close

Вход

Забыли?

вход по аккаунту

?

Retrieval of geophysical parameters over oceans from spaceborne microwave radiometers using artificial neural network

код для вставкиСкачать
RETRIEVAL OF GEOPHYSICAL PARAMETERS OVER OCEANS
FROM SPACEBORNE MICROWAVE RADIOMETERS USING
ARTIFICIAL NEURAL NETWORK
Thesis submitted to the
Department of Physics, Gujarat University
In partial fulfillment of the requirement for the Degree of
DOCTOR OF PHILOSOPHY
in
Physics
by
Bintu G. Vasudevan
Oceanic Sciences Division
Meteorology & Oceanography Group
Space Applications Centre (ISRO)
AHMEDABAD - 380015
June 2006
ProQuest Number: 3736632
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.
ProQuest 3736632
Published by ProQuest LLC (2015). Copyright of the Dissertation is held by the Author.
All rights reserved.
This work is protected against unauthorized copying under Title 17, United States Code
Microform Edition © ProQuest LLC.
ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106 - 1346
Dedicated to My Parents
DECLARATION
I hereby declare that the thesis entitled "Retrieval of Geophysical
Parameters over Oceans from Spaceborne Microwave Radiometers
using Artificial Neural Network” is a genuine record of the research
work carried out by me and no part of this thesis has been submitted to
any University or institution for the award of any degree or diploma.
Date: June 2006
Signature of Author
(Bintu G. Vasudevan)
Oceanic Sciences Division,
Meteorology and Oceanography Group,
Space Applications Centre (ISRO),
Ahmedabad-380015
1
CERTIFICATE
This is to certify that the thesis entitled “Retrieval of
Geophysical
Microwave
Parameters
Radiometers
over
using
Oceans
Artificial
from
Neural
Spaceborne
Network’'
is a bonafide record of research work done by Bintu G. Vasudevan in
the Oceanic Sciences Division, Meteorology and Oceanography Group,
Space Applications Centre (ISRO). He carried out the study reported in
this thesis, independently under my supervision. I also certify that the
subject matter of the thesis has not formed the basis for the award of
any Degree or Diploma of any University or Institution.
Dated: June 2006
Research Supervisor
r.
11
A cknow ledgem ents
First of all I would like to express my sincere gratitude to Dr Vijay K. Agarwal for
providing his guidance, encouragement and motivation m carrying out this research
work I specially thank Dr B S Gohil for his guidance through the course of research
work. He helped me m interpreting results from various experiments performed during
the study I am extremely grateful to him for all the help and support
I thank Dr Abhyit Sarkar for his keen interest m my work and encouragement
from time to time I express my sincere thanks to all scientists of Meteorology and
Oceanography Group for useful diseursion and suggestions durmg my work I would
like to thank Dr M S Narayanan, former Group Director, MOG for his constant
encouragement
During the course of my work, I received tremendous support from MOG computer
facility for which I sincerely thank Shri H I Andharia, Smt S B. Kartikeyan, and
Smt Yagna K. Mankad for their cooperation Support from CMC staff is thankfully
acknowledged
I would like to thank, Director, Space Applications Centre (ISEO) and Deputy Di­
rector, Remote Sensing Applications and Image Processing Area, for kindly allowing
me to register with Gujarat University
I have a deep sense of appreciation for Neeraj and his family who have been so
nice and considerate to for me all these years
in
I am indebted to Indram and Sushil for being there with me all the time, providing
moral support It was fun havmg friends hke Randhir, Jayshankar, Yaswant, Babu,
Falgum, Thomas Kutty and Sriram with whom I had several memorable moments.
My parents have always been a source of inspiration to me. This work is a result
of their blessings, love, encouragement and good wishes with which they brought me
up
Finally I would hke to thank God, the almighty, for giving me strength, tolerance,
patience and above all a very good luck that helped me to complete this work
I should also mention that my research in Gujarat Umversity India was supported
m part by Space Applications Center, Ahmedabad, and by the Department of Ocean
Development (DOD) India, under SATCORE-I project
Bmtu G. vasudevan
June 12, 2006
IV
Table of Contents
Acknowledgements
iii
Table of Contents
v
List of Tables
viii
List of Figures
xi
Abbreviations
xvii
Acknowledgements: Data and Software support
xviii
Summary
xix
1 Introduction
11
12
13
14
t
Introduction...........................
