Such differences on one hand
demonstrated the superiority of hyperspectral data and on the other hand revealed
implicitly the insufficiency of broadband sensors in many applied studies, especially
vegetation studies.
Compared with broadband remote sensing, imaging spectrometry has brought a
number of benefits towards analyses and applications. Since imaging spectrometry
enable measurement of radiation in many narrow wavebands, the full spectral
shape, narrow and contiguous waveband characteristics allow the identification of
spectral features of interest such as the absorption wells, reflectance peaks in
different materials (Schmidt and Skidmore, 2003). The dramatic variation of
reflectance with wavelength of vegetation can definitely benefit from the sufficient
spectral details of hyperspectral data for analysis, classification and monitoring (Gao,
1999). Many past researches have identified the advantages and potential of
imaging spectrometry for vegetation studies. First and foremost, it allows much
better discrimination between vegetations in terms of health status (De Jong and
Epema, 2001) and species (Kumar et al., 2001). Besides, it separates vegetation
from soil background effectively (De Jong and Epema, 2001). Moreover, the spectral
features and shapes are good estimators of biophysical and biochemical variables
and processes, which are important for vegetation condition assessment (Baret and
Jacquemoud, 1994, Curran, 1994, Wessman, 1994b, Baret et al., 1995, Asner et al.,
2000). The subtle change of reflectance with wavelength indicates the possible
change in physical or chemical variables of vegetations (Curran, 1994). Wessman
(1994b) argued that the spectral shapes of absorption features can provide more
24
information about canopy physiology and biochemistry than the broadband data.
Baret and Jacquemoud, (1994) suggested that the high spectral resolution data can
provide information on leaf and soil biochemical composition, which are extracted
from specific absorption features.
Some studies compared the accuracy of both broadband and narrowband data in
mapping different vegetation species (Gao, 1999, Belluco et al., 2006). Some studies
argued that the hyperspectral data are not always superior to broadband data.
Green et al. (1998a) showed that the imaging spectrometer (CASI) is better in
discriminating different mangrove categories than Landsat TM and SPOT XS. Gao
(1999) compared the merits of spatial and spectral resolution in vegetation
mapping and found that spectral resolution is more important than fine spatial
resolution in mangrove mapping. Belluco et al. (2006) compared the accuracy of
several space-borne multispectral and airborne hyperspectral sensor data in
classifying inter-tidal halophytic vegetation in Italy. Although the results from
hyperspectral data are better, the similar performance of both types of sensor
suggest that spatial resolution has much more prominent effect on classification
accuracy than spectral resolution. Even if discussion on broadband vs. narrowband,
spatial vs. spectral resolution continues, hyperspectral data, both airborne or spaceborne, are used to map different kinds of vegetations including coastal vegetations
(Schmidt et al., 2004, Filippi and Jensen, 2006); wetland vegetations (Zhang et al.,
1997, Hirano et al., 2003, Li et al., 2005, Rosso et al., 2005); non-native species
(Underwood et al., 2003); invasive plant (Ramsey et al., 2005a, Ramsey et al., 2005b,
Pengra et al., 2007, Tsai et al., 2007); and wetland degradation detection (Schmid et
al., 2004).
Apart from board category differentiation, hyperspectral data can be applied on
vegetation species classification through detecting subtle variations across
wavelength. Early studies measured the spectral reflectance of different vegetation
species in laboratory and compared their differences through examining the shape
of spectral curves (Elvidge, 1990, Vogelmann and Moss, 1993). Table 2.1
summarizes the latest studies involving hyperspectral feature selection techniques
for land cover discrimination. The majority of studies used airborne data while
25
others involve laboratory spectral measurement and only a few relied on satelliteborne sensors. The targets of interest ranged from general land covers to individual
agricultural crops and tree species. The techniques of feature selection are so
diverse involving both traditional techniques such as band-correlation analysis,
random forest and sequential selection as well as advanced selection algorithms
such as SVM-recursive feature elimination and genetic algorithm. The primary
purpose of these studies is to explore the potential of hyperspectral data in
discriminating different land covers and vegetation species.
26
Table 2.1. Summary of study using feature selection technique for hyperspectral feature selection
Data (Sensor)
Targets
Feature Selection Method
Author
Year
1
AVIRIS
Agricultural site
Fast Constrained Search
Maghsoudi et al.
2011
2
AVIRIS
Mainly agricultural
SVM-recursive feature elimination
Li et al.
2011
3
FieldSpec Pro FR
Paddy rice
Inter-band Correlation Analysis
Song et al.
2011
4
Hyperion
Tropical vegetation
Stepwise discriminant analysis
Vyas et al.
2011
5
AVIRIS/ ROSIS-3
General land cover
Random feature/ SVM
Waske et al.
2010
Pal and Foody
2010
6
AVIRIS/ DAIS
General land cover
SVM-recursive feature elimination/
Correlation-based feature selection/
Minimum-Redundancy-MaximumRelevance/ Random forest
7
AVIRIS
Agricultural site
SVM-recursive feature elimination
Zhang and Ma
2009
8
PHI/ AVIRIS
Agricultural site/
general land cover
Sequential forward selection/
Sequential forward floating selection/
Clonal selection feature selection/
Fast clonal selection feature selection
Zhong and Zhang
2009
9
Hyperion
General land cover
Multiobjective search strategy
Bruzzone and
Persello
2009
27
10
AISA Eagle sensor
Tree types and land cover
Sequential forward floating selection
Dalponte et al.
2009
11
AVIRIS/ DAIS
General land cover
Greedy features flip/
Iterative search margin-based algorithm
Pal
2009
12
AVIRIS
Mixed agriculture/ forest
Decision tree/ Sequential forward selection
Wang and Li
2008
13
AVIRIS
Land cover
(mainly vegetation)
Greedy optimization strategy/
Steepest ascent
Guo et al.
2008
14
HYDICE
Lianas and trees
Sequential forward selection
Kalacska et al.
2007
15
DAIS
General land cover
SVM/ Genetic algorithm/ Random forest
Pal
2007
16
CASI
Salt-marsh vegetation
Principal Component Analysis (PCA)/
Spectrum reconstruction
Wang et al.
2007
17
ROSIS/ CASI/ MIVIS
Salt-marsh vegetation
Sequential forward selection
Belluco et al.
2006
18
SE-590 spectroradiometer
Coastal wetland vegetation Second derivative approximation techniques
Becker et al.
2005
19
GER 3700 spectrometer
Salt-marsh vegetation
Mann-Whitesy U test
Schmidt and
Skidmore
2003
20
ASD FS UV/VNIR
spectroradiometer
Tropical vegetation species
Spectral shape filtering
Cochrane
2000
28
For instance, Cochrane (2000) retrieved the reflectance spectra of 11 tropical forest
species under laboratory environment and investigate the potential of tropical tree
species level discrimination. Schmidt and Skidmore (2003) studied the potential of
spectral discrimination of 27 saltmarsh vegetation types of the Dutch Waddenzee
wetland through field reflectance spectra measurement. Their findings identified
key regions locating at 404, 628, 771, 1398, 1803 and 2183 nm that are critical to
differentiate wetland species. Becker at al. (2005) took a step further to resample
the field-measured spectral radiance into band configurations of hyperspectral
imagery and used derivative technique in extracting narrow spectral bands in visible
and near infrared regions which can optimally distinguishing coastal wetland
vegetation. Their findings suggested wavelengths at 685.5, 731.5, 939.9, 514.9,
812.3, 835.5, 823.9 and 560.1 nm are significant to discriminate coastal wetland
species (Becker et al., 2005). Studies also tried to locate the spectral intervals, which
are effective in vegetation type discrimination. Wavelength interval around 680,
850, 1650 and 2200 nm are reported to be most useful to differentiate various kinds
of leaves (Kumar et al., 2001). The near infrared plateau located at 771 nm is also an
important region to discriminate vegetation types (Schmidt and Skidmore, 2003).
Vyas et al. (2011) used stepwise discriminant analysis to select bands from Hyperion
data to differentiate eight tropical vegetation types and their results revealed 22
bands locating at 437, 489, 509, 611, 631, 783, 916, 953, 973, 1003, 1170, 1478,
1498, 1578, 1709, 2062, 2153, 2164, 2183, 2235, 2264 and 2275 nm are significant.
Results from the studies advocated the utilization of high spectral resolution data
for species-based vegetation discrimination. The identified spectral regions with
high discriminability of vegetation species can act as a reference for analysis of
hyperspectral images.
Studies using remote sensing data for mangrove mapping exists, however those
using imaging spectrometry are scare in the literatures (Blasco et al., 1998,
Chauvaud et al., 1998, Green et al., 1998a, Green et al., 1998b, Gao, 1999, Manson
et al., 2001, Vaiphasa et al., 2005, Vaiphasa et al., 2006). For instances Blasco et al.
(1998) identified a few factors affecting spectral discrimination of mangrove. Since
mangroves are coastal vegetations, tidal effects on soils and the high air humidity in
29
the atmosphere are two prominent factors concerning mangrove discrimination.
Chauvaud et al. (1998) used high-resolution aerial photographs to map the tropical
ecosystem including mangrove, sea-grass and coral reef. Rasolofoharinoro et al.
(1998) used SPOT 1 and 2 data to estimate the components of mangrove ecosystem
in Madagascar. Green et al. (1998b) used high spatial resolution multispectral CASI
data for classification of mangrove categories and also compared the results with
Landsat TM and SPOT XS data. Gao (1999) also compared the merits of spatial and
spectral resolution in mangrove forest mapping and found that spectral resolution
is more important than fine spatial resolution in mangrove mapping because
through the use of high spectral resolution data, non-mangrove species are
identified. Manson et al. (2001) further compared the accuracy of different remote
sensing techniques including topographic maps, high-resolution aerial photographs
and Landsat TM images in mapping the extent of mangroves in northern Australia.
Vaiphasa et al. (2005) measured the spectral information of tropical mangrove
species under the laboratory conditions and used statistical distance measures to
explore the potential for discriminating mangroves at species level. But the study is
restricted in the laboratory environment without taking a further step to correlate
the measurements with airborne or space-borne hyperspectral data.
Vegetation classification studies can be benefited from such technology owing to
the unique spectral characteristics of vegetation and its dramatic variation of
reflectance with wavelength compared with other surface features (Gao, 1999).
Relationship between class discriminability and wavelength can be uncovered
through pattern recognition and classification processes. When coupled with the
knowledge on vegetation spectral properties, further understandings on the factors,
such foliar constitutes or canopy structural parameters affecting the results of
classification can probably be drawn.
30
2.4. Radar Imaging and Vegetation Classification
Apart from hyperspectral sensor, radar (radio detection and ranging) remote
sensing is also gaining popularity in vegetation-related studies. Radar occupies the
microwave region of electromagnetic spectrum extending from 1 mm to 1 m in
terms of wavelength or 0.3 GHz to 300 GHz in terms of frequency. Radar sensors
operate in different wavelengths including P, L, S, C, X, Ku and W (in ascending
order of wavelength). Irrespective of wavelength, radar signals can be transmitted
and/ or received in either horizontal or vertical polarization modes. Polarization of
electromagnetic wave refers to signals are “filtered in such a way that its electrical
wave vibrations are restricted to a single plane perpendicular to the direction of
wave propagation (Lillesand et al., 2008, p.649)”. For multipolarization radar
systems, there are in four typical polarization combinations including HH (horizontal
transmit and horizontal receive); VV (vertical transmit and vertical receive); HV
(horizontal transmit and vertical receive); and VH (vertical transmit and horizontal
receive). HH and VV are like-polarized while HV and VH are referred to as crosspolarized (Chand and Badarinath, 2007, Lillesand et al., 2008). Synthetic aperture
radar (SAR) is the most common system used to acquire data. SAR makes use of the
sensor motion along track to acquire a successive array of data using a single
physically short antenna, which are then mathematically synthesized into long
synthetic antenna (Lillesand et al., 2008). A number of airborne and spaceborne SAR
sensors in capturing different radar bands are available and they are shown in Table
2.2.
31
Table 2.2. Microwave bands and SAR sensors
(Adopted from Soergel, 2010)
Band
P
L
S
C
X
Ku
W
Center frequency
(GHz)
0.35
1.3
3.1
5.3
10
35
95
Wavelength (cm)
85
23
9.6
5.66
3
0.86
0.32
E-SAR
RAMSES
ERS-1
ERS-2
ENVISAT
RADARSAT-1
RADARSAT-2
SIR-C
E-SAR
EMISAR
AIRSAR
RAMSES
SIR-C
TERRASA
R-X
SRTM
PAMIR
E-SAR
Pi-SAR
RAMSES
MENPHIS
RAMSES
MENPHIS
RAMSES
SAR spaceborne
and airborne
sensors
E-SAR
AIRSAR
RAMSES
ALOS
E-SAR
EMISAR
Pi-SAR
AIRSAR
RAMSES
One of the distinctive appearances of radar images is the presence of speckles,
which looks like random pattern of bright and dark pixels throughout the image.
The phenomenon due to the fact that the receiving signal is in phase or out of phase
(of varying degrees) when compared with the transmitted signal depending on the
exact distances the wave travelled. The different degrees of in phase or out of
phase give radar image a grainy appearance (Lillesand et al., 2008). The presence of
speckles suggested that proper image processing techniques such as filtering
techniques must be applied prior to further image analysis.
Since radar operates in a side-looking manner, the relief displacement found in
radar images is different from images or photographs taken in vertical manner.
Typically, there are three types of image distortions including layover effect,
foreshortening effect and radar shadow caused by topographic variations (Tso and
Mather, 1999a). The magnitude of each type of distortion is the function of the
incident angles of radar pulse, height of feature and the distance of feature relative
to the radar trajectory (near or far range). The layover effect appears when return
signals from the top of a vertical feature reach the antenna before the returns from
the base. The effect is more severe at near range with steeper incident angle. When
the incident angle is less steep and/or the terrain feature is flatter, the radar pulse
32
reaches the base of the feature before it reaches the top and no layover happens.
However, whenever the top and base of the feature are not imaged simultaneously,
the foreshortening effect signifying the failure to reflect the true size and shape of
terrain feature occurs. Foreshortening is less extreme when moving from near to far
range. Radar shadows are more apparent when the slope facing away from the
antenna is steeper than the incident angle of radar pulse causing no illumination
and return signal from the areas. This results in a complete black shadow zone with
its length increase with range distance (Lillesand et al., 2008). However, the effect
of relief displacement on flat surface is negligible.
Compared with optical remote sensing, radar remote sensing has merits including
weather and darkness independent, cloud penetration ability (Pasqualini et al.,
1999, Kushwaha et al., 2000, Filho and Paradella, 2002, Miles et al., 2003); and
vegetation canopy penetration capability (Bourgeau-Chavez et al., 2001). As active
sensor emits its own energy for detection, data acquisition is not restricted to
daytime. The canopy penetration property allows acquisition of structural and
biophysical information under the canopy layer including the dielectric or moisture
properties of both vegetation and soil (Kasischke and Bourgeau-Chavez, 1997,
Baghdadi et al., 2001, Lu and Meyer, 2002); variation in vegetation density, canopy
structure (Patel et al., 2006), target roughness (Cohen and Lara, 2003). It is
suggested that due to the high sensitivity of radar backscattering to the change in
soil moisture content, radar images are more suitable for mangrove mapping and
monitoring (Miles et al., 2003). Although SAR data is free from the impact of
atmospheric condition; the magnitude of radar backscattering is complicated by a
number of factors including wavelength, polarization, angle of incidence, terrain,
and canopy parameters (Patel et al., 2006).
Radar has been actively used for vegetation studies because the signal is sensitive
to a variety of vegetation parameters such as canopy structure, orientation, size,
moisture level of leaves, branches and trunks (Patel et al., 2006). Generally,
vegetation canopy interact with SAR signal in volumetric manner. Theoretically, four
distinct layers including canopy, trunk, aerial root and ground surface layer can be
identified specifically for mangrove forest. A canopy layer comprises small branches
33
and leaves; a trunk layer makes up by large branches and trunks; an aerial root layer
is specific for some mangrove species such as A. marina with extensive
pneumatophore and a ground surface layer can either be free of or be
overwhelmed by water in mangrove stands. The major sources of scattering from
vegetation canopies contributing to the total radar backscatter from woody
vegetation,  tw , can be expressed after Kasischke and Bourgeau-Chavez (1997) as
 tw   c   c2 t2  m   t   a   s   d 
Eq. 2.1
where
 c is backscatter from multiple/ volume scattering of crown layer
 m is multiple-path scattering between ground and canopy layer
 t is direct scattering from the trunk layer
 s is direct scattering from ground surface
 d is double-bounce scattering between trunk and ground
 a is double-bounce scattering between aerial root and ground
 c is the transmission coefficient of vegetation canopy
 t2 is the transmission coefficient of trunk layer
For non-woody herbaceous vegetation, the total radar backscatter,  th , can be
simplified by removing all the terms related to the trunk layer as
 tw   c   c2  m   s 
Eq. 2.2
Figure 2.4 shows the schematic diagram of scatterers from woody and herbaceous
vegetation. Whenever interaction with ground occurs, soil moisture, presence of
surface water layer and surface roughness are important contributing factors to
backscattering (Patel et al., 2006).
34
a. Woody mangrove species
σ°m
σ°c
σ°d
σ°m
σ°a
σ°s
σ°t
b. Herbaceous mangrove species
σ°s
σ°c
σ°m
Figure 2.4. The backscattering model components of (a) woody and (b) herbaceous
mangrove
(Modified based on Kasischke and Bourgeau-Chavez, 1997 & Wang and Imhoff,
1993)
The magnitude of SAR backscatter attenuated by the vegetative volume and each of
the backscatter components is a function of wavelength/ frequency, polarization of
transmitted and received signal, angle of incidence and numerous terrain and
canopy characteristics (Patel et al., 2006). The contribution of the canopy layer in
backscatter depends highly on size distribution and orientation of vegetative
elements including foliages, branches and trunks (Patel et al., 2006).
With lots of combination and interaction between the variables, researchers tried
to come up with generalization rules of individual factor based on existing theories,
analyses, empirical observations and statistical modeling. In terms of radar
wavelength, shorter wavelengths such as X and C bands (3 and 5 cm wavelengths)
tend to interact most actively with the leaves and branches of the upper canopy. L
35
and P band with relatively longer wavelengths (68 and 24 cm) have higher
penetration ability and interact with trunk and ground (Dobson et al., 1992). Hence,
at a fixed angle of incidence, the depth of penetrability of radar signal into
vegetation canopy is dependent on the frequency, which in turn determines the
number of scatterers available for interaction with the incoming signal. With
penetration depth of a few centimeters, C band SAR is responsive to the structure
of top canopy reflected in leaf area index (Chand and Badarinath, 2007) though the
SAR backscatter from longer wavelengths (P and L band) is more sensitive to
biophysical parameters related to above-ground biomass than that of shorter
wavelengths owing to their high saturation level (Le Toan et al., 1992, Beaudoin et
al., 1994, Rauste et al., 1994, Imhoff, 1995, Rignot et al., 1995, Green et al., 1998a,
Green et al., 1998b). However, due to technical hindrance, L-band ALOS-PALSAR is
the only space-borne sensor providing the longest wavelength nowadays.
As for polarization, Dobson et al. (1995) summarized that HH is dominated by
ground-trunk scattering, HV is sensitive to volume scattering with the canopy and
VV tends to interact with crown–layer attributes. Besides, cross-polarized radar
signal is confirmed to be more sensitive to vegetation volume than like-polarized
radar signal at a given frequency (Patel et al., 2006) and tend to show better
discrimination between different cover classes (Baghdadi et al., 2001). This
attributes to depolarization of incoming radar signal due to multiple reflection while
interacting with vegetation components within the canopy (Patel et al., 2006). The
magnitude of signal return from like-polarized signal depends on the structure and
orientation vegetative components. Based on (Henderson and Lewis, 1998), strong
coupling takes place when the electrical field vector of the incoming wave is parallel
to the orientation of dielectric components. That is, vertically oriented branches
produce strong coupling with vertically oriented electrical field vector of incoming
radar signal.
The angle of incidence also affects the penetration ability of radar signal (Ozesmi
and Bauer, 2002). SAR data with low angle of incidence is more sensitive to
vegetation owing to the longer passage through the vegetation (Patel et al., 2006).
36
As shown in Figure 2.4 and Eq. 2.1, the total amount of backscatter  tw , is a
simplified function of five components including direct backscattering from the
ground, trunk-ground backscattering, direct backscattering from trunk, crownground backscattering and crown backscattering including multiple scattering with
the crown. The magnitude of each element depends on vegetation structure while
the interaction between ground and magnitude of radar backscattering depends on
surface roughness and moisture (Patel et al., 2006).
If several strata of canopy layers are present, the radar signals are high due to
multiple scattering effects while low radar responses were observed for
homogeneous canopy. Stand of different ages causing variation in height and
density exhibited different magnitude of radar signal response. Normally, mature
and homogeneous stands has relatively low backscattering signal compared with
young regenerations (Aschbacher et al., 1995)
Finally, water is an important factor affecting the dielectric constant. The increase
of soil moisture tends to increase the magnitude of radar backscatter (Lu and Meyer,
2002). However, the effect varies with different vegetation structure or types as
Kasischke and Bourgeau–Chavez (1997) pointed out that the situation will be
inversed if applied to herbaceous vegetation rather than woody vegetation.
Factors affecting SAR backscattering are inter-dependent. Due to local
environmental and biological settings such as vegetation structure, terrain
roughness, the interpretation of radar backscattering with biomass level is further
complicated.
Single use of radar backscattering data for image classification and vegetation
delineation existed (Kasischke et al., 1995, Kasischke and Bourgeau-Chavez, 1997,
Hashim and Kadir, 1999, Baghdadi et al., 2001, Bourgeau-Chavez et al., 2001, Miles
et al., 2003). Among those, Kasischke et al. (1995) and Kasischke and BourgeauChavez (1997) discriminated wetland and non-wetland vegetations using ERS-1 Cband SAR images. Hashim and Kadir (1999) examined the possibility of JERS-1 and
RADARSAT data in discriminating mangrove species and for biomass mapping.
37
Based on radar’s characteristics, approach of combining multi-polarized data, multiincidence angle radar images (Cimino et al., 1986) have been explored.
For single-band single-polarization images, time sequential information or multitemporal SAR data has also been explored for the potentials in vegetation
classification. (Wooding et al., 1993, Schotten et al., 1995, Kushwaha et al., 2000).
Multi-temporal SAR data have been applied to crop discrimination by its capability
in detecting the possible change in backscattering over the development stage of
crop (Wooding et al., 1993, Schotten et al., 1995). Although mangroves are
evergreen, the species exhibits distinct flowering and fruiting periods throughout
the year. The life cycle does not only influence the spectral characteristics, but also
the coherent structure of vegetation canopy that can be reflected in radar wave
backscatter. Baghdadi et al. (2001) examined multi-temporal and multi-polarization
SAR image to differentiate various wetland classes.
Apart from single use of achromatic radar image, merging with multi-spectral
optical sensor data (Kuplich et al., 2000, Kushwaha et al., 2000, Miles et al., 2003) or
combining both multi-spectral and multi-temporal (Zhu and Tateishi, 2006) have
been realized by many researchers. The varieties are reviewed in Pohl and Van
Genderen (1998). The aims of data integration from different sensors are to retrieve
information that is unattainable or to reduce uncertainty associated with data from
single or individual sensor and with the ultimate objective of minimizing the
classification errors (Kushwaha et al., 2000). Taking the advantage of image
integration, optical images were combined with SAR image for land use/ land cover
classification (Imhoff et al., 1987, Solberg et al., 1994) and vegetation classification
(Pasqualini et al., 1999, Le Hegarat-Mascle et al., 2000, Bourgeau-Chavez et al.,
2001, Held et al., 2003, Filho et al., 2006, Chand and Badarinath, 2007). Aschbacher
et al. (1995) found that the sole dependence of SAR image failed to classify
mangrove and non-mangrove areas. By combining SPOT and JERS-1 images, even
and uneven-aged stands were identified due to their variations in height and
density, which in turns affect surface roughness. Chand and Badarinath (2007)
found that the merge of optical data of LISS-III data and ENVISAT-ASAR C-band data
resulted in better discrimination of forest types. Examples specifically for mangrove
38
studies includes Pasqualini et al. (1999) who used both SPOT and SIR-C data
separately to map and monitor the mangrove habitats which is under exploitation;
Filho and Paradella (2002) who combined the Landsat TM data with RADARSAT-1
data to map coastal features related to mangrove; Bourgeau-Chavez et al. (2001)
compared multi-frequency and multi-polarization SIR-C data for wetland mapping
and monitoring; and Held et al. (2003). In general, the synergistic approach
integrates disparate and complementary information in vegetation mapping and
monitoring (Held et al., 2003) and allows more precise discernability (Leckie, 1990,
Kachhwaha, 1993, Aschbacher et al., 1995, Wilkinson et al., 1995) which in turns
improves accuracy in land use classification (Nezry et al., 1993, Solberg et al., 1994,
Ramsey III et al., 1998, Kuplich et al., 2000, Zhu and Tateishi, 2006, Chand and
Badarinath, 2007) and wetland classification (Baghdadi et al., 2001, Ozesmi and
Bauer, 2002).
The hyperspectral data retrieve information related to principal constitutes and
chemical composition of vegetation while the radar data offer structural
information of canopy and below-canopy. The distinctive information extracted
from the two different types of data exhibits potential for better vegetation
classification accuracy. However, the high volume of data offers opportunities but
also poses challenges to classification process. In the next section, the problem of
high dimensionality and methods of dimensionality reduction are discussed.
2.5. Pattern Recognition for Vegetation Classification
2.5.1. The Hughes Phenomenon and Dimensionality Reduction
Hyperspectral data and its integration with other data sources form high
dimensional vector data structure. In pattern recognition terminology, the
observations (pixels) are described by a large number features (bands). When the
number of features is over 100, it becomes a high dimensional problem. Generally,
it is believed that the increase of features should provide more information that can
improve the classification process. However, this is true if and only if additional
39
feature/ band adds independent information to the classification process and better
classification accuracy can be achieved (Kavzoglu and Mather, 2002). The high
correlation among the adjacent contiguous bands in hyperspectral data suggests
that the assumption of independent information is not necessarily valid.
From statistical perspective, provided that the performance of a statistical classifier
is a function of the interrelationship between sample sizes, number of features and
classifier complexity (Jain et al., 2000), high dimensional data can pose significant
impacts on statistical classifier. Theoretically, if the class-conditional probability
density is entirely known, the increase in number of features would not degrade the
classification accuracy of a classifier. Such condition can only be achieved when
sufficiently large samples are provided to reveal the underlying density. Since this is
rarely the case, the classifier’s performance exhibited a reverse trend with features
added. This paradoxical behavior is related to the “curse of dimensionality” coined
by Bellman (1961), which suggests the number of training samples should be
increased with dimensionality (Figure 2.5a). This leads to the so-called “peaking
phenomenon” or “Hughes phenomenon” which advocates the classification
accuracy improves in the beginning and to a point when the instability of estimators,
mean vector and variance-covariance matrices for a particular class, is larger than
the additional information content provided by extra feature (Hughes, 1968). This
eventually leads to the decline in accuracy as the number of dimension continue to
increase while the number of training sites remains unchanged as shown in Figure
2.5b. Given a fixed number of samples, as the number of features increases, a
corresponding increase in unknown parameters which curtail the reliability and
accuracy of estimation of statistical parameter is observed (Jain et al., 2000). The
class separation results may be seriously degraded when the ratio of sample size to
the number of features is low (Meisel, 1972, Raudys and Pikelis, 1980, Jain and
Chandrasekaran, 1982, Raudys and Jain, 1991), and therefore fail to reveal the
underlying structure of class distributions (Dundar and Landgrebe, 2003).
40
Figure 2.5. The classification accuracy as a function of training samples and
dimensionality
(Adopted from Duin and de Ridder, 1997)
The low ratio of sample size to the number of features results in a poor
generalization ability of a classifier (Jain and Chandrasekaran, 1982). The
generalization ability of a classifier is measured by its performance in classifying test
patterns which are independent from the training set. Under small sample size, a
classifier is more likely to be over-trained or over-fitted on the training data that
minimize the apparent error rate (Jain et al., 1987, Jain et al., 2000).
Dundar and Landgrebe (2003) pointed out that the lack of adequate training data
makes the use of high order statistics, other than the mean vector and covariance
matrix (second order statistics) to precisely estimate parametric class density
function impractical. When nonparametric class density estimation is considered,
the limited samples which fail to estimate the true density function, are less reliable
and more bias than the parametric estimation (Fukunaga, 1990). Indeed, the
amount of training data ascends exponentially with the increase in dimensionality
(Bishop, 1995, Jain et al., 2000). The generally accepted rule-of-thumb is
NT  10 N F
Eq. 2.3
41
Where NT is the number of training samples for each class while NF is the number
of features (Foley, 1972, Jain and Chandrasekaran, 1982, Raudys and Jain, 1991,
Casasent and Chen, 2000) though the practice is regarded as more than sufficient
(Jain et al., 2000). Besides, the more complex the algorithm is, the ratio of sample
size to dimensionality should be increased accordingly (Casasent and Chen, 2000).
Intuitively, the sample size can be increased to circumvent the ‘curse of
dimensionality’. However, ‘sufficient’ training samples are always prohibitively
expensive in terms of time and cost. Indeed, classification is always conducted
under the deficiency of training samples when hyperspectral data are used as tens
of thousands of training samples are needed (Hsu, 2007). Even the sample size is
met, the computational complexity is intolerable (Wang and Li, 2008).
It is apparent that the “curse of dimensionality” is realized with the advent of
hyperspectral remote sensing instruments as the size of training set can rarely
satisfy the classification requirement (Casasent and Chen, 2000, Hsu, 2007). Apart
from hyperspectral images, the integration with other sources of data such as
textural variables extracted from optical and radar images is becoming popular. The
combination of these data together for vegetation classification poses a great
challenge to the statistical classifiers and subsequently influences the classification
result. This urges for measures to reduce the dimensionality efficiently and
effectively.
Although an observation can be described by as many features/ bands as possible,
not all features are equally significant to a specific classification problem. Some of
them may be redundant due to high correlation or even irrelevant (Verikas and
Bacauskiene, 2002). This is especially true for hyperspectral data as high
correlations are found between the adjacent spectral bands. When training samples
are finite, better performance may be achieved by reducing the number of
dimensions (Fukunaga, 1990, Lee and Landgrebe, 1993a, Benediktsson et al., 1995,
Landgrebe, 2001, Hsu, 2007). The objective of dimension reduction is to reduce the
number of features substantially without sacrificing significant information (Goetz
et al., 1985, Shaw and Manolakis, 2002). It is a process of projecting data from high
dimensional space to low dimensional subspace consisted of fewer but salient and
42
relevant features (Tadjudin and Landgrebe, 1998, Hsu et al., 2002, Wang and Li,
2008). However, the elimination of features may probably result in loss of the
discrimination power and thus low classification accuracy (Jain et al., 2000). An
important concern of dimensionality reduction is to preserve relevant/ useful
information and to discard redundant and irrelevant in the process.
The method of dimensionality reduction can be broadly classified into two
approaches – feature selection and feature extraction (Young and Fu, 1986,
Tadjudin and Landgrebe, 1998, Hsu et al., 2002). Feature selection or band selection
is to select a subset of feature directly from the original feature space (Hsu et al.,
2002). Feature extraction involves the transformation of all the features in the
original feature space to lower-dimension subspace consisting of effective features
(Hsu et al., 2002). The choice of approach stems from the application domain and
the availability of specific training data (Jain et al., 2000, Pudil et al., 2002). For all
dimensionality reduction problems, the process can be divided into four important
components: (1) original feature set consists of D feature; (2) criterion function J ;
(3) effective search strategy; and (4) selected feature subset consisting of d feature.
The criterion function is used to evaluate the quality of selected subset of feature
and formulated the stopping rules. The search strategy is to find the optimal/
suboptimal feature subset or the optimal transformation (Pudil et al., 2002). In
determining the final subset d , there are basically two methods termed dparametrizing and d-optimizing. d-parametizing is to determine d arbitrarily and
then to find the best subset from the original D features yielding the highest score
in the criterion function J . d-optimizing is to find the smallest subset which result
in an tolerable discrimination error rate within a given threshold (Foroutan and
Sklansky, 1987).
Dimensionality reduction is critical for hyperspectral images prior to classification
for two main reasons. One arises from the fact that high correlation are always
found between adjacent bands, which suggests that a subset of selected bands can
be able to represent other bands without much loss in information (Tu et al., 1998).
Another reason is that the amount of information possessed by the bands is
different while most information may be dominated by a few major bands.
43
Prioritizing bands according to the significance of their information allows the
removal of redundant and irrelevant bands containing least information (Lee and
Landgrebe, 1993a, Tu et al., 1998). On the whole, the success of discovering the
band subset to trim down the data volume while maintaining the desired
information is highly appreciated for classification problems with limited samples.
To summarize, the goal of dimensionality reduction is four-fold: alleviating the
Hughes phenomenon (Jain et al., 2000, Hsu et al., 2002); enhancing efficiency by
reducing computation time/ measurement cost (Jain et al., 2000, Hsu et al., 2002,
Guyon and Elisseeff, 2003); improving the prediction performance/ classification
accuracy (Kittler, 1986, Kudo and Sklansky, 2000, Guyon and Elisseeff, 2003); and
better understanding the underlying concepts in the model (Bi et al., 2003, Guyon
and Elisseeff, 2003).
2.5.2. Statistical Pattern Recognition and Feature Selection
A typical statistical pattern recognition problem is to classify a pattern described by
a real D -dimensional feature vector x  ( x1 ,..., xD ) to one of C possible classes  ,
  1,..., C (Pudil et al., 2002). Feature selection has been intensively research since
1970s in statistical pattern recognition (Jain and Zongker, 1997, Mitra et al., 2002),
data mining (Dash and Liu, 1997) and machine learning (Kohavi and John, 1997).
Given a set of D features, feature selection is to select a feature subset of size d ,
d  D , that gives the best performance, where evaluation criterion J () is used to
compare the performance of the selected feature subset (Pudil et al., 2002,
Nakariyakul and Casasent, 2007). If the input patterns are labeled, relevant features
are selected while redundant and irrelevant features are disposed by their relations
with the corresponding class labels, i.e. supervised feature selection. The
unsupervised methods applied when labeled data is unavailable select relevant
features by exploration of data variance and separability (Dy and Brodley, 2004,
Wolf and Shashua, 2005).
The supervised methods are relatively more
computationally intensive due to the complexity of model formation process. The
unsupervised methods can sometimes be used as a pre-filter to rank the features
44
based on their information content regardless of their class label (Bajcsy and Groves,
2004). Kudo and Sklansky (2000) classified feature selection problem into different
scales according to the number of D considered. When D ranges from 1-19, 20-49
and 50-∞ , it refers to small, moderate and large scale respectively.
The process of feature selection is shown in Figure 2.6. Generally, feature selection
consists of four steps including subset generation, subset evaluation, stopping
criterion, and result validation (Liu and Yu, 2005).
Figure 2.6. The procedure of feature selection, a) filter approach and b) wrapper
approach.
Feature selection begins by subset generation which is a search process that
produce lower dimensional candidate feature subset (Liu and Motoda, 1998a). The
search procedure can be separated into two basic strategies – optimal and
45
 D
suboptimal methods. The optimal method requires the examination of all  
d 
possible subsets of size d . The exhaustive search is the only algorithm that
guarantees optimal solution (Jain et al., 2000). Branch and bound (BB) algorithm is a
non-exhaustive and accelerated search procedure that can attain optimal subset
based on monotonic criterion function. The suboptimal methods use optimality as a
trade-off for computational efficiency (Pudil et al., 1994). The sequential search
(Devijver and Kittler, 1982), floating search (Pudil et al., 1994), beam search (Aha
and Bankert, 1995), neural network, decision tree and genetic algorithm (Goldberg,
1989) are algorithms of suboptimal search.
After the subset was generated by the search algorithm, it is then evaluated by the
chosen objective criterion function. The criterion functions are divided into three
groups – filters, wrappers and hybrid (Liu and Yu, 2005). The filter approach
assesses feature subsets by their information content using interclass or
probabilistic distance measures, probabilistic dependence measures divergence or
information-theoretic measures. In other words, the evaluation is independent of
classifiers. The aim is to discard redundant or irrelevant features before the
classification procedure (Liu and Motoda, 1998a).
The wrapper approach evaluated the quality of feature subsets using the prediction
performance of a given classification algorithm through statistical resampling or
cross-validation (Kohavi and John, 1997, Kavzoglu and Mather, 2002, Guyon and
Elisseeff, 2003). It is apparent that the approach aims at minimizing the
classification error/ enhancing classification performance of the classifier of interest,
but at the expense of high computational complexity (Kohavi and John, 1997, Liu
and Yu, 2005). The expensive computational property prohibits its general
application on some classification algorithms (Blum and Langley, 1997). The
generalization ability of the selected features on other classifiers is substantially
curtailed (Peng et al., 2005).
The stopping criterion determines when the feature selection process should halt.
The criteria usually applied include, (1) all search process is completed; (2) the
46
defined bound such as the number of features, maximum iteration is reached; and
(3) no improvement of criterion function with additional feature added; and (4) a
classification error rate less than the allowable error rate is attained (Liu and Yu,
2005).
The advantage of feature selection is that it can retain the physical meaning of the
original features which is sometimes vital to the understanding of the underlying
process that generates the patterns (Jain et al., 2000, Pudil et al., 2002). The search
method, evaluation criterion and stability measures are reviewed below.
2.5.2.1. Search Method
The search process is to generate subsets. Jain and Zongker (1997) has provided a
taxonomy of commonly used search algorithms shown in Figure 2.7. Generally,
feature selection algorithms are split into those using statistical pattern recognition
technique and those employing artificial neural network. The statistical technique is
divided into algorithms offering optimal and suboptimal solution feature set. Based
on the characteristics in subset generation, the suboptimal methods are further
divided into single and multi-solution. Algorithms that keep one existing feature
subset and modify it at a time refers to the former while the latter uphold a
population of subsets. Given the same problem, if the same subset of features is
produced on every run, it is deterministic; if different subsets are generated every
time, it is stochastic (Jain and Zongker, 1997).
47
Figure 2.7. The commonly used search algorithm for feature selection
(Adopted from Jain and Zongker, 1997)
2.5.2.1.1. Exhaustive search
Exhaustive search evaluates all possible subsets of size d of the original feature set
and it is the only optimal search method with non-monotonic criteria (Shapira and
Gath, 1999, Pudil et al., 2002, Pudil and Somol, 2005). Supposed the number of
spectral and textural bands is D , an optimal subset of d , d  D , is to be selected.
The number of feature combinations required to be examined equals
D!
D  d !d!
Eq. 2.4
As the total number of features increases, the number of subsets to be evaluated
grows excessively (Yu and Yuan, 1993). For example, when 12-dimensional subset is
to be selected from 24 available features, it requires the evaluation of 2.7 million
possible subsets. The exhaustive search is computationally impractical and
prohibitive as the number of possible sets expands combinatorially (Narendra and
48
Fukunaga, 1977, Lee and Landgrebe, 1993b). It is only applicable for low
dimensional problem with D less than 30 (Nakariyakul and Casasent, 2007). Even if
the hyperspectral features are considered alone, the computation is unacceptable
(Hsu et al., 2002).
2.5.2.1.2.Branch and bound
The supervised feature selection method that circumvents exhaustive enumeration
but guarantees the selected feature subset gives globally best value of certain
feature selection criterion function that is monotonic is called the branch and
bound (BB) algorithm (Yu and Yuan, 1993, Somol et al., 2004). The algorithm was
first adopted by Narendra and Fukunaga (1977) for feature selection purpose. The
monotonicity is the essential condition that accelerates the optimal search process
by identifying the feature space within which an optimal solution cannot be possibly
found. Follows the notation of Somol et al. (2004), let X k be the subset of features
attained by discarding k features y1 , y2 ,, yk from the original set Y of all D
features, that is,
X k  Y { y1 , y2 ,, yk }
Eq. 2.5
The monotonicity condition assumes that, for feature subsets X 1 , X 2 , , X k ,
where
X 1  X 2  X k ,
Eq. 2.6
The feature selection criterion function J fulfills
J ( X 1)  J ( X 2 )    J ( X k ) ,
Eq. 2.7
*
The feature selection algorithm is to find the optimum subset X k such that it yields
the maximum criterion value, i.e.,
 
*
J X k  max X k
Eq. 2.8
49
The BB algorithm builds a search tree with the root represents the original set of D
features while the leaves characterize the target subsets of d features. When
traversing down the tree from root to leaves, the BB algorithm discards single
feature, i.e. y1 , y2 ,, yk , consecutively from the current set of feature candidates
in each level until the target number of features d is achieved. The algorithm stores
two important information while trespassing down the tree – (1) the current best
feature subset X and (2) the corresponding criterion value X * the best subset
yields denoted as the bound, that is,
 
*
X*  J Xk
Eq. 2.9
As long as the criterion value for any internal node is lesser than the current bound
X * , subject to the monotonic condition, all the nodes that are successor of that
node will also have criterion values less than X * (Narendra and Fukunaga, 1977).
That is,
 
J X h  X * , where h  k
Eq.2.10
Then by Eq. 2.7,


J X 1 , X h , X h1 ,, X k  X * ,
Eq. 2.11
for all possible { X h1 , X k }
Therefore, the whole sub-tree is trimmed as no optimum solution can be found
with those feature subsets, which saves lots of computational iterations (Narendra
and Fukunaga, 1977, Devijver and Kittler, 1982, Fukunaga, 1990, Somol et al., 2004).
An example for D = 5 and d =2 is illustrated in Figure 2.8.
50
Figure 2.8. Example of Branch and Bound solution, where 2 features are to be
selected from the original feature with 5 features (D =5 and d =2). The aim is to
maximize the illustrative criterion function
(Adopted from Somol et al., 2004)
The BB algorithm needs to simply consider a fraction of all possible feature subsets
by rejecting suboptimal subsets using the monotonicity property of criterion
function, which tends to be more efficient in finding the optimal subset than
exhaustive search counterpart (Pudil et al., 2002). It is regarded as the powerful
combinatorial optimization tool that offer significant performance gains (Narendra
and Fukunaga, 1977). Bruzzone and Serpico (2000) applied the optimal BB algorithm
to select TM bands and their derivatives to classify five agricultural classes. Due to
moderate feature size, optimal solution was obtained effectively.
One of the drawbacks is that BB requires the feature selection criterion to be
monotonic in nature (Jain and Zongker, 1997, Pudil et al., 2002). This means that
the addition of new features into the feature subset can only increase (never
decrease) the value of the criterion. Figure 2.9 shows the concept of different
degrees of monotonicity for classification problem as described in Kudo and
Sklansky (2000).
51
Figure 2.9. The concept of monotonicity for classification problems
(Adopted form Kudo and Sklansky, 2000)
If feature selection criterion function is totally non-monotonic, the feasible part of
the search space is likely to be pruned. And to the worst scenario, the whole
feasible region may be detached as the BB procedure has no means to explore the
detached parts of the feature selection space (Siedlecki and Sklansky, 1989). The
monotonicity property may not be applicable when small sample size is
encountered as recognized from the curse of dimensionality (Narendra and
Fukunaga, 1977, Jain and Zongker, 1997). Besides, if the feature set is too large, the
complexity of the algorithm will expand exponentially (Jain and Zongker, 1997).
However, some authors showed that BB also work well when the selection criterion
is non-monotonic (Hamamoto et al., 1990). Besides, the use of classification error
rate, which shows slightly non-monotonic with respect to subset inclusion plus
some modifications to the backtracking rules would allow the algorithm to explore
the whole feasible region (Foroutan and Sklansky, 1987). Another constraint that
limits the performance of BB is that the algorithm usually spends most of the
resources evaluating the less promising feature subsets near the tree root. Both the
criterion value computation efficiency as well as probability of subtree being pruned
near the tree root is low, especially when d  D (Somol et al., 2004).
Over the years, substantial efforts have been spent on accelerating the classical BB
algorithm. The research areas mainly focus on reducing redundant criterion J
evaluation through feature ordering, effective initial bound determination, begin
52
search level, and modification of search methods (Yu and Yuan, 1993, Casasent and
Chen, 2003, Chen, 2003, Somol et al., 2004, Nakariyakul and Casasent, 2007).
Fukunaga (1990) improved the basic BB algorithm by ordering the nodes within the
tree accorded to the significance of each feature in the feature set. The ordered BB
algorithm arranges the successor node of a given node with smaller J value to the
left-most of the tree while those with higher J value are allocated to the right.
Analysis starts from the rightmost path of the tree, where large initial bound is
expected. More criterion value calculation is anticipated when compared with basic
BB when d  D as the algorithm search the tree while designing it (Somol et al.,
2004).
Foroutan and Sklansky (1987) proposed a relaxed branch and bound (RBB) method
which introduced the concept of approximate monotonicity. A margin,  , is
introduced to allow the criterion value, J , to have a certain degree of violation of
monotonicity threshold, within which RBB can still operate to come up with the
optimal solution. But the problem is the determination of the value of  .
Chen (2003) developed an improved branch and bound (IBB) algorithm to
accelerate the search by reducing redundant J calculations. It uses the right-left
search strategy to increase the bound value X * as fast as possible and tends to
facilitate more effective branch pruning in later stages. However, the additional
computational cost of heuristic is strongly discouraging (Chen, 2003, Pudil and
Somol, 2005).
The modified branch and bound (MBB) method, developed by Casasent and Chen
(2003) makes use of the sequential forward selection (SFS) to order all D features
from best to worst (and from left to right) and then to determine the initial
estimate of bound X * using the d best features ordered by SFS. Instead of starting
the tree search in the root, the MBB use a jump search algorithm to begin search a
quarter of way down the tree at level ( D  d ) / 4 where J  X * is more likely to
happen. If J  X * for all nodes at level ( D  d ) / 4 , the calculation of J for all nodes
above this level are saved and the search jumps further down to level ( D  d ) / 2 .
53
However, if any nodes has J  X * at level ( D  d ) / 4 , all the node descendants
below that node require no evaluations and are pruned directly. If some nodes have
J  X * at level ( D  d ) / 4 , all the node descendants below these nodes are
examined (Casasent and Chen, 2003). However, the choice of starting search level
and sequential jump level are criticized being arbitrary (Nakariyakul and Casasent,
2007).
Somol et al. (2004) developed a fast branch and bound (FBB) algorithm with aim to
reduce the number of criterion value computations during the tree search process.
Instead of computing the real criterion value in every node, the FBB predicts the
criterion values based on the past feature-dependent criterion value decreases. If
the predicted criterion value remains considerably higher than the current bound, it
is expected that the true criterion value should not be lower and the corresponding
sub-tree could not be pruned. In that case, the FBB keeps on constructing the
successive tree levels. But if the predicted value is close to the bound, the real
criterion value must be computed. The sub-tree may be trimmed only if the “real”
criterion values are lower than the current bound X * . This suggested that the
prediction mechanism would not affect optimality of the results attained. Possible
inaccurate predictions may incur construction of redundant sub-trees, but the
reduction in the amount of computational resources for criterion value of internal
nodes, especially close to the root, would be substantially compensated (Somol et
al., 2004, Pudil and Somol, 2005). However, the drawback is that there is no
theoretical proof that the overall speed won’t be affected when excessive
prediction failure happens (Pudil and Somol, 2005).
The adaptive brand and bound (ABB) method was introduced by Nakariyakul and
Casasent (2007). ABB first orders the full set of D features by significance and
construct the full tree from left to right. The initial bound was then computed from
the best d features out of D using the sequential forward floating selection (SFFS).
ABB selected the start search level by comparing the J value after the removal of
the minimum number of most significant feature from the full feature set Y and
the J value of the initial bound obtained by SFFS. The jump search strategy is
54
applied to search the nodes. If J at some node is  X * , the search will jump
further down to high levels since it is very unlikely that the node descendants in
next level will have a J value less than X * . However, if J at some node is only
minimally larger than X * , the next immediate level of the tree will be examined. It
is expected that the jump search algorithm is able to find a level at which at least
one descendant node to be less than X * , then the descendant nodes beneath that
node can be pruned (Nakariyakul and Casasent, 2007).
Although the BB search algorithm and its variations significantly improve the
computation efficiency, they are more applicable to problems with low to moderate
feature size. Suboptimal search should be the only solution to high dimensional
problems.
2.5.2.1.3.Sequential forward/ backward selection
It is sometimes argued that execution time of algorithm is not as significant as the
optimality of feature subset since the computation is more likely conducted in an
offline manner. However, this is only applicable to feature subset of moderate scale
(Pudil and Somol, 2005). As the feature size grows bigger, subsets evaluation will
become computationally prohibitive (Jain et al., 2000). The data acquired by
hyperspectral sensors constitute a large scale feature selection problem, where an
optimal solution can hardly be obtained even by the non-exhaustive BB search
algorithm (Serpico and Bruzzone, 2001).
Besides, the most commonly used
criterion functions are not monotonic in nature (Pudil et al., 2002). In this case, suboptimal search techniques, which choose to tradeoff optimality for computational
efficiency has been proposed.
The sequential forward selection (SFS) put forward by (Whitney, 1971) constructs a
subset by incrementing one feature at a time into an empty set, X 0 with a view to
maximize J and terminates until the desired d features are retained. The
sequential backward selection (SBS) introduced by (Marill and Green, 1963) starts
with a complete set Y with D features, and remove redundant and irrelevant
55
features one by one from the set so that the resultant subset maximize J . It
terminates until the required d feature is selected. SFS is known as a bottom-up
approach while SBS is the top-down approach (Pudil et al., 1994, Pudil and Somol,
2005). The sequential search which is regarded as the simplest greedy search
algorithm is explained below.
Let Y be the set of D available features, i.e. Y  { yi :1  i  D} and X k be the subset
of k features selected from the set Y , i.e. X k  {xi :1  i  k , xi Y } . The criterion
function J ( X k ) is to be maximized. For SFS algorithm,
(1) Starting with an empty set X 0  { ∅} ;
(2)
the
best
feature
xi
is
selected
from
set
Y
if
the
Xk
if
J ( X k  xi )  max J ( X k  xi ) ;
1i  D
(3) the subset X k 1  X k  xi is updated and set k  k  1 ;
(4) Returns to Step (2).
As for SBS algorithm,
(1) Starting with the complete set X k  Y ;
(2)
the
worst
feature
xj
is
removed
from
J ( X k  x j )  max J ( X k  x j ) ;
1i k
(3) the subset X k 1  X k  x j is updated and set k  k  1;
(4) Returns to Step (2).
The SFS and SBS are commonly used owing to their simplicity and speed. However,
they suffer from so-called “nesting effect” which is regarded a critical drawback.
With nesting effect, the selected feature cannot be discarded later in the bottom up
search approach (SFS) while the feature cannot be re-selected again in the top
56
down approach (SBS) (Pudil et al., 1994). In other words, the subset with best five
features must contain the subset of the best four features, and so on. However, it is
likely that the best five features may not contain any of the best four features in
real practice (Jain and Zongker, 1997).
Early efforts to solve the nesting problem was firstly proposed by (Michael and Lin,
1973). Their ideas were improved and developed into Plus- l -Take-Away- r , search
algorithm by (Stearns, 1976). The algorithm goes forward and add l features by SFS
and then go backward and delete r features by SBS (Kudo and Sklansky, 2000).
Although it was argued that there lacks theoretical way to determine the values of
l and r to obtain the best feature set and the result is suboptimal, the principles
has been extended and formed the basis for sequential floating search.
2.5.2.1.4.Sequential Floating search
The novel concepts of sequential floating forward search (SFFS) and the sequential
backward floating selection (SBFS) methods were proposed by Pudil et al. (1994). It
enhances the SFS and SBS methods by its dynamic backtracking ability after each
sequential forward step so as to find a better subset than the existing subset
obtained so far of the same size (Jain et al., 2000, Kudo and Sklansky, 2000, Peng et
al., 2005). The backtracking mechanism will activate as long as improvement of
criterion J is made to the previous feature set of the same size (Pudil et al., 1994,
Pudil et al., 2002). Instead of changing monotonously by pre-specified values of l
and r in Plus- l -Take-Away- r process, the algorithm allows the flexibility by
“floating” these values up and down so as to approximate the optimal as much as
possible (Pudil et al., 1994). Figure 2.10 shows the sequential forward floating
selection algorithm.
57
Figure 2.10. The sequential forward floating selection (SFFS) algorithm
(Adopted from Pudil and Somol, 2005)
The SFFS selects features found by applying the basic SFS algorithm from the
current feature set D . This is followed by successive conditional exclusion of the
worst feature found by applying one step SBS algorithm in the newly updated set if
an increase of criterion function can be attained when compared with the previous
sets (Pudil et al., 1994). The procedure is shown below:
(1) Starts with an empty set X k  { ∅} ;
(2) apply the SFS algorithm, select the most significant feature xi from the
set Y if J ( X k  xi )  max J ( X k  xi ) ;
1i  D
(3) add xi into the subset and form X k 1  X k  xi ;
(4) apply the SBS algorithm, find the least significant feature xm from the
subset X k 1 ;
(5) Conditional exclusion:
if J ( X k 1  xm )  J ( X k 1  x j ) j  1,2,..., k , set k  k  1 and return
to Step (2); otherwise
if J ( X k 1  xm )  J ( X k ) , remove feature xm and form a subset
X 'k  X k 1  xm ;
58
(6) if k  2 , then set X k  X 'k and J ( X k )  J ( X 'k ) , and return to Step (2),
else go to Step (7);
(7) apply the SBS algorithm, find the least significant feature xn from the
subset X ' k ;
(8) Conditional exclusion continues:
if J ( X 'k  xn )  J ( X k 1 ) , then set X k  X 'k and J ( X k )  J ( X 'k ) ,
and return to Step (2); otherwise
if J ( X 'k  xn )  J ( X k 1 ) ,
remove xn and form a reduced
subset X 'k 1  X 'k  xn , and set k  k  1;
(9) if k  2 , then set X k  X 'k and J ( X k )  J ( X 'k ) , and return to Step (2),
otherwise continues the conditional exclusion process in Step (8) if the
target subset of d features is not achieved, else, the process terminates.
The SBFS excludes features found by applying the basic SFS algorithm from the
current feature set D . This is followed by successive conditional inclusion of the
most significant feature from the available features provided an enhancement of
criterion value can be obtained when compared with the previous sets (Pudil et al.,
1994). The procedure is shown below:
(1) Starts with a full set X k  Y ;
(2) apply the SBS algorithm, select the least significant feature xi from the
set Y if J ( X k  xi )  max J ( X k  xi ) ;
1i  D
(3) remove xi from the subset and form a reduced subset X k 1  X k  xi ;
(4) apply the SFS algorithm, find the most significant feature xm from the
subset X k 1 ;
(5) Conditional inclusion:
59
if J ( X k 1  xm )  J ( X k 1  x j ) j  1,2,..., k , set k  k  1 and return
to Step (2); otherwise
if J ( X k 1  xm )  J ( X k ) , add feature xm and form a subset
X 'k  X k 1  xm ;
(6) if k  2 , then set X k  X 'k and J ( X k )  J ( X 'k ) , and return to Step (2),
else go to Step (7);
(7) apply the SBS algorithm, find the most significant feature xn from the
subset X 'k ;
(8) Conditional inclusion continues:
if J ( X 'k  xn )  J ( X k 1 ) , then set X k  X 'k and J ( X k )  J ( X 'k ) ,
and return to Step (2); otherwise
if J ( X 'k  xn )  J ( X k 1 ) ,
add xn and
form
an
enlarged
subset X 'k 1  X 'k  xn , and set k  k  1;
(9) if k  2 , then set X k  X 'k and J ( X k )  J ( X 'k ) , and return to Step (2),
otherwise continues the conditional exclusion process in Step (8) if the
target subset of D  2 features is not achieved, else, the process
terminates.
A number of researches have compared the classification error and computation
time of several feature selection algorithms (Ferri et al., 1994, Jain and Zongker,
1997). The floating search techniques were found superior than other algorithms.
During the backtracking process, the criterion values are always compared only with
those of the same cardinality of feature subset. Hence, it allows the possible
decrease in value of the criterion function when a new feature is added (Pudil et al.,
1994). With the backtracking capability, not only can the floating search methods
solve the nesting problem, their solutions are comparably promising as the brandand-bound method (Ferri et al., 1994, Kudo and Sklansky, 2000). Although there is
no guarantee of optimal solutions, the close-to-optimal solutions make the
computation cost more affordable for most feature selection problems (Pudil and
60
Somol, 2005). An adaptive version of floating search, proposed by (Shapira and Gath,
1999), which concentrates on finding a better solution with desired dimensionality
also exhibited superior performance but at the cost of higher computational
demand (Somol et al., 1999, Pudil and Somol, 2005). Another advantage of floating
search is their tolerance to the divergences from monotonic behavior of the feature
selection criterion, which is superior to the basic BB method.
2.5.2.1.5.Oscillating Search
Somol and Pudil (2000) proposed the oscillating search (OS) method which is based
on replicated modification of the existing subset X d of d features through
alternative down- and up-swings. During the swings, some of the existing features
in the set X d are replaced by better ones evaluated by selected evaluation criterion
(filter-based) or accuracy of classifier (wrapper-based). The down-swing first
removes some worst features from the set X d and then adds back some good
features while the up-swing does the opposite by adding good features and
removing the bad ones. Two successive opposite swings refers to an oscillation
cycle. While sequential floating search requires no parameter setting, a parameter
called the maximum oscillation cycle depth,  which determines the number of
features to be replaced in one swing should be defined in oscillating search. The
search will stop once the maximum cycle depth is reached. Higher  value means
more thorough search at the expense of computational time (Somol and Pudil,
2000). Figure 2.11 shows the process of oscillating search.
61
Figure 2.11. The process of oscillating search
(Adopted from Somol et al., 2010)
The search process is described below:
(1) Starts with initial set X d of d features and set cycle depth, o  1 ;
(2) remove the worst features from X d and form a new subset X d o ;
(3) if X d o is currently the best one among the subset with cardinality d  o ,
add the best features from Y \ X d o to X d o and form a new subset X ' d .
62
If the criterion value J ( X 'd ) is currently the best among the subsets of
the required size, set X d = X ' d , set o  1 ; otherwise
(4) add the best features from Y \ X d to X d and form a new subset X d o ;
(5) if X d o is currently the best one among the subset with cardinality d  o ,
remove the worst features from X d o and form a new subset X ' d . If the
criterion value J ( X 'd ) is currently the best among the subsets of the
required size, set X d = X ' d , set o  1 and go to Step (2);
(6) termination condition:
if o   , stop the algorithm.
One of the advantages of oscillating search is its focus on finding better solution of
given cardinality without spending much time in evaluating subset size that are of
large disparity from the target ones. Besides, the criterion value improves fast in the
beginning of the search due to low initial cycle depth. This is advantageous because
the interim result is still acceptable in terms of criterion value even though user opt
to terminate the search in the middle of the search process (Somol et al., 2010a).
Figure 2.12 shows the graphical comparison of concepts of SFS, SFFS and OS
algorithms.
Figure 2.12. Graphical comparison of a) Sequential forward selection, b) Sequential
forward floating selection, and c) Oscillating search
(Adopted from Somol et al., 2010)
63
2.5.2.1.6.Genetic algorithm
The development of the genetic algorithm (GA) was stimulated by the hypothetical
natural selection process, which suggests “the fittest individuals at one generation
are more likely to survive and produce the new generation (Kavzoglu and Mather,
2002) p.2927”. It is classified as a suboptimal, stochastic and multi-solution feature
selection method (Jain and Zongker, 1997, Serpico and Bruzzone, 2001). The
application of genetic algorithm in feature selection problem was proposed by
Siedlecki and Sklansky (1989).
In the GA approach, a given feature subset is represented as a binary string of finite
length D (total number of features available) termed chromosome with zero or
one corresponding to the discarded or selected feature in the set (Siedlecki and
Sklansky, 1989, Jain and Zongker, 1997). The algorithm maintains a population of
chromosomes and each of them are evaluated for their likeliness to survive and
breed in the next generation (Jain and Zongker, 1997). Resembling the natural
evolution process, the chromosomes are allowed to crossover and mutate. The two
algorithms can be regarded as parallel and randomized process (Kudo and Sklansky,
2000). By crossover, components from two different parent chromosomes are
blended to produce a pair of offspring chromosomes. A crossover point, at which
the two chromosomes exchange their components, can be selected. For instances,
given two chromosomes of 8-bit binary strings, 01001100 and 10101110 with
crossover point in the middle, two new chromosomes produced by crossover
mechanism are 01001110 and 10101100 (Siedlecki and Sklansky, 1989). A mutation
process is to randomly alter one or more components in the chromosome to
produce a child chromosome. The objective is to augment the variability of the
population (Siedlecki and Sklansky, 1989).
The optimization process is conducted in cycles called generations. In each
generation, a new chromosome is reproduced through crossover, mutation and
evaluation. Parameters including population size, maximum number of generations,
probability of crossover and probability of mutation rate affect the execution time
and performance of GA algorithms (Kudo and Sklansky, 2000). When error rate of a
64
classifier is used as the evaluation criterion for performance measurement, a subset
with error rate lower than the defined feasibility threshold is regarded as
acceptable. Feature subset with the smallest number of features among feasible
feature subsets will be selected (Siedlecki and Sklansky, 1989). However, since GA
cannot handle constrained optimization problem using error rate, a penalty
function is required.
p(e) 
exp((e  t ) / m)  1
exp(1)  1
Eq. 2.12
where e is the classification error rate, t is the feasibility threshold and m is the
tolerance margin. As e approach zero, p(e) slowly approaches its minimal value.
When e  t , a negative penalty or a small reward is received. When e  t but below
t  m , a small penalty between 0 and 1 is received. Feature subsets with error rate
exceed t  m receive comparatively higher penalty (larger than one), which means
they cannot compete with subsets at the next higher level (Siedlecki and Sklansky,
1989). The generalized procedure is shown in Figure 2.13.
Figure 2.13. Flowchart showing the generalized procedure used in a genetic
algorithm
65
(Adopted from (Kavzoglu and Mather, 2002)
A number of researches have compared the performance of genetic algorithm with
other feature selections techniques. Siedlecki and Sklansky (1989) showed that
genetic algorithm outperformed SFS, SBS and non-optimal variation branch and
bound in terms of both classification accuracy and computational complexity. Ferri
et al. (1994) compared genetic algorithm methods with SFS and SFFS on dataset
exceeds 300 dimensions. The performance of genetic algorithm was comparable to
SFFS on moderate feature size (about 20-30 dimensions), but degraded as
dimensionality increases. The SFFS showed better performance even on very high
dimensional problems. Kavzoglu and Mather (2002) find that GA offers better
solution than SFS algorithm in terms of the critical value of the evaluation criterion
considered, but not always in terms of classification performance. A major
constraint of GA is the arbitrary setting of the four main parameters as well as the
initial population of chromosomes (Kudo and Sklansky, 2000).
2.5.2.2. Evaluation criteria
Evaluation criteria measure the statistical separability of classes in a feature space.
Most of the authors broadly categorized the evaluation criteria into two types
including filter and wrapper (Kohavi and John, 1997, Han and Kamber, 2006) while
Liu and Yu (2005) proposed the third type, hybrid approach which is the
combination of the former two (Liu and Yu, 2005). Typically, the filter approaches
use an evaluation criterion independent of classification performance. It evaluates
the relevance of a feature or feature subset by investigating the inherent properties
of the training data without engaging any classifiers during the feature selection
process (Liang et al., 2008). The effects of the selected feature subsets on the
performance of classification algorithm are totally ignored (Bi et al., 2003). The
most commonly used filter methods are using
inter-class separability index
(Kavzoglu and Mather, 2002) such as distance measures, dependency measures,
consistency measures and information measures (Ben-Bassat, 1982, Almauallium
and Dietterich, 1991, Liu and Motoda, 1998b). The wrapper approaches evaluate
66
the goodness of feature subsets by the estimated accuracy of a chosen classifier.
The resultant selected feature subset is therefore classifier-specific (Liang et al.,
2008).
The filters are usually computationally more efficient than wrappers though the
wrappers are often offering better classification accuracy because it is optimized for
a specific learning algorithm (Han and Kamber, 2006). However, wrappers must be
trained from classifier to classifiers (Liang et al., 2008).
2.5.2.2.1.Distance measure
Distance measures are used to determine the separability within or between two
different classes and they are also regarded as divergence or discrimination
measures. In a feature selection problem, provided with two classes, the objective
is to find features that can maximally separate the two classes. Feature x is more
preferable to feature y if x stimulates a greater difference between the two-class
conditional probabilities than y (Liu and Yu, 2005). In other words, the larger the
distance is, the better the solution will be in terms of class separability.
For remote sensing applications, distance measures such as divergence,
transformed divergence, Bhattacharyya distance, and Jeffries-Matusita (J-M)
distance are extensively used (Goodenough et al., 1978, Swain and Davis, 1978,
Mausel et al., 1990, Richards, 1993, Aha and Bankert, 1996, Jensen, 1996, Dutra and
Huber, 1999, Mather, 1999b, Tso and Mather, 1999a, Bruzzone and Serpico, 2000).
The average Jeffries-Matusita (JM) distance is one of the common and well-known
distance measures for multi-class problems (Narendra and Fukunaga, 1977, Swain
and Davis, 1978, Bruzzone and Serpico, 2000). The reason is that JM distance
exhibits a saturating behavior with increasing class separability, which is analogy to
the error probability behavior (Bruzzone and Serpico, 2000, Kavzoglu and Mather,
2002). This makes JM distance more effective for interclass separability
measurement than divergence and Bhattacharyya distance (Swain and Davis, 1978)
67
though it tends to exaggerate low separability values while restraining high
separability values (Kavzoglu and Mather, 2002).
Although statistical distance measures have been widely and successfully applied
for feature selection problem, the increase in dimensionality tends to increase the
iterative cost exponentially. Therefore, suboptimal solution instead of optimal
solution is preferred when distance measures are used. Besides, a major challenge
for the distance measures is that they require the distribution of class to be
Gaussian in nature though the assumption is unrealistic for the majority of remotely
sensed data. The non-Gaussian data distribution would curtail the reliability of using
distance measures for feature selection.
2.5.2.2.2.Information measure
The amount of information contained in a band/ feature is an important factor for
band evaluation (Jiang et al., 2004). Information measures use information gain as
evaluation criterion to select feature. In terms of probability theory, information
gain from feature X is defined as the difference between the prior uncertainty and
expected posterior uncertainty after using X (Liu and Yu, 2005). Feature X is
more attractive than feature Y if information gain from X is larger than that from
Y.
According to Shannon’s information theory, entropy H (C ) measures uncertainty of
a discrete variable C . Given two variables X and C , the conditional entropy
H (C X ) calculates the uncertainty about C when X is known. The mutual
information I ( X ; C ) computes the certainty about C that is explained by X . Their
relationship can be expressed as (Peng et al., 2005)
H (C )  H (C X )  I ( X ; C )
Eq. 2.13
I ( X ; C )  H (C )  H (C X )
Eq. 2.14
and equivalently
68
The goal of a feature selection process for classification is to maximize values of
I ( X ; C ) with the smallest possible size of feature subsets (Chow and Huang, 2005).
With information gain after adding feature X , the uncertainty, H (C X ) is lower
than the prior uncertainty H (C ) . The larger the reduction of uncertainty is after
adding a feature, the greater the information gain or mutual information, i.e.
I ( X ; C ) . By comparing the information gain of features, features can be ranked in
descent order of I ( X ; C ) . The success of feature selection algorithm relies on the
amount of information about the output class is retained by the selected feature
subset (Kwak and Choi, 2002a).
In terms of vegetation classification problems, given the input spectral band X , the
objective of training is to minimize the uncertainty about prediction on class labels
C through maximizing I ( X ; C ) as much as possible. The prior entropy of C with c
discrete distinctive classes computed using probability density function p(c)
expressed as (Kwak and Choi, 2002a)
H (C )   p(c) ln p(c)
cC
Eq. 2.15
The features, i.e. spectral and textural bands, are usually continuous variables, the
differential entropy is defined as
H ( X )   p( x) ln p( x)dx
Eq. 2.16
The conditional entropy H (C X ) is


H (C X )    p( x)  p(c x) ln p(c x) dx
 cC

X
Eq. 2.17
Then, the mutual information I ( X ; C ) between X and C is defined as
I ( X ; C )    p(c, x) ln
cC x
p(c, x)
dx
P (c ) p ( x )
Eq. 2.18
69
From the above equations,
estimation of MI is not easy as it involves the
knowledge of the underlying probability density functions of the data as well as the
integration of these probability density functions (Chow and Huang, 2005). The
probability density functions estimation is usually simplified using histogram such as
the mutual information feature selection (MIFS) (Battiti, 1994) and its variants
(Kwak and Choi, 2002b), but they are only applicable in low-dimensional data space.
The accuracy of estimation decrease with increasing dimensionality, especially
under relatively small size of training data due to the sparse distribution of data
(Young et al., 1995, Yang and Moody, 1999). This will lead to larger error in
approximating the mutual information. Another solution is to use density
estimation method such as Parzen windows to estimate mutual information (Kwak
and Choi, 2002a). Chow and Huang (2005) improved the Parzen windows by
introducing a supervised data compression algorithm so as to raise the
computational efficiency of direct mutual information estimation for large data set.
Besides, when at least one of the variable is continuous, it is very difficult to
estimate pdfs p(x) , p( y ) and p( x, y) and to execute the integrations in Eq.2.18
(Kwak and Choi, 2002a, Peng et al., 2005). The continuous space is usually
discretized into several partitions, so that the computation of the entropy and
mutual information can follow the discrete cases (Kwak and Choi, 2002a, Guyon and
Elisseeff, 2003), i.e.,
I ( X ; C )   p(c, x) ln
xX cC
p(c, x)
P (c ) p ( x )
Eq. 2.19
However, the discretization of space would likely cause possible inherent errors.
The mutual information (MI) is considered as higher ordered statistics. The prime
advantage of MI is the robustness to noise and data transformation. Details can be
found in (Battiti, 1994). It has been used to identify significant features (Battiti,
1994, Kwak and Choi, 2002a, Kwak and Choi, 2002b, Peng et al., 2005). Given
mutual information is a good indicator of relevance between two random variables
(Cover and Thomas, 2006), Pent et al. (2005) derived mRMR criterion by applying
70
mutual information to combine maximum relevance and minimum redundancy to
select relevant features.
2.5.2.2.3.Classification error
The use of classification error is the wrapper-based feature selection approach.
Prior to feature selection, a classification algorithm should be specified. The
classification accuracy/ error rate tied to the chosen classifier acts as the evaluation
criterion to select best possible features in the feature space. The process requires a
relatively large number of training samples which is sometimes unachievable in
many classification problems (Kudo and Sklansky, 2000). When encountering small
sample problem, the correct classification rate is usually estimated by leave-one-out
or k-fold cross validation techniques (Kudo and Sklansky, 2000). The leave-one-out
technique requires a training sample to be left behind as independent testing
sample while the remaining samples are used to train the classifier. The leave-out
testing sample is then classified by the trained classifier. The process repeated until
all the training samples are independently tested and the classification accuracies
are summed up. If the size of training samples is large enough, the k-fold cross
validation technique where each of k non-overlapping comparable-sized subsets is
used to substitute the testing sample in the leave-one-out technique (Kudo and
Sklansky, 2000).
Features selection through wrapper approach is more pertinent to classification
problems though it is criticized that the selected features are classifier-specific and
the generalization is sometimes lost. However, the growing interests in powerful
machine-learning classifiers such as support vector machines which is designed to
provide robust classification to unseen data not only offers more choices, but also
capable to achieve robust feature selection results. Besides, the cross-validation
measure renders the approach applicable under small sample condition.
71
2.5.2.3. Feature Selection Stability
Apart from model performance, stability or robustness of the feature selection
process/ method is another important area of concern. Unstable feature selection
performance may lead to the selection of wrong features. Moreover, during the
feature selection process, problem of feature over-selection, similar to classifier
over-training sometimes occurs. Such over-fitting phenomenon can hamper the
generalization ability and thus may seriously affect its performance on independent
data, especially under small training sample and high dimensional situation (Raudys,
2006). In order to circumvent such problem, major works have been focused on
developing stability indices based on distance measures (Dunne et al., 2002,
Kalousis et al., 2007), correlation coefficients (Kalousis et al., 2007), consistency
(Kuncheva, 2007, Somol and Novovicova, 2008) and entropy (Křížek et al., 2007).
Feature selection stability is defined as “the robustness of feature preferences it
produces to differences in training sets drawn from the same generating
distribution (Somol et al., 2010a, p.16)”. Suppose Y  { f1 , f 2, , f Y } be the set of
all features and S  {S1 ,, S n } be a system of n feature subsets attained from n
runs
of
the
selected
feature
selection
S j  { f i i  1,, d j , f i  Y , d j {1,, Y }} ,
algorithm
such
that
for
stability quantifies the effect on
feature preference in system S  {S1 ,, S n } by varying the training sets drawn
from the same generating distribution n times.
Kalousis et al. (2007) proposed the similarity measure based on Tanimoto distance
which is used to measure the amount of overlap between two subsets of features
S i and S j . It is calculated as the size of intersection of two subsets divided by the
size of union of the two subsets, S i and S j and defined as follows (Duda et al.,
2001):
S k (Si , S j ) 
Si  S j
Si  S j
Eq. 2.20
where Sk (Si , S j ) is the Tanimoto index.
72
The Tanimoto index was extended to measure the similarity over all pairs of feature
subset within system S of feature subsets through the averaging process as follows
(Somol and Novovicova, 2008):
2 n1 n
ATI ( S ) 
  S K ( S i ,S j )
n(n  1) i 1 j i 1
Eq. 2.21
The Average Tanimoto Index ATI (S ) has value range 0,1 with 0 indicating zero
intersection between all pairs of subsets while 1 indicates all subsets within the
system S are identical.
Apart from pairwise similarity measure, consistency measures proposed by Somol
and Novovicova (2008) allow stability assessment for subsets of different sizes in an
overall manner. Suppose
F f be the frequency of occurrences of feature f in system S ;
X be the subset of Y appear in system S ; and
n
N be the number of all features in system S , i.e. N   S i ,
i 1
three consistency measures including consistency C (S ) , weighted consistency
CW (S ) , and relative weighted consistency CWrel (S , Y ) can be computed for
stability measurement. The simplest consistency C (S ) of system S is defined as
(Somol and Novovicova, 2008):
C (S ) 
1
X
F f  Fmin
F
f X
max
 Fmin

1
X

f X
Ff  1
n 1
Eq. 2.22
The calculation is based on the frequency of occurrence of each individual features
F f in S after n trials. The weighted consistency CW (S ) of system S simply
weighted the calculation by the frequency of occurrence of feature, F f divide by
the total number of all features in the system N , which is defined as (Somol and
Novovicova, 2008):
73
CW ( S ) 
w
f X
F f  Fmin
f
Fmax  Fmin

Ff
N

f X
Ff  1
n 1
Eq. 2.23
Both consistency indices have value range 0,1 , that is
0  C (S )  1 and 0  CW (S )  1,
Eq. 2.24
with 0 indicates all subsets are disjuncted from each other while 1 indicates all
subsets are identical in S . However, Somol and Novovicova (2008) pointed out that
both C (S ) and CW (S ) tend to be affected by the sizes of subsets in S related to
the size of Y . They are likely to over-estimate the consistency if the subset sizes in
S approach the total number of features Y and produce misleading stability
indicator. In order to get rid of this unnecessary behavior, the relative weighted
consistency CWrel (S , Y ) was developed (Somol and Novovicova, 2008). CWrel (S , Y )
of system S characterized by N , n and for given Y is defined as (Somol et al.,
2010a):
CWrel ( S , Y ) 
CW ( S )  CWmin ( N , n, Y )
CWmax ( N , n)  CWmin ( N , n, Y )
Eq. 2.25
where
CWmin ( N , n, Y ) is the minimal possible value of CW () but possibly greater than
zero that can be yielded in system S ;
CWmax ( N , n, Y ) is the maximal possible value of CW () but possibly smaller than
one that can be yielded in system S ; and
CWrel (S , Y )  CW (S ) for CWmax ( N , n)  CWmin ( N , n, Y ) .
Similar to the other two consistency measure, CWrel (S , Y ) has value range 0,1 , i.e.
0  CWrel (S )  1 . But the difference is relative weighted consistency not only tells
how much the selected subsets overlap, but also illustrates the magnitude of
randomness inherent in the feature selection process. In this regard, 0 indicates
74
totally random occurrence of features while 1 represents the most stable process
and outcomes (Somol and Novovicova, 2008).
2.5.3. Feature extraction
Feature extraction reduces the high dimensional feature space by producing new
features through combinations or transformations of the original feature set (Jain et
al., 2000) and with classification accuracy is the primary criterion for feature
extraction (Lee and Landgrebe, 1993b). The transformation can be either linear or
non-linear. Among the linear feature extraction methods, principal component
analysis and linear discriminant analysis are two representative algorithms.
Principal components transformation (PCT) also known as the Karhunen-Loéve
transformation in communication theory is the most widely used method of feature
extraction (Fukunaga, 1990, Jain et al., 2000, Landgrebe, 2001). It is completely an
unsupervised method as it disregard the class information of input data (Li et al.,
2008a). PCT transforms the original feature spaces orthogonally to a new space
comprising mutually uncorrelated features called principal components. The
principal components are ordered in descending sequence based on the
eigenvalues of covariance matrix. Usually, the first few components are selected as
they retain the largest variance of the original data. Consequently, the
dimensionality of the original space can be dramatically reduced but without loss of
much information (Chen, 2001). However, PCT is criticized to be sensitive to noise
(Hsu, 2007) and is designed for data representation rather than discrimination.
Linear discriminant analysis (LDA) also know as canonical analysis (Richards, 1993)
develops based on the Fisher’s linear discriminant analysis (Duda et al., 2001).
Unlike PCA, LDA make full use of the class labels for class discrimination. It aims at
minimizing the within-class covariance matrix while maximizing the between-class
covariance matrix with a view to achieve maximum possible class separability (Tu et
al., 1998, Hsu, 2007). The new feature space is transformed based on the
maximizing the ratio of between-class and the within-class covariance matrix (Hsu,
75
2007). LDA has been widely used in pattern classification (Schowengerdt, 1983, Lee
and Landgrebe, 1993b, Richards, 1993, Duda et al., 2001).
Generally, LDA can improve class separability more than PCA. However, there are
also drawbacks of LDA. The criticism to LDA as a feature extraction method is that
the maximum number of features it can generate is equal to the number of classes
minus one. Besides, when the mean values of classes are close to each other, the
feature extracted becomes unreliable. On the other hand, if the mean vector of a
particular class is very different from other classes, the between-class variancecovariance matrix will tends to bias this class resulting in ineffective features
(Tadjudin and Landgrebe, 1998). Moreover, the within-class covariance matrix is
always ill-conditioned for high dimensional data. It is singular if the number of
samples is less than the dimensionality (Chen, 2001). In other words, the size of
training samples has to be large enough in order to come up with reliable
estimations of within-class and between-class covariance matrices. In fact, when
hyperspectral data is considered, the criterion of sufficient training samples is
always not satisfied.
Other feature extraction methods are developed to reduce the dimensionality of
hyperspectral data such as feature extraction using neural network (Lowe and
Webb, 1991); non-linear multidimensional scaling (Jain et al., 2000); wavelet-based
feature extraction method (Hsu et al., 2002); matching pursuit (Hsu, 2007);
constrained maximum variance mapping (Li et al., 2008a).
Transformed feature may offer a better discriminative power than the selected
feature subset, but the physical meaning of the transformed feature may not be
easily interpreted (Bruzzone and Serpico, 2000, Jain et al., 2000). In terms of
vegetation classification, the feature extraction method may hinder the
understanding of the spectral response of vegetation species.
76
2.6. Biophysical Parameters Measurement and Estimation
Data acquired from remotely-sensed images are not limited to vegetation
classification; it also acts as an effective scrutiny instrument for ecological
monitoring. Wessman (1994b) pointed out that only if two criteria are satisfied can
ecosystem or biophysical monitoring be undertaken. The first is the ability to
identify the primary structural and physiological features of ecosystems, which are
directly associated with underlying processes and functioning while the second is
the capability of measuring these features remotely. Ecosystem functional changes
are often expressed or reflected in the canopy biochemistry which is significant for
study of productivity, nutrient cycle and vegetation stress (Curran, 1994). The
advancement of imaging spectrometry provides sufficient information on
vegetation reflectance as a function of wavelength, which in turn reveals details in
biophysical and biochemical variables of vegetation (Wessman, 1994a). Estimation
of biophysical parameters from spectral reflectance can be classified into two types:
1) empirical relationship based on spectral vegetation indices; and 2) physicallybased canopy radiation models (Treitz and Howarth, 1999, Kumar et al., 2001).
After the launch of Terra satellite in the end of 1999, the Moderate-resolution
Imaging Spectroradiometer (MODIS) and multi-angle imaging spectroradiometer
(MISR) provide an effective tool for monitoring of the biophysical characteristics of
land surface such as leaf area index (LAI), fraction of absorbed photosynthetically
active radiation (fAPAR), albedo and gross primary vegetation productivity at
regional and global scale. The advancement of MODIS and MISR demonstrated the
ability of the technology in land surveying such as ecosystem and vegetation
monitoring. Other biophysical parameters having significance in environmental
physical and ecological studies including leaf inclination angle distribution (LAD),
canopy closure (CC), and albedo are also studied with the aids of spectral data.
Among those, LAI and fAPAR are frequently used to depict the status of a vegetated
surface; as input into different kinds of models as well as variable of different
studies since they describe important functioning and energy absorption capacity of
vegetation canopy (Myneni et al., 1997, Myneni et al., 2002). The following sections
provided a thorough review on the biophysical parameters usually used for the
77
ecosystem monitoring. In order to understand the functioning of current ecosystem,
in situ measurement of these parameters is indispensable. Efficient measurements
of these parameters through field work were examined. Finally, modeling
approaches combining the in situ measurements and data from remote-sensed
image to come up with estimation over the area was reviewed. As LAI is the focus of
this study, particular efforts were related to the parameters throughout the
discussion.
2.6.1. Leaf Area Index (LAI)
Foliage is one of the dominant components in the canopy. Their distribution,
surface area and orientation form the characteristics of canopy structure, which
shows high adaptation to subtle variations in the environment through time
(Fournier et al., 2003). One of the most commonly-used and significant parameters
in quantifying biophysical characteristics of vegetation is leaf area index (LAI) which
is defined as the total one-sided area of all leaves in the canopy within a defined
region (Gong et al., 2003). The significance of LAI found on its relationship with the
energy, mass and gas exchange process such as rate of photosynthesis, respiration,
transpiration, carbon and nutrient cycle as well as the net primary production of
terrestrial ecosystem (Pierce and Running, 1988, Price and Bausch, 1995, Gong et al.,
2003, Lee et al., 2004, Pu and Gong, 2004). Apart from being an ecologically
important variable, LAI is also a consistent variable applicable for studies of various
scales (Fournier et al., 2003). The success of LAI estimation can actually model the
ecological process and predict the response of ecosystem (Green et al., 1997, Green
and Clark, 2000). Since LAI is functionally linked to spectral reflectance (Baret and
Guyot, 1991), field measured data accompanied with remote sensing techniques
become an effective tool in retrieving LAI over large areas, especially for the general
inaccessibility of mangrove areas. In recent years, LAI mapping has become subject
of major interest. Since vegetation indices are most sensitive to the green leaves in
the canopy, they have been widely used to map LAI at different spatial scale (Asner
et al., 2003).
78
2.6.2. Fraction of Absorbed Photosynthetically Active Radiation (fAPAR)
Photosynthetically active radiation (PAR) limits to the part of the solar spectrum
ranging from 400 to 700 nm (Goetz and Huemmrich, 2007). In this region, strong
chlorophyll absorption takes place for photosynthesis with peaks located at
430/660 nm and 455/640 nm for chlorophyll a and b respectively (Verdebout et al.,
1994, Wessman, 1994a, Kumar et al., 2001). As incident PAR (arriving at the top of
canopy) travels through the canopy layers to the ground, they are intercepted and
absorbed by canopy layers. Fractional photosynthetically active radiation (fPAR)
calculates the proportion of the incident PAR that is either intercepted (fIPAR) or
absorbed (fAPAR) (Goetz and Huemmrich, 2007). The calculations of fIPAR and
fAPAR are expressed in Equations 2.26 and 2.27 respectively.
fIPAR 
IPAR
PARINT AC
Eq. 2.26
fAPAR 
APAR
PARINT AC
Eq. 2.27
where PARINT AC is the amount of PAR incident at canopy top; IPAR and APAR is the
amount of intercepted PAR and absorbed PAR respectively.
Fraction of absorbed photosynthetically active radiation (fAPAR), by definition,
refers to the ‘measurement of the proportion of available radiation at
photosynthetically active wavelengths absorbed by canopy (Roberts, 2001, p.485)’.
It differentiates itself from fIPAR by taking into consideration the portion of PAR
that is reflected off the canopy top back to the atmosphere as well as the part of
PAR that bounced back from the soil or background materials to the canopy (Goetz
and Huemmrich, 2007). By taking the two definitions into account, Equation 2.27
can therefore further elaborated and rewritten as follows (Gower et al., 1999):
fAPAR 
[(PARINT AC - PARREF AC ) - (PARINT BC - PARREF BC )]
PARINT AC
Eq. 2.28
where PARREF AC is the amount of PAR reflected off the canopy top back to the
atmosphere; PARINT BC is the quantity of incident PAR went down the canopy; and
79
PARREF
BC
is the amount of incident PAR bounced back from the background
materials. Therefore, the term (PARINT AC – PARREF AC) depicts the net amount of PAR
transferred through the canopy while (PARINT BC – PARREF BC) represents the net
quantity of PAR lost to background materials. Taking the difference of the two
represents the net amount absorbed by the canopy layers. Although difference
existed in fPAR and fAPAR as explained, cautions have to be paid when reading
literatures since the two terminologies are always used interchangeably as found in
the literatures (Goetz and Huemmrich, 2007).
Like LAI, fAPAR is a mostly studied biophysical parameter because it is a significant
indicator of energy absorption capacity of canopy (Fensholt et al., 2004). When
coupled with photosynthetic efficiency or light use efficiency, the assessment of net
primary production of vegetation can be conducted (Bicheron and Leroy, 1999).
Besides, fAPAR is an important input parameter to study hydrology and energy
budget of vegetated land surface (Gao et al., 2000).
Direct in situ measurement of fAPAR is a challenging job because the amount of PAR
absorbed varies diurnally, seasonally and with sky conditions (Gower et al., 1999,
Goetz and Huemmrich, 2007). Figure 2.14 shows the diurnal variation of fAPAR
within a spatially uniform canopy under different LAI. In order to have an accurate
computation of fAPAR, it requires instantaneous monitoring of APAR continuously
(Goetz and Huemmrich, 2007). Instead of taking direct measure, fAPAR is
sometimes estimated from LAI measurement (Gower et al., 1999). Besides, many
studies used IPAR instead of APAR because of easier measurement of IPAR (Gower
et al., 1999). Although the discrepancy between IPAR and APAR is dependent on a
number of factors including, canopy closure, density, composition, reflectance and
background materials (Goetz and Huemmrich, 2007), the IPAR and APAR will
produce similar results when the canopy is complete and dense (Goetz and
Huemmrich, 2007) as shown in Figure 2.14.
80
a
b
Figure 2.14. a) Diurnal variation of fAPAR within a spatial uniform canopy; b)
Photosynthetically active radiation (PAR) reflection, transmittance and absorption
characteristics varies across LAI
(Adopted from Goetz and Huemmrich, 2007)
2.6.3. In-situ Leaf Area Index Measurement
Accurate mapping of biophysical parameters through such technique is dependent
on the provision of reliable ground measured data. Several methodologies have
been proposed for in situ LAI measurement and they are broadly classified into two
categories – direct and indirect methods.
2.6.3.1. Direct and Indirect Methods
Direct methods refer to the direct measurement of size, area, shape and orientation
of plant organs such as leaves, branches manually (Norman and Campbell, 1989).
The mostly employed direct methods include the area harvest, litter collection,
allometry, and inclined point quadrats.
The area harvest approach involves destructive collection of leaves freshly cut down
from mangrove or litter collection on the forest floor within a designed quadrat
(Clough et al., 1997, Jensen and Binford, 2004). The sampled leaves are detached
81
from branches and the leaf areas are measured with a planimeter. The measured
leaves are then oven dried and established the leaf area to dry mass ratio with the
measured leaf areas. However, since the leaf area to dry mass ratio is species- and
site- specific, a sound leave sampling scheme should be established (Breda, 2003).
Another way to measurement LAI is through allometry or allometric equations
which are developed to relate the biophysical parameters of interest (dependent
variable) to direct measurement of biomass or stem diameter (independent
variable) through destructive sampling of trees (Chen et al., 1997, Gower et al.,
1999). However, an omnipotent equation does not exist since the abiotic and
abiotic factors affecting the allometric relations are species- and site-specific as well
(Chen et al., 1997, Clough et al., 1997, Comley and McGuinness, 2005). Using the
general allometric equations to estimate LAI of mangrove stands will result in errors
(Gower et al., 1999).
The inclined point quadrat method requires the insertion of a thin and sharp probe
into the canopy at a certain angle (known azimuth and inclination angle) and the
number of times the point contacts with leaves, stems and branches are counted
manually or with the aid of automatic contact detection system (Norman and
Campbell, 1989). However, such method is impractical for tall and dense canopies
(Chen et al., 1997).
Direct measurement of canopy structure is regarded as the most accurate method
(Jonckheere et al., 2004). For example, the development of allometric relationship
allows separate computation of areas of various plant organs and is more accurate
than using clumping index (Gower and Norman, 1990, Fassnacht et al., 1994, Chen
and Cihlar, 1995). However, they are criticized for being exceptionally timeconsuming, tedious and labour-intensive (Welles and Norman, 1991b, Jonckheere
et al., 2004, Weiss et al., 2004). Besides, apart from litter collection method, the
area harvesting and the development of allometric relationship is destructive to
canopy, which may not be authorized in protected areas governed by law (Norman
and Campbell, 1989, Chen et al., 1997). In addition, direct LAI measurement is
neither practical for large-scale implementation nor capable of achieving the goal of
82
long-term monitoring of leaf area development through space and time (Chason et
al., 1991, Jonckheere et al., 2004). Nevertheless, the measurement using direct
methods can be used to calibrate and evaluate the results from the indirect
methods (Breda, 2003, Jonckheere et al., 2004).
Unfavorable constraints, like high cost, disturbance to canopy and spatial sampling
difficulties posed by direct methods lead to active research on alternative methods
for LAI field measurement. Since the sixties, many studies put forward indirect
methods for canopy structure measurements (Welles and Norman, 1991b, Welles
and Cohen, 1996, Jonckheere et al., 2004, Weiss et al., 2004). The notion of indirect
method is to infer biophysical parameters from observation of another variable
(Jonckheere et al., 2004). According to Jonckheere et al. (2004), indirect methods of
in situ LAI measurement can be segregated into two types – indirect contact LAI
measurements and indirect non-contact measurements. Indirect contact LAI
measurement methods (while some authors named it as semi-indirect methods)
include inclined point quadrat and allometric techniques while indirect non-contact
LAI measurements methods refer mainly to optical method which measure light
transmission through canopies (Jonckheere et al., 2004). In most of the studies, the
indirect method is often referred to the latter category which concerns the use of
optical method for LAI estimation.
As aforementioned, canopy structure affects the characteristics of radiation
interception. A beam of radiation traveling inside a canopy is either blocked or
penetrated without any contact with the vegetative elements, which is totally
dependent on canopy structure or the arrangement of vegetative components. The
indirect non-contact LAI measurement method make use of the tight coupling
between canopy structure and radiation penetration (Welles and Norman, 1991b)
to infer biophysical parameters such as LAI from measurements of radiation
transmission through the canopy, combined with an appropriate radiative transfer
theory (Ross, 1981). In other words, simple measurement of radiative environment
above or from within the canopy can be used to estimate various biophysical or
structural parameters (Ross, 1981, Norman and Campbell, 1989, Welles, 1990). In
general, radiative transfer models are used to predict some elements of radiative
83
environment at a certain location within or outside the canopy dependent on the
canopy structural parameters using local spectral properties of incident radiation as
inputs. The indirect approach inverts the model to resolve the probable canopy
structure that caused the measured radiation below the canopy (Welles, 1990).
Basically, the optical methods employ statistical and probabilistic approach to
estimate the leaf distribution and arrangement in the canopy (Breda, 2003).
Norman and Campbell (1989) reviewed three groups of indirect methods including
spectral methods, gap fraction methods and bidirectional reflectance distribution
function (BRDF) with their inversion complexity ranging from trivial to almost
impossible. Among those, the gap-fraction method is relatively popular and wellresearched.
Since indirect methods rely upon the optical measurement of radiation above and
below the canopy, numerous ground-based optical devices and methods have been
developed to capture the characteristics of radiative environment and therefore to
estimate the canopy structure. They were all designed to use from ground level
looking upwards to the sky below or above the canopy and were classified into
three types. The first type takes measurement under diffuse light transmission or
using hemispherical view to capture the gaps of canopies and representative
techniques
includes LAI-2000
Plant
Canopy Analyzer
and hemispherical
photographic technique. The second type measures direct solar irradiance or
sunflecks at known solar angles and the techniques under this category includes
Quantum sensor, DEMON, and TRAC. The third type measures the vertical
distribution of canopy elements using the optical point quadrat method (Pu et al.,
2005).
Compared with direct methods, indirect methods are non-destructive to vegetation.
Besides, they provide fast and simple field measurement, which can tremendously
reduce the time for laborious fieldwork on one hand (Norman and Campbell, 1989)
and allow the sampling of larger area on the other (Jonckheere et al., 2004). That
constitutes the reason of surging importance and research in indirect methods.
Besides, many researches have been carried out to compare the LAI measurement
from indirect and direct measurements for crops (Brenner et al., 1995, Levy and
84
Jarvis, 1999) and forest (Chason et al., 1991, Smith et al., 1991, Fassnacht et al.,
1994, Dufrêne and Bréda, 1995, Chen et al., 1997, Küβner and Mosandl, 2000).
Owing to the popularity of gap fraction inversion method, the review that follows
will look into the theory of the method.
2.6.3.2. LAI Estimation through Gap Fraction Inversion
When optical methods are used to estimate LAI, they are all based on contact
frequency or its complement gap fraction (Ross, 1981), which are basically statistical
and probabilistic approaches (Breda, 2003). As a ray of radiation penetrates through
the canopy, there is a certain chance that it will be intercepted by different
vegetative elements. The probability of such interception is referred to contact
frequency (Weiss et al., 2004). On the contrary, gap frequency is the likelihood of
non-interception by the vegetative elements until it get to a reference level, usually
the ground level (Weiss et al., 2004). Gap fraction refers to the fraction of view from
a particular direction, defined by azimuth and zenith angles, that is not obstructed
by foliage or canopy elements (Ross, 1981, Welles, 1990) and is calculated by
considering different directions (Jonckheere et al., 2005). The measurement of gap
fraction is comparable to the measurement of transmittance at ground level (Weiss
et al., 2004). When coupled with appropriate radiative transfer model, LAI can be
computed from the inversion of the exponential expression of the gap fraction
(Breda, 2003).
The estimation of LAI from gap fraction has two very important characteristics. First,
it is assumed that the foliage or vegetative elements are randomly distributed in
terms of the azimuth angle (Chen et al., 1997). Second, the gap fraction-based
approach does not differentiate the photosynthetically active foliage from other
vegetative elements such as stems, branches, flowers or fruits (Jonckheere et al.,
2004) and also dead plant tissues. Because of the latter characteristics, alternative
nomenclatures for leaf area index such as “Plant Area Index (PAI)” (Neumann et al.,
1989), “Vegetation Area Index (VAI)” (Fassnacht et al., 1994), “Effective LAI (Le)”
85
(Chen and Black, 1992) have been proposed. These assumptions and characteristics
also posed constraints on LAI retrieval which will be discussed later.
The probability of non-interception or gap fraction T ( ,  ) is assumed to be
dependent upon the light incident angle , path length S ( ,  ) , foliage density (  )
and foliage orientation (Chen et al., 1991, Welles and Norman, 1991b). The
direction of an incident ray is described by zenith angle (  ) and azimuth angle (  ).
The path length is the distance that the ray traveled through the canopy in direction
( ,  ) . Foliage density (  ) describes the total leaf surface area per cubic meter
canopy, which is related to LAI and canopy height ( z ). The foliage orientation is
described by the leaf angle distribution (LAD) function in the G( ,  ) which is the
distribution of fractional foliage areas projected toward direction ( ,  ) (Welles,
1990, Welles and Norman, 1991b). Foliage orientation affects the nature of
radiation interception with direction and therefore the light extinction coefficient
(Welles, 1990). For examples, the probability of radiation interception is likely to be
greater when the leaves are facing the direction of the radiation than when they are
distributed perpendicular to the ray direction (Welles and Norman, 1991b).
Supposed g jk is the fraction of leaf area with leaf inclination angle (  ) class j and
azimuth angle (  ) class k (Norman and Campbell, 1989).
Nj Nk
 g
j 1 k 1
jk
1
Eq. 2.29
Equation 2.29 explains that the summation of the fractional leaf area of all the angle
classes should equal one. However, since the leaves of most canopies are about
azimuthal symmetry in nature (Ross, 1981), the leaf inclination angle class becomes
the only measured variable if the azimuthal symmetry is assumed. The equation is
simplified into:
Nj
g
j 1
j.
1
Eq. 2.30
86
On account of simplicity, hypothetical leaf orientation distribution, for instances,
spherical, extremophile, plagiophile, and planophile were sometimes assumed with
the spherical leaf orientation distribution as the most popular one because the
fractional area projected in any direction is 0.5 (Welles, 1990).
Follows Welles and Norman (1991), the gap fraction is the exponential of the
factors expressed in Equation 2.31,
T ( ,  )  eG ( , ) S ( , )
Eq. 2.31
Taking the natural logarithm, the equation is rearranged as follows:
G( )   
ln(T ( ))
 k ( )
S ( )
Eq. 2.32
Since most of the optical techniques, like the LAI-2000 Plant Canopy Analyzer,
averages the measurements over azimuth, the equation drops the azimuthal angle
specifier (  ). k ( ) is the contact frequency which is defined by Miller (1967) as the
average number of contacts per unit length of travel that a probe (a beam) would
make passing through the canopy at zenith angle  . The foliage density (  ) is then
calculated as:

 ln(T ( ))
sin d
S
(

)
0
2
 2
Eq. 2.33
Leaf area index (LAI) is related to foliage density (  ) by canopy height z , and path
length is related to the zenith angle (  ) and canopy height z . The two relationships
are expressed by Equations 2.34 and 2.35 respectively as follows (Welles and
Norman, 1991b):
L  z
Eq. 2.34
87
S ( ) 
z
cos 
Eq. 2.35
Substituting the two equations above into Equation 2.33 formulates

2
L  2   ln(T ( )) cos  sin d
Eq. 2.36
0
When gap fractions or probability of non-interception for indirect LAI estimation, a
few implications can be drawn from the sets of equations. First, the relationship
between gap fraction and LAI is logarithmic in nature. This implies that small errors
in gap fraction or canopy openness will produce large differences in LAI (Hale and
Edwards, 2002). Second, the gap fraction of a canopy is measured as a function of
zenith angle or towards the sun direction (Norman and Campbell, 1989). The former
refers to the angular measurement of gap fractions while the later one is usually
refers to the measurement of fractional sunfleck area (Welles, 1990, Welles and
Norman, 1991b). Third, with increasing zenith angle, the intensity of radiance would
decrease due to the increase in path length through the canopy as indicated by
cos  (Chen et al., 1991). Fourth, the weight, sin  in the equation indicates that
the gap fraction at larger zenith angle is more prominent than those at smaller
zenith angle (Chen et al., 1997). Fifth, the multiple scattering by foliage within the
canopy exists in reality though it is assumed that the leaves block all the light. The
multiple scattering effect is more intense at large zenith angles than at small zenith
angle (Chen et al., 1997). Sixth, with the consideration of the fractional distribution
of all leaf angle classes from 0 to 
2
, i.e. indicated by the integration function in
Equations 2.29 and 2.30, G( ) becomes unity and omitted in Equation 5 and 8
(Chen et al., 1997). Seventh, the range of zenith angles used in the equation is
between 0 and 
2
, the multiplication of the integration equation by two means
the whole canopy (180-degree) is considered with the assumption that the canopy
is symmetrical.
88
2.6.3.3. Gap Fraction Ground Measurement
Ground-based in situ gap fraction measurement is conducted beneath vegetation
canopy with sensors usually look skyward (Norman and Campbell, 1989, Pearcy,
1989, Welles, 1990, Welles and Norman, 1991b, Welles and Cohen, 1996, Gower et
al., 1999). The LAI-2000 Plant Canopy Analyzer (LAI-2000) and hemispherical
photography are two most frequently used techniques under diffuse light
environment.
2.6.3.3.1.LAI-2000 Plant Canopy Analyzer
LAI-2000 developed by Li-Cor is designed for real time LAI estimation of plant
canopies (Lincoln, NE) based on gap fraction measurement. By inverting the BeerLambert’s Law using the Miller’s (1967) theorem, the model computes leaf area
based on gap fraction method which calculates the probability that a light ray will
pass uninterruptedly through a plant canopy from a given zenith angle (Hicks and
Lascano, 1995). Figure 2.15 shows the structure of LAI-2000. Physically, the optical
sensor unit is installed with a fisheye lens to project the light ray hemispherically
onto silicon detectors arranged in five concentric rings (Welles and Norman, 1991b).
LAI-2000 measures diffuse radiation simultaneously in five 15-degree concentric
rings with mid-point zenith angles of 7º, 23º, 38º, 53º, and 68º. With the built-in
calibrated programs, gap fraction and LAI for five distinct zenith angle ranges are
computed in real time. In order to minimize the contribution of light scattering by
foliage, an optical filter is installed to restrict sensed radiation to wavelength
between 380-490nm (ultraviolet to blue) where leaf transmission is low (Welles,
1990, Welles and Norman, 1991b, Dufrêne and Bréda, 1995, Chen et al., 1997).
89
Figure 2.15. The optical sensor unit of LAI-2000 Plant Canopy Analyzer
(Adopted from LI-COR, 1992)
LAI estimation using LAI-2000 is based on the following assumptions (Welles and
Norman, 1991b, Hicks and Lascano, 1995, Fournier et al., 2003, Jonckheere et al.,
2004). (1) The foliage blocks or absorbs all the light with wavelength of 490nm or
below. (2) The foliage elements are small compared to the projected area of each
ring. (3) The foliage is randomly distributed in terms of azimuth angle. (4) The angle
of inclination of foliage is randomly distributed with respect to azimuth. (5) The sky
brightness is uniform. (6) The measurement of radiation beneath the canopy does
not include transmission through leaf or reflected radiation (Dufrêne and Bréda,
1995).
Operationally, LAI-2000 requires taking an above canopy reading and one or more
below canopy readings. The above-canopy reading acts as reference with sensor
pointing upward at the sky representing incident light value acquired in the open
sky. The below-canopy readings of each ring are normalized by the corresponding
ones taken above the canopy to estimate the gap fraction at five zenith angles
(Welles, 1990). Other related properties of the canopy such as the geometrical
90
structure, absorption properties, extinction coefficient and leaf angle distribution
are simultaneously obtained after the measurement (Cournac et al., 2002).
During the measurement, three operation modes can be chosen depending on field
condition. Normally, for short vegetation species such as agricultural crops or
brushes, a single sensor alternatively taking above- and below-canopy readings
provides the simplest solution. Another practice makes use of cross-calibrated
sensors connected to same control box with one sensor taking above-canopy
reading while another measuring intercepted radiation below-canopy. When tall
canopies are encountered and simultaneous above-canopy measurement is hard to
acquire, two cross-calibrated sensors and two synchronized control boxes are
deployed. This operation mode requires one set of control box and sensor being
placed in elevated ground free from any obstacles to record the above canopy
reference reading continuously. Another set is carried under the canopy. The data
logged in two control boxes are then calibrated with each other in the laboratory.
However, the cost of approach is relatively higher and it abandons the advantage of
real-time LAI computation (Breda, 2003). Besides, errors are likely induced when
local sky condition varied between the measuring sites (Cournac et al., 2002).
The Plant Canopy Analyser, LAI-2000 has been widely applied to biomass
monitoring for agricultural crops (Hicks and Lascano, 1995), coniferous stands
(Gower and Norman, 1991, Deblonde et al., 1994) and deciduous stands (Dufrêne
and Bréda, 1995, Cutini et al., 1998, Le Dantec et al., 2000). It is always used as an
indirect instrument for LAI estimation owing to its lightweight and waterproof
properties which is ideal for challenging environment such as mangrove areas
(Kovacs et al., 2004). Besides, the real time calculation and display of LAI, minimum
amount of laboratory work and good repeatability are also regarded as prime
advantages (Chen et al., 1997). However, there are also some potential weaknesses.
One of the practical limitations is the requirement of above canopy reading, which
is sometimes hard to acquire efficiently and accurately (Welles, 1990, Cournac et al.,
2002, Fournier et al., 2003). Another disadvantage is the relatively coarse
measurement of gap fraction with zenith angles grouped into the five annular rings.
91
The arbitrary integration of data in terms of sensor field of view hinders further
detailed spatial analyses (Jonckheere et al., 2004).
2.6.3.3.2.Hemispherical Photography
Hemispherical photography is a close-range remote imagery technique for plant
canopies study using photographs obtained through extreme wide-angle lens with
field of view equals to 180-degree either placed beneath the canopy looking
towards the zenith or above the canopy looking downward (Chen et al., 1997, Hale
and Edwards, 2002, Jonckheere et al., 2004). The technique originally designed for
cloud formation study (Hill, 1924) has been discovered by ecologists to capture the
characteristics of light environment under forest canopies.
Hemispherical photography has been widely adopted for indirect measurement and
characterization of canopy architecture such as leaf area, leaf angle distribution,
canopy openness, and light transmittance by projecting the hemisphere directions
onto a plane (Rich, 1990). Figure 2.16 shows the concept of the hemispherical
projection. Since the hemispherical lens captured the entire sky, the zenith is in the
centre while the horizon is at the edges in the circular image indicating the
maximum field of view (FOV). The sky direction is represented by angular
coordinates comprised of a unique zenith angle  (the angle between the zenith
and the sky direction) and a unique azimuth angle  . With the camera looking
upward, angle  is measured counterclockwise with respect to north towards the
east direction (Rich, 1990). The most commonly used fisheye lens geometry is the
polar or equi-angular/ equilinear projection which is expressed as
d


r  max
Eq. 2.37
where r is the radius of the projected circle; d is the distance from any specific
point to the center of circle. Under equi-angular projection, the zenith angle  is
projected directly proportional to the distance along a radial axis (Rich, 1990,
Jonckheere et al., 2005).
92
Image plane
Figure 2.16. The equilinear fisheye lens projection
(Modified from Régent Instrument Inc, 2008)
However, while real lenses do not possess the perfect linear projection property,
lens calibration is an essential step to ensure that the relationship between zenith
angles and calculated distances is exact and precise (Régent Instruments Inc., 2008).
In terms of azimuthal projection, an object locating at a  from the north will be
projected onto the image at the same azimuth angle a  relative the north direction.
As lens distortion in angular projection is insignificant, no compensation is
necessary.
Conventionally, film-based hemispherical photography, either using black-andwhite or color films (negatives or diapositives) is used to capture canopy
characteristics. However, the tedious procedures from photo scanning to LAI
computation rendered the technique being forsaken (Breda, 2003). The
development and widespread use of high-resolution digital cameras as well as
advancement in image processing software revive the interests in using the
technique for measurement of canopy structure. Digital optical imaging devices
93
recorded images using the couple charge device (CCD) which is a light-sensitive
integrated circuit placed at focal plane of the camera system. The CCD is assumed
to respond to light intensity in a linear manner (Leblanc et al., 2005). Film and
digital photographic techniques have been compared in characterizing the canopy
structure with different degrees of success (Frazer et al., 2001, Hale and Edwards,
2002). Frazer et al. (2001) found generally higher estimates of canopy openness and
correspondingly lower estimates of LAI with digital method when compared with
film. Hale and Edwards (2002) found that the film and digital hemispherical
photography produced comparable results in terms of LAI estimation. In fact, the
digital method save the monotonous photo scanning procedure and provide
comparable results.
When considering digital photography method, parameters such as image format,
quality, and dimension/ size were examined for precise measurement and
extraction of biophysical variables (Frazer et al., 2001, Inoue et al., 2004). For
instances, Inoue et al. (2004) examined the effects of different image quality, size
and camera type on estimating gap fraction and canopy openness. They found that
there are no significant differences in the estimates with different image qualities
but the estimates with different image sizes had significant effect on the estimation.
Frazer et al. (2001) found limited difference between LAI retrieved from TIFF and
JPEG digital fisheye photographs and concluded that the file compression ratio of
JPEG should be at least 1/4.
Apart from image parameters, procedures of photo acquisition, processing and
analysis have been extensively researched to come up with practical guidelines for
better and accurate canopy structure extraction. During image acquisition, the
ultimate goal is to attain high contrast photographs from which the canopy
elements can be effectively distinguished from the sky (Rich, 1990). The quality of
image is governed by factors including sky condition, lens calibration and the
exposure setting of camera. Identical to LAI-2000 PCA, the ideal condition for image
capture is under uniformly overcast sky or time period before sunrise or after
sunset, when diffuse light transmission or indirect site factors dominated (Chen et
al., 1991, Whitmore et al., 1993, Hale and Edwards, 2002). Photographs taken
94
under direct sunlight tend to produce strong brightness variation and glare effect,
which should be avoided. Precise lens calibration ensures correct measurement of
gap area and distribution (Herbert, 1986). The purpose of lens calibration is to
correct angular lens distortion of the fisheye converter as well as the view angle so
as to minimize the deviation from the theoretical projection and the resultant
errors (Jonckheere et al., 2005). Besides, significant measurement errors are always
caused by inappropriate camera exposure settings which are governed by two
parameters – aperture and shutter speed (Chen et al., 1991, Whitford et al., 1995,
Jonckheere et al., 2005). Previous works showed that underexposed photographs
tend to show better contrast between sky and canopy when compared with the
correctly automatic exposed ones (Chen et al., 1991). Chen et al. (1991) tested the
relationship between aperture value - f-stop, and corresponding LAI estimation and
compared with those determined by LAI-2000 PCA. They found that the level of
underexposure for correct LAI estimation is related to the proportion of the sky and
canopy under the hemispherical angle of view. When canopy is more open, the
amount of underexposure should be lower. Sometimes, exposure blanketing
technique under which several photographs with different exposure (one f-stop
above and one f-stop below the measured metered reading) is recommended (Rich,
1990). Fast shutter speed can freeze the motion of foliage under windy condition.
The minimum shutter speed of 1/125 second is suggested (Rich, 1990). Filters were
sometimes applied to enhance the contrast between canopy and sky either by
installing into the camera or using filter wheels (Rich, 1990, Chen et al., 1991). For
instances, blue filter should be mounted during sunny days while red filter should
be used during overcast days (Rich, 1990). However, Chen et al (1991) found
insignificant improvement in photographic results though a blue filter is used.
Photo processing and analysis involves determination of a brightness threshold that
can maximally distinguish canopy elements from sky background. A binary image
with vegetative components represent by grey level of 0 (black), whereas the sky is
assigned grey level of 256 (white) in an 8-bit photo. Gap fraction is then extracted as
the proportion of sky pixels to the total amount of pixels in the image. However,
effective threshold operation can be done only if the grey-levels of the vegetative
95
pixels and sky pixels are substantially different from each other (Sezgin and Sankur,
2004). And this reinforces the significance of quality photo capture. The
thresholding process is guided by analysis of grey-level histogram, which can either
be interactive or automatic.
Interactively, the analyst toggles the threshold
manually based on the histogram and observe the effect on the photo.
Unfortunately, the method is criticized for two reasons. First, appropriate value is
less apparent when uni- or multi-modal appears in the histogram. Under these
circumstances, consistent and repeatable of threshold selection is profoundly
subject to the judgment process and inter-operator variation (Jonckheere et al.,
2005). A well-trained analyst is therefore needed. Second, it becomes impractical
when a large numbers of images are to be dealt with simultaneously while factors
such as noise, inadequate contrast can be complicated factors that curtailed the
accuracy of threshold decision (Jonckheere et al., 2005). Automatic thresholding is
an alternative to interactive thresholding, which provides objective measure of
canopy parameters. Jonchkheere et al. (2005) proposed six automatic thresholding
method including histogram shape-based method, clustering-based thresholding
method, entropy-based method, object attribute-based method, spatial method,
and local method.
Digital photos enables instant preview of image quality in field, which in turn allows
optimal exposure setting subject to the chance in sky condition. Besides, it
circumvents the time and expense on traditional photographic film development
and scanning (Frazer et al., 2001). Moreover, instead of using pre-defined zenith
angle ranges, the analysts are given autonomy to define the zenith angles to be
analyzed (Chen et al., 1997). Compared with LAI-2000 that simply shows the
computation results, another merits of hemispherical photography is permanent
records of canopy geometry and structure (Rich, 1990). Furthermore, as no above
canopy reading is required, the method is especially suitable for tall canopy and
highly-constrained environment such as mangrove forest where fast and efficient
field work practice is crucial. Last but not the least, the time required for set up and
acquisition of each photograph is comparatively minimal, which suggest that many
samples can be obtained.
96
The main criticisms or challenges of hemispherical photography are divided into
hardware- and processing-specific. The hardware specific problems relates to the
inherent deficiency of digital camera system. The spatial and spectral qualities of
images acquired with digital cameras are affected by the errors related to the CCD
detector and supporting electronics, which further induced errors into LAI
estimation. Frazer et al. (2001) reviewed some of these errors including image
sharpness, chromatic aberration, and blooming. The image sharpness of digital
camera is comparatively poor than that of film camera. It deteriorates with
increased radial distance away from the zenith centre and most noticeable at zenith
angles larger than 45º (Frazer et al., 2001). The blurring effect curtails the ability to
define fine canopy structure in the image. Chromatic aberration is found in all
optical systems that results from the split of the light spectrum into separate
wavelengths due to the imperfection of lens optics (Leblanc et al., 2005). The
multiple color images or so-called halos are created concurrently with each spatially
separated by horizontal distance and size. The former is longitudinal chromatic
aberration while the latter refers to lateral chromatic aberration and both cause
color blur (Frazer et al., 2001). The effect of color blur introduces significant spectral
and spatial biases under varied sky conditions and curtails the color contrast
between canopy components and sky, which render accurate and consistent edge
detection of leaves and branches difficult (Chen et al., 1991). Besides, the artifacts
introduce substantial amount of mixed pixels in the digital images that modified the
shape and size of canopy gaps. Very small canopy gaps are lost, as they look darker
than other larger gaps. The sizes of large canopy gaps tend to increase and their
shapes are simplified (Frazer et al., 2001). The thresholding process can be
significantly affected by chromatic aberration. However, the negative effects can be
minimized when digital photos were taken under overcast sky. Finally, blooming
happens when a CCD element is saturated and spilled over to neighbouring
elements on the array. The problem is more severe near the zenith in sunlit
conditions or when the sky is bright during overcast days (Leblanc et al., 2005).
The process-specific problems mainly relates to the selection of optimal brightness
threshold to distinguish vegetative components from sky (Jonckheere et al., 2004).
97
Accuracy, reliability and consistency are important evaluation criteria determining
the success of thresholding. Problems such as uneven exposure, penumbral effect
and mixed pixels in the edge of foliage and sky results in misclassification of high
reflection foliage surface as sky openings or relatively dark sky regions as leaves
(Rich, 1990, Welles, 1990). Chen et al. (1997) found consistent overexposure at
large zenith angles (larger gap fraction) while underexposure at low zenith angle
(smaller gap fraction) by making comparison with LAI-2000 PCA and regarded this as
the inherent problem of hemispherical photography. Besides, small gaps tend to be
under-represented while large openings tend to be over-represented (Rich, 1990).
However, some argued that the choice of threshold value becomes less critical for
higher resolution digital camera because the problem of mixed pixel is reduced
when compared with low-resolution camera (Jonckheere et al., 2004).
The errors of gap fraction measurement can be minimized by considering the site
condition carefully and adopting proper post-processing procedures. For instances,
measurement under direct sunlight should be avoid because the increase in
transmittance of light and foliage scattering leads to an increase in below canopy
reading and therefore render LAI being underestimated (Hicks and Lascano, 1995).
It is very likely that canopies consist of dense and sparse regions in a sampling plot
instead of with ideally equal gap size. The sensor should not be set to detect the
two types of regions under the same below-canopy reading since the linear
averages of azimuthal fields of view by detectors will underestimate LAI (Hicks and
Lascano, 1995). Unfavourable field conditions such as small sampling plots, very
clumped canopies, uneven gap distribution, undesired objects or strong direct
sunlight can be masked by using snap-on view restrictors with a view to narrow the
azimuthal field of view (Welles, 1990) and ultimately reduce errors (Welles and
Norman, 1991a). Sometimes, light scattering at low sun angles are criticized for
increase diffuse radiation beneath the canopy. Besides, the LAI measured at larger
zenith angles are mainly contribute from stems when leaves are mostly distributed
in the upper part of the canopy (LI-COR, 1992). Hence, the outermost rings are
sometimes masked out either using view restrictors during field measurement or
are limited from LAI calculation during processing to minimize this error (Chasen et
98
al., 1991, Chen and Black, 1991). For instances, Bréda (2003) found that best
agreement between ground-truth LAI and LAI-2000 measured LAI are obtained
when the two outer zenith rings are negated from computation. Cutini et al. (1998)
found increased correlation between field LAI and LAI computed from LAI-2000 as
well as lower LAI underestimation after omitting the fifth sky sector from
calculation.
The indirect gap fraction inversion method offers a fast, practical, theoreticallysound and non-destructive solution for LAI estimation. However, regardless of
instruments used, the method suffers from a few constraints that affect the result
of estimation. First, the method fails to differentiate live and dead tissue and
therefore cannot separate out photosynthetically active LAI (Welles and Norman,
1991a). Furthermore, despite using the term leaf area index, the actual radiation
interception is not limited to foliage, but also by all opaque objects, such as stems,
branches and fruit (Welles and Norman, 1991a) though it is argued that the effect
of branches is negligible for a full leaved canopy (Kucharik et al., 1998). Moreover,
accurate determination of direct and indirect sky factors is the foundation of light
penetration model. The assumption of random distribution of canopy elements
tends to induce errors during the model inversion process (Rich, 1990). As noted by
Clough et al. (1997), the leaves of a mangrove species, Rhizophora, tends to appear
in a clumped form, which violate the assumption of randomly distributed foliage.
Although the resultant underestimation of LAI is considered as insignificant, the
assumption reveals the potential error in LAI estimation. Last but not the least, the
number of measurements in the sampling plots depends on the heterogeneity of
the canopy. In fact, more measurements are required for heterogeneous canopy
than the homogeneous one (Gower et al., 1999).
2.6.3.4. Correction of Indirect LAI Measurement
The results derived from different methods of indirect LAI estimation are always
observed to have certain degree of deviation from the direct estimates as noted by
a number of studies (Neumann et al., 1989, Chason et al., 1991, Smith et al., 1993,
99
Fassnacht et al., 1994, Dufrêne and Bréda, 1995, Küβner and Mosandl, 2000). The
two main possible sources of error came from (1) the assumption of random
distribution of foliar elements within the canopy as well as (2) the inclusion of nonphotosynthetic plant elements in LAI estimation. The first limitation is theoretical
while the second one is regarded as technical (Fournier et al., 2003).
2.6.3.4.1.Clumping
Under most circumstances, the indirect methods tend to underestimate LAI
compared with direct measurement. The underestimation ranged from 25 – 50% in
different stands (Gower and Norman, 1991, Cutini et al., 1998, Gower et al., 1999) is
primarily due to the non-random distribution of foliar elements in the canopy.
Generally, the distribution of foliar elements does not follow the Possion model due
to the characteristics of vegetation structure (Fournier et al., 2003). The degree of
LAI underestimation depends on the deviation of vegetation structure from the
assumption of random distribution of the foliar elements, which is described by
clumping index ( 0 ) (Lang, 1987, Chen et al., 1997, Kucharik et al., 1997). The
relationship can be expressed by a simple equation as shown below,
Leff  0 L
Eq. 2.38
where 0 is the clumping index controlling the dispersion or aggregation of the
canopy (Chen and Black, 1992); Leff is the effective LAI which is the product of the
clumping index with the area index (Chen, 1996). When a canopy shows random
dispersion, 0 = 1; when a canopy is clumped, 0 is either larger (regularly
distributed) or smaller (leaves clumped together) than unity (Fassnacht et al., 1994,
Breda, 2003, Weiss et al., 2004). The expression by inverting Beer-Lambert’s Law
using Miller’s theorem (1967) is expressed as
100

lnP 
cos  sin d



0
2
Lt  2 
Eq. 2.39
Some studies have further divided the clumping index in different scales, e.g.
between-shoots clumping factor, within-shoot clumping factor (Chen et al., 1991,
Chen et al., 1997). For instances, an optical instrument called Tracing Radiation and
Architecture of Canopies (TRAC) and canopy gap size theory developed by Chen and
Cihlar (1995a) are used to measure LAI of clumped canopies. Similar to other
indirect optical devices, TRAC estimates effective LAI using gap fraction while the
clumping index is derived from the gap size distribution (Leblanc et al., 2005).
However, the suggested correction parameters are locally fit rather than universally
applicable (Weiss et al., 2004).
2.6.3.4.2.Mixture of Green and Non-green Elements
The LAI generally describes the functioning of canopy through measuring the
photosynthetic parts of the vegetation. Another discrepancy between indirect and
direct measurement found on the incapability to differentiate non-green tissues
such as stems, branches, trunks, flowers, fruits, senescent or dead leaves from
green tissues (Breda, 2003, Fournier et al., 2003, Weiss et al., 2004). Among those,
the significant source of error comes from the woody components (Whitford et al.,
1995, Chen et al., 1997). Hence, alternative terms such as plant area index (PAI)
(Neumann et al., 1989), foliage area index (Welles and Norman, 1991b), vegetation
area index (Fassnacht et al., 1994) have been suggested in the literatures. Their
relationship with LAI is expressed as:
LAI  PAI - WAI ; or
Eq. 2.40
LAI  PAI (1 -  )
Eq. 2.41
where WAI is wood area index;  is the ratio of WAI to PAI. Both indirect and direct
measurements of WAI have been researched (Chen, 1996, Kucharik et al., 1997,
101
Kucharik et al., 1998). For instances, Chen (1996) estimated  using destructive
sampling method. Kucharik et al (1998, 1997) made use of Multiband Vegetation
Imager (MVI) to capture images of two bands – visible 400-620nm and near infrared
720-950 nm using the filter exchange mechanism (Jonckheere et al., 2004).
Together with the k-means clustering algorithm – Balanced Iterative Reducing And
Clustering using Hierarchies (BIRCH) in order to distinguish sunlit and shaded leaves,
sunlit and shaded branch area, clouds, and blue sky (Weiss et al., 2004). The results
showed that the contribution of woody components to PAI varies from 5 – 35%
(Gower et al., 1999). However, the direct subtraction of WAI from PAI was criticized
as branches are very likely overlapped by foliages (Dufrêne and Bréda, 1995, Gower
et al., 1999).
2.6.4. Empirical Relationship with Spectral Vegetation Indices
The field measured LAI acted as reliable ground observations are use to calibrate
the remotely-sensed data to provide a spatially continuous estimation of LAI status
of the study area. Since the use of single spectral band is typically insufficient to
portray vegetation properties, empirical spectral indices is the popular method to
estimate biophysical parameters (Zhao et al., 2007). The method computes an
empirical model through associating the measured biophysical variables, e.g. LAI
with parameters derived from remote sensing data. The parameters derived from
remote sensing data can be raw radiance data, spectral vegetation indices or
derivatives and the relationship can be linear or non-linear. If the relationship of the
model is strong, data extracted from the image can be used to predict the
biophysical parameters over the areas of interest. And if strong relationships could
be held consistently with the biophysical variables, they would be useful for
monitoring long-term environmental changes (Treitz and Howarth, 1999).
102
2.6.4.1. Traditional Vegetation Indices
Traditionally, vegetation indices (VIs) derived from the red and near infrared (NIR)
portions of electromagnetic spectrum were extensively used to estimate canopy
characteristics such as percent green cover, biomass, leaf area and fAPAR, which
can in turn analyze vegetation biophysical and biochemical processes such as
photosynthesis, and evapotranspiration (Wessman, 1994b, Elvidge and Chen, 1995,
Treitz and Howarth, 1999). Vegetation indices function by contrasting the high
chlorophyll pigment absorption in red against the high reflectance in the near
infrared owing to the multiple scattering effect (Elvidge and Chen, 1995, Todd et al.,
1998). The most widely used spectral vegetation indices are the simple ratio (SR)
and normalized difference vegetation index (NDVI) (Mutanga and Skidmore, 2004,
Zhao et al., 2007). SR is simply the ratio of red and near infrared reflectance while
NDVI is expressed as:
 NIR   RED
 NIR   RED
Eq. 2.42
Where  NIR and  RED is the reflectance from near infrared red band respectively. In
the equation, the unnormalized difference vegetation index  NIR   RED  is
normalized by dividing the sum  NIR   RED  that is closely connected to the
albedo of surface (Verstraete, 1994a).
Using VIs to estimate biophysical parameters of vegetation offers a few advantages.
Firstly, it is easy and fast computation by simple combinations of channels and do
not involve extra ancillary data. It involves very little expertise and provide a fast
access to general state of vegetation information, which may be useful in many
applications (Verstraete, 1994a). By taking the difference or normalization of two
channels, another advantage is the tendency to overcome irrelevant spectral
variability due to soil background, canopy geometry, topographic effects and
atmospheric conditions, which can enhance biophysical properties measurement of
vegetation such as LAI, biomass, at canopy scale (Wessman, 1994b, Elvidge and
Chen, 1995, Eastwood et al., 1997, Blackburn and Steele, 1999, Gao et al., 2000, De
Jong and Epema, 2001, Lu et al., 2004, Mutanga and Skidmore, 2004, Lu, 2006).
103
However, NDVI is still sensitive to internal and external factors such as atmospheric
and soil variables which renders them fail to provide accurate estimation (Goel and
Qin, 1994, Asner et al., 2003). Innovative spectral indices like Soil-Adjusted
Vegetation Index (SAVI) (Huete, 1988), Transformed Soil-Adjusted Vegetation Index
(TSAVI) Soil, Modified Soil-Adjusted Vegetation Index (MSAVI), Atmospherically
Resistant Vegetation Index (ARVI) (Kaufman and Tanre, 1992) and Enhanced
Vegetation Index (EVI) (Huete et al., 1994) were developed with prime objective to
improve their sensitivity to spectral contribution of green vegetations signals while
to suppress the contribution from soil, understory and atmospheric effects (Qi et al.,
1994b, Rondeaux et al., 1996, Karnieli et al., 2001, Sims and Gamon, 2002,
Rautiainen, 2005), which can enhance their responsiveness to canopy biophysical
parameters such as LAI (Nemani et al., 1993, Brown et al., 2000). VI that can be
totally insensitive to other variables and only sensitive to the desired parameters is
difficult if not impossible to develop (Govaerts et al., 1999). Therefore, VIs are
regarded as best estimator if they can maximally suppress the confounding factors
affecting the spectral reflectance (Haboudane et al., 2004).
Sensitivity analyses have been conducted to explore the effects of confounding
factors in estimating biophysical factors with VIs (Goel and Qin, 1994, Myneni and
Williams, 1994, Gao et al., 2000, Meza and Blackburn, 2003). Through three
dimensional computer simulation of tree canopies, Goel and Qin (1994) explored
the relationships between different VIs with LAI, fAPAR, percentage ground cover
under the influences of various factors such as soil brightness, leaf angle
distribution, sun-sensor geometry, optical properties of canopy elements, and
spacing distance between vegetation. Myneni and Williams (1994) specifically
explored the effects of confounding factors in influencing the relationship between
NDVI and fAPAR. Gao et al. (2000) compared the results of using NDVI and SAVI in
estimation of fAPAR and LAI with the SAVI showed stronger relationship. Using
laboratory simulation of mangrove canopy over different background, (Meza and
Blackburn, 2003) compared a number of different VIs and derivative-based indices
in correlating with LAI and canopy cover. They also investigated the sensitivity of
these indices to variations of background reflectance and concluded that Difference
104
Vegetation Index (DVI), which showed robust linear relationships with LAI can be
used to estimate biophysical variables of mangroves.
Spectral vegetation indices were applied to map biophysical variables such as LAI,
canopy closure, and fAPAR of mangrove though there are limited numbers (Elvidge
and Chen, 1995, Green et al., 1997, Davis and Jensen, 1998, Green et al., 1998b,
Satyanarayana et al., 2011). For instances, Green et al. (1997) used NDVI derived
from Landsat TM and SPOT satellite data to generate linear model in estimating
spatial distribution of LAI of mangrove ecosystem. Green et al. (1998b) established
linear relationship between NDVI derived from high resolution CASI and mangrove
LAI/ canopy closure. Davis and Jensen (1998) correlated SR and NDVI derived from
airborne high-resolution sensor with different field-collected biophysical variables in
mangrove areas such as canopy closure, tree height, basal area, diameter at breast
height and average leaf area. Satyanarayana et al (2011) derived NDVI from
Quickbird image to estimate stem density and basal area of mangrove in Malaysia.
Although spectral VIs based on broadband sensors are commonly applied to
biophysical parameters estimation, they are criticized for a number of reasons. The
fundamental one is the failure to exploit all the information contained in the
original data since most vegetation indices consider only two spectral wavebands
(Verstraete, 1994a, Danson et al., 2003). Another criticism is that the empirical
relationship between vegetation indices and biophysical parameters is site specific.
The lack of generality characteristics and physical base renders the necessity of new
model calibration when species or site characteristics are changed (Roberts, 2001,
Danson et al., 2003). Furthermore, vegetation indices fail to take the characteristics
of canopy structure such shadowing and multiple scattering into account (Asner et
al., 2003). Despite the fact that vegetation indices are claimed to be independent of
confounding background variations, Demetriades-Shah et al. (1990) argued that it is
applicable only when the level of background signal of the two wavelengths is
similar or relatively constant from one sample to another. This is endorsed by
Verstraete (1994a) that it is almost impossible to derive a vegetation index that is
completely insensitive to all possible atmospheric and soil effects. Traditional VIs,
such as SR and NDVI are very likely to be affected by background optical properties
105
not related to the state of vegetation (Hall et al., 1990, Verstraete, 1994a).
Moreover, other external variables such as sun-sensor relationship, canopy
architecture and biochemical properties of canopy are not considered (Verstraete
and Pinty, 1996, Danson et al., 2003, Haboudane et al., 2004, Rautiainen, 2005).
Another major limitation of VIs using red and near infrared portion of spectrum for
biophysical estimation is the saturation problem after LAI or biomass density
reaches a certain level (Todd et al., 1998, Gao et al., 2000, Thenkabail et al., 2000,
Mutanga and Skidmore, 2004). With the asymptotic addition of foliage, the NIR
reflectance will continuously surge due to multiple scattering (Kumar et al., 2001),
however, the capability of red light absorptions by leaves is reaching the peak
(Thenkabail et al., 2000). The asymptotically narrowing difference between NIR and
red reflectance results in less sensitivity of NDVI, thus, failing to detect change of
biomass (Mutanga and Skidmore, 2004). Generally, when LAI exceeds 2 to 5,
vegetation indices asymptotically approach saturation depending on the type of
vegetation index used (Haboudane et al., 2004). Danson and Plummer (1995)
reported that when LAI is greater than 6, no correlation between LAI and NDVI is
found. The problem is especially apparent during peak growing season when
biomass level increases significantly (Thenkabail et al., 2000, Mutanga and Skidmore,
2004). Criticisms of NDVI derived from broadband in providing reliable estimation of
LAI of different vegetation types such forest, grass, crops are presented (Danson
and Plummer, 1995, Pu et al., 2003).
2.6.4.2. Leaf Area Index Estimation from Hyperspectral and Radar Images
Apart from modifying the VIs based on the common spectral bands, the possibility
of narrowband hyperspectral data are explored for their prediction power and
stability in estimating biophysical parameters, especially LAI (Broge and Leblanc,
2000). Compared with broadband multispectral remote sensing which focus only on
the red and NIR bands, hyperspectral remote sensing offers opportunities of
examining vegetation indices based on narrow bands covering the whole
electromagnetic spectrum, typically from 350 to 2500 nm (Mutanga and Skidmore,
106
2004). A number of studies have compared the results of broadband and
narrowband vegetation indices for LAI and other biophysical parameters estimation
of different vegetation types. Some reported that the narrowband VIs were
marginally better than those derived from broadband (Elvidge and Chen, 1995,
Thenkabail et al., 2000, Lee et al., 2004, Pu et al., 2005, Schlerf et al., 2005) and with
greater flexibility (Zhao et al., 2007) while others observed no differences between
them (Broge and Leblanc, 2001, Broge and Mortensen, 2002). For instances, Elvidge
and Chen (1995) compared the broad-band and narrow-band red and near-infrared
vegetation indices in predicting LAI and percentage of green cover under different
background effects. The results revealed that first-order derivative green vegetation
index derived from narrow-band data obtained the best result and most effective in
reducing background effects. Pu et al. (2005) correlated the surface reflectance
extracted from three sensors Hyperion, Advanced Land Imager (ALI) and Landsat 7
ETM+ with measured forest LAI and crown closure (CC). They found that Hyperion
sensor do better than ALI and ETM+ and the SWIR spectral region of Hyperion
performs the best. Broge and Leblanc (2001) tested the prediction power of a
number of broadband and hyperspectral vegetation indices derived from simulated
canopy reflectance model to estimate LAI and Canopy Chlorophyll Density (CCD).
Their results showed that vegetation indices derived from hyperspectral data are
not essentially superior in predicting LAI and CCD, rather, the selection of suitable
VIs affects the results.
Taking the advantages of a large number of narrow bands in hyperspectral data, a
number of new vegetation indices have been proposed with the aim to compensate
the influences of soil background, atmospheric effects and biochemical components.
These vegetation indices make uses of different proportions of the spectrum to
investigate various features of interest. For examples, the chlorophyll absorption
ratio index (CARI) calculates chlorophyll concentration, which is useful for
agricultural studies (Liang, 2004). The variation of CARI is the chlorophyll absorption
continuum index (CACI) which make use of the spectral continuum characteristics of
hyperspectral data instead of using discrete bands (Broge and Leblanc, 2000). The
triangular vegetation index (TVI) is characterized by three areas of interests
107
including the green peak, the red chlorophyll absorption well and the shoulder of
red edge, which form a triangular spectral space (Broge and Leblanc, 2000). The size
of triangle acts as indicator of change in biophysical and biochemical status of
vegetation canopies (Liang, 2004). The photochemical reflectance index (PRI)
measures the efficiency of radiation usage (Gamon et al., 1992). The water index
(WI) with different combinations of water absorption bands estimates foliar water
content (Peñuelas et al., 1993, Ceccato et al., 2001). Haboudane et al. (2004)
developed the hyperspectral VIs based on modifications of CARI and TVI that can
accurately estimate LAI with minimum influence by leaf chlorophyll concentration.
Although these vegetation indices are developed with hyperspectral data, they are
confirmed by limited number of studies. Besides, they are liable to the influence of
canopy architecture and sun-sensor geometry (Liang, 2004).
Apart from VIs, the red edge parameter is the commonly used spectral region for
study and estimation of biophysical parameters because of several reasons. First,
the red edge position is comparatively irresponsive to changes of biophysical factors
such as soil background, canopy architecture, atmospheric effects and sun-sensor
relationship when compared to traditional vegetation indices (Horler et al., 1983,
Goetz, 1992, Danson and Plummer, 1995, Pu et al., 2003). Pu et al. (2003) used red
edge parameter to estimate forest LAI and proved that the model is relatively
robust even under canopy structure and soil background variation. Second, studies
have shown that narrow bands located in the red edge are sensitive to canopy
characteristics such as biomass, LAI, foliar inclination and chlorophyll concentration.
The biophysical and biochemical characteristics of canopy control the spectral shift
and movement of the red edge (Guyot et al., 1992, Baret and Jacquemoud, 1994,
Filella and Penuelas, 1994, Danson and Plummer, 1995, Todd et al., 1998, Blackburn
and Pitman, 1999, Liang, 2004, Mutanga and Skidmore, 2004). Among the factors,
LAI is the primary factor influencing the red edge position (REP) or red edge
inflection point (REIP) (Guyot et al., 1992). The shift of REP to shorter wavelength
(blue-shift) signifies the decline in vegetation density while the shift to longer
wavelengths (red-shift) is liable to increase in vegetation density (Liang, 2004). As
REP is sensitive to LAI estimation, a number of studies have explored the potential
108
of red edge in estimation of biophysical parameters, especially LAI. For instances,
Danson and Plummer (1995) found a strong non-linear correlation between red
edge and LAI. Using airborne CASI data, Lucas et al (2000) reported strong linear
correlation between inflection point of red edge and LAI of conifer forest. Mutanga
and Skidmore (2004) overcame the saturation problem in extracting biomass
information in dense vegetated areas using narrow band indices and the red edge
showed high correlation with biomass. Most of these studies not only confirm
strong relationship between the red edge and LAI, they also revealed great
potential for narrow wavebands to estimate biomass level at high canopy density
when compared to vegetation indices derived from broadband.
Derivative technique is another popular approach for empirical estimation of
canopy characteristics. The advantage of derivatives is their insensitivity to nonphotosynthetically active materials such as soil and litter (Hall et al., 1990, Wessman,
1994b) as well as other confounding factors such as atmospheric effects (Baret and
Jacquemoud, 1994). As a result, purer green vegetation information can be
acquired. The second derivative for three discrete wavebands is defined by Liang
(2004) as:
Eq. 2.43
For example, some studies found that the value of second derivative is highly
sensitive to canopy reflectance around the red edge regardless of different soil
background (Li et al., 1993). However, the derivative techniques are sometimes
problematic because of their sensitivity to noise (Wessman, 1994b).
Hyperspectral sensors record spectral bands in shortwave near-infrared (SWIR).
New VIs such as Reduced Simple Ratio (RSR) and the Moisture Stress Index (MSI)
incorporating SWIR were developed. RSR combines the red, NIR, and SWIR bands
while MSI is computed from NIR and SWIR (Rautiainen, 2005).
109
Reduce Simple Ratio (RSR) =
Moisture Stress Index =
Eq. 2.44
Eq. 2.45
where SR is the simple ratio (NIR/ RED). SWIRmax and SWIRmin is the SWIR reflectance
obtained from an open canopy and totally closed canopy respectively. The
maximum and minimum value of SWIR reflectance are acquired from the whole
image or within specific areas of the image (Brown et al., 2000, Rautiainen, 2005).
The SWIR band is proposed to be useful in mapping LAI due to its capability in
reducing the contribution from the understory vegetation in conifer forest (Brown
et al., 2000). For instances, Nemani et al. (1993) found leaf areas of forest were
better estimated after the inclusion of SWIR bands into NDVI.
As mentioned, radar is capable to acquire under-canopy vegetation parameters and
sensitive to dielectric constant related to water availability. The backscatter
coefficients from radar data can be good estimators for LAI and other biophysical
parameters. Mangroves flourish in tide inundation zone where water availability
and quantity change seasonally. With longer wavelength such as P-band, radar
signal can penetrate to the forest floor and extract the surface characteristics.
Besides, the all-weather capability enables regular data capture through time,
which allows effective monitoring of mangrove monthly or seasonally.
Many studies have made use of SAR data to retrieve forest biophysical parameters
(Dobson et al., 1992, Le Toan et al., 1992, Dobson et al., 1995, Wang et al., 1995,
Green, 1998, Kellndorfer et al., 2003), LAI (Chand and Badarinath, 2007, Kovacs et
al., 2008) and biomass retrieval (Kasischke et al., 1994, Ferrazzoli et al., 1997,
Imhoff et al., 1998, Kurvonen et al., 1999). For instance, Kasischke et al (1994)
found a significant correlation between backscattering coefficient from ERS-1 and
different components of biomass. Chand and Badarinath (2007) also established a
110
significant relationship between backscatter derived from ENVISAT-ASAR with
ground measured leaf area index. Kovacs et al. (2008) acquired a set of
multipolarized ASAR for health monitoring of mangrove forest. Biophysical
parameters including LAI, height, stem density, basal area and mean DBH are
correlated with backscatter. While biomass estimation is one of the keen research
fields using SAR data, similar to optical data, SAR data exhibited the saturation
problem as observed by investigators. It occurs when the relationship between
radar backscatter intensity and biomass level off after a certain level of biomass
reached. Normally, the longer the wavelength is, the higher the saturation level will
be. The biomass sensitivity increases from 30 t/ha for C-band with HH or VV
polarization to as high as 200 t/ha for P-band with HV polarization (Patel et al.,
2006).
2.6.5. Physically-based Canopy Reflectance Model Inversion
Apart from empirical modeling using various VIs, biophysical parameters can be
estimated through inversion of radiative transfer model. The radiative transfer
models were mostly designed for the normalization of bidirectional effects of sunviewing geometry and correction of atmospheric effects while some of them are
invertible to approximate vegetation biophysical properties (Qi et al., 1995).
2.6.5.1. Canopy Reflectance Model
The interaction of the vegetative elements in a canopy with electromagnetic
radiation can be simulated through canopy reflectance model based on the
radiative transfer theory. ‘A canopy reflectance (CR) model provides the logical
connection between the botanical and biophysical features of the canopy, the
geometry of the radiometric interaction and the resulting alteration to the reflected
radiation (Goel, 1988, p.2)’. CR model simulates the processes of light transfer and
interaction with vegetation canopies, i.e. scattering, transmittance and absorption,
based on radiative transfer theory (Goel, 1988). The scattering and absorption
mechanisms of vegetation are functions of intrinsic optical and structural factors
111
such as canopy architecture, foliar optical properties as well as spatially specific and
temporally dependent measurement conditions such as illumination-viewing
geometry, background reflectance, atmospheric effects (Verstraete, 1994a, Fourty
and Baret, 1997, Baret et al., 2000, Danson et al., 2003). Among the factors, leaf
optical property is the prominent input to all the CR models (Jacquemoud et al.,
2000). Vegetation canopy reflectance can be simulated when all the factors are
taken into the model for computation (Goel, 1988, Chen and Leblanc, 1997,
Jacquemoud et al., 2000, Kuusk and Nilson, 2000). Unlike statistical methods, the
physically based canopy reflectance model is not restricted to location, sensor or
seasonal change.
CR models are essentially developed based on similar set of input parameters based
on which pixel reflectance as a function of wavelength (λ) are computed. They can
be classified into several categories including sun-sensor geometry (GEOM),
atmospheric properties (ATMO), canopy structural parameters (STRCANOPY),
landscape structural parameters (STRLANDSCAPE), optical properties of leaf tissues
(LEAFOPT) and background soil reflectance (ρSOIL). Equation 2.46 represents the
parameters as a function modified based on Anser et al. (2003) and Goel (1988).
The description of each input parameters is shown in Table 2.3.
R(λ) = f(GEOM, ATMO, STRCANOPY, STRLANDSCAPE, LEAFOPT, ρSOIL) Eq. 2.46
where GEOM = f(θSUN, θVIEW, φSUN, φVIEW)
ATMO = f(CONAEROSOL, CONWATER, CONOZONE)
STRCANOPY = f(LAI, NPVAI, LAD, NPVAD, TISSUE A:H, CLUMPING)
STRLANDSCAPE = f(CRWNHGT, CRWNSHP, STANDDEN)
LEAFOPT = ρLEAF, τLEAF, ρNPV, τNPV = f(Cw, Cab, N,…)
112
Table 2.3. The description of input parameters in CR models
(Modified based on Anser et al., 2003 & Goel, 1988)
Parameter
Symbol
θSUN
θVIEW
φSUN
φVIEW
LAI
Parameter Description
CONAEROSOL
CONWATER
CONOZONE
NPVAI
LAD
NPVAD
TISSUE A:H
CLUMPING
CRWNHGT
Atmospheric concentration of aerosol
Atmospheric concentration of water vapor
Atmospheric concentration of ozone
Non-photosynthetic vegetation area indices
Leaf angle distribution
Non-photosynthetic vegetation angle distribution
Size of foliar relative to their height distribution in the canopy
Foliage dispersion index
Height of individual tree crowns
CRWNSHP
STANDDEN
ρLEAF
τLEAF
ρNPV
τNPV
Shape of individual tree crowns
Stand Density
Reflectance characteristics of foliar tissue
Transmittance characteristics of foliar tissue
Reflectance characteristics of non-photosynthetic vegetation
Transmittance
characteristics
of
non-photosynthetic
vegetation
Water content/ depth within foliage
Chlorophyll a + b concentration within foliage
Foliar mesophyll structure
Cw
Cab
N
Sun illuminating zenith angle
Sensor viewing zenith angle
Sun illuminating azimuth angle
Sensor viewing azimuth angle
Leaf area index
The majority of canopy modeling studies have centered on the simulation of
reflectance through the use of hypothetical or measured foliar and canopy
parameters. Such practice is called ‘forward modeling’ (Asner et al., 2003). Forward
modeling economically assists understanding the effect of interactions of structural
and biochemical components in canopy on vegetation reflectance signature (Bacour
et al., 2002, Rautiainen et al., 2004). Based on the variations of structural and
biochemical components, simulation of reflectance change in terms of age and
seasonal effect becomes viable (Rautiainen, 2005). Sensitivity analysis of different
113
biophysical parameters can be undertaken to build up a knowledge base related to
their effects on reflectance (Eriksson et al., 2006). Besides, through establishing
quantitative relationships between remote sensing data and vegetation attributes,
it provides an interface for reflectance model inversion, which can be used to
estimate vegetation biophysical parameters (Chen et al., 2000, Rautiainen, 2005).
CR model ranging from simple one-dimensional light scattering simulation (based
on radiative transfer model) to complex three dimensional ray tracing models, have
been developed in the past decades (Asner et al., 2003). They are generally
categorized into four types including geometrical models, turbid medium or
analytical models, hybrid models and ray-tracing or three-dimensional computer
simulation models (Goel, 1988, Roberts, 2001, Garcia-Haro and Sommer, 2002,
Schlerf and Atzberger, 2006). The geometrical models (e.g. (Li and Strahler, 1986))
are the simplest, which computed vegetation reflectance based on linear
combination of light interception and shadowing by simple geometric shapes such
as cones, cylinders and ground surface reflectance. However, the geometric models
ignore the multiple scattering effects within the canopy (Roberts, 2001, Garcia-Haro
and Sommer, 2002), which is applicable to sparse canopies such as shrubs, sparse
forest (Goel, 1988). The turbid medium models consider vegetative elements as
small scattering and absorbing particles of a given density with known optical
properties. During the modeling process, LAI and LAD are input parameters. This
kind of model is more lucrative for denser and horizontally uniform canopies (Goel,
1988, Roberts, 2001). The hybrid models consider the overall geometric shape of
the canopy and the microstructure within the canopy within which multiple
scattering within and between crowns, trunk and ground takes place (Chen and
Leblanc, 1997, Kuusk and Nilson, 2000, Garcia-Haro and Sommer, 2002). This type
of model is suitable for both sparse and dense canopy (Goel, 1988). Finally, the raytracing technique is the most structurally complex and computationally demanding
method because it simulates the trajectories of solar photons in the vegetation
structure and requires the detailed description of canopy geometry and the
interaction processes (Goel, 1988, Jacquemoud et al., 2000, Asner et al., 2003,
Schlerf and Atzberger, 2006).
114
Many CR models were developed in the past few decades. The Suits (1972) model
was one of the earliest models assuming a canopy is a composition of horizontal
and vertical leaves only. Simple canopy parameters such as canopy structure and
solar-viewing geometry were considered (Liang, 2004). By allowing variations of leaf
angles, Verhoef (1984) extended the Suits model to SAIL (Scattering by Arbitrarily
Inclined Leaves) canopy model and its variations SAILH with hot spot effect added.
Forest Light Interaction Model (FLIM) was developed to model discontinuous
canopy layer and the reflectance is computed using the probability of viewing a
crown (Rosema et al., 1992). When the foliar positional relationship with adjacent
layers was considered instead of assuming a homogeneous canopy, the Markov
chain canopy reflectance model (MCRM) was developed (Kuusk, 1995b). With the
additions of anisotopic soil reflectance, specular reflection of sun and hot spot
effect, MCRM is modified into ACRM which became a two-layer model (Kuusk,
2001). Recently developed models tend to couple other models such as leaf
reflectance models, atmospheric models, and soil reflectance models. For instances,
the Forest Reflectance and Transmittance (FRT) Model coupled with the broadleaf
reflectance model (PROSPECT), radiative transfer model 6S and the two-layer
understory reflectance model MCRM2 (Kuusk and Nilson, 2000). The Invertible
Forest Reflectance Model (INFORM), an extension of FLIM, incorporated SAILH and
needleleaf reflectance model (LIBERTY) (Zarco-Tejada et al., 2001, Schlerf and
Atzberger, 2006).
2.6.5.2. Model Inversion and Biophysical Parameters Extraction
The term, model inversions or inverse modeling offers an alternative to retrieve
optical and structural variables of vegetation canopies, provided that remotely
sensed reflectance observations and measuring conditions are provided (Verstraete,
1994a, Asner et al., 2003). ‘The purpose of inversion is to retrieve the values of
unknown parameters characterizing a surface from a given dataset using functional
dependence of these parameters (Rahman, 2001, p.1250)’. The idea can be
expressed by switching the factors in Equation 2.47 into the following equation:
115
STRCANOPY, LEAFOPT = f(GEOM, ATMO, R(λ), STRLANDSCAPE, , ρSOIL)
Eq. 2.47
The positional switches of pixel reflectance as a function of wavelength R(λ) and
canopy structural parameters (STRCANOPY) and leaf optical properties (LEAFOPT)
indicates that the canopy structural parameters and leaf optical properties are
functions of modeled or measured reflectance (Goel, 1988). GEOM, ATMO and ρSOIL
are regarded as confounding factors and their influences should be minimized
while the parameters of interest, STRCANOPY and LEAFOPT, are to be accurately infer
from remote sensing data (Goel and Qin, 1994). The model inversion approach
based on a primary assumption that canopy, leaf and soil background
characteristics would influence the spectral variations of canopy reflectance (Baret
and Jacquemoud, 1994). In other words, only those parameters adequately affect
the reflectance at canopy top can be accurately retrieved from the measured
reflectance data during the inversion process (Kimes et al., 2000). For example, the
Scattering by Arbitrarily Inclined Leaves (SAIL) model developed by Verhoef in 1984
requires input variables such as LAI, background soil reflectance, leaf reflectance
and leave transmittance. LAI can therefore be estimated through inversion of the
SAIL model (Qi et al., 1995). Prior to using spectral reflectance for biophysical or
biochemical parameter mapping, the relationship between spectral reflectance and
a number of factors such as canopy architecture, leaf properties, atmospheric and
background influences have to be accounted for (Houborg and Boegh, 2008). The
incorporation of sub-models such as leaf optical model, understory vegetation
model and atmospheric model into the main canopy reflectance model provide
detailed information about factors affecting canopy reflectance (Eriksson et al.,
2006).
A good inversion of a reflectance model should possess three criteria including a
sound canopy reflectance model, an appropriate inversion procedure and a set of
well-calibrated reflectance (Jacquemoud et al., 2000). Different model inversion
techniques are available such as the traditional iterative optimization technique
(Goel, 1988), look-up-table (Combal et al., 2002), artificial neural networks (Gong et
al., 1999), and generic algorithm (Fang et al., 2003).
116
Many studies have used model inversion to retrieve biophysical and biochemical
parameters like LAI (Ramsey and Jensen, 1995, Braswell et al., 1996, Kuusk, 1998,
Bicheron and Leroy, 1999, Jacquemoud et al., 2000, Rahman, 2001, Meroni et al.,
2004, Rautiainen, 2005, Eriksson et al., 2006, Schlerf and Atzberger, 2006, Houborg
and Boegh, 2008), leaf angle distribution (LAD) (Kuusk, 1995a, Ramsey and Jensen,
1995, Kuusk, 1998), leaf chlorophyll content (Kuusk, 1998, Bicheron and Leroy, 1999,
Jacquemoud et al., 2000, Zarco-Tejada et al., 2001, Houborg and Boegh, 2008),
canopy closure (Schlerf and Atzberger, 2006), fAPAR (Braswell et al., 1996, Bicheron
and Leroy, 1999), and albedo (Braswell et al., 1996). Recently, reflectance extracted
from hyperspectral data have been studied as input to model inversion (Ramsey
and Jensen, 1995, Jacquemoud et al., 2000, Zarco-Tejada et al., 2001, Meroni et al.,
2004, Schlerf and Atzberger, 2006). Ramsey and Jensen (1995) used high spectral
resolution data to study mangroves communities in southwest Florida. They applied
the light-canopy interaction model to predict LAI, LAD, and leaf reflectance factor
through optimization of difference between measured and predicted canopy
reflectance. Houborg and Boegh (2008) coupled the turbid medium Markov chain
canopy reflectance model (ACRM) with the PROSPECT leaf optics model to retrieve
LAI and leaf chlorophyll content using data from SPOT. Schlerf et al. (2006)
employed hybrid-type canopy reflectance model INFORM to estimate forest
structural characteristics of LAI, stem density and crown cover.
Although CR model inversion has been used extensively for parameter estimation,
they are not without criticisms. The first criticism is related to computational
efficiency. When more parameters are added to the models, a more accurate and
robust model can be formulated (Kimes et al., 2000). However, intensive computer
iterations and thus more processing time are required (Verstraete, 1994a, Roberts,
2001). For instances, the iterative optimization approach is not suitable for the
pixel-based regional and global studies because it is computationally prohibitive
(Schlerf and Atzberger, 2006). Models can be simplified to raise computational
efficiency but with the sacrifice on accuracy and robustness of the estimates (Kimes
et al., 2000). Neural network approach is also criticized for its lengthy training stage
(Schlerf and Atzberger, 2006). Secondly, inversion approach such as the iterative
117
optimization requires initial guess of values in the parameter space. Poor guess of
initial value can lead to greater risk of false minima instead of a globally optimal and
stable solution (Kimes et al., 2000). Thirdly, different combinations of canopy
variables will probably come up with similar reflectance spectra (Houborg and
Boegh, 2008). The problem refers to ill-posed nature of model inversion (Combal et
al., 2002, Atzberger, 2004). Fourth, the number of parameters can be retrieved
from CR model is limited by the number of measurements. Goel (1998) suggested
that the number of canopy parameters to be estimated must be less than the
number of measurements. Roberts (2001) also proposed that the number of model
parameters have to be small in order to have a robust inversion. For detailed
discussion on the advantages and major drawbacks of the methods of model
inversion, please refer to Schlerf et al. (2006) and Kimes et al. (2000).
2.7. Summary
Remote sensing has long been a practical tool for vegetation studies due to
distinctive spectral response of vegetation to light. With the advancement of sensor
technology, capturing narrow and continuous bands along the light spectrum by
hyperspectral sensor and collecting canopy structural information using synthetic
aperture radar (SAR) offers additional dimensions for effective vegetation
classification and ecological monitoring. Provided with more information about the
canopy, intuitively, it is believed that finer classes (species-based instead of board
category) and better classification accuracy can be attained. However, under
insufficient sample problems, proper pattern recognition measures are required to
select relevant and non-redundant features/ bands so as to avoid the curse of
dimensionality, enhance computational efficiency, and further understanding on
features governing vegetation discriminability. Apart from classification problem,
effective ecological monitoring through the biophysical parameters estimation is
another important field of application. Leaf area index (LAI) is the biophysical
parameter receiving most attention due to its significant relationship with many
biophysical and biochemical processes in an ecosystem. A number of field
118
techniques are proposed to provide an accurate measure of LAI. Indirect optical
method based on radiative transfer theory is highly efficient though it suffers from
some inherent constraints when compared with direct method. However, it is the
only non-destructive method that is applicable in conservation areas. Empirical
relationships using different vegetation indices are usually developed with the field
measured LAI to provide spatially continuous LAI estimation in the area of interest.
Compared with the traditional VIs derived from broadband sensors, the SWIR from
hyperspectral sensor allows the formulation of new VIs that are sensitive to LAI
estimation. Another biophysical parameter estimation method is through inversion
of canopy reflectance models. Although hyperspectral and radar techniques have
been actively studied in the past few decades, no extensive researches using these
new sensors were found on mangrove species-based classification and LAI mapping.
All in all, the integration of hyperspectral and radar coupled with efficient field
measurement techniques provide innovative insights for mangrove study and
monitoring.
119
CHAPTER 3
METHODOLOGY
3.1. Introduction
This chapter begins with the description of the study area and an overview of
research methodology. This is followed by presenting data acquisition and
processing procedures for field measured data and satellite acquired data. Finally,
the analytical processes including feature selection, classification and validation and
finally leaf area index modeling are explained.
3.2. Study Area Description
Mai Po, the largest mangrove sites in Hong Kong was chosen as the study area. Mai
Po Nature Reserve is located at in the northwestern part of Hong Kong (22°29'N 22°31'N, 113°59'E - 114°03'E), which borders the Shenzhen Special Economic Zone
of Guangdong Province, China. Figure 3.1 shows the location of Mai Po and
mangrove stands in the Hong Kong Territory. Embraced by the inner Deep Bay, Mai
Po consistently receives sediment inputs from inter-tidal flushing and freshwater
discharge from the Pearl River Delta, Shenzhen River as well as from streams in
Hong Kong and Shenzhen. The intertidal mudflats derived from the sediments
flourish the largest mangrove stand in the Pearl River estuary and continues to
expand towards the Deep Bay steadily every year. Looking back to the natural
history of Hong Kong, not only did mudflats and mangroves provide a unique
breeding ground for many marine and terrestrial species, they also offered
economic benefits for human beings. The intervening natural and human processes
shaped the present setting and condition of Mai Po.
120
114°20’0”E
Mai Po in Inner Deep Bay
22°30’0”N
Deep Bay
New Territories
Lantau Island
Hong Kong Island
Figure 3.1. The location of Mai Po Nature Reserve and the distribution of mangrove
stand in Hong Kong
In the past, Mai Po has sustained a number of agricultural and industrial activities.
Agricultural activities included rice cultivation, market gardening, fishing, oyster
cultivation and shrimp cultivation while tradition industries comprised salt panning,
lime burning and brick making. Figure 3.2 shows the change of land use in Deep Bay
from 1924-1985.
121
Figure 3.2. Land use change in Deep Bay, 1924-1985
(adopted from Irving and Morton, 1988)
Noticed from the land use change between 1924 and 1945, reclamation of
mangroves and swamps into gei wais was dominant activities at that time that
formulated current landscape. Gei wai is an area of intertidal marshland surrounded
by clay dyke. Using a sluice gate locating at the seaward side, water is flooded into
the gei wai at high tide and is drained away at low tide (Irving and Morton, 1988).
By 1945, an extensive part of mangrove and swamp were reclaimed and
transformed into gei wais for intertidal shrimp farming ponds (Melville and Morton,
1983, Irving and Morton, 1988). Most of the original mangroves were destroyed at
that time except those located in the shadow platform found in the centre of the
pond were left undisturbed as food source for the shrimps (Irving and Morton,
1988). The extensive gei wai system with shallow brackish ponds remained an
essential ground for fish and shrimp production. It is also an important habitat for a
wide variety of resident and migrating birds (Melville and Morton, 1983). Up until
now, it still serves as one of the important over-wintering stations of the East Asian122
Australian Shorebird Flyway. Besides, over 60,000 waterbirds have wintered around
Mai Po and the Inner Deep Bay areas including an endangered species – blackfacespoonbill (Young, 1999). Outside the gei wai bund wall are undisturbed mangrove
fringe with the decreasing rate of seaward encroachment. Documented by Duke
and Ajmal Khan (1999), the rate of seaward encroachment was about 7.6 meters
per year for the period of 1949 – 1987 and slowing rate of encroachment was
noticed for the period 1987 – 1993.
Mai Po was declared as nature conservation area in 1975 (Irving and Morton, 1988)
and it was then designated as Site of Special Scientific Interest (SSSI) in the year
followed. It became a nature reserve in 1984 and under the management of World
Wide Fund. The significance of Mai Po was not limited to local authority, the
wetlands around the sites were designated as Wetland of International Importance
under the Ramsar Convention in 1995 (WWF Hong Kong, 2007). The international
recognition was mainly due to its role as home for a wide range of resident birds
with some of them being classified as rare species (Irving and Morton, 1988).
Although Mai Po has been studied in the past, they are restricted to small areas in a
certain point location or along a line transect because of limited accessibility and
penetrability as well as its politically-sensitive location bordering with China. For
instance, Duke and Ajmal Khan (1999) studied structure and composition of the
fringe mangrove along a single transect.
Mangrove refers to ‘flora from several different plant families which share a unique
ability to grow in areas which, for some of the time, are submerged by the sea
(Irving and Morton, 1998, p.25)’. They are commonly found in the intertidal zone of
tropical estuaries, extending northward reaching Hong Kong and the southern shore
of Japan. Due to the relatively cool temperature in winter, their heights were
shorter when compared with the tropics counterparts (Irving and Morton, 1988).
The fringe mangrove was thrived from mangroves survived from the excavation,
which stayed mostly untouched by direct human disturbance. The dense fringe is
about 3 kilometers long and 500 meters wide divided by eight to nine channels and
has an area of about 172 hectares (Duke and Ajmal Khan, 1999). Mai Po nurtures
four dominant native species comprising Kandelia obovata and Avicennia marina
123
emerging to the canopy stratum while Acanthus ilicifolius and Aegiceras
corniculatum, occupy the understorey and situated along the coastal edges
(Agriculture, Fisheries and Conservation Department, unpublished data). Two less
common native species, Bruguiera gymnorrhiza and Excoecaria agallocha, are also
found. Apart from the natives, two exotic tropical species Sonneratia apetala and S.
caseolaris were propagated from the plantation along the Shenzhen River and in
Futian National Nature Reserve in Shenzhen. In the past, mangroves were harvested
for a variety of purposes, such as providing timber and fuel resources. In present,
the value of mangroves has altered from economic and consumption-based to
ecological-based. Their roles in sustaining and expanding the ecosystem become the
primary objective for conservation.
3.3. Methodological Flow
Figure 3.3 shows an overview of methodology which is broadly divided into four
main parts including remote sensing data acquisition, processing, field data
acquisition/ processing and data analysis. This study collected primary data through
in situ field survey and acquisition of satellite images. Field surveys involved leaf
collection of mangrove species and also hemispherical photographs taking along
planned routes. Satellite images captured by hyperspectral and microwave sensors
were acquired by satellite programming and through archived database provided by
the Institute of Space and Earth Information Science, CUHK respectively. Apart from
primary data, secondary data about species location were extracted from
government database.
The data acquired have gone through appropriated data processing steps. For data
from field surveys, the cut leaves were brought back to laboratory where the
spectral responses of mangrove species were measured and based on which the
spectral discriminability of different species are compared and used as a reference
to the reflectance extracted from the Hyperion. Biophysical parameter, Leaf area
index (LAI) was computed from the hemispherical photos using WinSCANOPY
system.
124
For satellite data, the hyperspectral image has undergone essential correction
procedures including uncalibrated bands elimination, bad pixels/ lines correction,
destripping, atmospheric correction, noise reduction geometric correction in order
to come up with georeferenced reflectance data. Similarly, a number of correction
procedures were applied to radar dataset to convert the original data to spatially
corrected backscattering coefficients and spatial filtering techniques were used to
reduce. Textural parameters were generated from the master radar image
(acquired in November) for LAI estimation.
After data collection and processing, data analysis involving feature selection,
species-based mangrove classification and LAI estimation followed. The field
measured and archived data together with the satellite data were used to conduct
mangrove species-based classification and leaf area index modeling and estimation
in the study area. Prior to mangrove classification, feature selection measures were
applied to select bands/ features from combined hyperspectral and multi-temporal
radar data that are influential to the classification results. The best selected features
were used as input into various classification algorithms. The accuracy of
classification was evaluated and compared visually and statistically.
The second part of the analysis involves LAI estimation in the study area. Simple
linear regression and stepwise multiple regression analyses were implemented to
explore
the
relationship
between
field
measured
LAI
and
vegetation
indices/textural variables. The models were evaluated and compared. Based on the
results from the analyses, the complementarlity of spectral and radar data for
mangrove species classification and LAI estimation was evaluated.
125
Figure 3.3. The methodological overview of the study
126
3.4. Remote Sensing Data Acquisition and Processing
3.4.1. Hyperion - EO-1
The Hyperion sensor onboard NASA Earth Observing 1 (EO-1) satellite was the first
spaceborne hyperspectral instrument launched in 2000 to acquire visible/ near
infrared (VNIR) and shortwave infrared (SWIR) spectral data covering a complete
spectrum ranging 357 – 2576 nm with spectral interval of 10 nm (a total of 242
spectral bands). It collects data in pushbroom mode with a spatial resolution of 30
meters using two spectrometers with one operating in VNIR range (70 bands
between 356 – 1058nm) while another in SWIR (172 bands between 852 – 2577 nm)
(Eckert and Kneubuhler, 2004). The pushbroom Hyperion sensor built with area
array detector uses one dimension of an array for spatial imaging while another for
spectral imaging (Gao et al., 2004).
A Level 1 Radiometric Hyperion image centered at 22.452°N, 114.014°E, was
acquired on 21st, November, 2008 at 02:48:53 GMT containing 256 pixels x 3400
lines x 242 bands. The L1R product has 242 bands in total, but with only 198 bands
including band 8 – 57 (426.82 to 925.41 nm) in VNIR, and 77-224 (912.45 to 2395.5
nm) in SWIR calibrated. Due to the low responsivity of the detector, bands 1 – 7 and
58 – 76 were not calibrated and set to zero (Beck, 2003). Among the 198 calibrated
bands, four bands are overlapped between the two spectrometers including the
VNIR bands 56 (915.7 nm) and 57 (925.9 nm) and SWIR bands 77 (912.5 nm) and 78
(922.6 nm). The image file was stored in BIL format. The image is of high quality
with minimal cloud effect and the study area is totally cloud-free. A sequence of
correction processes were implemented to transform the digital number to spatially
referenced reflectance value summarized in Figure 3.3.
3.4.1.1. Radiometric correction
Prior to further processing, the uncalibrated bands and the two overlapped bands
(band 77 and 78) were removed and the remaining 196 bands were retained. This is
followed by radiometric correction involving the elimination of bad lines/ pixels and
127
vertical strips commonly found in pushbroom sensor system. Corrected DN of each
spectral band were then converted into radiance L (measured in W/m2SR nm) using
the gain and offset values specific for each spectral band. After that, atmospheric
correction algorithms were applied to convert the radiance to at-sensor reflectance
by considering the solar illumination reaching the ground and atmospheric
modulation effects. Noise removal algorithm was then implemented to enhance
signal-to-noise ratio (SNR). Finally, the image was ortho-rectified to local coordinate
system.
3.4.1.1.1. Vertical strips removal
Vertical strips refer to the abnormal column digital number DN values when
compared with the adjacent columns. They appear when detector arrays are poorly
calibrated (Datt et al., 2003). Vertical lines with no DN values are referred to bad
lines (Li et al., 2008b). In Hyperion image, the DV values in problematic vertical
strips were consistently lower than those in the neighboring columns and appeared
in grey tone (Li et al., 2008b). The image bands were visually examined for the
magnitude of stripping and categorized into three groups – serious, moderate and
negligible as Table 3.1. The sample images with different degree of stripping are
shown in Figure 3.4.
Table 3.1. The magnitude of stripping in Hyperion data
Stripping magnitude
Bands
Serious
8-10, 54-57, 78-82, 94, 97-101, 134-137, 187-212
Moderate
11-15, 27-28, 83-88, 102, 138-154
Negligible
16-26, 29-53, 89-93, 95-96, 103-133, 155-186, 213-224
Atmospheric Water
Absorption
79-81, 98-101, 121-133, 165-182
128
Serious (Band 8)
Moderate (Band 13)
Negligible (Band 21)
Figure 3.4. The sampled bands with different degree of stripping
The vertical strips were removed using the DESTRIP module in PCI Geomatica.
Kernel sizes of 9 and 7 with first order polynomial function were applied to bands
with serious and moderate stripping effects respectively.
3.4.1.1.2. Atmospheric correction
The purpose of atmospheric correction is to eliminate the effects of solar
illumination and atmospheric modulation (mainly gas molecular absorption and
Rayleigh and aerosol scattering) from the measured radiance so as to obtain an
accurate estimate of surface reflectance (Griffin and Burke, 2003, Gao et al., 2004).
At wavelengths below 2.5 μm, the incident solar radiation flux is affected by water
vapor absorption and absorptions by well mixed-gases including ozone, oxygen,
methane and carbon dioxide while aerosol absorption is comparatively negligible.
The magnitude of absorption varies strongly with wavelength and properties of
gases.
The effect of atmospheric scattering by molecules and aerosols is observed to be
inversely related to wavelength, i.e., larger scattering effects in visible than in
longer wavelengths. Rayleigh Scattering by gas molecules is limited to 0.75 μm
while aerosol scattering can continues up to 1.3 μm (SWIR) because of large aerosol
particle size (Griffin and Burke, 2003).
129
Atmospheric correction can be implemented in two ways, namely, relative
atmospheric correction methods and physics-based absolute atmospheric
correction models (Kawishwar, 2007). The relative methods such as internal
average reflectance (IAR) and flat field correction (FFC) are empirical approaches
which do not require priori information about the surface and atmospheric
characteristics; rather they use information within the image itself (Gao et al., 2006).
Under the assumptions that the radiance at top-of-atmosphere and at ground level
is linearly related and there are no other important sources of error, the relative
methods can retrieve surface spectra under situation when information about the
scene is limited. They are computationally efficient and suitable for arid area
without vegetation (Griffin and Burke, 2003) but they are criticized for not involving
any atmospheric parameters in the calculation.
Absolute atmospheric correction modeling uses radiative transfer models, which
simulate the atmospheric-surface radiation interaction to derive ground reflectance.
It requires a full and synchronous depiction of atmospheric constitutes such as the
amount of water vapor, composition of mixed gases so as to quantify the effects of
scattering and absorption while the image was captured. Therefore, the absolute
correction method can be applied under any atmospheric condition, altitude and
sun-satellite geometry (Kawishwar, 2007). However, since the atmospheric
parameters are hard to acquire for precise calculation, current efforts have been
resorted to retrieving these parameters from the hyperspectral images. Frequentlyused radiative transfer models are LOWTRAN, MODTRAN, Code 5S and 6S. Among
those, the Moderate Resolution Transmittance (MODTRAN) code (with the latest
MODTRAN-4) is the most popular and publicly available code developed by Spectral
Science, Inc. and Air Force Research Laboratory.
In the study, atmospheric correction was conducted using two calibration programs
– Atmospheric and Topographic Correction model 2 (ATCOR-2) and Fast Line-of-Site
Atmospheric Analysis (FLAASH) in PCI Geomatica V10.3 and ENVI V4.7 respectively.
The reflectance curves resulted from the two algorithms were compared to
laboratory-measured spectra for their accuracy and fitness.
130
The ATCOR algorithm was developed by Dr. Richter of the German Aerospace
Center (DLR). Up until now, three versions namely, ATCOR-2, ATCOR-3 and ATOCR-4
were developed for different purposes. ATCOR-2 is mostly used for flat terrain;
ATCOR-3 is suitable for mountainous areas as it takes terrain height into calculation;
and ATCOR-4 is the airborne version. Since the study area is relatively flat and lacks
rugged topographic in the surroundings, ATCOR-2 was selected to correct the
Hyperion image. The parametric settings are shown in Table 3.2.
Table 3.2. The parametric setting for ATCOR-2 atmosphere correction
Input Parameter
Value
Average elevation (km)
0.01
Solar zenith angle
43.26
Ground visibility
40
Adjacency effect (km)
1
Correction type
Varying
The Fast Line-of-Site Atmospheric Analysis of Spectral Hypercubes (FLAASH)
algorithm was developed by Spectral Sciences Inc. sponsored by the US Air Force
Research Laboratory (Felde et al., 2003). Incorporated with new MODTRAN4
radiation transfer code which calculates the scattering and absorption of molecules
and particulates in the lower and middle atmosphere (Anderson et al., 1999), both
FLAASH and ATCOR aims at retrieving at-surface reflectance by removing the
atmospheric effects caused by scattering and absorption from at-sensor radiance
(Felde et al., 2003). The MODTRAN4 also takes into account the adjacency effects,
i.e. the scattering from neighboring pixels into the line of sight of current pixel.
Besides, FLAASH is also capable to retrieve the aerosol optical depth, the column
vapor amount for each image pixel, terrain elevation and cloud masks inherent in
the spectral information (Griffin and Burke, 2003). The basic steps converting sensor
radiance to surface reflectance using FLAASH algorithm are shown in Figure 3.5.
131
Figure 3.5. The schematic flow of FLAASH algorithm
(Modified based on Griffin & Burke, 2003)
The image was input into FLAASH and converted into floating-point radiances (in
W/m2SR nm) by dividing the digital number by the scaling factor of individual
spectrometer as follows:
VNIR radiance = Digital Number / 400
SWIR radiance = Digital Number / 800
The scene and sensor-related information such as scene center location, sensor
type and altitude, ground elevation, pixel size, flight date and time were set. In
order to derive an accurate estimation of surface reflectance, the effects of water
vapor, aerosols, and other mixed gases must be considered (Griffin and Burke,
2003). The parameter settings are shown in Table 3.3.
132
Table 3.3. The FLAASH parameters setting for atmospheric correction
FLAASH Parameters
Values
Scene and Sensor Information
22.45223°, 114.01455°
Hyperion
705
November, 21 2008
02:45:53
0
30
Column Water Vapor Retrieval
Tropical
Yes
Scene center location
Sensor type
Sensor altitude (km)
Flight date
Flight time (GMT)
Ground elevation (m)
Pixel size (m)
Atmospheric model
Water retrieval
Water absorption
feature
1135 nm
Aerosol Optical Depth Retrieval
Urban
2-Band (K-T)
40 (default)
Aerosol model
Aerosol retrieval
Initial visibility (km)
Spectral Polish
Spectral polishing
Width (number of bands)
Wavelength
Recalibration
Spectrograph
Aerosol Scale Height
(km)
CO2 Mixing Ratio (ppm)
Use Square Slit Function
Use Adjacency
Correction
Reuse MODTRAN
Calculations
MODTRAN Resolution
MODTRAN Multiscatter
Model
Calculated Visibility (km)
Calculated Average
Water Amount (cm)
Yes
7
Wavelength Recalibration
Yes
spectrograph = {762.6, 0.5, 1, 49}
spectrograph = {942.7, 0.75, 50, 196}
Modeling Parameters
1.00
390.00 (default)
NO
YES
NO
5cm-1
Scaled-DISORT-8
6.7535
1.2567
133
FLAASH estimates the column water amount for each pixel using a three-band ratio
technique. The three-band ratio method allows user to select water absorption
features centered in 820 nm, 940 nm and 1135 nm with spectral resolution of 15
nm or better (Kaufman and Sendra, 1988, Kaufman and Gao, 1992, ENVI, 2009).
Usually, three to five hyperspectral bands are selected to characterize the strongest
portion of the water vapor band while two bands located on each side of the
absorption region representing the high transmission are chosen for water vapor
retrieval process. The absorption feature of 1135 nm was selected as recommended
by ENVI (2009). The MODTRAN atmospheric model was selected based on the
latitudinal and seasonal dependence of surface temperature in the help manual of
ENVI. The tropical (T) model with standard column water vapor amount of 5119
atm-cm was set (ENVI, 2009). The average water amount of 1.2567 cm was
computed.
Standard rural MODTRAN aerosol model was selected to represent the scene area
as the area is not strongly affected by urban or industrial sources. However, as
suggested by ENVI (2009) the choice of model is not critical if visibility is high, i.e.
greater than 40 kilometers. The aerosol amount was estimated using the dark pixel
reflectance ratio method proposed by Kaufman et al. (1997). If no suitable dark
pixels are found, which means no aerosol can be retrieved, the initial visibility of 40
kilometers will be used. The calculated visibility was 6.7535 kilometers.
The purpose of spectral polishing is to reduce spectral artifacts in the hyperspectral
data using linear transformation. FLAASH polishing algorithm smoothes the spectral
curve using the running average over n adjacent bands, where n is the spectral
width defined by user (ENVI, 2009). A wide polishing width of 7 was set to eliminate
large-scale artifacts around the absorption bands. The details of computation can
be referred to the ENVI help manual.
In FLAASH, three MODTRAN multiscatter models ordered in ascending accuracy and
computer resources – Isaacs, Scaled-DISORT and DISORT are available. The
multiscattering model tends to affect the shorter wavelengths (visible) while the
effect on longer wavelengths (near infrared) is minimal (ENVI, 2009). The Scaled134
DISORT MODTRAN multiscatter model with 8 scattering directions was selected
since it can balance the accuracy and speed. The finest MODTRAN Resolution of
5cm-1 was set.
FLAASH model allows the incorporation of atmosphere scattering into reflectance
calculation. The model distinguishes radiance reflected from surface directly
reaching the sensor as well as radiance from surfaces that is scattered by the
atmosphere into the sensor. The latter is removed by switching on adjacency
correction in FLAASH. Other parameters such as the aerosol scale height (1.5km),
CO2 mixing ratio (390 ppm) were set by default.
3.4.1.1.3. Wavelength recalibration
Image captured by pushbroom Hyperion sensor has an inherent problem caused by
variation of detectors’ spectral response function across the 256 pixels of a line. A
“smile” or “frown” effect was resultant from the intrinsic light dispersion properties
across the two detector arrays and small misalignment of optical elements (Jupp et
al., 2002, Felde et al., 2003, Gao et al., 2004). The smile effect causes a slight
difference in the wavelengths for pixels near the center of an array and those near
the edges of the same array (Mourioulis et al., 2000, Davis et al., 2002). Normally,
the effect is most prominent in the VNIR with about 2 – 4 nm wavelength shift and
effectively negligible in the SWIR with 0.5 – 1 nm shift. Bands in various
wavelengths were sampled and the shift of center wavelength across the 265 pixel
columns of selected bands with respect to the wavelength at the center pixel
column at 128 of the acquired image was shown in Figure 3.6. Similar to others’
observations, the wavelength shift was more significant in VNIR with absolute
maximum shift ranged between 2.76nm in band 46 and 3.14nm in band 12. The
wavelength shift is comparatively less significant in SWIR with absolute maximum
wavelength shift ranged between 0.49nm in band 203 and 0.68nm in band 90.
135
Figure 3.6. The center wavelength shift of selected bands with respect to the
wavelength at the center pixel column at 128 of the acquired Hyperion image
The shift of center wavelengths is likely to introduce considerable errors in the
water retrieval process, which in turn lower the accuracy of retrieved surface
reflectance (ENVI, 2009). The effects of wavelength shift are more apparent in
strong atmospheric absorption regions where spectral artifacts were introduced
after the atmospheric correction (Felde et al., 2003). In order to minimize the error
from spectral artifacts, the central wavelength was adjusted.
FLAASH provides an option called wavelength calibration to automatically adjust the
wavelength shift before water vapor retrieval. It utilizes the strong atmospheric
absorption features of oxygen notch at 760nm and carbon dioxide region at 2059
nm to quantify the wavelength shift error for VNIR and SWIR spectrometer
respectively. In order to correct the wavelengths, a model is computed by fitting the
shape of the absorption bands to the measured spectra (Felde et al., 2003). A
spectrograph is used to define the absorption features used through describing the
wavelength and the full-width half-maximum (FWHM) factor of the referenced
absorption features. The FWHM factor defines the wavelength span on each side of
the reference absorption feature for identifying the feature. The spectrograph for
Hyperion data is defined as
136
spectrograph = {762.6, 0.5, 1, 49}
spectrograph = {942.7, 0.75, 50, 196}
The two lines define the selected absorption features in VNIR spectrometer and in
SWIR spectrometer respectively. The wavelength in the center pixel of a band, i.e.
pixel 128, was assumed to be valid. It was checked by spatially dividing the 256
columns into 64 columnar subsets, with 4 columns wide each. The wavelength
recalibration algorithm was executed on each of these spatial subsets. On average,
the output wavelengths were adjusted by 1.5 and 0.75 for VNIR and SWIR
spectrometer respectively.
After atmospheric correction, bands with strong water absorptions were removed.
A total of 158 bands including 8-56 (428.4 – 916.8nm), 78 (923.3nm), 82-97 (963.7 –
1114.9nm), 102-120 (1165.4 – 1347nm), 134-164 (1488.3 – 1791nm), 183-224
(1982.6 – 2396.2) were retained for further processing.
3.4.1.1.4. SNR enhancement through MNF
The minimum noise fraction (MNF) transform was utilized to segregate noise
inherent in the data. The MNF transform derived as an analogue of the principal
component transform projects input data into a new feature space where
components are ordered by the strength of signal-to-noise ratio (SNR) instead of
data variance (Green et al., 1988, Lee et al., 1990, Amato et al., 2009). The noisiest
components can either be removed or smoothed in order to raise the SNR. MNF
was performed in ENVI. The noise statistics and eigenvalues were estimated during
the transformation. The MNF eigenvalues were used to remove noise components
which were plotted in Figure 3.7. A zoom plot was embedded in the Figure to show
the eigenvalues from 30 components onwards.
137
Zoom plot
Figure 3.7. The eigenvalues of 158 components after MNF transformation. The
embedded zoom plot shows eigenvalues from 30 components onwards.
The first component which contains mainly signals has eigenvalue of 60 and
eigenvalues of subsequent components decreased gradually. An eigenvalue of 1.5
was chosen to segregate noise from signal as indicated by the red dotted line in the
zoom plot in Figure 3.7. 77 bands were retained and transformed back to the
original feature space by inverse MNF transform. Figure 3.8 shows the effect of
noise removal before and after MNF transformation was applied to band 8 at
428.4nm. The salt and pepper noise was largely reduced after the transformation.
138
a
b
Figure 3.8. The effect of MNF noise removal of band 8 (428.4nm)
a) before MNF and b) after MNF
3.4.1.2. Geometric correction
Geometric rectification was performed so that reflectance values extracted from
the image are geographically accurate and can match with the existing database
and field measured data. The ortho-rectification was conducted in PCI Geomatica. A
total of 28 ground control points (GCPs) were collected with reference to the
geocoded coastline and road network vectors. The x, y and overall root-meansquare error were 0.04 (1.2 m), 0.06 (1.8 m) and 0.07 (2.1 m) respectively. The GCP
residual report is shown in Table A1 in Appendix A. The earth model of “TM Transverse Mercator” was used with the true origins defined as 114d10'42.8000"E
(longitude) and 22d18'43.6800"N (latitude); and local origins defined as 836694.050
(easting) and 819069.800 (northing). The ellipsoid of International 1924 was set.
The simple polynomial mathematical model was used to rectify the image. A digital
elevation model with spatial resolution of 5 meters was input during orthorectification process. Nearest neighbour resampling method was selected in order
to retain the original reflectance value for subsequent analytical procedures.
139
3.4.1.3. Atmospheric correction algorithms comparison
The reflectance obtained from the two algorithms, ATCOR-2 and FLAASH, were
compared with averaged laboratory-measured spectral signatures (described in
Section ) to test the performance of the two algorithms. By overlaying known
species location acquired through field surveys, reflectance curves of 6 species were
extracted from the two atmospherically-corrected images. The uncalibrated bands
and heavy water absorption bands were removed with 158 bands left for analysis.
The wavelength intervals of laboratory-measured spectra were resampled to match
with that of the Hyperion data. As the wavelengths were calibrated in the FLAASH
process, the laboratory spectra were resampled to calibrated wavelength
accordingly.
Correlation analysis and spectral feature fitting were used to generate statistics for
quantitative testing. Pearson correlation coefficient between mangrove species
spectrum of atmospherically corrected image and corresponding species laboratory
spectrum were calculated for the full spectrum as well as eight separate spectral
regions including blue, green, red, near infrared and four short-wave infrared (SWIR)
regions in SPSS 15.0. The correlation coefficients for the same species were
averaged and compared based on different spectral regions.
The spectral feature fitting compares the fitness of image spectra to laboratory
spectra based on the absorption features in the spectrum in ENVI. Continuum
removed was firstly applied to the mangrove spectra extracted from image and
laboratory-measured. The absorption depth was then measured for the image with
reference to the laboratory spectra using least-square method. The fitting process
produces two images, the scale image and the RMS image. A better match is
indicated by brighter pixels in the scale image approaching but less than one and
dark pixel in the RMS image indicating low error. Fit images can be derived by
dividing the scale image by RMS image. For each species, fit values were extracted
after the fitting process and they were plotted as scatterplot for comparison.
140
3.4.2. ASAR - ENVISAT
3.4.2.1. Data Acquisition
The advanced Synthetic Aperture Radar (ASAR) is on-board the ENVISAT satellite
launched by the European Space Agency (ESA) in March 2002. ASAR operates at Cband (5.3GHz or 5.66cm) and is designed to work in five data acquisition modes
including image mode, wave mode, alternating polarization mode, wide swath
mode and global monitoring mode. The image mode provides images at 30m with
look angles ranging from 15° to 45° covering swath widths ranging from 58 to
109km with HH or VV polarization. The alternating polarization mode produces
images that operate at two polarization modes at any time (HH and VV; VV and VH;
or HH and HV) at 30 m (Lillesand et al., 2008). The radiometric resolution is about
1.5
–
3.5
dB
(ENVISAT
website
http://envisat.esa.int/earth/www/area/index.cfm?fareaid=6).
A total of eight archived ASAR SAR images captured in year 2008 were acquired
from the Institute of Space and Earth Information Science (ISEIS) for mangrove
classification as shown in Table 3.4. The temporal series were all in VV-polarization
with relatively steep incidence angle of 23° at mid-swath and captured in ascending
orbit along the same track.
Table 3.4. The SAR data list for ASAR-ENVISAT
Date
Product
Mode
Polarization
Incident Angle
(degree)
Track/
Frame
Pass
Direction
13 Feb 2008
IM
C-VV
23.15
25/ 433
Ascending
19 Mar 2008
IM
C-VV
22.39
25/ 435
Ascending
23 Apr 2008
IM
C-VV
23.16
25/ 435
Ascending
02 Jul 2008
IM
C-VV
23.16
25/ 434
Ascending
06 Aug 2008
IM
C-VV
23.16
25/ 434
Ascending
15 Oct 2008
IM
C-VV
23.17
25/ 433
Ascending
19 Nov 2008
IM
C-VV
23.14
25/ 435
Ascending
24 Dec 2008
IM
C-VV
23.15
25/ 435
Ascending
141
Table 3.5 summarized the general weather conditions including daily maximum and
minimum temperature, total rainfall (rainfall three days before radar data
acquisition, if any), and amount of sunshine on day of image acquisition acquired
from weather database of the Hong Kong Observatory. Besides, the highest and the
lowest observed tide levels and at the time of data capture were checked at the
nearest meteorological station in Tsim Bei Tsui which is about 2 kilometers in
distance to the mangrove stand as a reference to the flooding condition. On 23 April,
a minimal amount of 0.4mm precipitation was recorded while on 06 August, an
amount of 74.1mm daily rainfall was recorded due to typhoon. No precipitation was
recorded for the other acquisition days. However, the precipitation on 30 June and
01 July would affect the soil moisture on data acquired on 02 July. Ebb tide was
observed on 13 February, 23 April, 06 August and 19 November while high tide was
recorded on 19 March, 15 October and 24 December. No observed tide information
was available on 02 July due to missing data.
142
Table 3.5. The general weather conditions and tide information for multi-temporal
ASAR data
Date
Local
Time
Daily
Min/Max
Temperature
(°C)
Total amount of
daily rainfall
(mm)
Amount
of
Sunshine
(hours)
Lowest/
Highest/
Acquisition Time
Tide level (m)
13 Feb 2008
22:19
9.7/ 14.2
Negligible
2.5
0.64/ 2.32/ 0.78
19 Mar 2008
22:19
19.8/ 28.0
No
5.8
0.01/ 2.39/ 2.20
23 Apr 2008
22:19
20.7/ 24.7
0.4
0.0
0.14/ 2.48/ 0.92
02 Jul 2008
22:19
27.0/ 32.2
9.2
No data
06 Aug 2008
22:19
25.2/ 27.0
0.1
0.79/ 2.88/ 1.00
15 Oct 2008
22:19
24.6/ 29.8
Negligible
9.7
0.28/ 2.68/ 2.67
19 Nov 2008
22:19
16.7/ 21.4
No
8.1
0.37/ 2.50/ 1.59
24 Dec 2008
22:19
15.0/ 20.8
No
2.8
0.56/ 2.54/ 2.09
Negligible/
30 Jun (48.5)/
01 Jul (4.0)
74.1
(Typhoon Signal
no. 8 hoisted)
3.4.2.2. Data Processing
The eight multi-temporal SAR data were processed using BEST (Basic Envisat SAR
Toolbox) version 4.2.2-b and PCI Geomatica. BEST is the software developed by
European Space Agency (ESA) to handle SAR products obtained from ASAR. It was
mainly used to perform radiometric correction for SAR data. Geometric correction,
signal-to-noise ratio (SNR) enhancement and feature extraction will be conducted in
PCI Geomatica. The image processing procedures are shown in Figure 3.9 and the
steps are described in details below.
143
Figure 3.9. The processing procedure of ASAR-ENVISAT data
144
3.4.2.2.1. Radiometric and Geometric Correction
Prior to image processing, the header parameters were decoded automatically in
Header Analysis. The header information is essential for full resolution image
generation in BEST. The full resolution amplitude images (in 16 bit integer) were
then extracted by defining the spatial extension of top left (22.60, 113.74) and
bottom right (22.25, 114.42) corner covering the Hong Kong territory. The intensity/
power images were computed by squaring the amplitude images and stored in 32bit floating point format. The intensity images were then calibrated to correct the
radiometric effects including the factor of incidence angle and absolute calibration
constant (European Space Agency, 2009). The resultant images were the
backscattering coefficient or sigma nought,   expressed in both linear and log
scale. The 32-bit floating backscatter images were exported to GeoTiff format for
subsequent processing.
ASAR image acquired on November 19, 2008 was defined as the master image in
the temporal series. It was geometrically rectified to Hong Kong 1980 Grid System
using the Transverse Mercator (Gauss-Krueger) projection with ellipsoid defined by
“International 1924”. An ortho-rectified SPOT 5 image acquired on November 11,
2008 together with geocoded vector data were used as references based on which
GCPs were collected. The root-mean-square error (RMSE) was used to determine
the accuracy of the 1st order polynomial transformed model. The x, y and overall
RMSE were 0.03 (0.9 m), 0.04 (1.2 m) and 0.035 (1.05 m) respectively. The GCP
residual report is shown in Table A2 in Appendix A. The cubic convolution method
which determines the grey levels of image by weighted average of 16 nearby pixels
was used to resample the image. The geometrically rectified image was resampled
to 30-meter in order to be compatible with the spatial resolution of Hyperion data.
Provided that the eight multi-temporal ASAR images were captured with the same
geometry in terms of satellite track, polarization mode, incident angle and pass
direction as shown in Table 3.4, their original geometries only experienced the
translation shifting. The seven images were automatically co-registered with the
geometry of the master image (geometrically rectified ASAR image acquired on
145
November 19, 2008) in PCI Geomatica. The translation vectors were derived using
tie points automatically generated by matching the master image and each of the
seven SAR images based on mean correlation method in the AUTOGCP module. A
minimum acceptance score of 0.7 was set to ensure quality GCPs were generated.
The GCPs were saved as individual GCPs segments in the .pix file, which were then
used for image registration using the REG module. The seven SAR images were
geometrically registered with the cubic convolution method using the 1 st order
polynomial. Each SAR image was visually examined for their geometric accuracies by
overlaying geocoded vectors such as roads and coast boundary. Since the
mangroves locating near sea level are the main area of the study, a more gentle
tolerance of errors was applied to the mountainous areas.
3.4.2.2.2. Speckle Filtering
One of the apparent differences of SAR image data from optical imagery is the
presence of distinctly grainy appearance or speckle (Lillesand et al., 2008). This
seemingly random pattern of darker and brighter pixels in radar image is resulted
from the coherent interaction of incident radiation with a very large number of
incremental scatterers within a pixel (Ndi Nyoungui et al., 2002, Richards, 2009).
Generally, speckles obscure the correct interpretation of SAR images, limits the
accuracy in classification and ground parameter extraction (Huang and Liu, 2007).
Therefore, speckle reduction methods are essential prior to further image
processing and analysis. An ideal speckle filter can lessen speckle variance by
smoothing homogeneous regions and can simultaneously retain high frequency
features such as edges and boundaries though it is always hard to balance between
the speckle suppression and detail preservation (Tso and Mather, 1999b, Ndi
Nyoungui et al., 2002, Richards, 2009). Over the years, considerable efforts have
been spent on developing numerous speckle suppression filters (Lee, 1980, Frost et
al., 1982, Kuan et al., 1987, Lopes et al., 1990, Baraldi and Parmiggiani, 1995,
Fukuda and Hirosawa, 1998). Early studies have compared the performance of
various filters (Durand et al., 1987, Baraldi and Parmiggiani, 1995, Ndi Nyoungui et
146
al., 2002). For instance, Ndi Nyoungui et al (2002) compared and evaluated 10 most
commonly-used speckle filters for land cover classification.
According to Lee (1986), there are two types of speckle reduction techniques –
multi-look processing and moving window. The multi-look processing or incoherent
averaging reduce noise variance and enhances the estimation of   by averaging
several frames obtained along the azimuth spectral bandwidth (Ndi Nyoungui et al.,
2002, Oliver and Quegan, 2004). However, the spatial resolution is sacrificed for
radiometric resolution and the high frequency information is lost after the
averaging. The moving window technique reduces speckles by running a moving
template of defined size over the image. Each pixel is replaced in turn by the
average value of pixels within the template (Richards, 2009). Commonly used
speckle filter techniques such as the Gamma filter (Kuan et al., 1987), the Lee filter
(Lee, 1980), the Enhanced Lee filter (Lopes et al., 1990), the Frost filter (Frost et al.,
1982) and the Enhanced Frost filter (Lopes et al., 1990) can suppress speckle noise
without losing significant high frequency features such as boundaries or edges.
Five speckle reduction filters including mean filter, median filter, Enhanced Lee
adaptive filter, Enhanced Frost adaptive filter, and Gamma filter were applied to the
entire master image (backscatter coefficient expressed in linear scale). The mean
and median filters are general low pass filters which smooth the image by
substituting the pixels with the average and median pixel value within the defined
window size. The Enhanced Lee adaptive filter reduces speckle as a function of
heterogeneity measured by the local coefficient of variation and classifies each pixel
into one of the three classes including homogeneous, heterogeneous and isolated
point target classes. For homogeneous class, the pixel value is replaced by the
average of the moving window while a weighted average value is applied to pixels
in heterogeneous class. The isolated point target class keeps the value unchanged
(Lopes et al., 1990). Similar to Enhanced Lee adaptive filter, the Enhanced Frost
adaptive filter splits the pixels into three classes. Each pixel value is computed
based on the distance from the filter center, the damping factor and local variance
(Lopes et al., 1990). By assuming data with gamma distribution, the Gamma filter
147
involves the transformation of multiplicative noise model to additive noise model
(Kuan et al., 1987).
The results achieved by different filters are dependent on window size. Speckle
reduction is inadequate if the kernel size is too small while over-large window size
can cause loss of image detail (Tso and Mather, 1999b, Ndi Nyoungui et al., 2002).
For every speckle filter, three moving kernel sizes including 3 x 3, 5 x 5 and 7 x 7
were tested. Different combinations of filter and kernel size were compared and
evaluated by two methods. First, homogeneous features with different backscatter
regimes, namely, dark grey, medium grey and light grey were picked from the raw
and filtered images. Then, statistical parameters including mean and standard
deviation were extracted from these homogeneous areas of various regimes. The
efficiency of speckle suppression was evaluated based on their ability to maintain
the mean backscattering coefficients of raw data while minimizing the standard
deviation within the homogeneous features (Wakebayashi and Arai, 1996). Besides,
the filters were compared by their performance in separating the species measured
by the transformed divergence which is defined as