Historical Perspective .
.
.
Neural Network based Analysis
Problem Statement .
.
1
1
.
.
. .
.
.
2 Basic Passive Microwave Radiometry and Retrieval
21
2.2
.
. . . .
3
7
10
11
Introduction . . .
.
. .
. .
. .
. . 11
Thermal Radiation and Radiative Transfer
. .
.
. 1 1
2 2 1 Brightness Temperature
.
.
................
.
14
2 2 2 Radiometric Brightness Temperature of a Scene
.
15
2 3 Passive Microwave Remote Sensing of the Ocean and Atmosphere . . 17
2 31 Sea Surface Emissivity .
. . . .
.
.
17
2 3 2 Overview of the Sea Surface Emissivity Modelling ...............
24
2 4 Microwave Radiometer Instruments and Satellite Missions , .
. . 28
2 41 Microwave Radiometer
.
.
28
2 4 2 Description of Satellite Microwave Radiometers
28
2 5 Retrieval Methods
.
. . . . .
.
32
v
2 5 1 Statistical methods
. . . .
2 5.2 Physical / Physio-statistical methods
2 5 3 Neural Network method
. .
,
.
.
.
. .
33
...................34
. .
.3 6
3 Development of Artificial Neural Network for Geophysical Parame­
ter Retrieval
38
31 Introduction
. . .
.
.
,
.
. .
3 2 Artificial Neural Network .
.
.
..................................... 38
3 2 1 Apphcation of Neural Network
. .
. . .
.
3 2 2 Advantages and Limitations of Neural Network . . .
3.2 3 Building Blocks of Neural Network
...............................
3 2 4 Learning Law .
.
.................................................
3 3 The Back-Propagation Neural Network
. .
. . . .
. 4
3 3 1 Learning Algorithms for Single P e rce p tro n .........................
3 3.2 Function Minimization Evaluation.
.
. . .
3 4 Development of NN Algorithm for Geophysical Parameter Retrieval .
38
41
42
43
46
8
49
52
53
4 Neural Network Approach and Retrieval of Geophysical Parameters
from Satellite Data
59
41 Introduction . .
. .
.
. .
. .
.
.
4 2 Simulation of Brightness Temperature over O c e a n s ......................
4 21 Simulation of Geophysical Parameters
. .
. .
4 2 2 Simulation of Brightness Temperature
............................64
4 3 Preparation of Training database
...............................
.
4 4 Analysis of Different Neural Network M o d e l s ............................
4.4 1 Selection of Optimum NN Configuration
............................68
4 4 2 Impact of Noise on NN Models
...................
...............
4.5 Retrieval of Geophysical Parameters from MSMR . .
.
. .
4 6 Analysis of Across Track Distribution of Brightness Temperature . . .
4.6.1 In-Orbit Data Quality Evaluation . . . .
.
.
4 7 Retrieval Algorithm and Analysis of Derived Parameters TMI case .
5 Comparison and Validation of Derived Geophysical Parameters
59
61
61
66
68
71
79
85
93
96
100
51 Introduction
.
. . . .
.
.
. .
100
5.2 Description of Data set
.
.
............................101
5 2 1 Validation Data
............
. .
.
101
5 2 2 Inter-compaxison Data
.
.
. . . 101
5 3 Validation and Comparison of MSMR Derived Parameters
.
102
5 3 1 Water Vapour Content
. .
.
.
.
. 102
5 3 2 Cloud Liquid Water Content . .
......................
106
vi
533
534
54
Sea Surface Temperature and Wind Speed
...
.
Comparison of MSMR NN derived parameters with Finished
Products . . .
.
...................
TMI Validation . . . .
.
.
.
. .
5 4 1 Sea Surface Temperature and Wind Speed
. .