  Dij
TDij  c 1  exp 
 8









1
1
T
Dij  tr Ci  C j  Ci1  C j 1  tr Ci1  C j 1 M i  M j M i  M j 
2
2
Eq. 3.1

Eq. 3.2
Prior to computation of transformed divergence, the linear unit of backscattering
coefficient was converted to decibels (dB) by applying the following equation:
10  log10 ( in linear unit )
Eq. 3.3
If the divergence among the species is the highest, the optimal filter and kernel size
were applied to the temporal series.
148
3.5. Field Measurements and Data Processing
3.5.1. Species Distribution
A mangrove species database was acquired from the Agriculture, Fisheries and
Conservation Department, Hong Kong SARG for species classification. The database
provides ground survey records for the distribution of different mangrove species in
the territory. The surveys were conducted in years between 2004 and 2007. The
point data was recorded in local HK1980 Grid System. Other than location, the
approximate spatial extent and tree height were stored in the attributes. As no
systematic survey scheme was used, the point data were examined and edited in
order to fit with the resolution of hyperspectral and SAR data, i.e. 30m. A 30 x 30 m
spatial grid covering the study area was generated. The point data were overlaid
onto the grids. If more than one survey points were found falling into the same grid,
the point closest to the center of the grid was retained while other points within the
same grid were removed. After this examination, the number of ground survey
points has been greatly reduced for some species.
Additional point data were obtained from two sources. Provided with constant
water and sediment supply, there is negligible change with respect to the structure
and pattern of mangrove species classes within the mangrove stand in Mai Po. A
high-resolution pansharpened Quickbird image acquired in 2006 was used to
facilitate extra sample point selection with assistance of an experienced ecologist.
Besides, additional species point data was acquired from LAI field survey that will be
discussed in Section 3.5.3.2. Figure 3.10 shows the spatial distribution of samples
points. Table 3.6 shows the total number of sample points of different species
extracted from different sources and also indicates the color scheme for the
different species shown in Figure 3.10.
149
Shenzhen River
Deep Bay
Figure 3.10. The spatial distribution of sample points of different mangrove species
Table 3.6. Combination of in-situ and manually collected sample point of different
mangrove species
After
Elimination
of overlaps
(30m)
Manually
selected
(30m)
Elimination
of impure
sample
points
Total
sample
points
(30m)
7
30
2
35
5
48
3
50
22
5
12
2
15
Avicennia marina
79
37
34
0
71
Kandelia obovata
1
Kandelia obovata
2
75
45
8
112
135
0
34
3
31
Sonneratia spp.
0
0
13
0
13
Species
Acanthus
ilicifolius 1
Acanthus
ilicifolius 2
Aegiceras
corniculatum
Field
survey
85
150
3.5.2. Leaf Spectra Measurement
While canopy reflectance recorded by remote sensors are mostly contributed by
leaves when compared with background or branches (Guyot et al., 1992, Asner,
1998), the collection of leaf spectra can be used to compared the canopy spectra
captured from remote platforms. Further researches have studied the relationship
between leaf reflectance and biophysical/ biochemical properties of leaves
(Buschmann and Nagel, 1993, Blackburn, 1998, Blackburn and Pitman, 1999). The
success of these studies has contributed to species discrimination and biophysical
modeling. Leaf spectra can be either measured in situ or in the laboratory while the
latter required leaves to be in fresh state. In situ field measurements of leaf
reflectance are influenced by many external factors such as the highly-variable
lighting conditions; wind condition, accessibility, which often renders promising
measurement results impractical (Foley et al., 2006).
An alternative solution is to transport the leaves to laboratory within variable time
periods, where measurements can be conducted (Horler et al., 1983, Asner, 1998).
The laboratory provides an ideal environment that minimizes the unnecessary
ambient impact on leaf spectra extraction. For instances, as the light environment is
controlled in the laboratory, the photochemical effects such as xanthophylls cycle
pigment changes and chloroplast movement due to variable light conditions can be
substantially reduced (Haupt and Scheuerlein, 1990, Gamon and Surfus, 1999).
However, the most significant concern of this approach is to minimize the change in
reflectance properties between the sampling and measurement time (Richardson
and Berlyn, 2002). Water loss of leaf is one of the main factors causing spectral
change as it influences the absorption of near infrared and short-wave infrared and
the scattering properties of leaves (Ripple, 1986, Hunt and Rock, 1989). As a result,
it is critical to maintain the water volume within the leaf samples, and therefore
requires sound leaf handling techniques. Many methods have been tested. For
instances, leaf samples are supplied with water and placed within plastic bags in
order to preserve a humid microenvironment where vapour pressure gradients
between air and leaf is reduced (Horler et al., 1983). The artificial cooling of leaves
with the purpose of reducing transpiration has been mostly applied in many
151
researches (Lacaze and Joffre, 1994, Sims and Gamon, 2003). Richardson and Berlyn
(2002) explored the effectiveness of water loss reduction using moist paper towel
and plastic bags through analyzing various vegetation indexes while Foley et al.
(2006) used the same technique to explore the spectral change between 350 nm
and 2500 nm and found that the leaf water content tends to affect the spectral
shape rather than the magnitude of raw reflectance.
3.5.2.1. Leaf Collection and Handling
Leaf samples of six mangrove species were collected for laboratory-based spectral
measurement during Oct – Nov 2007. Leaf samples of the native species including
Acanthus ilicifolius, Aegiceras corniculatum, Avicennia marina and Kandelia obovata
were collected along the bridges so as to shorten the time between leaf collection
and measurement. Since the two exotic species, Sonneratia caseolaris and
Sonneratia apetala can only be found in the outskirt of Mai Po protected area and it
was only accessible with the aid of hovercraft. The locations where the leaves were
sampled were recorded with the Trimble GPS GeoXH Handheld as shown in Figure
3.11. During leaf collection, the whole branches were cut and put into plastic bags.
All the leaves collected from the non-active growing part of the tree, except
Acanthus ilicifolius.
152
By Hovercraft
Along the bridge
Figure 3.11. The map showing the locations of leaf sample collection
In order to reduce the transpiration which can in turn lessen the rate of spectral
change, the leaf samples were cooled with ice and stored in the thermal insulating
containers after detaching from the main branches as shown in Figure 3.12. They
153
were then transported back to the laboratory within 2 hours and refrigerated in
laboratory freezers.
Figure 3.12. Collected leaves cooled with ice in a thermal insulating box
3.5.2.2. ASD FieldSpec 3 Setup
FieldSpec® 3 Portable Spectroradiometer (FS3) from Analytical Spectral Devices, Inc.
was used to measure the spectral response of the seven mangrove species under
laboratory setup. FS3 is equipped with one Silicon photodiode array detecting
visible and near infrared (VNIR) spectra ranging 350 – 1000 nm and also two
separate InGaAs photodiode arrays sensing the spectral response of shortwave
infrared (SWIR) from 1000 to 2500 nm. A fiber optic sensor with 25° field of view
was used. FS3 is designed to capture spectral data ranging from 350 – 2500 nm with
fine sampling interval of 1.4 nm and 2 nm at 350 – 1000 nm and 1000 – 2500 nm
respectively. The spectral resolution (full width half maximum) is 3nm for the
wavelength interval 350 nm – 1000 nm, and 10 nm for the wavelength interval 1000
nm – 2500 nm.
FS3 was warmed up for 90 minutes before any calibration or measurement began.
Since the three photodiode arrays warm up at different rates, the warm up is
154
necessary to avoid the spectral steps that are very likely to occur at the overlap
regions at wavelength of 1000nm and 1800nm as shown in Figure 3.13.
Spectral steps
Figure 3.13. Spectral steps at the wavelength overlap regions (i.e. 1000nm and
1800nm)
Prior to spectra measurement, the sensor was optimized for dark current and white
reflectance scan. Dark current refers to the electrical current generated by thermal
electrons within FS3 that was added to the measured incoming photons of light.
The electrical component is likely to falsify the measurement. The two SWIR
detectors can automatically eliminate the dark current while dark current has to be
subtracted from the VNIR array on a channel-by-channel basis. Dark current
optimization was conducted in every twenty minutes during the measurement.
The white reflectance scan is to transform the raw digital number (DN) to relative
DN. A 99% reflectance panel (Spectralon, Lapsphere, North Sutton) was used to
calibrate the absolute reflectance factor. The ratio between DN of mangrove foliage
relative to the DN of the Spectralon white reflectance was displayed for real-time
quality control as well as stored in computer memory. During the white reflectance
155
calibration, the fiber optic sensor was placed on top of the Spectralon panel.
Successful calibration will result in a stable 100% line of Spectralon as shown in
Figure 3.14. The Spectralon panel was used to check the quality of 100% line every
twenty minutes in order to ensure correct absolute reflectance is attained.
Figure 3.14. The 100% line during white reflectance calibration
(DN of Spectralon/ DN of Spectralon = 1)
The spectra were then sampled and saved. The larger the number of samples for
each scan is, the higher the signal-to-noise ratio. It is especially important to the
short-ware infrared (SWIR) bands because the spectra will be dominated by noise if
the number of samples is below 25. However, if more sampling is conducted, the
time of scan will be lengthened. In order to balance the SNR and computation
efficiency, the average number of dark current subtraction, white reference
calibration and spectrum measurement was set to 30.
3.5.2.3. Laboratory setup
In order to reveal the true spectral signal of the leaves unaffected by the ambient
light sources, all the spectra measurements were taken place in dark room at room
temperature of 25°C. Two halogen lamps (ProLamp©) were setup to provide a
stable electro-magnetic energy between 400nm – 1800nm. The halogen lamp was
selected because the energy range was reconciled with the majority of
hyperspectral sensors (Lillesand et al., 2008). The two halogen lamps were fixed on
the tripods with one illuminating on the left while another on the right of the
156
measurement table at an angle of approximately 45° from the normal to the surface.
Each lamp was warmed up for five minutes before measurement and a diffuse light
source was set.
A rotating plate with a dimension of 33.5cm x 33.5cm was designed to allow the
measurement of leaf spectra in different directions in order to average out the
differences in orientation of each measurement on the plate. The measurement
table and rotating plate were covered with black cloths in order to provide equal
illumination and minimize the reflectance from other objects. The laboratory setup
is shown in Figure 3.15.
Figure 3.15. The laboratory setting for spectra measurement
The optical fibre sensor was fastened on the pistol grip which was then stabilized on
a tripod. With 25° field of view, the optical sensor was mounted on a height of 70
cm (27.56 inches) above the rotating plate at nadir position giving a projection of
754.39 cm2. The projection cone is illustrated in Figure 3.16.
157
Rotating Plate
33.5 x 33.5cm
(13.2 x 13.2inches)
Figure 3.16. The illustrative diagram showing 25° projection cone
3.5.2.4. Spectra Measurement
All sample leaves were collected in the morning while spectral measurement was
conducted in the afternoon on the same day in order to prevent the dehydration
and structural changes of the collected leaves. Prior to measurement, the leaves of
each mangrove species were shed off from the branches and were randomly
divided into 30 piles with 20 – 30 leaves per pile. The leave samples were spread
randomly and covered the rotating plate completely. For each measured spectrum,
it was averaged from 25 samples. For each sampling plate, five spectra were
recorded at every 90° horizontal angle, i.e. 0°, 90°, 180° and 270° using RS3 Spectral
Acquisition software provided by ASD. As a result, a total of 20 spectra were
recorded for a sampling plate and they were subsequently averaged to construct a
spectrum curve representing the reflectance spectrum of the mangrove species
measured on that sample plate. The leaves were then removed from the plate and
replaced with new leave samples. The steps above were repeated for all the leaf
plates. For each measurement, information including the starting time, time since
spectroradiometer optimization (dark current and white referencing), measured
spectra name and number, species, and photo numbers were marked on the
designed measurement log sheet as shown in Appendix Z. In addition, photos were
158
taken for every sampling plate. Sample photos were shown in Figure 3.17. For each
mangrove species, at least 30 reflectance curves were constructed. The spectral
data of each species were then exported and a matrix was formed for subsequent
analysis.
Figure 3.17. The sampling plate during spectra measurement
3.5.2.5. Spectral similarity and variability
The reflectance spectra of the six mangrove species were statistically analyzed to
test if the variance of reflectance between species is significantly greater than
within species. The Mann-Whitney U-test was chosen to test the null hypothesis for
all pairs of species that there is no signification difference between the median
reflectance of each individual band. Mathematically, the null hypothesis and
alternative hypothesis are expressed as:
H 0   n (i)   n1 (i)
Eq. 3.4
H1   n (i)   n1 (i)
Eq. 3.5
159
where  n is the median reflectance of species n  1,2,3,( N  1) , and i  1,2,3,I
is the wavebands. The Mann-Whitney U-test is non-parametric, i.e. no assumption
of normal distribution for the reflectance in the wavebands for all species-pairs. The
observations from the species group should satisfy two very basic assumptions,

groups of observations are random and independent of each other; and

The observations are numeric (ordinal, interval or ratio)
The observations in a tested species-pair under a specific wavelength were
combined and ranked in ascending order. The ranks for each species was then
added up and U -values for two species groups, n and n  1 , were calculated for
each species,
U n  PQ 
P( P  1)
  Rank (n)
2
U n1  PQ 
Q(Q  1)
  Rank (n  1)
2
Eq. 3.6
Eq. 3.7
where P and Q are total number of samples in species groups n and n  1
respectively;
 Rank (n) and  Rank (n  1) are
summation of ranks in species groups n and
n  1 respectively. The smaller of these 2 U -values were then compared with the
lookup tables for significance testing. The null hypothesis was tested at significant
level of   0.01. If   0.01 for a species-pair in a particular wavelength, the null
hypothesis was rejected and the alternative hypothesis in Eq. 3.5 was accepted
stating that the median reflectance between the species-pair are not equal.
Significance tests were computed for each species pair. The total number of
possible pairs from 6 species is:
N!
6!
720
N


 15
 
 2  ( N  2)!2! (6  2)!2! 48
Eq. 3.8
160
After the computation of 2151 wavebands for the 15 species pairs, the wavelength
positions at which the species-pair are significantly different were examined.
Besides, the number of significantly different species-pairs was counted according
to wavebands in order to identify wavelengths that are most prominent in resolving
spectral differences among mangrove species. Apart from pairwise comparison,
wavelength at which the median reflectance of each individual mangrove species is
significantly different to the median reflectance of all other five species was
identified. The reflectance at these wavelengths characterizes spectral property of
individual species that are potentially different to other species, which can be
considered for species discrimination.
3.5.3. In situ Leaf Area Index Measurement
In situ leaf area index (LAI) was acquired by optical measurement of diffuse
radiation under the canopy through hemispherical photo capture using the
WinSCANOPY imaging system (Régent Instruments Inc., 2008). The commonly used
LAI-2000 Plant Canopy Analyzer was not chosen as its applicability is highly limited
by site characteristics and accessibility. Mangroves in some locations are too tall
(about 8-10 meters) for efficient acquisition of above-canopy reading in two-sensor
operation mode. Besides, although the above reading can be measured using a
separate set of sensor, it is difficult to find an open area large enough for
simultaneous sky radiation measurement that are free from the effect of
environmental scattering.
3.5.3.1. The optical instrument
The WinSCANOPY imaging system consists of five main components including a
digital camera, a fisheye lens, an O-Mount, a NorthFinder and a remote control for
effective and efficient image capture under the canopy as shown in Figure 3.18. The
Nikon Coolpix P5000 digital camera with effective pixels of 10 million was attached
161
with a fisheye lens (FC-E8) calibrated for 180° field of view and position of
hemispherical projection by Régent Instruments Inc. The O-Mount is a self-leveling
unit used to automatically hold the camera in horizontal position during image
acquisition. Not only can the automatic leveling save time especially when the
mangrove substratum is unstable, it also ensures quality image and reduce
manipulation errors. The Northfinder provides a real time indication of magnetic
north direction using a LED around the fisheye lens. During photo taking, the LED is
captured and presented as a red dot on the edge of the photo. The north direction
is ideal to calculate the sun paths and radiation levels during the time of image
acquisition. Finally, the remote controller which allows image capture at a distance
up to 25 meters offers two main advantages. First, it minimizes undesirable manual
distortion to the horizontal leveling of the camera. Second, operating on radio
wavelengths, the remote control is not affected by ambient light and obstacles like
tree stems, branches and leaves.
FC-E8 Fisheye Lens
Northfinder
Remote Control
O-Mount
Nikon Coolpix P5000
Figure 3.18. The WinSCANOPY Imaging System
During field survey, the imaging system was mounted on tripod. With unstable
substratum, it was hard to set a consistent height level for every measurement. The
approximate height for photo taking is about 1.5 meters above the ground.
162
The camera parameter settings including image quality, image size and exposure
level are critical to image quality which affects the precision of subsequent LAI
extraction (Frazer et al., 2001, Inoue et al., 2004). Optimal camera settings are
essential to obtain high contrast photographs from which the canopy elements can
be effectively distinguished from sky (Rich, 1990). The image quality was set to
JPEG-fine which is the best image quality attained by the camera. The image size
was set to 3648 x 2736 pixels, which is the highest resolution available. For
exposure settings, ISO 64 (the finest available) was used; the shutter speed of 1/125
second or faster was used to freeze the motion of leaves caused by wind (Rich,
1990); and the focus distance was set to infinity. In order to enhance the contrast
between vegetative components and the sky, photos underexposed by 1.0 f-stop
and 2.0 f-stops were taken using the exposure blanketing technique. For each
sampling point, a total of three photos were captured (normal exposure, -1.0EV and
-2.0EV).
3.5.3.2. The LAI survey campaign
LAI sampling was conducted from mid-October to end of November in 2008 which
was scheduled to coincide with the acquisition of the Hyperion data. Four transects
with lengths varied from 550 meters to 1300 meters were designed to cover a total
of 95 sampling quadrats as shown in Figure 3.19. Two species, Kandelia obovata and
Avicennia marina dominating the top canopy were covered during the survey.
Although Acanthus ilicifolius and Aegiceras corniculatum exist, they mainly occupy
in the form of undergrowth. Besides, Acanthus ilicifolius are short and denselygrown herbaceous which make photo taking impossible even though they are
largely found along the mangrove fringe.
From theoretical perspective, photos should be captured under diffuse light
condition with uniform sky luminance. However, it is always hard to achieve as the
sky luminance is high variable during some measurement periods. In order to
acquire image data with reasonable quality, the camera settings have been checked
before each measurement so as to minimize exposure errors such as image blur.
163
Besides, no photographs will be taken during windy conditions. Last but not the
least, sun-blocker supplied with the imaging system was sometimes used when
strong direct sunlight was encountered.
Figure 3.19. The LAI sampling transects and locations in Mai Po
164
Each sampling quadrat is about 30 x 30 meters covering an area of approximately
900 square meters. Within each quadrat, photos were taken at four locations with
one sitting in the centre of the quadrat as illustrated in Figure 3.20. A is the central
measurement point while B, C and D are sub-sampling points chosen to have
representative spatial coverage within the quadrat. The coordinates of the central
location of each quadrat was recorded with the Trimble GPS GeoXH Handheld
attached with a hurricane antenna for signal magnification. The ArcPadTM installed
in GPS recorded the plant species of the quadrat and types of understory (e.g.
vegetation types, soil and water). Figure 3.21 shows some of the field photos during
the survey.
Figure 3.20. The illustrative diagram showing the ideal measurement locations
within a sampling quadrat
165
Figure 3.21. Field photos during the survey
3.5.3.3. Data processing and canopy analysis
The photos were taken back to laboratory for computer analysis. After downloading
the photos, they were systematically renamed according to their transect number,
quadrat number and exposure level. The images of three different exposure levels
were pre-screened for their qualities in terms of image contrast. After visual
examination, the set of image files with -2.0EV were selected for further processing
and canopy analysis in WinSCANOPY 2006a due to relatively higher contrast
between sky and vegetative elements. Three main steps were involved including
image defects removal, areal definition for analysis and pixel classification before
LAI and other related biophysical parameters can be computed.
While the quality of images has direct impact on the accuracy of LAI extraction, the
first step was to try to repair image defects caused by undesirable illuminating
166
conditions such as the presence of lens flare and uneven luminance using image
editing techniques. Figure 3.22 shows some sample images in which possible
illumination defects were identified. Figure 3.22a illustrated the perfect and
uniform diffuse conditions under which the sky and vegetative elements can be
distinguished confidently. In Figures 3.22b, the strong and direct sunlight causes an
extreme bright object in a large opening and obscure part of the image. Figure 3.22c
shows the varied sky condition and Figure 3.22d illustrates the presence of lens
flare due to internal reflection.
Figure 3.22. The sampled images acquired under different illumination conditions
a) ideal uniform sky condition; b) presence of strong light source; c) varied sky
condition; d) presence of lens flare.
The defects were repaired using different photo-editing measures such as contrast
enhancement, histogram editing and exposure adjustment in Adobe Photoshop.
However, some flaws required fine manual adjustments. For instances, the direct
strong light source and lens flare were replaced by surrounding textures and colors.
167
The tree trunks or large branches shined by direct sun were painted in black so as to
avoid being classified as gap.
After image repair, they were loaded into WinSCANOPY for hemispherical definition
and projection. The purpose was to define the area for analysis and to project the
sky grids defined by zenith annuli and azimuth angles onto the images. The
equilinear projection, which is the most common method for fisheye lens projection,
was adopted by WinSCANOPY. Since the FC-E8 fisheye lens was well-calibrated by
Régent, no further adjustment was needed to ensure precise relationship between
zenith angles and projected distance.
First, the area for analysis was specified through the interactive hemisphere
identification method while the north direction was located by pointing the radius
line to the red LED indicated by the NorthFinder. Then, the sky grids and suntracks
were automatically generated and overlaid on the image as shown in Figure 3.23.
The red circle delimits the maximum hemispherical boundary within which the
canopy analysis will be conducted. The sky grids in yellow circles divide the
hemisphere into regions in terms of zenith and azimuth for canopy measurement.
The number of zenith rings with equiangular views was set to 18 while the number
of azimuth slices was set 8.
168
Figure 3.23. The hemispheric photo overlaid with suntracks (rainbow in the middle
of image) sky grids (yellow lines) divided according to self-defined azimuthal and
zenithal divisions
The final and most critical step was to select a threshold transforming the grey
levels of image pixels into binary, 0 (black) or 1(white), representing canopy and sky
respectively. The blue channel was suggested for pixel classification because the
effects of multiple scattering and chromatic aberration are minimized in the band
(Zhang et al., 2005). Hence, the vegetative elements appear a relatively darker tone
in blue band than that in other bands. However, due to limited functions available
in the regular package of WINSCANOPY, the option of selecting single band for
analysis was not available. Instead, a threshold was manually selected by visual
interpretation with a view to maximally separate the vegetative elements from the
sky for each photo. In the process of visual inspection, the analyst toggled between
the original image and classified image as well as to make reference to the grey
level/ color histogram for more precise threshold selection. In order to have a
169
consistent analysis and result, an experienced analyst was responsible to classify all
photos (95 sampling quadrats x 4 photos = 380 photos). Figure 3.24 shows the
interactive threshold selection interface in deriving the binary image. After a
reasonable threshold was selected, parameters describing the canopy structure
including gap fraction, LAI (computed with different method), leaf inclination angle
distribution, clumping index were calculated automatically.
Figure 3.24. The transformation from raw to binary image through interactive
threshold selection
3.5.3.4. Canopy parameter computation – gap fraction, LAI, clumping index, mean
inclination angle
After the canopy analysis, a number of canopy measurements were derived from
the WinSCANOPY program. Table 3.7 shows the six categories of canopy
measurements including sky openness, LAI derived from linear and log methods,
clumping index, site factor and radiation data adopted from Régent Instruments Inc.
(2008).
170
Table 3.7. The canopy measurements extracted from hemispherical photos
(Adopted from Régent Instruments Inc., 2008)
Categories
Canopy
measurement
Gap fraction
Sky
Openness
Openness
LAI(Bonhom)-Lin
LAI(2000)-Lin
LAI(2000G)-Lin
LAI Linear
Method
LAI(Sphere)-Lin
LAI(Ellips)-Lin
MeanLeafAngle-Lin
LAI(Bonhom)-Log
LAI(2000)-Log
LAI(2000G)-Log
LAI Log
Method
LAI(Sphere)-Log
LAI(Ellips)-Log
MeanLeafAngle-Log
NofZeroGapFrRgn
Clumping
Index
CI(Bonhom)
CI(2000)
CI(2000G)
CI(Sphere)
CI(Ellips)
DirectSiteFactor
Site Factor
IndirectSiteFactor
TotalSiteFactor
Radiation
Data
PPFDDirectOver
2
[MJorMol/m day]
PPFDDiffuseOver
2
[MJorMol/m day]
PPFDTotalOver
Definition and Descriptions
The fraction of pixels classified as open sky in a sky grid
region in the image (in a two dimensional space)
The fraction of pixels classified as open sky in a specific
region of the real canopy above the lens (taking
elevation into consideration)
Leaf Area Index calculated with the Bonhomme and
Chartier linear method (at 57.5°)
Leaf Area Index calculated with Licor LAI2000 linear
method (5 rings with 148° field of view)
Leaf Area Index calculated with Licor LAI2000
generalized linear method using the same zenith rings
as the sky grid
Leaf Area Index calculated with a spherical leaf angle
distribution and linear method using the same zenith
rings as the sky grid
Leaf Area Index calculated with the ellipsoid linear
method using the same zenith rings as the sky grid
Mean Leaf Angle calculated from the ellipsoid linear
method (in degree)
Leaf Area Index calculated with the Bonhomme and
Chartier log method (at 57.5°)
Leaf Area Index calculated with Licor LAI2000 log
method (5 rings with 148° field of view)
Leaf Area Index calculated with Licor LAI2000
generalized log method using the same zenith rings as
the sky grid
Leaf Area Index calculated with a spherical leaf angle
distribution and log method using the same zenith rings
as the sky grid
Leaf Area Index calculated with the ellipsoid log method
using the same zenith rings as the sky grid
Mean Leaf Angle calculated from the ellipsoid log
method (in degree)
Number of regions of the diffuse sky with a gap fraction
equal to zero
Clumping Index = LAI(Bonhom)-Lin / LAI(Bonhom)-Log
Clumping Index = LAI(2000)-Lin / LAI(2000)-Log
Clumping Index = LAI(2000G)-Lin / LAI(2000G)-Log
Clumping Index = LAI(Sphere)-Lin / LAI(Sphere)-Log
Clumping Index = LAI(Ellips)-Lin / LAI(Ellips)-Log
Proportion of direct radiation relative to that in the
open sky
Proportion of diffuse radiation relative to that in the
open sky
Proportion of direct + diffuse radiation relative to that in
the open sky
Direct photosynthetically active flux density over canopy
average for the growing season
Diffuse photosynthetically active flux density over
canopy average for the growing season
Total (direct + diffuse) photosynthetically active flux
171
2
[MJorMol/m day]
PPFDDirectUnder
2
[MJorMol/m day]
PPFDDiffuseUnder
2
[MJorMol/m day]
PPFDTotalUnder
2
[MJorMol/m day]
density over canopy average for the growing season
Direct photosynthetically active flux density under
canopy average for the growing season
Diffuse photosynthetically active flux density under
canopy average for the growing season
Total (direct + diffuse) photosynthetically active flux
density under canopy average for the growing season
WINSCANOPY uses the gap fraction method to compute LAI. Gap fraction is defined
as the proportion of pixels classified as open sky in a sky grid region. Based on gap
fraction, WinSCANOPY provides five LAI computation methods including Bonhomme
and Chartier method (Bonhomme and Cihlar, 1995), LiCor’s LAI-2000 method
(Welles and Norman, 1991b), generalized LiCor’s LAI-2000 method, spherical
method, and ellipsoid method (Campbell, 1985). The five LAI estimates can be
compensated with or without leaf clumping calculation which is referred to log and
linear method respectively. The clumping index of different LAI estimates was
computed by dividing the LAI-linear by LAI-log. The radiation parameters including
site factors, photosynthetically active flux density are also reported.
Regarding different LAI estimates, the Bonhomme and Chartier method,
abbreviated as LAIBon, measures gap fraction at zenith angle of 57.5° at which gap
fraction is assumed to be insensitive to leaf angle distribution. LAI was calculated
using a simple formula (Bonhomme and Chartier, 1972)
LAI  1.1ln[T (57.5)]
Eq. 3.9
The Li-Cor’s LAI-2000 canopy analyzer computation method, abbreviated as LAI2000,
measures gap fraction at five default range of zenith rings (  ) centered at 7°, 23°,
38°, 53°, and 68° with a total field of view 148°, which was then converted to LAI
using the Miller’s (1967) theorem (Welles and Norman, 1991b). The equation is
expressed as
LAI  2  ln(T ( )) cos  sin d

Eq. 3.10
172
The generalized Li-Cor’s LAI-2000 canopy analyzer computation method, LAI2000G,
applies the same radiation transfer theory as the LAI2000 method does. However, it
has been generalized for any number of zenith rings and field of view defined by
user, i.e. 18 equilinearly projected zenith rings and 180° field of view in this study
(Régent Instruments Inc., 2008).
The spherical method, LAI(Sphere), assumes that the leaf angle distribution is
identical to those of a sphere with the leaf projection coefficient equal to 0.5 in any
direction. LAI(Sphere) uses the same sky grid definition as in LAI(2000G) (Régent
Instruments Inc., 2008).
Finally, the ellipsoid method, LAI(Ellips), assumes the leaf area density distribution is
ellipsoidal and computes parameters including LAI, mean leaf angle and leaf angle
distribution through non-linear elimination curve fitting (Campbell, 1985).
Conventionally, LAI estimation through gap fraction assumes that the foliar
elements are randomly distributed in terms of the azimuth angle (Chen et al., 1997).
However, the assumption tended to underestimate LAI ranged from 25 – 50% in
different stands due to the non-random or discontinued canopy (Gower and
Norman, 1991, Cutini et al., 1998, Gower et al., 1999). The degree of LAI
underestimation depends on the deviation of vegetation structure from the
assumption of random distribution of the foliar elements, which is described by the
clumping index (Lang, 1987, Chen et al., 1997, Kucharik et al., 1997, Van Gardingen
et al., 1999). WinSCANOPY adopted a simple log-average method for clumping
compensation described in Van Gardingen (1999) and initially proposed by Lang and
Xang (1986). The method divides the zenith annuli,  into n segments delimited by
the azimuth slices of the sky grid. Individual segments under a specific zenith
annulus are represented by n as shown in Figure 3.25. Instead of computing the
logarithm of averaged gap fraction of the whole zenith annuli under conventional
practice, the logarithms of individual gap fractions for each segment of the zenith
173
annuli, ln T (n) were computed, summed and averaged by the number of segment
in the annulus, S . It is expressed as
 n1 ln T (n)
S
ln T ( ) 
S
Eq. 3.11
The clumping indices are calculated by dividing the linear LAI by the log LAI. The
notion behind the log-average method is that the foliage distribution should be
random in a relatively small segment. Rather than applying a constant correction
factor, the log-average method can significantly reduce LAI underestimation of
canopies where the scale of clumping is not uniform (Van Gardingen et al., 1999).
Figure 3.25. An illustrative diagram showing the log-average method
As the foliar distribution was observed to be non-random during field survey and
from the hemispherical photo, the clumping compensation was taken into account
during LAI computation. For the 18 zenith annuli, each was divided into 8 azimuth
segments. The LAI of each measurement point was computed using the log-average
174
method. Finally, the LAI of each quadrat was calculated as the average of the four
measurement points within the quadrat.
3.5.3.5. Field LAI and Their Correlation with Reflectance and Backscattering
Coefficient Data Exploration
Descriptive statistics including mean, standard deviation of field measured LAI were
examined. The correlation between field-measured LAI and reflectance of each
band as well as the backscattering coefficient were investigated.
3.6. Feature Selection
The curse of dimensionality suggests that high dimensional input data will
deteriorate rather than increase the classification accuracy if training samples are
insufficient. Therefore, feature selection is an indispensible process to remove
irrelevant and redundant parts of data and reveal data bands that can truly
influence the classification. Not only can feature selection save processing time, it
can also enhance understandings of physiology of mangrove species based on their
responses to spectral bands or textural data.
Prior to further analysis, bands derived from hyperspectral and radar backscattering
datasets were selected using wrapped-based and hybrid feature selection methods.
The feature selection process was conducted using the C++ feature selection library
– Feature Selection Toolbox 3.1 (FST3) (Somol et al., 2010b). Developed by the
Institute of Information Theory and Automation (UTIA), Academy of Sciences of the
Czech Republic, Prague, FST3 is a free and widely applicable toolbox specialized in
dimensionality reduction and stability assessment. The toolbox provides a number
of optimal and sub-optimal search algorithms, evaluation criteria (wrapper- and
filter based; normal and multinomial) and model stability and overfitting
measurement tools to reduce data dimensionality and maximize the accuracy of
models. The details can be found the website published at http://fst.utia.cz/.
175
FST3 is a collection of C++ source codes. The executables for feature selection were
compiled after their integration with external libraries Boost version 1.44 and
LibSVM version 3.0 in Visual Studio 2010. Both Boost and LibSVM are free external
libraries. The former is C++ standard library while the later is a source code of
library C++ for support vector machines (SVM) developed by Chang and Lin (2011)
which support the use of SVM as classifier during wrapper-based feature selection.
Feature selection programs were compiled using different evaluation criteria and
search algorithms provided in FST3. Figure 3.26 shows the procedural flow of
feature selection based on the architectural diagram in Somol et al. (2010).
Provided that the purpose of feature selection was to achieve high classification
accuracy of mangrove classes, the wrapper-based and hybrid approaches were
adopted. Besides, given the high dimensionality which prohibits optimal search,
only suboptimal search algorithm was applied. The processing details are described
below. The scripts written in C++ language for feature selection are documented in
Appendix B.
176
Figure 3.26. The procedural flow of feature selection in FST3
(Based on Somol et al. 2010)
177
3.6.1. Data Preprocessing and Preparation
Spectral reflectance, backscattering coefficients and geo-referenced locations (in
Hong Kong 1980 Grid System) were extracted from satellite images covering the
area of the study. A pixel described by a set of bands/ features/ dimensions is called
an observation/ a pattern using pattern recognition terminology. For pixels at which
the mangrove species were known (as illustrated in Section 3.4.1), species labels
were attached to these patterns. These labeled patterns were used to train the
feature selection algorithms. There were a total of 190 features and 327 labeled
patterns. Prior to feature selection, the feature and observations were examined
and appropriate data conversion measures were applied in order to come up with a
more reliable dataset as input to feature selection process.
In terms of features, the hyperspectral reflectance and backscattering coefficients
were treated separately due to different nature of data. The atmospherically
corrected reflectance exhibited negative values mainly in blue bands with some in
SWIR bands. The negatives were all changed to zero for all the labeled patterns. It
was observed that the majority of observations have zero value in five bands at the
wavelengths from 428.44 to 469.14nm. Considering that these five bands provide
no significant information to the any process, they were discarded for further
processing. After that, the reflectance was normalized to a range between 0 and 1.
No adjustments were made to backscattering coefficients. Table 3.8 and 3.9
summaries the final feature set, D extracted from the hyperspectral and radar
dataset respectively. With a total number of 185 features, 153 were spectral
reflectance bands extracted hyperspectral data atmospherically corrected by
FLAASH while the remaining 32 features were multi-temporal SAR backscattering
coefficients expressed in linear and dB scale. As the number of features/ bands
exceeds 100, it is a high dimensionality problem.
178
Table 3.8. Summary of hyperspectral features (0-152)
Feature Number
Band number in
original Hyperion
Wavelength (nm)
Spectral bands
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
479.31
489.49
499.65
509.82
519.98
530.13
540.29
550.44
560.59
570.74
580.89
591.04
601.19
611.34
621.50
631.66
641.83
652.00
662.17
672.35
682.53
692.72
702.92
713.11
723.30
733.49
743.68
753.86
764.04
774.23
784.43
794.64
804.84
815.04
825.24
835.42
845.59
855.77
865.94
876.12
886.29
896.47
Blue
Blue
Blue
Blue
Green
Green
Green
Green
Green
Green
Green
Green
Green
Green
Red
Red
Red
Red
Red
Red
Red
Red-edge
Red-edge
Red-edge
Red-edge
Red-edge
Red-edge
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
NIR
179
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
55
56
78
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
134
135
136
137
138
139
140
141
906.64
916.82
923.30
963.67
973.76
983.84
993.93
1004.06
1014.06
1024.16
1034.26
1044.35
1054.45
1064.55
1074.65
1084.74
1094.84
1104.94
1114.93
1165.41
1175.51
1185.61
1195.70
1205.80
1215.90
1225.89
1235.99
1246.08
1256.18
1266.28
1276.38
1286.48
1296.57
1306.67
1316.77
1326.77
1336.87
1346.97
1488.29
1498.40
1508.50
1518.60
1528.71
1538.71
1548.81
1558.92
NIR
NIR
NIR
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR1
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR2
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
180
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
1569.02
1579.12
1589.22
1599.33
1609.43
1619.53
1629.63
1639.63
1649.73
1659.84
1669.94
1680.04
1690.14
1700.24
1710.33
1720.43
1730.53
1740.53
1750.62
1760.72
1770.82
1780.91
1791.01
1982.60
1992.70
2002.80
2012.90
2023.00
2033.10
2043.19
2053.19
2063.29
2073.38
2083.48
2093.58
2103.67
2113.77
2123.86
2133.96
2144.05
2154.05
2164.14
2174.24
2184.33
2194.43
2204.52
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR3
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
181
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
2214.62
2224.72
2234.81
2244.91
2254.90
2265.00
2275.10
2285.20
2295.29
2305.39
2315.49
2325.59
2335.69
2345.78
2355.88
2365.88
2375.98
2386.08
2396.18
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
SWIR4
Table 3.9. Summary of multi-temporal SAR features (153-184)
Feature Number
Captured date
Processing status
Scale
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
19 Nov
13 Feb
19 Mar
23 Apr
02 Jul
06 Aug
15 Oct
24 Dec
19 Nov
13 Feb
19 Mar
23 Apr
02 Jul
06 Aug
15 Oct
24 Dec
19 Nov
13 Feb
19 Mar
23 Apr
Raw
Raw
Raw
Raw
Raw
Raw
Raw
Raw
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Raw
Raw
Raw
Raw
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
dB
dB
dB
dB
182
173
174
175
176
177
178
179
180
181
182
183
184
02 Jul
06 Aug
15 Oct
24 Dec
19 Nov
13 Feb
19 Mar
23 Apr
02 Jul
06 Aug
15 Oct
24 Dec
Raw
Raw
Raw
Raw
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
dB
dB
dB
dB
dB
dB
dB
dB
dB
dB
dB
dB
The 327 samples were checked for possible outliers using the PRTools version 4.1.
The PRTools 4.1 developed by (Duin et al., 2007) is a statistical pattern recognition
toolbox run in Matlab environment. Prior to the analysis, the training data were reorganized into mandatory matrix format with the size m x k for m row representing
the samples, each described by k column feature values. Together with feature and
class labels, the training data were defined using dataset command. Then, the
remoutl command was called to detect and remove outliners through calculation of
class by class distance matrix satisfying the following two criteria:
1) Sample having a fraction of 0.05% of their distances larger than the mean
distance in the class; and
2) Sample having within-class distance larger than 3 times the standard
deviation.
While 18 samples met the two criteria, they were removed from the list of training
patterns and 309 samples were retained for feature selection.
3.6.2. Data Format and Split
Feature selection in FST3 requires the patterns to be firstly re-organized in a
mandatory structure and saved as .TRN file format as shown in Figure 3.27 below.
183
#datafile
#title
Spectral reflectance and radar backscattering of 7 groups of mangrove species
;
; The 7 classes are Aegiceras corniculatum (ac), Acanthus ilicifolius Group 1 (ai-g1), Acanthus ilicifolius Group 2
(ai-g2), Avicennia marina (am),
; Kandelia Obovata Group 1 (ko-g1), Kandelia Obovata Group 2 (ko-g2) and Sonneratia (sc)
#features 185
#classes 7 15,31,49,69,104,30,11
#data
185 features
309 observations
309x185 matrix
Figure 3.27. The .TRN file format required by FST3
The file starts with #datafile, follows by the title and description of the file. Then, the
number of features (#features), number of classes (#classes) and number of samples in
each class were defined in the header of TRN file. Starting from #data onwards is the
data matrix formed by 309 observations arranged by rows and 185 features in
columns.
The prepared dataset was then assessed by Data_Accessor_Mem_TRN. The Data_Accessor
object was used to pre-processes and split the data. The pre-processing was mainly
referred to normalization of data values using Data_Scaler object. Since no
normalization was needed, the class Data_Scale_Void was called. Data_Splitter object
allows multi-level sample separation into training, validation and testing subsets
and different subsets can be accessed during various steps of feature selection
process. Two nesting levels of split was used, namely outer and inner as shown in
Figure 3.28.
184
ALL DATA
Train
Test
Train
Train
Test
Train
Test
block
block
block
block
block
block
block
1
1
2
3
2
4
3
Outer Split
TEST
TRAIN
Training : Testing = 70:30
0 0 0 0 0 0 0 0 0 1
1 2 3 4 5 6 7 8 9 0
Inner Split
Train blocks = 9
TEST
TRAIN
10-fold cross validation
Figure 3.28. Data splitting scheme: outer and inner split for model training,
validation and testing.
The outer splitting level first divided the whole dataset randomly into training and
testing in proportion of 70 to 30 using the random sampling (class
Data_Splitter_Randrand).
The testing subset was used as an independent dataset to
test the final result of feature selection. During the learning phase, the training
subset was further split for estimation of classifier accuracy in wrapper-based and
hybrid feature selection process using 10-fold cross validation (class Data_Splitter_CV).
For 10-fold cross validation, the training subset was randomly partitioned into 10
blocks of subsamples. Every time, one single block was held for classification
accuracy validation while the remaining nine blocks were used as training data. The
cross validation was repeated ten times, with each of the ten blocks used once as
the validation subset. The 10 resultant classification accuracies were averaged to
provide accuracy estimation for the selected feature subset.
3.6.3. Wrapper-based Approach
The wrapper-based feature selection approach assesses the quality of feature
subsets using the prediction performance of a chosen classifier through statistical
185
resampling or cross-validation. Subset with comparatively higher classification
accuracy (or lower classification error) is regarded as the best solution. Two
classification algorithms, namely, the k-nearest neighbor (kNN) and support vector
machines (SVM) were used as wrapper criterion for feature selection.
kNN is the simplest machine learning algorithm under which an unknown sample is
assigned to the class that most commonly occurred amongst its k nearest
neighbors. In this study, k  3 was used. The Euclidean distance ( d ) was used as to
measure the distance to the neighbor. And the inverse of Euclidean distance ( 1/ d )
was used as weights so that distant neighbors contribute lesser to the average while
than the nearer ones.
SVM is a complex classifier which transforms the non-linear function into linear
ones using the kernel function. The kernel function transforms the functional space
into high-dimensional space where hyperplanes were found to separate the classes.
The radial basis function (RBF) kernel function of 2 was set. It is known that the
performance of SVM strongly depends on two parameters, penalty coefficient C
and width of RBF kernel  . In the course of search, the two parameters were
optimized through repeated consecutive process. As long as better SVM
performance/ accuracy on the training subset were achieved, the parameter
optimization continued. The parameter optimization stopped when the accuracy
difference was zero. The optimized values of C and  were used by SVM classifier
for feature selection. Detail description of SVM classification algorithm can be found
in Section 3.6.4.
In FST3, the two classifiers were applied using class classifier_knn and class
classifier_LIBSVM.
The classifier_LIBSVM provides an interface to link the external SVM
library of LIBSVM developed by (Chang and Lin, 2011). The implementation of class
Criterion_Wrapper adapts the chosen classifier to serve as feature selection criterion.
186
3.6.4. Search Algorithm
Owing to the exponential time complexity, optimal search methods such as the
exhaustive search and variations of branch and bound are only applicable to
problems of low dimensionality with approximately D  50 (Somol et al., 2010b).
Having D  185 in this study, optimal search is not a feasible option. Instead,
suboptimal search methods including sequential forward selection (class Search_SFS),
sequential forward floating selection (class Search_SFFS) and oscillating search (class
Search_OS)
were adopted. The Sequential_Step governs the basic mechanism of feature
addition and removal from a current subset. For oscillating search, the maximum
oscillation cycle depth,  which determines the number of features to be replaced
in one swing was set to five. In order to avoid local maxima, oscillating search was
run repeatedly from various random initial points. When there is no improvement
for the criterion value in five consecutive runs, the subset with the best criterion
value was the final subset. For all the search algorithms, they can be implemented
in either d-parametrized or d-optimizing mode. d-parametrized requires the analyst
to set the desired cardinality should the finalized subset retained while d-optimizing
search will optimize the subset size and its content at once provided that a nonmonotonous criterion such as classification accuracy is chosen (Somol et al., 2010b).
The d-parameterized mode with four targeted subset size including 5, 10, 15 and 20
were implemented.
3.6.5. Stability Evaluation
During the search, the object Result_Tracker was linked to the search algorithm to
provide alternative solutions rather than a single one from standard search. When
using together with the class Result_Tracker_Stablieval, various statistics and similarity
measures were collected for stability evaluation after a defined number of runs
(Somol et al., 2010b). In order to assess the stability of various search algorithms,
pairwise similarity measures including Average Tanimoto Index ATI (S ) , Overall
consistency measures including consistency C (S ) , weighted consistency CW (S ) ,
and relative weighted consistency CWrel (S , Y ) were computed and extracted after
187
n  20 trials. Different search algorithms were compared based on the similarity
measures.
3.6.6. Feature Frequency analysis
With combinations of three search algorithms (sequential forward selection,
sequential forward floating selection and oscillating search), two classifiers (knearest neighbor and support vector machines) and four subset sizes (5, 10, 15, 20),
after twenty trials, there are a total of (3x2x4x20) 480 feature subsets. In every
feature selection, training accuracy, independent testing accuracy, selected feature
subset and consistency statistics were logged in the text file. Frequency analysis is
used to examine and sort out features that are potentially significant for
classification.
Prior to the analysis, the global mean and standard deviation of training and testing
accuracy were computed according to different subset size. For each individual trial,
the training and testing accuracy are evaluated based the global mean and standard
deviation with the corresponding subset size. Trials producing subsets with accuracy
below the limit which is one standard deviation from the global mean accuracy are
removed. By maintaining classification accuracy to relatively high level, quality
feature subsets are retained for further evaluation. For the remaining subsets,
frequency analysis was conducted to count the frequency of occurrence of every
single feature in the remaining subsets. Features with the highest frequency count
based on six criteria below are regarded as the best features for classification.
I.
Approximate 20 features having the highest frequency count after removal
of feature subsets with accuracy below one standard deviation from mean in
subset size of 5;
II.
Approximate 20 features having the highest frequency count after removal
of feature subsets with accuracy below one standard deviation from mean in
subset size of 10;
188
III.
Approximate 20 features having the highest frequency count after removal
of feature subsets with accuracy below one standard deviation from mean in
subset size of 15;
IV.
Approximate 20 features having the highest frequency count after removal
of feature subsets with accuracy below one standard deviation from mean in
subset size of 20;
V.
Approximate 20 features having the highest frequency count in feature
subsets with the highest training and testing accuracy in different subset
sizes under different search algorithm-classifier combination; and
VI.
Approximate 20 features having the highest frequency count in feature
subsets with the highest training and testing accuracy in different subset
sizes regardless of search algorithm-classifier.
Features meeting all above criteria will have a maximum score of 6. The features
are ranked accordingly and those features have a score value of 4 or above are
retained. The extracted features are likely to correlate with each other especially
when adjacent features in the hyperspectral bands are considered, Pearson
correlation analysis was used to examine the correlation among the extracted
features in SPSS 15.0. Features having correlation of 0.8 or above with significant
level p<0.01 were discarded.
Based on the final feature set, various subsets were formed based on the
characteristics of the features, i.e. spectral or radar features as well as feature size.
3.7. Mangrove Species Classification
Supervised classifiers were trained to classify seven mangrove spectral classes using
different subsets generated from the final feature set. The observations were
divided using stratified random approach into training and testing partitions in a
ratio of 70 to 30. The random seed was set to ensure the same set of training and
testing data are used for every classification. Hence, out of 309 observations, 212
189
are used for training and the remaining 97 are used for testing. The number of
training and testing observations for each mangrove spectral class is shown in Table
3.10.
Table 3.10. Number of training and testing observations for the mangrove spectral
classes
Species
Training
Testing
A. corniculatum
12
3
A. ilicifolius G1
22
9
A. ilicifolius G2
32
17
A. marina
46
23
K. obovata G1
70
34
K. obovata G2
23
7
Sonneratia spp.
7
4
Procedures including classification, prediction and validation are conducted through
proper organization and manipulation of data streams. Figure 3.29 illustrates the
operation streams for different processes in classification. The basic element of a
stream is node. Specifically for the study, five types of node are used including the
source node, field operation node, modeling/ classifier node, model node and
output node and they are represent by different shapes as indicated in Figure 3.29.
The nodes are connected and formed operation streams of different purposes. The
whole process is realized using four streams.
190
1. Model Training Stream –
Training dataset is divided into
training and testing partitions and
trained by user defined specific
classifier
2. Model Prediction Stream –
Model computed from the training
stream is used to predict the
mangrove spectral classes for the
observations in the training dataset
3. Accuracy Analysis Stream –
Class prediction in the training dataset
is assessed for training and testing
accuracy and computation of
confusion matrix
4. Model Prediction Stream –
Model is applied to unseen mapping
dataset to predict the mangrove
spectral classes in the study area
Source node
Field operation node
Modeling node
Figure 3.29. The node, stream and workflow in Visual Classifier and Clementine 12.0
Model node
Output node
191
First, selected feature vectors after feature selection analysis are extracted from the
hyperspectral and multi-temporal SAR datasets covering the study area. They are
then organized in form of matrix with features in columns and observations in rows.
The matrix is imported as mapping data and stored in the source node in the fourth
stream. In the first stream, observations with known mangrove spectral classes are
sorted out and attached with class labels from 1 to 7. This data matrix is imported
as training data and stored in the source node. The source node is then connected
with two field operation nodes in are used to define the properties of the dataset.
The type node is used to define the properties of the features including data type
and direction. For data type, it tells whether the data are stored as strings, integers,
real numbers or other specific types. Direction specifies whether the fields are input
as predictors (features) or output as predicted (classes) for the modeling node. The
function of the partition node is to separate the data into training and testing
samples in proportion of 70 to 30 using random seed. By specifying the starting
value in the random number generator, the same observations are assigned to
training and testing samples in each execution. The last node in the first stream is
modeling node which is a specific classifier. Four classifiers including Gaussian
Maximum Likelihood (ML), Decision Tree (DT), Artificial Neural Network (ANN) and
Support Vector Machines (SVM) were chosen and the classification accuracy was
compared among the classifiers. ML and SVM were conducted in the visual
classification software developed in-house while DT and ANN were conducted in
Clementine 12.0. Specific parametric settings for different classifiers are discussed
in individual sections below.
The classifier is trained using the training dataset and produces a model. The
execution of the first stream is completed when a model node is generated for the
classifier. The model node stores the model information and acts as input to the
next three streams for classification and validation procedures. The second stream
is the model prediction stream for the training dataset. The training dataset is
connected to the model which is then is connected to the Partition node to
separate the dataset into training and testing samples. The output node produces
192
the class prediction results for each observation in the training dataset in the form
of table.
The third stream follows the same flow as the second stream to evaluate the
performance of the model prediction by generation of the confusion matrix and
estimation of the classification accuracy for the training and independent testing
samples in the output node.
The fourth stream applies the classification model to the mapping data to predict
mangrove spectral classes in the study area. Similar to the second stream, the
output nodes generate a table for each sample in the mapping data.
3.7.1. Species Separability
Prior to classification, the selected features were examined for their ability to
separate the seven spectral classes of mangrove species. Transformed divergence
(TD) which measures pairwise distance between spectral classes was computed in
ENVI 4.7. Features were entered one by one into TD calculation based on their
criterion scores. TD is the scaled divergence with maximum value equal to 2. A value
of 1.9 or above indicates good separability while TD below 1.7 suggests poor
separability between species. TD was plotted in graph to indicate the change of
value when features were added gradually.
3.7.2. Gaussian Maximum Likelihood Classifier
Gaussian Maximum likelihood (ML) classifier is the conventional algorithm that
assumes the training data of different classes follows the Gaussian or normal
distribution. Under this assumption, the distribution of a class can be totally
described by its mean vector and covariance matrix based on which the probability
density functions for each class are computed. Given an unknown pixel, the
probability of this pixel belongs to each of the classes are computer and the pixel
would be allocated to the class with the highest probability value (Lillesand et al.,
193
2008). ML is simple, fast and provides reliable results if the assumption of normal
distribution was met. However, the parametric algorithm was criticized being
unrealistic for many classification applications and will therefore lower the accuracy
and reliability of the results. In recent years, a number of non-parametric or
machine learning classifiers were gaining importance due to their free from
assumption of data distribution, capable to work well under small samples. The
conventional ML classifier was used as a benchmark to compare the performance of
several machine learning techniques.
3.7.3. Decision Tree Classifier
A hierarchical decision tree classifier is by definition a classification procedure that
recursively divides a dataset into smaller and increasingly homogeneous partitions
based on a set of splitting criteria defined at each branch (or node) in the tree
((Friedl and Brodley, 1997, Li and Chen, 2005)). Figure 3.30 shows the structure of a
binary-split decision tree made up by three types of nodes including root, internal
(splits), and terminal (leaf) nodes. The classification process is executed by a set of
decision rules which determines the path to be followed all the way from root to
leaf node. The classification result is therefore strongly rested on the selection of
splitting criteria and tree induction algorithm which governs the structure of the
tree (Tso and Mather, 2009). Information gain (Quinlan, 1993) and Gini impurity
index (Breiman et al., 1984) are two commonly used splitting criteria.
194
Figure 3.30. A simple binary-split decision tree classifier
(Based on Yang et al., 2003)
Instead of designing tree manually, many automatic tree induction methods have
been developed rapidly with promising performance in recent years. Classification
and Regression Tree [CART] ((Breiman et al., 1984)), Iterative Dichotomizer 3 [ID3]
(Quinlan, 1986), C4.5 (Quinlan, 1993) and C5.0 algorithms are commonly used
algorithms. Except CART which allows binary splits at non-terminal nodes, all other
algorithms allow generation of different number of branches at each node.
In Clementine 12, the latest C5.0 decision using entropy (information gain) as
splitting criteria was adopted in the study. At each split, an individual variable (band)
with the maximum normalized information gain is chosen to divide the data (Rogan
et al., 2003). The process continues until each observation was assigned to a specific
195
mangrove species class. Compared with its previous counterparts (C3.0 and C4.5),
C5.0 introduces an important concept called boosting to enhance classification
accuracy (Tso and Mather, 2009). The boosting technique initially assigns equal
weights to each observation in the training dataset. After the tree was built, the
predicted output class is compared with the known class for each individual training
sample. The incorrectly classified samples are indentified and assigned with a larger
weight. The increased weight of these training samples will induce a heavier penalty
on any misclassification further and force the classifier to focus on these samples in
the next iteration (Tso and Mather, 2009). 10 boosting iterations are set to improve
classification accuracy. The generality mode is used to enhance the generality of the
mode and reduce the susceptibility of over-fitting.
Decision tree classifier offers many advantages over the procedure of traditional
supervised classification. The most prominent one is the great flexibility on data
input. As a non-parametric classifier, decision tree classifier does not need to satisfy
any a priori assumption regarding the distributions of the input data (Friedl and
Brodley, 1997). A diverse data sources such as categorical information can therefore
be input into calculation and the classifier will automatically choose the best data
layers from those provided by the analyst (Lawrence et al., 2004). Besides, decision
tree is conceptually simple and computationally efficient (Pal and Mather, 2003)
while the results from dichotomous tree structure are easily interpreted (Lawrence
et al., 2004). Moreover, when compared with its counterpart, decision tree are
trained faster than neural network and there is no ‘black box’ hidden in the process
as the tree structure is clearly shown after execution (Pal and Mather, 2003). Finally,
the classification accuracy using decision tree classifier is usually higher than other
methods in many comparison studies (Pal and Mather, 2003). On account of these
advantages, decision tree classifier is gaining popularity in remotely-sensed data
classification problems. A number of studies compared decision trees and other
classifiers (Brown et al., 1993, Gahegan and West, 1998, Pal, 2005).
196
3.7.4. Artificial Neural Network Classifier
Artificial neural networks are biological metaphors developed by imitating the basic
elements and structure of a neuron of the brain (Anderson and McNeil, 1992). A
neural network is a composition of layers of nodes with each node represents a
single neuron. The organization of the individual nodes into layers forms the
structure of a neural network (Anderson and McNeil, 1992). A typical neural
network will be composed of at least three distinctive layers, namely input layer,
hidden layer and output layer and the number of nodes within each layer is
problem specific. According to Tso and Mather (2009), there are five fundamental
neural network architecture, namely the multilayer perceptron with backpropagation (Rumelhart et al., 1986), self-organized feature map (SOM) (Kohonen,
1982), counter-propagation networks, Hopfield networks (Hopfield, 1982) and
systems evolved from adaptive resonance theory (ART) (Carpenter and Grossberg,
1987b, Carpenter and Grossberg, 1987a, Carpenter et al., 1991a, Carpenter et al.,
1991b). Among those, the multilayer perceptron with back-propagation is the most
popular ANN models in remote sensing (Benediktsson et al., 1990, Kanellopoulos et
al., 1992, Paola and Schowengerdt, 1995, Atkinson and Tatnall, 1997). This study
will also apply this ANN to classify mangrove species. A typical three-layer
multilayer perceptron ANN is shown in Figure 3.31.
Figure 3.31. A typical three-layer multilayer perceptron artificial neural network
(Adopted from Sherrod, 2007)
197
External signals are received from various bands or features ( xi ) in the form of pixel
value by the neurons in the input layer. No computations are performed in the
input layer. The signal is then passed through the hidden layer and to the output
layer where the results of classification are produced. There can be more than one
hidden layer. No interconnections are found between the neurons in the same layer,
however, all neurons in a given layer are entirely connected to the neurons in the
adjacent layer with a connection weight (  i ). The weighed signals from the input
layer are summed (  ) before they are transferred to neurons the hidden layer. The
neurons then perform their mapping functions (σ) and pass the transformed signals
from the hidden layer to the neurons in the output layer after the signals are
attached with a connection weight and summed. Another mapping function (σ) is
executed and output to classes ( yi )(Anderson and McNeil, 1992). The idea can be
expressed mathematically as
n
yi  (  xi wi  )
Eq. 3.12
i 1
where  is the constant and the weights (  i ) determine the intensity of input
signals (Nelson and Illingworth, 1991, Kulkarni, 1994). Once this forward pass is
completed, the actual outputs from the neurons are compared with their expected
outcomes and the differences signify in network error (Tso and Mather, 2009). The
network learns by minimizing the error between actual and expected outputs
through weight adjustment by backward pass, i.e. starts from the output layer (Pal
and Mather, 2003). The back-propagation mechanism allows the network learns the
patterns inherent in the training data. And based on these patterns, it constructs
rules and applies to the unseen data (Tso and Mather, 2009).
The feedforward back-propagation neural network classification is conducted in
Clementine 12. The network topology consists of two hidden layers with the first
contains 20 nodes and second layer possesses 15 nodes. Apart from the topology,
two important parameters, namely momentum (  ) and learning rate ( ) which
control neuron weight updating mechanism across the layers in the network during
198
training. Both parameters have values ranged between 0 and 1. The momentum
helps to speed up convergence and escape from local minima in the backpropagation learning process. High  value of 0.9 is adopted as it stabilizes weight
change and realizes faster convergence by amplifying the learning rate if all weight
changes are in the same direction. The learning rate governs the magnitude of
weight adjustment at each update. As training progresses,  starts from high value
and gradually decreases to low value and the cycle repeats until training is complete.
Although high  allows faster learning, it is likely to produce poor results. The initial,
high and low  are set to 0.3, 0.1 and 0.01 respectively. The rate of decay is set to
30, which specifies the number of cycles change from high to low  . In order to
prevent overtraining in the modeling building process, data in the training partition
are further split into training and testing sets in proportion of 70 to 30. The training
process stops when it attains classification accuracy of 90%. During the training
process, the network performance is visualized in form of graph. If the accuracy
cannot be achieved, the training process will be terminated manually by judging the
accuracy fluctuation between the training and testing partitions. The network with
the best accuracy is saved.
The high computation rate, which therefore permits real-time processing of large
datasets, is one of the advantages of ANN. Besides, it is nonparametric where no
assumptions of statistical distribution of the data are necessary. Therefore, it can be
applied to classification problems with limited training samples, i.e. reliable
estimation of statistical parameters are hard to get
(Tso and Mather, 2009).
However, the process of ANN is always criticized a black box.
3.7.5. Support Vector Machines Classifier
In recent years, there were growing interests in using support vector machines
(SVM) for solving pattern recognition and image classification problems mainly for a
few reasons. First, it is highly applicable when limited samples are available for
multi-class problems. Unlike the traditional probability-based classifier which
required sufficiently large training samples to compute reliable probability density
199
functions for each class, SVM uses representative observations called support
vectors to determine the optimum hyperplane while the other non-support vectors
have small effects on classification (Zhao et al., 2005). Second, SVM is designed to
minimize the so-called structural risk. According to Tso and Mather (2009), SVM
takes this concept to minimize “the probability of misclassifying a previously unseen
data point drawn randomly from a fixed but unknown probability distribution
(p.125)”.
Third, SVM has been tested for comparatively better classification
accuracy over other classification algorithms such as maximum likelihood and
neural network (Huang et al., 2002, Melgani and Bruzzone, 2004, Pal and Mather,
2005).
The basic element of SVM is vector, which is defined as a set of brightness values in
a finite number of bands of a sample point. Figure 3.32a plots hypothetically the
training samples of two mangrove classes. SVM analysis tries to come up with a 1dimensional hyperplane, i.e. a line that can separate the two mangrove classes.
Obviously there are infinite choices of lines in this idealized example. However, SVM
finds an optimal hyperplane oriented in a direction so that it can maximally
separate the clusters of vector (Burges, 1998, Melgani and Bruzzone, 2004), which is
line 3 in this case. The dashed lines (parallel to the separating hyperplane) in Figure
3.32b mark the distance between the closest vectors and the separating hyperplane.
The distance between the two dashed lines is called the margin. Support vectors are
those vectors located near the hyperplane constraining the width of the margin
(Sherrod, 2007) highlighted in solid blue color.
200
a
1
2
3
Optimal separating hyperplane:
b
w xb  0
T
4
Class 1
w
b
w
Class 2
Origin

w
c
Y
Margi
n
Non-linear
d
Boundary
X
w
b
w
Y
Mapping Φ
z
Origin
Margi
Separating Hyperplane
X
n
Figure 3.32. The graphical description of support vector machine (SVM) for a
hypothetical two-class case
a) optimal hyperplane and support vectors; b) linearly separable scenario; c) linearly
non-separable scenario with introduction of slack variable; d) non-linear separable
scenario and high dimensional separable hyperplane (Based on Tso and Mather,
2009)
Figure 3.32b is an ideal example in which two classes can be separated by a linear
boundary. An optimal hyperplane is derived by the following boundary function:
wT x  b  0
Eq. 3.13
201
where x is a point on the hyperplane, w is perpendicular to the hyperplane,
T represent the transpose of matrix, and b is the bias parameter. However, the
classes derived from the remotely-sensed data are not always linearly separable as
described in Figure 3.32c. In this case, slack variables i , i  1, , n , are introduced
to relax the constraints (Cortes and Vapnik, 1995). Mathematically, the linear model
is to minimize the following function:
min  w, i  
N
1 2
w  C  i
2
i 1
Eq. 3.14
Subject to
yi w   ( xi )  b  1  i , i  1,..., N And
Eq. 3.15
Eq. 3.16
i  0, i  1,..., N
where  i is the slack variable initiated to justify non-separable data; and C is the
penalty coefficient to control the tolerance of classification error. SVM model tries
to strike a balance between maximizing the margin controlled by
1 2
w and
2
N
penalize misclassification error controlled by C   i . The higher the value of C ,
i 1
the larger the penalty is given to misclassified samples and thus curtails the
generalization capability (Tso and Mather, 2009).
When a linear hyperplane is unable to separate the classes properly, non-linear
SVM strategy is adopted. The strategy projects the training samples into a higher
dimensional space so as to spread the distribution of the training samples further
apart and facilitate the fitting of a linear hyperplane (Tso and Mather, 2009). The
idea is shown in Figure 3.32d. Specifically, kernel function Κ(x,xi ) is introduced to
transform the inner product  xi  x   R n in original low dimension space to
  ( xi )   ( x)   R n ' in the high dimensional feature space, where n '  n . The
optimum hyperplane is obtained through the discriminative function
202
N
f ( x)  sign( i yi K ( x, xi )  b), x  R n
i 1
Eq. 3.17
where xi is the samples used for training, yi is the corresponding class label,  i is
the Lagrangian variables and the calculated weighted vector of the bands/ features
N
is denoted as w   i yi ( xi ) . The training set give an estimation of the two
i 1
parameters  i and b . Given the non-linear nature of remote sensing data, SVM
uses the kernel method to map the original feature space into high dimensional
Hilbert space (Zhang and Ma, 2009). The kernel function is an alternative to the
nonlinear vector mapping function to simplify and save the demanding
computational process. Commonly used kernel functions include the polynomial,
radial basis function (RBF), Gaussian radial basis function and sigmoid (Tso and
Mather, 2009).
SVM was originally designed to tackle binary classification problems. Several
approaches have been proposed to extend SVM to multiclass classifications
including one-against-one (Knerr et al., 1990), one-against-many (Vapnik, 1995),
direct acyclic graph (DAG) (Platt et al., 2000) and multiclass SVM (Vapnik, 1998).
The RBF SVM classification is conducted using LIBSVM developed by Chang and Lin
(2011). The LIBSVM was incorporated into the visual classification software built inhouse. Two hyperparameters, namely the penalty coefficient C and width of RBF
kernel  are defined prior to classification. The two parameters governed the
tradeoff between approximation and generalization capability of the resultant
model. When  is changing from large to small value, approximation capability is
gradually in exchanged for generalization performance. That is, the support vectors
are expanding from training samples that are closest to other classes (good
generalization but poor approximation) to a situation when all samples become
support vectors (poor generalization but good approximation). As with smaller C ,
which means lenient penalty, the approximation capability is scarified for
generalization capability, vice versa.
203
The two hyperparameters were optimized using cross-validation method which
check the performance of different combinations of C and  . The minimum value,
maximum value and the step of change are specified prior to the optimal search of
parameters as shown in Table 3.11.
Table 3.11. The range and step of change for search of optimal C and 
Parameters
Minimum
Maximum
Step of change
C
-3
500
2

-2
10
0.05
The pair yielding the least cross validated error is defined as the optimal parameters
and used as the final model. As a multi-class problem, a one-to-one approach was
adopted to train the machine to classify one class against all others.
3.7.6. Accuracy Assessment
In order to compare the results more efficiently, visual-interpretation analysis and
accuracy statistics were used for accuracy assessment. A total of twelve sites were
picked from the study area for visual inspection as shown in Figure 3.33. These sites
are known mangrove stands with predominant and mixed species as recorded in
archived studies and from the experiences of field surveys. Sites G, K and L are
active growing region where pioneer species A. ilicifolius are predominant species.
Site C is also active growing area but mixed species of A.ilicifolius, A. corniculatum
and K. obovata are found here. The remaining eight sites are predominant stands of
specific species. The description for each individual site is adhered with the Figure
3.33.
204
Site
A
B
C
D
E
F
G
H
I
J
K
L
Description
Pure Sonneratia spp stand by plantation in Shenzhen
Mixed stand with A. ilicifolius as the predominant species
Active growing stand with species mix of A.ilicifolius, A. corniculatum
and K. obovata
Predominant K. obovata G1 stand distributed in north-south direction
Predominant A. marina stand distributed in north-south direction
Predominant K. obovata G2 stand distributed in north-south direction
Active growing stand with A. ilicifolius G2 as the predominant species
Predominant A. ilicifolius G1 grow along the river channels. The
distribution is in north-south direction with east-west strips cut across.
Predominant K. obovata G1 stand
Predominant K. obovata G1 stand
Predominant A. ilicifolius G2 stand distributed in NW-SE direction
Predominant A. ilicifolius G2 stand follow the V-shape fringe boundary
Figure 3.33. The twelve sites of visual inspection for accuracy assessment with
description of each site adhered.
Apart from visual-interpretation analysis, the classification accuracy was assessed
based on the training and independent validation dataset. The performance was
evaluated and quantified through the computation of confusion matrix based on
which the overall accuracy, the user’s and producer’s accuracy of each species were
205
computed. The combination of classifiers and feature subsets were evaluated
accordingly based on the two methods.
3.8. Leaf Area Index Modeling
The second part of the analysis involves LAI modeling using vegetation index
derived from hyperspectral data as well as textural variables extracted from the
ASAR image acquired on 19th November 2008.
3.8.1. Preliminary Exploration of Relationship between Hyperspectral bands
and LAI
Linear regression was conducted to reveal the relationship between the
hyperspectral bands and LAI using SPSS 15.0. LAI computed with clumping
compensation (log-average) using different methods including Bonhom’s, LAI2000
and LAI2000 generalized were explored. The coefficients of determination (r2)
which determines the proportion of variance in LAI that is predictable from each
hyperspectral bands was extracted and plotted. Through this preliminary
exploration exercise, hyperspectral bands that are significant to LAI prediction were
identified. By comparing the r2 derived from linear and log-average method, the
ability of clumping factor in enhancing the relationship with spectral reflectance
was revealed.
3.8.2. Vegetation Index Derived from Hyperspectral Data
A total of seven spectral vegetation indices (VIs) derived from narrowband
hyperspectral data were used to predict LAI in the study area. Table 3.12 shows the
vegetation indices, their corresponding formulation and related references used in
the study. In the equations,  represents the radiance of a specific band in
206
reflectance units. The vegetation indices involve three discrete wavelengths locating
in green peak (570nm), red absorption (702nm) and near infrared (764nm) and they
were selected based on the best features resultant from feature selection analysis.
Table 3.12. The spectral vegetation indices (VIs) for LAI modeling
Vegetation Index
Formulation
Reference
Normalized Difference
Vegetation Index
(NDVI)
764  702  764  702 
Rouse et al.
(1974)
Renormalized Difference
Vegetation Index
(RDVI)
764  702  764  702 
Rougean and
Breon (1995)
Soil-Adjusted Vegetation
Index (SAVI)
1  L764  702  764  702  L
Huete (1988)
Modified Soil-Adjusted
Vegetation Index
(MSAVI)
1
2  764  1 
2 
2764  12  8764  702 

Qi et al. (1994)
Triangular Vegetation
Index (TVI)
0.5120764  570   200702  570 
Broge and
Leblanc (2000)
Modified Chlorophyll
Absorption Ratio Index
(MCARI 1)
1.22.5764  702   1.3764  570 
Haboudane et
al. (2004)
Modified Chlorophyll
Absorption Ratio Index
(MCARI 2)
1.52.5 764   702   1.3 764   570 
2 764  1
2


 6  764  5  702  0.5
Haboudane et
al. (2004)
The Normalized Difference Vegetation Index (NDVI) is most commonly used
broadband indices in vegetation studies. It relates to leaf pigment content, canopy
properties and biomass. NDVI is effective in predicting canopy properties in
moderate vegetation density (Rouse et al., 1974). Renormalized Difference
Vegetation Index (RDVI) combines the advantages of NDVI and Difference
Vegetation Index (DVI) which are responsive for high and low vegetation density
respectively. It has been tested with high linearity with canopy parameters
207
(Rougean and Breon, 1995). These two indices are derived based on normalized
difference aiming to enhance the linear relationship with LAI.
The Soil-Adjusted Vegetation Index (SAVI) and Modified Soil-Adjusted Vegetation
Index (MSAVI) were developed to minimize the influence of soil background. SAVI
requires a prior knowledge of vegetation density which determines the effect of soil
on vegetation reflectance (Huete, 1988). The soil effect was accommodated by
setting factor L  0.5 in the formula, which was the optimal value suggested by
Huete (1988). MSAVI is a modification of SAVI which embedded the soil-adjustment
factor in the formulation (Qi et al., 1994a).
The remaining three VIs including Triangular Vegetation Index (TVI), Modified
Chlorophyll Absorption Ratio Index (MCARI 1 and 2) were developed based on
modifications from Haboudane et al. (2004). They are relatively less sensitive to
change of biochemical effect, mainly chlorophyll that influences LAI prediction. TVI
is defined by the area formed by green peak, red absorption and near infrared
shoulder. It characterizes energy absorption by leaf pigments which relates to LAI.
TVI will increase as chlorophyll absorption increases (decrease in red reflectance) as
well as increase in abundance of leaf tissues (increase in near infrared reflectance)
(Broge and Leblanc, 2000). The MCARI 1 and 2 was modified by Haboudane et al.
(2004) for LAI prediction based on MCARI developed by (Daughtry et al., 2000).
Both indices integrate near infrared wavelength to enhance its sensitivity to LAI
changes. MCARI 2 has added a soil adjustment term to further restrain soil
background effect (Haboudane et al., 2004).
3.8.3. Radar Backscatter and Derived Textural Parameters
Generally speaking, tone and texture of objects are commonly used parameters for
visual interpretation of aerial photos or satellite images in describing and assessing
object surfaces. Tone describes the grey level of pixels while texture measures the
tonal variations at the neighborhood of the pixels (Haralick et al., 1973, Mather,
1999a). Vegetations canopy structure difference due to variations in species,
208
growing status, maturity or healthiness tends to exhibit distinctive textural
variations. By considering spatial relationship of a pixel to its neighbor, information
on the local variability and texture extracted from radar images have been applied
to improve classification (Shanmugham et al., 1981, Kurosu et al., 2001) and
retrieve vegetation/ forest parameters (Kasischke et al., 1994, Chand and
Badarinath, 2007). It is suggested that classes with highly confusing probability
density functions in terms of tonal information of pixel can be distinguished using
textural analysis (Kurosu et al., 1999). With the ability to penetrate into the canopy
layer, the C-band ASAR data can capture canopy characteristics based on which
textural analysis can extract significant information for vegetation classification and
biophysical parameter estimation. Textural analysis can be conducted in two
approaches, namely structural and statistical (Haralick, 1979) with the latter
commonly applied to remote sensing data. Based on spatial distribution of grey
levels of an image, statistical approach employs a set of local statistical measures to
generate textures and the second-order statistics based on Haralick’s Gray Level Cooccurrence Matrices (GLCM) is the common technique (Haralick et al., 1973). GLCM
is a two dimensional matrix of joint probabilities measuring the likelihood of
occurrence of two grey levels separated by a given distance in a given direction
(Mather, 1999a). Specific textural characteristics were extracted from GLCM
effectively for a number of SAR applications including crop differentiation (Soares et
al., 1997), land cover mapping (van der Sanden and Hoekman, 1999). Table 3.13
summarizes the equations of textures computed based on Haralick’s (1973) (GLCM).
The GLCM textural features were computed based on Hall-Beyer (2000) and
grouped into three categories measuring contrast, orderliness and descriptive
statistics of the image. In the equations, the element, Pi, j ,where, i is the row
number and j is the column number represents the relative frequency of two
neighboring pixels. A short description is provided in below. A more completed
theoretical description can be found in Haralick et al. (1973) and Hall-Beyer (2000).
209
Table 3.13. GLCM-derived textural variables and their corresponding equations
Textural
Group
Texture
Abbrev.
Homogeneity
HOMO
Equation
P[i, j ]
 1  i  j 
i
Contrast
Measure
Orderliness
Measure
Descriptive
Measure
2
j
Contrast
CT
 P(i  j)
Dissimilarity
DSM
 P i  j
Entropy
ENT
i
2
P[i, j ]
j
i
j
  P[i, j ] ln P[i, j ]
i
j
 Pi, j 
2
Angular Second
Moment
ASM
Mean
MN
Standard Deviation
SD
Correlation
CORR
i
j
i   iP[i, j ]  j   jP[i, j ]
 i2   i 2 P[i, j ]  i2
 2j   j 2 P[i, j ]   2j
 ijP[i, j]   
i
i
j
 i j
j
210
Homogeneity or inverse difference measures local similarity. Large values indicate
small grey tone differences in the pairs. Homogeneous elements locate around the
main diagonal in GLCM. Contrast is inversely correlated with homogeneity. Entropy
measures the degree of disorder. Large entropy value indicates that the image is
heterogeneous or non-uniform. When reaching a maximum, the image is
completely random pixels. The angular second moment or energy measures
textural uniformity. It reaches its maximum value when the pixels in the sampling
window have similar grey levels. It is inversely correlated with entropy. The mean
and standard deviation texture measures describe the central tendency and spread
of the grey levels in a local window similar to the descriptive statistics. Correlation
measures the linear dependencies of grey level. High correlation values represent a
linear relationship between grey levels of pixel pairs. A homogeneous area will have
correlation equal to unity (Haralick et al., 1973, Hall-Beyer, 2007).
Textural variables were generated from the filtered master image for LAI mapping
purpose. Parameters including window size, spatial distance between pixel pairs,
texture direction and number of GLCM grey levels were defined for texture
computation. A window size of 5 x 5 pixels was chosen and spatial relationship
computed at four directions (0, 45, 90, 135 degrees) were averaged, i.e. directional
invariant. Correspondingly, the parameters can be expressed using displacement
vectors d ,  where d is the spatial distance and  is the angle. Displacement
vectors of (1, 0°), (1, 45°), (1, 90°) and (1, 135°) were used to obtain the grey level
co-occurrence matrix (GLCM) based on which the textural features were extracted.
128 grey levels were used for constructing the GLCM matrix and the extraction of
texture measures was conducted in PCI Geomatica.
3.8.4. Regression Analysis
In order to understand the sensitivity of vegetation indices, radar backscatter and
textural bands to LAI variation, two types of regression analysis, simple linear and
stepwise multiple were used to identify the relationship between different remote
sensing parameters extracted from the images at the sampling locations and field211
measured LAI from corresponding sampling locations. Simple linear regression and
multiple regression model are expressed in Eq. 3.18 and Eq. 3.19 respectively.
Y  a  1 x1
Y  a  1 x1   2 x2     p x p
Eq. 3.18
Eq. 3.19
where Y is the dependent variable of field-measured LAI, a is the constant and
X p are independent variables of vegetation index, radar backscatter and radar-
extracted textural measures and  p are unstandardized beta coefficients of the
corresponding independent variable. A simple linear regression was performed with
individual VI acts as single independent predictor variable and regressed with the
dependent variable, LAI.
In multiple regression, each individual VI was combined the full set of radar
parameters. The stepwise method is designed to determine the most prudent set of
predictors that are most effective in predicting the dependent variable. Stepwise
selection was used to find the set of bands that are most effective in predicting LAI.
The aim was to select radar parameters that can potentially merged with VI which
can improve the model and therefore LAI estimation. The regression analysis was
conducted in SPSS v.15.0 and the workflow is shown in Figure 3.34. LAI derived from
each method, including Bohomme and Chartier’s, LAI2000, and LAI2000 generalized
were regressed against the independent variables separately. Hence, a total of
three sets of model will be generated.
212
Figure 3.34. The workflow of regression analysis
213
Regression analysis has two basic underlying assumptions with which the
dependent and independent variables have to follow in order to have a reliable
report on the strength of the relationship. The first assumption states that the
variables (dependent and independent) are normally distributed while the second
requires the relationships between each pair of dependent-independent variables
to be linea. In order to ensure the strength of resultant relationship, the normality
and linearity of dependent and independent variables were explored prior to
regression analysis.
With sample size larger than 50, the Kolmogorov-Smirnov statistical test was used
to test the normality of the dependent and independent variables. The null
hypothesis of Kolmogorov-Smirnov test is
H 0 : Actual distribution of variable  Expected distribution of variable
A significant level less than or equal to the level of significance (p<0.01) will reject
the null hypothesis stating that the variable is not normally distributed. Besides,
normality plot were used to examine the skewness of the distribution.
Correlation analysis using scatterplot and statistical diagnosis was used for linearity
testing between each pair of dependent and independent variable. A significant
linearity was indicated when the significant level of correlation coefficient was less
than 0.05, i.e. p<0.05.
If the dependent variables fail to meet the assumptions, logarithmic and square
root transformation functions were applied in successive order to transform the
variable. If the transformed variable satisfies the normality assumption, it will
replace the original variable. However, if no transformation satisfies normality, the
untransformed variable is retained with caution for violation of assumption added.
Similarly, if the independent variables fail the assumptions, logarithmic and square
root functions were applied in successive order to transform the variables. If the
transformed variable satisfies the normality assumption, it will replace the original
variable. If no transformation satisfies both criteria, the untransformed variables
are retained with caution for violation of assumption added.
214
After normality and linearity testing, possible outliners were identified through
residual and distance computation in SPSS 15. Three types of outliners including
univariate outliners (UO) on dependent variables, multivariate outliners (MO) on
independent variables, and influential outliner (IO) tended to distort the regression
results. They were identified using the studentized residual, Mahalanobis distance
D2 score and Cook’s distance respectively. UO on dependent variable has values of
studentized residual greater than or equal to 3.0. Cumulative probability density of
Mahalanobis distance D2 score was computed and was subtracted by one in order
to indicate the probability in the upper tail of the distribution. A case was identified
as MO has the cumulative probability less than 0.001. An IO case has a large impact
on the regression solution relative to other cases and it was measured by Cook’s
distance. The critical value of Cook’s distance was computed as:
4
(n  k  1)
Eq. 3.20
where n is the total number of LAI sampling sites/ cases and k is the number of
independent variables. In this case, n = 95 and k = 7, which suggested that the
critical value should be equaled 0.0460. If a case has a value of Cook’s distance
larger than the critical, it is regarded as an IO. The 95 observed samples were
checked for the three types of outliners. To summarize, a case that satisfies the
following logic would be retained for analysis, otherwise, they will be examined for
exclusion.
“ABS(Studentized residual) < 3.0 AND
(1 - Cumulative probability density of
Mahalanobis distance D2 score) > 0.001 AND (Cook’s distance) <= 0.0460”
If the resultant regression model can be improved with better coefficients of
determination (r2), the outliners will be discarded.
215
After outliner removal, the regression analyses were executed in SPSS 15. For
simple linear regression, no specific model setting was required. As for stepwise
regression, it requires the setting of the significance of probability of F value that
determines the entry or removal of a variable in the equation one at a time. The
minimum significance level of F value for variable entry was set to 0.05 while that
for variable removal was equal to 0.10. An independent variable with the
significance of F value less than the entry value is entered while independent
variable with the significance of F value sufficiently greater than the significance of F
value is removed from the equation. The analysis stops when there is no statistically
significant improvement in coefficient of determination (r2) can be made through
addition of predictor variable.
Statistics governing the performance including r2, adjusted r2 of the models were
extracted to evaluate the models. Hypothesis testing based on level of significance
was used to examine the validity of regression models.
Regarding the overall performance, the significance level of the probability of the Fvalue, coefficient of determination (r2), and adjusted R were examined. The
probability of the F value in the ‘ANOVA’ table was used to describe the overall
regression relationship between the set of independent variables and dependent
variable. A level of significance less than or equal to 0.05 (at 95% confidence level)
was used to reject the null hypothesis that there is no significant relationship
between the set of independent variables and the dependent variable. The
coefficient of determination (r2) describes the proportion of variance in the
dependent variable explained by the set of selected independent variables. The
adjusted r2 measure the strength of relationship by considering the number of
independent variables as well as the number of cases. A robust model should have
little discrepancy between r2 and adjusted r2.
The performance of individual independent variables to dependent variable as well
as multicollinearity was indicated in the ‘Coefficients’ table. The  coefficients
indicate the direction of relationship between each of the independent variable and
the dependent variable. The probability of the t-statistic for  coefficients was
216
used to indicate if there is a significant statistical relationship between dependent
variable and each of individual independent variable. A level of significance less
than or equal to 0.05 (at 95% confidence level) was used to reject the null
hypothesis that there is no significant relationship between an independent variable
and the dependent variable. The spatial pattern of LAI over the study area was
generated based on the constant and unstandardized  coefficients from the
regression models.
Specifically for stepwise regression model, the multicollinearity of selected
independent predictor variables was checked based on the tolerance value.
Multicollinearity occurs when two independent variables are highly correlated, r =
0.90 or higher in a regression analysis. Such strong relationship between the
independent variables tends to distort the final result significantly. In order to avoid
the problem, multicollinearity was detected by checking the value of tolerance
which describes the amount of variability in one independent variable that is not
explained by the other predictors. Collinearity presents when the tolerance values
are less than 0.10. Whenever collinearity happens, the variable was omitted from
the analysis and model with the second highest r2 was selected.
3.8.5. Error Estimation
Instead of splitting the sample into training and validation sets due to small sample
size, 3-fold cross validation was used to estimate the prediction errors of the linear
and stepwise regression models. Retained samples after outliner removal were
randomly partitioned into three sets of more or less equal size using random seed.
A single set was retained for model testing and the remaining two sets were used to
formulate the models. The cross-validation repeated 3 times until each set was used
exactly once for validation. The unstandardized predicted values and residuals were
saved for every run in SPSS. The root-mean-square-error (RMSE) was computed
from the unstandardized residuals using Eq. 3.21 below:
217
 x
n
RMSE 
i 1
i
n
 xi'