. .
Ill
125
134
134
6 Conclusions and Future Scope
142
Bibliography
144
V ll
List orn Tables
t
•
j
m
1 1
2 1 Sensitivity of SMME brightness temperature Tg, to geophysical para­
meter Prabhakara et al. (1982)
. .
27
22
Major characteristic of the different microwave satellite sensors
3.1
Neural network model classification
3 2 Different NN Models
...
. .
.
. . .
used30
.
.
.
4.1
Regional statistics of simulated geophysical parameters
42
Global statistics of simulated geophysical parameters .
46
.
55
.
63
63
4 3 Range of SST (K) used m MSMR operational algorithms
.
.
65
4 4 Theoretical RMS retrieval accuracies of geophysical parameters by
MSMR operational SR algorithm
. .
.
. .
. .
65
4 5 Statistics of simulated geophysical parameters and brightness temper­
atures used in MSMR NN training (M_S-I dataset)
. .
.
.6 7
4 6 Initial noise figures m MSMR channel based on limited data and sim­
ulation
.
.
...
. .
. .
67
47
Statistics of simulated training dataset used m TMI NN model
48
Impact of Noise on class-I, 3HL-15N configuration for WVC
. .
.
68
71
4 9 Impact of noise on (8Tb —AGP) NN Model, performed on test dataset
(M_S-II)
.
vm
.
.
.
72
4 10 Theoretical RMS error of different NN Models
.
, .
. ,
75
411 MSMR bias RMS deviation with respect to RT model for all the channels 95
4 12 TMI bias and RMS deviation with respect to RT model for low reso­
lution channels
. . .
. .
...
. .
96
4.13 Theoretical RMS error for TMI (77 b —4G.P) retrieval on testing data
points
.
. .
. .
. .
. .
. .
. .
99
5 1 Statistics of validation of MSMR derived water vapour with Vaisala
102
5 2 Inter-comparisons of water vapour of MSMR derived and MSMR finish
products with TMI and SSM/I
.
...
. . 104
5 3 Inter-comparisons of cloud liquid water of MSMR derived and MSMR
finish products with TMI and SSM/I .
.
.
.
109
5.4 Validation of MSMR NN and SR derived SST and SSW with m-siiu
(ICOADS Ships, « 10900, collocated data p o in ts ) ...............................112
5 5 Vahdation of MSMR NN/SR derived SST and SSW, with m-situ (ICOADS
Buoys, number of collocated points « 2569).
. . . .
.117
5 6 Vahdation of MSMR NN/SR derived SST and SSW, with m-situ (NIOT
buoys, number of collocated points as 250) . .
...
. .
121
5 7 Inter-comparisons of SSW and SST for MSMR NN/SR derived with
TMI and SSM/I finished products . .
. .
...................
123
5 8 Comparisons of MSMR derived geophysical parameters with MSMR
finish products
59
.
. .
.
. . . .
132
Comparison of TMI (7TB-4GP) NN derived SSW and SST with ICOADS
ships data with m ±1 hour with m 0.25° x 0.25° grid .
.
. .
135
5 10 Comparison of TMI (7TB - 4GP) NN derived SSW and SST with
ICOADS buoys data with m 1 hour m 0 25° x 0.25° grid . . . .
rx
137
5 11 Comparisons of TMI NN derived geophysical parameters with TMI
finish products for 2 cays 15-16, Jan 2000, 0.25° x 0 25° (latitudelongitude)average,
.
.
.
x
.
.
.
139
List of Figures
2 1 The major components of the radiation received by a radiometer . .
12
2.2
Planck’s black body radiation at different temperatures
13
2.3
A typical variation of Ocean surface emissivity with vary mg angle
18
2.4
A typical variations of emissivity at 6 6 GHz (a) SST and (b) SSW
19
2.5
Effect of wind speed (surface roughness) on ocean brightness temper­
ature, SSW W3 > W2 > W1
2.6
................................................................ 20
..................................
2 7 MSMR Viewing Geometry
Neurons
. . . .
.
.