2
Eq. 3.21
where n is the number of samples, xi is the measured LAI, xi' is the predicted LAI,
the residuals equals ( xi  xi' ) . The residuals of each case were computed and saved
in SPSS during model computation. Better regression model has lower RMSE.
3.9. Summary
To summarize, this study combined satellite-borne hyperspectral and multitemporal radar data with field data acquired through in situ field survey to conduct
mangrove species-based classification and LAI mapping in Mai Po. Provided with
such a large data volume, irrelevant and redundant features were removed through
feature selection process. Retained features were used as input for species and LAI
mapping.
The hyperspectral data was captured by the Hyperion sensor onboard EO-1 satellite
through programming on November 21, 2008. To convert to geographicallyreferenced reflectance spectra, the hyperspectral data has undergone a number of
rectification processes including bad line/ pixel removal, destriping, atmospheric
correction, noise removal through MNF and geometric correction in ENVI 4.5 and
PCI Geomatica 10.1. The multi-temporal radar data were captured by ASAR sensor
onboard ENVISAT satellite. They were extracted from the archived database
provided by the Institute of Space and Earth Information Science, CUHK. The multitemporal data were converted to backscatter coefficient using BEST 4.2.2b. They
were then geo-rectified in PCI Geomatica 10.1.
During field survey, two types of data including leaf samples and hemispherical
photos were collected. GPS was used to record the geographical locations for each
type of data. The leaf samples were transported back and processed in the
laboratory to extract the spectra of different mangrove species using Fieldspec 3.
LAI was computed from the hemispherical photos through indirect gap fraction
218
analysis using WinSCANOPY system. LAI was computed using three methods,
namely Bonhomme and Chartier’s (LAIBon), LAI2000 plant canopy analyzer (LAI2000)
and LAI2000 generalized (LAI2000G) methods. Apart from field measured data,
archived species data were extracted from the government database.
Wrapper-based feature selection was conducted to select from the hyperspectral
and multi-temporal datasets the best set of features for species discrimination and
LAI mapping with the aid of FST 3.0. By combining various search algorithms
including sequential forward selection (SFS), sequential forward floating selection
(SFFS) and oscillating search (OS) and classifiers (k-nearest-neighbor and support
vector machines) based on which classification accuracy was used as evaluation
criterion, hyperspectral and radar features were ranked by frequency analysis.
Features with the highest criterion values were retained as input for further
analyses.
After feature selection, the best features were combined into various subsets based
on their characteristics such as spectral alone, radar alone or combination of both
and act as input to species-based classification. The Classification was conducted
using classifiers including maximum likelihood (ML), decision tree C5.0 (DT), artificial
neural network (ANN) and support vector machines (SVM) in Clementine 12.0 as
well as the visual classification software developed in-house. The accuracy of
classifier and feature combinations was evaluated through visual-interpretation
analysis of twelve representative sites as well as through comparison of accuracy
statistics such as overall, producer’s and user’s accuracy.
The second part of the study involves LAI estimation through simple linear and
stepwise multiple regression analyses in SPSS 15. Based on the feature rank resulted
from the feature selection analysis, three discrete bands – red, green, near infrared
were used to compute vegetation indices including normalized difference
vegetation index (NDVI), renormalized difference vegetation index (RDVI), soiladjusted vegetation index (SAVI), modified soil-adjusted vegetation index (MSAVI),
triangular vegetation index (TVI), modified chlorophyll absorption ratio index 1 and
index 2 (MCARI 1 & 2). They act as independent variables to regress with field
219
measured LAI (dependent variable) through simple regression analysis. Stepwise
multiple regression was conducted to select radar parameters that can potentially
improve the regression model and therefore LAI estimation. Regression models
having the highest coefficients of determination (r2) and lowest root-mean-squareerror (RMSE) were regarded as the best model.
220
CHAPTER 4
RESULTS AND DISCUSSION (I) – FEATURE
SELECTION AND MANGROVE SPECIES
CLASSIFICATION
4.1. Introduction
This chapter firstly discussed the results from data processing and exploration
including comparison of reflectance extracted from two atmospheric correction
algorithms as well as comparison of radar speckle filtering techniques. The models/
techniques yielding the best reflectance and backscatter cross-section based on the
evaluation criteria are selected for further processing. Statistical examination of
species discriminability under the hyperspectral spectrum and multi-temporal radar
data are followed. The presentation of feature selection results of various
combinations of search methods (SFS, SFFS and OS) and wrapper-based evaluation
criteria/ classifiers (KNN and SVM) forms the first main section of this chapter.
Feature selected by different algorithms are compared and the final best feature set
is identified and used as input to the subsequent classification process. The second
key section of this chapter focus on species-based classification results using
different classifiers, namely maximum likelihood, decision tree C5.0, artificial neural
network, and support vector machines under different combinations of features
from the final feature set. The classification results are visually interpreted and
statistically compared in terms of feature subsets and classifiers. The discussion and
implication session followed provides a more in-depth discussion drawn from the
two main sessions. This chapter ends by summarizing the main findings in the
feature selection and species classification.
4.2. Data Processing and Exploration
The results presented in this section involve comparison of two atmospheric
algorithms in retrieving surface reflectance; comparison of various speckle
221
reduction techniques in terms of filter and size; and exploration of mangrove
species separability in the spectral and radar backscatter data.
4.2.1. Atmospheric correction algorithms comparison
The reflectance spectrum of mangrove species computed from the two atmospheric
correction algorithms, ATCOR-2 and FLAASH were compared with the laboratorymeasured spectrum using correlation analysis and spectral feature fitting analysis.
The Pearson correlation coefficients of different species across different spectral
regions are shown in Figure 4.1. The correlation coefficients of ATCOR-2 and
FLAASH are represented by blue and red color respectively.
A. corniculatum
A. ilicifolius
b
1.0
1.0
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
Correlation
Correlation
a
0.5
0.4
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.0
0.0
Full
Blue Green Red
Red
Edge
NIR SWIR SWIR SWIR SWIR
1
2
3
4
Full
Spectral Region
A. marina
1.0
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.4
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.0
0.0
Full
Blue Green Red
Red
Edge
NIR SWIR SWIR SWIR SWIR
1
2
3
4
Full
Spectral Region
K. obovata Group 2
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
Correlation
Correlation
Red
Edge
NIR SWIR SWIR SWIR SWIR
1
2
3
4
Sonneratia spp.
f
1.0
0.4
Blue Green Red
Spectral Region
1.0
0.5
NIR SWIR SWIR SWIR SWIR
1
2
3
4
K. obovata Group 1
d
1.0
e
Red
Edge
Spectral Region
Correlation
Correlation
c
Blue Green Red
0.5
0.4
222
0.3
0.3
0.2
0.2
Sonneratia spp.
f
0.1
0.1
1.0
0.0
Full
0.0
Blue Green Red
0.9
0.8
NIR SWIR SWIR SWIR SWIR
1
2
3
4
Full
Spectral Region
0.9
0.8
0.7
K. obovata Group 2
0.5
0.2
0.4
0.3
0.1
0.2
0.0
0.1
0.9
0.8
0.4
0.3
0.7
Full
0.6
0.5
0.4
0.3
0.2
Full
0.0
NIR SWIR SWIR SWIR SWIR
1
2
3
4
1.0
0.5
0.6
Red
Edge
Sonneratia spp.
f
0.6
Correlation
Correlation
1.0
Blue Green Red
Spectral Region
0.7
e
Correlation
Red
Edge
Blue Green Red
Red
Edge
Blue Green Red
NIR SWIR SWIR SWIR
1
2
3
Spectral Region
Red
Edge
SWIR
0.1
NIR SWIR SWIR SWIR SWIR
0.0
1
2
3
4
Full Blue Green Red Red NIR SWIR
4
Spectral Region
ATCOR-2
Edge
1
SWIR SWIR SWIR
2
3
4
Spectral Region
FLAASH
Figure 4.1. The correlations of between atmospherically corrected images (by
ATCOR-2 and FLAASH) and laboratory-measured spectrum for full range and regions
of the spectrum for various species including (a) Aegiceras corniculatum; (b)
Acanthus ilicifolius; (c) Avicennia marina; (d) Kandelia obovata Group 1; (e) Kandelia
obovata Group 2; and (f) Sonneratia spp.
When the full spectrum was considered, the correlation coefficients between the
reflectance of the two algorithms are similar across the species though FLAASH
shows a slightly higher correlation. When the spectrum was broken down into
different regions, distinctive correlation patterns were observed. Generally,
reflectance from FLAASH has a higher correlation with the laboratory spectrum in
almost all spectral regions except in near infrared and SWIR 1. For A. ilicifolius,
spectra derived from ATCOR-2 have a relatively higher correlation in green, red and
SWIR 4 spectral regions. Similarly, spectra from ATCOR-2 have a slightly higher
correlation with laboratory spectrum in SWIR 4 for A. marina; and in red spectral
regions for K. obovata group 1.
The magnitude of correlation coefficient difference varies across the spectrum.
Highest correlation differences were found in SWIR 2 for all species with FLAASH
has a much higher correlation with the laboratory spectrum. The range of
difference is between 0.7 – 0.8. The correlation differences between the two
algorithms ranges from 0.1 – 0.4 and 0.1 – 0.3 for NIR and SWIR 1 under which
ATCOR-2 shows a higher correlation. Spectral regions including green, red edge,
SWIR 3, SWIR 4 all show little correlation difference.
223
Apart from correlation, fit value from spectral feature fitting analysis was also used
to compare the performance of two atmospheric algorithms. Figure 4.2 shows the
fit value of image pixels according to species.
Fit
a
A. corniculatum
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
0
2
4
6
8
10
12
14
16
No. of Samples
Fit
b
A. ilicifolius
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
No. of Samples
f
Fit
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
A. marinaspp.
Sonneratia
5.50
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
0
0
5 2 10
15
4
20
25
30 6 35
40
8
45
50
55 10 60
65
12
70
75
No.
No.ofofSamples
Samples
ATCOR-2
d
Fit
Fit
c
FLAASH
K. obovata Group 1
5.50
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
224
14
2.00
1.50
1.00
0.50
0.00
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
No. of Samples
Fit
d
K. obovata Group 1
5.50
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
0
10
20
30
40
50
60
70
80
90
100
110
120
No. of Samples
Fit
e
K. obovata Group 2
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
0
5
10
15
20
25
30
35
No. of Samples
f
Sonneratia spp.
Sonneratia
spp.
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
Fit
Fit
f
5.00
4.50
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
0
0
2
2
4
4
6
6
8
8
10
10
12
12
14
No.
No.ofofSamples
Samples
ATCOR-2
FLAASH
Figure 4.2. The scatterplots showing fit value from spectral feature fitting analysis of
image pixels for (a) Aegiceras corniculatum; (b) Acanthus ilicifolius; (c) Avicennia
marina; (d) Kandelia obovata Group 1; (e) Kandelia obovata Group 2; and (f)
Sonneratia spp.
225
14
From the scatterplots, spectra extracted from ATCOR-2 shows two extremes, large
fit values due to low RMS error and zero and low fit values due to unrecognized
spectra. The average RMS errors ranged between 0.21 (A.marina & Sonneratia spp.)
and 0.31 (A. corniculatum) while the mean fit values ranged between 2.18
(Sonneratia spp.) and 3.86 (A. corniculatum) among the species. A total of 2, 28, 26,
35, 13 and 6 image pixels have been unrecognized in the species groups A.
corniculatum, A. ilicifolius, A.marina, K. obovata Group 1, K. obovata Group 2,
Sonneratia spp. respectively. For spectra computed from FLAASH, the fit values are
comparatively lower due to large RMS errors. The average RMS errors have a range
between 1.36 (A. corniculatum) and 11.92 (Sonneratia spp.) while the mean fit
values ranged between 0.43 (Sonneratia spp.) and 2.49 (A. corniculatum) among the
species. However, almost all the image pixels can be mapped using the laboratory
spectrum without undefined value. Although negative scale value appear in species
A. ilicifolius, A.marina, K. obovata Group 1 and Sonneratia spp., the proportion is
very small, i.e. one sample pixel in the each species group.
The spectral feature fitting measures the fitness of the continuum-removed image
spectra based on the absorption features along the spectrum. Vegetation features
exhibits a number of absorption features along the spectrum. However, it is hard to
conclude which algorithm has the best match with the laboratory spectrum from
the results. Although ATCOR-2 shows higher fit values, the large number of null
value in fitting process probably indicates unstable and inconsistent computation
across the scene. As for FLAASH, the mean fit values are comparatively lower for all
species, however, no null fit value was found across the scene. Besides, for some
pixels having null value in ATCOR-2, FLAASH provides a relatively acceptable fit
value indicated by large vertical distances of the same sample in Figure 4.2. It
seemed that pure comparison of absorption features cannot differentiate the
performance of the algorithms. It is concluded that none of the atmospheric
correction result is superior to another.
226
4.2.2. Radar Data Speckle Reduction
Various speckle reduction techniques and window sizes were evaluated based on
two criteria. First, their ability to maintain mean backscatter value and
simultaneously lower the standard deviation of feature with homogeneous
backscatter value when compared with the non-filtered or raw data was measured
and compared. Second, each combination of technique and window size was
measured for its ability to separate different species was quantified by transformed
divergence.
By comparing with the raw backscatter, the percentage of change in terms of mean
backscatter value extracted from four tonally homogeneous features, i.e. dark grey,
medium grey, light grey and bright of various filters and window sizes were plotted
in Figure 4.3. The lesser the change of mean backscatter value in different grey
tones, the better is the performance of the filter. Generally, the speckle filters
regardless of window sizes can maintain the mean backscatter value as compared
with the raw data except the median filter. The homogeneous regions of medium
grey and light grey exhibited a more distinctive pattern. From the figures, distinctive
drop in mean value after median filter of various sizes were applied to the image.
The maximum change of mean backscatter (about 45%) was found in bright objects
using median filter with window size of 7x7 and the drop in mean value is
progressive with increasing window size. Other filtering techniques were managed
to maintain more or less the same mean as the raw backscatter value. Less than
0.1% change in mean backscatter was observed in medium and light grey objects
regardless of window size while there were fluctuations of mean value among dark
grey and bright objects with the maximum change found in average filter with
window size of 7x7. The variation of window size suggested a negligible deviation of
mean value from the original non-filtered data. A sound filter should maintain the
mean value of homogeneous objects as much as possible. Under this principle, the
median filter was excluded as it had the worst performance.
With reference to standard deviation, all speckle reduction techniques regardless of
window sizes showed a sharp drop in standard deviation of backscatter value as
227
shown in Figure 4.4. Again, the trend was more obvious in medium grey and light
grey homogeneous features. By examining the medium and light grey features, the
decrease in standard deviation regardless of filtering technique was about 55%,
65% and 70% for window size of 3x3, 5x5 and 7x7 respectively. With increase in
window size, there was a clear drop in standard deviation for all filtering techniques.
Although the median filter showed a slight advantage over others techniques in
terms of the drop in standard deviation especially for the dark grey and bright
objects, the significant drop in mean backscatter value renders the filter not to be
considered.
Apart from comparing the mean and standard deviation, different combinations of
filtering technique and window size were measured and compared for their ability
to separate mangrove species. The average transformed divergence (TD) for each
species using different combinations of filter and window size was plotted in Figure
4.5. In terms of window size, size of 5x5 (represented in green color) showed a
better TD for most of the species (except Ko_G2). Regarding the filtering technique,
both the average and enhanced adaptive Lee filters offered a better discriminability
among species but with the latter one performs comparatively better. By comparing
the average and Lee filters with 5x5 window size, distinctively higher TD was found
in Am. It may be due to the fact that the adaptive Lee filter can retain high
frequency features which are exhibited by the crown structure of species while it is
likely for low-pass average filter to remove the structural information by smoothing.
228
50%
Percentage change in mean backscattering coefficient
45%
40%
35%
30%
25%
20%
15%
10%
5%
0%
Speckle reduction filter (window size)
DG-Diff in Mean
MG-Diff in Mean
LG-Diff in Mean
B-Diff in Mean
Figure 4.3. The percentage of change of mean backscatter value of various
homogeneous objects with different combinations of filter and window size
Speckle reduction filter (window size)
0%
Percentage change in standard deviation
-10%
-20%
-30%
-40%
-50%
-60%
-70%
-80%
DG-SD % Diff
MG-SD % Diff
LG-SD % Diff
B-SD % Diff
Figure 4.4. The decrease of standard deviation of various homogeneous objects
with different combinations of filter and window size
229
1.2
Average transformed divergence (TD)
1
0.8
0.6
0.4
0.2
0
Ai_G1
Ai_G2
Ac
Am
Ko_G1
Ko_G2
SSpp
Mangrove Spectral Class
3x3 Average
3x3 Median
3x3 Gamma
3x3 Lee
3x3 Frost
5x5 Average
5x5 Median
5x5 Gamma
5x5 Lee
5x5 Frost
7x7 Average
7x7 Median
7x7 Gamma
7x7 Lee
7x7 Frost
Figure 4.5. The average transformed divergence of species using different
combinations of filter and window size
4.2.3. Statistical Discrimination of Mangrove Spectral Classes
The mean reflectance and backscatter coefficient of the mangrove species classes
are shown in Figures 4.6a – d. Figure 4.6a shows the mean spectral reflectance of
2151 bands at 1nm spectral interval extracted from laboratory-measured leave
spectra covering 350 – 2500nm. Figure 4.6b illustrates the mean spectral
reflectance of 157 bands at 10nm spectral interval extracted from the Hyperion
dataset covering 480 – 2400nm. Figures 4.6c and d shows the mean raw and filtered
backscatter coefficient of the radar temporal series respectively. It should be noted
that the species groups from the laboratory measurement are slightly different
from that of satellite datasets. On one hand, the leaf samples did not cover A.
ilicifolius G2 (AiG2) while the two Sonneratia species were distinguished from each
other.
230
Backscattering coefficient (in dB scale)
428.4
448.8
469.1
489.5
509.8
530.1
550.4
570.7
591.0
611.3
631.7
652.0
672.3
692.7
713.1
733.5
753.9
774.2
794.6
815.0
835.4
855.8
876.1
896.5
916.8
963.7
983.8
1004.1
1024.2
1044.3
1064.6
1084.7
1104.9
1165.4
1185.6
1205.8
1225.9
1246.1
1266.3
1286.5
1306.7
1326.8
1347.0
1498.4
1518.6
1538.7
1558.9
1579.1
1599.3
1619.5
1639.6
1659.8
1680.0
1700.2
1720.4
1740.5
1760.7
1780.9
1982.6
2002.8
2023.0
2043.2
2063.3
2083.5
2103.7
2123.9
2144.1
2164.1
2184.3
2204.5
2224.7
2244.9
2265.0
2285.2
2305.4
2325.6
2345.8
2365.9
2386.1
Reflectance
Reflectance
0.80
0.70
(a)
0.60
0.50
0.40
0.30
0.20
0.10
0.00
350
450
550
19 Nov
650
750
850
Ac-mean
13 Feb
Ac-mean
950
1050
Ac Mean
AiG1-mean
AiG1-mean
1150
1250
Wavelength (nm)
Ai Mean
19 Mar
AiG2-mean
1350
AiG2-mean
1450
Am Mean
Am-mean
23 Apr
Am-mean
1550
Ko Mean
1650
KoG1-mean
KoG1-mean
1750
Sa Mean
1850
KoG2-mean
02 Jul
KoG2-mean
1950
06 Aug
2050
2150
15 Oct
2250
2350
2450
Sc Mean
0.500
(b)
0.400
0.300
0.200
0.100
0.000
Wavelength (nm)
SSpp-mean
-5.00
-6.00
(c)
-7.00
-8.00
-9.00
-10.00
-11.00
Multi-temporal SAR dataset (captured date)
24 Dec
SSpp-mean
231
-5.00
(d)
Backscattering coefficient (in dB scale)
-6.00
-7.00
-8.00
-9.00
-10.00
-11.00
19 Nov
13 Feb
19 Mar
23 Apr
02 Jul
06 Aug
15 Oct
24 Dec
Multi-temporal SAR dataset (capture date)
Ac-mean
AiG1-mean
AiG2-mean
Am-mean
KoG1-mean
KoG2-mean
SSpp-mean
Figure 4.6. The mean reflectance and backscatter signature of the mangrove
spectral classes – (a) mean spectral reflectance extracted from laboratory-measured
leave samples; (b) mean spectral reflectance extracted from Hyperion data; (c)
mean raw backscatter coefficient in dB scale; and (d) mean filtered backscatter
coefficient in dB scale.
Preliminary exploration of the mean spectral curves in Figure 4.6a shows that with
the exception of A. ilicifolius, the mangrove species have relatively similar spectral
response in green and red (500-680nm). Distinct difference of mean reflectance
among species is found in the infrared regions, especially in the near infrared
plateau at 700-1100nm and the shortwave infrared at 1150-1200nm. The
separability among the species varies across wavelengths. For instance, A. ilicifolius
and S. caseolaris have close mean spectral response at 750-1000nm and the
separability improved after 1000nm. It is noted that spectral step at 1000nm cannot
be eliminated though FS3 was sufficiently warmed up prior to the measurement.
In Figure 4.6b, the mean reflectance of species extracted from Hyperion data show
distinct separability across the spectrum except in the shortwave infrared regions at
1982-2386nm. For instances, species-pair K. obovata G1 and Sonneratia spp., shows
substantial mean spectral difference in visible (509-672nm) and shortwave infrared
(973-1064nm and 1165-1306nm), but the difference narrowed in near infrared
(733-886nm). The two spectral groups of A. ilicifolius show distinct difference in all
spectral regions except in the visible bands. The two K. obovata groups have mean
232
spectral reflectance difference found in all spectral regions except in the shortwave
infrared at 1498-1982nm. By comparing the mean reflectance across the species,
the worst spectral region in terms of separability is located at the visible region
where the species have close reflectance indicated by the overlap curves. For
instance, A. corniculatum, the two A. ilicifolius spectral classes and A. marina all
have similar mean spectral reflectance in green peak at 550nm. The observations
are similar to the laboratory-measured spectra.
The mean raw and filtered backscatter coefficient of multi-temporal radar data in
decibel scale are shown in Figures 4.6c and d. Signature difference among the
species varies among different dates. For instances, the two K. obovata groups have
maximum backscatter difference of about 1.5dB in data captured on 19 November
while the difference reduces to a minimum of 0.1db on 23 rd Apr in Figure 4.6c. The
A. ilicifolius G2 is barely separable with A. marina in data captured on 19 March and
23 April; however, large difference of about 1dB was observed on 19 November,
15th Oct and 24th Dec. The range of mean backscatter difference among the species
was reduced after the application of filter, but the signature difference was
maintained as shown in Figure 4.6d. Signature separbility between some species are
enhanced (e.g. Sonneratia spp. and others on 19th Mar and 06th Aug) while some are
curtailed (e.g. Sonneratia spp. and others on 23rd Apr).
Apart from comparison of mean differences, the non-parametric Mann-Whitney Utest was used to examine the potential separability of species. The Mann-Whitney
U-test was used to test if the median reflectance of each individual band is
statistically significantly different between two mangrove species or two spectral
classes. For 7 mangrove species and classes, there are 21 possible species pairs.
Figures 4.7 shows the summary of spectral wavelength positions at which the
species pairs are significantly different for the laboratory-measured leave spectra
and Hyperion data. Wavelengths at which the species pairs are significantly
different are indicated by colors. In Figure 4.7a, two species pairs including Ai-Ko
and Ai-Sa have over 90% of wavelength positions that are significantly different
from each other. Ac-Ai, Ai-Sc, Am-Ko and Sa-Sc have over 80% of wavebands that
are significantly different from each other. In Figure 4.7b, significant difference
233
coincides for the majority species pairs. Three species pairs including Ac-Sc, Ai1-Sc
and Am-Sc have over 70% of wavelength positions that are significantly different
from each other. Species pairs Ai1-Ko1, Ai1-Ko2, Ai2-Sc, Am-Ko1 and Ko1-Ko2 have
over 60% of wavebands that are significantly different from each other. The species
pairs Ac-Ai1 and Ai2-Ko2 have the lowest of 7% and 16% of wavelength positions
that are significantly different from each other respectively.
Figure 4.8 shows the summary of the backscatter time series in which species pairs
show statistically significant differences in median backscatter coefficient. The
suffixes 1 and 2 indicate the raw and Lee-filtered backscatter coefficient in linear
unit respectively. The species pair Am-Ko2 is significantly different in backscatter
coefficient in the whole time series. Out of 16 temporal dataset, Ai1-Ko2 and AmKo2 have 15 and 13 raw or filtered dataset that are significantly different from each
other respectively. The species pairs, Ai1-Am and Ai1-Sc have the lowest number of
1 and 3 temporal dataset that are significantly different from each other.
Species pairs Ai2-Ko2 and Am-Ko2 having low percentage of significant difference in
hyperspectral data are having comparatively high percentage of significant
difference in the radar temporal series. On the other hand, species pairs having
relatively high percentage of significant difference in hyperspectral data such as Ai1Sc, Am-Sc, Ai1-Am and Ac-Sc experiences relatively low percentage of significant
difference in radar temporal data. The simple comparison shows that hyperspectral
and radar data can be complement in terms of species pairwise distinguish.
234
Species Pair
15
Sa - Sc
14
Ko - Sc
13
Ko - Sa
12
Am - Sc
11
Am - Sa
10
Am- Ko
9
Ai - Sc
8
Ai - Sa
7
Ai- Ko
6
Ai - Am
5
Ac - Sc
4
Ac - Sa
3
Ac - Ko
2
Ac - Am
Ac - Ai
1
350
450
550
650
750
850
950
1050
1150
1250
1350
1450
1550
1650
1750
1850
1950
2050
2150
2250
2350
2450
Wavelength (nm)
Mean Reflectance of A. corniculatum
a
235
Species pair
21
Ko2- Sc
20
Ko1 - Sc
19
Ko1 - Ko2
18
Am - Sc
17
Am - Ko2
16
Am- Ko1
15
Ai2- Sc
14
Ai2 - Ko2
13
Ai2 - Ko1
12
Ai2 - Am
11
Ai1- Sc
10
Ai1 - Ko2
9
Ai1 - Ko1
8
Ai1- Am
7
Ai1 - Ai2
6
Ac - Sc
5
Ac - Ko2
4
Ac - Ko1
3
Ac - Am
2
Ac - Ai2
Ac - Ai1
1
479
579
679
779
879
979
1079
1179
1279
1379
1479
1579
1679
1779
1879
1979
2079
2179
2279
2379
Wavelength (nm)
Mean reflectance of A. corniculatum
b
236
Figure 4.7. Summary of the wavelength at which different species pairs show statistically significant differences in median reflectance for (a)
leave samples (b) Hyperion. The mean reflectance curve of A. corniculatum is overlaid as a reference to vegetation reflectance features.
Species pair
21
Ko2- Sc
20
Ko1 - Sc
19
Ko1 - Ko2
18
Am - Sc
17
Am - Ko2
16
Am- Ko1
15
Ai2- Sc
14
Ai2 - Ko2
13
Ai2 - Ko1
12
Ai2 - Am
11
Ai1- Sc
10
Ai1 - Ko2
9
Ai1 - Ko1
8
Ai1- Am
7
Ai1 - Ai2
6
Ac - Sc
5
Ac - Ko2
4
Ac - Ko1
3
Ac - Am
2
Ac - Ai2
1
Ac - Ai1
ASAR Temporal Series (Month-Day)
Figure 4.8. Summary of multi-temporal radar data in which species pairs show statistically significant differences in median backscatter
coefficient (1=raw σ in linear unit; 2=Lee filtered σ in linear unit; 3=raw σ in db unit; 4=Lee filtered σ in db unit).
237
In order to reveal the spectral bands that are most prominent in determining the
spectral difference among mangrove species, the number of species pair that is
significantly different
(   0.01 ) in each wavelength position was counted,
summarized and plotted in Figure 4.9. Figures 4.9a and b are the frequency plots of
significantly different species pairs extracted from leaf spectra measurement and
Hyperion data respectively overlaid with a typical mean vegetation reflectance
curve of A. corniculatum. The histogram indicates the frequency of mangrove
species pairs that are significantly different with reference to the wavelength in
Hyperion image. In Figure 4.9a, for instance, at 750nm, 14 species pairs (out of the
maximum 15 possible pairs) have significantly different reflectance. The three
highest frequencies (14, 13 and 12) were colored by red, green and purple
respectively. Out of 2151 wavebands, 1802 have a significant frequency of equal to
and more than 10. 58 wavebands having the maximum frequency of 14, mainly
concentrated in two spectral regions. The first is located at the blue absorption at
455-486nm while another located near the shoulder of red edge at 740-765nm.
When considering frequency of 12 and above, the number of wavebands expands
to 650 (30%) and a more general pattern can be observed. Major local maxima are
located in blue absorption at 380-506nm, red edge shoulder and near infrared
plateau at 740-840nm, shortwave-infrared spectrum at 1741-1800nm and also at
1315-1223nm. Apart from the maximum frequency, minimum frequency of 5 is
found in two spectral regions – near the green peak at 531-538nm and the red edge
at 727-729nm. Key local minima with frequency ranged between 5 and 8 extended
from green peak at 519nm to red absorption at 659nm and the upper part of red
edge at 723-733nm.
In Figure 4.9b, the four highest frequencies (16 – 19) are highlighted by orange,
purple, green and red in order of ascending frequency. Out of 158 wavebands, 110
have frequency of significance equal to and higher than ten. 2 wavebands having
the maximum frequency of 19, mainly concentrates in short-wave infrared region at
1740-1750nm. When considering frequency of 16 and above, the number of
significant wavebands increases to 30 and a more general pattern can be observed.
Major local maxima locate in the red edge at 713-723nm and shortwave-infrared
238
spectrum at 1165-1316nm and also at 1599-1770nm. Low frequency count of
significance is found in the latter part of shortwave infrared region. At 1488nm,
1982-2013nm, 2144-2154nm, 2295-2315nm and 2365-2386nm, no species pairs
show significantly different. That may probably be due to the presence of noise in
latter part of SWIR. Other key local minima with frequency ranged between 5 and 8
locate in the blue absorption at 479-489nm, near infrared at 923nm, shortwave
infrared at 1346nm, 1498nm, 2023-2154nm, 2184-2194nm, 2234-2254nm and
2285-2396nm.
Similarly, the frequency of significantly different species pairs after Mann-Whitney
U-Test at   0.01 for multi-temporal radar backscatter data is plotted in Figure
4.10. The highest frequency of 17 is highlighted in purple. Out of 16 dataset, half of
them have frequency of significance equal to and more than thirteen and they all
cluster in Lee-filtered dataset. Radar data captured on 19 November 2008 has the
highest frequency count of 17 while data capture in February, March, August and
October have 15 species pairs that are significantly different from each other.
Lowest frequencies are found in the raw backscatter dataset. The distinctive
difference suggests the filtered dataset provide more information on potential
discriminability of mangrove species when compared with the raw dataset.
239
15
0.7
14
13
0.6
11
0.5
10
9
0.4
8
7
0.3
Reflectance
Mann-Whitney U-Test Significance (Frequency)
12
6
5
0.2
4
3
0.1
2
1
0
0
350
450
550
650
750
850
950
1050
1150
1250
1350
1450
1550
1650
1750
1850
1950
2050
2150
2250
2350
2450
Wavelength (nm)
Frequency < 12
Frequency = 12
Frequency = 13
Frequency = 14
Mean Reflectance of A. corniculatum
a
240
20
0.6
19
18
17
0.5
15
14
0.4
13
12
11
10
0.3
9
Reflectance
Mann-Whitney U-Test Significance (Frequency)
16
8
7
0.2
6
5
4
0.1
3
2
1
0
0
479
530
580
631
682
733
784
835
886
963
1014
1064
1114
1205
1256
1306
1488
1538
1589
1639
1690
1740
1791
2023
2073
2123
2174
2224
2275
2325
2375
Wavelength (nm)
Frequency < 16
Frequency = 19
Frequency = 18
Frequency = 17
Frequency = 16
Mean reflectance of A. corniculatum
b
Figure 4.9. Frequency plot of significantly different species pairs after Mann-Whitney U-Test at   0.01 for (a) leave samples (b) Hyperion.
The mean reflectance curve of A. corniculatum is overlaid as a typical vegetation curve.
241
18
17
16
Mann-Whitney U-Test Significance (Frequency)
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
Wavelength (nm)
Frequency < 15
Frequency = 17
242
Figure 4.10. Frequency plot of significantly different species pairs after Mann-Whitney U-Test at   0.01 for multi-temporal radar backscatter
data (1=raw σ in linear unit; 2=Lee filtered σ in linear unit; 3=raw σ in db unit; 4=Lee filtered σ in db unit). The time series attaining the highest
frequency of 17 are highlighted.
The hyperspectral wavebands at which the median reflectance of individual
mangrove species is significantly different to the median reflectance of all other
species were summarized in Figure 4.11. Figure 4.11a and b are summaries from the
measured leaf spectra and Hyperion data respectively. In Figure 4.11a, all six species
have wavebands that are significantly different from all other species in median
reflectance. In terms of species, A.corniculatum is statistically different from all
other species in 40 wavebands locating within 385-456nm in the visible spectrum. A.
ilicifolius is statistically different from all other species in 1130 wavebands covering
spectral regions including visible spectrum at 420-705nm, near the red edge
shoulder at 740-778nm, the short-wave infrared spectrum and the water
absorption regions at 1399-1534nm/ 1866-2500nm. A. marina exhibits statistical
difference in median reflectance to all other species in 378 wavebands. They locate
near the red edge shoulder at 740-778nm, first half of the short-wave infrared
spectrum at 1318-1337nm/ 1550-1795nm/ 2135-2200nm. K. obovata is statistically
different to all other species in 135 wavebands covering spectral regions of very
narrow visible spectrum at 384-385nm, the first half of the short-wave infrared
spectrum at 1612-1643nm/ 1741-1759nm/ 2080-2124nm/ 2333-2400nm. S. apetala
is statistically different to all other species in 692 wavebands locating mainly in the
visible spectrum at 448-506nm, the short-wave infrared spectrum and the water
absorption regions at 1001-1387nm/ 1595-1840nm. Finally, S. caseolaris exhibits
statistical difference in median reflectance to all other species in 106 wavebands
with positions in the visible spectrum at 357-373nm/ 435-486nm, the red edge
shoulder at 740-765nm, and shortwave infrared at 2395-2466nm.
Along the wavelength, four wavebands are commonly identified by four species
having significant different median reflectance to other species and they locate in
the blue absorption at 435nm, 448nm, 449nm, 456nm. Besides, 107 wavebands are
commonly identified by three species having significantly different median
reflectance to other species and they locate in the blue absorption at 455-486, the
red edge shoulder at 740-765nm and the shortwave infrared at 1612-1643nm,
1741-1759nm, 2399-2400nm. These spectral regions are unique signature of the
mangrove species acquired from the leave samples.
243
In Figure 4.11b, expect A. corniculatum and A. ilicifolius group 2, the other five
species have wavebands that are significantly different from all other species in
median reflectance. In terms of species, A.corniculatum is statistically different from
all other species in 40 wavebands locating within 385-456nm in the visible spectrum.
A. ilicifolius group 1 is statistically different from all other species in 47 wavebands
covering the short-wave infrared spectrum at 1165-1780nm. A. marina exhibits
statistical difference in median reflectance to all other species in 38 wavebands.
They are located at the red edge and its shoulder at 713-794nm and the short-wave
infrared spectrum at 1165-1336nm/ 1558-1770nm. K. obovata group 1 is
statistically different to all other species in 28 wavebands covering the green peak
at 540-570nm, the center of the red edge at 723 and the first half of the short-wave
infrared spectrum at 963-1024nm/ 1165-1316nm. K. obovata group 2 exhibits
statistical difference in median reflectance to all other species in visible spectrum at
601-662nm and red edge at 702nm. Sonneratia spp is statistically different to all
other species in 56 wavebands locating mainly in the visible spectrum at 601-662nm,
the red edge at 702-723nm, the short-wave infrared spectrum at 963-1024nm,
1114-1336nm, 1579-1760nm and 2204-2214nm.
Ten wavebands locating in the shortwave infrared spectrum at 1165nm, 11851225nm, 1276nm, 1296-1316nm are commonly identified by four species having
significant different median reflectance to other species. When at least two species
are considered, 47 wavebands are commonly identified by three species having
significant different median reflectance to other species and they locate in visible
spectrum at 601-662nm, centre red edge at 702-723nm and shortwave infrared at
963-1024nm, 1165-1336nm and 1558-1770nm. To summarize, the first half of
shortwave infrared up to 1780nm is the unique signature in distinguishing the
mangrove species based on the Hyperion data.
Correspondingly, the temporal radar series at which the median backscatter
coefficient of each mangrove species is significantly different to the median
backscatter coefficient of all other six species were obtained and summarized in
Figure 4.12. No species is statistically different from all other species in the raw
backscatter coefficient dataset and also the filtered dataset acquired in December.
244
Except A. ilicifolius group 1 and A. ilicifolius group 2, the other five species indicate
radar datasets that are significantly different from all other species in median Leefiltered backscatter coefficient. In terms of species, A.corniculatum is statistically
different from all other species in February and April dataset. A. marina exhibits
statistical difference in median backscatter coefficient to all other species in 19
November. K. obovata group 1 is statistically different from all other species in 13
February. K. obovata group 2 exhibits statistical difference in median backscatter
coefficient to all other species in 5 datasets acquired in February, March, July,
August and October. Finally, Sonneratia spp is statistically different from all other
species in March and August datasets.
The February dataset is commonly identified by three species having significant
different median backscatter coefficient to other species while the March and
August datasets are commonly identified by two species.
245
6
S. caseolaris
5
S. apetala
K. obovata
Species
4
A. marina
3
A. ilicifolius
2
A. corniculatum
1
350
450
550
650
750
850
950
1050
1150
1250
1350
1450
1550
1650
1750
1850
1950
2050
2150
2250
2350
2450
Wavelength (nm)
Mean Reflectance of A. corniculatum
a
246
Species
7
Sonneratia spp.
6
K. obovata G2
5
K. obovata G1
A. marina
4
A. ilicifolius G2
3
A. ilicifolius G1
2
A. corniculatum
1
479
579
679
779
879
979
1079
1179
1279
1379
1479
1579
Wavelength (nm)
Mean reflectance of A. corniculatum
1679
1779
1879
1979
2079
2179
2279
2379
b
Figure 4.11. The wavelength at which individual mangrove species having statistically significant different median reflectance from all other
species under Mann-Whitney U-Test at   0.01 for (a) leave samples (b) Hyperion. The mean reflectance curve of A. corniculatum is overlaid
as a reference to vegetation reflectance features.
247
Sonneratia spp.
7
K. obovata G2
6
K. obovata G1
Species
5
A. marina
4
3
A. ilicifolius G2
2
A. ilicifolius G1
A. corniculatum
1
ASAR Temporal Series (Month-Day)
Figure 4.12. The multi-temporal radar backscatter coefficient of species with statistically significant different median reflectance to all other 6
mangrove species under Mann-Whitney U-Test at   0.01 (1=raw σ in linear unit; 2=Lee filtered σ in linear unit; 3=raw σ in db unit; 4=Lee
filtered σ in db unit).
248
Comparison of spectral regions between the laboratory spectra and Hyperion data
at which most species are statistically different suggested that they do not coincide
with each other. For instances, the red edge have been identified as an important
region to differentiate the majority of the species in both datasets, however, results
from leave spectra measurement indicated the significant regions located at the red
edge shoulder (740-765nm) while those from Hyperion pointed to the red edge well
(702-723nm). Similarly, the shortwave infrared has been recognized as another
critical region of species discrimination. Nevertheless, the laboratory spectra and
Hyperion data do not show correspondence. Results from the laboratory spectra
identified the latter half of SWIR (1612-1643nm, 1741-1759nm, 2399-2400nm)
while the Hyperion found mainly the first half of SWIR (963-1024nm, 1165-1336nm
and 1558-1770nm) as regions of significant spectral difference among species.
The lack of correspondence between spectra extracted from laboratory and
Hyperion sensor can be due to many factors such as variations of crown structure,
leaf angle, presence of non-vegetative elements and other factors. Although the
exploration of difference is not the prime objective of the study, such discrepancy
pinpoints two things. First, as spectra derived from the control environment cannot
reflect the true environment, they are highly unlikely to be used as surrogate for
spectra derived from satellite sensors. Second, the number of bands from the
laboratory-measured spectra is so enormous that band selection is still prohibitive
due to the requirement of substantial computational and time resources. Therefore,
feature selection was conducted based on the hyperion dataset.
4.3. Feature Selection
The results of different wrapper-based feature selection algorithms including
sequential forward selection (SFS), sequential floating forward selection (SFFS), and
oscillating search (OS) are summarized and presented in this section. The algorithms
are compared in terms of accuracy (training and testing) of classification algorithms
(k-nearest neighbor and support vector machines), selected features in spectral and
multi-temporal backscatter data as well as consistency with respect to four feature
249
subset sizes (d=5, 10, 15, 20). The final feature subsets deduced from frequency
analysis are further analyzed for their correlations and separability between classes.
4.3.1. Sequential Forward Selection (SFS)
For each combination of classifier and subset size, the training and testing accuracy
as well as different consistency indices after the 20 trials are summarized and
shown in Table 4.1.
In terms of classifier, KNN performs better than SVM with generally higher mean
training and testing classification accuracy. Besides, the standard deviation of
accuracy for SVM is much higher than that of KNN for all subset sizes. This is
confirmed by the large range (maximum - minimum) of accuracy difference. In
other words, the SVM classifier tends to produce results in more extreme ends
while the KNN is much more stable in the 20 trials.
With an increased subset size, both mean training and testing accuracy improve
gradually with the exception of SVM with15 features (SFS-SVM-15). The maximum
accuracy also enhances with subset size; however, there is no further improvement
in maximum accuracy for KNN after subset size of 15, the maximum training and
testing accuracy attained within 20 trials halt at 0.915 and 0.854 respectively. As for
SVM, the maximum accuracy keeps improving as more features are added to the
subset. The highest training and testing accuracy of 0.929 and 0.896 are attained in
a trial when SVM is used to select 20 features (SFS-SVM-20).
The consistency indices generally show low values indicating that selected subsets
do not show significant overlap in the 20 trials. The highest relative weighted
consistency (Cwrel) of 0.245 is found in KNN classifier with 20 features (SFS-KNN-20).
250
Table 4.1. Accuracy statistics and consistency indices after SFS feature selection
Training Accuracy
Testing Accuracy
Consistency
Classifiersubset size
Mean
SD
Max
Min
Mean
SD
Max
Min
Cwrel
ATI
CW
C
SFS-KNN-5
0.809
0.029
0.856
0.745
0.742
0.040
0.792
0.656
0.086
0.05
0.08
0.04
SFS-SVM-5
0.763
0.063
0.853
0.660
0.701
0.061
0.823
0.615
0.242
0.15
0.24
0.07
SFS-KNN-10
0.850
0.020
0.894
0.814
0.788
0.029
0.844
0.740
0.128
0.07
0.13
0.07
SFS-SVM-10
0.811
0.044
0.889
0.709
0.750
0.048
0.823
0.635
0.157
0.09
0.16
0.06
SFS-KNN-15
0.863
0.022
0.915
0.813
0.796
0.028
0.854
0.740
0.174
0.12
0.20
0.10
SFS-SVM-15
0.800
0.055
0.896
0.678
0.731
0.076
0.844
0.583
0.222
0.15
0.25
0.09
SFS-KNN-20
0.872
0.020
0.915
0.822
0.796
0.027
0.854
0.750
0.245
0.17
0.29
0.13
SFS-SVM-20
0.822
0.069
0.929
0.669
0.763
0.080
0.896
0.583
0.186
0.14
0.23
0.10
251
Prior to examination of individual feature subsets, the accuracy of each trial is
compared with the global mean accuracy computed based on subset size. The mean
accuracy and standard deviation are shown in Table 4.2. Accuracy of individual trial
is assessed based on the low limit of training and testing accuracy.
Table 4.2. Global mean and standard deviation of classification accuracy calculated
based on subset size
Training Accuracy
Testing Accuracy
d
Global
mean
Global SD
5
0.794
0.068
1 SD
below
mean
0.726
Global
mean
Global SD
0.724
0.063
1 SD
below
mean
0.661
10
0.842
0.055
0.787
0.770
0.063
0.708
15
0.847
0.056
0.790
0.770
0.066
0.704
20
0.849
0.066
0.782
0.771
0.072
0.699
After checking with the training and testing accuracy for individual trial, subsets
with accuracy lower than one standard deviation from mean are removed for
further consideration. The number of trials removed for the combinations is shown
in Table 4.3 below.
Table 4.3. Number of trial with accuracy below 1 standard deviation from mean (SFS
algorithm)
Classifier-subset size
Number of trial below the limit
SFS-KNN-5
SFS-SVM-5
SFS-KNN-10
SFS-SVM-10
SFS-KNN-15
SFS-SVM-15
SFS-KNN-20
SFS-SVM-20
2
9
0
5
0
9
0
5
252
A total of 30 trials are removed from different combinations of classifier-subset size
with the majority in SVM classifier. Almost half of the trials are removed under
subset of 5 and 15 using SVM. This further verifies the existence of extremely low
accuracy in some trials.
After the removal of subsets of low accuracy, features from the remaining trials are
examined with frequency analysis. Figures 4.13 and 4.14 show the frequency plots
of selected features under different subset sizes using k-nearest neighbor (KNN) and
support vector machines (SVM) respectively. The x-axis shows the feature number
including both hyperspectral and multi-temporal radar features and the y-axis
represents the frequency after 20 trials. The colors represent different subset sizes.
The mean reflectance curve of A. corniculatum is overlaid as a reference to indicate
typical vegetation reflectance features.
253
SFS-KNN
40
35
30
Frequency
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFS-KNN-5
SFS-KNN-10
SFS-KNN-15
SFS-KNN-20
Multi-temporal SAR
Mean Reflectance of A. corniculatum
Figure 4.13. Frequency plots of selected features using SFS wrapped by KNN in different subset sizes
254
SFS-SVM
40
35
30
Frequency
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFS-SVM-5
SFS-SVM-10
SFS-SVM-15
SFS-SVM-20
Multi-temporal SAR
Mean Reflectance of A. corniculatum
Figure 4.14. Frequency plots of selected features using SFS wrapped by SVM in different subset sizes
255
Generally, KNN exhibits distinctive high frequency in some features when compared
with SVM. For KNN, feature with the maximum frequency is found in SWIR at
1488nm (feature 80). Other features with high frequency locates in visible blue at
479nm (feature 0), green peak at 570-591nm (features 9-11), red absorption at 641
and 672nm (features 16 and 19), red-edge at 702-713nm (features 22-23),
shortwave infrared region at 1276nm, 1296nm and 1692nm (features 72, 74, 94)
and linear backscatter coefficient in Mar 19 (feature 163). For SVM, backscatter
coefficient in dB scale captured on Mar 19 (feature 179) is the most frequently
selected feature. Other important features position at visible blue at 479nm
(feature 0), red-edge at 713nm (feature 23) and also radar features acquired on Feb
13 in both linear and dB scales (features 162 and 178).
4.3.2. Sequential Floating Forward Selection (SFFS)
The training and testing accuracy and various consistency indices after the 20 trials
are summarized and shown in Table 4.4.
Similar to SFS, KNN performs better than SVM with generally higher mean training
and testing classification accuracy summarized from the results of 20 trials. The
range and standard deviation of accuracy for SVM is much higher than that of KNN
for all subset sizes. Therefore, the SVM classifier tends to produce more extreme
results or feature subsets while the KNN classifier is comparatively more stable in
terms of accuracy.
The increase subset size gradually improves the mean training and testing accuracy
for both classifiers. However, the trend reserves after feature size stepped from 15
to 20 for SVM classifier. The maximum accuracy (training and testing) also exhibits
the same reserve trend when SVM is used as a wrapper criterion. The maximum
accuracy improves in the beginning and decreases after subset size of 15. For KNN,
there is no further improvement in maximum accuracy for KNN after subset size of
15. The highest training and testing accuracy of 0.924 and 0.865 are attained in a
trial when KNN is used to select 20 features (SFFS-KNN-20).
256
The consistency indices generally show low values indicating that selected subsets
do not show significant overlap in the 20 trials. The highest relative weighted
consistency (Cwrel) of 0.311 is attained using SVM classifier with 5 features (SFFSSVM-5).
257
Table 4.4. Accuracy statistics and consistency indices after SFFS feature selection
Training Accuracy
Testing Accuracy
Consistency
Classifiersubset size
Mean
SD
Max
Min
Mean
SD
Max
Min
Cwrel
ATI
CW
SFFS-KNN-5
0.826
0.023
0.865
0.779
0.750
0.035
0.813
0.656
0.098
0.057
0.098
0.047
SFFS-SVM-5
0.741
0.085
0.885
0.628
0.672
0.075
0.844
0.563
0.311
0.220
0.311
0.082
SFFS-KNN-10
0.871
0.018
0.906
0.838
0.801
0.035
0.865
0.740
0.142
0.084
0.148
0.066
SFFS-SVM-10
0.832
0.048
0.916
0.711
0.754
0.072
0.875
0.552
0.212
0.129
0.218
0.077
SFFS-KNN-15
0.886
0.018
0.924
0.852
0.800
0.028
0.865
0.750
0.170
0.116
0.203
0.095
SFFS-SVM-15
0.851
0.057
0.929
0.705
0.758
0.070
0.917
0.594
0.157
0.109
0.191
0.094
SFFS-KNN-20
0.894
0.015
0.924
0.864
0.806
0.026
0.865
0.760
0.176
0.132
0.229
0.123
SFFS-SVM-20
0.828
0.050
0.915
0.745
0.742
0.059
0.833
0.646
0.173
0.132
0.226
0.108
C
258
After checking with the training and testing accuracy for individual trial, subsets
with accuracy below the limit which is defined in Table 4.2 are removed for further
consideration. The number of trial removed for the combinations is shown in Table
4.5.
Table 4.5. Number of trial with accuracy below one standard deviation from mean
(SFFS algorithm)
Classifier-subset size
Number of trial below the limit
SFFS-KNN-5
1
SFFS-SVM-5
12
SFFS-KNN-10
0
SFFS-SVM-10
4
SFFS-KNN-15
0
SFFS-SVM-15
4
SFFS-KNN-20
0
SFFS-SVM-20
9
A total of 30 trials are removed from different combinations of classifier-subset size
with the majority in SVM classifier. Half of the trials are removed under subset of 5
and 20 using SVM.
After the removal of subsets of low accuracy, features from the remaining trials are
examined with frequency analysis. Figures 4.15 and 4.16 show the frequency plots
of selected features under different subset sizes using k-nearest neighbor (KNN) and
support vector machines (SVM) respectively. The mean reflectance curve of A.
corniculatum is overlaid as a reference to indicate typical vegetation reflectance
features.
259
SFFS-KNN
45
40
35
Frequency
30
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFFS-KNN-5
SFFS-KNN-10
SFFS-KNN-15
Multi-temporal SAR
SFFS-KNN-20
Mean Reflectance of A. corniculatum
Figure 4.15. Frequency plots of selected features using SFFS wrapped by KNN in different subset sizes
260
SFFS-SVM
40
35
30
Frequency
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFFS-SVM-5
SFFS-SVM-10
SFFS-SVM-15
Multi-temporal SAR
SFFS-SVM-20
Mean Reflectance of A. corniculatum
Figure 4.16. Frequency plots of selected features using SFFS wrapped by SVM in different subset sizes
261
For KNN, a few distinctive features exhibit high frequency count. Feature
experiences the maximum frequency count is the filtered radar backscatter
coefficient acquired on Mar 19 (feature 163). Other features with high frequency
locates in green peak at 570-611nm (features 9-13), red-edge at 702-713nm
(features 22-23), the first half of shortwave infrared region at 1276nm (features 72)
and another linear backscatter coefficient acquired on Feb 13 (feature 162). For
SVM, regions of important features exhibiting high frequency count are located
though it is not as obvious as that of KNN. Identical to KNN, feature having the
highest frequency count is the radar backscatter coefficient in dB scale captured on
Mar 19 (feature 179) for SVM trials. Other important features position at red-edge
at 713nm (feature 23) and also radar backscatter features acquired on Nov 19 and
Feb 13 (features 162 and 163) in dB scales.
4.3.3. Oscillating Search (OS)
The training and testing accuracy and the consistency indices after 20 trials are
computed and shown in Table 4.6.
KNN performs better than SVM with generally higher mean training and testing
classification accuracy. The range and standard deviation of accuracy for SVM is
much higher than that of KNN for all subset sizes. In some subset size, for instance
20, the standard deviation of accuracy in SVM trials is more than four times than
that of KNN trials. Hence, the SVM classifier tends to produce more fluctuating
results or feature subsets while the KNN classifier is comparatively more stable in
terms of accuracy.
The increase subset size generally improves the mean training and testing accuracy
when KNN classifier is used while the situation is more fluctuated when SVM
classifier is applied. For KNN, the mean accuracy improves in the beginning for 5 to
10 features and exhibits a slight drop as feature size enlarges to 15. The accuracy
improves again in feature size of 20. For SVM, the improvement of average training
accuracy ceases after subset size of 10 features and it drops as features are added.
262
The mean testing accuracy drops in the beginning from 5 to 10 feature size and then
it improves a bit. However, it drops to the lowest level (0.697) which is even lower
than that of subset size of 5. The maximum accuracy (training and testing) also
exhibits similar trend. For KNN, there is no further improvement in maximum
accuracy for KNN after subset size of 15. For SVM, The maximum accuracy improves
in the beginning and decreases after subset size of 10. The highest training accuracy
of 0.951 is attained in a trial when SVM is used to select 20 features (OS-SVM-20).
However, the subset having the highest training accuracy does not exhibit the
highest testing accuracy. The highest testing accuracy of 0.906 is found in a SVM
wrapped search of 10 subset size (OS-SVM-10).
The consistency indices generally show low values indicating that selected subsets
do not show significant overlap in the 20 trials. The highest relative weighted
consistency (Cwrel) of 0.240 is attained using SVM classifier with 5 features (OSSVM-10).
263
Table 4.6. Accuracy statistics and consistency indices after OS feature selection
Training Accuracy
Testing Accuracy
Consistency
Classifiersubset size
Mean
SD
Max
Min
Mean
SD
Max
Min
Cwrel
ATI
CW
C
OS-KNN-5
0.848
0.015
0.882
0.818
0.753
0.042
0.823
0.656
0.093
0.053
0.093
0.045
OS-SVM-5
0.776
0.081
0.888
0.638
0.728
0.069
0.854
0.573
0.216
0.152
0.216
0.076
OS-KNN-10
0.894
0.011
0.919
0.875
0.808
0.027
0.854
0.750
0.142
0.085
0.148
0.062
OS-SVM-10
0.791
0.075
0.891
0.643
0.721
0.086
0.906
0.594
0.240
0.151
0.246
0.086
OS-KNN-15
0.893
0.016
0.934
0.873
0.805
0.055
0.885
0.677
0.102
0.078
0.139
0.068
OS-SVM-15
0.788
0.048
0.879
0.711
0.732
0.070
0.885
0.635
0.173
0.119
0.207
0.081
OS-KNN-20
0.901
0.018
0.928
0.873
0.820
0.035
0.885
0.750
0.108
0.093
0.166
0.088
OS-SVM-20
0.775
0.079
0.951
0.640
0.697
0.090
0.875
0.542
0.117
0.098
0.174
0.088
264
The inspection of training and testing accuracy for individual trial identifies subsets
with accuracy below the limit defined in Table 4.2. They are eliminated for further
analysis. The number of trial removed for the combinations is shown in Table 4.7.
Table 4.7. Number of trial with accuracy below one standard deviation from mean
(SFFS algorithm)
Classifier-subset size
Number of trial below the limit
OS-KNN-5
1
OS-SVM-5
7
OS-KNN-10
0
OS-SVM-10
10
OS-KNN-15
2
OS-SVM-15
11
OS-KNN-20
0
OS-SVM-20
13
A total of 44 trials are removed from different combinations of classifier-subset size
with the majority in SVM classifier. Half of the trials are removed under subset of 10,
15 and 20 using SVM as wrapper.
Features of the remaining trials after the removal of subsets of low accuracy are
examined with frequency analysis. Figures 4.17 and 4.18 show the frequency plots
of selected features under different subset sizes using k-nearest neighbor (KNN) and
support vector machines (SVM) respectively. The mean reflectance curve of A.
corniculatum is overlaid as a reference to indicate typical vegetation reflectance
features.
265
OS-KNN
50
45
40
35
Frequency
30
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
OS-KNN-5
OS-KNN-10
OS-KNN-15
OS-KNN-20
Multi-temporal SAR
Mean Reflectance of A. corniculatum
Figure 4.17. Frequency plots of selected features using OS wrapped by KNN in different subset sizes
266
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
Frequency
25
OS-SVM
20
15
10
5
0
Hyperspectral Features
OS-SVM-5
OS-SVM-10
OS-SVM-15
OS-SVM-20
Multi-temporal SAR
Mean Reflectance of A. corniculatum
Figure 4.18. Frequency plots of selected features using OS wrapped by SVM in different subset sizes
267
For KNN, a few distinctive regions exhibit high frequency count. Feature
experiences the maximum frequency count is the filtered radar backscatter
coefficient acquired on Mar 19 (feature 163). Other features with high frequency
clusters in boundary between green peak and red absorption at 601-621nm
(features 12-14), red-edge at 702-713nm and 733nm (features 22-23, 25), the
second half of shortwave infrared region at 1609nm and 1629-1639nm (features 92,
94-95) and filtered backscatter coefficient in linear scale acquired on Nov 19, Feb 13,
Aug 06, and Dec 24 (feature 161, 162, 166 and 168). For SVM, features having the
highest frequency count include filtered backscatter features in dB scale acquired
on Nov 19 and Mar 19 (features 177 and 179). Other regions of important features
exhibiting high frequency count locates in near infrared plateau at 753-764nm and
784nm (features 27-28 and 30) and radar backscatter features acquired on Feb 13
and Mar 19 (features 162 and 163) in linear scale as well as in dB scale.
4.3.4. Search Algorithms comparison
SFS is the most computational efficient among the three search methods. This is
followed by OS and SFFS. In terms of classifier, KNN is much more efficient than
SVM as parameter optimization is required for every trial when wrapped with SVM.
Figures 4.19 and 4.20 show the comparison of mean training and testing accuracy
across different search methods. Different colors represent different search
methods while the wrapped KNN and SVM classifiers are symbolized by solid and
dashed lines respectively. As reflected from the graph, KNN outperforms SVM in all
search algorithms and subset sizes in terms of accuracy and stability. Distinctive
difference is more obvious between search methods when wrapped with KNN than
using SVM as classifier. Under KNN, OS is consistently better than SFFS in both
training and testing accuracy in all subset sizes while SFS is the worst though the
accuracy is still better than results produced by SVM classifier. As subset size
increases, the resultant accuracy difference between OS and SFFS narrows.
268
Mean Training Accuracy
1.00
0.95
Accuracy
0.90
0.85
0.80
0.75
0.70
5
10
15
20
Subset Size (d)
SFFS-KNN
OS-KNN
SFS-KNN
SFFS-SVM
OS-SVM
SFS-SVM
Figure 4.19. Comparison of mean training accuracy of different search algorithms
Mean Testing Accuracy
1.00
0.95
0.90
Accuracy
0.85
0.80
0.75
0.70
0.65
0.60
5
10
15
20
Subset Size (d)
SFFS-KNN
OS-KNN
SFS-KNN
SFFS-SVM
OS-SVM
SFS-SVM
Figure 4.20. Comparison of mean testing accuracy of different search algorithms
When SVM is used as a wrapper, the mean training and testing accuracy indicates
SFFS performs the worst in subset size of 5. However, as d increases, SFFS becomes
269
the best among the three algorithms and reaching the maximum of 0.85, which is
the highest training accuracy attained in SVM wrapped results though it drops when
feature size increase from 15 to 20. SFS is relatively fluctuated than others as
reflected from both training and testing accuracy. The abrupt drops and rises of
accuracy in subset sizes of 15 and 20 respectively indicate the feature subsets are
not stable. The OS has the best performance in subset size of 5 but the accuracy
deteriorates afterwards and becomes the search algorithms with the worst
performance. This is a big contrast when compared with the results wrapped with
KNN classifier.
Based on the results from frequency analysis, a few features yield high frequency
count in most of the search algorithms. These common features locates in green
peak at 570-580nm and 601nm (features 9-10, 12), red-edge at 702-713nm
(features 22-23), shortwave infrared region at 1276nm, 1316nm, 1488nm (features
72, 76 and 80) and the filtered multi-temporal backscatter data acquired on Feb 13,
Mar 19, Aug 06 and Nov 19 (features 161, 162, 163, 166).
4.3.5. Final Subset Selection
Figures 4.21 – 4.26 plots the frequency count of best features for each criterion and
the mean reflectance of A. corniculatum is overlaid as a typical vegetation spectrum
for the hyperspectral features. When compared the frequency graphs for the first
four criteria, similar features of high frequency are observed. Three distinctive
frequency peaks position around red-edge (713nm), the shortwave infrared
(1276nm) and the filtered multi-temporal radar backscatter in linear scale.
Secondary peaks occur around regions including green peak (580nm), near infrared
plateau (774mm), the latter half of shortwave infrared (1629nm). In criterion V and
VI, similar high frequency regions are observed, but the highest frequency features
are found in the multi-temporal radar backscatter in linear scale, especially for
feature 162, i.e. filtered backscatter acquire on Feb 13. The distinctiveness is much
more obvious in criterion VI.
270
Criterion I
35
30
Frequency
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFFS-KNN-5
OS-KNN-5
SFS-KNN-5
SFFS-SVM-5
Multi-temporal SAR
OS-SVM-5
SFS-SVM-5
Mean Reflectance of A. corniculatum
Figure 4.21. Criterion I. Frequency count of features with subset size of 5 after removal of feature subsets having one standard deviation below
mean training and testing accuracy
271
Criterion II
45
40
35
Frequency
30
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFFS-KNN-10
OS-KNN-10
SFS-KNN-10
SFFS-SVM-10
Multi-temporal SAR
OS-SVM-10
SFS-SVM-10
Mean Reflectance of A. corniculatum
Figure 4.22. Criterion II. Frequency count of features with subset size of 10 after removal of feature subsets having one standard deviation
below mean training and testing accuracy
272
Criterion III
50
45
40
35
Frequency
30
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFFS-KNN-15
OS-KNN-15
SFS-KNN-15
SFFS-SVM-15
Multi-temporal SAR
OS-SVM-15
SFS-SVM-15
Mean Reflectance of A. corniculatum
Figure 4.23. Criterion III. Frequency count of features with subset size of 15 after removal of feature subsets having one standard deviation
below mean training and testing accuracy
273
Criterion IV
60
50
Frequency
40
30
20
10
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
SFFS-KNN-20
OS-KNN-20
SFS-KNN-20
SFFS-SVM-20
Multi-temporal SAR
OS-SVM-20
SFS-SVM-20
Mean Reflectance of A. corniculatum
Figure 4.24. Criterion IV. Frequency count of features with subset size of 20 after removal of feature subsets having one standard deviation
below mean training and testing accuracy
274
Criterion V
35
30
Frequency
25
20
15
10
5
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
Multi-temporal SAR
Mean Reflectance of A. corniculatum
Figure 4.25. Criterion V. Frequency count of features with the best training and testing accuracy in different subset sizes under different search
algorithm-classifier combination
275
Criterion VI
8
7
6
Frequency
5
4
3
2
1
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
0
Hyperspectral Features
Multi-temporal SAR
Mean Reflectance of A. corniculatum
Figure 4.26. Criterion VI. Frequency count of features with the best training and testing accuracy size in different subset sizes regardless of
search algorithm-classifier
276
The six criteria are combined and the final score is plotted in Figure 4.27. Table 4.8
shows the best 45 features arranged in descending criterion score and their
frequency counts in each criterion. Five features having the highest criterion score
of 6 locates in the center of the green peak at 570nm (feature 9), middle of rededge at 713nm (feature 23) and three filtered radar backscatter captured on Nov 19
(feature 161), Feb 13 (feature 162) and Mar 19 (feature 163). Five features position
in green peak at 580nm and 601nm (features 10 and 12), shortwave infrared at
1276nm and 1316nm (features 72 and 76) and filtered radar backscatter acquired
on Aug 06 (feature 166) have scores of 5. The other five features locates in green
peak at 591nm (feature 11), red-edge at 702nm (feature 22), near infrared shoulder
at 764-774nm (features 28 and 29) and shortwave infrared at 1629nm (feature 94)
have criterion score of 4. Five features with criterion score of 3 position in
shortwave infrared at 1205-1215nm and 1488nm (features 65, 66 and 80) and the
multi-temporal radar backscatter data acquired on Feb 13 and Mar 19 in dB scale
(feature 178 and 179). The remaining seven and eighteen features have score of 2
and 1 respectively. Features with criterion score of 4 or above are carried forward
for further analysis and processing.
277
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
46
48
50
52
54
56
58
60
62
64
66
68
70
72
74
76
78
80
82
84
86
88
90
92
94
96
98
100
102
104
106
108
110
112
114
116
118
120
122
124
126
128
130
132
134
136
138
140
142
144
146
148
150
152
154
156
158
160
162
164
166
168
170
172
174
176
178
180
182
184
Score
6
Final score computed from six criteria
5
4
3
2
1
0
Hyperspectral Features
Mean Reflectance of A. corniculatum
Figure 4.27. The criterion score computed from frequency count of features satisfying the six criteria
Multi-temporal SAR
278
Table 4.8. Best 45 features arranged in descending criterion score
Criteria (Frequency)
Feature
Number
Feature Name
9
Criterion
Score
I
II
III
IV
V
VI
Green_570
10
17
21
23
9
4
6
23
Rededge_713
18
42
45
43
25
5
6
161
LeeBSLin_1119
9
25
27
29
22
4
6
162
LeeBSLin_0213
20
38
37
40
30
7
6
163
LeeBSLin_0319
32
40
39
46
26
4
6
10
Green_580
9
16
22
25
9
0
5
12
Green_601
8
20
27
35
15
1
5
72
SWIR2_1276
21
20
27
20
9
1
5
76
SWIR2_1316
8
22
23
18
14
3
5
166
LeeBSLin_0806
9
13
16
26
11
4
5
11
Green_591
8
11
17
21
9
0
4
22
Red_702
7
15
23
33
11
1
4
28
NIR_764
8
11
17
15
9
4
4
29
NIR_774
9
20
20
17
10
0
4
94
SWIR3_1629
10
13
17
20
8
0
4
65
SWIR2_1205
8
9
8
21
9
1
3
66
SWIR2_1215
9
12
11
17
11
2
3
80
SWIR3_1488
1
14
21
39
7
0
3
178
LeeBSdB_0213
8
17
20
15
5
0
3
179
LeeBSdB_0319
13
30
29
19
5
0
3
0
Blue_479
0
8
29
51
6
1
2
14
Red_621
4
10
14
22
9
0
2
26
Rededge_743
10
13
11
15
4
0
2
41
NIR_896
4
6
9
11
10
2
2
75
SWIR2_1306
6
17
16
17
10
1
2
168
LeeBSLin_1224
2
12
17
14
7
2
2
177
LeeBSdB_1119
9
13
15
12
2
0
2
1
Blue_489
0
4
16
40
5
0
1
2
Blue_499
1
3
9
22
3
1
1
8
Green_560
7
9
10
21
1
0
1
13
Green_611
6
10
12
26
5
0
1
25
Rededge_733
13
8
7
17
2
0
1
27
NIR_753
5
12
13
14
8
2
1
279
30
NIR_784
3
15
12
10
2
1
1
58
SWIR1_1094
1
3
9
8
3
2
1
63
SWIR2_1185
3
4
11
20
4
0
1
68
SWIR2_1235
1
2
5
6
3
2
1
71
SWIR2_1266
4
6
10
12
6
2
1
74
SWIR2_1296
8
12
9
15
2
0
1
77
SWIR2_1326
4
5
12
16
9
1
1
93
SWIR3_1619
4
7
10
16
5
3
1
101
SWIR3_1700
0
10
6
9
5
2
1
123
SWIR4_2103
0
1
3
6
6
2
1
124
SWIR4_2113
0
1
0
3
4
2
1
165
LeeBSLin_0702
2
11
17
13
3
0
1
4.3.6. Correlation Analysis
By inspecting the final feature subset, it is very likely to find correlated features,
especially for the narrow bands in the hyperspectral dataset. Correlation analysis is
applied to the final subset to explore the potential correlation between different
features. The correlation matrix is plotted in graphical format in Figure 4.28. Highly
correlated feature pairs have data distribution along the 45° diagonal as highlighted
by red box in the figure.
280
Figure 4.28. Correlation matrix of the subset of 15 features
The correlation coefficient and the significance level (p-value) are checked for every
feature pair. Correspondingly, features with high correlation coefficient (greater
than 0.8) and p<0.01 are shown in Table 4.9.
281
Table 4.9. List of significantly highly correlated features in the final feature subset
Feature pair
9-10
9-11
9-12
9-22
10-11
10-12
10-22
11-12
11-22
12-22
28-29
72-76
Correlation coefficient
(*p<0.01)
0.993*
0.973*
0.930*
0.974*
0.973*
0.933*
0.943*
0.936*
0.943*
0.926*
0.995*
0.923*
High correlation is mainly found in features of adjacent wavelengths in the
hyperspectral dataset as well as between the linear- and dB-scaled filtered radar
dataset. In the final subset, spectral regions including the green peak, red-edge,
near-infrared and shortwave infrared have more than two being identified as
important features. It is well-understood that adjacent bands are highly correlated
in hyperspectral spectrum. Apart from hyperspectral data, high correlation of radar
feature expressed in different scales are expected. As highly-correlated features
provide no extra information to improve classification, they are redundant features
that should be discarded.
The twenty features in the final subset are examined for their correlation in
descending order of criterion score. The first five features yielding the highest
criterion score do not show significant and high correlation among the features,
hence, they are all retained. For the next five features having criterion score of 5,
two features, 72 and 166 are retained while others are discarded due to high
correlation. Lastly, for criterion score of 4, features 29 and 94 are retained and the
other 3 features are removed due to high correlation with previously retained
features. To summarize, 9 features are carried forward as final features for species
classification as shown in Table 4.10.
282
Table 4.10. The finalized features having the highest criterion score after feature
selection analysis
Feature Number
Feature Name
Criterion Score
9
Green_570
6
23
Rededge_713
6
161
LeeBSLin_1119
6
162
LeeBSLin_0213
6
163
LeeBSLin_0319
6
72
SWIR_1276
5
166
LeeBSLin_0806
5
29
NIR_774
4
94
SWIR_1629
4
4.4. Image Classification
Based on the results of feature selection as well as frequency analysis, the final
feature set are used to map the mangrove species using four different classifiers
namely, Gaussian Maximum Likelihood (ML), Decision Tree (DT), Artificial Neural
Network (ANN) and Support Vector Machines (SVM) in the study area. The final
feature set is combined into four different subsets in Table 4.11 as follows:
Table 4.11. Feature subsets formulated by the combining best features in the final
feature set
Subset
Feature Subset Description
Feature Number
1
Best multi-temporal radar features
161, 162, 163, 166
2
Best hyperspectral features only
9, 23, 29, 72, 94
3
Best 5 features in the final set
9, 23, 161, 162, 163
4
All best features in the final set
9,23,29,72,94,161,162,163,166
Prior to the classification exercise, the separability between the mangrove spectral
classes is examined using transformed divergence (TD). The classification results
283
and accuracy assessment of different classifiers are presented in the following
sections.
4.4.1. Mangrove Spectral Class Separability
The TD results of the seven mangrove spectral classes are shown in Figures 4.29 and
4.30. In Figure 4.29, features are added incrementally to TD calculation according to
the feature scores shown in Table 4.10. TD of different species pairs increase
gradually as features are added. When the subset sizes increase to 7 features (9, 23,
29, 72, 161, 162, 163, 166), all species pairs has TD over 1.7. After two more
features are added, TD has improved to above 1.9 which suggests good separability
between all species pairs. Out of 21 species pairs, 13 pairs have TD reaching the
maximum of 2.0. In the figure, species pair Ac-AiG2 shows the most distinctive
improvement with feature addition. Huge improvement of TD from 0.333 to 1.277
is observed after inclusion of the fourth feature (162) which is filtered backscatter
radar data captured on Feb 13. This signifies that the use of multi-temporal radar
data can enhance the separability of species. Two species pairs, AiG2-SSpp and AcSSpp exhibit the highest TD even under situation of two features.
Figure 4.30 is the TD plot with feature subset follows the combinations in Table 4.11.
TDs computed for the first subset are purely from radar features (161, 162, 163 and
166). The pure use of radar feature shows species separability of two extremes,
poor and good separability. Out of 21 species pairs, 12 pairs achieve good
separability with TD above 1.9 while TD of 6 pairs including Ac-AiG1, Ac-SSpp, AiG2SSpp, KoG1-SSpp, Am-SSpp and KoG2-SSpp have reached 2.0. The species pair AiG1KoG1 yields the poorest TD of 0.702. The average and standard deviation of TD for
all species pairs is 1.682 and 0.396 respectively.
TDs calculated from the second subset are computed from the combination of 5
best hyperspectral features according to the criterion score. Out of 21 species pairs,
8 pairs are of good separability with TD above 1.9 while only one species pair AmSSpp has achieved this highest TD of 2.0. The lowest separability of 0.947 is found in
284
the species pair Ac-AiG2. The low TD suggested that the two species, i.e. Ac and
AiG2 are non-separable based on spectral features alone. The average and standard
deviation of TD for all species pairs is 1.843 and 0.215 respectively.
The third subset which is the 5 features with the best criterion score combines both
hyperspectral and multi-temporal radar features. 13 species pairs have TD above
1.9, out of which 5 pairs reach the maximum of 2.0. Species pair, Ac-AiG2 yields the
lowest TD of 1.501. The average and standard deviation of TD for all species pairs is
1.874 and 0.157 respectively.
Finally, TD of the fourth subset is computed from all 9 best features. All species
pairs have TD above 1.9, out of which 13 pairs reach the maximum of 2.0. Although
species pair Ac-AiG2 still has the lowest separability, TD of 1.953 indicates good
separability. The average and standard deviation of TD for all species pairs is 1.989
and 0.016 respectively.
285
2.000
1.800
Transformed Divergence (TD)
1.600
1.400
1.200
1.000
0.800
0.600
0.400
0.200
0.000
2
3
4
5
6
7
8
9
Number of Features
Ac-AiG1
Ac-AiG2
Ac-Am
Ac-KoG1
Ac-KoG2
Ac_SSpp
AiG1-AiG2
AiG1-Am
AiG1-KoG1
AiG1-KoG2
AiG2-Am
AiG2-KoG1
AiG2-KoG2
AiG2-SSpp
Am-KoG1
Am-KoG2
Am-SSpp
KoG1-KoG2
KoG1-SSpp
KoG2-SSpp
AiG1-SSpp
Figure 4.29. Transformed divergence of 7 mangrove spectral classes under incremental increase of subset size (feature input in descending
order of criterion score)
286
2.000
Transformed Divergence (TD)
1.800
1.600
1.400
1.200
1.000
0.800
0.600
1
2
3
4
Subset
Ac-AiG1
Ac-AiG2
Ac-Am
Ac-KoG1
Ac-KoG2
Ac_SSpp
AiG1-AiG2
AiG1-Am
AiG1-KoG1
AiG1-KoG2
AiG2-Am
AiG2-KoG1
AiG2-KoG2
AiG2-SSpp
Am-KoG1
Am-KoG2
Am-SSpp
KoG1-KoG2
KoG1-SSpp
KoG2-SSpp
AiG1-SSpp
Figure 4.30. Transformed divergence of 7 mangrove spectral classes under different feature subsets selected from the final feature set
287
Based on TD results, a few observations are drawn. First, the species pair Ac-AiG2 is
the least separable in all feature combinations except when pure radar feature
subset is considered. Second, Sonneratia spp. (SSpp) is the only species that
experiences good separability with all other species in all four subsets in Table 4.11.
Third, as the number of features increase, species separability improves accordingly.
Fourth, the use of pure radar features yields the lowest average TD for the species
pairs. Although highly separable species pairs exist, a number of species pairs have
TD far below 1.7 on the other extreme causing high deviation of TD. Fifth, the pure
use of hyperspectral features show better results when compared with the pure use
of radar features. However, it seems that the feature subset cannot separate some
species pairs, especially the least separable one Ac-AiG2 causing extremely low TD.
Sixth, with the same feature size of 5, the combination of hyperspectral and radar
features has further improved the average TD, the number pairs having maximum
TD achieved as well as absence of pairs having extremely low TD as compared with
the pure hyperspectral feature subset. Seventh, when the best 8 features are input,
all species are well separable except Ac-AiG2. As the number of features increases
to the maximum of 9, all species pairs are of good separability. The gradual
improvement of separability evidences that the feature set deduced from feature
selection process is highly valid for classification purpose.
4.4.2. Gaussian Maximum Likelihood (ML)
The resultant classification maps using ML classifier under different subsets are
shown in Figures 4.31 – 4.34. Through visual-interpretation analysis of the twelve
sites, result from subset 3 indicates Sonneratia spp. as the predominant species in
site A while results from other subsets all provide inaccurate prediction. The
predominant A. ilicifolius in sites B and G are captured by the predictions from all
subsets except from subset 1. Similarly, species mix of A.ilicifolius and A.
corniculatum in site C is revealed in the results of all subsets except in subset 1.
Sites D, E and F are dominated by K. obovata G1, A. marina and K. obovata G2
respectively with elongated shape extending in north-south direction. The
288
predictions in three sites are consistent in all subsets though result from subset 1
exhibits a mix of the two K. obovata spectral groups in site F and the shrinked
distribution of A. marina in site E. For site H with predominant species is A. ilicifolius
G1, the predictions vary across subsets with the most satisfactory estimation under
subsets 2 and 4. Distribution in north-south direction with east-west strips cut
across the vertical distribution by the pioneer species is clearly observed from the
resultant classified map of the two subsets. Sites I and J are predominant by K.
obovata G1. The predictions from all four subsets have identified the species as K.
obovata but the accuracy on the spectral group varies. For site I, results from all
subsets are correct about the spectral group except that from subset 1 in which
there is a mix between the two groups. The same happens in site J from the results
of subsets 1 and 3 with spectral groups vary while the predictions are more
accurate in subsets 2 and 4. Finally, for sites K and L, they are predominant sites of
A. ilicifolius G2. The ML classification model predict site K correctly under subset 3
while results from subset 2 is accurate about the species but not the spectral group.
As for site L, except subset 1, predictions from all subsets reveal the accurate
distribution of A. ilicifolius G2 though result from subset 3 indicates minor mix with
A. corniculatum in the adjacent areas.
289
Figure 4.31. ML classification result using four filtered radar features (161, 162, 163
& 166) – subset 1
Figure 4.32. ML classification result using five narrow spectral features (9, 23, 29, 72
& 94) – subset 2
290
Figure 4.33. ML classification result using the first five features with the best
criterion score (9, 23, 161, 162 & 163) – subset 3
Figure 4.34. ML classification result using all nine features with the best criterion
score (9, 23, 29, 72, 94, 161, 162, 163 & 166) – subset 4
291
The confusion matrix of training and testing sets under different subsets is shown in
Tables 4.12 – 4.15. Individual mangrove spectral classes were investigated based on
the producer’s and user’s accuracy under different combinations of features, i.e.
subsets below.
In terms of producer’s accuracy, A. corniculatum attains the highest training
accuracy of 92% under subset 4. Subsets 1 and 3 yield the same training accuracy of
83% while subset 2 has the lowest accuracy of 75%. Except subset 1 which has a
testing accuracy of 67, other subsets have relatively low testing accuracy (0-33%). A.
ilicifolius G1 attains the highest accuracy of 91% in subset 4 while has the lowest
accuracy of 32% in the training phase. Same training accuracy of 77% is attained in
subsets 2 and 3. Subsets 2, 3 and 4 have the same testing accuracy of 67% while
subset 1 has the lowest of 22%. Generally, A. ilicifolius G2 attains better training and
testing accuracy in all subsets compared with A. ilicifolius G1. All subsets have
training accuracy above 80% except subset 1 which is 47%. The testing accuracy is
at maximum for subset 4 while the lowest (47%) is found in subset 1. A. marina
yields over 85% of accuracy for all subsets in both training and testing phases except
in subset 1. The highest achievable is 93% in subsets 2 and 4 in training phase while
subset 4 has the highest testing accuracy of 96%. For the two K. obovata groups,
they exhibits similar accuracy in the training phase with K. obovata G2 has slightly
higher user’s accuracy in the subsets. K. obovata G1 has high training accuracy
above 86% in all subsets except in subset 1 with the highest (97%) and lowest (61%)
found in subset 4 and 1 respectively. The testing phase follows the same trend as
the training one. The highest training accuracy of 100% for K. obovata G2 is
observed in subset 4 while the lowest (74%) is in subset 1. However, there is a sharp
drop in accuracy to 43% in subset 4 in the testing phase. The testing accuracy for
subsets stays close to the training accuracy. Sonneratia spp. has the maximum
accuracy in the first three subsets in the training phase. But the producer’s accuracy
is at minimum (0%) for all subsets in the testing phase.
In terms of user’s accuracy, A. corniculatum yields relatively high user’s accuracy of
over 90% derived from the training dataset in all subsets. However, the user’s
accuracy derived from the testing data set fluctuates over the subsets with 100%
292
accuracy in subsets 1 and 3 and 0% in subsets 2 and 4. A. ilicifolius G1 has extreme
user’s accuracy across different subset. All feature subsets has accuracy over 80%
from the training dataset except subset 1. And the user’s accuracy from
classification using subset 2 (5 features) is higher than that of subset 4 (9 features)
in terms of the training dataset though the reserve is observed in the testing
dataset. The accuracy of the A. ilicifolius G2 is relatively stable across different
subsets with the highest user’s accuracy reaching 94% in subset 4. Similar patterns
are observed in A. marina and K. obovata G2, but with higher user’s accuracy of
over 95% in subset 4 and lower user’s accuracy in subsets 1 and 3. For K. obovata
G1, both subsets 2 and 3 (5 features in the subset) have user’s accuracy of over 90%
which is higher than that of subset 4 (9 features). Subset one with pure radar
features performs the worst with 56% user’s accuracy. As for Sonneratia spp.,
classification of subset 4 has no correctly classified pixels in terms of user’s accuracy.
While the accuracy from subsets 1 and 3 reaches the maximum of 100%, the pure
use of hyperspectral features in subset 2 only has 70% correctly classified pixels.
Table 4.12. The confusion matrix after maximum likelihood (ML) classification using
feature combination of subset 1
Training dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
10
0
1
0
0
0
0
11
0
7
1
1
5
0
1
15
1
0
15
2
3
1
0
22
0
7
1
30
8
1
0
47
1
8
9
12
43
4
0
77
0
0
5
1
11
17
0
34
0
0
0
0
0
0
6
6
12
22
32
46
70
23
7
212
90.9
46.7
68.2
63.8
55.8
50.0
100.0
Producer's
accuracy (%)
83.3
31.8
46.9
65.2
61.4
73.9
85.7
60.4
293
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
2
0
0
0
0
0
0
2
0
2
1
0
1
0
2
6
1
0
8
0
2
0
0
11
0
4
2
19
3
0
0
28
0
3
5
4
21
2
2
37
0
0
1
0
7
5
0
13
0
0
0
0
0
0
0
0
3
9
17
23
34
7
4
97
100.0
33.3
72.7
67.9
56.8
38.5
0
Producer's
accuracy (%)
66.7
22.2
47.1
82.6
61.8
71.4
0.0
58.8
Table 4.13. The confusion matrix after maximum likelihood (ML) classification using
feature combination of subset 2
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
9
0
1
0
0
0
0
10
0
17
0
0
0
0
0
17
2
0
27
0
0
0
0
29
1
4
0
43
1
1
0
50
0
1
1
2
67
0
0
71
0
0
2
1
0
22
0
25
0
0
1
0
2
0
7
10
12
22
32
46
70
23
7
212
90.0
100
93.1
86.0
94.4
88.0
70.0
Producer's
accuracy (%)
75.0
77.3
84.4
93.5
95.7
95.7
100.0
90.6
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
0
0
3
0
0
0
0
3
0
6
0
1
0
0
0
7
3
1
13
1
1
1
0
20
0
0
0
21
0
0
0
21
0
1
0
0
31
0
4
36
0
1
1
0
0
6
0
8
0
0
0
0
2
0
0
2
3
9
17
23
34
7
4
97
0
85.7
65.0
100
86.1
75.0
0
Producer's
accuracy (%)
0
66.7
76.5
91.3
91.2
85.7
0
79.4
294
Table 4.14. The confusion matrix after maximum likelihood (ML) classification using
feature combination of subset 3
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
10
0
1
0
0
0
0
11
0
17
0
2
2
0
0
21
2
0
28
1
0
1
0
32
0
5
0
39
4
0
0
48
0
0
0
4
60
2
0
66
0
0
3
0
4
20
0
27
0
0
0
0
0
0
7
7
12
22
32
46
70
23
7
212
90.9
81.0
87.5
81.3
90.9
74.1
100
Producer's
accuracy (%)
83.3
77.3
87.5
84.8
85.7
87.0
100
85.4
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
1
0
0
0
0
0
0
1
0
6
0
0
1
1
2
10
2
1
13
0
0
0
0
16
0
1
2
21
1
0
0
25
0
1
1
2
30
0
2
36
0
0
1
0
2
6
0
9
0
0
0
0
0
0
0
0
3
9
17
23
34
7
4
97
100
60.0
81.3
84.0
83.3
66.7
0
Producer's
accuracy (%)
33.3
66.7
76.5
91.3
88.2
85.7
0
79.4
Table 4.15. The confusion matrix after maximum likelihood (ML) classification using
feature combination of subset 4
Training dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
11
0
0
0
0
0
0
11
0
20
0
0
1
0
0
21
1
0
31
0
1
0
0
33
0
2
0
43
0
0
0
45
0
0
1
3
68
0
7
79
0
0
0
0
0
23
0
23
0
0
0
0
0
0
0
0
12
22
32
46
70
23
7
212
100
95.2
93.9
95.6
86.1
100
0
Producer's
accuracy (%)
91.7
90.9
96.9
93.5
97.1
100
0
92.5
295
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
0
0
0
0
0
0
0
0
0
6
0
0
0
0
0
6
3
2
17
1
1
1
0
25
0
0
0
22
0
0
0
22
0
1
0
0
33
3
4
41
0
0
0
0
0
3
0
3
0
0
0
0
0
0
0
0
3
9
17
23
34
7
4
97
0
100
68
100
80.5
100
0
Producer's
accuracy (%)
0
66.7
100
95.7
97.1
42.9
0
83.5
296
4.4.3. Decision Tree (DT)
The resultant classification maps using DT classifier under different subsets are
shown in Figures 4.35 – 4.38. Through visual-interpretation analysis of the twelve
sites, predictions from subsets 1 and 4 have identified Sonneratia spp. as the
predominant species in site A while results from other subsets have predictions mix
with other species. The predominant A. ilicifolius stand in site B is revealed from the
results of all subsets except in subset 1. However, the subsets show different
degrees of mix with other species. Similarly, various levels of species mix of
A.ilicifolius, A. corniculatum and K. obovata are revealed from the predictions of all
subsets in site C. Although result from subset 1 indicates species mix, the
predominant species is identified as A. marina which is inaccurate. Sites D, E and F
are dominated by K. obovata G1, A. marina and K. obovata G2 respectively
occupying the central part of the study area in north-south direction. The
predictions in three sites are consistently revealed in all subsets though subset 1 has
mixed up the two K. obovata spectral groups in site F and the shrink distribution of
K. obovata G1 and mix with A. marina in site D. For site H with predominant A.
ilicifolius G1 stand, model predictions vary from little occurrence in subset 1 to
reasonable distribution in subsets 2 and 4. Distinctive north-south direction is
clearly indicated; however, the east-west strips along the old river channels are
under-estimated in both subsets. Predominant K. obovata G1 stand in sites I and J
are predicted with different degrees of success under the four subsets. For site I,
model prediction from all subsets are correct about the species and the spectral
group except subset 1 under with there is a mix between the two spectral groups. In
site J, similar observations are found from the results of subsets 1 and 3 with
spectral groups vary while more accurate predictions are revealed from the
prediction of subsets 2 and 4. Finally, for the predominant A. ilicifolius G2 stand in
sites K and L, the model predicts site K correctly under subset 3 while other subsets
fail to indicate the species predominance. In site L, results from subsets 3 and 4
reveal the accurate distribution of A. ilicifolius G2 except though the prediction from
subset 3 indicates minor mix with A. corniculatum around areas. Result from subset
4 reveals the best scene for this site.
297
Figure 4.35. DT classification result using four filtered radar features (161, 162, 163
& 166) – subset 1
Figure 4.36. DT classification result using five narrow spectral features (9, 23, 29, 72
& 94) – subset 2
298
Figure 4.37. DT classification result using the first five features with the best
criterion score (9, 23, 161, 162 & 163) – subset 3
Figure 4.38. DT classification result using all nine features with the best criterion
score (9, 23, 29, 72, 94, 161, 162, 163 & 166) – subset 4
299
The confusion matrix of training and testing sets under different subsets is shown in
Tables 4.16 – 4.19. Individual mangrove spectral classes were investigated based on
the producer’s and user’s accuracy under different combinations of features below.
For producer’s accuracy, A. corniculatum performs satisfactorily in all subsets in the
training phase. However, poor performance is observed in the testing phase with
the first three subsets yields 0% accuracy. As for A. ilicifolius G1, the producer’s
accuracy is poor (32%) for subset 1 with only radar features added. Subsets 2 and 4
have better performances with producer’s accuracy over 85% though accuracy from
the testing data is far below the training ones. A. ilicifolius G2 has similar pattern of
producer’s accuracy as A. ilicifolius G1 but with improved accuracy in subset 2. A.
marina has high producer’s accuracy of over 90% in all subsets in the training phase
while that from the testing phase is also comparable with the lowest of 74% in
subset 1. K. obovata G1 and G2 show similar patterns in producer’s accuracy with
the 9-feature subset performs the best in the training phase. This is followed by
subset 2, 3 and 1. One disparity between the two species originates from the testing
phase with K. obovata G2 performs poorer in subsets 1 and 4. Sonneratia spp. has
higher producer’s accuracy of over 85% in subset 4, 3 and 1 while the pure
hyperspectral data input yields (subset 2) yields the lowest (57%). However, for all
subsets, they yield the same producer’s accuracy in the testing set.
For user’s accuracy, A. corniculatum performs satisfactorily in all subsets in the
training phase. However, poor performance is observed in the testing phase with
the first three subsets yields 0% accuracy while that of the fourth subset reaches
the maximum of 100%. The user’s accuracy from the training phase for A. ilicifolius
G1 is generally high as 3 subsets including 1, 2 and 4 has 100% accuracy. However,
the accuracy from the testing phase fluctuates with 0% found in subset 1, lower
accuracy in subset 3 and 4 (about 60%) and 100% in subset 2. A. ilicifolius G1 and A.
marina show similar accuracy pattern in the training phase with subset 4 performs
the best and follows by subsets 2, 3 and 1. However, A. marina experiences much
higher testing accuracy in all subsets except subset 1. Both the training and testing
accuracy have exceeded 90% for subsets 2 and 4. For K. obovata G1, the training
accuracy is over 90% in all subsets except subset 1. With the same number of
300
features in subsets 2 and 3, the performance of subset 3 is slightly better than that
of subset 2 in this mangrove spectral class. The accuracy from the testing phase
follows the same trend and close to the accuracy from the training phase. K.
obovata G2 has better user’s accuracy (96%) in subsets 2 and 4 in the training phase.
However, the testing accuracy is about 30% lower. Subset 1 has the poorest
performance in both training and testing phases. Similar observations are found in
the training phase for Sonneratia spp. with user’s accuracy reaching the maximum
in subsets 2 and 4 and subset 1 yields the lowest accuracy of 50%. The accuracy in
the testing phase also reaches the maximum in subsets 2 and 3 while that for subset
4 is only 75%.
Table 4.16. The confusion matrix after decision tree (DT) classification using feature
combination of subset 1
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
10
0
2
0
0
0
0
12
0
7
0
0
0
0
0
7
0
0
14
2
1
0
1
18
0
8
5
44
11
1
0
69
1
4
10
0
52
5
0
72
0
0
1
0
4
17
0
22
1
3
0
0
2
0
6
12
12
22
32
46
70
23
7
212
83.3
100
77.8
63.8
72.2
77.3
50.0
Producer's
accuracy (%)
83.3
31.8
43.8
95.7
74.3
73.9
85.7
70.8
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0
1
4
2
1
0
0
8
2
5
7
17
8
0
0
39
1
2
3
4
19
5
0
34
0
0
1
0
5
2
0
8
0
1
2
0
1
0
3
7
3
9
17
23
34
7
4
97
0
0
50.0
43.6
55.9
25.0
42.9
Producer's
accuracy (%)
0.0
0.0
23.5
73.9
55.9
28.6
75.0
46.4
301
Table 4.17. The confusion matrix after decision tree (DT) classification using feature
combination of subset 2
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
11
0
0
0
0
0
0
11
0
19
0
0
0
0
0
19
1
0
30
0
0
1
0
32
0
2
0
44
2
0
0
48
0
1
1
2
68
0
3
75
0
0
1
0
0
22
0
23
0
0
0
0
0
0
4
4
12
22
32
46
70
23
7
212
100
100
93.8
91.7
90.7
95.7
100
Producer's
accuracy (%)
91.7
86.4
93.8
95.7
97.1
95.7
57.1
93.4
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
0
0
2
0
0
0
0
2
0
5
0
0
0
0
0
5
1
1
10
1
0
0
0
13
0
0
0
22
1
0
0
23
1
2
3
0
33
0
1
40
1
1
2
0
0
7
0
11
0
0
0
0
0
0
3
3
3
9
17
23
34
7
4
97
0
100
76.9
95.7
82.5
63.6
100
Producer's
accuracy (%)
0
55.6
58.8
95.7
97.1
100
75.0
82.5
Table 4.18. The confusion matrix after decision tree (DT) classification using feature
combination of subset 3
Training dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
11
0
2
0
0
0
0
13
0
15
2
0
0
0
0
17
1
0
22
1
0
2
1
27
0
6
1
43
3
0
0
53
0
0
2
2
64
1
0
69
0
0
3
0
2
20
0
25
0
1
0
0
1
0
6
8
12
22
32
46
70
23
7
212
84.6
88.2
81.5
81.1
92.8
80.0
75.0
Producer's
accuracy (%)
91.7
68.2
68.8
93.5
91.4
87.0
85.7
85.4
302
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
0
0
0
0
0
0
0
0
0
5
2
0
1
0
0
8
2
0
8
3
1
0
0
14
1
3
4
19
2
0
0
29
0
0
2
1
29
0
1
33
0
1
1
0
1
7
0
10
0
0
0
0
0
0
3
3
3
9
17
23
34
7
4
97
0
62.5
57.1
65.5
87.9
70.0
100
Producer's
accuracy (%)
0
55.6
47.1
82.6
85.3
100
75.0
73.2
Table 4.19. The confusion matrix after decision tree (DT) classification using feature
combination of subset 4
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
12
0
0
0
0
0
0
12
0
21
0
0
0
0
0
21
0
0
31
0
0
0
0
31
0
1
0
46
0
0
0
47
0
0
0
0
70
0
0
70
0
0
1
0
0
23
0
24
0
0
0
0
0
0
7
7
12
22
32
46
70
23
7
212
100
100
100
97.9
100
95.8
100
Producer's
accuracy (%)
100
95.5
96.9
100
100
100
100
99.1
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
1
0
0
0
0
0
0
1
0
6
2
0
0
1
0
9
2
2
12
2
0
1
0
19
0
0
1
21
1
0
0
23
0
1
0
0
32
0
1
34
0
0
2
0
0
5
0
7
0
0
0
0
1
0
3
4
3
9
17
23
34
7
4
97
100
66.7
63.2
91.3
94.1
71.4
75.0
Producer's
accuracy (%)
33.3
66.7
70.6
91.3
94.1
71.4
75.0
82.5
303
4.4.4. Artificial Neural Network (ANN)
The resultant classification maps using ANN classifier under different subsets are
shown in Figures 4.39 – 4.42. Through visual-interpretation analysis of the twelve
sites, model predictions from all subsets except subset 2 have identified Sonneratia
spp. as predominant species in site A. The predominant A. ilicifolius stand in site B is
captured by the prediction from all subsets except in subset 1 but results tend to
mix with other species in different degrees. For instance, result from subset 2 tends
to mix with A. marina, K. obovata and A. corniculatum, result from subset 3 shows
the mix with A. marina and K. obovata only and result from subset 4 shows the
largest predominance. In site C, different levels of species mix of A.ilicifolius, A.
corniculatum and K. obovata are observed from the predictions of all subsets except
in subset 1. Although subset 1 indicates species mix, the predominant species is
identified as A. marina and A. corniculatum which is largely departed from the
observation. The predominant species of K. obovata G1, A. marina and K. obovata
G2 in sites D, E and F respectively show distinctive and consistent distribution from
the results of all subsets except from subset 1. Result from subset 1 has mixed up
the two K. obovata spectral groups in site F and the shrink distribution of K. obovata
G1 in site D and mix with A. marina and A. corniculatum. The prediction of
predominant A. ilicifolius G1 stand in site H shows different degrees of success. The
distinctive north-south direction is clearly shown in the results of subsets 2 – 4 while
the prediction is inaccurate under subset 1. However, the east-west strips along the
old river channels are most distinctive in subset 4. For the predominant K. obovata
G1 stand in site I, predictions from all subsets are correct about the species and the
spectral group except subset 1 under with there is a mix between the two spectral
groups. For site J, the worst result is found in prediction under subset 2 in which
there is a mix of K. obovata, A. marina, A. ilicifolius and Sonneratia spp. Although
the predictions about the species are correct under subsets 1 and 3, the two
spectral groups mix with each other. More accurate and concise prediction is
presented in results under subset 4. Finally, for the predominant A. ilicifolius G2
stand in sites K and L, the model predicts site K correctly under subset 3 while other
subsets fail to indicate species predominance in the site. In site L, results in all
304
subsets reveal the accurate distribution of A. ilicifolius G2 except subsets 1 though
results under subsets 3 and 4 indicate minor mix with A. corniculatum around areas.
Result from subset 3 reveals the best scene for this site.
305
Figure 4.39. ANN classification result using four filtered radar features (161, 162,
163 & 166) – subset 1
Figure 4.40. ANN classification result using five narrow spectral features (9, 23, 29,
72 & 94) – subset 2
306
Figure 4.41. ANN classification result using the first five features with the best
criterion score (9, 23, 161, 162 & 163) – subset 3
Figure 4.42. ANN classification result using all nine features with the best criterion
score (9, 23, 29, 72, 94, 161, 162, 163 & 166) – subset 4
307
The confusion matrix of training and testing sets under different subsets is shown in
Tables 4.20 – 4.23. Individual mangrove spectral classes were investigated based on
the producer’s and user’s accuracy under different combinations of features below.
In terms of producer’s accuracy, A. corniculatum has the highest accuracy of 92%
and 100% in training and testing phases respectively when subset 4 is input while
the lowest accuracy of 25% and 0% results from subset 2. The other two subset
input provides a moderate accuracy in training and testing datasets. A. ilicifolius G1
attains the highest accuracy of 91% and 87% in training and testing respectively in
subsets 3 and 4. Subset 2 also gives a good accuracy of 86% while subset 1 provides
the worst results of 55%. A. ilicifolius G2 attains high accuracy of over 80% all
subsets except subset 1. The accuracy in the training and testing phases are the
same. The performance for A. marina is relatively satisfactory in all subsets with
accuracy higher than 80% in both training and testing phases. Particularly for this
species, the best subset is 2. For the two K. obovata spectral groups, the accuracy is
high of over 90% in all subsets except subset 1. Sonneratia spp. has been totally
classified correctly in terms of producer’s accuracy when pure radar features are
use in subset 1. Subsets 3 and 4 also demonstrate satisfactory results with accuracy
of 86% and 100% in training and testing phases respectively. While subset 2
produces the worst result of 57% in the training phase, it has 100% accuracy in the
testing phase.
In terms of user’s accuracy, A. corniculatum attains the maximum accuracy in the
training phase for all subsets except subset 4. However, the testing accuracy
fluctuates across different subsets with the largest discrepancy found in the second
subset having 0% testing accuracy. The most consistent one is found for subset 1
which has same 100% correctly classified pixels as the training phase. A. ilicifolius
G1 attains relatively high training and testing accuracy of about 90% in all subsets
except subset 1. The maximum accuracy is achieved when subset 4 when all
features are input. As for A. ilicifolius G2, the species achieves higher accuracy of
over 90% in both training and testing phases for subsets 3 and 4. Comparatively
lower accuracy is found in the other subsets. A. marina is satisfactorily classified
with over 85% of classification accuracy in subsets 2, 3 and 4 with the highest
308
attained in subset 4. For the two K. obovata spectral groups, the accuracy is high of
over 80% in both training and testing phases in all subsets except subset 1. While K.
obovata G2 has better performance in the training phase, K. obovata G1 is more
accurate in the testing phase except in subset 4. Sonneratia spp. attained the
maximum accuracy in both training and testing phases in almost all subsets except
the subset 1 in the training phase and subset 3 in the testing phase.
Table 4.20. The confusion matrix after artificial neural network (ANN) classification
using feature combination of subset 1
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
9
0
0
0
0
0
0
9
0
12
1
0
2
0
0
15
2
1
19
0
1
2
0
25
0
4
1
42
10
0
0
57
1
5
7
3
54
7
0
77
0
0
4
0
2
14
0
20
0
0
0
1
1
0
7
9
12
22
32
46
70
23
7
212
100
80.0
76.0
73.7
70.1
70.0
77.8
Producer's
accuracy (%)
75.0
54.5
59.4
91.3
77.1
60.9
100
74.1
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
2
0
0
0
0
0
0
2
0
4
0
2
2
0
0
8
1
1
10
0
1
0
0
13
0
2
2
20
4
0
0
28
0
1
3
1
25
2
0
32
0
1
2
0
2
5
0
10
0
0
0
0
0
0
4
4
3
9
17
23
34
7
4
97
100
50.0
76.9
71.4
78.1
50.0
100
Producer's
accuracy (%)
66.7
44.4
58.8
87.0
73.5
71.4
100
72.2
309
Table 4.21. The confusion matrix after artificial neural network (ANN) classification
using feature combination of subset 2
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
3
0
0
0
0
0
0
3
0
19
0
0
1
1
0
21
8
0
28
0
2
1
1
40
1
2
1
44
3
0
0
51
0
1
2
2
63
0
2
70
0
0
1
0
1
21
0
23
0
0
0
0
0
0
4
4
12
22
32
46
70
23
7
212
100
90.5
70.0
86.3
90.0
91.3
100
Producer's
accuracy (%)
25.0
86.4
87.5
95.7
90.0
91.3
57.1
85.8
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
0
0
1
0
0
0
0
1
0
8
0
0
0
0
0
8
2
1
15
0
0
1
0
19
1
0
0
23
3
0
0
27
0
0
0
0
31
0
0
31
0
0
1
0
0
6
0
7
0
0
0
0
0
0
4
4
3
9
17
23
34
7
4
97
0
100
78.9
85.2
100
85.7
100
Producer's
accuracy (%)
0
88.9
88.2
100.0
91.2
85.7
100.0
89.7
Table 4.22. The confusion matrix after artificial neural network (ANN) classification
using feature combination of subset 3
Training dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
10
0
0
0
0
0
0
10
0
20
0
1
0
0
0
21
1
0
28
0
1
0
0
30
0
2
1
38
3
0
0
44
1
0
2
7
64
1
1
76
0
0
1
0
2
22
0
25
0
0
0
0
0
0
6
6
12
22
32
46
70
23
7
212
100
95.2
93.3
86.4
84.2
88.0
100
Producer's
accuracy (%)
83.3
90.9
87.5
82.6
91.4
95.7
85.7
88.7
310
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
2
0
1
0
0
0
0
3
0
7
0
1
0
0
0
8
1
0
15
0
0
0
0
16
0
0
1
22
0
0
0
23
0
1
0
0
32
0
0
33
0
1
0
0
1
7
0
9
0
0
0
0
1
0
4
5
3
9
17
23
34
7
4
97
66.7
87.5
93.8
95.7
97.0
77.8
80.0
Producer's
accuracy (%)
66.7
77.8
88.2
95.7
94.1
100
100
91.8
Table 4.23. The confusion matrix after artificial neural network (ANN) classification
using feature combination of subset 4
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
11
0
1
0
0
0
0
12
0
20
0
0
0
0
0
20
1
0
30
0
0
1
0
32
0
2
1
43
0
0
0
46
0
0
0
3
70
0
1
74
0
0
0
0
0
22
0
22
0
0
0
0
0
0
6
6
12
22
32
46
70
23
7
212
91.7
100
93.8
93.5
94.6
100
100
Producer's
accuracy (%)
91.7
90.9
93.8
93.5
100
95.7
85.7
95.3
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
3
0
1
0
0
0
0
4
0
7
0
0
1
0
0
8
0
1
16
0
0
0
0
17
0
0
0
23
0
0
0
23
0
1
0
0
33
0
0
34
0
0
0
0
0
7
0
7
0
0
0
0
0
0
4
4
3
9
17
23
34
7
4
97
75.0
87.5
94.1
100
97.1
100
100
Producer's
accuracy (%)
100
77.8
94.1
100
97.1
100
100
95.9
311
4.4.5. Support Vector Machines (SVM)
Table 4.24 shows the optimized hyperparameter combinations, i.e. the penalty
coefficient ( C ) and kernel (  ) after accuracy assessment using cross-validation
method in different subsets.
Table 4.24. The optimized hyperparameters combinations and their corresponding
accuracy in different subsets
Subset
Penalty coefficient
(C )
Kernel (  )
Training
Accuracy
Testing
Accuracy
1
0.55
633
60.8
59.8
2
0.15
577
92.5
85.6
3
0.45
1282
88.2
76.3
4
0.1
7
93.9
90.7
The resultant classification maps using SVM classifier under different subsets are
shown in Figures 4.43 – 4.46. Through visual-interpretation analysis of the twelve
sites, model predictions in all subsets except subset 2 have identified Sonneratia
spp. as predominant species in site A and the prediction under subset 4 is the most
satisfactory one. The predominant A. ilicifolius stand in site B is captured by all
subsets except in subset 1 but predictions tend to mix with other species in
different degrees. For instance, result from subset 2 tends to mix with A. marina
and A. corniculatum, result from subset 3 shows the mix with A. marina and K.
obovata only and result from subset 4 shows the largest predominance of A.
ilicifolius. In the active growing zone in site C, results from subsets 2 and 4 show
clearly the mix of A.ilicifolius, A. corniculatum and K. obovata while those from
subset 1 and 3 indicates A. corniculatum as the predominant species in the site. The
elongated distribution of predominant species of K. obovata G1, A. marina and K.
obovata G2 in sites D, E and F respectively show distinctive and consistent
distribution from the predictions of all subsets. Similar to other algorithms, the two
K. obovata spectral groups are mixed together in site F from the result of subset 1.
The predominant A. ilicifolius G1 stand in site H is shown in the results from all
312
subsets except from subset 1 in which A. marina is identified instead. The distinctive
north-south direction is clearly shown from the results of subsets 2 – 4. However,
the east-west strips along the old river channels are most distinctive in subset 4.
Besides, there is over-estimation of the extent in the result of subset 2. The
prediction of predominant K. obovata G1 stands in sites I and J vary across the
results from different subsets. For site I, the predominant species is correctly
predicted in subsets 2 and 4 while the results from the other two subsets indicate a
mix with A. ilicifolius. For site J, the worst result is found in prediction under subset
2 in which a mix of K. obovata, A. marina, A. ilicifolius and Sonneratia spp. is found.
Although the prediction results from subsets 1, 3 and 4 in site J is correct about the
species, the two spectral groups mix with each other as indicated by mix of dark and
light green. Finally, for the predominant A. ilicifolius G2 stand in sites K and L, result
from subset 3 indicates clearly a predominant A. ilicifolius G2 stand in the center
though a minor mix with K. obovata and A. corniculatum is found. Results under
other subsets fail to indicate species predominance in the site. As for site L, results
from all subsets reveal the accurate distribution of A. ilicifolius G2 except that from
subsets 1 though result from subset 3 shows certain mix with A. corniculatum
around areas. Results from subsets 2 and 4 capture the best scene for this site.
313
Figure 4.43. SVM classification result using four filtered radar features (161, 162,
163 & 166) – subset 1
Figure 4.44. SVM classification result using five narrow spectral features (9, 23, 29,
72 & 94) – subset 2
314
Figure 4.45. SVM classification result using the first five features with the best
criterion score (9, 23, 161, 162 & 163) – subset 3
Figure 4.46. SVM classification result using all nine features with the best criterion
score (9, 23, 29, 72, 94, 161, 162, 163 & 166) – subset 4
315
The confusion matrix of training and testing sets under different subsets is shown in
Tables 4.25 – 4.28. Individual mangrove spectral classes were investigated based on
the producer’s and user’s accuracy under different combinations of features below.
For producer’s accuracy, A. corniculatum attains the highest accuracy of 83% in
subsets 1, 3 and 4 while the other subset 2 produce a relatively low accuracy of 67%
in the training phase. The testing accuracy for all subsets is lower than 70% while
the lowest for this species is 33% in subset 2 and subset 4. For the two A. ilicifolius
groups, the accuracy exceeds 80% in the training phase in all subsets except in
subset 1. The worst is found in subset 1 with about 30-40% accuracy attained. The
testing accuracy fluctuates in different subsets. For A. ilicifolius G1, the highest
testing accuracy of 78% locates in subset 2 while the highest of 94% for A. ilicifolius
G2 locates in subset 4. A. marina yields the highest accuracy of 98% and 100% in the
training and testing phases in subset 4 respectively. The worst accuracy of 63% and
78% under pure radar features subset in training and testing phases respectively.
For the two K. obovata groups, the average producer’s accuracy is relatively higher
for G2 than G1. K. obovata G1 attains high accuracy of over 90% in both training
and testing phases under all subsets except with subset 1. Besides, the training and
testing accuracy for G1 is comparable. K. obovata G2 attains the maximum training
accuracy of 100% under subset 2. Subsets 3 and 4 have accuracy over 90% while the
lowest accuracy of 48% is found in subset 1. K. obovata G2 has a relatively large
discrepancy between training and testing phase. The producer’s accuracy fluctuates
in different subsets for Sonneratia spp. Subsets 3 and 4 attain the maximum
accuracy. Subset 2 produces the lowest accuracy of 57% in the training phase,
however, it achieves the maximum accuracy in testing.
In terms of user’s accuracy, A. corniculatum attains relatively satisfactory results
with over 90% accuracy in both training and testing phases for all subsets except
under subset 2. Subset 2 has the maximum training accuracy, but the testing
accuracy is only 17%. For the two A. ilicifolius groups, the accuracy for A. ilicifolius
G2 is comparatively more stable under the subsets. For A. ilicifolius G1, two subsets,
2 and 4 produces maximum accuracy in training phase while the minimum accuracy
of 44% is found in subset 1. The testing accuracy follows the same trend with the
316
highest (over 75%) found in subset 2 and 4 while the lowest (33% and 50%) in
subsets 1 and 3 respectively. A. ilicifolius G2 attains the highest accuracy of 94%
during the training phase in subset 2. This is followed by subsets 4, 3 and 1. The
lowest accuracy of 70% in the training phase is found in subset 1. In the testing
phase, accuracy produces from subset 4 (89%) is relatively higher than that of
subset 2 (77%) and the lowest (63%) is found in subset 1. A. marina attains the
highest accuracy in subset 4 in both training (92%) and testing (96%) phases. The
lowest accuracy of 66% is found in subset 1 while the other two subsets have
accuracy around 77-96%. For the two K. obovata groups, the highest accuracy
attained for K. obovata G1 is 96% and 90% under subset 4 and 3 respectively while
the lowest is 55% under subset 1. Subsets 2 and 4 have the highest testing accuracy
of over 94% while the lowest (60%) found under subset 1. K. obovata G2 has the
highest accuracy of 96% and 100% under subset 2 while the lowest accuracy of 55%
and 67% is found under subset 1 during the training and testing phases respectively.
K. obovata G1 has shown comparatively better results than G2 in subsets 3 and 4
with around 80-90% accuracy achieved. Sonneratia spp. has 100% training accuracy
in all subsets. Maximum accuracy is achieved under all subsets for Sonneratia spp.
during the training phase. Except subset 1, all other subsets have the maximum
accuracy Sonneratia spp. in the testing phase.
Table 4.25. The confusion matrix after support vector machines (SVM) classification
using feature combination of subset 1
Training dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
10
0
1
0
0
0
0
11
0
7
1
2
5
0
1
16
1
0
14
1
2
2
0
20
0
7
1
29
7
0
0
44
1
8
10
14
52
10
0
95
0
0
5
0
4
11
0
20
0
0
0
0
0
0
6
6
12
22
32
46
70
23
7
212
90.9
43.8
70.0
65.9
54.7
55.0
100
Producer's
accuracy (%)
83.3
31.8
43.8
63.0
74.3
47.8
85.7
60.8
317
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
2
0
0
0
0
0
0
2
0
2
2
1
1
0
0
6
1
0
5
0
2
0
0
8
0
3
4
18
5
0
1
31
0
4
5
4
24
3
0
40
0
0
1
0
1
4
0
6
0
0
0
0
1
0
3
4
3
9
17
23
34
7
4
97
100
33.3
62.5
58.1
60.0
66.7
75.0
Producer's
accuracy (%)
66.7
22.2
29.4
78.3
70.6
57.1
75.0
59.8
Table 4.26. The confusion matrix after support vector machines (SVM) classification
using feature combination of subset 2
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
8
0
0
0
0
0
0
8
0
19
0
0
0
0
0
19
2
0
29
0
0
0
0
31
1
2
1
44
1
0
0
49
1
1
1
2
69
0
3
77
0
0
1
0
0
23
0
24
0
0
0
0
0
0
4
4
12
22
32
46
70
23
7
212
100
100
93.5
89.8
89.6
95.8
100
Producer's
accuracy (%)
66.7
86.4
90.6
95.7
98.6
100
57.1
92.5
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
1
0
4
0
0
1
0
6
1
7
0
0
1
0
0
9
1
2
13
0
0
1
0
17
0
0
0
21
1
0
0
22
0
0
0
2
32
0
0
34
0
0
0
0
0
5
0
5
0
0
0
0
0
0
4
4
3
9
17
23
34
7
4
97
16.7
77.8
76.5
95.5
94.1
100
100
Producer's
accuracy (%)
33.3
77.8
76.5
91.3
94.1
71.4
100
85.6
318
Table 4.27. The confusion matrix after support vector machines (SVM) classification
using feature combination of subset 3
Training dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
10
0
1
0
0
0
0
11
0
18
1
1
0
0
0
20
2
0
27
1
1
0
0
31
0
4
0
39
3
0
0
46
0
0
1
5
64
1
0
71
0
0
2
0
2
22
0
26
0
0
0
0
0
0
7
7
12
22
32
46
70
23
7
212
90.9
90.0
87.1
84.8
90.1
84.6
100
Producer's
accuracy (%)
83.3
81.8
84.4
84.8
91.4
95.7
100
88.2
Testing dataset
Observed
Predicted
1
2
3
4
5
6
7
Total
User's
accuracy (%)
1
2
3
4
5
6
7
Total
2
0
0
0
0
0
0
2
0
5
3
0
2
0
0
10
1
1
8
0
0
0
0
10
0
1
4
20
1
0
0
26
0
1
0
3
30
1
1
36
0
1
2
0
1
6
0
10
0
0
0
0
0
0
3
3
3
9
17
23
34
7
4
97
100
50.0
80.0
76.9
83.3
60.0
100
Producer's
accuracy (%)
66.7
55.6
47.1
87.0
88.2
85.7
75.0
76.3
Table 4.28. The confusion matrix after support vector machines (SVM) classification
using feature combination of subset 4
Training dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
10
0
0
0
0
0
0
10
0
20
0
0
0
0
0
20
2
0
28
0
1
0
0
31
0
2
0
45
2
0
0
49
0
0
1
1
67
1
0
70
0
0
3
0
0
22
0
25
0
0
0
0
0
0
7
7
12
22
32
46
70
23
7
212
100
100
90.3
91.8
95.7
88.0
100
Producer's
accuracy (%)
83.3
90.9
87.5
97.8
95.7
95.7
100
93.9
319
Testing dataset
Observed
Predicted
Total
User's
accuracy (%)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
Total
1
0
0
0
0
0
0
1
0
6
0
0
1
1
0
8
1
1
16
0
0
0
0
18
0
1
0
23
0
0
0
24
1
0
0
0
33
1
0
35
0
1
1
0
0
5
0
7
0
0
0
0
0
0
4
4
3
9
17
23
34
7
4
97
100
75.0
88.9
95.8
94.3
71.4
100
Producer's
accuracy (%)
33.3
66.7
94.1
100
97.1
71.4
100
90.7
320
4.4.6. Algorithm Comparison
The overall classification accuracy and kappa coefficient (  ) using different
classifiers under different subsets are compared in Figure 4.47. Generally, the
overall accuracy and kappa coefficient follows the same pattern.
DT has the highest classification accuracy of 99.1% (  = 0.99) achieved under subset
4 during training. However, the independent testing accuracy drops to 82.5% (  =
0.77) which is the lowest among the counterparts using the same feature subset.
The large discrepancy implies the model was likely overtrained and the
generalization capability of the model was curtailed. ANN is the second best
classifier under the same subset with overall training and testing accuracy of 95.3%
(  = 0.94) and 95.9% (  = 0.95) respectively. This is followed by SVM classifier with
overall training and testing accuracy of 93.9% (  = 0.92) and 90.7% (  = 0.88)
respectively. The relatively small accuracy difference between the training and
testing stages suggests the model is likely more robust than DT. ML has the lowest
overall training accuracy of 92.5% (  = 0.9) under the same subset; however, the
overall testing accuracy of 83.5% (  = 0.78) is higher than DT.
The sole dependency on radar feature (subset 1) yields the lowest overall accuracy
regardless of classifier. Among the algorithms, ANN yields the highest overall
training and testing accuracy of 74.1% (  = 0.67) and 72.2% (  = 0.64) respectively.
DT has the second best overall training accuracy of 70.8% (  = 0.63) but at same
time, yields the lowest accuracy of 46.4% (  = 0.29) when evaluated independently
from unseen dataset. The low kappa value suggests poor agreement between the
observed and predicted classes. SVM and ML presents close overall training (testing)
accuracy of 60.8% (59.8%) and 60.4% (58.8%) respectively.
The remaining two subsets (subset 2 and 4) have equal number of features. The
overall accuracy from input of hyperspectral features (subset 2) and combined input
of hyperspectral and radar features (subset 3) fluctuate across classifiers. Subset 2
has better performance under ML, DT and SVM classifiers while subset 3 has lower
error rate under ANN classifier. Under subset 2, the highest overall training
accuracy of 93.4% (  = 0.92) is achieved when DT classifier is applied. However, the
321
testing accuracy has been lowered to 82.5% (  = 0.77). SVM is the classifier with the
second best overall training accuracy of 92.5% (  = 0.90) while the overall accuracy
from the independent dataset is 85.6% (  = 0.81). This is followed by ANN having
overall training and testing accuracy of 85.8% (  = 0.82) and 89.7% (  = 0.87)
respectively. ML yields the lowest overall training and testing accuracy of 90.6%
(  = 0.88) and 79.4% (  = 0.73) respectively.
ANN classifier yields the highest overall training and testing accuracy of 88.7% (  =
0.86) and 91.8% (  = 0.89) respectively under subset 3. This is followed by SVM and
MLC with training (testing) accuracy of 88.2% (76.3%) and 85.4% (79.4%)
respectively. DT has same training accuracy as ML but yield the lowest accuracy of
73.2% (  = 0.65) under the same subset.
322
(a)
100%
DT, 99.1%
NN, 95.9%
NN, 95.3%
90%
NN, 91.8%
NN, 89.7%
NN, 88.7%
MLC, 85.4%
MLC, 83.5%
80%
SVM, 93.9%
SVM, 92.5%
SVM, 90.7%
DT, 93.4%
MLC, 92.5%
MLC, 90.6%
SVM, 88.2%
NN, 85.8%
DT, 85.4%
SVM, 85.6%
DT, 82.5% DT, 82.5%
MLC, 79.4% MLC, 79.4%
Accuracy
SVM, 76.3%
NN, 74.1%
NN, 72.2%
DT, 73.2%
DT, 70.8%
70%
60%
SVM, 60.8%
SVM, 59.8%
MLC, 60.4%
MLC, 58.8%
50%
DT, 46.4%
40%
MLC
Subset 1 Training
DT
Subset 1 Testing
Subset 2 Training
Classification Algorithm
Subset 2 Testing
Subset 3 Training
ANN
Subset 3 Testing
SVM
Subset 4 Training
Subset 4 Testing
323
(b)
1
DT, 0.99
NN, 0.95
NN, 0.94
0.9
0.8
DT, 0.92
MLC, 0.9
MLC, 0.88
MLC, 0.82
DT, 0.82
MLC, 0.78
MLC, 0.73
SVM, 0.92
SVM, 0.9
SVM, 0.88
NN, 0.89
NN, 0.87
NN, 0.86
DT, 0.77
SVM, 0.85
NN, 0.82
SVM, 0.81
DT, 0.77
MLC, 0.73
0.7
SVM, 0.69
Kappa Value
NN, 0.67
DT, 0.65
DT, 0.63
NN, 0.64
0.6
0.5
SVM, 0.49
MLC, 0.5
SVM, 0.47
MLC, 0.46
0.4
0.3
DT, 0.29
0.2
MLC
Subset 1 Training
DT
Subset 1 Testing
Subset 2 Training
Classification Algorithm
Subset 2 Testing
Subset 3 Training
ANN
Subset 3 Testing
SVM
Subset 4 Training
Subset 4 Testing
Figure 4.47. Comparison of (a) classification accuracy and (b) kappa coefficient using different classification algorithms under different subsets
324
4.5. Discussion and Implication
4.5.1. Feature Selection
In terms of computational efficiency, search algorithm SFS combined with KNN
classifier as wrapper criterion is the most efficient while SFFS wrapped with SVM
classifier takes the longest time to evaluate the features. SFS is regarded as greedy
search approach under which discarded features cannot be re-selected again. When
coupled with the simplest KNN classifier, the selection process is exceptionally rapid
in sacrifice of accuracy as indicated by the lowest mean training and testing
accuracy among the three search algorithms under KNN classifier in all subset sizes
in Figure 4.19 and 4.20. Under the same KNN wrapper evaluation criterion, feature
subsets from OS have the best mean training and testing accuracy in all subset sizes.
OS allows user-defined maximum oscillation cycle to control the completeness of
search. Unlike the entirely automatic backtracking mechanism in SFFS, the search
efficiency can be adjusted in OS. SFFS is renowned for producing close-to-optimal
solution in many pattern recognition studies. In this study, the resultant feature
subsets from OS have demonstrated reasonable search duration and better
accuracy when coupled with KNN classifier.
When feature selection algorithm is wrapped with SVM classifier, the search
duration increases substantially due to the optimization process of two
hyperparameters of SVM classifier prior to the search every time. Although the
computation time is much longer, the mean classification accuracy of the subsets is
comparatively worse than that when wrapped with KNN classifier under all subsets.
The search algorithms have mean accuracy fluctuation across different subset sizes.
The OS search algorithm has the worst performance in all subset sizes except in
subset size of 5. Besides, 98 out of 480 features subsets (total number of trials in all
subset sizes) under SVM-wrapped search are discarded from further consideration
due to poor accuracy performance when compared with only 6 features subsets
removed under KNN-wrapped search. Many of the poorly-performed subsets (41)
are found under OS and this explains the relatively poor results from OS. On one
hand, search algorithms wrapped with SVM generates feature subsets with the
325
worst training and testing classification accuracy while individual feature subsets
with the best testing accuracy are also generated under all subset sizes. OS search
wrapped with SVM classifier generates individual feature subset having the highest
classification accuracy under subset sizes of 5 and 20. The large contrast of accuracy
suggests SVM classifier produce feature subsets of highly varied quality. This is
particularly valid under OS algorithm with the accuracy of relatively large number of
feature subsets below mean accuracy when compared with SFS and SFFS search
algorithm. This is due to the limit of swing in the oscillating search process. The
accuracy of search can be improved by increasing the maximum oscillation cycle at
the expense of computation time. Comparatively, feature subsets produce under
KNN classifier is more stable in terms of accuracy.
Generally, the mean training and testing accuracy of different search algorithms
wrapped with KNN classifier improves with increase subset size. The only exception
is the marginal decrease of 0.001% in both training and testing accuracy in OS
search when feature size expands from 10 to 15. Large improvement in accuracy is
found when subset size expands from 5 to 10 and the accuracy improvement slows
down afterwards. Same accuracy improvement is observed when subset size
increases from 5 to 10 under SVM classifier using various search algorithms.
However, as the subset sizes increases, no obvious trend is observed as different
search algorithms have diverse accuracy performance under different subset sizes.
With less than 1% increase in accuracy from 15 to 20 features under KNN classifier,
it suggested that the ‘optimal’ number of features should fall between 10 and 15.
The low consistency indices (below 0.5) reveal little co-occurrence of features in the
twenty trials for all algorithm-classifier-size combinations. Theoretically, the low
consistency indicates selected features are unstable across different trials in a given
combination regardless of the classification accuracy attained by the feature subset.
After removal of feature subsets below defined accuracy threshold, the selected
features are further analyzed through frequency analysis. The first four criteria
based on which the final set of features is generated reveal the best features are
commonly selected regardless of search algorithms and wrapped classifier. The
commonly selected features reveal the specific wavelengths at which are significant
326
for species discrimination. For instance, under criterion I shown in Figure 4.21, high
frequency features 23, 72, 94 and 163 representing by different colors along the
bars indicated these features are commonly selected under different searchclassifier combinations.
Apart from features of maxima, features cluster around a specific narrow
wavelength region can be of high significance as well. In Figure 4.21, for instance,
the narrow band of green peak and the transition to red band do not have the
highest frequency; however, relatively high frequency of occurrence across the
regions signifies the potential of green bands for species classification. Instead of
pinpointing to a particular feature of narrowband, it is the cluster of high frequency
that shows some light on feature importance. As hyperspectral data/ features are
continuous and features in adjacent wavelengths are highly-correlated, selection of
green band in a specific wavelength or the one locates adjacently may have minimal
impact on the result. However, this fact does not reflect by the consistency indices
as they count the frequency of occurrence of particular feature in the subsets
concerned. The indices alone are not able to capture the distinctive characteristics
of hyperspectral features.
As feature size increases from 5 to 20 (criterion I to criterion IV), it is clear that the
selected features gradually spread across different wavelengths and multi-temporal
radar data until almost all features have been selected at least once under subset
size of 20. The wide spread of selected features in the plots explains the low
consistency of selected features in different trials. However, the features with the
highest frequency as well as those clustered in particular wavelength regions
remain similar through different subset sizes. In other words, the co-occurrence of
high frequency features under various subset sizes signifies that these co-occurred
features are significant in distinguishing the mangrove spectral classes.
Figure 4.48 compares the spectral features selected by this study with other studies.
Out of the best fifteen features in Table 4.8, eleven are spectral features.
327
2500
2400
2300
2200
2100
2000
1900
1800
Wavelength (nm)
1700
1600
1500
1400
1300
1200
1100
1000
900
800
700
600
500
400
Current
Study
Li et al. (2011)
Mainly agricultural crops
Song et al. (2011)
Agricultural
crops
Vyas et al. (2011)
Tropical
vegetations
Wang et al. (2007)
Salt-marsh
vegetations
Figure 4.48. Comparison of spectral bands selected by this study and other studies
Belluco et al. (2006)
Salt-marsh
vegetations
Becker et al. (2005)
Coastal wetland
vegetations
Schmidt &
Skidmore (2003)
Coastal wetland
vegetations
328
Although the types of vegetation were different in Figure 4.48, similar bands were
selected between 570 – 780nm. Specifically, spectral regions located at or near
580nm, 702nm, 774nm were commonly selected to discriminate vegetation
regardless of vegetation types. Comparatively, the SWIR regions did not show much
correspondence among the studies. The selected features were relatively scattered
across the SWIR region. Current study showed the four selected SWIR bands limited
to the first half of the SWIR prior to 1650nm while some studies (Li et al., 2011,
Song et al., 2011, Vyas et al., 2011) have indicated bands in the latter half of SWIR
from 1800nm onwards are important bands for vegetation discrimination.
In the final feature set shown in Table 4.8 and Figure 4.27, the 15 features not only
identify individual narrow wavebands, but also general spectral regions that are in
fact important for species discrimination. Among the most important features, four
bands span across the green peak from 570 to 601nm, two bands locates in the
transition from red to red-edge at 702 – 713nm, two near infrared bands at 764774nm and three shortwave infrared bands at 1276-1316nm and 1629nm. The
green peak, red-edge and near infrared are characteristic vegetation regions, which
are highly related to and affected by the concentration of major leaf pigments and
leaf cellular structure. The green peak is characterized by relatively low absorption/
high reflectance of green light when compared with the high absorption of violetblue and red light by chlorophyll pigments for photosynthesis in the visible region.
According to reflectance simulation using PROSPECT leaf model from other studies,
the magnitude of green peak is determined by the concentration of biochemical
constituents such as chlorophyll a and b, dry matter and structural parameters
(Zarco-Tejada and Sepulcre-Cantó, 2007). Wavelengths between 702 and 713nm
characterize the transition from strong chlorophyll absorption in red to strong
scattering of near infrared. This well-known red-edge position has been known to
be particularly sensitive to leaf chlorophyll concentration (Miller et al., 1990,
Dawson and Curran, 1998). The lateral shift of the red-edge indicates the
chlorophyll concentration, i.e. increase in chlorophyll content border the absorption
in red and shift the red-edge to longer wavelength while decrease in chlorophyll
concentration moves the red-edge to short wavelength. Leaf reflection, particularly
340
in near infrared is primarily controlled by leaf structure which refers to the
arrangement of cells, airspace and water in the spongy mesophyll layer. The
selected near infrared features at 764-774nm locates next to the shoulder of rededge. Usually, the number of layers ( N ) is used to indicate the separation of
absorption tissues and airspaces in which substantial reflection takes place. From
the model PROSPECT, as N increases, the reflectance increases though it saturates
rapidly when N goes beyond about 8 (Jacquemoud and Baret, 1990). Although leaf
structure is important factor to reflection/ transmission, significant fraction of
infrared reflection takes place in the upper cuticular surface (Jones and Vaughan,
2010). Comparatively low reflectance in the shortwave infrared region compared
with the near infrared region is mainly due to the presence of various biochemical
constitutes such as lignin, cellulose, starch, proteins, nitrogen and water acting as
absorption agents in the wavelengths (Kumar et al., 2001). The mean spectral
reflectance of different mangrove species classes shown in Figure 4.6a has
correspondingly indicated clear separability in the near infrared and shortwave
infrared while the separability in green and red-edge are not as obvious as the
infrared regions.
Referring to the radar backscatter features, the non-filtered backscatter features
are negligibly selected in a few trials under SVM classifier as indicated in Figures
4.13 – 4.18, which results in very low frequency count in every subset sizes.
However, after speckle reduction using 3x3 Enhanced Lee filter, four filtered radar
backscatter features acquired in different seasons (spring, summer and winter) are
selected in the final feature set. This proves the effectiveness of Lee filter in
suppressing the salt-and-pepper noise and enhancing the information inherent in
radar features. Besides, the feature selection algorithms are more sensitive to
backscatter coefficient in linear scale than that expressed in decibel scale.
341
4.5.2. Mangrove Classification
Among the classifiers, DT and ANN have comparatively better performance in
overall accuracy. The boosting option in the C5.0 with the ability to enhance
accuracy through penalizing the misclassified records does improve the
classification accuracy as shown in Table 4.29 below. Except subset 1, all other
subsets demonstrate better accuracy after boosting.
Table 4.29. Comparison of classification accuracy of DT models with and without
boosting option
Boosting
Subset 1
Subset 2
Subset 3
Subset 4
Training (with boosting)
70.8%
93.4%
85.4%
99.1%
Testing (with boosting)
46.4%
82.5%
73.2%
82.5%
Training (without boosting)
79.7%
87.8%
84.0%
93.9%
Testing (without boosting)
45.4%
81.4%
69.1%
75.3%
However, the generality performance of DT is much worse than that of ANN.
Results from DT regardless of feature subset shows large discrepancy in training and
testing accuracy. The largest difference of 24.4% is found in subset 1 followed by
16.6% difference in subset 4, 12.2% in subset 3 and 10.9% in subset 2. The same
happens to the results when boosting option is not applied. The tree models are
likely suffered from the problem of overtraining though generality measure was
applied in the learning process. Comparatively, ANN with two hidden layers
presents more robust models with relatively high accuracy in the training and
testing phases. Given appropriate learning rate, momentum and sufficient learning
time, the network was optimized with stability. However, details of the training
process and the model cannot be retrieved and the relationship between the input
training data and output classes is a blackbox which is one of the main criticisms of
ANN. SVM classifier performs better than ML classifier in terms of overall training
accuracy though the difference is relatively small (2-3%). However, relatively large
accuracy discrepancy was found in testing phase for ML classifier. The SVM models
are more robust except in subset 3 in which both hyperspectral and radar features
are combined.
342
In terms of feature subsets, interesting observations can be drawn from the
response of classifier to feature size and individual features. It is clear that subset 4
with nine features give the best accuracy. Use of radar features alone results in the
worst accuracy in all classifiers and this is supported by other studies, for instance,
Leckie (1990) The sole input of spectral features in subset 2 as well as mixed
spectral and backscatter features in subset 3 have resultant accuracy fall between
the best and the worst subsets though there is fluctuation among classifiers.
However, it is very obvious that better classification accuracy is resulted when
spectral bands were used alone as opposed to only radar bands. ANN classifier is
more adaptive to the radar features as it produces higher accuracy when radar
features are input (in subset 1 and 3). For other classifiers, the performance of input
of purely hyperspectral features is comparatively better.
When compared the mean reflectance and backscatter of different species of
selected features in Figure 4.6, it reveals the reason for the complementary input of
both spectral and radar features. Under the five selected spectral wavelengths,
similar mean reflectance is found among some of the mangrove species, which
cause confusion during species classification. For instances, A. corniculatum, A.
ilicifolius G1 and G2 have very close mean reflectance in 570nm. Similar mean
reflectance is also found for A. ilicifolius G2, K. obovata G2 and A. corniculatum in
1276nm; A. ilicifolius G2, K. obovata G2 in 1629nm. However, the selected radar
features show significant separability for these species indicated by the large
difference in backscatter response. The integration of spectral and radar features
reduces the confusion and therefore enhances the classification accuracy of these
species.
The relative importance of individual features obtained in DT and ANN models
under different subsets are plotted in Figure 4.49 within which a – d are from DT
classifier while e – h are from ANN classifier. Feature importance is expressed in xaxis and sum of feature importance in each plot equals to unity.
343
a – subset 1
b – subset 2
c – subset 3
d – subset 4
e – subset 1
f – subset 2
g – subset 3
h – subset 4
Figure 4.49. Relative feature importance in DT and ANN models
a – d are feature importance from DT classifier while e – h are feature
importance from ANN classifier under different subsets.
344
Although the two classifiers have comparable accuracy performance, the
contribution of features in the training process is diverse in terms of magnitude of
difference and feature rank. Immediate comparison of the plots between the two
classifiers in Figure 4.49 reveals that features under DT classifier have larger
difference in magnitude of importance while feature importance under ANN
classifier is of relatively lesser difference under all subsets. In other words, features
are of more or less the same importance in making prediction in ANN models while
the DT models are more dependent on a few features for the prediction. Apart from
magnitude difference, the rank of importance between the two models under the
same feature subsets is different. For instances, when compared the relative
feature importance for DT and ANN models under subset 3 in Figure 4.49c and g
respectively, the DT regards the radar backscatter (feature 161) as the most
significant feature in its prediction model while ANN ranks the feature of the least
importance in the model. The backscatter feature 162 ranks the second best in ANN
model is of negligible importance in the DT model. As the feature importance is
computed based on the testing partition, the ANN models are therefore more
reliable as they have relatively higher testing accuracy. As shown in Figure 4.47, the
testing accuracy of three subsets (2-4) is even better than the training accuracy.
Feature size is another major determinant of accuracy. Normally, accuracy improves
with number of feature increases. When feature size increases from five in subsets
2 and 3 to nine in subset 4, the magnitude of accuracy improvement varies among
the classifiers. Relatively large improvement is found in DT models and the largest
accuracy improvement of 13.7% is found between subset 3 and 4 from 85.4% to
99.1% in the training phase. In ANN, the largest improvement is found between
subset 2 and 4 with about 10% accuracy increases from 85.8% to 95.3% in the
training phase. Accuracy improvement of 7.1% and 5.7% after inclusion of
additional features (from subset 3 to 4) are found for ML and SVM respectively
which is comparatively lesser than DT and ANN. Better accuracy is also observed in
the testing phase, but the magnitude of improvement is smaller than that found in
the training phase.
345
Apart from the overall classification accuracy, the accuracy of individual species/
mangrove spectral classes is examined through producer’s and user’s accuracy.
Generally, A. corniculatum, A. ilicifolius G1 and Sonneratia spp. have wider variation
across different classifier-subset combinations and the variation is more prominent
in the testing phase. For instances, sonneratia spp. was completely misclassified in
the ML models under all subsets in the testing phase as shown in the confusion
matrix in Tables 4.12 – 4.15. A 100% misclassification error was unexpectedly found
in the training phase when all nine features were used. However, sonneratia spp.
was mapped reasonably well in other classification models. The inaccurate
estimation for sonneratia spp. is therefore classifier-specific. Similar observation is
found for A. ilicifolius G1 in DT model compute using pure radar features (subset 1)
in which the species yields zero producer’s and user’s accuracy in the testing phase.
However, it is observed that classifier in combination with features in subset 1 all
yields relatively low accuracy. A. corniculatum has reasonable accuracy in the
training phase, but enormous errors are found across classifiers in the testing phase.
Complete misclassification appears in ML, DT and ANN classifier under various
subsets. The error is most prominent in DT models in which 100% misclassification
errors are found in all subsets except in subset 4 while the same happens in ML and
ANN models under subset 2. Although SVM performs slightly better, the accuracy in
subset 2 is also relatively low when compared with other subsets. The erroneous
estimation for A. corniculatum can be identified as feature-specific. The sole use of
spectral features (subset 2) cannot satisfy accurate classification for A. corniculatum.
The accuracy of other mangrove spectral classes in different classifier-subset
combinations is comparatively more stable.
The relatively low classification accuracy and unstable estimation for the two
species A. corniculatum and Sonneratia spp. is mainly caused by inadequate
learning and testing samples as well as the locational characteristics of the species.
A total of 15 and 11 samples were collected for A. corniculatum and Sonneratia spp.
respectively for model training process. There are a few reasons for insufficient
samples of the two species. For A. corniculatum, the majority of them occupy the
area as undergrowth and the rest of them distributed in a scattered manner in the
346
study area. Given the coarse resolution of the satellite images, it is hard to locate
homogeneous samples. Besides, homogeneous patch of A. corniculatum tends to
locate along the coastal fringe and it is probable for its spectral signature to be
mixed with that of water. Constrained by the physical characteristics of the area, it
is hard to locate sufficient samples. As for Sonneratia spp., the known
homogeneous plantation in Futain National Nature Reserve across the border was
the only set of training samples selected for the training process. The isolated
distribution of the species in Mai Po makes it hard to select the training sites with
sufficient confidence. Similarly, the coarse resolution has limited the selection of
sufficient training samples. Although the samples are extracted from reliable
sources and also through field surveys, the inadequate amount of sample curtails
the reliability of model learning and building process. Traditional parametric
classifiers such as maximum likelihood is particularly sensitive to the amount of
training samples because they needs sufficient samples to come up with reliable
estimates of class statistics, i.e. mean and variance-covariance matrix and based on
which a robust parametric model is computed. Deficiency of training samples
hinders good estimation of class statistics which therefore degrades the accuracy
and affects the generalization ability of classifier. It is found that the sample
deficiency problem was not only limited to ML, other machine learning algorithms
such as DT and ANN are also suffered though the magnitude is less severe. SVM is
the best classifier in facing limited data problem. In Tables 4.25 – 4.28, the SVM
models have comparatively better accuracy for A. corniculatum in the testing phase.
The close user’s accuracy in the training and testing phases for the majority of
subsets suggest the model is of high generality and robustness. Sonneratia spp.
show similar performance in the SVM models though the accuracy (user’s and
producer’s) are relatively better. Having the overall better accuracy, ANN classifier
has comparable performance for Sonneratia spp. though the accuracy for A.
corniculatum is not as good as SVM classifier.
Visual-interpretation analysis of the classification results is supplemented to the
statistical accuracy analysis to review the reliability of the classifiers in different
subsets. The visual analysis is based on field experiences and review of past
347
literatures on mangrove researches related to the areas. The sole input of radar
features yields the lowest classification in all classifiers and this is also reflected in
the classified map. Sonneratia spp. is over-estimated in all classifiers with the worst
found in DT. Areas such as sites L and K which should be the habitats of A. ilicifolius
was misclassified as other species by all classifiers. Although ANN and SVM models
have similar accuracy about site A, result from ANN is visually more promising as a
homogeneous area of Sonneratia spp. is observed. The multi-temporal yields the
lowest classification accuracy implies that the backscatter do not provide additional
data variance for classification. Normally, the multi-temporal data are able to
capture the seasonal canopy change such as leaf-on, leaf off, flowering, fruiting,
which will alter the backscatter return. As mangroves are evergreen, significant
natural structural change does not occur. The variation of backscatter due to
flowering and fruit may be too small to be detected by radar signal or the spatial
resolution of the image is unable to record the subtle change. This results in
relatively poor accuracy with the input of radar features only.
With the same number of features, the effect of radar combining with spectral
features is revealed by comparing the difference the results between subset 2
(spectral bands only) and 3 (spectral and radar bands). Both subset inputs result in
comparable overall accuracy though use of spectral features alone performs slightly
better than that of spectral-radar feature mix in all algorithms except ANN. The
accuracy of individual species is also comparable though classifiers’ performance in
species with deficiency of sample is relatively better when spectral and radar
features were combined particularly when tested with independent data. An
apparent difference is found in the result of A. corniculatum under ANN classifier in
Tables 4.20 – 4.23 in which pure spectral feature subset yields zero percentage of
user’s and producer’s accuracy while subset 3 has 67% testing samples being
classified correctly. Through visual examination of results from ANN, combination of
spectral and radar features has successfully identified the species in certain
locations that were absence in pure spectral feature subset. For instances, a small
area of A. ilicifolius G1 locating between the homogeneous K. obovata G1 and A.
marina in sites D and E respectively appears in the results of spectral-radar feature
348
mix while it is the result is not distinct in pure spectral feature as shown in Figure
4.50 below.
Spectral features alone
Spectral and radar
features
Figure 4.50. Comparison of ANN classification results using pure spectral features
and combination of spectral and radar features – A. ilicifolius G1
Another interesting area locates in site K where A. ilicifolius G2 is found. The sole
dependence of spectral features has the area identified as A. ilicifolius G1 while the
input of radar features has identified the correct class G2 as shown in Figure 4.51.
Generally, canopy texture is coarser as the age of the stand increases. Therefore, A.
ilicifolius stands along the active growing fringe as well as in site K have a smoother
texture when compared with those located towards the land boundary. The age of
the stand affects radar backscatter by their differences in stand parameters such as
crown closure, height, etc. These parameters are effectively captured by radar and
used to eliminate the confusion between the two groups.
Spectral features alone
Spectral and radar features
Figure 4.51. Comparison of ANN classification results using pure spectral features
and combination of spectral and radar features – A. ilicifolius G2
349
However, combination of spectral and radar features tends to over-estimate A.
corniculatum and Sonneratia spp. In fact, overestimation of the two species is
observed in the majority of classifiers when radar features were used. Figure 4.52
shows the prediction results of ANN classifier in the northern part of the study area.
A. corniculatum should have a scattered distribution similar to the one predicted by
sole use of spectral features. However, estimation from spectral-radar feature mix
results in a large homogeneous distribution of the species as shown within the red
circle. Similarly, overestimation of Sonneratia spp. is apparent within the black circle
in Figure 4.52. The extent of Sonneratia stands across the Shenzhen River and is
over-predicted while those located on the landward side next to the fishpond are
misclassified.
As A. corniculatum is predicted to be coastally located close to the fringe area, the
tidal levels would affect the estimation of the species as compared to the species
locates towards the land boundary. The multi-temporal radar data were captured
under varied tide levels as documented in Table 3.5, which affects the backscatter.
The effect is more prominent in the fringe zone than on the landward side. The tide
controls the submergence of the trunk, branches or even canopy under water.
Spectral features alone
Spectral and radar features
Figure 4.52. Comparison of ANN classification results using pure spectral features
and combination of spectral and radar features – A. corniculatum and Sonneratia
spp.
The combination of the best nine features resultant from feature selection analysis
yields the best overall accuracy in all classifiers as well as the best producer’s and
user’s accuracy in the majority of species. Similar classification results are found in
350
the four classifiers through visual comparison while the largest variation lies in two
species, A. corniculatum and Sonneratia spp. where training samples are insufficient.
The estimation of A. corniculatum is highly varied across different classifiers. ML, DT
and SVM classifiers identify the main clusters of A. corniculatum in the northern
part of the mangrove area as well as some scattered over along the coastal fringe of
the mangrove area. Predictions from ANN suggest a much wider and abundant
distributions of the species. Although results from ANN model have the most
promising accuracy (producer’s and user’s) for the species, over-estimation is likely
occurred. A. corniculatum is hard to be estimated with sufficient confidence
because of a few reasons. First, there is an absence of homogeneously large and
representative samples of the species under the 30m resolution. The distribution of
A. corniculatum is comparatively more scattered than other species, a sufficiently
large pure area to meet the resolution of the satellite image is difficult to be
identified. Therefore, training areas are likely to mix with signature of other species,
e.g. K. obovata or signature of seawater. Besides, the number of training samples is
far below requirement for the computation a robust model.
In fact, Sonneratia spp. is expected to be found in the active-growing fringe areas
especially in the area close to the border area due to the proximity to seed source.
Besides, they tend to distribute in a scattered manner rather than in large clusters.
ML had no pixels classified as Sonneratia spp.; DT, ANN and SVM have Sonneratia
spp. estimated along the land boundary similar to the estimate distribution in
subset 3, which was misclassified. At the same time, the three models have
predicted small and scattered distribution in areas close to Shenzhen River, which is
reasonably valid.
4.6. Summary
To summarize, this chapter shows the feature selection results in terms of various
combinations of search algorithm and wrapped-based classifier. Feature selected
using oscillating search (OS) coupled with K-nearest neighbor (KNN) produces better
criterion value, i.e. classification accuracy with reasonable searching time. Six
351
criteria formulated based on the classification accuracy extracts the best final
feature set through frequency analysis. The final feature set consists of nine
features, with five from hyperspectral data including green at 570nm, Red-edge at
713nm, near infrared at 774nm and two shortwave infrared at 1276 and 1629nm
and four are from multi-temporal radar time series captured on 13 February, 19
March, 06 August and 19 November. Features in the final set are recombined to
form four subsets, namely (1) radar features alone, (2) spectral feature alone, (3)
the first five best spectral-radar feature mix and (4) all nine spectral-radar feature
mix and acted as input for species classification.
Mangrove species-based classification is conducted based on the four best feature
subsets using different classifiers including maximum likelihood (ML), decision tree
C5.0 (DT), artificial neural network (ANN) and support vector machines (SVM). In
terms of classifier, ANN performs better due to consistently high overall accuracy in
different subsets and robust model indicated by compatible training and testing
accuracy. This is followed by SVM, DT and ML. DT and ML classifier have a sign of
overtraining as there is large discrepancy between training and testing accuracy.
The largest discrepancy of 24.4% is found in DT when radar features are the sole
input in the classification. The accuracy of individual species is expressed in
producer’s and user’s accuracy. Two species, A. corniculatum and Sonneratia spp.
shows large accuracy oscillation in the classifiers due to deficiency of training
samples. Serious over-estimation of the two species is sometimes observed. In
terms of feature subset, with radar data alone, it yields the lowest classification
accuracy (60 – 73%) while the use of all nine best features results in the highest
classification accuracy (90 – 96%). With same number of features, sole input of
spectral features performs slightly better than the spectral-radar feature mix in all
classifiers except in ANN. Visual interpretation of different classification results
suggest spectral and radar features are complements that can enhance mangrove
species separability due to the capture of different characteristics of the canopy
layer.
352
CHAPTER 5
RESULTS AND DISCUSSION (II) - LEAF AREA
INDEX MODELING
5.1. Introduction
This chapter presents the results of mangrove leaf area index (LAI) after regression
modeling. Firstly, the characteristics of field measured LAI is explored. This is
followed by examining the relationship between LAI and various predictor variables
extracted from remotely-sensed data including the hyperspectral bands, vegetation
indices, radar backscatter and textures extracted from radar backscatter data.
Through data exploration process, two basic assumptions of regression analysis
namely, normality and linearity are inspected and possible outliners are detected
before regression analysis. After that, LAI estimation based on chosen VIs using
simple linear regression analysis is presented. This is followed by examining the
results of LAI estimation from stepwise multiple regression models after
incorporating radar parameters. The discussion and implication session provides a
more in-depth examination of the results. This chapter ends by summarizing the
major findings of LAI estimation.
5.2. Data Exploration
The results presented in this session cover dependent variable (LAI) and
independent variables (hyperspectral, VI and radar parameters) exploration. The
results of normality and linearity check between the independent and dependent
variables as well as the existence of possible outliners followed the presentation.
5.2.1. Dependent Variable: Field measured LAI
After the analysis of hemispherical photographs, LAI of the 95 sample sites were
computed using three methods including Bonhomme and Cartier’s, LAI2000 and
LAI2000G. The frequency distribution of the 95 samples is shown in Figure 5.1.
353
Descriptive statistics including minimum, maximum, mean and standard deviation
of the 95 sample sites are shown in Table 5.1. The Bonhom and Charter’s method,
LAI(Bon) has the relatively higher LAI compared with the other two methods with
mean LAI of 2.27. This is followed by LAI(2000) methods and LAI2000 generalized
method, LAI(2000G) has relatively low LAI. The standard deviation of LAI calculated
using the three methods ranges between 0.42 and 0.49.
As observed in Table 5.1, LAI computed by Bonhomme and Chartie’s method has
the largest range of field LAI (2.25) followed by LAI2000 method (1.73) and LAI2000
generalized method (1.65). Besides, Bonhomme and Chartier’s method also yields
the highest maximum measured LAI of 3.56 and LAI2000 generalized method has
the lowest maximum computed LAI of 2.91.
Figure 5.1. Frequency plots of field measured LAI computed using different methods
354
Table 5.1. The descriptive statistics of leaf area index (LAI) computed using
Bonhomme and Chartier’s, LAI2000 and LAI2000 generalized methods
Minimum
Maximum
Mean
Standard Deviation
N
LAI(Bon)
1.31
3.56
2.27
0.49
95
LAI(2000)
1.27
3.00
2.15
0.45
95
LAI(2000G)
1.26
2.91
2.08
0.42
95
5.2.2. Independent Variables: Vegetation Index and texture measure
The first order statistics of seven vegetation indices (VIs) and radar parameters
computed based on the 95 LAI sample sites are summarized in Table 5.2.
Table 5.2. The descriptive statistics of VI and radar parameters
Abbrev.
Min.
Max.
Mean
Std.
Deviation
Normalized Difference
Vegetation Index
NDVI
0.77
0.94
0.83
0.03
Renormalized Difference
Vegetation Index
RDVI
0.48
0.63
0.54
0.03
Soil-Adjusted Vegetation
Index
SAVI
0.50
0.67
0.57
0.03
MSAVI
0.50
0.77
0.60
0.05
TVI
17.66
25.63
21.28
1.98
Modified Chlorophyll
Absorption Ratio Index 1
MCARI_1
0.42
0.62
0.51
0.05
Modified Chlorophyll
Absorption Ratio Index 2
MCARI_2
0.46
0.78
0.57
0.05
RBS
-16.25
-4.46
-8.44
2.13
Vegetation Index
Modified Soil-Adjusted
Vegetation Index
Triangular Vegetation Index
Raw Backscatter
355
Filtered Backscatter
FBS
-9.97
-5.43
-7.99
0.85
GLCM-Homogeneity
HOMO
0.12
0.66
0.42
0.11
CT
1.05
59.00
7.31
9.71
GLCM-Dissimilarity
DSM
0.75
6.10
1.88
0.91
GLCM-Entropy
ENT
2.56
3.58
3.18
0.26
GLCM-Angular Second
Moment
ASM
0.03
0.09
0.05
0.02
GLCM-Mean
MN
47.57
72.65
57.90
5.07
GLCM-Standard Deviation
SD
1.08
10.91
3.15
2.12
CORR
0.02
0.95
0.63
0.17
GLCM-Contrast
GLCM-Correlation
5.2.3. Hyperspectral bands and LAI
The coefficients of determination (r2) showing the relationship between the
hyperspectral bands and LAI computed using different methods is plotted in Figure
5.2. Spectral bands in near infrared and first portion of shortwave infrared have
relatively higher association with LAI. According to r2, the largest linear association
of LAI is found in bands locating in the red-edge shoulder and the near infrared
plateau at 723-916nm. The highest r2 of 0.644 locates at 825nm associated with LAI
computed from LAI2000 method. Another spectral region of comparatively high
association with LAI locates in the shortwave infrared at 993-1104nm. The r2 (0.604)
peaks at 1005nm with LAI2000. Spectral regions in the visible and other portion of
shortwave infrared have low association with LAI. The difference in r2 between LAI
computed from different methods is most apparent in the near infrared and
shortwave-infrared spanning from 702nm to 1347nm. Negligible difference is found
356
in other spectral regions. Besides, the difference of strength of association with LAI
between LAI2000 and LAI2000G methods is small.
357
428
449
469
489
510
530
550
571
591
611
632
652
672
693
713
733
754
774
795
815
835
856
876
896
917
964
984
1004
1024
1044
1065
1085
1105
1165
1186
1206
1226
1246
1266
1286
1307
1327
1347
1498
1519
1539
1559
1579
1599
1620
1640
1660
1680
1700
1720
1741
1761
1781
1983
2003
2023
2043
2063
2083
2104
2124
2144
2164
2184
2205
2225
2245
2265
2285
2305
2326
2346
2366
2386
Coefficients of determination (r2)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Wavelength (nm)
LAI(Bonhom)Log
LAI(2000)Log
LAI(2000G)Log
Figure 5.2. Coefficients of determination (r2) between 158 hyperspectral bands and LAI computed using different methods; solid and dashed
lines represent LAI computed using log-average and linear method respectively
358
5.2.4. Normality testing
The normality of each (dependent and independent) variable was tested using the
normality plots and Kolmogorov-Smirnov statistics. The normality plots are shown
in Figure 5.3 while the significant levels after Kolmogorov-Smirnov test for the
dependent and independent variables were shown in Table 5.3.
In Figure 5.3, a normally-distributed variable should have data distribution follow
the straight line in the Q-Q plot closely. The three LAI and various VIs fit the line
quite closely though extremely large values in the plots of some independent
variables would likely cause skewing. Extreme values are more obvious in plots of
VIs including NDVI, RDVI, SAVI, MSAVI and MCARI 2. In terms of parameters
extracted from radar data, textural variables including contrast (CT), dissimilarity
(DSM) and standard deviation (SD) show obvious departure from solid line in the QQ plots, which suggest they suffered from normality problem.
Apart from normality plots, the Kolmogorov-Smirnov (K-S) significant test was used
to indicate the normality of variables. In Table 5.4, all three LAIs have level of
significance larger than 0.01 which rejects the null hypothesis and suggests that the
actual distribution is equal to the expected normal distribution. Independent VI
variables are all found to be normal as the level of significance is above 0.01.
However, six independent variables extracted from radar data including filtered
backscatter (FBS), contrast (CT), dissimilarity (DSM), mean (MN), standard deviation
(SD) and correlation (CORR) have level of significance less than 0.01, which suggests
that they are not normally distributed. Besides, some large values are observed in
the Q-Q plot, which indicate possible outliners may exist. The results of outliner
detection will be discussed in Section 5.2.6.
359
LAI(Bon)
LAI(2000)
LAI(2000G)
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI 1
MCARI 2
360
FBS
HOMO
CT
DSM
MN
SD
ENT
ASM
CORR
Figure 5.3. The normality Q-Q Plot for
the independent (LAI) and dependent
(VIs) variables
361
Table 5.3. The level of significance after Kolmogorov-Smirnov test
Dependent
variables
Independent
variables –
Vegetation Index
Independent
variables –
Radar parameters
Variables
Significant level of
K-S Test
Transformation
LAI(Bon)
0.2
-
LAI(2000)
0.011
-
LAI(2000G)
0.015
-
NDVI
0.2
-
RDVI
0.042
-
SAVI
0.011
-
MSAVI
0.088
-
TVI
0.024
-
MCARI_1
0.014
-
MCARI_2
0.149
-
RBS
0.2
-
FBS
0.003
Caution added
HOMO
0.2
-
CT
0.0001
LOG (0.138)
DSM
0.0001
LOG (0.2)
MN
0.004
Caution added
SD
0.0001
Caution added
ENT
0.011
-
ASM
0.012
-
CORR
0.0001
Caution added
Logarithmic and square root transformations were applied to the six radar
parameters. Only two textural parameters – contrast and dissimilarity becomes
normally distributed after logarithmic transformation with level of significance
equals to 0.138 and 0.2 respectively. The normality Q-Q plots for the two
parameters are shown in Figure 5.4. No transformation can successfully adjust the
non-normal distribution of the remaining four parameters and caution is added
during interpretation of relationship if these variables are selected in the regression
modeling.
362
T-CT
T-DSM
Figure 5.4. The normality Q-Q Plot for the two radar parameters – contrast and
dissimilarity after logarithmic transformation
5.2.5. Linearity testing
The linearity between dependent variable and individual independent variable was
evaluated by correlation analysis. Table 5.4 shows the result of correlation analysis
between LAI and VIs/ radar parameters. Flagged and bolded coefficients are
significant either at 0.05 or 0.01 level. In terms of VIs, it was evident that a linear
relationship between each of the three dependent variables (LAI) and the
independent variables (VIs) existed as indicated by the significance of the
correlation coefficient of either p<0.05 or p<0.01. The null hypothesis of no
relationship between the variables was rejected and all the variables satisfy the
linearity assumption. Besides, strong correlation is found between LAI and various
VIs as indicated by the high correlation coefficients. Except with NDVI, the
correlation coefficients between LAI and other VIs are all over 0.66. The strongest
correlation is found between LAI and TVI with correlation around 0.8. Comparing
with individual band in Figure 5.2, VIs generally has better association with LAI.
The result of correlation analysis between LAI and independent variables extracted
from radar image (backscatter and GLCM-derived textures). Flagged and bolded
coefficients are significant either at 0.05 or 0.01 level. For the three LAI dependent
variables, the evidence of linearity were found with filtered backscatter (FBS) and
textural variables derived from GLCM including homogeneity (HOMO), mean (MN),
entropy (ENT) and angular second moment (ASM) as indicated by the statistical
363
significance of correlation coefficient with p<0.05. The null hypothesis of no
relationship between the variables was rejected. No linearity was found for the
other five radar-derived independent variables including raw backscatter (RBS),
transformed contrast (T-CT), transformed dissimilarity (T-DSM), standard deviation
(SD) and correlation (CORR) with the LAI as signified by statistical insignificance of
correlation coefficient with p>0.05 which failed to reject the null hypothesis of no
relationship between the variables. Caution for violation of assumption was added
if the variable was selected in the final regression model.
In terms of magnitude of correlation with radar parameters, the absolute
correlation coefficient ranges from 0.25 to 0.37 after removing the insignificant
ones. Among the parameters, two textural parameters namely homogeneity and
angular second moment have comparatively stronger positive and negative
relationship with LAI respectively as indicated by higher correlation coefficient of
above 0.35. Other radar parameters show a negative association with LAI. However,
it is obvious that correlations of LAI with radar parameters are comparatively much
weaker than that with VI.
Table 5.4. Correlation between dependent variables (LAI) and independent
variables (Vegetation index and radar parameter)
Independent
Variables
Vegetation
Index
Radar
Parameter
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI1
MCARI2
RBS
FBS
HOMO
T-CT
T-DSM
LAI(Bon)
LAI(2000)
LAI(2000G)
0.236*
0.762**
0.771**
0.754**
0.785**
0.781**
0.665**
-0.06
0.194*
0.774**
0.785**
0.764**
0.814**
0.808**
0.657**
-0.01
0.217*
0.779**
0.789**
0.769**
0.810**
0.805**
0.668**
-0.03
-0.30**
-0.27**
-0.28**
0.37**
0.36**
0.36**
-0.07
-0.06
-0.04
-0.2
-0.19
-0.17
364
MN
SD
ENT
ASM
CORR
-0.28**
-0.25*
-0.26**
-0.05
-0.05
-0.04
-0.31**
-0.31**
-0.32**
-0.35**
-0.35**
-0.35**
-0.15
-0.15
-0.13
*Coefficients are significant at 0.05 level or better
** Coefficients are significant at 0.01 level or better
5.2.6. Outliner detection
Different types of outliner were detected after computation of Mahalanobis
distance D2 score and Cook’s distance as shown in Table 5.5. However, no univariate
outliner is found for LAI parameters. As for multivariate outliner (MO) on
independent variables, two cases with site ID of 307 and 314 are found as outliners
in all three LAI measurements as they had a cumulative probability Mahalanobis
distance D2 score in the upper tail of the distribution smaller than 0.001. Four cases
are identified as influential outliner (IO) in each of the LAI measurements. Cases 307
and 314 are identified as both MO and IO. Correlation analysis between the LAI and
VIs was performed again and the results are shown in Table 5.6. LAI and VIs are all
significantly correlated at p< 0.05 or p<0.01 except the correlation with NDVI. When
compared with the correlation coefficient before outliner removal in Table 5.4, the
associations between LAI and VIs are strengthened as indicated by higher
correlation coefficients except that with NDVI which showed less relationship with
the LAI. The outliners were excluded from the database before regression analysis.
The number of remaining valid cases after outliner removal equals 90, 91 and 91 for
LAI (Bon), LAI (2000) and LAI (2000G) respectively.
365
Table 5.5. The types and numbers of outliners detected
LAI(Bon)
*Types of
Outliner
Site ID
LAI(2000)
LAI(2000G)
MO
IO
MO
IO
MO
IO
307
314
120
131
220
307
307
314
216
307
314
323
307
314
216
219
307
314
No. of
Remaining
Cases
90
91
91
*MO = Multivariate outliner; IO = Influential outliner
Table 5.6. The Pearson correlation coefficient between LAI and VIs after outliner
removal
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI
1
MCARI
2
LAI(Bon)
0.197*
0.769**
0.778**
0.768**
0.783**
0.775**
0.675**
LAI(2000)
0.122
0.788**
0.800**
0.784**
0.829**
0.824**
0.681**
LAI(2000G)
0.131
0.790**
0.802**
0.786**
0.830**
0.826**
0.690**
*Coefficients are significant at 0.05 level or better
** Coefficients are significant at 0.01 level or better
5.3. Simple Linear Regression Analysis
Simple linear regression analysis was used to develop the relationship between LAI
(dependent variable) and various VIs (independent variables) after outliner removal.
Table 5.7 shows the results of linear regression analysis including the coefficients of
determination (r2), root-mean-square-error (RMSE) and regression model between
LAI and various VIs.
366
Results show that the r2 ranges from 0.015 to 0.689. TVI explains the highest
amount of variance in LAI with r2 equal 0.613, 0.687 and 0.689 for LAIBon, LAI2000
and LAI2000G respectively. Regression models computed from MCARI 1 also have
relatively high r2 of 0.601, 0.679 and 0.682 for LAIBon, LAI2000 and LAI2000G
respectively. SAVI is the third VI that has relatively better model fit with LAI as
indicated by r2 of 0.606, 0.640 and 0.643 for LAIBon, LAI2000 and LAI2000G
respectively. Similar trends of r2 are observed for the VI predictors and variances
explained by RDVI, MSAVI, MCARI 2 and NDVI decreases gradually. NDVI has the
worst fit to LAI among the VIs with r2 equal 0.039, 0.015 and 0.017 for LAIBon,
LAI2000 and LAI2000G respectively. The difference between r2 and adjusted r2 for
all models are less than 1% with the exception of regression model derived from
NDVI and MCARI 2.
LAI estimation using various VIs has RMSE ranged from 0.237 to 0.468. The RMSE
has an inverse relationship with the r2, i.e. the higher the r2 is, the lower is the RMSE.
Regression model using TVI as predictor produces the lowest error in LAI estimation
(RMSE=0.237 – 0.299) while models from NDVI results in the largest error
(RMSE=0.416 – 0.468). The lowest RMSE is found in TVI model regressed with
LAI2000G while the largest RMSE results from the NDVI model regressed with
LAIBon. Regression models using MCARI 1 have comparable prediction error with
TVI (RMSE=0.240 – 0.304). Prediction models from MCARI 2 also exhibit relatively
large errors among the VIs with RMSE vary from 0.307 to 0.355. For regression
models from other VIs including RDVI, SAVI and MSAVI, the prediction errors are
quite close to each other.
367
Table 5.7. Simple linear regression models computed from different VIs
LAI
LAI(Bon)
LAI(2000)
LAI(2000G)
368
VIs
r2
Adjusted r2
RMSE
Regression Model
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI 1
MCARI 2
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI 1
MCARI 2
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI 1
MCARI 2
0.039
0.592
0.606
0.590
0.613
0.601
0.455
0.015
0.621
0.640
0.614
0.687
0.679
0.464
0.017
0.625
0.643
0.618
0.689
0.682
0.476
0.028
0.587
0.602
0.586
0.609
0.596
0.449
0.004
0.617
0.636
0.610
0.684
0.675
0.458
0.006
0.620
0.639
0.614
0.685
0.678
0.470
0.468
0.305
0.300
0.307
0.299
0.304
0.355
0.454
0.282
0.275
0.285
0.260
0.263
0.339
0.416
0.258
0.252
0.261
0.237
0.240
0.307
-0.402+3.162(NDVI)
-4.66+12.739(RDVI)
-4.366+11.549(SAVI)
-2.728+8.304(MSAVI)
-1.770+0.188(TVI)
-1.748+7.792(MCARI 1)
-1.799+7.116(MCARI 2)
0.557+1.908(NDVI)
-4.598+12.477(RDVI)
-4.318+11.325(SAVI)
-2.688+8.105(MSAVI)
-1.851+0.188(TVI)
-1.842+7.811(MCARI 1)
-1.765+6.916 (MCARI 2)
0.518+1.886(NDVI)
-4.096+11.433(RDVI)
-3.837+10.372(SAVI)
-2.347+7.427(MSAVI)
-1.584+0.173(TVI)
-1.583+7.185(MCARI 1)
-1.560+6.444(MCARI 2)
LAI computed using LAI2000G method modeled with all VIs yields the highest r2 and
lowest RMSE except with NDVI. This is followed by LAI2000 while LAIBon have
comparatively weaker relationship with the VIs. Detail results of LAI estimation
(LAI2000G method) with individual VI are discussed below. Results from the
LAI2000 and LAIBon methods are excluded here. However, maps of predicted LAI
for the two methods can be found under Appendix C.
5.3.1. LAI2000 Generalized method
The relationship between measured LAI and seven VIs are plotted in Figure 5.5.
LAI variation explained by the NDVI model is the lowest among the VIs as indicated
by very low r2 (0.017). The relationship between NDVI and LAI is very poor and this
is not surprised due to the low correlation with LAI (0.131) in Table 5.6. The F value
(1.56) has the significance level of 0.216, which fail to reject the null hypothesis and
there is no significant relationship between LAI and NDVI. Besides, the significance
level of  coefficient for NDVI (0.681) is also insignificant at p<0.05.
Regression model formulated using RDVI has r2 and adjusted r2 equals 0.625 and
0.620 respectively. The relationship between RDVI and LAI is strong and small
discrepancy between the two r2 indicates the model is robust. The F value (148.04)
has a significance level less than 0.01, which suggest a significant relationship
between LAI and RDVI. Besides, the  coefficient for RDVI is also significant at
p<0.05.
The amount of LAI variations explained by SAVI is high as indicated by relatively high
r2 of 0.643 and adjusted r2 of 0.639. The relationship between SAVI and LAI is strong
while the model is robust indicated by little discrepancy between r 2 and adjusted r2.
The F value (159.98) has the significance level less than 0.01, which rejects the null
hypothesis and signify a significant relationship between LAI and SAVI. The 
coefficient for SAVI is also significant at p<0.05.
369
Regression model computed from MSAVI has r2 and adjusted r2 equals 0.618 and
0.614 respectively. The relationship between MSAVI and LAI is strong while small
discrepancy between the two r2 signifies the model is robust. The F value (144.07)
with significance level less than 0.01 rejects the null hypothesis and suggests a
significant relationship between LAI and MSAVI. The  coefficient for MSAVI is also
significant at p<0.05.
Regression model computed from TVI has the highest r2 and adjusted r2 equals
0.689 and 0.685 respectively among the VIs. The relationship between TVI and LAI is
very strong and small discrepancy between the two r2 proves the model is robust.
The F value (197.17) having the level of significance less than 0.01 rejects the null
hypothesis and a significant relationship is found between LAI and TVI. The 
coefficient for TVI is also significant at p<0.05.
Regression model formulated from MCARI 1 has similar amount of LAI variation
compared with TVI as indicated by the relatively high r2 and adjusted r2 of 0.682 and
0.678 respectively. The F value (190.66) with significance level less than 0.01 rejects
the null hypothesis and suggests a significant relationship between LAI and MCARI 1.
The  coefficient for MCARI 1 is also significant at p<0.05.
Finally, the amount of LAI variations explained by MCARI 2 is comparatively lower as
indicated by relatively low r2 of 0.476 and adjusted r2 of 0.47. The relationship
between MCARI 2 and LAI is moderate and the model is robust indicated by little
discrepancy between r2 and adjusted r2. The F value (80.88) has the significant level
less than 0.01, which rejects the null hypothesis and signify a significant relationship
between LAI and SAVI. The  coefficient for MCARI 2 is also significant at p<0.05.
370
2
2
r =0.625
Measured LAI
r =0.017
2
2
r =0.618
Measured LAI
r =0.643
2
2
r =0.682
Measured LAI
r =0.689
371
2
Measured LAI
r =0.476
Figure 5.5. The linear regression model between field-measured LAI and VIs
The relationship between measured and predicted LAI using various VIs are shown
in Figure 5.6. LAI estimations that follow closely to the solid line indicates accurate
prediction. The dotted line indicates ±1 standard deviation (σ=0.41) of the
measured LAI after outliner removal. Points distributed on the left hand side of the
solid line are overestimated LAI while those on the right are underestimated cases.
The majority of predictions by the seven VIs fall within ±1 standard deviation of
measured LAI except from the NDVI model in which only 63% of field samples fall
within ±1σ. The best models are produced by two VIs namely, TVI and MCARI1.
Over 93% of field samples fall within ±1σ. Models from RDVI, SAVI and MSAVI
perform similarly with about 90% of field samples within ±1σ.
NDVI
RDVI
372
SAVI
MSAVI
TVI
MCARI 1
MCARI 2
Figure 5.6. Scatterplot of measured and predicted LAI (LAI2000G) using different VIs
after linear regression analysis
373
Figures 5.7 – 5.13 shows the spatial distribution of predicted LAI after applying the
linear regression models of different VIs and Table 5.8 shows the description
statistics including maximum, minimum, average and standard deviation of LAI over
the study area. From the statistical descriptive, the maximally predicted LAI is
different across VI models. Model of NDVI has the lowest LAI (2.48) while the
highest is found in the MCARI2 (5.15). Linear regression models using RDVI, SAVI,
TVI and MCARI1 as predictors have maximum LAI stays around 4 (range between
4.00 and 4.08) and the MSAVI model has slightly higher maximum LAI of 4.23.
Similar variation is also observed in LAI computed from the other two methods. The
minimal predicted LAI for all VI models stays very close to zero. Examining the
spatial distribution of prediction in the figures, the minimal represented by magenta
color, distribute along the coastal fringe, adjacent to the land boundary as well as
on boundary along the river channels. Besides, as their distributions are not
continuous, it is believed that the effect of mixed pixel in these boundary locations
causes impure vegetation signature which in turn influences LAI estimation when
reflectance retrieved from the image are used for prediction. With the exception of
NDVI model prediction, distinct pattern of LAI is observed over the area. Apparent
pattern of LAI is observed under all models (except NDVI), however, the most
significant one found in the TVI and MCARI1 models in Figures 5.11 and 5.12
respectively. Three areas with relatively higher LAI indicated by brown and orange
colors locate in the northern part, the central part as well as along the old river
channels in the southern part of the study area. Mangroves in the northern coastal
fringe are characterized by the active growing species mix of A. corniculatum, A
ilicifolius and K. obovata. Region of high LAI in the central part of the area is
dominated by K. obovata G2 while those locates along the old river channels in the
southern part is mainly A. ilicifolius G1. Mangroves locate towards the land
boundary has relatively lower LAI value as indicated by dark and light blue color.
Apparent low LAI is found for K. obovata G1.
374
Figure 5.7 Predicted LAI(2000G) from regression model using NDVI as predictor
Figure 5.8. Predicted LAI(2000G) from regression model using RDVI as predictor
375
Figure 5.9. Predicted LAI(2000G) from regression model using SAVI as predictor
Figure 5.10. Predicted LAI(2000G) from regression model using MSAVI as predictor
376
Figure 5.11. Predicted LAI(2000G) from regression model using TVI as predictor
Figure 5.12. Predicted LAI(2000G) from regression model using MCARI 1 as
predictor
377
Figure 5.13. Predicted LAI(2000G) from regression model using MCARI 2 as
predictor
Table 5.8. Descriptive statistics for LAI (2000G) predicted over the study area
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI
1
MCARI
2
Maximum
2.48
4.05
4.00
4.23
4.05
4.08
5.15
Minimum
1.24E01
1.83E02
2.36E03
4.33E03
1.52E04
3.18E03
2.34E04
Mean
1.92
2.06
2.05
2.02
1.98
1.97
1.99
Standard deviation
0.36
0.74
0.74
0.79
0.74
0.74
0.81
Similarly, simple linear regression was used to regress LAI against individual radar
parameters. The relationship between measured LAI and radar parameters are
plotted in Figure 5.14. LAI variations explained by various radar parameters are very
low VIs as indicated by very low r2 ranged between 0.000304 – 0.12. The
relationship between the radar parameters and LAI is very poor. It is not surprised
378
as the correlation between is also low shown in Table 5.4. Predictor variables
including filtered backscatter, homogeneity, angular second moment and entropy
all have the significance level of  coefficient less than 0.05 which indicates that
they are significant. Other radar predictor variables have significance level of 
coefficient greater than 0.05 which suggest they are insignificant. As radar
parameters have poor relationship with LAI, it is obvious that use of radar
parameter as independent predictor alone would not produce reliable LAI
estimation.
2
2
r =0.061
Measured LAI
r =0.000304
2
2
r =0.006
Measured LAI
r =0.12
379
2
2
r =0.054
Measured LAI
r =0.03
2
2
r =0.09
Measured LAI
r =0.003
2
2
r =0.034
Measured LAI
r =0.12
Figure 5.14. The linear regression model between field-measured LAI and radar
parameters
380
5.4. Stepwise Multiple Regression Analysis
LAI was regressed against VI combined with backscatter and GLCM-texture
parameters using stepwise multiple regression. Important independent variables
were selected if they are significant to LAI estimation through the stepwise process.
Depending on combination of independent variables, multiple stepwise regression
were conducted in two ways:
i.
Each VI was combined individually with all radar parameters
ii.
All VIs and radar parameters were combined
Table 5.9 shows the results of stepwise linear regression based on the first set of
independent variables. The coefficients of determination (r2), root-mean-squareerror (RMSE) and regression model between dependent variables (LAI) and
independent variables (VIs and radar parameters) are shown in the Table. The
coefficients of determination for the full model before stepwise selection are
incorporated as a reference to the effect of variable removal by stepwise process.
The NDVI and MSAVI models are excluded from further analysis. During the variable
selection process, NDVI was dropped and the GLCM textural variable, Angular
Second Moment (ASM) was the only variable selected with resultant r2 equals 0.16.
For MSAVI model, no radar parameter was selected during stepwise regression
process. For the remaining models, ASM was the only radar predictor retained
during stepwise regression.
Results show that the r2 has range between 0.54 and 0.79. Vegetation indices, TVI
and MCARI 1 coupled with ASM explains the highest amount of variance of LAI with
r2 equal 0.71, 0.79 and 0.79 for LAIBon, LAI2000 and LAI2000G respectively.
Multiple regression model combining SAVI and ASM has relatively better model fit
as indicated by r2 of 0.673, 0.712 and 0.723 for LAIBon, LAI2000 and LAI2000G
respectively. Model obtained from regression model combining RDVI and ASM has
comparable variances explained with r2 ranged between 0.66 - 0.71. Combination of
MCARI 2 and ASM has the worst model fit to LAI with r 2 equal 0.54, 0.54 and 0.56
for LAIBon, LAI2000 and LAI2000G respectively.
381
LAI estimation has RMSE ranged from 0.20 to 0.33. The RMSE has an inverse
relationship with the r2 that is the higher the r2 is, the lower is the RMSE. Models
obtained using TVI and MCARI 1 separately combined with ASM as predictor
produces the lowest error in LAI estimation (RMSE=0.20 - 0.27) while model
combining MCARI 2 and ASM results in the largest error (RMSE=0.28 - 0.33).
Prediction models of RDVI and SAVI combine individually with ASM results in similar
prediction errors ranged from 0.23 (with LAI2000G) to 0.29 (with LAIBon).
Based on the same variable entry and removal setting, stepwise regression using all
VIs and radar parameters (second set) has selected TVI and ASM as independent
variable predictors in all three LAI models. The r2, RMSE and the regression equation
are therefore the same as the TVI-ASM model shown in Table 5.9.
LAI computed using LAI2000G method yields the highest r2 and lowest RMSE.
Similarly, LAI computed using LAI2000 and LAIBon methods have comparatively
weaker relationship with various VIs combined with ASM. Results of LAI estimation
from the LAI2000G method are presented below. Estimation using methods of
LAI2000 and LAIBon are organized under Appendix D.
382
Table 5.9. Stepwise regression models computed from different VIs and radar textural variables
LAI
LAI(Bon)
LAI(2000)
LAI(2000G)
Full model r2
r2
Adjusted r2
RMSE
Regression Model
RDVI
0.715
0.660
0.652
0.290
-4.523+11.924(RDVI)+5.98(ASM)
SAVI
0.723
0.673
0.665
0.285
-4.238+10.791(SAVI)+6.001(ASM)
*TVI
0.756
0.713
0.707
0.267
-1.961+0.179(TVI)+7.970(ASM)
MCARI 1
0.756
0.712
0.705
0.267
-1.99+7.429(MCARI 1)+8.5(ASM)
MCARI 2
0.64
0.544
0.534
0.330
-1.901+6.642(MCARI 2)+7.247(ASM)
RDVI
0.759
0.694
0.687
0.254
-4.483+11.888(RDVI)+3.801(ASM)
SAVI
0.771
0.712
0.706
0.247
-4.216+10.792(SAVI)+3.794(ASM)
*TVI
0.824
0.790
0.785
0.212
-2.005+0.182(TVI)+5.765(ASM)
MCARI 1
0.823
0.789
0.784
0.212
-2.038+7.564(MCARI 1)+6.285(ASM)
MCARI 2
0.651
0.542
0.531
0.310
-1.821+6.54(MCARI 2)+5.062(ASM)
RDVI
0.776
0.706
0.699
0.233
-4.173+11.24(RDVI)+3.34(ASM)
SAVI
0.788
0.723
0.717
0.227
-3.915+10.193(SAVI)+3.334(ASM)
*TVI
0.835
0.788
0.783
0.200
-1.794+0.17(TVI)+5.21(ASM)
MCARI 1
0.835
0.789
0.784
0.199
-1.827+7.084(MCARI 1)+5.692(ASM)
MCARI 2
0.679
0.563
0.552
0.283
-1.7+6.266(MCARI 2)+4.491(ASM)
VIs
#
# Full model contains 11 predictor variables including single VI, RBS, FBS, HOMO, CT, DSM, MN, SD, ENT, ASM, CORR
383
* Model formed by using the second set of independent variables (all VI and radar parameters) as input
5.4.1. LAI2000 Generalized method
Three predictor variables namely RDVI, ASM and ENT were selected in the final
regression model combining RDVI and radar parameters after stepwise analysis.
However, the two radar parameters, ASM and ENT, have tolerance value of 0.082
which is smaller than the critical value of 0.1. This suggests the two variables are
highly correlated. The correlation analysis shows that the two variables have very
high correlation of -0.958 significant at p<0.01. The presence of two variables
together
in
the
model
suggested
information
redundancy
and
cause
multicollinearity problem. Therefore, the second best model retaining two variables
namely, RDVI and ASM is selected as the final model. The final model has r 2 and
adjusted r2 equals 0.706 and 0.699 respectively. The relationship between LAI and
the two independent variables is very strong and small discrepancy between the
two r2 (less than 1%) indicates the model is robust. The F value (100.90) has a
significance level less than 0.01. Besides, the  coefficients for RDVI and ASM have
level of significance smaller than 0.05. A significant relationship between LAI and
the predictors is observed.
Similarly, stepwise analysis combining SAVI and radar parameters select three
predictor variables namely SAVI, ASM and ENT in the final model. However, the two
radar parameters have tolerance value of 0.082 which is smaller than the critical
value of 0.1. Consequently, the second best model retaining two variables namely,
RDVI and ASM is selected as the final model. The amount of LAI variations explained
by SAVI is high as indicated by relatively high r2 of 0.723 and adjusted r2 of 0.717.
The relationship between LAI and the independent variables is very strong while the
model is robust shown by small discrepancy between r2 and adjusted r2. The F value
(109.79) has the significance level less than 0.01, which rejects the null hypothesis.
Besides, the  coefficients for RDVI and ASM are both significant at p<0.05. These
all imply the relationship between LAI and the predictors is significant.
Stepwise regression has selected two variables, TVI and ASM when LAI was
regressed with TVI and radar parameters. The resultant model has high r2 and
adjusted r2 equals 0.788 and 0.783 respectively. The relationship between LAI and
384
the two independent predictors is very strong and small discrepancy of less than 1%
between the original and adjusted r2 proves the model is robust. The F value
(156.43) having the level of significance less than 0.01 rejects the null hypothesis.
Besides, the  coefficients for the two predictors are also significant at p<0.05. A
significant relationship between LAI and the predictors is observed.
Stepwise analysis combining MCARI 1 and radar parameters results in selection of
two variables namely MCARI 1 and ASM in the final model. The resultant model has
comparatively high r2 and adjusted r2 equals 0.789 and 0.784 respectively. The F
value (156.72) with significance level less than 0.01 rejects the null hypothesis.
Besides, the  coefficient for MCARI 1 and ASM have level of significance smaller
than 0.05. These all suggest a significant relationship between LAI and the two
predictor variables.
The final model has similarly selected two three variables namely MCARI 2, ASM
and ENT when combining MCARI 2 with the radar parameters in stepwise analysis.
The high and significant correlation between the two selected radar predictor
variables results in tolerance value less than 0.1. The ENT was discarded from the
model and the final model has two predictors, MCARI 2 and ASM. The amount of
LAI variations explained by the predictors is comparatively lower as indicated by
relatively low r2 of 0.563 and adjusted r2 of 0.552. The relationship between LAI and
the two predictors is moderate and the model is not as robust as the other VI
predictors as the discrepancy between r2 and adjusted r2 is about 2%. The F value
(54.04) is significant at p<0.01, which rejects the null hypothesis. Besides, the 
coefficients for MCARI 2 and ASM are significant at p<0.05. A significant relationship
between LAI and the predictors is observed.
The relationship between measured and predicted LAI (LAI2000G method)
combining individual VI and ASM are shown in Figure 5.15. LAI estimations that
follow closely to the solid line indicates accurate prediction. The dotted line
indicates ±1 standard deviation (σ=0.41) of the measured LAI after outliner removal.
Comparing with measured LAI, estimations through combining individual VI and
ASM mostly fall within ±1σ of measured LAI. The best models are produced by
385
combining TVI and MCARI 1 separately with ASM having over 96% of field samples
fall within ±1σ. Models from RDVI and SAVI combined with ASM also perform well
with about 94% of field samples within ±1 standard deviation. The worst model fit is
by combining MCARI 2 and ASM with 87% samples within ±1σ.
RDVI+ASM
SAVI+ASM
TVI+ASM
MCARI1+ASM
MCARI2+ASM
Figure 5.15. Scatterplot of measured and predicted LAI (LAI2000G) combining VIs
and ASM after stepwise regression model
386
Figures 5.16 – 5.20 shows the spatial distribution of predicted LAI after applying the
stepwise model combining individual VIs with ASM as predictor variables. Table
5.10 shows the first order descriptive statistics including maximum, minimum,
average and standard deviation of predicted LAI summarized over the study area.
From the descriptives, the maximally predicted LAI is different across VI models.
Model of SAVI-ASM has the lowest maximum of 4.65 while the highest maximum is
found under the combined MCARI1-ASM model (5.98). The minimum predicted LAI
for all VI models stays very close to zero. Areas of minimum LAI are represented by
magenta color in figures and they are found along the coastal fringe, adjacent to the
land boundary as well as on boundary along the river channels. Similar to the linear
regression model, the effect of mixed pixel along the boundaries causes such
variation. All models show distinct LAI pattern over the study area. The models have
highlighted a few local hotspots of extremely high LAI colored by brown and orange
colors locating in the northern part, the central part and in the southern part of the
study area. These hotspots and areas of relatively high LAI are most distinctive
under the TVI-ASM and MCARI1-ASM models in Figures 5.18 and 5.19 respectively.
The additional hotspots in the northern part are dominant by A. ilicifolius G2.
Region of high LAI in the central part of the area is dominated by K. obovata G2 as
well as A. ilicifolius G2. A. ilicifolius G1 locating along the old river channels in the
southern part remains of high LAI. K. obovata G1 located adjacent to the land
boundary has relatively lower LAI value as indicated by dark and light blue color.
387
Figure 5.16. Predicted LAI(2000G) from regression model using RDVI and ASM as
predictors
Figure 5.17. Predicted LAI(2000G) from regression model using SAVI and ASM as
predictors
388
Figure 5.18. Predicted LAI(2000G) from regression model using TVI and ASM as
predictors
Figure 5.19. Predicted LAI(2000G) from regression model using MCARI 1 and ASM as
predictors
389
Figure 5.20. Predicted LAI(2000G) from regression model using MCARI 2 and ASM as
predictors
Table 5.10. Descriptive statistics for LAI(2000G) predicted over the study area
RDVI-ASM
SAVI-ASM
TVI-ASM
MCARI 1ASM
MCARI 2ASM
Maximum
4.66
4.65
5.72
5.98
5.40
Minimum
3.53E-03
7.28E-03
3.16E-03
5.24E-03
1.72E-03
Mean
1.79
1.80
1.95
1.98
1.89
Standard
deviation
1.10
1.09
1.03
1.05
1.09
390
5.5. Discussion and Implication
Based on the predicted LAI from the linear and stepwise regression models, further
analyses were conducted to have an understanding on the effects of VI, backscatter
and GLCM-textural variables on LAI estimation. LAI computed using different
methods are compared first. Second, relationship between LAI and species
composition is then explored. The two sessions followed reveal the relationship of
LAI with VIs, backscatter and radar parameters by computing statistics of the whole
study area and also along pre-defined profiles. Except the first part, all other
comparisons were based on LAI computed using LAI2000 generalized (LAI2000G)
method. Last but not the least, the complementary of VI and radar parameters for
LAI mapping is discussed.
5.5.1. LAI model comparison
The study has compared three LAI computation methods, namely the Bonhomme
and Chartier (LAIBon), LAI2000 Plant Canopy Analyzer (LAI2000) and LAI2000
generalized (LAI2000G) methods. The main difference among the three methods
lies on the field of view (in terms of degree) considered. The Bonhomme and
Chartier’s method is the simplest in terms of computation as it considers the only
zenith angle at 57.5°. The method yields the largest range of LAI, 2.25 as well as the
highest maximum measured LAI of 3.56 among the three methods. However, LAI
prediction based on this set of measured LAI shows the largest error. LAI derived
from LAI2000 generalized methods has the minimum range of LAI of 1.65 as well as
the lowest maximum LAI of 2.91 among the three methods. However, LAI
estimation based on this set of field measured LAI yields the smallest error. Figure
5.21 shows the comparison LAI prediction of the three methods based on the
simple TVI regression model. The spatial variation and pattern of LAI are almost the
same in the area. The difference lies in the predicted maximum which are 4.35, 4.27
and 4.05 for LAI(Bon), LAI(2000) and LAI(2000G) respectively.
391
LAI(Bon)
LAI(2000)
LAI(2000G)
Figure 5.21. Comparison of results for LAI computed using three different methods
based on simple linear TVI regression model
The accuracy of LAI estimation from the simple linear regression varies across
individual VI. NDVI has the worst performance in predicting LAI as indicated by low
r2, high divergence from measured LAI (large RMSE). NDVI is the most commonly
adopted broadband vegetation index contrasting the maximum absorption in red
and the maximum reflection in near infrared. NDVI has been reported to saturate
when LAI reaches a certain limits, for instance exceeds 2 as experienced in many
studies (Lillesaeter, 1982, Baret and Guyot, 1991). In Figure 5.6, regression model
show no distinct linear relationship with field-measured LAI. Compared with NDVI,
RDVI was developed to enhance the linearity with biophysical variables in both low
and high vegetation density. By simply taking the square root of the summation of
red and near infrared reflectance, apparent improvement of relationship with LAI is
observed in the RDVI model. As field-measured LAI varies from 1.26 to 3.56,
application of RDVI is particularly suitable for diverse LAI estimation.
Apparent LAI variation is observed across the study area. Under low LAI, the soil
optical properties will contribute to the spectral reflectance of canopy, which would
392
curtail the effectiveness of using remotely sensed data for LAI estimation. The two
VIs namely SAVI and MSAVI was chosen because they are less sensitive to the effect
of soil background, which can therefore improve the prediction power of LAI. This is
proved by the relatively high r2 (0.59 - 0.64) and low RMSE (0.25 - 0.31) for the two
models. However, arbitrary setting of the soil adjustment L factor in SAVI results in
better estimation than the embedded soil adjustment in MSAVI.
Apart from effect of soil background, leaf biochemical variation, especially
chlorophyll concentration is another important external factor influencing the
responsiveness of VI to LAI variations. The three VIs including TVI, MCARI 1 and 2
were selected as they are resistant to chlorophyll change. Regression models from
TVI and MCARI 1 have the best estimation with the highest r2 (0.6 - 0.69) and the
lowest RMSE (0.24 - 0.30). It is suggested that the suppression of chlorophyll
influence has enhanced the sensitivity of VIs to LAI variations. And the
enhancement is even more prominent than the suppression of soil effect. However,
the addition of soil adjustment factor in MCARI 2 has negative effect on the model
performances with r2 reduced by 25-30% and RMSE increased by 16-22%. The soil
adjustment factor may not be applicable the distinctive soil in inter-tidal mangrove
area which is always in saturated condition or inundated under water.
5.5.2. Species composition and LAI
In order to have a more thorough understanding towards the relationship between
species and LAI, LAI of different species classes were extracted from the VI models.
The species distribution was adopted from the classification result of artificial
neural network (ANN) as the algorithm was comparatively more accurate and stable.
Only the central part of the mangrove zone was considered as it shows the most
distinctive pattern of species distribution. Besides, as the distribution of A.
corniculatum and Sonneratia spp. is not accurately predicted, A. corniculatum was
eliminated from consideration while Sonneratia spp. was located manually by
selecting the representative site in the Futian area. Moreover, boundary pixels
locating along the coastal fringe and close to land boundary with LAI lower than
393
0.01 were excluded from the computation. A total of 1,352 samples were extracted
and the locations are shown in Figure 5.22. The mean LAI plots of the species for the
simple linear VI model and multiple regression model are shown in Figures 5.23 and
5.24 respectively.
Figure 5.22. Location of mangrove species class where LAI were extracted for
comparison
In Figure 5.23, LAI difference is observed among the mangrove species classes. With
the exception of estimation from the NDVI model, other VI models show similar LAI
variations across species. K. obovata G2 has the highest mean LAI ranged between
2.26 and 3.26. This is followed by A. ilicifolius G1 (2.07 - 2.65), A. ilicifolius G2 (1.97 2.24), A. marina (2.01 - 2.18), K. obovata G1 (1.74 - 2.03) and Sonneratia spp. has
the lowest average LAI ranged between 1.31 and 2.10 across the VI models. The
largest mean LAI difference is found between K. obovata G2 and Sonneratia spp.
with range between 0.16 (NDVI model) and 1.51(MCARI2 model). The mean LAI
difference between the two groups of K. obovata is quite obvious with range
394
between 0.23 (NDVI model) and 1.38 (MCARI2 model). Comparatively, the two
groups of A. ilicifolius have relatively smaller mean LAI difference with range
between 0.01 (MCARI 2 model) and 0.26 (MCARI 1 model).
3.50
3.00
Leaf Area Index (LAI)
2.50
2.00
1.50
1.00
0.50
0.00
NDVI model
RDVI Model
SAVI Model
MSAVI Model
TVI Model
MCARI1 Model
MCARI2 Model
Regression Model
A. ilicifolius G1
A. ilicifolius G2
A. marina
K. obovata G1
K. obovata G2
Sonneratia spp.
Figure 5.23. Mean LAI (LAI2000G) plot of mangrove species classes under simple
linear VI models
Figure 5.24 shows LAI variation across species classes under the multiple VI-ASM
models. LAI variation among the species is identical to that from the simple VI
models but the mean values have increased for some species in specific models
while others remain the same values. K. obovata G2 has the highest mean LAI
ranged between 3.04 - 3.47 in all models. This is followed by A. ilicifolius G1 (2.43 2.69), A. ilicifolius G2 (2.18 - 2.48), A. marina (2.08 - 2.30), K. obovata G1 (1.75 - 1.96)
and Sonneratia spp. has the lowest average LAI (1.25 - 1.55). The mean LAI
difference between the two groups of K. obovata is still obvious with range
between 1.11 (TVI model) and 1.51 (MCARI 2 model). Comparatively, the two
groups of A. ilicifolius have relatively smaller mean LAI difference with range
between 0.07 (MCARI 2 model) and 0.27 (RDVI and SAVI models).
395
4.00
3.50
Leaf Area Index (LAI)
3.00
2.50
2.00
1.50
1.00
0.50
0.00
RDVI Model
SAVI Model
MSAVI Model
TVI Model
MCARI1 Model
MCARI2 Model
Regression model
A. ilicifolius G1
A. ilicifolius G2
A. marina
K. obovata G1
K. obovata G2
Sonneratia spp.
Figure 5.24. Mean LAI (LAI2000G) plot of mangrove species classes under multilple
VI-ASM models
To further explore the change of LAI for the species groups, comparison of mean LAI
in Figure 5.23 and 5.24 results in the difference of mean LAI before and after
incorporating the radar parameter ASM. Change of mean LAI of the species under
different VI models is plotted in Figure 5.25. All species have an increase in mean
LAI after adding ASM predictor variable in different VI models with the exception of
Sonneratia spp. The increase is consistent in different VI models with the largest
increase under MCARI 1 model followed by TVI, MCARI 2 and LAI increase in RDVI
and SAVI models is about the same. In terms of species, K. obovata G2 has the
largest increase in LAI (0.16 - 0.30) across different models. This is followed by A.
ilicifolius G2 (0.13 - 0.23), K. obovata G1 (0.10 - 0.18), A. marina (0.06 - 0.12), A.
ilicifolius G1 (0.02 - 0.05). Sonneratia spp. experiences a slight decrease in LAI by
0.04 - 0.06.
In simple linear regression analysis using VI alone, LAI of A. ilicifolius G1 is higher
than that of A. ilicifolius G1. After the input of ASM, increase in mean LAI of A.
ilicifolius G1 (0.02 - 0.05) is comparatively smaller than increase in mean LAI of A.
ilicifolius G2 (0.13 - 0.23) in different models. Therefore, LAI difference between the
396
two groups of A. ilicifolius is narrowed by a range of 29 - 66% (VI model dependent).
The maximum difference reduction of 66% is found in the MCARI 2 model.
Similarly, LAI of K. obovata G2 is generally higher than that of K. obovata G1. The
inclusion of ASM results in the largest LAI increase (0.16 - 0.30) for K. obovata G2.
The increase in LAI for K. obovata G1 (0.10 - 0.18) is not as high as its counterpart.
Therefore, LAI difference between the two groups of K. obovata is narrowed by a
range of 5 - 12% (VI model dependent). The maximum difference reduction of 12%
is found in the TVI model.
0.35
0.30
0.25
0.20
LAI
0.15
0.10
0.05
0.00
-0.05
-0.10
RDVI Model
A. ilicifolius G1
SAVI Model
A. ilicifolius G2
TVI Model
A. marina
K. obovata G1
MCARI1 Model
K. obovata G2
MCARI2 Model
Sonneratia spp.
Figure 5.25. Change of mean LAI of species after addition of radar parameter ASM in
different VI models
5.5.3. Hyperspectral Bands, Vegetation Indices and LAI
To further understand the relationship between LAI and hyperspectral data as well
as radar backscatter/ textural variables, predicted LAI covering the study area are
grouped into classes based on LAI values shown Table 5.11 below. Predicted LAI
from the linear TVI regression model are grouped into 5 classes while that from the
multiple regression (TVI-ASM) model are grouped into 6 classes because of
397
relatively higher predicted LAI in the multiple regression model. The number of
cases/ pixels in each LAI class is noted as well. It is noted that 90% and 80% of cases
have LAI values less than 3 in the TVI and TVI-ASM model respectively. As both
models have less than 1% of cases in the class of the highest LAI, that class was
excluded from consideration. LAI computed from simple linear and multiple
regression models are related to the hyperspectral bands, vegetation indices to
understand the response of various remotely-sensed data under low and high LAI.
Table 5.11. Group LAI classes of the study area
LAI Class
*LAI
(TVI model)
Number of
cases (%)
*LAI
(TVI-ASM
model)
Number of
cases
1
0.00 – 1.00
576 (13%)
0.00 – 1.00
545 (12%)
2
1.01 – 2.00
1341 (30%)
1.01 – 2.00
1136 (25%)
3
2.01 – 3.00
2312 (52%)
2.01 – 3.00
2154 (48%)
4
3.01 – 4.00
229 (5%)
3.01 – 4.00
575 (13%)
5
4.01 – 5.00
3 (<1%)
4.01 – 5.00
46 (1%)
5.01 – 6.00
5 (<1%)
6
*LAI are computed using the LAI2000G method
The mean reflectance across the hyperspectral spectrum of LAI classes for the
simple VI regression and stepwise multiple regression model are shown in Figure
5.26 and 5.27 respectively. As the number of cases in class 5 and 6 in the simple VI
and multiple regression model respectively is too small, they were excluded from
the calculation. In Figure 5.26, apparent differences among the LAI classes are
observed from the visible bands to the shortwave infrared bands at 1276nm. The
most distinct difference in mean reflectance between the LAI groups is observed in
the near infrared (753 – 906nm) as well as in shortwave infrared regions (1024 –
1084nm). Regions of low LAI have relatively higher mean reflectance in the visible
bands and mean reflectance decrease with increasing LAI. The reverse is observed
398
in the infrared regions with lower mean reflectance observed in areas of low LAI
and mean reflectance increases with increasing LAI. The relatively high reflectance
in low LAI is due to comparatively less significant absorption by leaf pigments than
absorption occurs in high LAI. Besides, sparse canopy in low LAI, effect of underlying
soil optical properties becomes more apparent. According to (Campbell and
Norman, 1998), wet soil has solar reflectivity of about 0.08 – 0.10. Although the
reflectivity is not high, it certainly contributes to the reflectance in the visible bands.
Increasing LAI results in higher absorption and reduces reflectance in the visible
band. Different from the visible region, vegetation canopy is characterized by high
reflectance, transmittance and low absorption in the near infrared region.
Governed by the leaf additive reflectance, increase in the number of leaf or leaf
layers, i.e. LAI would further enrich near infrared reflectance. Therefore, NIR
reflectance and LAI are in positive relationship.
In the visible regions, there is a large decrease in mean reflectance as LAI increases
from 1 to 2 and magnitude of mean reflectance reduction decreases as LAI
increases further from 2 to 4. Similarly, increase in mean reflectance in the infrared
regions is more pronounced in low LAI (from 1 to 2) and the increase diminishes in
magnitude towards high LAI. It implies that the change in mean reflectance is not
linearly related to LAI. This is supported by the radiative transfer simulation using
SAIL model by (Verhoef, 1984) in which the relationship follows an exponential
function.
In Figure 5.27, the addition of ASM parameters raises the maximum LAI by one class
represented by the orange line. Mean reflectance of LAI classes follows the same
variation across the spectrum in Figure 5.26. However, compared with the mean LAI
of the same class in Figure 5.26, mean reflectance of low LAI (class 1) decreases
while that of the high LAI (class 4) increases. Mean reflectance of classes 2 and 3
remains almost the same. The range of mean reflectance between low and high LAI
is enlarged. Class of maximum LAI (class 5) shows very little difference of mean
reflectance with class 4 across the whole spectrum. Presumably, it suggests
sensitivity of near infrared to LAI larger than 4 diminishes significantly.
399
0.50
0.45
0.40
Mean Reflectance
0.35
0.30
0.25
0.20
0.15
0.10
0.05
2375
2315
2345
2224
2254
2285
2164
2194
2073
2103
2133
2012
2043
1770
1982
1680
1710
1740
1619
1649
1528
1558
1589
1336
1498
1276
1306
1185
1215
1246
1084
1114
993
1024
1054
906
963
845
876
753
784
815
692
723
601
631
662
540
570
479
509
0.00
Wavelength (nm)
LAI Class 1
LAI Class 2
LAI Class 3
LAI Class 4
Figure 5.26. Mean spectral reflectance of LAI classes grouped according to the
predicted LAI (LAI2000G) from simple VI regression model
0.50
0.45
0.40
Mean Reflectance
0.35
0.30
0.25
0.20
0.15
0.10
0.05
2375
2285
2315
2345
2224
2254
2164
2194
2073
2103
2133
2012
2043
1740
1770
1982
1680
1710
1619
1649
1528
1558
1589
1336
1498
1246
1276
1306
1185
1215
1084
1114
993
1024
1054
906
963
845
876
753
784
815
692
723
601
631
662
540
570
479
509
0.00
Wavelength (nm)
LAI Class 1
LAI Class 2
LAI Class 3
LAI Class 4
LAI Class 5
Figure 5.27. Mean spectral reflectance of LAI classes grouped according to the
predicted LAI (LAI2000G) from multiple regression model
400
The idea of using VIs instead of individual bands for LAI estimation originated from
the understanding that combination of individual bands are resistant to undesired
external effect from atmospheric conditions, soil optical properties and illumination
geometry and enhance the sensitivity to LAI only. The contrast of high absorption in
red and high reflectance and transmittance in near infrared has been used to
compute various vegetation index and they show different success for LAI
estimation. Figure 5.28 and 5.29 show the comparison of mean value of seven VIs
for the LAI classes calculated from linear and multiple regression model respectively.
TVI is scaled by dividing the original index by 100 for presentation purpose. In Figure
5.28, various VIs have a positive relationship with increasing LAI (from class 1 to 4).
Sensitivity of all VIs is high when LAI is low (class 1 and 2) indicated by the steep
slope. However, the sensitivity to LAI is lessened after LAI class 2 described by the
reduced slope for all VIs. However, the magnitude of decline varies across VIs. NDVI
has the most pronounced change of the slope which suggests significant sensitivity
decline occurs. Difference in mean value lowers to 0.035 when LAI increases from
class 3 to 4 compared with 0.219 when increases from class 1 and 2. The asymptotic
trend at high LAI suggests occurrence of saturation effect which has been reported
by many studies using NDVI for biophysical parameter mapping. RDVI and SAVI also
experiences decreasing sensitivity with increasing LAI though the magnitude is not
as significant as NDVI. Other VIs including MSAVI, TVI, MCARI 1 and 2 are able to
maintain their sensitivity from low to high LAI and MCARI 1 are much more
responsive to increasing LAI than the others.
In Figure 5.29, a pronounced saturation effect appeared in all VIs. The sensitivity to
the change of LAI begins to level off after LAI greater than 3.0 for NDVI, RDVI and
SAVI. The other four VIs show better trend and reach saturation level asymptotically
when LAI greater than 4.0 (class 4). This proves that the resistance to soil
reflectance and change of chlorophyll concentration can effectively enhance the
sensitivity of VI to LAI up to 4.
401
Figure 5.28. Mean value of various VIs for four LAI classes grouped according to the
predicted LAI (LAI2000G) from simple VI regression model (*TVI is scaled)
Figure 5.29. Mean value of various VIs for five LAI classes grouped according to the
predicted LAI (LAI2000G) from multiple regression model (*TVI is scaled)
402
In order to understand local variations between VIs and LAI, data were extracted
along two selected profiles shown in Figure 5.30. Both profiles locate in the central
part of the mangrove area where distinct species variation is found. Profile A was
chosen because of relatively homogeneous species variation across the region.
Classification result shows three species groups including A. marina, K. obovata G1
and G2 are found along profile A. Profile B is characterized by high species variation