3 3 Single neuron
.
. .
.
.
.
.....................................
3 2 A feed-forward neural network
34
.
Transmission function of the cloud free atmosphere in the microwave
spectral range . .
31
.
. .
. .
Single perceptron learning
.
.
.
.
31
. .
. . .
.
.
.
. .
.2 3
.
40
. .
. .
.
43
.
. .
.
44
. 4 9
3 5 NN architecture for the standard backprogation algorithm of a threelayer perceptron (Cichocki et al 1993)
3 6 Sensitivity of
Tb
.
. .
.
,
52
to geophysical parameters, the arrow indicates the
SMMR frequencies, Wilheit (1979a)
.
3 7 Class-I neural network model configurations .
.
.
. .
.
.
.
54
...
56
3 8 Class-II neural network model configurations..................................
xi
57
4 1 Block diagram of retrieval algorithm development, testing and data
quality evolution
...
. .
...
60
4 2 Frequency distribution of simulated G P’s m training dataset (M_S-I,
29500 points), (a) WVC g/cm2, (b) CLW g/cm2, (c) SSW m /s, (d)
SST K
..............................................
. .
. .
62
4 3 Evolution of RMS error minimization for the training dataset for 4Tg —
1GP NN, configuration for WVC
.
.
.
.
.
4.4 Impact of noise in (8Tb - 4GP) NN Model, RMS Vs Epochs
70
. .
4 5 Impact of noise m (8TB - 4GP) NN Model, RMS Vs Global Iterations
73
73
4 6 Error distribution of class-II NN model (8Tjg —4GP) on test dataset
(MJS-II) for. (a) WVC (g/cm2), (b) CLW (g/cm2), (c) SSW (m/s),
(d) SST(K)
.
.
.
.
.
.
76
4.7 Scatter plot of simulated and retrieved parameter using (8Tb — 4GP)
NN model, trained with M-S-I (29500 data points) and apphed on test
dataset (M.S-II, 132710 data p o i n t s ) ........................................
77
4 8 Global distribution of MSMR retrieved WVC (g/cm2) using various
NN models, MSMR operational products and Wentz products for SSM/I
and T M I .....................
4.9
.
.
. . . .
80
Global distribution of MSMR retrieved CLW (g/cm2) using various NN
models, MSMR operational products and Wentz products for SSM/I
and TMI
. .
.
. .
. .
.
.
81
4 10 Global distribution of MSMR retrieved SSW (m/s) using various NN
models, MSMR operational products and Wentz products for SSM/I
and T M I ........................................
xii
. . .
.
82
411 Global distribution of MSMR retrieved SST (K) using various NN mod­
els, MSMR operational products and Wentz products for SSM/I and
TMI .
.
.
.
.
. 8
3
4 12 Monthly average of MSMR TB at 2° x 2° grid for Jul-1999 (a) TB06V
(b) T b06H (c) Tb 10V (d) TB10H (e) TB1SV (f) TS18H (g) TB21V (h)
r B21H
.
.
.
.
.
. 8
4
4.13 Across swath MSMR TB, six month (Jun - Oct, 1999) data histogram
of all channels
.
.
.
.
. .
.
.
.8 7
4 14 Across swath TMI TB, six month (Jun - Oct, 1999) data histogram of
low resolution channels
.
.
.
.
.
88
4 15 Across swath SSM/I TB, six month (Jun - Oct, 1999) data histogram
of low resolution channels .
.
.
89
4 16 Instant across track SSW, (a) before and (b) after across track bias
correction
.
.
.
.
.
.
91
4 17 SSW retrieval using (82jj —AGP) NN model, (a) before and (b) after
across track bias correction and (c) the differences . . .
4 18 MSMR m-orbit data quality evaluation
. .
. .
. .
92
. .
93
4 19 Global distribution TMI (7TB-A G P ) NN derived parameters for 15-16
Jul, 1999 ...................................................................
. .
4 20 Global distribution Wentz finished product for 15-16 Jul, 1999
5.1
.9 7
. .