with A. ilicifolius G1 and G2, K. obovata G1 and G2 and A. marina. The ranges of LAI
variation are 1.09-4.11 and 1.53-4.15 for profile A and B respectively.
Figure 5.30. The profiles cut across the central part of the mangrove area
403
Scatterplots depicting the relationship between LAI and various VIs extracted under
the two profiles are shown in Figure 5.31. Species based on the classification results
are represented by different colors. Along profile A, the sensitivity of VIs to LAI
change is strong represented by the strong r2 in the plots. Except NDVI and MCARI2,
all other VIs have strong positive relationship with r2 exceeded 0.9 significant at
p<0.01. Apparent variation of among the species is observed. K. obovata G1 has the
lowest LAI. LAI increases with A. marina and K. obovata G2 has the highest LAI. As
LAI is closely linked with VIs, such variation is also captured by various VIs. Besides,
a stronger linear relationship with VIs is observed when LAI is low except with NDVI.
As LAI increases, its strength of relationship with VIs decreases as shown by more
dispersed distribution in the scatterplots.
The strength of relationship between LAI and various VIs along profile B is not as
strong as that in profile A indicated by lower r2. NDVI has the worst sensitivity to
change of LAI while RDVI, SAVI, MSAVI and MCARI2 shows relatively strong
association with LAI as indicated by high r2 exceeding 0.8 significant at p<0.01. With
respective to species, A. ilicifolius G2 has high variability in terms of LAI while K.
obovata and A. marina have LAI pattern similar to profile A.
The relationship between LAI and VIs are quite different along the two profiles. A
more apparent relationship is found along profile A than along profile B. With more
uniform/ homogeneous species variation and therefore reflectance characteristics
along profile A, LAI estimation using VIs tends to be more reliable. Heterogeneity or
high variation of species causes significant variation of spectral reflectance.
404
Profile A
Profile B
r2=0.682**
r2=0.57**
r2=0.95**
r2=0.82**
r2=0.945**
r2=0.833**
r2=0.934**
r2=0.833**
405
r2=0.97**
r2=0.677**
r2=0.963**
r2=0.682**
r2=0.843**
r2=0.821**
Figure 5.31. Relationship between LAI and VIs across defined profiles A and B
406
5.5.4. Backscatter, texture measures and LAI
The relationships between radar parameters and LAI are examined by boxplots of
backscatter (in dB) and GLCM-based textural variables with change of LAI in Figure
5.32. LAI was extracted from the best multiple regression model combining TVI and
ASM with the lowest RMSE. The center line within the box represents the median of
the radar variable in a particular LAI class. The box length indicates the interquartile
range and the whiskers represent reasonable distance computed from the end of
the box. Median radar backscatter before and after filtering are shown in Figure
5.23 a and b respectively. No distinct trend is observed for raw backscatter while
the filtered backscatter shows a slightly downward trend with increasing LAI. The
difference of the median backscatter between LAI class 1 and 5 is about 3dB.
Besides, backscatter variation of the filtered backscatter is less significant than raw
backscatter in the LAI classes indicated by the shorter box length. LAI class 5 has the
lowest variation in terms of filtered backscatter.
It would be hard to explain the difference in backscatter between low and high LAI
due to the complex backscatter interaction within the canopy. A more reasonable
explanation is due to the contribution of background effect. As about 37% of
predicted LAI is below 2 in TVI-ASM model, the effects of soil background or
presence of undergrowth tend to be more significant in these areas. Radar
backscatter is sensitive to dielectric constant which is governed by the water
content. Theoretically, the relatively short wavelength of C-band (5cm) typically
interacts with the leaves, and small and secondary branches. The majority of
backscatter returns are contributed from volume scattering of canopy than from
soil surface and trunk. However, with relatively small incident angle (around 23°) of
ASAR data, the reduced path-length enhances the penetration ability and may
probably contribute to higher returns. Although the effect of standing water under
mangrove stands is more obvious with longer wavelengths due to greater
penetration through canopy, studies have shown backscatter enhancement in all
radar wavelengths (Ford and Casey, 1988, Drieman et al., 1989). Mangroves flourish
in the inter-tidal zone; soil is either fully saturated or flooded with water due to
daily inundation of seawater. In either situation, the soil contains high volumetric
407
moisture content causing high dielectric constant. Based on the closet tide
monitoring station in Tsim Bei Tsui, the tide level was 1.59 meters at the time of
data acquisition on 19 November and no rainfall was recorded on and 3-day before
the acquisition. It is highly probable that the soil was flooded with water. Besides,
the co-polarized signal has been found to have very significant trunk-ground
double-bounce (  d ) when LAI is low causing relatively high backscatter returns.
Wang and Imhoff (1993) have demonstrated that relatively higher backscatter
returns from flooded surface in smaller incident angle (less than 35°) in L-band is
observed and modeled when compared with non-flooded surface. The flooded
surface of high dielectric constant, which enhances the double-bounce trunkground interaction as it acts as a dihedral corner reflector. And the enhanced
backscatter tends to be stronger in like-polarization. The double-bounce scatter can
be greater than the volume backscatter from the canopy at small incidence angles
(Wang and Imhoff, 1993). Backscatter and LAI is negatively related with each other
along the two profiles. Stronger backscatter in sites of low LAI is partially due to
enhanced signal returns from the flooded soil surface. However, as demonstrated in
some studies, no bright returns were found from flooded forest. For instance, due
to extensive undergrowth causing significant attenuation, (Waite et al., 1981) found
that double-bounce returns were not apparent under flooded area.
Relationships between GLCM-derived textural variables and LAI are shown in Figure
5.23 c – j. Comparatively, the textural variables show much apparent median
variation with increasing LAI and the relationship tends to be non-linear. Among the
variables, homogeneity and angular second moment are positively associated with
increasing LAI while other textural variables have an inverse relationship with
increasing LAI. Apart from GLCM-contrast, all other variables show no saturation
behavior to high LAI.
For GLCM-homogeneity, the median increases with increasing LAI. Areas of low LAI
have relatively large backscatter differences resulting in low homogeneity (0.44)
while areas of high LAI are relatively more homogeneous (0.82). The relationship
between homogeneity and LAI exhibits an exponential trend.
408
GLCM-contrast is inversely related to LAI. High contrast of backscatter difference is
found in low LAI and the contrast decreases rapidly with increasing LAI. Besides, the
change begins to level off after LAI equals five. The variability of median GLCMdissimilarity across shows similar inverse trend with increasing LAI as GLCMcontrast but no saturation at high LAI value is found for the textural variable.
Both GLCM-entropy and GLCM-angular second moment are associated with the
orderliness of backscatter. The two textural variables are inversely related to each
other as indicated by highly negative correlation coefficient in correlation analysis
and this is also reflected in their relationship with LAI. With increasing LAI, entropy
shows an exponential decrease in median value while the median value of angular
second moment increases exponentially. It implies that areas of low LAI have a
much non-uniform backscatter and the non-uniformity reduces substantially with
increasing LAI. The two measures of orderliness were the only significant predictors
selected during stepwise regression modeling.
Certainly, a distinct difference of median of various textural variables is observed
with increasing LAI. However, most variables have a wide spanning of values across
the LAI classes. For instance, GLCM-homogeneity has a wide spread of values
between 0.34 and 0.54 in the interquartile range in LAI class 1 and the spread
narrows with increasing LAI (between between 0.78 and 0.84 in the interquartile
range in LAI class 5). The angular second moment textural parameter shows a
reverse trend with small variation in low LAI (interquartile range equals 0.03 in LAI
class 1) and the spread expands with increasing LAI (interquartile range equals 0.2
in LAI class 5). GLCM-entropy show consistently large spread of value regardless of
LAI class. Such a large variation or spread indicates vegetation/ canopy structures
other than LAI contribute to the backscatter properties and therefore textural
variation.
409
a
b
c
d
e
f
g
Figure 5.32. Boxplots of radar parameters for five LAI classes grouped based on the
predicted LAI (LAI2000G) from multiple regression model
410
In order to understand local variations of radar backscatter/ textural variables with
LAI, radar parameters were extracted along the same representative profiles shown
in Figure 5.30.
Scatterplots describing the relationship between LAI and selected radar parameters
extracted under the two profiles are shown in Figure 5.33. The points were colored
to represent species found along the profiles. In profile A, a strong negative
relationship is found between backscatter and LAI with r2 equals 0.609 significant at
p<0.01. Backscatter decreases with LAI. Besides, the three species groups have
distinct LAI variation and therefore variation with backscatter. K. obovata G1 has
relatively low LAI (high σ), followed by A. marina and K. obovata G2 has the highest
LAI (low σ) across the profile. LAI association with the three GLCM-derived radar
parameters is much weaker than with the backscatter. GLCM-angular second
moment is the only parameter that is positively related to LAI with r2 equals 0.306
significant at p<0.01. The other two have low and insignificant relationship with LAI.
Scatterplots from profile B shows similar trend but the strength of negative
relationship between LAI and backscatter is comparatively lower with r 2 equals
0.352 significant at p<0.01. LAI variation with the species with is still noticeable for K.
obovata G1, G2 and A. marina. However, the two A. ilicifolius groups have large
range of LAI variation. LAI along profile B has stronger relationship with GLCMderived textural variables described by high r2 equal 0.529, 0.62 and 0.608 for
angular second moment, homogeneity and entropy respectively and they are all
significant at p<0.01. The textural variables indicate backscatter spatial variation
becomes more a uniform and homogeneous as LAI increases. Species including K.
obovata G2 and A. ilicifolius have relatively high LAI among the species and appear
smoother while K. obovata G1 has low LAI and the texture is therefore rougher.
The two profiles show similar trend and relationship among species, LAI and radar
parameters. However, the strength of relationship varies. With large and
homogeneous distribution of species along profile A, the sensitivity of backscatter
to change of LAI is relatively better than that of the GLCM-derived variables. Along
profile B, backscatter return is also negatively related to LAI and the relationship
411
with LAI is enhanced in the GLCM-derived texture variables. The relatively higher
sensitivity of textural variables to LAI along profile B may be due to the high spatial
backscatter variation resulted from the diverse species composition. With
homogeneous species distribution along profile A, local backscatter variation is less
prominent, which curtail the sensitivity of the derived textural variables to LAI
change. Sites with high spatial variability of backscatter have better sensitivity to LAI
change in terms derived textural parameters. In other words, sensitivity to LAI is
enhanced by incorporating the textural parameters derived from backscatter. And
the spatial variability is largely related to change of canopy structure which is
partially resulted from change of species composition.
412
Profile A
Profile B
2
r =0.352**
2
r =0.609**
2
r =0.306**
2
r =0.529**
2
r =0.620**
2
r =0.103
2
2
r =0.12
r =0.608**
Figure 5.33. Relationship between LAI and selected radar parameters across defined
profiles A and B
413
5.5.5. Complementarity of Vegetation Index and Radar Parameters
Results from the stepwise regression model suggested that textural variables
derived from radar backscatter when coupled with VIs have significant impact on
LAI prediction. Two GLCM-derived variables – angular second moment (ASM) and
entropy (ENT) with F value significant at 0.05 level were selected as predictors in
stepwise regression. However, as the two variables are significantly correlated, ASM
is the only radar parameters retained in the multiple regression models. Generally,
multiple regression models have higher maximum predicted LAI compared with
simple linear regression model using VI alone. The absolute and percentage of r 2
and RMSE differences between the simple linear regression and multiple regression
analysis are shown in Table 5.12. LAI variations explained by regression models with
radar texture increase by 11 - 20% under combinations with various VIs and RMSE
reduced by 5 - 19% correspondingly. In terms of absolute LAI, the under or overprediction are reduced by 0.01 - 0.05. The largest decline of estimation error is
found in combining ASM with TVI (11 - 18%) and also with MCARI 1 (12 - 19%). The
two VIs also produce the best prediction models with the lowest RMSE in the linear
regression analysis.
Table 5.12. Coefficients of determination (r2) and RMSE differences between simple
linear and stepwise models
Absolute r2
difference
LAI (Bon)
LAI
(2000)
RDVI
SAVI
TVI
MCARI 1
MCARI 2
RDVI
SAVI
TVI
MCARI 1
MCARI 2
0.07
0.07
0.10
0.11
0.09
0.07
0.07
0.10
0.11
0.08
Absolute
RMSE
difference
0.02
0.01
0.03
0.04
0.03
0.03
0.03
0.05
0.05
0.03
% r2
difference
% RMSE
difference
11%
11%
16%
18%
20%
12%
11%
15%
16%
17%
5%
5%
11%
12%
7%
10%
10%
18%
19%
9%
414
LAI
(2000G)
RDVI
SAVI
TVI
MCARI 1
MCARI 2
0.08
0.08
0.10
0.11
0.09
0.02
0.02
0.04
0.04
0.02
13%
12%
14%
16%
18%
10%
10%
16%
17%
8%
Apart from statistical comparison, predicted LAI (LAI2000G) from linear and multiple
regression models with TVI are compared visually. Generally, the spatial variations
of LAI do not have significant change after addition of ASM to the VI model. Areas of
significant change are clipped and the corresponding LAI profiles are extracted from
the TVI model and compared in Figure 5.34. Extract A locates in the northern part of
the Mai Po. The incorporation of ASM texture raises significantly the LAI of some
locations. An apparent high LAI zone locates in the center of the image. The profile
shows that an area has LAI reaches above 4.5 in A2 while LAI in the corresponding
area is below 2.5 in A1. The enhanced area has high value in terms of ASM and
therefore uniformity. Other than this specific area, spatial variation of LAI in other
areas has no significant difference. Extract B locates in the central part of the
mangrove area where large homogeneous species distribution is found. In B1, the
relationship between LAI and species distribution is again apparent. Addition of
ASM into the model has increased LAI value K. obovata by around 0.5 while LAI of A.
marina stays about the same as shown in the profile in B2. LAI variation between A.
marina and K. obovata G1 is narrowed while difference between K. obovata G2 and
the other two species is enhanced. Extract C locates in the southern part of the
mangrove area. Similar to A and B, the chosen profile cut through the distinct high
LAI areas. LAI prediction from multiple regression analysis has increased the LAI of
A. ilicifolius G2 by a maximum of 1.5 while LAI of A. ilicifolius G1 remains about the
same as the prediction from the simple linear model. Comparing the profiles in C1
and C2, A. ilicifolius G2 has relatively higher LAI than A. ilicifolius G1 in this particular
location.
415
Simple Linear Regression with TVI
Multiple Regression with TVI and ASM
A2
A1
5
5
4.5
4.5
4
4
3.5
3.5
3
3
2.5
2.5
2
1
1.5
Ko 1
0.5
B1
B2
4.5
4.5
4
4
3.5
3.5
3
3
2.5
2.5
Ko 2
1.5
1
Am
2
Ko 1
1
0.5
0
0
C1
C2
4.5
4.5
4
4
3.5
3.5
3
3
2.5
2.5
2
2
1.5
1.5
Ai 2
Ai 1
Ko 2
1.5
0.5
1
Ko 1
1
0.5
0
0
2
Ai 2
2
Ai 2
1.5
1
0.5
0.5
0
0
Ai 2
Am
Ko 1
Ai 1
Figure 5.34. LAI profiles extracted from TVI linear and multi-regression model
416
Based on the above results, a few observations can be drawn from the use of VIs
and radar parameters for LAI estimation. In terms of VIs, they have strong and
robust relationship with LAI as indicated by the high correlation coefficient and
amount of variance explained in model development. The strong association was
also proved by high r2 from the extracted profiles after model prediction. Although
the performance of VI is satisfactory, major constraints of using remotely-sensed
data or their derivatives for accurate LAI estimation are still inevitable. Similar to
other studies, three major constraints are found. First, the relationship between LAI
and VIs are non-linear indicated by the changing slope. Generally, LAI increases with
VI but comparatively steeper slope is found in low LAI while the slope diminishes
gradually with increasing LAI. As a result, use of simple linear model for LAI
estimation would probably induce significant errors when LAI varies across a certain
range.
Second, when LAI exceed a certain levels, vegetation indices would approach
saturation level asymptotically. NDVI is the most well-known and intensively used VI
that contrasting the lowest red reflectance caused by maximum pigment absorption
and the highest near infrared reflectance due to the cellular structure. In current
study, NDVI approaches saturation levels when LAI exceed 3, which is comparatively
inferior to other VIs. The poor relationship between NDVI and mangrove LAI is
primarily because the mangrove habitats are always characterized by high biomass
and leaf areas. Saturated at low biomass, NDVI is not a recommended predictor for
mangrove. However, indices involving the combination of red and near infrared for
LAI estimation would face the problem of saturation no matter how they are
modified or transformed though some of them can raise sensitivity to higher LAI.
For instances ,the chosen indices including MSAVI, TVI, MCARI 1 and 2 have better
sensitivity to high LAI than NDVI, RDVI and SAVI. It proves that modification of VI to
be resistant to soil and change of chlorophyll content can enhance sensitivity to LAI
indicated by steeper slopes in Figures 5.28 and 5.29 and can lessen the saturation
behavior.
The third constraint revolves in the impossibility to design a VI that can be totally
resistant to all other vegetation parameters and sensitive only to LAI or any other
417
interested variables. In other words, the relationship between LAI and the VIs in this
study is not unique due to the fact canopy reflectance extracted from remotelysensed data is a function of many factors such as viewing geometry, biochemical
properties, LAI, atmospheric effect, etc. VIs derive based on canopy reflectance are
designed to suppress the effects of some of these confounding factors. However,
variation of other factors causes a scatter relationship between VI and LAI. For
instances, Haboudane et al. (2004) pointed out that effect of LAI and chlorophyll
content are similar in spectral region between 550nm and 750nm. Therefore,
MCARI 1 was designed to uncouple the jointed effect of LAI and chlorophyll
variation. MCARI 2 has even taken a step further to uncouple the contamination
influence of soil background. However, results from the study shows that MCARI 1
produces better regression model than MCARI 2 indicated by higher r 2 and lower
RMSE. MCARI 2 that considers soil property together with chlorophyll content
deteriorates its relationship with LAI.
From radar perspective, the prime advantage of radar for LAI estimation is because
of its ability to penetrate into vegetation canopy. The C-band SAR data carry mainly
information about the top canopy layer including leaves and small and secondary
branches. With vertical-polarized backscatter, it is supposed that the radar signal is
particular sensitive to LAI estimation (Leckie and Ranson, 1998). However, data
exploration using correlation analysis shows that field-measured LAI has very low
and insignificant correlation with raw backscatter. With enhanced contrast between
features after spatial filtering, the filtered backscatter shows slightly better trend
with LAI though the correlation is also low though significant. It is expected that the
structural change related to tree height, species composition, canopy thickness
would induce significant change in canopy roughness, which would consequently
reflect by textural images. However, textural variables including homogeneity,
mean, entropy and angular second moment extracted from radar backscatter also
have low but significant correlation with LAI in the model building. The low
correlation implies the use of radar data alone would not be able to produce
accurate estimation of LAI as indicated by low r2 and insignificant  coefficient in
418
Figure 5.14. Negligible relationship between LAI and C-VV backscatter is also
observed by other studies (Kovacs et al., 2008, Gao et al., 2010).
Although using radar parameters alone cannot retrieve LAI with satisfactory
accuracy, combining VI and GLCM-derived textural measure, ASM has produced the
best model with the lowest RMSE and the highest r2. Based on LAI predicted by the
best model, apparent relationships are found with the radar parameters.
Relationship between LAI and GLCM-derived textural variables from radar data
shows non-linear relationship as shown in Figure 5.32. Areas of low LAI have
relatively higher contrast and lower uniformity suggesting a rougher surface. With
low LAI, the penetrability of radar signal tends to be deeper and interaction with
undergrowth, aerial root, trunk or even the ground would probably occur. Trunkground double bounce, canopy-ground multi-path scattering, direct scatter from
ground surface and interaction with undergrowth complicate the understanding of
total backscatter. Moreover, leaves and other vegetative elements act as scatterers
and attenuators to radar signals and the magnitude of scattering and attenuation is
dependent on many factors such as incident angle, orientation or leaves and
branches, polarization of radar signal, etc. The resultant backscatter has a much
higher variability due to the interaction of many components and factors. It is not
the objective of this study to decompose the contribution of each scattering
components. However, as observed in Figure 5.32b, the backscatter is slightly
higher when LAI is low. The scatterplots of extracted profiles in Figure 5.33 also
showed the same negative relationship. With denser foliage, the penetrability is
comparatively less prominent with radar signal interacts mainly with the canopy
layer composed of leaves and small branches. Backscatter is dominated primarily by
volume scattering and subject to less or negligible influence of vegetative elements
beneath the canopy, the backscatter is therefore less variable and resulting in a
smoother surface. However, while scattering by other components are unimportant
due to weak penetrability and strong canopy attenuation of the incident wave, the
backscatter tends to be lower when LAI is high.
GLCM-derived angular second moment (ASM) was selected to improve the accuracy
of LAI in stepwise multiple regression process. When LAI was regressed against VI
419
and ASM, predicted LAI has generally augmented by a range of 0.02 - 0.3 dependent
on species. Sonneratia spp. is only species experience LAI drop. The effect of ASM
on K. obovata G2 and A. ilicifolius G2 is more prominent than on other species. In
general, LAI is positively related to the uniformity of backscatter value as indicated
by increasing ASM. It is believed that the associations between LAI and ASM or
other textural measures are probably due to change of canopy characteristics/
structure. As measures of local backscatter variability, GLCM-derived textural
variables are able to capture the spatial variation of mangrove canopy structure. As
LAI changes, canopy structure, backscatter and derived texture varies accordingly.
However, the extracted profiles in Figure 5.33 demonstrated the strength of
relationship between predicted LAI and ASM are not consistent throughout the area.
In order to understand the importance of textural measures derived from
backscatter (BS), GLCM-derived angular second moment (ASM) was computed for
the seven vegetation indices (VIs) using the same displacement vectors. Simple
correlation analysis was conducted to explore the relationship between the two
types of data. Table 5.13 shows the Pearson correlation between ASM derived from
the BS and VIs.
Table 5.13. Correlation between GLCM-derived Angular Second Moment (ASM)
from backscatter (BS) data and different vegetation indices (VIs)
ASM (BS)
ASM
(NDVI)
ASM
(RDVI)
ASM
(SAVI)
ASM
(MSAVI)
ASM
(TVI)
0.003
-0.185
-0.188
-0.224*
-0.164
ASM
ASM
(MCARI 1) (MCARI 2)
0.117
-0.137
* Correlation is significant at the 0.05 level
The low correlation between ASM derived from BS and VIs reinforced the
independence in terms of information inherent in radar and spectral data. Apart
from correlation, stepwise regression analyses were conducted with additional
input of ASM derived from VIs to the initial list of independent variable. The
resultant models were exactly the same as those shown in Table 5.9, i.e. individual
420
VI and ASM derived from BS were selected as predictors for every single model.
These two simple analyses revealed that textural variable extracted from radar data
offer extra information for LAI estimation. Such textural information is unique and
they tend to be more significant in explaining the variation of LAI than the textural
variables derived from the spectral VIs.
Similar to VIs, relationship between LAI and backscatter/ GLCM-derived parameters
is prone to be non-linear. However, the majority of GLCM-derived parameters are
less susceptible to saturation problem. For instances, LAI is associated with ASM in
an exponential manner. The sensitivity of ASM to increase of LAI is relatively small
when LAI is low and it augments with increasing LAI. The phenomenon is just
opposite to VI which experiences diminishing sensitivity as LAI increases. From this
perspective, hyperspectral and SAR data are compatible with each other for LAI
estimation. Integration of textural parameters with VI has the potential to raise the
accuracy of LAI estimation.
5.6. Summary
To summarize, this chapter demonstrates the results of LAI estimation through both
linear and multiple regression model analyses. Gap fraction measured from field
survey were converted to LAI using three methods including Bonhomme and
Chartier’s (LAIBon), LAI2000 Plant Canopy Analyzer (LAI2000) and LAI2000
Generalized (LAI2000G). Among the three methods, the LAI2000G has better
relationship with all VIs except with NDVI.
Simple linear regression models formulated using different VIs as independent
variables showed various degrees of success. In terms of error and strength of
relationship with LAI, simple NDVI model has the worst performance in LAI
estimation with RMSE ranged from 0.416 to 0.468 and r2 varied between 0.004 and
0.028 under different methods of LAI computation. The TVI model has
421
comparatively better performance than its counterparts with RMSE varied between
0.237 and 0.299 and r2 ranged from 0.613 to 0.689.
As correlations between LAI and individual backscatter/ GLCM-derived textural
parameters are low, use of radar parameters alone results in poor estimation.
Individual VI was combined with the set of radar parameters based on which
stepwise regression was run to select variables that are significant to predict LAI
variation. For all models, the VI and GLCM-derived angular second moment (ASM)
were selected as predictor variables except when NDVI and MSAVI were input.
Generally, the overall predicted LAI have been increased for the species except
Sonneratia spp. Two models namely, TVI-ASM and MCARI1-ASM have the best
performance with comparably high r2 ranged from 0.71 to 0.79 and low RMSE
varied between 0.2 and 0.27. Compared with simple linear model, r2 has been
increased by 11-20% while RMSE has been reduced by 5 -19%. With enhancement
through textural extraction, ASAR is promising for LAI estimation.
By relating the predicted LAI over the study area to hyperspectral bands, VIs,
filtered backscatter and GLCM-derived textural variables, some interesting
observations were revealed. First, a negative relationship was observed between
LAI and mean reflectance in the visible bands while the relationship becomes
positive in the infrared regions. Second, the saturation problem of various VIs is
quite apparent when LAI is greater than 4. The effect is more prominent for NDVI
with saturation begins at around 3. Third, the sensitivity of filtered backscatter to
change of LAI is not as strong as GLCM-derived textural variables. Textural variables
showed generally non-linear relationship with LAI but with no effect of saturation
found except the GLCM-contrast variable. However, the strength of association is
not consistent in the area. Fourth, different species show apparent LAI variation.
Species ranked in descending order of LAI are K. obovata G2, A. ilicifolius G1, A.
ilicifolius G2, A. marina, K. obovata G1 and Sonneratia spp.
422
CHAPTER 6
CONCLUSION
6.1. Summary of the Study
This study combined in situ field measurements and remotely-sensed data to study
the largest mangrove stands in Mai Po. The study has two main focuses. The first
concerns mangrove species-based classification through integrating hyperspectral
and multi-temporal SAR data. Wrapper-based feature selection was conducted first
to select bands/ features that are prominent to discriminating species. Based on
feature ranking according to the pre-defined criteria, four feature subsets were
generated. Species-based classification was performed by combining field data and
various feature subsets using four classification algorithms. Classification accuracy
in terms of feature subsets and classifiers was assessed and compared. The second
focus involves leaf area index (LAI) estimation through regression using vegetation
indices (VIs), radar backscatter and GLCM-derived textural measures. Hemispherical
photographs were captured through field survey campaign based on which LAI was
computed as dependent variable. As for independent variables, spectral bands
based on the feature ranks after feature selection process were combined to form
seven VIs while GLCM-derived textural measures were computed from the radar
backscatter (σ). Simple linear regression was first conducted to regress LAI against
various VIs. Stepwise multiple regression was then implemented to select radar
parameters combining with individual VIs. Based on the results from classification
and LAI modeling, the complementarity of spectral and radar data was explored.
Laboratory-measured leaf spectra and spectra extracted from Hyperion data
showed low correspondence of wavelengths at which the mangrove species are
effectively discriminated. Early studies have primarily used laboratory spectra to
reveal the potential of narrowbands in differentiating vegetation species. However,
the low correspondence between the two sources of data in current study
suggested the sole dependence of laboratory-measured spectra is unable to match
with the satellite data. External factors such as leaf clumping and orientation are
423
influencing the spectral responses of the species, which in turn affect the
discriminability as compared with measurement taken from individual leaf.
The presence of more than two hundred bands in hyperspectral data renders it
impossible to apply optimal search approaches such as exhaustive search or branch
and bound to select relevant bands for vegetation classification. Facing such a high
data volume, suboptimal search approaches become the most plausible solution to
deal with the curse of dimensionality. Current study has demonstrated the use of
simple search algorithms developed from computer science including SFS, SFFS and
OS to effectively single out spectral bands that can maximally discriminate the
mangrove spectral classes based on evaluating the classification accuracy of feature
subsets. Feature selection shows that narrowbands locating in spectral regions of
green at 570nm, 580nm, 591nm, 601nm, red at 702nm, red-edge at 713nm, near
infrared at 764nm, 774nm 1119nm and 1276nm and shortwave infrared at 1316nm
and 1629nm. The process of feature selection is fast and straightforward. The mean
training and testing accuracy of subsets reach 90% and 80% respectively suggesting
the selected features are promising in terms of species classification. In terms of
spectral data, results of feature selection contributed to the understanding of
wavelengths that are critical to mangrove species discrimination on one hand and
to identify common wavelengths that are essential for vegetation discrimination
regardless of vegetation types on the other. Vegetation studies involving feature
selection of high-dimensional spectral space and sophisticated search algorithms
have identified common bands locating at or near 580nm, 702nm, 774nm for
effective vegetation discrimination, which were also recognized in current study.
Apart from common wavelengths, bands locating at the steepest climb of the red
edge at 713nm and those in the NIR and SWIR regions (1119nm, 1276nm, 1316nm
and 1629nm) are unique bands identified to distinguish different mangrove species
from this study. Selected wavelengths at red edge and NIR imply that leaf internal
structure is an important characteristic for species discrimination. The two selected
SWIR bands locating in the former portion of SWIR region indicated water status
can be another major factor discriminating the species. Apart from the spectral data,
multi-temporal filtered backscatter captured on 19 November, 13 February, 19
424
March and 06 August were also identified as sensitive to mangrove species
discrimination.
Classification accuracy varies across feature subsets and classifiers. In terms of
feature subsets, use of multi-temporal radar backscatter features alone results in
the poorest accuracy regardless of classifiers. Under same feature size, pure input
of spectral features have better overall accuracy than mixed input of spectral and
backscatter features in all classifiers except under artificial neural network (ANN).
The entry of all nine features resulted in the best overall accuracy. In terms of
classification algorithm, all classifiers had similar variations of training accuracy
under the same feature subset. However, accuracy based on independent testing
dataset showed particularly large discrepancy in training accuracy under decision
tree C5.0 (DT) classifier. Results from maximum likelihood (ML) and support vector
machines (SVM) also experienced relatively lower testing accuracy though the
magnitude was not as strong as DT. ANN was the only classifier having testing
accuracy better than training accuracy. Performance of ANN was comparatively
more stable and robust than other classifiers while overtraining occurred for the DT
classifier. Among the species, A. corniculatum and Sonneratia spp. showed large
oscillation of accuracy in the classifiers due to deficiency of training samples. For
other species, they showed satisfactory results as revealed by the producer’s and
user’s accuracy.
Simple linear regression model with VIs revealed that triangular vegetation index
(TVI) and modified chlorophyll absorption ration index 1 (MCARI1) had the best
relationship with LAI indicated by the highest amount of variance explained and the
lowest estimation error. These two vegetation indices are designed to be less
sensitive to change of biochemical effect, mainly chlorophyll that influences LAI
prediction. Instead of arbitrarily selected individual narrowbands in the
hyperspectral spectrum, current study computed the vegetation indices based on
the result from feature selection. Although the primary objective of feature
selection was not designed for LAI estimation, the relationship between the VIs
formulated using the selected features and LAI is promising. Conversely, the worst
relationship was found between the widely-used normalized difference vegetation
425
index (NDVI) and LAI. This is mainly due to poor sensitivity of NDVI to high LAI which
is found in the mangrove ecosystem. For all VIs, they are positively related to LAI.
The simple linear regression also showed that radar parameters including C-VV
backscatter and GLCM-derived textural measures had weak association with field
measured LAI. Use of radar parameters alone did not result in good LAI estimation.
Results from stepwise multiple regression suggested that combination of individual
VIs with GLCM-derived angular second moment (ASM) enhanced LAI prediction by
reducing the estimation error. Instead of using the raw or filtered backscatter (BS)
data, the derived textural parameter, ASM describing the spatial variability of BS
provided additional information for accurate LAI estimation. It is well-understood
that backscattering coefficients provide particular details about the vegetation
canopy structure while ASM extracted from BS data are also unique compared with
ASM computed from VIs. The additional details can therefore improve the
explanation of LAI variation. LAI estimation from the best model integrating TVI and
ASM revealed the adjusted VIs shown saturation behavior when LAI is higher than 4
while NDVI begins to approach saturation level asymptotically after LAI larger than
3. In terms of radar parameters, LAI and backscatter returns are negatively related,
i.e. lower LAI results in higher backscatter due to the contribution of scattering from
ground and trunk beneath the canopy. Besides, areas of low LAI tend to have
relatively higher contrast and lower uniformity of backscattering suggesting a
rougher surface captured by the textural measures. By comparing LAI variation and
species distribution, the species have apparent LAI variation with K. obovata G2
have the highest LAI, followed by A. ilicifolius G1, A. ilicifolius G2, A. marina, K.
obovata G1 and Sonneratia spp.
The study has demonstrated spectral and radar data are complementarity to each
other for accurate species discrimination and LAI mapping. Although radar
possesses important characteristics such as weather-independent and signal
penetration capability which is superior to spectral data for broad-scale mangrove
monitoring, the results suggested that it would be difficult to produce confident
species map and LAI estimation with radar backscatter data and the textural
derivatives alone. Data derived from radar sensor contribute to better species
426
classification and accurate LAI estimation only if they are used complementarily
with the optical satellite sensor. The multi-temporal radar data improves
classification accuracy as it captures the variability of some species pairs that are
not found in the spectral data. Besides, as C-band is sensitive to canopy structural
variations, the multi-temporal SAR data capture the seasonal variation due to
structural change of various species which can somehow enhance the mapping
accuracy. Similarly, derived textural measure, ASM, lowers the estimation error and
improves the relationship with measured LAI. It also highlights some areas/ species
where enhancement can be made. However, as both spectral and radar data are
necessary for accurate estimation, the role of satellite radar as continuous and
single data source for monitoring has been largely curtailed.
6.2. Limitation of the Study
The limitation of the study falls into three areas concerning data acquisition, data
processing and data analysis.
The problem of data acquisition is related to both field measurement and remotelysensed SAR data. The study has mainly conducted field campaigns for leaf collection
as well as LAI measurement. Due to highly-restricted accessibility to the mangrove
stand in Mai Po, field data acquisition has suffered from tremendous restrictions.
Site remoteness is the very first constraint, which hinders the representativeness of
leaf samples collected. In order to minimize the time between leaf sample
collection and spectral measurement in laboratory, instead of collecting leaf
samples along well-planned transects, they were collected along the two sides of
the only accessible footbridge. It may argue that leaf samples collected along the
active-growing edge were not representative enough. However, if leaf samples
were collected deep in the mangrove stands, the leaves can hardly be kept in fresh
status which would affect the results of spectral measurement. Besides, additional
tools were required to get the leaves from the top canopy. As the required number
of leaf samples was quite substantial, collection along the most assessable path is
427
the only solution that can compensate site remoteness due to resource constraint.
Besides, the results also show distinct difference among species.
The accessibility was largely constrained by the tidal level and unstable substratum.
The variation of tide level has restricted the time when field work can be carried out.
During medium or high tide, the floor was inundated and LAI measurement was
prohibited. Besides, the fully-saturated substratum was so unstable that rendered
the setup of hemispherical camera perfectly level to the ground difficult and
inefficient. In some areas, the presence of extensive undergrowth especially along
the old river channels and the species boundary has hampered the access. No
measurement was taken in sites of K. obovata G2 close to the fringe area due to the
accessibility and safety concerns. Besides, A. ilicifolius was not covered in the LAI
campaign because they are mostly presented as undergrowth and relatively short in
height. Under such circumstance, the hemispherical camera was limited in its
capability to capture the gap fraction of the species in an efficient and reliable
manner. Restricted accessibility to areas located along the old river channels and
the coastal fringe, where A. ilicifolius is presented also hinders effective LAI
measurement.
Apart from field measurement, the unavailability of multi-polarization SAR data has
hindered further research into the understanding of relationship between
backscatter and mangrove. C-VV is the only polarization mode available in archived
database from ISEIS, CUHK that has similar incident angle and along the same
satellite track. Low correlation with field measured LAI suggested that VV may not
be the best polarization to detect LAI change.
Two possible limitations are related to data processing. The first problem related to
atmospheric correction of Hyperion image. The atmospheric correction process has
introduced negative apparent reflectance in blue bands and in some of the
shortwave infrared bands. The two atmospheric correction algorithms, FLAASH and
ATCOR2 have reflected the same negativity problem. Tremendous efforts have been
spent on testing and adjusting the parameter settings in FLAASH. It was observed
that the lower of CO2 mixing ratio to 370-380 ppm can improve the values, but
428
negativity still exists. As the negative reflectance in blue bands between 428 –
469nm were the most serious, they were removed from further consideration.
Another possible problem related to data processing falls into the determination of
gap fraction during hemispherical image processing. Precise determination of gap
fraction is crucial to LAI computation. However, finding an optimal threshold
separating sky from vegetative elements in the photographs has been vigorously
discussed in the literatures. The task is not simple and is further complicated by
possible errors on image sharpness, chromatic aberration, and blooming introduced
by digital camera during the image capture process. The commonly used LAI-2000
Plant Canopy Analyzer (LAI-2000 PCA) have been considered and tested prior to
using the hemispherical technique. However, it was not adopted due to site
limitations posed by the study area. Acquisition of above-canopy reading required
by LAI-2000 PCA was considered to be very difficult as the mangroves are quite tall
and it is also hard to find an open area for simultaneous sky radiation measurement.
As the constraints were well-understood, an analyst was trained to define the
threshold as consistent as possible.
Further limitation was posed by incapability to offer a further understanding and
validation of predicted results. Spectral reflectance derived from remotely-sensed
data is a function of many factor related to leaf composition, canopy structure and
sensor geometry. LAI is just one of the many factors affecting the received
reflectance. Although it was found that the sensitivity of enhanced VIs after
suppressing the soil background or chlorophyll variation to LAI is comparatively
better, the accurate effect or magnitude was still not understood through
regression modeling. Correspondingly, backscatter derived from SAR also influenced
by a number of factors such as canopy structure, water content and surface
background. Theoretically, C-band radar signals interact mainly with the canopy
layer, which is believed to be of high correlation with LAI. However, the relatively
low correlation between radar backscatter and LAI suggests other factors are
contributing to backscatter. Parameters relating to canopy structure, sensor
geometry and ground environment all affect the magnitude of backscatter making
the understanding of relationship between mangrove and backscatter difficult.
429
Under low LAI and steep incidence angle (20° - 30°), radar penetration can be
significant even in C-band. The penetration capability causes radar interaction with
trunks, aerial roots and ground surface and renders the understanding of
relationship between radar backscatter /textural parameters and LAI of vegetation
canopy never be straightforward. Besides, locating in tidally inundated region,
constant flooded and non-flooded variations beneath the canopy presents a major
challenge for this study and in many studies as well. Another possible complication
lies in the presence of salt-excreting glands in the leaves of some mangrove species,
which would affect canopy extinction coefficient and therefore the penetration
ability of radar signal. However, it is beyond the scope of this study to resolve the
causes of backscatter variation though they are important to understand the
connection between mangrove and radar backscatter.
Spatial resolution is another major constraint facing the study as it would affect the
accuracy of species and LAI estimation as well as hinders clear interpretation of
results. With a spatial resolution of 30-meter, problem of mixed pixels are
commonly found along boundaries along the coastal fringe, river channels, close to
the land boundary and where species change. Species locating along the coastal
fringe and the river channels are highly affected by the water. With the effect of
water, spectral reflectance would be lower while radar backscatter returns may be
enhanced due to significant dihedral corner scattering for mangrove locating along
the coastal fringe. Its Impacts on species classification and LAI estimation were
expected. For instances, predicted LAI along coastal and land boundaries were
comparatively lower. In areas where heterogeneous species are found or along
boundary where species alter, uncertainty caused by mixed pixel in species
classification is relatively higher than that of LAI estimation. Species having
scattered distribution such as A. corniculatum and Sonneratia spp. cannot be
mapped with sufficient confidence.
430
6.3. Recommendation
Based on the results, recommendations are offered in two areas. The first concerns
the refinement of methodology to further enhance the understanding of feature
selection techniques and relationship between LAI, backscatter and spectral data.
The second concerns the implications and suggestions based on the results of the
study.
As more and more information is available, pattern recognition becomes an
important and indispensible tool to sort out redundant details and retain the
relevant ones. In the majority of pattern recognition problems, the original feature
space ( D ) would be too large for optimal search to be applied. As sub-optimal
feature search would always become the only solution, selection of search
algorithms and evaluation criteria is critical to the final feature subset. The relatively
fast and straight forward search algorithms coupled with classification algorithms
have demonstrated the success of selecting spectral and radar features for species
classification and LAI modeling in current study. It is also worthwhile to compare
other feature selection techniques such as backward search, SVM-recursive feature
elimination and genetic algorithms in terms of their efficiency, characteristics of the
feature subset(s) and accuracy of classification. Apart from feature selection,
feature extraction is another realm that received much attention though the
extracted features are sometimes difficult to be interpreted.
Radar presents both opportunities and challenges to mangrove species
classification and LAI estimation in this study. It is worthwhile to explore further the
potential of multi-polarization and multi-sensor SAR data in improving classification
and estimation results while in the meantime to enhance understanding of
mangroves’ response to radar signal. It is understood that different polarizations
such as HH, VH and HV have varied sensitivity to LAI. Some studies showed that HH
has strong correlation with LAI. Besides, with multi-polarization data, formulation of
polarization ratios is possible. These polarization ratios are able to lessen the effect
of forest structure as demonstrated by some studies. Apart from multi-polarization,
further enhancement can be achieved by incorporating multi-sensor data. The
431
PALSAR sensor onboard the Advanced Land Observing Satellite (ALOS) operating at
L-band is a potential option. The longer wavelength allows a much higher
penetration capability and sensitivity to above-ground biomass. Besides, its finer
spatial resolution offers a more detail estimation. Provided with the results from
this study, mangrove forest can be a good testing ground for multi-polarization and
multi-sensor data.
Another possible enhancement to existing study is to compare LAI prediction from
the regression model with the estimation through physically-based canopy
reflectance (CR) model inversion method. Another approach is to validate the
estimated LAI by using it as an input parameter in forward modeling. The simulated
reflectance from the CR model can be compared with the reflectance derived from
the hyperspectral data. A typical CR model consists of both leaf optical model and
canopy reflectance model. For instance, the PROSAIL model which is the
combination of PROSPECT leaf optical model and SAIL (Scattering by Arbitrarily
Inclined Leaves) canopy bidirectional reflectance model has long been used to
retrieve the vegetation biophysical properties including LAI. However, the success
of LAI retrieval or validation using CR model is dependent on accurate estimation of
required model parameters. Parameters related to sun-sensor geometry are
relatively straightforward while the biochemical inputs such as water content,
chlorophyll a and b concentration, carotenoid concentration requires either precise
measurement or reasonable and valid assumptions. Applications of CR model in
broad-leaf or needle forest have been abundant but its application is relatively
limited in mangrove forest. Results from this study can act as foundation to test the
applicability of CR model in mangrove biophysical parameters estimation.
The second part of recommendation tries to draw some implications from the
results from the perspective of mangrove conservation and management. As
pointed out in the previous session, the mangrove stand in Mai Po is not only the
largest mangrove stand within Hong Kong’s border; it is also listed as Wetlands of
International Importance under the Ramsar Convention. This study has successfully
mapped the species distribution and estimated LAI as one of the most important
biophysical parameters with promising accuracy. The map of species distribution
432
does not only provide an up-to-date and complete inventory of mangrove species, it
also reveals the history of mangrove succession. For instance, for K. obovata, the
species mapping exercise has clearly separated the mature stands (close to the land
boundary) from the young stands (close to coastal fringe) from their spectral
characteristics and canopy structure captured by the spectral and radar data
respectively. The same happens to A. ilicifolius in which those found along the river
channels have different spectral and canopy characteristics from those located
along the coastal fringe. Although the spatial resolution is not high, boundary
delineating the species distribution is clear. When incorporated species distribution
with the result of LAI estimation, it is clear that species have distinct LAI variation. K.
obovata and A. ilicifolius locating close to the coastal fringes as well as A. ilicifolius
found along the river channels are of high LAI. This proves that the active-growing
relatively young species have high LAI. As LAI links closely with many parameters in
an ecosystem, higher leaf area will probably induce higher rate of photosynthesis,
respiration, transpiration and therefore net primary production of the ecosystem.
Therefore, prioritized and active conservation measures should be granted to
protect these areas from disturbance. With low LAI found close to the land
boundary, they may probably suffer from deterioration or in a declining stage.
Although the decline may be a natural process, human-induced pollution such as
accumulation of heavy metal may have unforeseeable consequences. It would be
necessary to have a continuous monitoring of mangrove areas on yearly base. As
demonstrated in this study, data acquired by spaceborne sensors offers an effective
solution to monitor and understand the status of mangrove stands in Mai Po.
433
Reference
Aha, D. W. & R. L. Bankert. 1995. A comparative evaluation of sequential feature
selection algorithms. In Proceedings of the 5th International Workshop on
Artificial Intelligence and Statistics, eds. D. Fisher & H. Lenz, 1-7. Ft.
Lauderdale, FL.
Aha, D. W. & R. L. Bankert. 1996. A comparative evaluation of sequential feature
selection algorithms. In Learning from Data: Artificial Intelligence and
Statistics V, eds. D. Fisher & J.-H. Lenz, 199-206. New York: Springer Verlag.
Almauallium, H. & T. G. Dietterich. 1991. Learning with many irrelevant features. In
Proceedings of Ninth National Conference on Artificial Intelligence, 547-552.
Anaheim, CA: AAAI Press.
Amato, U., R. M. Cavalli, A. Palombo, S. Pignatti & F. Santini. 2009. Experimental
Approach to the Selection of the Components in the Minimum Noise
Fraction. IEEE Transactions on Geoscience and Remote Sensing, 47, 153-160.
Anderson, D. & G. McNeil. 1992. Artificial Neural Networks Technology. Rome: Data
& Analysis Center for Software.
Anderson, G. P., B. Pukall, C. L. Allred, L. S. Jeong, M. Hoke, C. J. H, S. M. AdlerGolden, A. Berk, L. S. Bernstein & S. C. Richtsmeier. 1999. FLAASH and
MODTRAN 4 - State-of-the-art atmospheric correction for hyperspectral data
In IEEE Aerospace Conference, 177-181. Aspen, CO; UNITED STATES.
Aschbacher, J., R. Ofren, J. P. Delsol, T. B. Suselo, S. Vibulsresth & T. Charrupat. 1995.
An integrated comparative approach to mangrove vegetation mapping using
advanced remote sensing and GIS technologies: preliminary results.
Hydrobiologia, 295, 285-294.
Asner, G. P. 1998. Biophysical and biochemical sources of variability in canopy
reflectance. Remote Sensing of Environment, 64, 234-253.
Asner, G. P., J. A. Hicke & D. B. Lobell. 2003. Per-pixel analysis of forest structure:
Vegetation indices, spectral mixture analysis and canopy reflectance
modeling. In Remote sensing of forest environments : concepts and case
studies, eds. M. A. Wulder & S. E. Franklin, 209-254. Boston: Kluwer
Academic Publishers.
Asner, G. P., C. A. Wessman, C. A. Bateson & J. L. Privette. 2000. Impact of tissue,
canopy, and landscape factors on the hyperspectral reflectance variability of
arid ecosystems. Remote Sensing of Environment, 74, 69-84.
Atkinson, P. M. & A. R. L. Tatnall. 1997. Introduction: Neural networks in remote
sensing. International Journal of Remote Sensing, 18, 699-709.
434
Atzberger, C. 2004. Object-based retrieval of biophysical canopy variables using
artificial neural nets and radiative transfer models. Remote Sensing of
Environment, 93, 53-67.
Bacour, C., S. Jacquemoud, Y. Tourbier, M. Dechambre & J.-P. Frangi. 2002. Design
and analysis of numerical experiments to compare four canopy reflectance
models. Remote Sensing of Environment, 79, 72-83.
Baghdadi, N., M. Bernier, R. Gauthier & I. Neeson. 2001. Evaluation of C-band SAR
data for wetlands mapping. International Journal of Remote Sensing, 22, 7188.
Bajcsy, P. & P. Groves. 2004. Methodology for hyperspectral band selection.
Photogrammetric Engineering and Remote Sensing, 70, 793-802.
Baraldi, A. & F. Parmiggiani. 1995. A refined Gamma MAP SAR speckle filter with
improved geometrical adaptivity. IEEE Transactions on Geoscience and
Remote Sensing, 33, 1245-1257.
Baret, F., J. G. P. W. Clevers & M. D. Steven. 1995. The robustness of canopy gap
fraction estimates from red and near- infrared reflectances: a comparison of
approaches. Remote Sensing of Environment, 54, 141-151.
Baret, F. & G. Guyot. 1991. Potentials and limits of vegetation indices for LAI and
APAR asssessment. Remote Sensing of Environment, 35, 161-173.
Baret, F. & S. Jacquemoud. 1994. Modeling canopy spectral properties to retrieve
biophysical and biochemical characteristics. In Imaging Spectrometry - A
Tool for Environmental Observations, eds. J. Hill & J. Megier, 145-167.
Dordrecht: Kluwer Academic Publishers.
Baret, F., M. Weiss, D. Troufleau, L. Precot & B. Combal. 2000. Maximum
information exploitation for canopy characterization from remote sensing:
radiative transfer model inversion and assimilation into canopy functioning
model. Aspects of Applied Biology, 60, 71-82.
Battiti, R. 1994. Using Mutual Information for Selecting Features in Supervised
Neural Net Learning. IEEE Transactions on Neural Networks, 5, 537-550.
Beaudoin, A., T. Le Toan, S. Goze, E. Nezry, A. Lopes, E. Mougin, C. C. Hsu, H. C. Han,
J. A. Kong & R. T. Shin. 1994. Retrieval of forest biomass from SAR data.
International Journal of Remote Sensing, 15, 2777-2796.
Beck, R. 2003. EO-1 User Guide v.2.3. U.S. Geological Survey, U.S. Government.
http://eo1.usgs.gov/documents/EO1userguidev2pt320030715UC.pdf (last
accessed September 10, 2007).
435
Becker, B. L., D. P. Lusch & J. Qi. 2005. Identifying optimal spectral bands from insitu measurements of Great Lakes coastal wetland using second derivative
analysis. Remote Sensing of Environment, 97, 238-248.
Belluco, E., M. Camuffo, S. Ferrari, L. Modenese, S. Silvestri, A. Marani & M. Marani.
2006. Mapping salt-marsh vegetation by multispectral and hyperspectral
remote sensing. Remote Sensing of Environment, 105, 54-67.
Belward, A. S. 1991. Spectral characteristics of vegetation, sol and water in the
visible, near infrared and middle-infrared wavelengths. In Remote Sensing
and Geographical Information Systems for Resource Management in
Developing Countries., eds. A. S. Belward & C. R. Valenzuela, 506. Dordrecht:
Kluwer Academic.
Ben-Bassat, M. 1982. Pattern Recognition and Reduction of Dimensionality. In
Handbook of Statistics Vol.2, eds. P. R. Krishnaiah & L. N. Kanal, 773-791.
Amsterdam: North-Holland Pub. Co.
Benediktsson, J. A., J. R. Sveinsson & K. Arnason. 1995. Classification and feature
extraction of AVIRIS data. IEEE Transactions on Geoscience and Remote
Sensing, 33, 1194-1205.
Benediktsson, J. A., P. H. Swain & O. K. Esroy. 1990. Neural network approaches
versus statistical methods in classification of multisource remote sensing
data. IEEE Transactions on Geoscience and Remote Sensing, 28, 540-552.
Bi, J., K. P. Bennet, M. Embrechts, C. M. Breneman & M. Song. 2003. Dimensionality
reduction via sparse support vector machines. Journal of Machine Learning
Research, 3, 1229-1243.
Bicheron, P. & M. Leroy. 1999. A method of biophysical parameter retrieval at
global scale by inversion of a vegetation reflectance model. Remote Sensing
of Environment, 67, 251-266.
Bishop, C. M. 1995. Neural Networks for Pattern Recognition. New York: Oxford
University Press.
Blackburn, G. A. 1998. Quantifying chlorophylls and carotenoids at leaf and canopy
scales: an evaluation of some hyperspectral approaches. Remote Sensing of
Environment, 66, 273-285.
Blackburn, G. A. & J. I. Pitman. 1999. Biophysical controls on the directional spectral
reflectance properties of bracken (Pteridium aquilinum) canopies: results of
a field experiment. International Journal of Remote Sensing, 20.
Blackburn, G. A. & C. M. Steele. 1999. Towards the remote sensing of Matorral
vegetation physiology: relationships between spectral reflectance, pigment,
and biophysical characteristics of semiarid bushland canopies. Remote
Sensing of Environment, 70, 278-292.
436
Blasco, F., T. Gauquelin, M. Rasolofoharinoro, J. Denis, M. Aizpuru & V. Caldairou.
1998. Recent advances in mangrove studies using remote sensing data.
Marine and Freshwater Research, 49, 287-296.
Blum, A. L. & P. Langley. 1997. Selection of relevant features and examples in
machine learning. Artificial Intelligence, 97, 245-271.
Boegh, E., H. Soegaard, N. Broge, C. B. Hasager, N. O. Jensen, K. Schelde & A.
Thomsen. 2002. Airborne multispectral data for quantifying leaf area index,
nitrogen concentration, and photosynthetic efficiency in agriculture. Remote
Sensing of Environment, 81, 179-193.
Bonan, G. B. 1993. Importance of leaf area index and forest type when estimating
photosynthesis in boreal forests. Remote Sensing of Environment, 43, 303314.
Bonhomme, R. & P. Chartier. 1972. The interpretation and automatic measurement
of hemispherical photographs to obtain sunlit foliage area and gap
frequency. Israel Journal of Agricultural Research, 22, 53-61.
Bonhomme, R. & J. Cihlar. 1995. The interpretation and automatic measurement of
hemispherical photographs to obtain sunlit foliage area and gap frequency.
israel Journal of Agricultural Research, 22, 53-61.
Boochs, F., G. Kupfer, K. Dockter & W. Kuhbauch. 1990. Shape of the red edge as
vitality indicator for plants. International Journal of Remote Sensing, 11,
1741-1753.
Bourgeau-Chavez, L. L., E. S. Kasischke, S. M. Brunzell, J. P. Mudd, K. B. Smith & A. L.
Frick. 2001. Analysis of space-borne SAR data for wetland mapping in
Virginia riparian ecosystems. International Journal of Remote Sensing, 22,
3665-3687.
Boyer, M., J. Miller, M. Belanger, E. Hare & J. Wu. 1988. Senescnce and spectral
reflectance in leaves in Northern Pin Oak (Quercus palustris Muenchh.).
Remote Sensing of Environment, 25, 71-87.
Braswell, B. H., D. S. Schimel, J. L. Privette, B. Moore, W. J. Emery, E. W. Sulzman &
A. T. Hudak. 1996. Extracting ecological and biophysical infomation from
AVHRR optical data: an integrated algorithm based on inverse modelling.
Journal of Geophysical Research, 101(D18), 335-348.
Breda, N. J. J. 2003. Ground-based measurements of leaf area index: a review of
methods, instruments and current controversies. Journal of Experimental
Botany, 54, 2403-2417.
Breiman, L., J. H. Friedman, R. A. Olsen & C. J. Stone. 1984. Classification and
Regression Trees. Belmont, CA: Wadsworth.
437
Brenner, A. J., M. Cueto Romero, J. Garcia Haro, M. A. Gilabert, L. D. Incoll, J.
Martinez Fernandez, E. Porter, F. I. Pugnaire & M. T. Younis. 1995. A
comparison of direct and indirect methods for measuring leaf and surface
areas of individual bushes. Plant, Cell and Environment, 18, 1332-1340.
Broge, N. H. & E. Leblanc. 2000. Comparing prediction power and stability of
broadband and hyperspectral vegetation indices for estimation of green leaf
area index and canopy chlorophyll density. Remote Sensing of Environment,
76.
Broge, N. H. & E. Leblanc. 2001. Comparing prediction power and stability of
broadband and hyperspectral vegetation indices for estimation of green leaf
area index and canopy chlorophyll density. Remote Sensing of Environment,
76, 156-172.
Broge, N. H. & J. V. Mortensen. 2002. Deriving crop area index and canopy
chlorophyll density of winter wheat from spectral reflectance data. . Remote
Sensing of Environment, 81, 45-57.
Brown, D. E., V. Corruble & C. L. Pittard. 1993. A comparison of decision tree
classifiers with backpropagation neural networks for multimodal
classification problems. Pattern Recognition, 26, 953-961.
Brown, L., J. M. Chen, S. G. Leblanc & J. Cihlar. 2000. A shortwave infrared
modification to the simple ratio for LAI retrieval in boreal forests: An image
and model analysis. Remote Sensing of Environment, 71, 16-25.
Bruzzone, L. & S. B. Serpico. 2000. A technique for feature selection in multiclass
problems. International Journal of Remote Sensing, 21, 549-563.
Burges, C. J. C. 1998. A tutorial on support vector machines for pattern recognition.
Data Mining and Knowledge Discovery, 2, 121-167.
Buschmann, C. & E. Nagel. 1993. In vivo spectroscopy and internal optics of leaves
as basis for remote sensing of vegetation. International Journal of Remote
Sensing, 14, 711-722.
Campbell, G. S. 1985. Extinction coefficients for radiation in plant canopies
calculated using an ellipsoidal inclination angle distribution. Agricultural and
Forest Meteorology, 36, 317-321.
Campbell, G. S. & J. M. Norman. 1998. An introduction to environmental biophysics.
New York: Springer.
Carpenter, G. A. & S. Grossberg. 1987a. ART2: Stable self-organisation of pattern
recognition codes for analogue input patterns. Applied Optics, 26, 49494930.
438
Carpenter, G. A. & S. Grossberg. 1987b. A massively parallel architecture for a
selforganising neural pattern recognition machine. Computer Vision,
Graphics and Image Processing: Graphical Models in Image Processing, 37,
54-115.
Carpenter, G. A., S. Grossberg & J. H. Reynolds. 1991a. ARTMAP: Supervisor realtime learning and classification of non-stationary data by a self-organising
neural network. Neural Networks, 4, 565-588.
Carpenter, G. A., S. Grossberg & D. B. Rosen. 1991b. Fuzzy ART: Fast stable learning
and categorisation of analogue patterns by an Adaptive Resonance system.
Neural Networks, 4, 759-771.
Casasent, D. & X.-W. Chen. 2000. Hyperspectral data discrimination methods. In
Proceedings of the Society of Photo-optical Instrumentation Engineers.
Casasent, D. & X.-W. Chen. 2003. Waveband selection for hyperspectral data:
Optimal feature selection. Proceeding of SPIE 5106, 259-270.
Ceccato, P., S. Flasse, S. Tarantola, S. Jacquemoud & J. M. Gregoire. 2001. Detecting
vegetation leaf water content using reflectance in the optical domain.
Remote Sensing of Environment, 77, 22-33.
Chand, T. R. K. & K. V. S. Badarinath. 2007. Analysis of ENVISAT ASAR data for forest
parameter retrieval and forest type classification—a case study over
deciduous forests of central India. International Journal of Remote Sensing,
28, 4985-4999.
Chang, C. C. & C. J. Lin. (2011). "LIBSVM: A library for support vector machines."
Retrieved 15 April, 2011, from http://www.csie.ntu.edu.tw/~cjlin/libsvm/.
Chasen, J. W., D. D. Baldocchi & M. A. Huston. 1991. A comparison of direct and
indirect methods for estimating forest canopy leaf area. Agricultural and
Forest Meteorology, 57, 107-128.
Chason, J., D. Baldocchi & M. Hutson. 1991. A comparison of direct and indirect
methods for estimating forest leaf area. Agricultural and Forest Meteorology,
57, 107-128.
Chauvaud, S., C. Bouchon & R. Maniere. 1998. Remote sensing techniques adapted
to high resolution mapping of tropical coastal marine ecosystems (coral
reefs, seagrass beds and mangrove). International Journal of Remote Sensing,
19, 3625-3639.
Chen, J. M. 1996. Optically-based methods for measuring seasonal variation of leaf
area index in boreal conifer stands. . Agricultural and Forest Meteorology, 80,
135-163.
439
Chen, J. M. & T. A. Black. 1991. Measuring leaf area index of canopies with branch
architecture. Agricultural and Forest Meteorology, 57, 1-12.
Chen, J. M. & T. A. Black. 1992. Defining leaf-area index for non-flat leaves. Plant,
Cell and Environment, 15, 421-429.
Chen, J. M., T. A. Black & R. S. Adams. 1991. Evaluation of hemispherical
photography for determining plant area index and geometry of a forest
stand. . Agricultural and Forest Meteorology, 56, 129-143.
Chen, J. M. & J. Cihlar. 1995. Plant canopy gap size analysis theory for improving
optical measurements of leaf area index. Applied Optics, 34, 6211-6222.
Chen, J. M. & S. G. Leblanc. 1997. A four-scale bidirectional reflectance model based
on canopy architecture. IEEE Transactions on Geoscience and Remote
Sensing, 35, 1316-1337.
Chen, J. M., S. G. Leblanc, J. R. Miller, J. Freemantle, S. E. Loechel, C. L. Walthall, K. A.
Innanen & H. P. White. 1999. Compact airborne spectrographic imager used
for mapping biophysical parameters of boreal forests. Journal of Geophysical
Research, 104, 945-958.
Chen, J. M., x. Li, T. Nilson & A. Strahler. 2000. Recent advantages in geometrical
optical modelling and its applications. Remote Sensing Reviews, 18, 227-262.
Chen, J. M., P. M. Rich, S. T. Gower, J. M. Norman & S. Plummer. 1997. Leaf area
index of boreal forests: theory, techniques, and measurements. Journal of
Geophysical Research, 102, 429-443.
Chen, X.-W. 2001. Feature reduction and classification of high dimensional
hyperspectral data. In Electrical and Computer Engineer, 209. Pittsburgh:
Carnegie Mellon University.
Chen, X.-W. 2003. An improved branch and bound algorithm for feature selection.
Pattern Recognition Letters, 24, 1925-1933.
Chow, T. W. S. & D. Huang. 2005. Estimating optimal feature subsets using efficient
estimation of high-dimensional mutual information. IEEE Transactions on
Neural Networks, 16, 213-224.
Cimino, J., A. Brandani, D. Casey, J. Rabassa & S. D. Wall. 1986. Multiple incidence
angle SIR-B experiment over Argentina: mapping of forest units. IEEE
Transactions on Geoscience and Remote Sensing, 24, 498-509.
Clevers, J. G. P. W. 1988. The derivation of a simplified reflectance model for
estimation of Leaf Area Index. Remote Sensing of Environment, 25, 53-69.
440
Clough, B. F., J. E. Ong & G. W. Gong. 1997. Estimating leaf area index and
photosynthetic production in canopies of the mangrove rhizophora
apiculata. Marine Ecology Progress Series, 159, 285-292.
Cohen, M. C. L. & R. J. Lara. 2003. Temporal changes of mangrove vegetation
boundaries in Amazonia: Application of GIS and remote sensing techniques.
Wetlands Ecology and Management, 11, 223-231.
Combal, B., F. Baret, M. Weiss, A. Trubuil, D. Macé & A. Pragnère. 2002. Retrieval of
canopy biophysical variables from bidirectional reflectance using prior
information to solve the ill-posed inverse problem. Remote Sensing of
Environment, 84, 1-15.
Comley, B. W. T. & K. A. McGuinness. 2005. Above– and below–ground biomass,
and allometry, of four common northern Australian mangrove. Australian
Journal of Botany, 53, 431-436.
Cortes, C. & V. Vapnik. 1995. Support vector networks. Machine Learning, 20, 273297.
Cournac, L., M.-A. Dubois, J. Chave & B. Riera. 2002. Fast determination of light
availability and leaf area index in tropical forests. Journal of Tropical Ecology,
18, 295-302.
Cover, T. M. & J. A. Thomas. 2006. Elements of Information Theory. Hoboken, N.J.:
Wiley-Interscience.
Curran, P. J. 1994. Imaging spectrometry - its present and future role in
environmental research. In Imaging Spectrometry - A Tool for Environmental
Observations, eds. J. Hill & J. Megier, 1-23. Dordrecht: Kluwer Academic
Publishers.
Curran, P. J., J. L. Dungan & H. L. Gholz. 1991. Exploring the relationship between
reflectance red-edge and chlorophyll content in slash pine. Tree Physiology,
7.
Curran, P. J., J. L. Dungan, B. A. Macler, S. E. Plummer & D. L. Peterson. 1992.
Reflectance spectroscopy of fresh whole leaves for the estimation of
chemical concentration. Remote Sensing of Environment, 39, 153-166.
Cutini, A., G. Matteucci & G. S. Mugnozza. 1998. Estimation of leaf area index with
the Li-Cor LAI 2000 in deciduous forests. Forest Ecology and Management,
105, 55-65.
Danielsen, F., M. K. Sorensen, M. F. Olwig, V. Selvam, F. Parish, N. D. Burgess, T.
Hiralshi, V. M. Karunagaran, M. S. Rasmussen, L. B. Hansen, A. Quarto & N.
Suryadiputra. 2005. The Asian Tsunami: a protective role for coastal
vegetation. Science, 310, 643.
441
Danson, F. M. & S. E. Plummer. 1995. Red-edge response to forest leaf area index.
International Journal of Remote Sensing, 16, 183-188.
Danson, F. M., C. S. Rowland & F. Baret. 2003. Training a neural network with a
canopy reflectance model to estimate crop leaf area index. International
Journal of Remote Sensing, 24, 4891-4905.
Dash, M. & H. Liu. 1997. Feature selection for classification. Intelligent Data Analysis,
3, 131-156.
Datt, B., T. R. McVicar, T. G. Van Niel, D. L. B. Jupp & J. S. Pearlman. 2003.
Preprocessing EO-1 Hyperion hyperspectral data to support the application
of agricultural indexes. IEEE Transactions on Geoscience and Remote Sensing,
41, 1246-1259.
Daughtry, C. S. T., C. L. Walthall, M. S. Kim, E. Brown de Colstoun & J. E. McMurtrey
III. 2000. Estimating corn leaf chlorophyll concentration from leaf and
canopy reflectance. Remote Sensing of Environment, 74, 229-239.
Davis, B. A. & J. R. Jensen. 1998. Remote sensing of mangrove biophysical
characteristics. Geocarto International, 13, 55-64.
Davis, C. O., J. Bowles, R. A. Leathers, D. Korwan, T. V. Downes, W. A. Snyder, W. J.
Rhea, W. Chen, J. Fisher, W. P. Bissett & R. A. Reisse. 2002. Ocean PHILLS
hyperspectral imager: design, characterization, and calibration. Optics
Express, 10, 210-221.
Dawson, T. P. & P. J. Curran. 1998. Technical note: A new technique for
interpolating the reflectance red edge position. International Journal of
Remote Sensing, 19, 2133-2139.
De Jong, S. M. & G. F. Epema. 2001. Imaging spectrometry for surveying and
modelling land degradation. In Imaging Spectrometry: Basic Principles and
Prospective Applications, eds. F. D. Van der Meer & S. M. De Jong, 65-86.
Dordrecht: Kluwer Academic Publishers.
Deblonde, G., M. Penner & A. Royer. 1994. Measuring leaf area index with Li-Cor
LAI-2000 in pine stand. Ecology, 75, 1507-1511.
Demetriades-Shah, T. H., M. D. Steven & J. A. Clark. 1990. High resolution derivative
spectra in remote sensing. Remote Sensing of Environment, 33, 55-64.
Devijver, P. A. & J. Kittler. 1982. Pattern Recognition: A Statistical Approach.
Englewood Cliffs: Prentice-Hall.
Dobson, M. C., F. T. Ulaby, T. LeToan, A. Beaudoin, E. S. Kasischke & N. Christensen.
1992. Dependence of Radar Backscatter on Coniferous Forest Biomass. IEEE
T ransactions on Geoscience and Remote Sensing, 30, 412-415.
442
Dobson, M. C., F. T. Ulaby, L. E. Prierce, T. L. Sharik, K. M. Bergen, J. Kellndorfer, J. R.
Kendra, E. Li, Y. C. Lin, A. Nashashibi, K. Sarabandi & P. Siqueira. 1995.
Estimation of Forest Biophysical Characteristics in Northem Michigan with
SIR-C/X-SAR. IEEE T ransactions on Geoscience and Remote Sensing, 33, 877895.
Drieman, J. A., D. G. Leckie & F. J. Ahern. 1989. Multitemporal C-sar for forest typing
In Eastern Ontario. Proceeding of the 12th International Geoscience and
Remote Sensing Symposium 1989, 3, 1376-1378.
Duda, R. O., P. E. Hart & D. G. Stork. 2001. Pattern Classification. New York: Wiley.
Dufrêne, E. & N. Bréda. 1995. Estimation of deciduous forest leaf area index using
direct and indirect methods. Oecologia, 104, 156-162.
Duin, R. P. W., P. Juszczak, P. Paclik, E. Pekalska, D. de Ridder, D. M. J. Tax & S.
Verzakov. 2007. PRTools4.1, A Matlab Toolbox for Pattern Recognition. Delft
University of Technology.
Duke, N. C. & M. Ajmal Khan. 1999. Structure and composition of the seaward
mangrove forest at the Mai Po Marshes Nature Reserve, Hong Kong. In The
Mangrove Ecosystem of Deep Bay and the Mai Po Marshes, Hong Kong, ed. S.
Y. Lee, 83-104. Hong Kong: Hong Kong University Press.
Dundar, M. M. & D. Landgrebe. 2003. Toward An Optimal Analysis of Hyperspectral
Data. In School of Electrical and Computer Engineering, 94. West Lafayette,
Indiana: Purdue University.
Dunne, K., P. Cunningham & F. Azuaje. 2002. Solutions to instability problems with
sequential wrapper-based approaches to feature selection. In Technical
Report TCDCD-2002-28. Dublin, Ireland: Dept. of Computer Science, Trinity
College.
Durand, J. M., B. J. Gimonet & J. R. Perbos. 1987. SAR data filtering for classification.
IEEE T ransactions on Geoscience and Remote Sensing, 25, 629-637.
Dutra, L. V. & R. Huber. 1999. Feature extraction and selection for ERS-1/2 InSAR
classification. International Journal of Remote Sensing, 20, 993-1016.
Dy, J. & C. E. Brodley. 2004. Feature selection for unsupervised learning. Journal of
Machine Learning Research, 5, 845-889.
Eastwood, J. A., M. G. Yates, A. G. Thomson & R. M. Fuller. 1997. The reliability of
vegetation indices for monitoring saltmarsh vegetation cover. International
Journal of Remote Sensing, 18, 3901-3907.
Eckert, S. & M. Kneubuhler. 2004. Application of HYPERION data to agricultural land
classification and vegetation properties estimation in Switzerland. In
443
Proceedings of the XXth ISPRS Congress on Geo-Imagery Bridging Continents.
Istanbul, Turkey.
Elvidge, C. D. 1990. Visible and near-infrared reflectance characteristics of dry plant
materials. International Journal of Remote Sensing, 11, 1775-1795.
Elvidge, C. D. & Z. Chen. 1995. Comparison of broad-band and narrow-band red and
near-infrared vegetation indices. Remote Sensing of Environment, 54, 38-48.
ENVI. 2009. The Environment for Visualizing Images (ENVI) Help Manual. ITT Visual
Information Solutions.
Epiphanio, J. C. N. & A. R. Huete. 1995. Dependence of NDVI and SAVI on
sun/sensor geometry and its effect of fAPAR relationships in alfalfa. Remote
Sensing of Environment, 51, 351-360.
Eriksson, H. M., L. Eklundh, A. Kuusk & T. Nilson. 2006. Impact of understory
vegetation on forest canopy reflectance and remotely sensed LAI estimates.
Remote Sensing of Environment, 103, 408-418.
European Space Agency. (2009). "BEST - Basic Envisat SAR Toolbox User Manual
(Version 4.2.2)." Retrieved August 25, 2010, from http://earth.esa.int/best/.
Fang, H., S. Liang & A. Kuusk. 2003. Retrieving leaf area index using a genetic
algorithm with a canopy radiative transfer model. Remote Sensing of
Environment, 85, 257-270.
FAO. 2007. FAO forestry paper 153: the world's mangrove 1980-2005. Rome: Food
and Agriculture Organization of the United Nations.
Fassnacht, K. S., S. T. Gower, J. M. Norman & R. E. McMurtrie. 1994. A comparison
of optical and direct methods for estimating foliage surface area index in
forests. Agricultural and Forest Meteorology, 71, 183-207.
Felde, G. W., G. P. Anderson, T. W. Cooley, M. W. Matthew, S. M. Adler-Golden, A.
Berk & J. Lee. 2003. Analysis of Hyperion data with the FLAASH atmospheric
correction algorithm. In Geoscience and Remote Sensing Symposium, 2003.
IGARSS '03. Proceedings. 2003 IEEE International, 90-92.
Fensholt, R., I. Sandholt & M. S. Rasmussen. 2004. Evaluation of MODIS LAI, fAPAR
and the relation between fAPAR and NDVI in a semi-arid environment using
in situ measurements. Remote Sensing of Environment, 91, 490-507.
Ferrazzoli, P., S. Paloscia, P. Pampaloni, G. Schiavon, S. Sigismondi & D. Solimini.
1997. The potential of multi-frequency polarimetric SAR in assessing
agricultural and arboreous biomass. IEEE Transactions on Geoscience and
Remote Sensing, 35, 5-17.
444
Ferri, F., P. Pudil, M. Hatef & J. Kittler. 1994. Comparative study of rechniques for
large scale feature selection. In Pattern Recognition in Practice IV - multiple
paradigms, comparative studies and hybrid systems: proceedings of an
international workshop held in Vlieland, eds. E. Gelsema & L. Kanal, 403-413.
Filella, I. & J. Penuelas. 1994. The red edge position and shape as indicators of plant
chlorophyll content, biomass and hydric status. International Journal of
Remote Sensing, 15, 1459-1470.
Filho, P. W. M. S., E. d. S. F. Martins & F. R. da Costa. 2006. Using mangroves as a
geological indicator of coastal changes in the Braganca macrotidal flat,
Brazilian Amazon: A remote sensing data approach. Ocean and Coastal
Management, 49, 462-475.
Filho, P. W. M. S. & W. R. Paradella. 2002. Recognition of the main geobotanical
features along the Braganca mangrove coast (Brazilian Amazon region) from
Landsat TM and RADARSAT-1 data. Wetlands Ecology and Management, 10,
123-132.
Filippi, A. M. & J. R. Jensen. 2006. Fuzzy learning vector quantization for
hyperspectral coastal vegetation classification. Remote Sensing of
Environment, 100, 512-530.
Foley, D. H. 1972. Considerations of sample and feature size. IEEE Transactions on
Information Theory, IT-18, 618-626.
Foley, S., B. Rivard, G. A. Sanchez-Azofeifa & J. Calvo. 2006. Foliar spectral
properties following leaf clipping and implications for handling techniques.
Remote Sensing of Environment, 103, 265-275.
Ford, J. P. & D. J. Casey. 1988. Shuttle radar mapping with diverse incidence angles
in rainforest of Borneo. International Journal of Remote Sensing, 9, 927-943.
Foroutan, I. & J. Sklansky. 1987. Feature selection for automatic classification of
non-Gaussian data. IEEE Transactions on Systems, Man and Cybernetics, 17,
187-198.
Fournier, R. A., D. Mailly, J.-M. N. Walter & K. Soudani. 2003. Indirect measurement
of forest canopy structure from in situ optical sensors. In Remote Sensing of
Forest Environments: Concepts and Case Studies eds. M. A. Wulder & S. E.
Franklin, 519. Boston: Kluwer Academic Publishers.
Fourty, T. & F. Baret. 1997. Vegetation and dry matter contents estimated from topof-the-atmosphere reflectance data: a simulation study. . Remote Sensing of
Environment, 61, 34-45.
Fourty, T., F. Baret, S. Jacquemoud, G. Schmuch & J. Verdebout. 1996. Leaf opeical
properties with explicit descritpion of its biochemical composition: direct
and inverse problems. Remote Sensing of Environment, 56, 104-117.
445
Frazer, G. W., R. A. Fournier, J. A. Trofymow & R. J. Hall. 2001. A comparison of
digital and film fisheye photography for analysis of forest canopy structure
and gap light transmission. Agricultural and Forest Meteorology, 109, 249263.
Friedl, M. A. & C. E. Brodley. 1997. Decision tree classification of land cover from
remotely sensed data. Remote Sensing of Environment, 61, 399-409.
Frost, V. S., J. A. Stiles, K. S. Shanmugan & J. C. Holtzman. 1982. A model for radar
images and its application to adaptive digital filtering of multiplicative noise.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 4, 157-166.
Fukuda, S. & H. Hirosawa. 1998. Suppression of speckle in synthetic aperture radar
images using wavelet. International Journal of Remote Sensing, 19, 507-519.
Fukunaga, K. 1990. Introduction to Statistical Pattern Recognition. Boston; London:
Academic Press.
Fung, T., Y. Lenung, F. K. K. Wong, M. W. Dai & C. K.S. Under review. Species-based
Mangrove Mapping Using High Resolution QuickBird Image in Inner Deep
Bay, Hong Kong. International Journal of Remote Sensing.
Gahegan, M. & G. West. 1998. The classification of complex geographic datasets: An
operational
comparison of artificial neural networks and decision tree classifiers. In Proceedings
of the 3rd International Conference on GeoComputation University of Bristol,
United Kingdom.
Gamon, J. A., J. Penuelas & C. B. Field. 1992. A narrow waveband spectral index that
tracks diurnal changes in photosynthetic efficiency. Remote Sensing of
Environment, 41, 35-44.
Gamon, J. A. & J. S. Surfus. 1999. Assessing leaf pigment content and activity with a
reflectometer. New Phytologist, 143, 105-117.
Gao, B.-C., M. J. Montes & C. O. Davis. 2004. Refinement of wavelength calibrations
of hyperspectral imaging data using a spectrum-matching technique.
Remote Sensing of Environment, 90, 424-433.
Gao, B. C., C. O. Davis & A. F. H. Goetz. 2006. A Review of Atmospheric Correction
Techniques for Hyperspectral Remote Sensing of Land Surfaces and Ocean
Color. In IEEE International Conference on Geoscience and Remote Sensing
Symposium, 2006. IGARSS 2006., 1979 - 1981 Denver, CO.
Gao, J. 1999. A comparative study on spatial and spectral resolutions of satellite
data in mapping mangrove forests. International Journal of Remote Sensing,
20, 2823-2833.
446
Gao, S., Z. Niu & C. Wu. 2010. Multi-polarization Envisat-ASAR images as a function
of leaf area index (LAI) of White Poplar and Desert Date plantations.
International Journal of Remote Sensing, 31, 1095-1102.
Gao, X., A. R. Huete, W. Ni & T. Miura. 2000. Optical-biophysical relationships of
vegetation spectra without background contamination. Remote Sensing of
Environment, 74, 609-620.
Garcia-Haro, F. J. & S. Sommer. 2002. A fast canopy reflectance model to simulate
realistic remote sensing scenarios. Remote Sensing of Environment, 81, 205227.
Gausman, H. W., W. A. Allen, R. Cardenas & A. J. Richardson. 1970. Relationship of
light reflectance to histological and physical evaluation of cotton leaf
maturity. Applied Optics, 9, 545-552.
Goel, N. S. 1988. Models of vegetation canopy reflectance and their use in the
estimation of biophysical parameters from reflectance data. Remote Sensing
Reviews, 4, 1-212.
Goel, N. S. & W. Qin. 1994. Influences of canopy architecture on relationships
between various vegetation indices and LAI and FPAR: a computer
simulation. Remote Sensing Reviews, 10, 309-374.
Goetz, A. F. H. 1992. Imaging spectrometry for earth remote sensing. In Imaging
Spectroscopy: Fundamentals and Prospective Applications, eds. F. Toselli & J.
Bodechtel, 1-20. Dordrecht: Kluwer Academic Publishers.
Goetz, A. F. H., G. Vane, J. E. Solomon & B. N. Rock. 1985. Imaging spectrometry for
earth remote sensing. Science, 228, 1147-1153.
Goetz, S. & K. F. Huemmrich. 2007. Plant and Vegetation Community Properties.
http://www.ccpo.odu.edu/SEES/veget/vg_class.htm (last accessed 14
December 2007).
Goldberg, D. E. 1989. Genetic Algorithms in Search, Optimization and Machine
Learning. Reading, Mass.: Addison-Wesley Pub. Co.
Gong, P., R. Pu, G. Biging & M. Larrieu. 2003. Estimation of forest leaf area index
using vegetation indices derived from Hyperion hyperspectral data. IEEE
Transactions on Geoscience and Remote Sensing, 41, 1355-1362.
Gong, P., D. X. Wang & S. Liang. 1999. Inverting a canopy reflectance model using a
neural network. International Journal of Remote Sensing, 20, 111-122.
Goodenough, D. G., P. M. Narenda & K. O’Neill. 1978. Feature subset selection in
remote sensing. Canadian Journal of Remote Sensing, 4, 143-148.
447
Govaerts, Y. M., M. M. Verstraete, B. Pinty & N. Gobron. 1999. Designing optimal
spectral indices: a feasibility and proof of concept study. International
Journal of Remote Sensing, 20, 1853-1873.
Gower, S. T., C. J. Kucharik & J. M. Norman. 1999. Direct and Indirect Estimation of
Leaf Area Index, fAPAR, and Net Primary Production of Terrestrial
Ecosystems. Remote Sensing of Environment, 70, 29-51.
Gower, S. T. & J. M. Norman. 1990. Rapid estimation of leaf area index in forests
using the LI COR LAI 2000. Ecology, 72, 1896-1900.
Gower, S. T. & J. M. Norman. 1991. Rapid estimation of leaf area index in conifer
and broad leaf plantations. . Ecology, 72, 1896-1900.
Green, A. A., M. Berman, P. Switzer & M. D. Craig. 1988. A transformation for
ordering multispectral data in terms of image quality with implications for
noise removal. IEEE Transactions on Geoscience and Remote Sensing, 26, 6574.
Green, E. P. & C. D. Clark. 2000. Assessing Mangrove Leaf Area Index and Canopy
Closure. In Remote Sensing Handbook for Tropical Coastal Management., ed.
A. J. Edwards. Paris: UNESCO.
Green, E. P., C. D. Clark, P. J. Mumby, A. J. Edwards & A. C. Ellis. 1998a. Remote
sensing techniques for mangrove mapping. International Journal of Remote
Sensing, 19, 935-956.
Green, E. P., P. J. Mumby, A. J. Edwards, C. D. Clark & A. C. Ellis. 1997. Estimating
leaf area index of mangroves from satellite data. Aquatic Botany, 58, 11-19.
Green, E. P., P. J. Mumby, A. J. Edwards, C. D. Clark & A. C. Ellis. 1998b. The
assessment of mangrove areas using high resolution multispectral airborne
imagery. Journal of Coastal Research, 14, 433-443.
Green, R. M. 1998. Relationships between polarimetric SAR backscatter and forest
canopy and sub-canopy biophysical properties. International Journal of
Remote Sensing, 19, 2395-2412.
Griffin, M. K. & H.-h. K. Burke (2003) Comparison of hyperspectral data for
atmospheric effects. Lincoln Laboratory Journal, 14, 29-54.
http://eo1.gsfc.nasa.gov/new/validationReport/Technology/Documents/Tec
h.Val.Report/Science_Summary_Goetz.pdf (last accessed 11 October 2007).
Guyon, I. & A. Elisseeff. 2003. An introduction to variable and feature selection.
Journal of Machine Learning Research, 3, 1157-1182.
Guyot, G., F. Baret & S. Jacquemoud. 1992. Imaging spectroscopy for vegetation
studies. In Imaging Spectroscopy: Fundamentals and Prospective
448
Applications, eds. F. Toselli & J. Bodechtel, 145-165. Dordrecht: Kluwer
Academic Publishers.
Haboudane, D., J. R. Miller, E. Pattey, P. J. Zarco-Tejada & L. B. Strachan. 2004.
Hyperspectral vegetation indices and novel algorithms for predicting green
LAI of crop canopies: modeling and validation in the context of precision
agriculture. Remote Sensing of Environment, 90, 337-352.
Hale, S. E. & C. Edwards. 2002. Comparison of film and digital hemispherical
photography across a wide range of canopy densities. Agricultural and
Forest Meteorology, 112, 51-56.
Hall-Beyer, M. (2007). "GLCM Texture: A tutorial." Retrieved August 24, 2010, from
http://www.fp.ucalgary.ca/mhallbey/tutorial.htm.
Hall, F. G., K. F. Huemmrich & S. N. Goward. 1990. Use of narrow-band spectra to
estimate the fraction of absorbed photosynthetically active radiation.
Remote Sensing of Environment, 34, 47-54.
Hamamoto, Y., S. Uchimura, Y. Matsuura, T. Kanaoka & S. Tomita. 1990. Evaluation
of the branch and bound algorithm for feature selection. Pattern
Recognition Letters, 11, 453-456.
Han, J. & M. Kamber. 2006. Data Mining: Concepts and Techniques. San Francisco,
California: Morgan Kaufmann.
Haralick, R. M. 1979. Statistical and structural approaches to texture. Proceedings of
the IEEE, 67.
Haralick, R. M., K. Shaunmugam & I. Dinstein. 1973. Textural features for image
classification. IEEE Transactions on Systems, Man and Cybernetics, 3.
Hashim, M. & W. H. W. Kadir (1999) Comparison of JERS-1 and RADARSAT synthetic
aperture radar data for mangrove mapping and its biomass. Asian
Conference
on
Remote
Sensing
(ACRS)
Proceeding,
1999.
http://www.gisdevelopment.net/aars/acrs/1999/ps1/ps1017.asp
(last
accessed 9 Sept 2010).
Haupt, W. & R. Scheuerlein. 1990. Chloroplast movement. Plant, Cell and
Environment, 13, 565-614.
Held, A., C. Ticehurst, L. Lymburner & N. Williams. 2003. High resolution mapping of
tropical mangrove ecosystems using hyperspectral and radar remote sensing.
International Journal of Remote Sensing, 24, 2739-2759.
Henderson, F. M. & J. Lewis. 1998. Principles and applications of imaging radar. In
Manual of Remote Sensing - Volume 2, ed. R. A. Ryerson, 461-465. New York:
Wiley.
449
Herbert, T. J. 1986. Calibration of fisheye lenses by inversion of area projections.
Applied Optics, 25, 1875-1976.
Hicks, S. & R. Lascano. 1995. Estimation of leaf area Index for cotton canopies using
the Li-Cor LAI 2000 plant canopy analyser. Agronomy Journal, 87, 458-464.
Hill, R. 1924. A lens for whole sky photographs. Quarterly Journal of the Royal
Meteorological Society 50, 227-235.
Hirano, A., M. Madden & R. Welch. 2003. Hyperspectral image data for mapping
wetland vegetation. Wetlands, 23, 436-448.
Holmes, P. R. 1988. Tolo Harbour - the case for integrated water quality
management in a coastal environment. Journal of the Institution of Water &
Environmental Management, 2, 171-179.
Hopfield, J. J. 1982. Neural networks and physical systems with emergent collective
computational abilities. In Proceedings of the National Academy of Sciences
of the USA, 2554-2558.
Horler, D. N. H., M. Dockray & J. Barber. 1983. The red edge of plant leaf reflectance.
International Journal of Remote Sensing, 4, 273-288.
Houborg, R. & E. Boegh. 2008. Mapping leaf chlorophyll and leaf area index using
inverse and forward canopy reflectance modeling and SPOT reflectance data.
Remote Sensing of Environment, 112, 186-202.
Hsu, P.-H. 2007. Feature extraction of hyperspectral images using wavelet and
matching pursuit. ISPRS Journal of Photogrammetry & Remote Sensing, 62,
78-92.
Hsu, P.-H., Y.-H. Tseng & P. Gong. 2002. Dimension reduction of hyperspectral
images for classification applications. Geographic Information Sciences, 8, 18.
Huang, C., L. S. Davis & J. R. G. Townshend. 2002. An assessment of support vector
machines for land cover classification. International Journal of Remote
Sensing, 23, 725-749.
Huang, S. & D. Liu. 2007. Some uncertain factor analysis and improvement in
spaceborne synthetic aperture radar imaging. Signal Processing, 87, 32023217.
Huemmrich, K. F. & S. N. Goward. 1997. Vegetation canopy PAR absorptance and
NDVI: an assessment for ten tree species with the SAIL model. Remote
Sensing of Environment, 61, 254-269.
Huete, A., C. Justice & H. Liu. 1994. Development of Vegetation and Soil Indexes For
Modis-Eos. Remote Sensing of Environment, 49, 224-234.
450
Huete, A. R. 1988. A soil-adjusted vegetation index (SAVI). Remote Sensing of
Environment, 25, 295-309.
Hughes, G. F. 1968. On the mean accuracy of statistical pattern recognizers. IEEE
Transactions on Information Theory, 14, 55-63.
Hunt, E. R. & B. N. Rock. 1989. Detection of changes in leaf water content using near
and middle infrared reflectances. Remote Sensing of Environment, 30, 43-54.
Hutchings, P. & P. Saenger. 1987. Ecology of Mangroves. St Lucia, Australia:
University of Queensland Press.
Imhoff, M. L. 1995. Radar Backscatter and Biomass Saturation: Ramifications for
Global Biomass Inventory. IEEE Transactions on Geoscience and Remote
Sensing, 33, 511-518.
Imhoff, M. L., S. Carson & P. Johnson. 1998. A low-frequency radar experiment for
measuring vegetation biomass. IEEE Transactions on Geoscience and Remote
Sensing, 36, 1988-1991.
Imhoff, M. L., C. Vermillion & M. H. Story. 1987. Monsoon flood boundary
delineation and damage assessment using space borne imaging radar and
landsat data. Photogrammetric Engineering and Remote Sensing, 53, 405413.
Inoue, A., K. Yamamoto, N. Mizoue & Y. Kawahara. 2004. Effects of image quality,
size and camera type on forest light environment estimates using digital
hemispherical photography. Agricultural and Forest Meteorology, 126, 89-97.
Irving, R. & R. Morton. 1988. A Geography of The Mai Po Marshes. Hong Kong: Hong
Kong University Press.
Jacquemoud, S., C. Bacour, H. Poilve & J.-P. Frangi. 2000. Comparison of four
radiative transfer models to simulate plant canopies reflectance: Direct and
inverse mode. . Remote Sensing of Environment, 74, 471-481.
Jacquemoud, S. & F. Baret. 1990. PROSPECT—a model of leaf optical-properties
spectra. Remote Sensing of Environment, 34, 75-91.
Jain, A. & D. Zongker. 1997. Feature selection: evaluation, application, and small
sample performance. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 19, 153-158.
Jain, A. K. & B. Chandrasekaran. 1982. Dimensionality and Sample Size
Considerations in Pattern Recognition Practice. In Handbook of Statistics
Vol.2, eds. P. R. Krishnaiah & L. N. Kanal, 835-855. Amsterdam: NorthHolland Pub. Co.
451
Jain, A. K., R. C. Dubes & C.-C. Chen. 1987. Bootstrap Techniques for Error
Estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence,
9, 628-633.
Jain, A. K., R. P. W. Duin & J. Mao. 2000. Statistical pattern recognition: a review.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 22, 4-37.
Jensen, J. R. 1996. Introductory Digital Image Processing: A Remote Sensing
Perspective. Upper Saddle River: Prentice-Hall, Inc.
Jensen, J. R. 2007. Remote Sensing of the Environment: An Earth Resource
Perspective. Upper Saddle River: Prentice Hall.
Jensen, R. R. & M. W. Binford. 2004. Measurement and comparison of Leaf Area
Index estimators derived from satellite remote sensing techniques.
International Journal of Remote Sensing, 25, 4251-4265.
Jiang, X., L. Tang, C. Wang & C. Wang. 2004. Spectral characteristics and feature
selection of hyperspectral remote sensing data. International Journal of
Remote Sensing, 25, 51-59.
Jonckheere, I., S. Fleck, K. Nackaerts, B. Muys, P. Coppin, M. Weiss & F. Baret. 2004.
Review of methods for in situ leaf area index determination Part I. Theories,
sensors and hemispherical photography. Agricultural and Forest
Meteorology, 121, 19-35.
Jonckheere, I., K. Nackaerts, B. Muys & P. Coppin. 2005. Assessment of automatic
gap fraction estimation of forests from digital hemispherical photography.
Agricultural and Forest Meteorology, 132, 96-114.
Jones, H. G. & R. A. Vaughan. 2010. Remote sensing of vegetation - principles,
techniques and applications. New York: Oxford University Press.
Jupp, D. L. B., B. Datt, T. R. McVicar, T. G. Van Niel, J. S. Pearlman, J. Lovell & E. G.
King. 2002. Improving the Analysis of Hyperion Red Edge Index from an
Agricultural area. In Proceedings of the SPIE. Hangzhou, China.
Kachhwaha, T. S. 1993. Temporal and multisensor approach in
forest/vegetationmapping and corridor identification for effective
management of Rajaji National Park, Uttar Pradesh, India. International
Journal of Remote Sensing, 14, 3105-3114.
Kalousis, A., J. Prados & M. Hilario. 2007. Stability of feature selection algorithms: a
study on high-dimensional spaces. Knowledge and Information Systems, 12,
95-116.
Kanellopoulos, I., A. Varfis, G. G. Wilkinson & J. Megier. 1992. Land-cover
discrimination in SPOT HRV imagery using an artificial neural network: A 20class experiment. International Journal of Remote Sensing, 13, 917-924.
452
Karnieli, A., Y. J. Kaufman, L. Remer & A. Wald. 2001. AFRI-aerosol
free vegetation index. . Remote Sensing of Environment, 77, 10-21.
Kasischke, E. S. & L. L. Bourgeau-Chavez. 1997. Monitoring South Florida Wetlands
Using ERS-1 SAR Imagery. Photogrammetric Engineering and Remote Sensing,
63, 281-291.
Kasischke, E. S., L. L. Bourgeau-Chavez & E. Haney. 1994. Observation on the
sensitivity of ERS-1 SAR image intensity to changes in aboveground biomass
in young loblolly pine forests. International Journal of Remote Sensing, 15, 316.
Kasischke, E. S., L. Morrissey, J. Way, N. H. F. French, L. L. Bourgeau, E. Rignot, J. A.
Stearn & G. P. Livingston. 1995. Monitoring seasonal variations in boreal
ecosystems using multi-temporal spaceborne SAR data. Canadian Journal of
Remote Sensing, 21, 96-109.
Kaufman, Y. J. & B.-C. Gao. 1992. Remote sensing of water vapor in the near IR from
EOS/MODIS. IEEE T ransactions on Geoscience and Remote Sensing, 30, 871884.
Kaufman, Y. J. & C. Sendra. 1988. Algorithm for automatic atmospheric corrections
to visible andnear-IR satellite imaging. International Journal of Remote
Sensing, 9, 1357-1381.
Kaufman, Y. J. & D. Tanre. 1992. Atmospherically resistant vegetation index (ARVI)
for EOSMODIS. IEEE Transactions on Geoscience and Remote Sensing, 30,
261-270.
Kavzoglu, T. & P. M. Mather. 2002. The role of feature selection in artificial neural
network applications. International Journal of Remote Sensing, 23, 29192937.
Kawishwar, P. 2007. Atmospheric Correction Models for Retrievals of Calibrated
Spectral Profiles from Hyperion EO-1 Data. In International Institute for Geoinformation Science and Earth Observation, 72. Enschede, The Netherlands.
Kellndorfer, J. M., M. C. Dobson, D. V. John & M. Clutter. 2003. Toward precision
forestry: plot-level parameter retrieval for slash pine plantations with JPL
AIRSAR. IEEE Transactions on Geoscience and Remote Sensing, 41, 15711582.
Kimes, D. S. 1984. Modeling the directional reflectance from complete
homogeneous vegetation canopies with various leaf-orientation
distributions. Journal of the Optical Society of America, A1, 752-737.
Kimes, D. S., Y. Knyazikhin, J. L. Privette, A. A. Abuelgasim & F. Gao. 2000. Inversion
methods for physically-based models. Remote Sensing Reviews, 18, 381-439.
453
Kittler, J. 1986. Feature selection and extraction. In Handbook of Pattern Recognition and
Image Processing, ed. T. Y. Y. a. K.-S. Fu. New York: Academic Press.
Knerr, S., L. Personnaz & G. Dreyfus. 1990. Single-layer learning revisited: A stepwise
procedure for building and training neural network. In Neurocomputing:
Algorithms, architectures and applications, ed. J. Fogelman, 41-50. Berlin:
Springer-Verlag.
Kohavi, R. & G. H. John. 1997. Wrappers for feature subset selection. Artificial
Intelligence, 97, 273-324.
Kohonen, T. 1982. Self-organized formation of topologically correct feature maps.
Biological Cybernetics, 43, 59-69.
Kovacs, J. M., V. F. Flores, J. Wang & L. P. Aspden. 2004. Estimating leaf area index
of a degraded mangrove forest using high spatial resolution satellite data.
Aquatic Botany, 80, 13-22.
Kovacs, J. M., C. V. Vandenberg, J. Wang & F. Flores-Verdugo. 2008. The use of
multipolarized spaceborne SAR backscatter for monitoring the health of a
degraded mangrove forest. Journal of Coastal Research, 24, 248-254.
Křížek, P., J. Kittler & V. Hlaváč. 2007. Improving stability of feature selection
methods. The 12th International Conference on Computer Analysis of Images
and Patterns, LNCS 4673, 929-936.
Kuan, D. T., A. A. Sawchuk, T. C. Strand & P. Chavell. 1987. Adaptive restoration of
images with speckle. IEEE Transactions on Acoustics, Speech, & Signal
Processing, 35, 373-383.
Kucharik, C. J., J. M. Norman & S. T. Gower. 1998. Measurements of branch area
and adjusting leaf area index indirect measurements. . Journal of
Agricultural and Forest Meteorology, 91, 61-88.
Kucharik, C. J. N., J.M., L. M. Murdock & T. S. Gower. 1997. Characterizing canopy
nonrandomness with a Multiband Vegetation Imager MVI. Journal of
Geophysical Research, 102, 455-473.
Kudo, M. & J. Sklansky. 2000. Comparison of algorithms that select features for
pattern classifiers. Pattern Recognition, 33, 25-41.
Kulkarni, A. D. 1994. Artificial neural networks for image understanding. New York:
Van Nostrand Reinhold.
Kumar, L., K. Schmidt, S. Dury & A. Skidmore. 2001. Imaging spectrometry and
vegetation science. In Imaging Spectrometry: Basic Principles and
Prospective Applications, eds. F. D. Van der Meer & S. M. De Jong, 111-155.
Dordrecht: Kluwer Academic Publishers.
454
Kuncheva, L. I. 2007. A stability index for feature selection. In IASTED International
Conference on Artificial Intelligence and Applications, part of the 25th MultiConference on Applied Informatics, ed. V. Devedzic, 421-427. Innsbruck,
Austria: ACTA Press.
Kupiec, J. A. & P. J. Curran. 1995. Decoupling the effect of the canopy and foliar
biochemical concentration in AVIRIS spectra. International Journal of Remote
Sensing, 16, 1731-1739.
Kuplich, T. M., C. C. Freitas & J. V. Soares. 2000. The study of ERS-1 SAR and Landsat
TM synergism for land use classification. International Journal of Remote
Sensing, 21, 2101-2111.
Kurosu, T., S. Uratsuka, H. Maeno & T. Kozu. 1999. Texture statistics for
classification of land use with multitemporal JERS-1 SAR single-look imagery.
IEEE Transactions on Geoscience and Remote Sensing, 37, 227-235.
Kurosu, T., S. Yokoyama & K. Chiba. 2001. Land use classification with textural
analysis and the aggregation technique using multi-temporal JERS-1 L-band
SAR images. International Journal of Remote Sensing, 22, 595-613.
Kurvonen, L., J. Pulliainen & M. Hallikainen. 1999. Retrieval of biomass in boreal
forests from multi-temporal ERS-1 and JERS-1 SAR images. IEEE Transactions
on Geoscience and Remote Sensing, 37, 198-205.
Kushwaha, S. P. S., R. S. Dwivedi & B. R. M. Rao. 2000. Evaluation of various digital
image processing techniques for detection of coastal wetlands using ERS-1
SAR data. International Journal of Remote Sensing, 21, 565-579.
Kuusk, A. 1995a. A fast, invertible canopy reflectance model. Remote Sensing of
Environment, 51, 342-350.
Kuusk, A. 1995b. A Markov chain model of canopy reflectance. Agricultural and
Forest Meteorology, 76, 221-236.
Kuusk, A. 1998. Monitoring of vegetation parameters on large areas by the
inversion of a canopy reflectance model. International Journal of Remote
Sensing, 19, 2893-2905.
Kuusk, A. 2001. A two-layer canopy reflectance model. Journal of Quantitative
Spectroscopy and Radiative Transfer, 71, 1-9.
Kuusk, A. & T. Nilson. 2000. A directional multispectral forest reflectance model.
Remote Sensing of Environment, 72, 244-252.
Küβner, R. & R. Mosandl. 2000. Comparison of direct and indirect estimation of leaf
area index in mature Norway spruce stands of eastern Germany. Canadian
Journal of Forest Research, 30, 440-447.
455
Kwak, N. & C.-H. Choi. 2002a. Input feature selection by mutual information based
on parzen window. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 24, 1667-1671.
Kwak, N. & C.-H. Choi. 2002b. Input feature selection for classification problems.
IEEE Transactions on Neural Networks, 13, 143-159.
Lacaze, B. & R. Joffre. 1994. Extracting biochemical information from visible and
near infrared reflectance spectroscopy of fresh and dried leaves. Journal of
Plant Physiology, 144, 277-281.
Landgrebe, D. A. 2001. Analysis of multispectral and hyperspectral image data. In
Introduction to Modern Photogrammetry, eds. E. M. Mikhail, J. S. Bethel & C.
McGlone. New York: John Wiley & Sons Inc.
Lang, A. R. G. 1987. Simplified estimate of leaf area index from transmittance of the
sun's beam. . Agricultural and Forest Meteorology, 41.
Lawrence, R., A. Bunn, S. Powell & M. Zambon. 2004. Classification of remotely
sensed imagery using stochastic gradient boosting as a refinement of
classification tree analysis. Remote Sensing of Environment, 90, 331-336.
Le Dantec, V., E. Dufrene & B. Saugier. 2000. Interannual and spatial variation in
maximum leaf area index of temperate deciduous stands. Forest Ecology and