98
Comparison of MSMR derived WVC with Vaisala (after bias removal)
(a)NN and (b)SR
.
xni
.
.
103
5 2 Inter-comparison of 2° x 2° averaged MSME WVC with TMI and
SSM/I finished products, (a) TMI versus MSMR NN, (b) SSM/I ver­
sus MSMR NN, (c) TMI versus MSMR SR, (d) SSM/I versus MSMR
SR
5.3
..................................
.
.
.
105
Comparison of histogram of CLW from MSMR (a) MSMR NN and (b)
MSMR finished product
. .
.
...................
. .
. 107
5 4 Inter-comparison of 2° x 2° averages of MSMR cloud liquid water with
TMI and SSM/I finished products, (a) TMI versus MSMR NN (b)
SSM/I versus MSMR NN (c) TMI versus MSMR SR (d) SSM/I versus
MSMR SR
.
.
.
.
......................108
5 5 Positions of collocated MSMR and ICOADS datasets (a) Ships (10,911
points) (b) buoys (2569 pomts)
.
. .
......................
Ill
5 6 Scatter plots of MSMR derived parameters with ICOADS ships (a) NN
SSW (b) SR SSW (c) NN SST (d) SR S S T ........................................... 113
5.7 The residual A SST, MSMR (NN/SR) SST - ship SST are plotted
against (a/b) MSMR NN/SR CLW, (c/d) MSMR NN/SR WVC, (e/f)
MSMR NN/SR SSW The solid lines (blue) on each figure indicate
the average residual, while the dashed fines (red) are ± one standard
deviation from the mean
................................................................... 114
5 8 Scatter plots of MSMR NN/SR derived parameters with ICOADS
buoys, (a) NN SSW (b) SR SSW (c) NN SST (d) SR S S T ............... 116
5 9 RMS error and bias, binned at every 1 m / s and 1 K for SSW and SST
with refrences to ICOADS buoys respectively, for MSMR NN derived
and finished products (a) SSW (b)SST
xiv
. .
.
.
...
118
5 10 Time series of SSW with different NIOT buoys, MSMR NN derived
and finished products
...
....
. .
119
5 11 (a) Time series of SST with different NIOT buoys, MSMR NN derived
and finished products, (b) Location of buoys
...
. .
120
5 12 Inter-comparison of MSMR NN/SR derived SST (K) and SSW (m/s)
with TMI finished products (a) TMI versus MSMR NN, SSW (b) TMI
versus MSMR NN, SST (c) TMI versus MSMR SR, SSW (d) TMI
versus MSMR SR, SST .
.
.
. . .
.
. . .
124
5.13 Inter-comparison of MSMR NN/SR derived SSW (m/s) with SSM/I
(Wentz) products (a) SSM/I versus MSMR NN, SSW (b) SSM/I versus
MSMR SR, S S W ..............................
...............................
124
5 14 Global distribution of MSMR derived WVC (g/crn2) for Oct 9-10 1999,
2 day, averaged 2° x 2° box, (a) NN (b) SR and (c) NN-SR . .
.
126
5 15 Global distribution of MSMR derived CLW (g/ctn2) for Oct 9-10 1999,
2 day, averaged 2° x 2° box, (a) NN (b) SR and (c) NN-SR
.
127
5.16 Global distribution of MSMR derived SSW (m /s) for Oct 9-10 1999,
2 day, averaged 2° x 2° box, (a) NN (b) SR and (c) NN-SR . . . .
128
5.17 Global distribution of MSMR derived SST (K) for Oct 9-10 1999, 2
day, averaged 2° x 2° box, (a) NN (b) SR and (c) NN-SR
.
129
5.18 Scatter plots of MSMR. NN derived geophysical parameter with that
of MSMR finish products, (a) WVC, (b) CLW, (c) SSW and (d) SST
131
5 19 Comparison of histogram of SSW from MSMR, (a) NN and (b) finished
product
............................................................................................
132
5 20 Positions of collocated TMI and ICOADS datasets (a) Ships (b) buoys 134
xv
5 21 Scatter plots of TMI NN derived parameters with ICOADS ships (a)
NN SSW (b) SR SSW 'c) NN SST (d) SR SST .