Management, 134, 71-81.
Le Hegarat-Mascle, S., A. Quesney, D. Vidal-Madjar, O. Taconet, M. Normand & C.
Loumagne. 2000. Land cover discrimination from multitemporal ERS images
and multispectral Landsat images: a study case in an agricultural area in
France. International Journal of Remote Sensing, 21, 435-456.
Le Toan, T., A. Beaudoin, J. Riom & D. Guyon. 1992. Relating forest biomass to SAR
data. IEEE T ransactions on Geoscience and Remote Sensing, 30, 403-411.
Le Toan, T., S. Quegan, I. Woodward, M. Lomas, N. Delbart & G. Picard. 2004.
Relating radar remote sensing of biomass to modelling of forest carbon
budgets. Climate Change, 67, 379-402.
Leblanc, S. G., J. M. Chen, R. Fernandes, D. W. Deering & A. Conley. 2005.
Methodology comparison for canopy structure parameters extraction from
digital hemispherical photography in boreal forests. Agricultural and Forest
Meteorology, 129, 187-207.
Leckie, D. G. 1990. Synergism of Synthetic Aperture Radar and Visible/Infrared Data
for Forest Type Discrimination. Photogrammetric Engineering and Remote
Sensing, 56, 1237-1246.
456
Leckie, D. G. & K. J. Ranson. 1998. Forestry applications using imaging radar. In
Principles and Applications of Imaging Radar (Manual of Remote Sensing,
Volume 2), eds. F. M. Henderson & A. J. Lewis, 435-509. New York: Wiley.
Lee, C. & D. A. Landgrebe. 1993a. Analyzing high-dimensional multispectral data.
IEEE Transactions on Geoscience and Remote Sensing, 31, 792-800.
Lee, C. & D. A. Landgrebe. 1993b. Feature extraction based on decision boundaries.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 15, 388-400.
Lee, J. B., A. S. Woodyatt & M. Berman. 1990. Enhancement of high spectral
resolution remote-sensing data by a noise-adjusted principal components
transform. IEEE Transactions on Geoscience and Remote Sensing, 28, 295304.
Lee, J. S. 1980. Digital image enhancement and noise filtering by use of local
statistics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2,
165-186.
Lee, K. S., W. B. Cohen, R. E. Kennedy, T. K. Maiersperger & S. T. Gower. 2004.
Hyperspectral versus multispectral data for estimating leaf area index in four
different biomes. Remote Sensing of Environment, 91, 508-520.
Levy, P. E. & P. G. Jarvis. 1999. Direct and indirect measurements of LAI in millet and
fallow vegetation in HAPEX-Sahel. Agricultural and Forest Meteorology, 97,
199-212.
LI-COR. 1992. LAI-2000 Plant Canopy Analyzer - Instruction Manual. Lincoln: LI-COR,
Inc.
Li, B., D.-S. Huang, C. Wang & K.-H. Liu. 2008a. Feature extraction using constrained
maximum variance mapping. Pattern Recognition, 41, 3287-3294.
Li, H., D. Zhang, Y. Zhang & Y. Xu. 2008b. Research of image preprocessing methods
for EO-1 Hyperion hyperspectral data in tidal flat area. In Proceedings of SPIE,
the International Society for Optical Engineering Geoinformatics 2008 and
Joint Conference on GIS and Built Environment, 71471G.1-71471G.8.
Guangzhou, China.
Li, J. & W. Chen. 2005. A rule-based method for mapping Canada’s wetlands using
optical, radar and DEM data. International Journal of Remote Sensing, 26,
5051-5069.
Li, L., S. L. Ustin & M. Lay. 2005. Application of multiple endmember spectral
mixture analysis (MESMA) to AVIRIS imagery for coastal salt marsh mapping:
a case study in China Camp, CA, USA. International Journal of Remote
Sensing, 26, 5193-5207.
457
Li, X. & A. H. Strahler. 1986. Geometric– optical bidirectional reflectance modeling
of a conifer forest canopy. IEEE Transactions on Geoscience and Remote
Sensing, GE-24, 906-919.
Li, Y., T. H. Demetriades-Shah, E. T. Kanemasu, J. K. Shultis & M. B. kirkham. 1993.
Use of second derivatives of canopy reflectance for monitoring prairie
vegetation over different soil background. Remote Sensing of Environment,
44, 81-87.
Liang, J., S. Yang & A. Winstanley. 2008. Invariant optimal feature selection: a
distance discriminant and feature ranking based solution. Pattern
Recognition, 41, 1429-1439.
Liang, S. 2004. Quantitative Remote Sensing of Land Surfaces. Hoboken: John Wiley
& Sons.
Lichtenthaler, H. K., A. Gotelson & M. Lang. 1996. Non-destructive determination of
chlorophyll content of leaves of a green and an aurea mutant tobacco by
reflectance measurements. Journal of Plant Physiology, 148, 483-493.
Lillesaeter, O. 1982. Spectral reflectance of partly transmitting leaves: laboratory
measurements and mathematical modeling. Remote Sensing of Environment,
12, 247-254.
Lillesand, T. M., R. W. Kiefer & J. W. Chipman. 2004. Remote sensing and image
interpretation. New York: Wiley.
Lillesand, T. M., R. W. Kiefer & J. W. Chipman. 2008. Remote sensing and image
interpretation. Hoboken, NJ: John Wiley & Sons.
Liu, H. & H. Motoda. 1998a. Feature Extraction Construction and Selection, A Data
Mining Perspective. Norwell, MA,: Kluwer Academic Publisher.
Liu, H. & H. Motoda. 1998b. Feature selection for knowledge discovery and data
mining. Boston: Kluwer Academic.
Liu, H. & L. Yu. 2005. Toward integrating feature selection algorithms for
classification and clustering. IEEE Transactions on Knowledge and Data
Engineering, 17, 491-502.
Lopes, A., R. Touzi & E. Nezry. 1990. Adaptive speckle filters and scene
heterogeneity. IEEE Transactions on Geoscience and Remote Sensing, 28,
992-1000.
Lowe, D. & A. R. Webb. 1991. Optimized Feature Extraction and the Bayes Decision
in Feed-Forward Classifier Networks. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 13, 355-364.
458
Lu, D. 2006. The potential and challenge of remote sensing-based biomass
estimation. International Journal of Remote Sensing, 27, 1297-1328.
Lu, D., P. Mausel, E. Brondizio & E. Moran. 2004. Relationships between forest stand
parameters and Landsat Thematic Mapper spectral responses in the
Brazilian Amazon basin. Forest Ecology and Management, 198.
Lu, Z. & D. J. Meyer. 2002. Study of high SAR backscattering caused by an increase
of soil moisture over a sparsely vegetated area: implications for
characteristics of backscattering. International Journal of Remote Sensing, 23,
1063-1074.
Mangrove at Shenzhen Halved. September 16, 2006. Ming Pao Daily News.
Manson, F. J., N. R. Loneragan, I. M. McLeod & R. A. Kenyon. 2001. Assessing
techniques for estimating the extent of mangroves: Topographic maps,
aerial photographs and Landsat TM images. Marine and Freshwater
Research, 52, 787-792.
Marill, T. & D. M. Green. 1963. On the effectiveness of receptors in recognition
system. IEEE Transactions on Information Theory, 9, 11-17.
Martin, M. E., S. D. Newman, J. D. Aber & R. G. Congalton. 1998. Determining forest
species composition using high spectral resolution remote sensing data.
Remote Sensing of Environment, 65, 249-254.
Mather, P. M. 1999a. Computer Processing of Remotely-Sensed Images. Chichester:
John Wiley and Sons.
Mather, P. M. 1999b. Computer Processing of Remotely-Sensed Images: An
Introduction. Chichester, UK: John Wiley.
Maunsell, G. 1991. New Airport Master Plan, Final Report, Environmental Impact
Assessment. Hong Kong: Provisional Airport Authority.
Mausel, P. W., W. J. Kramber & J. K. Lee. 1990. Optimum band selection for
supervised classification of multispectral data. Photogrammetric Engineering
and Remote Sensing, 56, 55-60.
Meisel, W. S. 1972. Computer-oriented approaches to pattern recognition. New York:
Academic Press.
Melana, D. M., J. Atchue III, C. E. Yao, R. Edwards, E. E. Melana & H. I. Gonzales.
2000. Mangrove management handbook. Cebu, Philippines: Department of
Environment and Natural Resources, Manila, Philippines through the Coastal
Resource Management Project.
459
Melgani, F. & L. Bruzzone. 2004. Classification of hyperspectral remote sensing
images with support vector machines. IEEE Transactions on Geoscience and
Remote Sensing, 42, 1778-1790.
Melville, D. S. & B. Morton. 1983. Mai Po marshes. Hong Kong: World Wide Fund
Hong Kong.
Meroni, M., R. Colombo & C. Panigada. 2004. Inversion of a radiative transfer model
with hyperspectral observations for LAI mapping in poplar plantations.
Remote Sensing of Environment, 92, 195-206.
Meza, D. B. & G. A. Blackburn. 2003. Remote sensing of mangrove biophysical
properties: Evidence from a laboratory simulation of the possible effects of
background variation on spectral vegetation indices. International Journal of
Remote Sensing, 24, 53-73.
Michael, M. & W. C. Lin. 1973. Experimental study of information and inter-intra
class distance ratios and feature selection and orderings. IEEE Transactions
on Systems, Man and Cybernetics, 3.
Miles, V. V., L. P. Bobylev, S. V. Maximov, O. M. Johannessen & V. M. Pitulko. 2003.
An approach for assessing boreal forest conditions based on combined use
of satellite SAR and multi-spectral data. International Journal of Remote
Sensing, 24, 4447-4466.
Miller, J. R., E. W. Hare & J. Wu. 1990. Quantitative characterization of the
vegetation red edge reflectance. 1. An inverted Guassian reflectance model.
International Journal of Remote Sensing, 11, 1755-1773.
Mitra, P., C. A. Murthy & S. K. Pal. 2002. Unsupervised feature selection using
feature similarity. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 24, 301-312.
Mourioulis, P., R. O. Green & T. G. Chrien. 2000. Design of pushbroom imaging
spectrometers for optimum recovery of spectroscopic and spatial
information. Applied Optics, 39, 2210-2220.
Mutanga, O. & A. K. Skidmore. 2004. Narrow band vegetation indices overcome the
saturation problem in biomass estimation. International Journal of Remote
Sensing, 25, 3999-4014.
Myneni, R. B., S. Hoffman, Y. Knyazikhin, J. L. Privette, J. Glassy, Y. Tian, Y. Wang, X.
Song, Y. Zhang, G. R. Smith, A. Lotsch, M. Friedl, J. T. Morisette, P. Votava, R.
R. Nemani & S. W. Running. 2002. Global products of vegetation leaf area
and fraction absorbed PAR from year one of MODIS data. Remote Sensing of
Environment, 83, 214-231.
460
Myneni, R. B., R. R. Nemani & S. W. Running. 1997. Estimation of global leaf area
index and absorbed PAR using radiative transfer model. IEEE Transactions on
Geoscience and Remote Sensing, 35, 1380-1393.
Myneni, R. B. & D. L. Williams. 1994. On the relationship between FAPAR and NDVI.
Remote Sensing of Environment, 49, 200-211.
Nagelkerken, I., S. Blaber, S. Bouillon, P. Green, M. Haywood, L. G. Kirton, J.-O.
Meynecke, J. Pawlik, H. M. Penrose, A. Sasekumar & P. J. Somerfield. 2008.
The habitat function of mangroves for terrestrial and marina fauna: a review.
Aquatic Botany, 89, 155-185.
Nakariyakul, S. & D. P. Casasent. 2007. Adaptive branch and bound algorithm for
selecting optimal features. Pattern Recognition Letters, 28, 1415-1427.
Narendra, P. M. & K. Fukunaga. 1977. A branch and bound algorithm for feature
subset selection. IEEE Transactions on Computers, 26, 917-922.
Ndi Nyoungui, A., E. Tonye & A. Akono. 2002. Evaluation of speckle . ltering and
texture analysis methods for land cover classi. cation from SAR images.
International Journal of Remote Sensing, 23, 1895-1925.
Nelson, M. M. & W. T. Illingworth. 1991. A Practical Guide to Neural Nets. Reading,
Mass.: Addison-Wesley.
Nemani, R., L. Pierce, S. Running & L. Band. 1993. Forest ecosystem processes at the
watershed scale: Sensitivity to remotely-sensed leaf area index estimates.
International Journal of Remote Sensing, 14, 2519-2534.
Neumann, H. H., G. Den Hartog & R. H. Shaw. 1989. Leaf area measurements based
on hemispheric photographs and leaf-litter collection in deciduous forest
during autumn leaf fall. Agricultural and Forest Meteorology, 45, 325-345.
Nezry, E., E. Mougin, A. Lopes & J. P. Gastellu-Etchegorry. 1993. Tropical vegetation
mapping with combined visible and SAR spaceborne data. International
Journal of Remote Sensing, 14, 2165-2184.
Norman, J. M. & G. S. Campbell. 1989. Canopy structure. In Plant physiological
ecology : field methods and instrumentation, eds. R. W. Pearcy, J. Ehleringer,
H. A. Mooney & P. W. Rundel, 301-325. London: Chapman and Hall.
Oliver, C. & S. Quegan. 2004. Understanding Synthetic Aperture Radar Images.
Raleigh: SciTech Publishing Inc.
Ozesmi, S. L. & M. E. Bauer. 2002. Satellite remote sensing of wetlands. Wetlands
Ecology and Management, 10, 381-402.
Pal, M. 2005. Random forest classifiers for remote sensing classification.
International Journal of Remote Sensing, 26, 217-222.
461
Pal, M. & P. M. Mather. 2003. An assessment of the effectiveness of decision tree
methods for land cover classification. Remote Sensing of Environment, 86,
554-565.
Pal, M. & P. M. Mather. 2005. Support vector machines for classification in remote
sensing. International Journal of Remote Sensing, 26, 1007-1011.
Paola, J. D. & R. A. Schowengerdt. 1995. A review and analysis of backpropagation
neural networks for classification of remotely-sensed multi-spectral imagery.
International Journal of Remote Sensing, 16, 3033-3058.
Pasqualini, V., J. Iltis, N. Dessay, M. Lointier, O. Guelorget & L. Polidori. 1999.
Mangrove mapping in North-Western Madagascar using SPOT-XS and SIR-C
radar data. Hydrobiologia, 413, 127-133.
Patel, P., H. S. Srivastava, S. Panigrahy & J. S. Parihar. 2006. Comparative evaluation
of the sensitivity of multi-polarized multifrequency SAR backscatter to plant
density. International Journal of Remote Sensing, 27, 293-305.
Pearcy, R. W. 1989. Radiation and light measurements. In Plant physiological
ecology : field methods and instrumentation, eds. R. W. Pearcy, J. Ehleringer,
H. A. Mooney & P. W. Rundel, 95-116. London: Chapman and Hall.
Peng, H., F. Long & C. Ding. 2005. Feature selection based on mutual information:
criteria of max-dependency, max-relevance, and min-redundancy. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 27, 1226-1238.
Pengra, B. W., C. A. Johnston & T. R. Loveland. 2007. Mapping an invasive plant,
Phragmites australis, in coastal wetlands using the EO-1 Hyperion
hyperspectral sensor. Remote Sensing of Environment, 108, 74-81.
Peñuelas, I. Fielella, C. Biel, L. Serrano & R. Save. 1993. The reflectance at the 950970 nm region as an indicator of plant water status. International Journal of
Remote Sensing, 14, 1887-1905.
Pierce, L. L. & S. W. Running. 1988. Rapid estimation of coniferous forest leaf index
using a portable integrating radiometer. Ecology, 69.
Platt, J. C., N. Cristianini & Shawe-Taylor. 2000. Large margin DAGs for multiclass
classification. In Advances in Neural Information Processing Systems, eds. M.
I. Jordan, Y. LeCun & S. A. Solla, 547-553. Cambridge, MA: MIT Press.
Price, J. C. & W. C. Bausch. 1995. Leaf Area Index estimation from visible and nearinfrared reflectance data. Remote Sensing of Environment, 52, 55-65.
Pu, R. & P. Gong. 2004. Wavelet transform applied to EO-1 hyperspectral data for
forest LAI and crown closure mapping. Remote Sensing of Environment, 91,
212-224.
462
Pu, R., P. Gong, G. S. Biging & M. R. Larrieu. 2003. Extraction of red edge optical
parameters from Hyperion data for estimation of forest Leaf Area Index.
IEEE Transactions on Geoscience and Remote Sensing, 41, 916-921.
Pu, R., Q. Yu, P. Gong & G. S. Biging. 2005. EO-1 Hyperion, ALI and Landsat 7 ETM+
data comparison for estimating forest crown closure and leaf area index.
International Journal of Remote Sensing, 26, 457-474.
Pudil, P., J. Novovičová & J. Kittler. 1994. Floating search methods in feature
selection. Pattern Recognition Letters, 15, 1119-1125.
Pudil, P., J. Novovičová & P. Somol. 2002. Feature selection toolbox software
package. Pattern Recognition Letters, 23, 487-492.
Pudil, P. & P. Somol. 2005. Current feature selection techniques in statistical pattern
recognition. In Proceedings of the 4th International Conference on Computer
Recognition Systems CORES' 05, eds. M. Kurzynski, E. Puchała, M. Wozniak &
A. Zołnierek, 53-68. Rydzyna Castle, Poland: Springer-Verlag Berlin
Heidelberg.
Qi, J., F. Cabot, M. S. Moran & G. Dedieu. 1995. Biophysical Parameter Estimations
Using Multidirectional Spectral Measurements. Remote Sensing of
Environment, 54, 71-83.
Qi, J., A. Chehbouni, A. R. Huete, Y. H. Keer & S. Sorooshian. 1994a. A modified soil
vegetation adjusted index. Remote Sensing of Environment, 48, 119-126.
Qi, J., A. Chehbouni, A. R. Huete, Y. H. Kerr & S. Sorooshian. 1994b. A modified soil
adjusted vegetation index. Remote Sensing of Environment, 48, 119-126.
Quinlan, J. R. 1986. Induction of decision trees. Machine Learning, 1, 81-106.
Quinlan, J. R. 1993. C4.5: Programs for machine learning. San Mateo, Calif: Morgan
Kaufmann.
Rahman, H. 2001. Influence of atmospheric correction on the estimation of
biophysical parameters of crop canopy using satellite remote sensing.
International Journal of Remote Sensing, 22, 1245-1268.
Ramsey, E., A. Rangoonwala, G. Nelson & R. Ehrlich. 2005a. Mapping the invasive
species, Chinese tallow, with EO1 satellite Hyperion hyperspectral image
data and relating tallow occurrences to a classified Landsat Thematic
Mapper land cover map. International Journal of Remote Sensing, 26, 16371657.
Ramsey, E., A. Rangoonwala, G. Nelson, R. Ehrlich & K. Martella. 2005b. Generation
and validation of characteristic spectra from EO1 Hyperion image data for
detecting the occurrence of the invasive species, Chinese tallow.
International Journal of Remote Sensing, 26, 1611-1636.
463
Ramsey, E. W. & J. R. Jensen. 1995. Modeling mangrove canopy reflectance using a
light interaction model and an optimization technique. In Wetland and
environmental applications of GIS, eds. J. G. Lyon & J. Mc Carthy. Boca Raton:
CRC Press.
Ramsey III, E. W., G. A. Nelson & S. K. Sapkota. 1998. Classifying coastal resources by
integrating optical and radar imagery and color infrared photography.
Mangroves and Salt Marshes, 2, 109-119.
Raudys, S. J. 2006. Feature over-selection. In Structural, Syntactic, and Statistical
Pattern Recognition (LNCS 4109), 622-631. Belin/ Heidelberg, Germany:
Springer-Verlag.
Raudys, S. J. & A. K. Jain. 1991. Small Sample Size Effects in Statistical Pattern
Recognition: Recommendations for Practitioners. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 13, 252-264.
Raudys, S. J. & V. Pikelis. 1980. On Dimensionality, Sample Size, Classification Error,
and Complexity of Classification Algorithms in Pattern Recognition. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 2, 243-251.
Rauste, Y., T. Hame, J. Pulliainen, K. Heiska & M. Hallikainen. 1994. Radar-based
forest biomass estimation. International Journal of Remote Sensing, 15,
2797-2808.
Rautiainen, M. 2005. Retrieval of leaf area index for a coniferous forest by inverting
a forest reflectance model. Remote Sensing of Environment, 99, 295-303.
Rautiainen, M., P. Stenberg, T. Nilson & A. Kuusk. 2004. The effect of crown shape
on the reflectance of coniferous stands. . Remote Sensing of Environment, 89,
41-52.
Régent Instruments Inc. 2008. WinSCANOPY 2008a For Canopy Analysis. . Regent
Instruments Canada Inc.
Rich, P. M. 1990. Characterizing plant canopies with hemispherical photographs. .
Remote Sensing Reviews, 5, 13-29.
Richards, J. A. 1993. Remote Sensing Digital Image Analysis: An Introduction. New
York: Springer-Verlag.
Richards, J. A. 2009. Remote Sensing with Imaging Radar. Heidelberg: Springer.
Richardson, A. D. & G. P. Berlyn. 2002. Changes in foliar spectral reflectance and
chlorophyll fluorescence of four temperate species following branch cutting.
Tree Physiology, 22, 499-506.
464
Rignot, E. J., R. Zimmermann & J. J. van Zyl. 1995. Spacebome Applications of P Band
Imaging Radars for Measuring Forest Biomass. IEEE T ransactions on
Geoscience and Remote Sensing, 33, 1162-1169.
Ripple, W. J. 1986. Spectral reflectance relationships to leaf water stress.
Photogrammetric Engineering and Remote Sensing, 52, 1669-1675.
Roberts, G. 2001. A review of the application of BRDF models to infer land cover
parameters at regional and global scales. Progress in Physical Geography, 25,
483-511.
Rondeaux, G., M. Steven & F. Baret. 1996. Optimization of soil-adjusted vegetation
indices. Remote Sensing of Environment, 55, 95-107.
Rosema, A., W. Verhoef, H. Noorbergen & J. J. Borgesius. 1992. A new forest light
interaction model in support of forest monitoring. Remote Sensing of
Environment, 42, 23-41.
Ross, J. 1981. The radiation regime and architecture of plant stands. The Hague, The
Netherlands: Dr Junk W.
Rosso, P. H., S. L. Ustin & A. Hastings. 2005. Mapping marshland vegetation of San
Francisco Bay, California, using hyperspectral data. International Journal of
Remote Sensing, 26, 5169-5191.
Rougean, J. L. & F. M. Breon. 1995. Estimating PAR absorbed by vegetation from
bidirectional reflectance measurements. Remote Sensing of Environment, 51,
375-384.
Rouse, J. W., R. H. Haas, J. A. Schell & D. W. Deering. 1974. Monitoring vegetation
systems in the Great Plains with ERTS. In Proceedings Third ERTS Symposium,
NASA SP-351, 309-317. Washington, DC.
Rumelhart, D. E., G. E. Hinton & R. J. Williams. 1986. Learning internal
representation by error propagation. In Parallel distributed processing:
Explorations in the microstructures of cognition, 318-362. Cambridge, MA:
MIT Press.
Satyanarayana, B., K. A. Mohamad, I. F. Idris, M.-L. Husain & F. Dahdouh-Guebas.
2011. Assessment of mangrove vegetation based on remote sensing and
ground-truth measurements at Tumpat, Kelantan Delta, East Coast of
Peninsular Malaysia. International Journal of Remote Sensing, 32, 1635-1650.
Schlerf, M. & C. Atzberger. 2006. Inversion of a forest reflectance model to estimate
structural canopy variables from hyperspectral remote sensing data. Remote
Sensing of Environment, 100, 281-294.
465
Schlerf, M., C. Atzberger & J. Hill. 2005. Remote sensing of forest biophysical
variables using HyMap imaging spectrometer data. Remote Sensing of
Environment, 95, 177-194.
Schmid, T., M. Koch, J. Gumuzzio & P. M. Mather. 2004. A spectral library for a semiarid wetland and its application to studies of wetland degradation using
hyperspectral and multispectral data. International Journal of Remote
Sensing, 25, 2485-2496.
Schmidt, K. S. & A. K. Skidmore. 2003. Spectral discrimination of vegetation types in
a coastal wetland. Remote Sensing of Environment, 85, 92-108.
Schmidt, K. S., A. K. Skidmore, E. H. Kloosterman, O. H. Van, L. Kumar & J. A. M.
Janssen. 2004. Mapping coastal vegetation using an expert system and
hyperspectral imagery. Photogrammetric Engineering and Remote Sensing,
70, 703-715.
Schotten, C. G. J., v. R. W. W. L. & L. L. F. Janssen. 1995. Assessment of the
capabilities of multi-temporal ERS-1 SAR data to discriminate between
agricultural crops. International Journal of Remote Sensing, 14, 2619-2637.
Schowengerdt, R. A. 1983. Techniques for image processing and classification in
remote sensing. New York: Academic Press.
Serpico, S. B. & L. Bruzzone. 2001. A new search algorithm for feature selection in
hyperspectral remote sensing images. IEEE Transactions on Geoscience and
Remote Sensing, 39, 1360-1367.
Sezgin, M. & B. Sankur. 2004. Survey over image thresholding techniques and
quantitative performance evaluation. Journal of Electronic Imaging, 13, 146165.
Shanmugham, K., V. Narayanan, V. S. Frost, J. A. Stiles & J. C. Holtzman. 1981.
Textural features on radar image analysis. IEEE Proceedings on Geoscience
and Remote Sensing, GE-19, 153-156.
Shapira, Y. & I. Gath. 1999. Feature selection for multiple binary classification
problems. Pattern Recognition Letters, 20, 823-832.
Shaw, G. & D. Manolakis. 2002. Signal processing for hyperspectral image
exploitation. IEEE Signal Processing Magazine, 19, 12-16.
Sherrod, P. H. (2007). "DTREG Software for Predictive Modeling and Forecasting."
Retrieved 22 December, 2007, from http://www.dtreg.com/index.htm.
Siedlecki, W. & J. Sklansky. 1989. A note on genetic algorithms for large-scale
feature selection. Pattern Recognition Letters, 10, 335-347.
Sims, D. A. & J. A. Gamon. 2002. Relationships between leaf pigment
466
content and spectral reflectance across a wide range of species, leaf
structures and developmental stages. . Remote Sensing of Environment, 81, 331-354.
Sims, D. A. & J. A. Gamon. 2003. Estimation of vegetation water content and
photosynthetic tissue area from spectral reflectance: a comparison of
indices based on liquid water and chlorophyll absorption features. Remote
Sensing of Environment, 84, 526-537.
Smith, F. W., A. D. Sampson & N. J. Long. 1991. Comparison of leaf area index
estimates from tree allometrics and measured light interception. Forest
Science, 37, 1682-1688.
Smith, N. J., J. M. Chen & T. A. Black. 1993. Effects of clumping on estimates of stand
leaf area index using the Li-Cor LAI-2000. Canadian Journal of Forest
Research, 23, 1940-1943.
Soares, J. V., C. D. Renno, A. R. Formaggio, C. C. F. Yanasse & A. C. Frery. 1997. An
investigation of the selection of texture features for crop discrimination
using SAR imagery. Remote Sensing of Environment, 59, 234-247.
Solberg, A. H. S., A. K. Jain & T. Taxt. 1994. Multisource Classification of Remotely
Sensed Data: Fusion of Landsat TM and SAR Images. IEEE T ransactions on
Geoscience and Remote Sensing, 32, 768-778.
Somol, P. & J. Novovicova. 2008. Evaluating the stability of feature selectors that
optimize feature subset cardinality. In Structural, Syntactic, and Statistical
Pattern Recognition (LNCS 5342), 956-966. Heidelberg, Germany: SpringerVerlag.
Somol, P., J. Novovicova & P. Pudil. 2010a. Efficient feature subset selection and
subset size optimization. In Pattern Recognition, Recent Advances, ed. A.
Herout, 75-97. Vukovar, Croatia: In-Tech.
Somol, P. & P. Pudil. 2000. Oscillating search algorithms for feature selection. In
International Conference on Pattern Recognition, 2406-2409. Barcelona,
Spain: IEEE Computer Society.
Somol, P., P. Pudil & J. Kittler. 2004. Fast branch and bound algorithms for optimal
feature selection. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 26, 900-912.
Somol, P., P. Pudil, J. Novovicova & P. Paclik. 1999. Adaptive Floating Search
Methods in Feature Selection. Pattern Recognition Letters, 20, 1157-1163.
Somol, P., P. Vacha, S. Mikes, J. Hora, P. Pudil & P. Zid. 2010b. Introduction to
Feature Selection Toolbox 3 - The C++ library for subset search, data
modeling and classification. In Technical Report No. 2287. UTIA.
467
Spanner, M. A., L. L. Pierce, S. W. Running & D. L. Peterson. 1990. The seasonality of
AVHRR data of temperate coniferous forests: relationship with leaf area
index. Remote Sensing of Environment, 33, 97-118.
Staenz, K., J. Secker, B.-C. Gao, C. Davis & C. Nadeau. 2002. Radiative transfer codes
applied to hyperspectral data for retrieval of surface reflectance. ISPRS
Journal of Photogrammetry & Remote Sensing, 57, 194-203.
Stearns, S. D. 1976. On selecting features for pattern classifiers. In Third
International Conference on Pattern Recognition. Coronado, USA.
Sulong, I., H. Mohd-Lokman, K. Mohd-Tarmizi & A. Ismail. 2002. Mangrove mapping
using Landsat imagery and aerial photographs: Kemaman District,
Terengganu. Malaysia Environment. Development and Sustainability, 4, 135152.
Swain, P. H. & S. M. Davis. 1978. Remote Sensing: The Quantitative Approach.
London ; New York: McGraw-Hill International Book Co.
Tadjudin, S. & D. Landgrebe. 1998. Classification of High Dimensional Data with
Limited Training Samples. In School of Electrical and Computer Engineering,
129. West Lafayette, Indiana.: Purdue University.
Tam, N. F. Y., Y. S. Wong, C. Y. Lu & R. Berry. 1997. Mapping and characterization of
mangrove plant communities in Hong Kong. Hydrobiologia, 352, 25-37.
Thenkabail, P. S., R. B. Smith & E. De Pauw. 2000. Hyperspectral vegetation indices
and their relationships with agricultural crop characteristics. Remote Sensing
of Environment, 71, 158-182.
Todd, S. W., R. M. Hoffer & D. G. Milchunas. 1998. Biomass estimation on grazed
and ungrazed rangelands using spectral indices. International Journal of
Remote Sensing, 19, 427-438.
Treitz, P., M. & p. J. Howarth. 1999. Hyperspectral remote sensing for estimating
biophysical parameters of forest ecosystems. Progress in Physical Geography,
23, 359-390.
Tsai, F., E. K. Lin & K. Yoshino. 2007. Spectrally segmented principal component
analysis of hyperspectral imagery for mapping invasive plant species.
International Journal of Remote Sensing, 28, 1023-1039.
Tso, B. & P. M. Mather. 1999a. Crop discrimination using multi-temporal SAR
imagery. International Journal of Remote Sensing, 20, 2443-2460.
Tso, B. & P. M. Mather. 1999b. Crop discrimination using multi-temporal SAR
imagery. International Journal of Remote Sensing, 20, 2443-2460.
468
Tso, B. & P. M. Mather. 2009. Classification Method for Remotely Sensed Data. Boca
Raton: CRC Press.
Tu, T.-M., C.-H. Chen, J.-L. Wu & C.-I. Chang. 1998. A fast two-stage classification
method for high-dimensional remote sensing. IEEE Transactions on
Geoscience and Remote Sensing, 36, 182-191.
Underwood, E., S. Ustin & D. Dipietro. 2003. Mapping non-native plants using
hyperspectral imagery. Remote Sensing of Environment, 86, 150-161.
Vaiphasa, C., S. Ongsomwang, T. Vaiphasa & A. K. Skidmore. 2005. Tropical
mangrove species discrimination using hyperspectral data: A laboratory
study. . Estuarine, Coastal and Shelf Science, 65, 371-379.
Vaiphasa, C., A. K. Skidmore & W. F. de Boer. 2006. A post-classifier for mangrove
mapping using ecological data. Journal of Photogrammetry and Remote
Sensing, Article in Press.
Valiela, I., J. L. Bowen & J. K. York. 2001. Mangrove Forests: One of the World's
Threatened Major Tropical Environments. BioScience, 51, 807-815.
van der Sanden, J. J. & D. H. Hoekman. 1999. Potential of airborne radar to support
the assessment of land cover in a tropical rain forest environment. Remote
Sensing of Environment, 68, 26-40.
Van Gardingen, P. R., G. E. Jackson, S. Hernandez-Daumas, G. Russel & L. Sharp.
1999. Leaf area index estimates obtained for clumped canopies using
hemispherical photography. Agricultural and Forest Meteorology, 94, 243257.
Vane, G. & A. F. H. Goetz. 1993. Terrestrial imaging spectrometry: current status,
future trends. Remote Sensing of Environment, 44, 117-126.
Vapnik, V. 1995. The nature of statistical learning theory. New York: Springer-Verlag.
Vapnik, V. 1998. Statistical learning theory. New York: John Wiley.
Verdebout, J., S. Jacquemoud & G. Schmuch. 1994. Optical properties of leaves:
modeling and experimental studies. In Imaging Spectrometry - A Tool for
Environmental Observations, eds. J. Hill & J. Megier, 169-191. Dordrecht:
Kluwer Academic Publishers.
Verhoef, W. 1984. Light scattering by leaf layers with application to canopy
reflectance modeling: the SAIL model. Remote Sensing of Environment, 16,
125-141.
Verikas, A. & M. Bacauskiene. 2002. Feature selection with neural networks. Pattern
Recognition Letters, 23, 1323-1335.
469
Verstraete, M. M. 1994a. Retrieving canopy properties from remote sensing
measurements. In Imaging Spectrometry - A Tool for Environmental
Observations, eds. J. Hill & J. Megier, 109-123. Dordrecht: Kluwer Academic
Publishers.
Verstraete, M. M. 1994b. Scientific issues and instrumental opportunities in remote
sensing and high resolution spectrometry. In Imaging Spectrometry - A Tool
for Environmental Observations, eds. J. Hill & J. Megier, 25-38. Dordrecht:
Kluwer Academic Publishers.
Verstraete, M. M. & B. Pinty. 1996. Designing optimal spectral indices for remote
sensing applications. IEEE Transactions on Geoscience and Remote Sensing,
34, 1254-1265.
Vogelmann, J. E. & D. M. Moss. 1993. Spectral reflectance measurement in the
genus Sphagnum. Remote Sensing of Environment, 45, 273-279.
Vogelmann, J. E., B. N. Rock & D. M. Moss. 1993. Red edge spectral measurements
from
sugar maple leaves. International Journal of Remote Sensing, 14, 1563-1575.
Waite, W. P., H. C. MacDonald, V. H. Kaupp & J. S. Demarcke. 1981. Wetland
mapping with imaging radar. In International Geoscience and Remote
Sensing Symposium (IGARSS'81), 794-799. Niagara Falls, New York.
Wakebayashi, H. & K. Arai. 1996. A new method for speckle noise reduction (CST
filter). Canadian Journal of Remote Sensing, 22, 190-197.
Wang, Y., F. W. Davis, J. M. Melack, E. S. Kasischke & J. N. L. Christensen. 1995. The
Effects of Changes in Forest Biomass on Radar Backscatter from Tree
Canopies. International Journal of Remote Sensing, 16, 503-513.
Wang, Y. & M. L. Imhoff. 1993. Simulated and observed L-HH radar backscatter from
tropical mangrove forests. International Journal of Remote Sensing, 14,
2819-2828.
Wang, Y. Y. & J. Li. 2008. Feature-selection ability of the decision-tree algorithm and
the impact of feature-selection/ extraction on decision-tree results based
on hyperspectral data. International Journal of Remote Sensing, 29, 29933010.
Weiss, M., F. Baret, G. J. Smith, I. Jonckheere & P. Coppin. 2004. Review of methods
for in situ leaf area index (LAI) determination part II - estimation of LAI,
errors and sampling. Agricultural and Forest Meteorology, 121, 37-53.
Welles, J. M. 1990. Some indirect methods of estimating canopy structure. Remote
Sensing of Environment, 5, 31-43.
470
Welles, J. M. & S. Cohen. 1996. Canopy structure measurement by gap fraction
analysis using commercial instrumentation. Journal of Experimental Botany,
47, 1335-1342.
Welles, J. M. & J. M. Norman. 1991a. Instrument for indirect measurement of
canopy architecture. Agronomy Journal, 83, 818-825.
Welles, J. M. & J. M. Norman. 1991b. Instrument for indirect measurement of
canopy architecture. Agronomy Journal, 83, 818-825.
Wessman, C. A. 1994a. Estimating canopy biochemistry through imaging
spectrometry. In Imaging Spectrometry - A Tool for Environmental
Observations, eds. J. Hill & J. Megier, 57-69. Dordrecht: Kluwer Academic
Publishers.
Wessman, C. A. 1994b. Remote sensing and the estimation of ecosystem
parameters and functions. In Imaging Spectrometry - A Tool for
Environmental Observations, eds. J. Hill & J. Megier, 39-56. Dordrecht:
Kluwer Academic Publishers.
Whitford, K. R., I. J. Colquhoun, A. R. G. Lang & B. M. Harper. 1995. Measuring leaf
area index in a sparse eucalypt forest: a comparison of estimates from direct
measurement, hemispherical photography, sunlight transmittance and
allometric regression. Agricultural and Forest Meteorology, 74, 237-249.
Whitmore, T. C., N. D. Brown, M. D. Swaine, D. Kennedy, C. I. Goodwin-Bailey & W.K. Gong. 1993. Use of hemispherical photographs in forest ecology:
measurement of gap size and radiation totals in a Bornean tropical rain
forest. Journal of Tropical Ecology, 9, 131-151.
Whitney, A. W. 1971. A direct method of nonparametric measurement selection.
IEEE Transactions on Computers, 20, 1100-1103.
Wilkinson, G. G., S. Folving, I. Kanellopoulos, N. McCormick, K. Fullerton & J. Mégier.
1995. Forest mapping from multi-source satellite data using neural network
classifiers: An experiment in Portugal. Remote Sensing Reviews, 12, 83-106.
Wolf, L. & A. Shashua. 2005. Feature selection for unsupervised and supervised
inference: the emergence of sparsity in a weight-based approach. Journal of
Machine Learning Research, 6, 1855-1887.
Wooding, M. G., A. D. Zmuda & G. H. Griffiths. 1993. Crop discrimination using
multi-temporal ERS-1 SAR data. In Proceedings of the Second ERS-1
Symposium - Space at the Service of our Environment, 51- 56. Hamburg,
Germany.
WWF Hong Kong. (2007). "Mai Po."
Retrieved September 29, 2007, from
http://www.wwf.org.hk/eng/maipo.
471
Yang, H. H. & J. Moody. 1999. Feature selection based on joint mutual information.
In Proceedings of International ICSC Symposium on Advances in Intelligent
Data Analysis, 342-349. Rochester, New York, USA.
Young, L. 1999. Management plan for the Mai Po marshes wildlife education centre
and nature reserve 2000-2004. Hong Kong: WWF Hong Kong.
Young, M., B. Rajagopalan & U. Lall. 1995. Estimation of mutual information using
kernel density estimators. Physical Review E, 52, 2318-2321.
Young, T. Y. & K. S. Fu. 1986. Handbook of Pattern Recognition and Image
Processing. San Diego: Academic Press.
Yu, B. & B. Yuan. 1993. A more efficient branch and bound algorithm for feature
selection. Pattern Recognition, 26, 883-889.
Zarco-Tejada, P. J., J. R. Miller, T. L. Noland, G. H. Mohammed & P. H. Sampson.
2001. Scaling-up and model inversion methods with narrow-band optical
indices for chlorophyll content estimation in closed forest canopies with
hyperspectral data. . IEEE Transactions on Geoscience and Remote Sensing,
39, 1491-1507.
Zarco-Tejada, P. J. & G. Sepulcre-Cantó. 2007. Remote sensing of vegetation
biophysical parameters for detecting stress condition and land cover changes.
http://www.zonanosaturada.com/publics/ZNS07/index.html (last accessed
18 July 2011).
Zhang, M., S. L. Ustin, E. Rejmankova & E. W. Sanderson. 1997. Monitoring Pacific
coast salt marshes using remote sensing. Ecological Applications, 7, 10391053.
Zhang, R. & J. Ma. 2009. Feature selection for hyperspectral data based on recursive
support vector machies. International Journal of Remote Sensing, 30, 36693677.
Zhang, Y., J. M. Chen & J. R. Miller. 2005. Determining digital hemispherical
photograph exposure for leaf area index estimation. Agricultural and Forest
Meteorology, 133, 166-181.
Zhao, D., L. Huang, J. Li & J. Qi. 2007. A comparative analysis of broadband and
narrowband derived vegetation indices in predicting LAI and CCD of a cotton
canopy. ISPRS Journal of Photogrammetry and Remote Sensing, 62, 25-33.
Zhao, W., G. Ji, R. Nian & C. Feng. 2005. SVM Classification method based marginal
points of representative sample sets. International Journal of Information
Technology, 11, 1-10.
Zhu, L. & R. Tateishi. 2006. Fusion of multisensor multitemporal satellite data for
land cover mapping. International Journal of Remote Sensing, 27, 903-918.
472
Appendix A
Geometric correction of hyperspectral data
Table A1. Residuals of ground control points collected for Hyperion Image
Point ID
Residual
Residual X
Residual Y
Photo X
Photo Y
G0001
0.08
0.08
0.01
19.50
2200.00
G0002
0.07
-0.06
-0.02
144.70
2212.40
G0003
0.11
0.00
0.11
106.70
1703.90
G0004
0.06
0.05
0.05
190.70
1736.80
G0005
0.08
0.01
-0.08
169.70
1583.70
G0006
0.05
0.04
-0.02
166.30
1505.10
G0008
0.09
-0.06
-0.07
42.70
2625.90
G0009
0.09
-0.03
0.09
218.80
2360.10
G0010
0.09
-0.03
-0.08
121.70
1978.80
G0012
0.04
-0.03
-0.02
216.80
1665.10
G0013
0.09
-0.05
-0.08
43.80
2017.30
G0014
0.10
0.00
-0.10
45.30
1988.50
G0015
0.03
-0.03
0.00
153.50
1623.30
G0016
0.06
0.05
-0.03
223.00
1565.80
G0017
0.04
0.02
0.03
217.30
2558.90
G0018
0.09
0.09
0.01
56.70
2787.20
G0019
0.03
-0.03
0.00
133.80
2505.10
G0020
0.07
-0.05
-0.05
186.80
2008.90
G0021
0.05
0.00
-0.05
46.30
1931.80
G0022
0.02
-0.02
0.00
54.90
1678.60
G0023
0.09
-0.05
0.08
50.60
1578.10
G0024
0.08
0.00
0.08
146.10
1535.70
G0025
0.07
0.06
-0.04
186.10
1565.60
G0026
0.05
0.03
0.04
68.40
2378.40
G0027
0.09
0.04
0.09
50.00
2313.60
G0028
0.03
-0.02
-0.02
219.90
2025.70
G0029
0.05
-0.02
0.05
54.30
1851.80
G0030
0.02
0.01
0.02
93.70
1764.70
Overall X RMS: 0.04 Overall Y RMS: 0.06
473
Table A2. Residuals of ground control points for SAR image acquired on 19
November 2008
Point ID
Residual
Residual X
Residual Y
Photo X
Photo Y
G0001
0.03
-0.02
-0.02
2272.20
1834.80
G0002
0.02
0.01
-0.01
2047.70
2132.40
G0003
0.07
0.04
-0.05
1457.60
2556.70
G0004
0.06
-0.05
0.03
1745.40
2974.10
G0005
0.03
0.03
-0.01
1619.40
3470.60
G0006
0.05
0.01
-0.05
4318.30
2175.90
G0007
0.07
0.01
0.07
3012.10
3293.90
G0008
0.02
-0.02
-0.01
4364.70
3892.40
G0009
0.07
-0.01
-0.07
2403.10
4134.10
G0010
0.09
0.02
0.09
2924.10
1784.10
G0011
0.05
-0.04
0.04
2729.20
1723.30
G0012
0.05
0.05
-0.01
2661.40
1901.40
G0013
0.03
-0.03
0.01
2279.50
2309.60
G0014
0.02
0.00
-0.02
2140.90
3409.80
G0015
0.09
0.03
0.08
3265.40
3319.50
G0016
0.02
0.02
-0.01
3203.90
3751.40
G0017
0.07
0.01
-0.07
3540.90
2022.00
G0018
0.05
-0.04
-0.01
3238.80
1640.60
G0019
0.04
0.01
0.04
4049.30
2591.50
G0020
0.06
0.05
-0.02
2479.10
2320.20
G0021
0.03
-0.02
-0.01
5037.10
1690.10
G0022
0.05
-0.04
0.02
887.40
4223.70
Overall X RMS: 0.03 Overall Y RMS: 0.04
474
Appendix B
Scripts Derived from Feature Selection Toolbox (FST) for Feature Selection
Search algorithm: Sequential Forward Selection (SFS)/ Evaluation criterion: k-nearest-neighbor (KNN)/ d=5
#include <boost/smart_ptr.hpp>
#include <exception>
#include <iostream>
#include <cstdlib>
#include <string>
#include <vector>
#include "error.hpp"
#include "global.hpp"
#include "subset.hpp"
#include "data_intervaller.hpp"
#include "data_splitter.hpp"
//#include "data_splitter_5050.hpp"
#include "data_splitter_cv.hpp"
//#include "data_splitter_holdout.hpp"
//#include "data_splitter_leave1out.hpp"
//#include "data_splitter_resub.hpp"
#include "data_splitter_randrand.hpp"
//#include "data_splitter_randfix.hpp"
#include "data_scaler.hpp"
#include "data_scaler_void.hpp"
//#include "data_scaler_to01.hpp"
//#include "data_scaler_white.hpp"
#include "data_accessor_splitting_memTRN.hpp"
#include "data_accessor_splitting_memARFF.hpp"
475
//#include "criterion_normal_bhattacharyya.hpp"
//#include "criterion_normal_gmahalanobis.hpp"
//#include "criterion_normal_divergence.hpp"
//#include "criterion_multinom_bhattacharyya.hpp"
#include "criterion_wrapper.hpp"
#include "criterion_wrapper_bias_estimate.hpp"
//#include "criterion_subsetsize.hpp"
//#include "criterion_sumofweights.hpp"
//#include "criterion_negative.hpp"
#include "distance_euclid.hpp"
//#include "distance_L1.hpp"
//#include "distance_Lp.hpp"
#include "classifier_knn.hpp"
//#include "classifier_normal_bayes.hpp"
//#include "classifier_multinom_naivebayes.hpp"
//#include "classifier_svm.hpp"
476
//#include "search_bif.hpp"
//#include "search_bif_threaded.hpp"
//#include "search_monte_carlo.hpp"
//#include "search_monte_carlo_threaded.hpp"
//#include "search_exhaustive.hpp"
//#include "search_exhaustive_threaded.hpp"
//#include "branch_and_bound_predictor_averaging.hpp"
//#include "search_branch_and_bound_basic.hpp"
//#include "search_branch_and_bound_improved.hpp"
//#include "search_branch_and_bound_partial_prediction.hpp"
//#include "search_branch_and_bound_fast.hpp"
#include "seq_step_straight.hpp"
//#include "seq_step_straight_threaded.hpp"
//#include "seq_step_hybrid.hpp"
//#include "seq_step_ensemble.hpp"
#include "search_seq_sfs.hpp"
//#include "search_seq_sffs.hpp"
//#include "search_seq_sfrs.hpp"
//#include "search_seq_os.hpp"
//#include "search_seq_dos.hpp"
//#include "result_tracker_dupless.hpp"
//#include "result_tracker_regularizer.hpp"
#include "result_tracker_feature_stats.hpp"
#include "result_tracker_stabileval.hpp"
int main()
{
try{
typedef double RETURNTYPE;
typedef double DATATYPE; typedef double REALTYPE;
typedef unsigned int IDXTYPE; typedef unsigned int DIMTYPE; typedef short BINTYPE;
typedef FST::Subset<BINTYPE, DIMTYPE> SUBSET;
typedef FST::Data_Intervaller<std::vector<FST::Data_Interval<IDXTYPE> >,IDXTYPE> INTERVALLER;
typedef boost::shared_ptr<FST::Data_Splitter<INTERVALLER,IDXTYPE> > PSPLITTER;
typedef FST::Data_Splitter_CV<INTERVALLER,IDXTYPE> SPLITTERCV;
typedef FST::Data_Splitter_RandomRandom<INTERVALLER,IDXTYPE,BINTYPE> SPLITTERRANDRAND;
typedef FST::Data_Accessor_Splitting_MemTRN<DATATYPE,IDXTYPE,INTERVALLER> DATAACCESSOR;
typedef FST::Distance_Euclid<DATATYPE,DIMTYPE,SUBSET> DISTANCE;
typedef FST::Classifier_kNN<RETURNTYPE,DATATYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR,DISTANCE> CLASSIFIERKNN;
typedef FST::Criterion_Wrapper<RETURNTYPE,SUBSET,CLASSIFIERKNN,DATAACCESSOR> WRAPPERKNN;
typedef FST::Criterion_Wrapper_Bias_Estimate<RETURNTYPE,SUBSET,CLASSIFIERKNN,DATAACCESSOR> WRAPPERBIAS;
typedef FST::Sequential_Step_Straight<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERKNN> EVALUATOR;
typedef FST::Result_Tracker_Stability_Evaluator<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKER;
typedef FST::Result_Tracker_Feature_Stats<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKERSTATS;
477
std::cout << "Wrapper-based (3NN) feature selection with Sequential Forward Search, 10-fold with Stability Evaluator d=5" << std::endl;
// print console output to log.txt
freopen("log_Wrapper01e_SFS_with stabilityeval_d=5.txt","w",stdout);
std::cout << "Wrapper-based (3NN) feature selection with Sequential Forward Search, 10-fold with Stability Evaluator d=5 \n" << std::endl;
// use randomly chosen 70% of data for training and keep the other 30% for independent testing of final classification performance
PSPLITTER dsp_outer(new SPLITTERRANDRAND(20/*splits=trials*/,70,30)); // (there will be one outer randomized split only)
// in the course of search use the first half of data by 10-fold cross-validation in wrapper FS criterion evaluation
PSPLITTER dsp_inner(new SPLITTERCV(10));
// do not scale data
boost::shared_ptr<FST::Data_Scaler<DATATYPE> > dsc(new FST::Data_Scaler_void<DATATYPE>());
// set-up data access
478
boost::shared_ptr<std::vector<PSPLITTER> > splitters(new std::vector<PSPLITTER>);
splitters->push_back(dsp_outer); splitters->push_back(dsp_inner);
boost::shared_ptr<DATAACCESSOR> da(new DATAACCESSOR("species_remoutl_normal_185_7_309.trn",splitters,dsc));
da->initialize();
// initiate access to split data parts
da->setSplittingDepth(0); if(!da->getFirstSplit()) throw FST::fst_error("70/30 data split failed.");
da->setSplittingDepth(1); if(!da->getFirstSplit()) throw FST::fst_error("10-fold cross-validation failure.");
// initiate the storage for subset to-be-selected
boost::shared_ptr<SUBSET> sub(new SUBSET(da->getNoOfFeatures())); sub->deselect_all();
// set-up 3-Nearest Neighbor classifier based on Euclidean distances
boost::shared_ptr<CLASSIFIERKNN> cknn(new CLASSIFIERKNN); cknn->set_k(3);
// wrap the 3-NN classifier to enable its usage as FS criterion (criterion value will be estimated by 10-fold cross-val.)
boost::shared_ptr<WRAPPERKNN> wknn(new WRAPPERKNN);
wknn->initialize(cknn,da);
// set-up the standard sequential search step object (option: hybrid, ensemble, etc.)
boost::shared_ptr<EVALUATOR> eval(new EVALUATOR);
// set-up Sequential Forward Selection search procedure
FST::Search_SFS<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERKNN,EVALUATOR> srch(eval);
srch.set_search_direction(FST::FORWARD); // try FST::BACKWARD
// set target subset size
DIMTYPE target_subsize=5;
// set-up tracker to gather data for computation
boost::shared_ptr<TRACKERSTATS> trackerstats(new TRACKERSTATS);
srch.enable_result_tracking(trackerstats);
// set-up result tracker to collect results of each trial
boost::shared_ptr<TRACKER> tracker(new TRACKER);
// run the trials
std::cout << "Feature selection setup:" << std::endl << *da << std::endl << srch << std::endl << *wknn << std::endl << *tracker << std::endl << std::endl;
RETURNTYPE critval_train, critval_test, biasval;
srch.set_output_detail(FST::SILENT); // set FST::SILENT to disable all text output in the course of search (FST::NORMAL is default)
da->setSplittingDepth(0);
unsigned int trial=0;
bool run=da->getFirstSplit(); if(!run) throw FST::fst_error("RandRand data split failed.");
while(run)
{
479
trial++; std::cout << std::endl<<"TRIAL "<<trial<< " ---------------------------------------------------------------------"<<std::endl;
da->setSplittingDepth(1);
if(!srch.search(target_subsize,critval_train,sub,wknn,std::cout)) throw FST::fst_error("Search not finished."); //d-parameterized (d=5)
tracker->add(critval_train,sub);
std::cout << std::endl << "(TRIAL "<<trial<<") Search result: " << std::endl << *sub << "Criterion value=" << critval_train << std::endl;
// estimate classifier bias on training data with respect to the selected feature subset
boost::shared_ptr<WRAPPERBIAS> wbias(new WRAPPERBIAS);
wbias->initialize(cknn,da);
//da->setSplittingDepth(0);
if(!wbias->evaluate(biasval,sub)) throw FST::fst_error("Bias estimation failed.");
std::cout << "Estimated bias=" << biasval << std::endl << std::endl;
// (optionally) validate result by estimating kNN accuracy on selected feature sub-space on independent test data
da->setSplittingDepth(0);
cknn->train(da,sub);
cknn->test(critval_test,da);
std::cout << "Validated "<<cknn->get_k()<<"-NN accuracy=" << critval_test << std::endl << std::endl;
// (optionally) list the best known solutions for each cardinality as recorded throughout the course of search
std::cout << "Best recorded solution for subset size:" << std::endl;
for(DIMTYPE d=1;d<=sub->get_n();d++)
if(srch.get_result(d,critval_train,sub)) std::cout << d << ": val="<< critval_train << ", "<<*sub << std::endl;
da->setSplittingDepth(0);
run=da->getNextSplit();
}
// evaluate stability using results collected by tracker
std::cout<<std::endl;
std::cout << "---------------------------------------------------------------------" << std::endl;
std::cout << "Resulting criterion values' mean: " << tracker->value_mean() << ", standard deviation: " << tracker->value_stddev() << std::endl;
std::cout << "Resulting subset sizes' mean: " << tracker->size_mean() << ", standard deviation: " << tracker->size_stddev() << std::endl;
std::cout << std::endl;
std::cout << "stability measure CWrel("<<da->getNoOfFeatures()<<")=" << tracker->stability_CWrel(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure ATI()=" << tracker->stability_ATI() << std::endl;
std::cout << "stability measure CW()=" << tracker->stability_CW() << std::endl;
std::cout << "stability measure ANHI("<<da->getNoOfFeatures()<<")=" << tracker->stability_ANHI(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure SH()=" << tracker->stability_SH() << std::endl;
std::cout << "stability measure PH("<<da->getNoOfFeatures()<<")=" << tracker->stability_PH(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure C()=" << tracker->stability_C() << std::endl;
// compute statistics
trackerstats->compute_stats();
// (optionally) print computation statistics
trackerstats->print_stats(std::cout);
}
catch(FST::fst_error &e) {std::cerr<<"FST ERROR: "<< e.what() << ", code=" << e.code() << std::endl;}
catch(std::exception &e) {std::cerr<<"non-FST ERROR: "<< e.what() << std::endl;}
return 0;
}
480
Search algorithm: Oscillating Search (OS)/ Evaluation criterion: k-nearest-neighbor (KNN)/ d=5
#include <boost/smart_ptr.hpp>
#include <exception>
#include <iostream>
#include <cstdlib>
#include <string>
#include <vector>
#include "error.hpp"
#include "global.hpp"
#include "subset.hpp"
#include "data_intervaller.hpp"
#include "data_splitter.hpp"
//#include "data_splitter_5050.hpp"
#include "data_splitter_cv.hpp"
//#include "data_splitter_holdout.hpp"
//#include "data_splitter_leave1out.hpp"
//#include "data_splitter_resub.hpp"
#include "data_splitter_randrand.hpp"
//#include "data_splitter_randfix.hpp"
#include "data_scaler.hpp"
#include "data_scaler_void.hpp"
//#include "data_scaler_to01.hpp"
//#include "data_scaler_white.hpp"
#include "data_accessor_splitting_memTRN.hpp"
#include "data_accessor_splitting_memARFF.hpp"
481
//#include "criterion_normal_bhattacharyya.hpp"
//#include "criterion_normal_gmahalanobis.hpp"
//#include "criterion_normal_divergence.hpp"
//#include "criterion_multinom_bhattacharyya.hpp"
#include "criterion_wrapper.hpp"
#include "criterion_wrapper_bias_estimate.hpp"
//#include "criterion_subsetsize.hpp"
//#include "criterion_sumofweights.hpp"
//#include "criterion_negative.hpp"
#include "distance_euclid.hpp"
//#include "distance_L1.hpp"
//#include "distance_Lp.hpp"
#include "classifier_knn.hpp"
//#include "classifier_normal_bayes.hpp"
//#include "classifier_multinom_naivebayes.hpp"
//#include "classifier_svm.hpp"
//#include "search_bif.hpp"
//#include "search_bif_threaded.hpp"
//#include "search_monte_carlo.hpp"
//#include "search_monte_carlo_threaded.hpp"
//#include "search_exhaustive.hpp"
//#include "search_exhaustive_threaded.hpp"
//#include "branch_and_bound_predictor_averaging.hpp"
//#include "search_branch_and_bound_basic.hpp"
//#include "search_branch_and_bound_improved.hpp"
//#include "search_branch_and_bound_partial_prediction.hpp"
//#include "search_branch_and_bound_fast.hpp"
#include "seq_step_straight.hpp"
//#include "seq_step_straight_threaded.hpp"
//#include "seq_step_hybrid.hpp"
//#include "seq_step_ensemble.hpp"
//#include "search_seq_sfs.hpp"
//#include "search_seq_sffs.hpp"
//#include "search_seq_sfrs.hpp"
#include "search_seq_os.hpp"
//#include "search_seq_dos.hpp"
//#include "result_tracker_dupless.hpp"
//#include "result_tracker_regularizer.hpp"
#include "result_tracker_feature_stats.hpp"
#include "result_tracker_stabileval.hpp"
482
int main()
{
try{
typedef double RETURNTYPE;
typedef double DATATYPE; typedef double REALTYPE;
typedef unsigned int IDXTYPE; typedef unsigned int DIMTYPE; typedef short BINTYPE;
typedef FST::Subset<BINTYPE, DIMTYPE> SUBSET;
typedef FST::Data_Intervaller<std::vector<FST::Data_Interval<IDXTYPE> >,IDXTYPE> INTERVALLER;
typedef boost::shared_ptr<FST::Data_Splitter<INTERVALLER,IDXTYPE> > PSPLITTER;
typedef FST::Data_Splitter_CV<INTERVALLER,IDXTYPE> SPLITTERCV;
typedef FST::Data_Splitter_RandomRandom<INTERVALLER,IDXTYPE,BINTYPE> SPLITTERRANDRAND;
typedef FST::Data_Accessor_Splitting_MemTRN<DATATYPE,IDXTYPE,INTERVALLER> DATAACCESSOR;
typedef FST::Distance_Euclid<DATATYPE,DIMTYPE,SUBSET> DISTANCE;
typedef FST::Classifier_kNN<RETURNTYPE,DATATYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR,DISTANCE> CLASSIFIERKNN;
typedef FST::Criterion_Wrapper<RETURNTYPE,SUBSET,CLASSIFIERKNN,DATAACCESSOR> WRAPPERKNN;
typedef FST::Criterion_Wrapper_Bias_Estimate<RETURNTYPE,SUBSET,CLASSIFIERKNN,DATAACCESSOR> WRAPPERBIAS;
typedef FST::Sequential_Step_Straight<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERKNN> EVALUATOR;
typedef FST::Result_Tracker_Stability_Evaluator<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKER;
typedef FST::Result_Tracker_Feature_Stats<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKERSTATS;
483
std::cout << "Wrapper-based (3NN) feature selection with Oscillating Search, 10-fold with Stability Evaluation d=5" << std::endl;
// print console output to log.txt
freopen("log_Wrapper01b_OS_with stabilityeval_d=5.txt","w",stdout);
std::cout << "Wrapper-based (3NN) feature selection with Oscillating Search, 10-fold with Stability Evaluation d=5 \n" << std::endl;
// use randomly chosen 70% of data for training and keep the other 30% for independent testing of final classification performance
PSPLITTER dsp_outer(new SPLITTERRANDRAND(20/*splits=trials*/,70,30)); // (there will be one outer randomized split only)
// in the course of search use the first half of data by 10-fold cross-validation in wrapper FS criterion evaluation
PSPLITTER dsp_inner(new SPLITTERCV(10));
// do not scale data
boost::shared_ptr<FST::Data_Scaler<DATATYPE> > dsc(new FST::Data_Scaler_void<DATATYPE>());
// set-up data access
boost::shared_ptr<std::vector<PSPLITTER> > splitters(new std::vector<PSPLITTER>);
splitters->push_back(dsp_outer); splitters->push_back(dsp_inner);
boost::shared_ptr<DATAACCESSOR> da(new DATAACCESSOR("species_remoutl_normal_185_7_309.trn",splitters,dsc));
da->initialize();
// initiate access to split data parts
da->setSplittingDepth(0); if(!da->getFirstSplit()) throw FST::fst_error("70/30 data split failed.");
da->setSplittingDepth(1); if(!da->getFirstSplit()) throw FST::fst_error("10-fold cross-validation failure.");
484
// initiate the storage for subset to-be-selected
boost::shared_ptr<SUBSET> sub(new SUBSET(da->getNoOfFeatures())); sub->deselect_all();
boost::shared_ptr<SUBSET> submax(new SUBSET(da->getNoOfFeatures()));
// set-up 3-Nearest Neighbor classifier based on Euclidean distances
boost::shared_ptr<CLASSIFIERKNN> cknn(new CLASSIFIERKNN); cknn->set_k(3);
// wrap the 3-NN classifier to enable its usage as FS criterion (criterion value will be estimated by 10-fold cross-val.)
boost::shared_ptr<WRAPPERKNN> wknn(new WRAPPERKNN);
wknn->initialize(cknn,da);
// set-up the standard sequential search step object (option: hybrid, ensemble, etc.)
boost::shared_ptr<EVALUATOR> eval(new EVALUATOR);
// set-up Oscillating Selection search procedure
FST::Search_OS<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERKNN,EVALUATOR> srch(eval);
srch.set_delta(5);
// target subset size must be set because Bhattacharyya is monotonous with respect to subset size (i.e., evaluates full set as the best)
DIMTYPE target_subsize=5;
// set-up tracker to gather data for computation
boost::shared_ptr<TRACKERSTATS> trackerstats(new TRACKERSTATS);
srch.enable_result_tracking(trackerstats);
// set-up result tracker to collect results of each trial
boost::shared_ptr<TRACKER> tracker(new TRACKER);
// run the trials
std::cout << "Feature selection setup:" << std::endl << *da << std::endl << srch << std::endl << *wknn << std::endl << *tracker << std::endl << std::endl;
RETURNTYPE critval_train, critval_trainmax, critval_test, biasval;
srch.set_output_detail(FST::NORMAL); // set FST::SILENT to disable all text output in the course of search (FST::NORMAL is default)
da->setSplittingDepth(0);
unsigned int trial=0;
bool run=da->getFirstSplit(); if(!run) throw FST::fst_error("RandRand data split failed.");
while(run)
{
trial++; std::cout << std::endl<<"TRIAL "<<trial<< " ---------------------------------------------------------------------"<<std::endl;
da->setSplittingDepth(1);
int non_improving_runs=-1; // -1 indicates first run
const int max_non_improving_runs=5;
do {
sub->make_random_subset(target_subsize);
485
if(!srch.search(target_subsize,critval_train,sub,wknn,std::cout)) throw FST::fst_error("Search not finished."); //d=5
if(non_improving_runs==-1 || critval_train>critval_trainmax) {
critval_trainmax=critval_train;
submax->stateless_copy(*sub);
non_improving_runs=0;
} else non_improving_runs++;
} while(non_improving_runs<max_non_improving_runs-1);
tracker->add(critval_train,sub);
std::cout << std::endl << "(TRIAL "<<trial<<") Search result: " << std::endl << *submax << "Criterion value=" << critval_trainmax << std::endl;
// estimate classifier bias on training data with respect to the selected feature subset
boost::shared_ptr<WRAPPERBIAS> wbias(new WRAPPERBIAS);
wbias->initialize(cknn,da);
//da->setSplittingDepth(0);
if(!wbias->evaluate(biasval,sub)) throw FST::fst_error("Bias estimation failed.");
std::cout << "Estimated bias=" << biasval << std::endl << std::endl;
// (optionally) validate result by estimating kNN accuracy on selected feature sub-space on independent test data
da->setSplittingDepth(0);
cknn->train(da,submax);
cknn->test(critval_test,da);
std::cout << "Validated "<<cknn->get_k()<<"-NN accuracy=" << critval_test << std::endl << std::endl;
da->setSplittingDepth(0);
run=da->getNextSplit();
}
// evaluate stability using results collected by tracker
std::cout<<std::endl;
std::cout << "---------------------------------------------------------------------" << std::endl;
std::cout << "Resulting criterion values' mean: " << tracker->value_mean() << ", standard deviation: " << tracker->value_stddev() << std::endl;
std::cout << "Resulting subset sizes' mean: " << tracker->size_mean() << ", standard deviation: " << tracker->size_stddev() << std::endl;
std::cout << std::endl;
std::cout << "stability measure CWrel("<<da->getNoOfFeatures()<<")=" << tracker->stability_CWrel(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure ATI()=" << tracker->stability_ATI() << std::endl;
std::cout << "stability measure CW()=" << tracker->stability_CW() << std::endl;
std::cout << "stability measure ANHI("<<da->getNoOfFeatures()<<")=" << tracker->stability_ANHI(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure SH()=" << tracker->stability_SH() << std::endl;
std::cout << "stability measure PH("<<da->getNoOfFeatures()<<")=" << tracker->stability_PH(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure C()=" << tracker->stability_C() << std::endl;
// compute statistics
trackerstats->compute_stats();
// (optionally) print computation statistics
trackerstats->print_stats(std::cout);
}
catch(FST::fst_error &e) {std::cerr<<"FST ERROR: "<< e.what() << ", code=" << e.code() << std::endl;}
catch(std::exception &e) {std::cerr<<"non-FST ERROR: "<< e.what() << std::endl;}
return 0;
}
Search algorithm: Sequential Forward Floating Selection (SFFS)/ Evaluation criterion: k-nearest-neighbor (KNN)/ d=5
486
#include <boost/smart_ptr.hpp>
#include <exception>
#include <iostream>
#include <cstdlib>
#include <string>
#include <vector>
#include "error.hpp"
#include "global.hpp"
#include "subset.hpp"
#include "data_intervaller.hpp"
#include "data_splitter.hpp"
//#include "data_splitter_5050.hpp"
#include "data_splitter_cv.hpp"
//#include "data_splitter_holdout.hpp"
//#include "data_splitter_leave1out.hpp"
//#include "data_splitter_resub.hpp"
#include "data_splitter_randrand.hpp"
//#include "data_splitter_randfix.hpp"
#include "data_scaler.hpp"
#include "data_scaler_void.hpp"
//#include "data_scaler_to01.hpp"
//#include "data_scaler_white.hpp"
#include "data_accessor_splitting_memTRN.hpp"
#include "data_accessor_splitting_memARFF.hpp"
//#include "criterion_normal_bhattacharyya.hpp"
//#include "criterion_normal_gmahalanobis.hpp"
//#include "criterion_normal_divergence.hpp"
//#include "criterion_multinom_bhattacharyya.hpp"
#include "criterion_wrapper.hpp"
#include "criterion_wrapper_bias_estimate.hpp"
//#include "criterion_subsetsize.hpp"
//#include "criterion_sumofweights.hpp"
//#include "criterion_negative.hpp"
487
#include "distance_euclid.hpp"
//#include "distance_L1.hpp"
//#include "distance_Lp.hpp"
#include "classifier_knn.hpp"
//#include "classifier_normal_bayes.hpp"
//#include "classifier_multinom_naivebayes.hpp"
//#include "classifier_svm.hpp"
//#include "search_bif.hpp"
//#include "search_bif_threaded.hpp"
//#include "search_monte_carlo.hpp"
//#include "search_monte_carlo_threaded.hpp"
//#include "search_exhaustive.hpp"
//#include "search_exhaustive_threaded.hpp"
//#include "branch_and_bound_predictor_averaging.hpp"
//#include "search_branch_and_bound_basic.hpp"
//#include "search_branch_and_bound_improved.hpp"
//#include "search_branch_and_bound_partial_prediction.hpp"
//#include "search_branch_and_bound_fast.hpp"
#include "seq_step_straight.hpp"
//#include "seq_step_straight_threaded.hpp"
//#include "seq_step_hybrid.hpp"
//#include "seq_step_ensemble.hpp"
//#include "search_seq_sfs.hpp"
#include "search_seq_sffs.hpp"
//#include "search_seq_sfrs.hpp"
//#include "search_seq_os.hpp"
//#include "search_seq_dos.hpp"
//#include "result_tracker_dupless.hpp"
//#include "result_tracker_regularizer.hpp"
#include "result_tracker_feature_stats.hpp"
#include "result_tracker_stabileval.hpp"
488
int main()
{
try{
typedef double RETURNTYPE;
typedef double DATATYPE; typedef double REALTYPE;
typedef unsigned int IDXTYPE; typedef unsigned int DIMTYPE; typedef short BINTYPE;
typedef FST::Subset<BINTYPE, DIMTYPE> SUBSET;
typedef FST::Data_Intervaller<std::vector<FST::Data_Interval<IDXTYPE> >,IDXTYPE> INTERVALLER;
typedef boost::shared_ptr<FST::Data_Splitter<INTERVALLER,IDXTYPE> > PSPLITTER;
typedef FST::Data_Splitter_CV<INTERVALLER,IDXTYPE> SPLITTERCV;
typedef FST::Data_Splitter_RandomRandom<INTERVALLER,IDXTYPE,BINTYPE> SPLITTERRANDRAND;
typedef FST::Data_Accessor_Splitting_MemTRN<DATATYPE,IDXTYPE,INTERVALLER> DATAACCESSOR;
typedef FST::Distance_Euclid<DATATYPE,DIMTYPE,SUBSET> DISTANCE;
typedef FST::Classifier_kNN<RETURNTYPE,DATATYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR,DISTANCE> CLASSIFIERKNN;
typedef FST::Criterion_Wrapper<RETURNTYPE,SUBSET,CLASSIFIERKNN,DATAACCESSOR> WRAPPERKNN;
typedef FST::Criterion_Wrapper_Bias_Estimate<RETURNTYPE,SUBSET,CLASSIFIERKNN,DATAACCESSOR> WRAPPERBIAS;
typedef FST::Sequential_Step_Straight<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERKNN> EVALUATOR;
typedef FST::Result_Tracker_Stability_Evaluator<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKER;
typedef FST::Result_Tracker_Feature_Stats<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKERSTATS;
489
std::cout << "Wrapper-based (3NN) feature selection with Sequential Forward Floating Search, 10-fold with Stability Evaluator d=5" << std::endl;
// print console output to log.txt
freopen("log_Wrapper01a_SFFS_with stabilityeval_d=5.txt","w",stdout);
std::cout << "Wrapper-based (3NN) feature selection with Sequential Forward Floating Search, 10-fold with Stability Evaluator d=5 \n" << std::endl;
// use randomly chosen 70% of data for training and keep the other 30% for independent testing of final classification performance
PSPLITTER dsp_outer(new SPLITTERRANDRAND(20/*splits=trials*/,70,30)); // (there will be one outer randomized split only)
// in the course of search use the first half of data by 10-fold cross-validation in wrapper FS criterion evaluation
PSPLITTER dsp_inner(new SPLITTERCV(10));
// do not scale data
boost::shared_ptr<FST::Data_Scaler<DATATYPE> > dsc(new FST::Data_Scaler_void<DATATYPE>());
// set-up data access
boost::shared_ptr<std::vector<PSPLITTER> > splitters(new std::vector<PSPLITTER>);
splitters->push_back(dsp_outer); splitters->push_back(dsp_inner);
boost::shared_ptr<DATAACCESSOR> da(new DATAACCESSOR("species_remoutl_normal_185_7_309.trn",splitters,dsc));
da->initialize();
// initiate access to split data parts
da->setSplittingDepth(0); if(!da->getFirstSplit()) throw FST::fst_error("70/30 data split failed.");
da->setSplittingDepth(1); if(!da->getFirstSplit()) throw FST::fst_error("10-fold cross-validation failure.");
// initiate the storage for subset to-be-selected
490
boost::shared_ptr<SUBSET> sub(new SUBSET(da->getNoOfFeatures())); sub->deselect_all();
// set-up 3-Nearest Neighbor classifier based on Euclidean distances
boost::shared_ptr<CLASSIFIERKNN> cknn(new CLASSIFIERKNN); cknn->set_k(3);
// wrap the 3-NN classifier to enable its usage as FS criterion (criterion value will be estimated by 10-fold cross-val.)
boost::shared_ptr<WRAPPERKNN> wknn(new WRAPPERKNN);
wknn->initialize(cknn,da);
// set-up the standard sequential search step object (option: hybrid, ensemble, etc.)
boost::shared_ptr<EVALUATOR> eval(new EVALUATOR);
// set-up Sequential Forward Floating Selection search procedure
FST::Search_SFFS<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERKNN,EVALUATOR> srch(eval);
srch.set_search_direction(FST::FORWARD); // try FST::BACKWARD
// set target subset size
DIMTYPE target_subsize=5;
// set-up tracker to gather data for computation
boost::shared_ptr<TRACKERSTATS> trackerstats(new TRACKERSTATS);
srch.enable_result_tracking(trackerstats);
// set-up result tracker to collect results of each trial
boost::shared_ptr<TRACKER> tracker(new TRACKER);
// run the trials
std::cout << "Feature selection setup:" << std::endl << *da << std::endl << srch << std::endl << *wknn << std::endl << *tracker << std::endl << std::endl;
RETURNTYPE critval_train, critval_test, biasval;
srch.set_output_detail(FST::SILENT); // set FST::SILENT to disable all text output in the course of search (FST::NORMAL is default)
da->setSplittingDepth(0);
unsigned int trial=0;
bool run=da->getFirstSplit(); if(!run) throw FST::fst_error("RandRand data split failed.");
while(run)
{
trial++; std::cout << std::endl<<"TRIAL "<<trial<< " ---------------------------------------------------------------------"<<std::endl;
da->setSplittingDepth(1);
if(!srch.search(target_subsize,critval_train,sub,wknn,std::cout)) throw FST::fst_error("Search not finished."); //d-parameterized (d=5)
tracker->add(critval_train,sub);
std::cout << std::endl << "(TRIAL "<<trial<<") Search result: " << std::endl << *sub << "Criterion value=" << critval_train << std::endl;
// estimate classifier bias on training data with respect to the selected feature subset
boost::shared_ptr<WRAPPERBIAS> wbias(new WRAPPERBIAS);
wbias->initialize(cknn,da);
491
//da->setSplittingDepth(0);
if(!wbias->evaluate(biasval,sub)) throw FST::fst_error("Bias estimation failed.");
std::cout << "Estimated bias=" << biasval << std::endl << std::endl;
// (optionally) validate result by estimating kNN accuracy on selected feature sub-space on independent test data
da->setSplittingDepth(0);
cknn->train(da,sub);
cknn->test(critval_test,da);
std::cout << "Validated "<<cknn->get_k()<<"-NN accuracy=" << critval_test << std::endl << std::endl;
// (optionally) list the best known solutions for each cardinality as recorded throughout the course of search
std::cout << "Best recorded solution for subset size:" << std::endl;
for(DIMTYPE d=1;d<=sub->get_n();d++)
if(srch.get_result(d,critval_train,sub)) std::cout << d << ": val="<< critval_train << ", "<<*sub << std::endl;
da->setSplittingDepth(0);
run=da->getNextSplit();
}
// evaluate stability using results collected by tracker
std::cout<<std::endl;
std::cout << "---------------------------------------------------------------------" << std::endl;
std::cout << "Resulting criterion values' mean: " << tracker->value_mean() << ", standard deviation: " << tracker->value_stddev() << std::endl;
std::cout << "Resulting subset sizes' mean: " << tracker->size_mean() << ", standard deviation: " << tracker->size_stddev() << std::endl;
std::cout << std::endl;
std::cout << "stability measure CWrel("<<da->getNoOfFeatures()<<")=" << tracker->stability_CWrel(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure ATI()=" << tracker->stability_ATI() << std::endl;
std::cout << "stability measure CW()=" << tracker->stability_CW() << std::endl;
std::cout << "stability measure ANHI("<<da->getNoOfFeatures()<<")=" << tracker->stability_ANHI(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure SH()=" << tracker->stability_SH() << std::endl;
std::cout << "stability measure PH("<<da->getNoOfFeatures()<<")=" << tracker->stability_PH(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure C()=" << tracker->stability_C() << std::endl;
// compute statistics
trackerstats->compute_stats();
// (optionally) print computation statistics
trackerstats->print_stats(std::cout);
}
catch(FST::fst_error &e) {std::cerr<<"FST ERROR: "<< e.what() << ", code=" << e.code() << std::endl;}
catch(std::exception &e) {std::cerr<<"non-FST ERROR: "<< e.what() << std::endl;}
return 0;
}
Search algorithm: Sequential Forward Selection (SFS)/ Evaluation criterion: Support Vector Machines (SVM)/ d=5
492
#include <boost/smart_ptr.hpp>
#include <exception>
#include <iostream>
#include <cstdlib>
#include <string>
#include <vector>
#include "error.hpp"
#include "global.hpp"
#include "subset.hpp"
#include "data_intervaller.hpp"
#include "data_splitter.hpp"
//#include "data_splitter_5050.hpp"
#include "data_splitter_cv.hpp"
//#include "data_splitter_holdout.hpp"
//#include "data_splitter_leave1out.hpp"
//#include "data_splitter_resub.hpp"
#include "data_splitter_randrand.hpp"
//#include "data_splitter_randfix.hpp"
#include "data_scaler.hpp"
#include "data_scaler_void.hpp"
//#include "data_scaler_to01.hpp"
//#include "data_scaler_white.hpp"
#include "data_accessor_splitting_memTRN.hpp"
#include "data_accessor_splitting_memARFF.hpp"
//#include "criterion_normal_bhattacharyya.hpp"
//#include "criterion_normal_gmahalanobis.hpp"
//#include "criterion_normal_divergence.hpp"
//#include "criterion_multinom_bhattacharyya.hpp"
#include "criterion_wrapper.hpp"
#include "criterion_wrapper_bias_estimate.hpp"
//#include "criterion_subsetsize.hpp"
//#include "criterion_sumofweights.hpp"
//#include "criterion_negative.hpp"
493
//#include "distance_euclid.hpp"
//#include "distance_L1.hpp"
//#include "distance_Lp.hpp"
//#include "classifier_knn.hpp"
//#include "classifier_normal_bayes.hpp"
//#include "classifier_multinom_naivebayes.hpp"
#include "classifier_svm.hpp"
//#include "search_bif.hpp"
//#include "search_bif_threaded.hpp"
//#include "search_monte_carlo.hpp"
//#include "search_monte_carlo_threaded.hpp"
//#include "search_exhaustive.hpp"
//#include "search_exhaustive_threaded.hpp"
//#include "branch_and_bound_predictor_averaging.hpp"
//#include "search_branch_and_bound_basic.hpp"
//#include "search_branch_and_bound_improved.hpp"
//#include "search_branch_and_bound_partial_prediction.hpp"
//#include "search_branch_and_bound_fast.hpp"
#include "seq_step_straight.hpp"
//include "seq_step_straight_threaded.hpp"
//#include "seq_step_hybrid.hpp"
//#include "seq_step_ensemble.hpp"
#include "search_seq_sfs.hpp"
//#include "search_seq_sffs.hpp"
//#include "search_seq_sfrs.hpp"
//#include "search_seq_os.hpp"
//#include "search_seq_dos.hpp"
//#include "result_tracker_dupless.hpp"
//#include "result_tracker_regularizer.hpp"
#include "result_tracker_feature_stats.hpp"
#include "result_tracker_stabileval.hpp"
494
int main()
{
try{
const unsigned int max_threads=2;
typedef double RETURNTYPE;
typedef double DATATYPE; typedef double REALTYPE;
typedef unsigned int IDXTYPE; typedef unsigned int DIMTYPE; typedef short BINTYPE;
typedef FST::Subset<BINTYPE, DIMTYPE> SUBSET;
typedef FST::Data_Intervaller<std::vector<FST::Data_Interval<IDXTYPE> >,IDXTYPE> INTERVALLER;
typedef boost::shared_ptr<FST::Data_Splitter<INTERVALLER,IDXTYPE> > PSPLITTER;
typedef FST::Data_Splitter_CV<INTERVALLER,IDXTYPE> SPLITTERCV;
typedef FST::Data_Splitter_RandomRandom<INTERVALLER,IDXTYPE,BINTYPE> SPLITTERRANDRAND;
typedef FST::Data_Accessor_Splitting_MemTRN<DATATYPE,IDXTYPE,INTERVALLER> DATAACCESSOR;
typedef FST::Classifier_LIBSVM<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR> CLASSIFIERSVM;
typedef FST::Criterion_Wrapper<RETURNTYPE,SUBSET,CLASSIFIERSVM,DATAACCESSOR> WRAPPERSVM;
typedef FST::Criterion_Wrapper_Bias_Estimate<RETURNTYPE,SUBSET,CLASSIFIERSVM,DATAACCESSOR> WRAPPERBIAS;
typedef FST::Sequential_Step_Straight<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERSVM> EVALUATOR;
typedef FST::Result_Tracker_Stability_Evaluator<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKER;
typedef FST::Result_Tracker_Feature_Stats<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKERSTATS;
495
std::cout << std::endl << "SVM-wrapper-based feature selection with Sequential Forward Selection, 10-fold and Stability Measurement d=)" << std::endl;
// print console output to log.txt
freopen("log_Wrapper04e_SFS_stabilityeval_d=5.txt","w",stdout);
std::cout << "SVM-wrapper-based feature selection with Sequential Forward Selection, 10-fold and Stability Measurement d=5 \n" << std::endl;
// use randomly chosen 70% of data for training and keep the other 30% for independent testing of final classification performance
PSPLITTER dsp_outer(new SPLITTERRANDRAND(20/*splits=trials*/,70,30)); // (there will be one outer randomized split only)
// in the course of search use the first half of data by 10-fold cross-validation in wrapper FS criterion evaluation
PSPLITTER dsp_inner(new SPLITTERCV(10));
// do not scale data
boost::shared_ptr<FST::Data_Scaler<DATATYPE> > dsc(new FST::Data_Scaler_void<DATATYPE>());
// set-up data access
boost::shared_ptr<std::vector<PSPLITTER> > splitters(new std::vector<PSPLITTER>);
splitters->push_back(dsp_outer); splitters->push_back(dsp_inner);
boost::shared_ptr<DATAACCESSOR> da(new DATAACCESSOR("species_remoutl_normal_185_7_309.trn",splitters,dsc));
da->initialize();
// initiate access to split data parts
da->setSplittingDepth(0); if(!da->getFirstSplit()) throw FST::fst_error("70/30 data split failed.");
da->setSplittingDepth(1); if(!da->getFirstSplit()) throw FST::fst_error("10-fold cross-validation failure.");
// initiate the storage for subset to-be-selected
boost::shared_ptr<SUBSET> sub(new SUBSET(da->getNoOfFeatures()));
boost::shared_ptr<SUBSET> sub_temp(new SUBSET(da->getNoOfFeatures()));
496
// set-up SVM (interface to external library LibSVM)
boost::shared_ptr<CLASSIFIERSVM> csvm(new CLASSIFIERSVM);
csvm->set_kernel_type(RBF); // (option: LINEAR, POLY, SIGMOID)
csvm->initialize(da);
// wrap the SVM classifier to enable its usage as FS criterion (criterion value will be estimated by 10-fold cross-val.)
boost::shared_ptr<WRAPPERSVM> wsvm(new WRAPPERSVM);
wsvm->initialize(csvm,da);
// set-up the sequential search step object (option: hybrid, ensemble, etc.)
boost::shared_ptr<EVALUATOR> eval(new EVALUATOR);
// set-up Sequential Forward Selection search procedure
FST::Search_SFS<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERSVM,EVALUATOR> srch(eval);
srch.set_search_direction(FST::FORWARD); // try FST::BACKWARD
// set target subset size
DIMTYPE target_subsize=5;
// set-up tracker to gather data for computation
boost::shared_ptr<TRACKERSTATS> trackerstats(new TRACKERSTATS);
srch.enable_result_tracking(trackerstats);
// set-up result tracker to collect results of each trial
boost::shared_ptr<TRACKER> tracker(new TRACKER);
// run the trials
std::cout << "Feature selection setup:" << std::endl << *da << std::endl << srch << std::endl << *wsvm << std::endl << *tracker << std::endl << std::endl;
RETURNTYPE bestcritval_train, critval_train, critval_test, biasval;
srch.set_output_detail(FST::SILENT); // set FST::SILENT to disable all text output in the course of search (FST::NORMAL is default)
da->setSplittingDepth(0);
unsigned int trial=0;
bool run=da->getFirstSplit(); if(!run) throw FST::fst_error("RandRand data split failed.");
while(run)
{
trial++; std::cout << std::endl<<"TRIAL "<<trial<< " ---------------------------------------------------------------------"<<std::endl;
//optimize SVM parameters using 10-fold cross-validation on training data on the full set of features
sub->select_all();
da->setSplittingDepth(1);
csvm->optimize_parameters(da,sub);
double best_svm_param_C=csvm->get_parameter_C();
double best_svm_param_gamma=csvm->get_parameter_gamma();
497
double best_svm_param_coef0=csvm->get_parameter_coef0();
bool stop=false;
sub->deselect_all();
if(!srch.search(target_subsize,bestcritval_train,sub,wsvm, std::cout)) throw FST::fst_error("Search not finished."); //set target_size d=5
sub_temp->stateless_copy(*sub);
while(!stop)
{
csvm->optimize_parameters(da,sub);
if(best_svm_param_C!=csvm->get_parameter_C() || best_svm_param_gamma!=csvm->get_parameter_gamma() ||
best_svm_param_coef0!=csvm->get_parameter_coef0())
{
if(!srch.search(target_subsize,critval_train,sub_temp,wsvm,std::cout)) throw FST::fst_error("Search not finished."); //set target_size
d=5
if(critval_train>bestcritval_train)
{
bestcritval_train=critval_train;
sub->stateless_copy(*sub_temp);
best_svm_param_C=csvm->get_parameter_C();
best_svm_param_gamma=csvm->get_parameter_gamma();
best_svm_param_coef0=csvm->get_parameter_coef0();
} else stop=true;
} else stop=true;
}
tracker->add(critval_train,sub);
std::cout << std::endl << "(TRIAL "<<trial<<") Search result: " << std::endl << *sub << "Criterion value=" << bestcritval_train << std::endl;
// estimate classifier bias on training data with respect to the selected feature subset
boost::shared_ptr<WRAPPERBIAS> wbias(new WRAPPERBIAS);
wbias->initialize(csvm,da);
//da->setSplittingDepth(0);
if(!wbias->evaluate(biasval,sub)) throw FST::fst_error("Bias estimation failed.");
std::cout << "Estimated bias=" << biasval << std::endl << std::endl;
// (optionally) validate result by estimating SVM accuracy on selected feature sub-space on independent test data
da->setSplittingDepth(0);
csvm->set_parameter_C(best_svm_param_C);
csvm->set_parameter_gamma(best_svm_param_gamma);
csvm->set_parameter_coef0(best_svm_param_coef0);
csvm->train(da,sub);
csvm->test(critval_test,da);
std::cout << "Validated SVM accuracy=" << critval_test << std::endl << std::endl;
// (optionally) list the best known solutions for each cardinality as recorded throughout the course of search
std::cout << "Best recorded solution for subset size:" << std::endl;
for(DIMTYPE d=1;d<=sub->get_n();d++)
if(srch.get_result(d,critval_train,sub)) std::cout << d << ": val="<< critval_train << ", "<<*sub << std::endl;
da->setSplittingDepth(0);
run=da->getNextSplit();
}
// evaluate stability using results collected by tracker
std::cout<<std::endl;
std::cout << "---------------------------------------------------------------------" << std::endl;
std::cout << "Resulting criterion values' mean: " << tracker->value_mean() << ", standard deviation: " << tracker->value_stddev() << std::endl;
std::cout << "Resulting subset sizes' mean: " << tracker->size_mean() << ", standard deviation: " << tracker->size_stddev() << std::endl;
std::cout << std::endl;
std::cout << "stability measure CWrel("<<da->getNoOfFeatures()<<")=" << tracker->stability_CWrel(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure ATI()=" << tracker->stability_ATI() << std::endl;
std::cout << "stability measure CW()=" << tracker->stability_CW() << std::endl;
std::cout << "stability measure ANHI("<<da->getNoOfFeatures()<<")=" << tracker->stability_ANHI(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure SH()=" << tracker->stability_SH() << std::endl;
std::cout << "stability measure PH("<<da->getNoOfFeatures()<<")=" << tracker->stability_PH(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure C()=" << tracker->stability_C() << std::endl;
// compute statistics
trackerstats->compute_stats();
// (optionally) print computation statistics
trackerstats->print_stats(std::cout);
}
catch(FST::fst_error &e) {std::cerr<<"FST ERROR: "<< e.what() << ", code=" << e.code() << std::endl;}
catch(std::exception &e) {std::cerr<<"non-FST ERROR: "<< e.what() << std::endl;}
return 0;
}
498
Search algorithm: Oscillating Search (OS)/ Evaluation criterion: Support Vector Machines (SVM)/ d=5
499
#include <boost/smart_ptr.hpp>
#include <exception>
#include <iostream>
#include <cstdlib>
#include <string>
#include <vector>
#include "error.hpp"
#include "global.hpp"
#include "subset.hpp"
#include "data_intervaller.hpp"
#include "data_splitter.hpp"
//#include "data_splitter_5050.hpp"
#include "data_splitter_cv.hpp"
//#include "data_splitter_holdout.hpp"
//#include "data_splitter_leave1out.hpp"
//#include "data_splitter_resub.hpp"
#include "data_splitter_randrand.hpp"
//#include "data_splitter_randfix.hpp"
#include "data_scaler.hpp"
#include "data_scaler_void.hpp"
//#include "data_scaler_to01.hpp"
//#include "data_scaler_white.hpp"
#include "data_accessor_splitting_memTRN.hpp"
#include "data_accessor_splitting_memARFF.hpp"
//#include "criterion_normal_bhattacharyya.hpp"
//#include "criterion_normal_gmahalanobis.hpp"
//#include "criterion_normal_divergence.hpp"
//#include "criterion_multinom_bhattacharyya.hpp"
#include "criterion_wrapper.hpp"
#include "criterion_wrapper_bias_estimate.hpp"
//#include "criterion_subsetsize.hpp"
//#include "criterion_sumofweights.hpp"
//#include "criterion_negative.hpp"
500
//#include "distance_euclid.hpp"
//#include "distance_L1.hpp"
//#include "distance_Lp.hpp"
//#include "classifier_knn.hpp"
//#include "classifier_normal_bayes.hpp"
//#include "classifier_multinom_naivebayes.hpp"
#include "classifier_svm.hpp"
//#include "search_bif.hpp"
//#include "search_bif_threaded.hpp"
//#include "search_monte_carlo.hpp"
//#include "search_monte_carlo_threaded.hpp"
//#include "search_exhaustive.hpp"
//#include "search_exhaustive_threaded.hpp"
//#include "branch_and_bound_predictor_averaging.hpp"
//#include "search_branch_and_bound_basic.hpp"
//#include "search_branch_and_bound_improved.hpp"
//#include "search_branch_and_bound_partial_prediction.hpp"
//#include "search_branch_and_bound_fast.hpp"
#include "seq_step_straight.hpp"
//include "seq_step_straight_threaded.hpp"
//#include "seq_step_hybrid.hpp"
//#include "seq_step_ensemble.hpp"
//#include "search_seq_sfs.hpp"
//#include "search_seq_sffs.hpp"
//#include "search_seq_sfrs.hpp"
#include "search_seq_os.hpp"
//#include "search_seq_dos.hpp"
//#include "result_tracker_dupless.hpp"
//#include "result_tracker_regularizer.hpp"
#include "result_tracker_feature_stats.hpp"
#include "result_tracker_stabileval.hpp"
501
int main()
{
try{
const unsigned int max_threads=2;
typedef double RETURNTYPE;
typedef double DATATYPE; typedef double REALTYPE;
typedef unsigned int IDXTYPE; typedef unsigned int DIMTYPE; typedef short BINTYPE;
typedef FST::Subset<BINTYPE, DIMTYPE> SUBSET;
typedef FST::Data_Intervaller<std::vector<FST::Data_Interval<IDXTYPE> >,IDXTYPE> INTERVALLER;
typedef boost::shared_ptr<FST::Data_Splitter<INTERVALLER,IDXTYPE> > PSPLITTER;
typedef FST::Data_Splitter_CV<INTERVALLER,IDXTYPE> SPLITTERCV;
typedef FST::Data_Splitter_RandomRandom<INTERVALLER,IDXTYPE,BINTYPE> SPLITTERRANDRAND;
typedef FST::Data_Accessor_Splitting_MemTRN<DATATYPE,IDXTYPE,INTERVALLER> DATAACCESSOR;
typedef FST::Classifier_LIBSVM<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR> CLASSIFIERSVM;
typedef FST::Criterion_Wrapper<RETURNTYPE,SUBSET,CLASSIFIERSVM,DATAACCESSOR> WRAPPERSVM;
typedef FST::Criterion_Wrapper_Bias_Estimate<RETURNTYPE,SUBSET,CLASSIFIERSVM,DATAACCESSOR> WRAPPERBIAS;
typedef FST::Sequential_Step_Straight<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERSVM> EVALUATOR;
typedef FST::Result_Tracker_Stability_Evaluator<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKER;
typedef FST::Result_Tracker_Feature_Stats<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKERSTATS;
502
std::cout << std::endl << "SVM-wrapper-based feature selection with Oscillating Search, 10-fold and Stability Measurement d=5" << std::endl;
// print console output to log.txt
freopen("log_Wrapper04b_OS_stabilityeval_d=5.txt","w",stdout);
std::cout << "SVM-wrapper-based feature selection with Oscillating Search, 10-fold and Stability Measurement d=5 \n" << std::endl;
// use randomly chosen 70% of data for training and keep the other 30% for independent testing of final classification performance
PSPLITTER dsp_outer(new SPLITTERRANDRAND(20/*splits=trials*/,70,30)); // (there will be one outer randomized split only)
// in the course of search use the first half of data by 10-fold cross-validation in wrapper FS criterion evaluation
PSPLITTER dsp_inner(new SPLITTERCV(10));
// do not scale data
boost::shared_ptr<FST::Data_Scaler<DATATYPE> > dsc(new FST::Data_Scaler_void<DATATYPE>());
// set-up data access
boost::shared_ptr<std::vector<PSPLITTER> > splitters(new std::vector<PSPLITTER>);
splitters->push_back(dsp_outer); splitters->push_back(dsp_inner);
boost::shared_ptr<DATAACCESSOR> da(new DATAACCESSOR("species_remoutl_normal_185_7_309.trn",splitters,dsc));
da->initialize();
// initiate access to split data parts
da->setSplittingDepth(0); if(!da->getFirstSplit()) throw FST::fst_error("70/30 data split failed.");
da->setSplittingDepth(1); if(!da->getFirstSplit()) throw FST::fst_error("10-fold cross-validation failure.");
// initiate the storage for subset to-be-selected
boost::shared_ptr<SUBSET> sub(new SUBSET(da->getNoOfFeatures()));
boost::shared_ptr<SUBSET> sub_temp(new SUBSET(da->getNoOfFeatures()));
boost::shared_ptr<SUBSET> submax(new SUBSET(da->getNoOfFeatures()));
503
// set-up SVM (interface to external library LibSVM)
boost::shared_ptr<CLASSIFIERSVM> csvm(new CLASSIFIERSVM);
csvm->set_kernel_type(RBF); // (option: LINEAR, POLY, SIGMOID)
csvm->initialize(da);
// wrap the SVM classifier to enable its usage as FS criterion (criterion value will be estimated by 10-fold cross-val.)
boost::shared_ptr<WRAPPERSVM> wsvm(new WRAPPERSVM);
wsvm->initialize(csvm,da);
// set-up the sequential search step object (option: hybrid, ensemble, etc.)
boost::shared_ptr<EVALUATOR> eval(new EVALUATOR);
// set-up oscillating search procedure
FST::Search_OS<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERSVM,EVALUATOR> srch(eval);
srch.set_delta(5);
// set target subset size
DIMTYPE target_subsize=5;
// set-up tracker to gather data for computation
boost::shared_ptr<TRACKERSTATS> trackerstats(new TRACKERSTATS);
srch.enable_result_tracking(trackerstats);
// set-up result tracker to collect results of each trial
boost::shared_ptr<TRACKER> tracker(new TRACKER);
// run the trials
std::cout << "Feature selection setup:" << std::endl << *da << std::endl << srch << std::endl << *wsvm << std::endl << *tracker << std::endl << std::endl;
RETURNTYPE bestcritval_train, critval_train, critval_trainmax, critval_test, biasval;
srch.set_output_detail(FST::NORMAL); // set FST::SILENT to disable all text output in the course of search (FST::NORMAL is default)
da->setSplittingDepth(0);
unsigned int trial=0;
bool run=da->getFirstSplit(); if(!run) throw FST::fst_error("RandRand data split failed.");
while(run)
{
trial++; std::cout << std::endl<<"TRIAL "<<trial<< " ---------------------------------------------------------------------"<<std::endl;
//optimize SVM parameters using 10-fold cross-validation on training data on the full set of features
sub->select_all();
da->setSplittingDepth(1);
csvm->optimize_parameters(da,sub);
double best_svm_param_C=csvm->get_parameter_C();
double best_svm_param_gamma=csvm->get_parameter_gamma();
504
double best_svm_param_coef0=csvm->get_parameter_coef0();
bool stop=false;
sub->deselect_all();
if(!srch.search(target_subsize,bestcritval_train,sub,wsvm, std::cout)) throw FST::fst_error("Search not finished."); //set target_size d=5
sub_temp->stateless_copy(*sub);
while(!stop)
{
csvm->optimize_parameters(da,sub);
if(best_svm_param_C!=csvm->get_parameter_C() || best_svm_param_gamma!=csvm->get_parameter_gamma() ||
best_svm_param_coef0!=csvm->get_parameter_coef0())
{
if(!srch.search(target_subsize,critval_train,sub_temp,wsvm,std::cout)) throw FST::fst_error("Search not finished."); //set target_size
d=5
if(critval_train>bestcritval_train)
{
bestcritval_train=critval_train;
sub->stateless_copy(*sub_temp);
best_svm_param_C=csvm->get_parameter_C();
best_svm_param_gamma=csvm->get_parameter_gamma();
best_svm_param_coef0=csvm->get_parameter_coef0();
} else stop=true;
} else stop=true;
}
int non_improving_runs=-1; // -1 indicates first run of the randomized procedure
const int max_non_improving_runs=5;
do {
sub->make_random_subset(target_subsize);
if(!srch.search(target_subsize,critval_train,sub,wsvm,std::cout)) throw FST::fst_error("Search not finished.");
if(non_improving_runs==-1 || critval_train>critval_trainmax) {
critval_trainmax=critval_train;
submax->stateless_copy(*sub);
non_improving_runs=0;
} else non_improving_runs++;
} while(non_improving_runs<max_non_improving_runs-1);
tracker->add(critval_train,sub);
505
std::cout << std::endl << "(TRIAL "<<trial<<") Search result: " << std::endl << *submax << "Criterion value=" << critval_trainmax << std::endl;
// estimate classifier bias on training data with respect to the selected feature subset
boost::shared_ptr<WRAPPERBIAS> wbias(new WRAPPERBIAS);
wbias->initialize(csvm,da);
//da->setSplittingDepth(0);
if(!wbias->evaluate(biasval,sub)) throw FST::fst_error("Bias estimation failed.");
std::cout << "Estimated bias=" << biasval << std::endl << std::endl;
// (optionally) validate result by estimating SVM accuracy on selected feature sub-space on independent test data
da->setSplittingDepth(0);
csvm->set_parameter_C(best_svm_param_C);
csvm->set_parameter_gamma(best_svm_param_gamma);
csvm->set_parameter_coef0(best_svm_param_coef0);
csvm->train(da,submax);
csvm->test(critval_test,da);
std::cout << "Validated SVM accuracy=" << critval_test << std::endl << std::endl;
da->setSplittingDepth(0);
run=da->getNextSplit();
}
// evaluate stability using results collected by tracker
std::cout<<std::endl;
std::cout << "---------------------------------------------------------------------" << std::endl;
std::cout << "Resulting criterion values' mean: " << tracker->value_mean() << ", standard deviation: " << tracker->value_stddev() << std::endl;
std::cout << "Resulting subset sizes' mean: " << tracker->size_mean() << ", standard deviation: " << tracker->size_stddev() << std::endl;
std::cout << std::endl;
std::cout << "stability measure CWrel("<<da->getNoOfFeatures()<<")=" << tracker->stability_CWrel(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure ATI()=" << tracker->stability_ATI() << std::endl;
std::cout << "stability measure CW()=" << tracker->stability_CW() << std::endl;
std::cout << "stability measure ANHI("<<da->getNoOfFeatures()<<")=" << tracker->stability_ANHI(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure SH()=" << tracker->stability_SH() << std::endl;
std::cout << "stability measure PH("<<da->getNoOfFeatures()<<")=" << tracker->stability_PH(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure C()=" << tracker->stability_C() << std::endl;
// compute statistics
trackerstats->compute_stats();
// (optionally) print computation statistics
trackerstats->print_stats(std::cout);
}
catch(FST::fst_error &e) {std::cerr<<"FST ERROR: "<< e.what() << ", code=" << e.code() << std::endl;}
catch(std::exception &e) {std::cerr<<"non-FST ERROR: "<< e.what() << std::endl;}
return 0;
}
Search algorithm: Sequential Forward Floating Selection (SFFS)/ Evaluation criterion: Support Vector Machines (SVM)/ d=5
506
#include <boost/smart_ptr.hpp>
#include <exception>
#include <iostream>
#include <cstdlib>
#include <string>
#include <vector>
#include "error.hpp"
#include "global.hpp"
#include "subset.hpp"
#include "data_intervaller.hpp"
#include "data_splitter.hpp"
//#include "data_splitter_5050.hpp"
#include "data_splitter_cv.hpp"
//#include "data_splitter_holdout.hpp"
//#include "data_splitter_leave1out.hpp"
//#include "data_splitter_resub.hpp"
#include "data_splitter_randrand.hpp"
//#include "data_splitter_randfix.hpp"
#include "data_scaler.hpp"
#include "data_scaler_void.hpp"
//#include "data_scaler_to01.hpp"
//#include "data_scaler_white.hpp"
#include "data_accessor_splitting_memTRN.hpp"
#include "data_accessor_splitting_memARFF.hpp"
//#include "criterion_normal_bhattacharyya.hpp"
//#include "criterion_normal_gmahalanobis.hpp"
//#include "criterion_normal_divergence.hpp"
//#include "criterion_multinom_bhattacharyya.hpp"
#include "criterion_wrapper.hpp"
#include "criterion_wrapper_bias_estimate.hpp"
//#include "criterion_subsetsize.hpp"
//#include "criterion_sumofweights.hpp"
//#include "criterion_negative.hpp"
507
//#include "distance_euclid.hpp"
//#include "distance_L1.hpp"
//#include "distance_Lp.hpp"
//#include "classifier_knn.hpp"
//#include "classifier_normal_bayes.hpp"
//#include "classifier_multinom_naivebayes.hpp"
#include "classifier_svm.hpp"
//#include "search_bif.hpp"
//#include "search_bif_threaded.hpp"
//#include "search_monte_carlo.hpp"
//#include "search_monte_carlo_threaded.hpp"
//#include "search_exhaustive.hpp"
//#include "search_exhaustive_threaded.hpp"
//#include "branch_and_bound_predictor_averaging.hpp"
//#include "search_branch_and_bound_basic.hpp"
//#include "search_branch_and_bound_improved.hpp"
//#include "search_branch_and_bound_partial_prediction.hpp"
//#include "search_branch_and_bound_fast.hpp"
#include "seq_step_straight.hpp"
//include "seq_step_straight_threaded.hpp"
//#include "seq_step_hybrid.hpp"
//#include "seq_step_ensemble.hpp"
//#include "search_seq_sfs.hpp"
#include "search_seq_sffs.hpp"
//#include "search_seq_sfrs.hpp"
//#include "search_seq_os.hpp"
//#include "search_seq_dos.hpp"
//#include "result_tracker_dupless.hpp"
//#include "result_tracker_regularizer.hpp"
#include "result_tracker_feature_stats.hpp"
#include "result_tracker_stabileval.hpp"
508
int main()
{
try{
const unsigned int max_threads=2;
typedef double RETURNTYPE;
typedef double DATATYPE; typedef double REALTYPE;
typedef unsigned int IDXTYPE; typedef unsigned int DIMTYPE; typedef short BINTYPE;
typedef FST::Subset<BINTYPE, DIMTYPE> SUBSET;
typedef FST::Data_Intervaller<std::vector<FST::Data_Interval<IDXTYPE> >,IDXTYPE> INTERVALLER;
typedef boost::shared_ptr<FST::Data_Splitter<INTERVALLER,IDXTYPE> > PSPLITTER;
typedef FST::Data_Splitter_CV<INTERVALLER,IDXTYPE> SPLITTERCV;
typedef FST::Data_Splitter_RandomRandom<INTERVALLER,IDXTYPE,BINTYPE> SPLITTERRANDRAND;
typedef FST::Data_Accessor_Splitting_MemTRN<DATATYPE,IDXTYPE,INTERVALLER> DATAACCESSOR;
typedef FST::Classifier_LIBSVM<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR> CLASSIFIERSVM;
typedef FST::Criterion_Wrapper<RETURNTYPE,SUBSET,CLASSIFIERSVM,DATAACCESSOR> WRAPPERSVM;
typedef FST::Criterion_Wrapper_Bias_Estimate<RETURNTYPE,SUBSET,CLASSIFIERSVM,DATAACCESSOR> WRAPPERBIAS;
typedef FST::Sequential_Step_Straight<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERSVM> EVALUATOR;
typedef FST::Result_Tracker_Stability_Evaluator<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKER;
typedef FST::Result_Tracker_Feature_Stats<RETURNTYPE,IDXTYPE,DIMTYPE,SUBSET> TRACKERSTATS;
509
std::cout << std::endl << "SVM-wrapper-based feature selection with Sequential Floating Forward Selection, 10-fold and Stability Measurement d=5" << std::endl;
// print console output to log.txt
freopen("log_Wrapper04a_SFFS_stabilityeval_d=5.txt","w",stdout);
std::cout << "SVM-wrapper-based feature selection with Sequential Floating Forward Selection, 10-fold and Stability Measurement d=5 \n" << std::endl;
// use randomly chosen 70% of data for training and keep the other 30% for independent testing of final classification performance
PSPLITTER dsp_outer(new SPLITTERRANDRAND(20/*splits=trials*/,70,30)); // (there will be one outer randomized split only)
// in the course of search use the first half of data by 10-fold cross-validation in wrapper FS criterion evaluation
PSPLITTER dsp_inner(new SPLITTERCV(10));
// do not scale data
boost::shared_ptr<FST::Data_Scaler<DATATYPE> > dsc(new FST::Data_Scaler_void<DATATYPE>());
// set-up data access
boost::shared_ptr<std::vector<PSPLITTER> > splitters(new std::vector<PSPLITTER>);
splitters->push_back(dsp_outer); splitters->push_back(dsp_inner);
boost::shared_ptr<DATAACCESSOR> da(new DATAACCESSOR("species_remoutl_normal_185_7_309.trn",splitters,dsc));
da->initialize();
// initiate access to split data parts
da->setSplittingDepth(0); if(!da->getFirstSplit()) throw FST::fst_error("70/30 data split failed.");
da->setSplittingDepth(1); if(!da->getFirstSplit()) throw FST::fst_error("10-fold cross-validation failure.");
// initiate the storage for subset to-be-selected
boost::shared_ptr<SUBSET> sub(new SUBSET(da->getNoOfFeatures()));
boost::shared_ptr<SUBSET> sub_temp(new SUBSET(da->getNoOfFeatures()));
510
// set-up SVM (interface to external library LibSVM)
boost::shared_ptr<CLASSIFIERSVM> csvm(new CLASSIFIERSVM);
csvm->set_kernel_type(RBF); // (option: LINEAR, POLY, SIGMOID)
csvm->initialize(da);
// wrap the SVM classifier to enable its usage as FS criterion (criterion value will be estimated by 10-fold cross-val.)
boost::shared_ptr<WRAPPERSVM> wsvm(new WRAPPERSVM);
wsvm->initialize(csvm,da);
// set-up the sequential search step object (option: hybrid, ensemble, etc.)
boost::shared_ptr<EVALUATOR> eval(new EVALUATOR);
// set-up Sequential Forward Floating Selection search procedure
FST::Search_SFFS<RETURNTYPE,DIMTYPE,SUBSET,WRAPPERSVM,EVALUATOR> srch(eval);
srch.set_search_direction(FST::FORWARD); // try FST::BACKWARD
// set target subset size
DIMTYPE target_subsize=5;
// set-up tracker to gather data for computation
boost::shared_ptr<TRACKERSTATS> trackerstats(new TRACKERSTATS);
srch.enable_result_tracking(trackerstats);
// set-up result tracker to collect results of each trial
boost::shared_ptr<TRACKER> tracker(new TRACKER);
// run the trials
std::cout << "Feature selection setup:" << std::endl << *da << std::endl << srch << std::endl << *wsvm << std::endl << *tracker << std::endl << std::endl;
RETURNTYPE bestcritval_train, critval_train, critval_test, biasval;
srch.set_output_detail(FST::SILENT); // set FST::SILENT to disable all text output in the course of search (FST::NORMAL is default)
da->setSplittingDepth(0);
unsigned int trial=0;
bool run=da->getFirstSplit(); if(!run) throw FST::fst_error("RandRand data split failed.");
while(run)
{
trial++; std::cout << std::endl<<"TRIAL "<<trial<< " ---------------------------------------------------------------------"<<std::endl;
//optimize SVM parameters using 10-fold cross-validation on training data on the full set of features
sub->select_all();
da->setSplittingDepth(1);
csvm->optimize_parameters(da,sub);
double best_svm_param_C=csvm->get_parameter_C();
double best_svm_param_gamma=csvm->get_parameter_gamma();
511
double best_svm_param_coef0=csvm->get_parameter_coef0();
bool stop=false;
sub->deselect_all();
if(!srch.search(target_subsize,bestcritval_train,sub,wsvm, std::cout)) throw FST::fst_error("Search not finished."); //set target_size d=5
sub_temp->stateless_copy(*sub);
while(!stop)
{
csvm->optimize_parameters(da,sub);
if(best_svm_param_C!=csvm->get_parameter_C() || best_svm_param_gamma!=csvm->get_parameter_gamma() ||
best_svm_param_coef0!=csvm->get_parameter_coef0())
{
if(!srch.search(target_subsize,critval_train,sub_temp,wsvm,std::cout)) throw FST::fst_error("Search not finished."); //set target_size
d=5
if(critval_train>bestcritval_train)
{
bestcritval_train=critval_train;
sub->stateless_copy(*sub_temp);
best_svm_param_C=csvm->get_parameter_C();
best_svm_param_gamma=csvm->get_parameter_gamma();
best_svm_param_coef0=csvm->get_parameter_coef0();
} else stop=true;
} else stop=true;
}
tracker->add(critval_train,sub);
std::cout << std::endl << "(TRIAL "<<trial<<") Search result: " << std::endl << *sub << "Criterion value=" << bestcritval_train << std::endl;
// estimate classifier bias on training data with respect to the selected feature subset
boost::shared_ptr<WRAPPERBIAS> wbias(new WRAPPERBIAS);
wbias->initialize(csvm,da);
//da->setSplittingDepth(0);
if(!wbias->evaluate(biasval,sub)) throw FST::fst_error("Bias estimation failed.");
std::cout << "Estimated bias=" << biasval << std::endl << std::endl;
// (optionally) validate result by estimating SVM accuracy on selected feature sub-space on independent test data
da->setSplittingDepth(0);
csvm->set_parameter_C(best_svm_param_C);
csvm->set_parameter_gamma(best_svm_param_gamma);
csvm->set_parameter_coef0(best_svm_param_coef0);
csvm->train(da,sub);
csvm->test(critval_test,da);
std::cout << "Validated SVM accuracy=" << critval_test << std::endl << std::endl;
// (optionally) list the best known solutions for each cardinality as recorded throughout the course of search
std::cout << "Best recorded solution for subset size:" << std::endl;
for(DIMTYPE d=1;d<=sub->get_n();d++)
if(srch.get_result(d,critval_train,sub)) std::cout << d << ": val="<< critval_train << ", "<<*sub << std::endl;
da->setSplittingDepth(0);
run=da->getNextSplit();
}
// evaluate stability using results collected by tracker
std::cout<<std::endl;
std::cout << "---------------------------------------------------------------------" << std::endl;
std::cout << "Resulting criterion values' mean: " << tracker->value_mean() << ", standard deviation: " << tracker->value_stddev() << std::endl;
std::cout << "Resulting subset sizes' mean: " << tracker->size_mean() << ", standard deviation: " << tracker->size_stddev() << std::endl;
std::cout << std::endl;
std::cout << "stability measure CWrel("<<da->getNoOfFeatures()<<")=" << tracker->stability_CWrel(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure ATI()=" << tracker->stability_ATI() << std::endl;
std::cout << "stability measure CW()=" << tracker->stability_CW() << std::endl;
std::cout << "stability measure ANHI("<<da->getNoOfFeatures()<<")=" << tracker->stability_ANHI(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure SH()=" << tracker->stability_SH() << std::endl;
std::cout << "stability measure PH("<<da->getNoOfFeatures()<<")=" << tracker->stability_PH(da->getNoOfFeatures()) << std::endl;
std::cout << "stability measure C()=" << tracker->stability_C() << std::endl;
// compute statistics
trackerstats->compute_stats();
// (optionally) print computation statistics
trackerstats->print_stats(std::cout);
}
catch(FST::fst_error &e) {std::cerr<<"FST ERROR: "<< e.what() << ", code=" << e.code() << std::endl;}
catch(std::exception &e) {std::cerr<<"non-FST ERROR: "<< e.what() << std::endl;}
return 0;
}
512
Appendix C
Predicted LAI(Bon) and LAI (2000) from
Simple Linear Regression Models
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI 1
513
MCARI 2
Figure C1. Scatterplot of measured and predicted LAI(Bon)
Figure C2. Predicted LAI(Bon) from regression model using NDVI as predictor
514
Figure C3. Predicted LAI(Bon) from regression model using RDVI as predictor
Figure C4. Predicted LAI(Bon) from regression model using SAVI as predictor
515
Figure C5. Predicted LAI(Bon) from regression model using MSAVI as predictor
Figure C6. Predicted LAI(Bon) from regression model using TVI as predictor
516
Figure C7. Predicted LAI(Bon) from regression model using MCARI 1 as predictor
Figure C8. Predicted LAI(Bon) from regression model using MCARI 2 as predictor
517
Table C1. Descriptive statistics for LAI(Bon) predicted over the study area
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI
1
MCARI
2
Maximum
2.88
4.42
4.36
4.63
4.35
4.40
5.61
Minimum
0.00
0.00
0.00
0.00
0.00
0.00
0.00
Average
1.98
2.21
2.21
2.18
2.11
2.12
2.14
Standard
deviation
0.54
0.81
0.80
0.86
0.79
0.80
0.88
518
NDVI
RDVI
SAVI
MSAVI
TVI
MCARI 1
MCARI 2
Figure C9. Scatterplot of measured and
predicted LAI (LAI2000)
519
Figure C10. Predicted LAI(2000) from regression model using NDVI as predictor
Figure C11. Predicted LAI(2000) from regression model using RDVI as predictor
520
Figure C12. Predicted LAI(2000) from regression model using SAVI as predictor
Figure C13. Predicted LAI(2000) from regression model using MSAVI as predictor
521
Figure C14. Predicted LAI(2000) from regression model using TVI as predictor
Figure C15. Predicted LAI(2000) from regression model using MCARI 1 as predictor
522
Figure C16. Predicted LAI(2000) from regression model using MCARI 2 as predictor
Table C2. Descriptive statistics for LAI(2000) predicted over the study area
MCARI MCARI
1
2
NDVI
RDVI
SAVI
MSAVI
TVI
Maximum
2.54
4.29
4.24
4.49
4.27
4.32
5.43
Minimum
0.16
0.00
0.00
0.00
0.00
0.00
0.00
Average
1.97
2.14
2.13
2.11
2.04
2.05
2.07
Standard
deviation
0.37
0.78
0.78
0.83
0.78
0.78
0.85
523
Appendix D
Predicted LAI(Bon) and LAI(2000) from
Multiple Stepwise Regression Models
RDVI+ASM
SAVI+ASM
TVI+ASM
MCARI1+ASM
MCARI2+ASM
Figure D1. Scatterplot of measured and predicted LAI(Bon) using different VIs
combined with ASM
524
Figure D2. Predicted LAI(Bon) from regression model using SAVI and ASM as
predictors
Figure D3. Predicted LAI(Bon) from regression model using RDVI and ASM as
predictors
525
Figure D4. Predicted LAI(Bon) from regression model using TVI and ASM as
predictors
Figure D5. Predicted LAI(Bon) from regression model using MCARI 1 and ASM as
predictors
526
Figure D6. Predicted LAI(Bon) from regression model using MCARI 2 and ASM as
predictors
Table D1. Descriptive statistics for LAI(Bon) predicted over the study area
RDVI+ASM
SAVI+ASM
TVI+ASM
MCARI1
+ASM
MCARI2
+ASM
Maximum
6.47
6.48
7.58
7.86
7.21
Minimum
9.51E-04
6.52E-03
5.62E-04
9.63E-04
7.25E-04
Mean
2.00
2.01
2.18
2.20
2.11
Standard
deviation
1.25
1.24
1.19
1.20
1.25
527
RDVI+ASM
SAVI+ASM
TVI+ASM
MCARI1+ASM
MCARI2+ASM
Figure D7. Scatterplot of measured and predicted LAI(2000) using different VIs
combined with ASM
528
Figure D8. Predicted LAI(2000) from regression model using RDVI and ASM as
predictors
Figure D9. Predicted LAI(2000) from regression model using SAVI and ASM as
predictors
529
Figure D10. Predicted LAI(2000) from regression model using TVI and ASM as
predictors
Figure D11. Predicted LAI(2000) from regression model using MCARI 1 and ASM as
predictors
530
Figure D12. Predicted LAI(2000) from regression model using MCARI 2 and ASM as
predictors
Table D2. Descriptive statistics for LAI(2000) predicted over the study area
RDVI+ASM
SAVI+ASM
TVI+ASM
MCARI1
+ASM
MCARI2
+ASM
Maximum
5.04
5.03
6.16
6.44
5.77
Minimum
4.83E-03
2.39E-03
1.76E-05
4.98E-04
6.33E-05
Mean
1.86
1.86
2.03
2.05
1.96
Standard
deviation
1.16
1.15
1.10
1.11
1.14
531
Документ
Категория
Без категории
Просмотров
0
Размер файла
32 461 Кб
Теги
sdewsdweddes
1/--страниц
Пожаловаться на содержимое документа