. . .
136
5 22 Scatter plots of TMI derived parameters with ICOADS buoys (a) NN
SSW (b) SR SSW (c) NN SST (d) SR S S T ............................
138
5.23 Scatter plots of TMI NN derived geophysical parameter with that of
TMI finish products, a) WVC, b) CLW, c) SSW and d) SST
xvi
. 140
Abbreviations
AMSR
Advanced Microwave Scanning Radiometer
BESEX
Bering Sea Experiment
BOBMEX
Bay of Bengal Monsoon Experiment
CLW
Cloud Liquid Water Content
Corrl
Correlation coefficient
ESMR
Electrically Scanned Microwave Radiometer
GOME
Global Ozone Monitoring Experiment
GP
Geophysical parameter
ICOADS
International Comprehensive Ocean-Atmosphere Data Set
MERIS
Medium Resolution Imaging Spectrometer
MSMR
Multi-frequency Scanning Microwave Radiometer
NDBC
National Data Buoy Center
NIOT
Institute of Ocean Technology
NN
Neural Network
RMS
Root-Mean-Square
SR
Statistical Regression
SSM/I
Special Sensor Microwave/Imager
SST
Sea Surface Temperature
SSW
Sea Surface Wind Speed
Tjg
Brightness Temperature
TMI
TRMM Microwave Imager
TRMM
Tropical Rainfall Measuring Mission
WVC
Total Water Vapour Content
XVII
Acknowledgem ents: D ata and
Software support
Several datasets, data visualization tools were used exhaustively m this thesis work
• SSM/I and TMI data were obtained courtesy Frank Wentz, from Remote Sens­
ing Systems, Santa Rosa, CA, U S A through ftp //f tp ssmi com/tmi
• The International Comprehensive Ocean-Atmosphere Data Set (ICOADS Re­
lease 2 1) http,//dss ucar edu/pub/coads
• The Climate Prediction Center using the method of Reynolds and Smith (1995)
and Smith and Reynolds (1998), ftp //f tp cpc ncep noaa gov/wd52yx/sstchm/
• ETOPO-5 data is courtesy of ’’Data Announcement 88-MGG-02, Digital relief
of the Surface of the Earth NOAA, National Geophysical Data Center, Boulder,
Colorado, 1988”
• Grid Analysis and Display System (GrADS) GNU Pubhc License
http-//www lges org/grads
• The Generic Mapping Tools (GMT) is an open source GNU General Pubhc
License http //gm t soest.hawan.edu
• The Microsoft Office etc http / / www.microsoft com and Latex
XVUl
Sum m ary
The thesis focuses on the development of Artificial Neural Network based retrieval
algorithms for various geophysical parameters using passive microwave radiometer
data The major contributions to the thesis work are the development of a new multi­
parameter retrieval algorithm based on the back-propagation Neural Network (NN)
approach for parameters like water vapour, cloud liquid water content, sea surface
wind speed and temperature from spaceborne microwave radiometers namely, IRSP4 Multi-frequency Scanning Microwave Radiometer (MSMR) and Tropical Rainfall
Measuring Mission (TRMM) TRMM Microwave Imager (TMI) The algorithms have
been developed using simulated database through radiative transfer models employ­
ing simulated atmospheric and surface conditions. NN is trained to establish the
relationship between water vapour content, cloud liquid water content, sea surface
wind speed and sea surface temperature, and brightness temperature for MSMR and
TMI radiometers The neural network architecture has been optimized through error
analysis suggesting the retrieval of multiple parameters simultaneously as the best
model having the configuration of 3 hidden layers with 15 neurons in each layer. Sep­
arate optimum models have been developed for MSMR (8 input - 4 output) and TMI
(7 input - 4 output) respectively. Study of impact of noise on NN based algorithms
has been carried out and it is found that retrieval errors are mostly m-variant with
noise indicating its effectiveness for retrieval The chapter wise summary of the work
XX
is as follows
Chapter-1 discusses the historical perspective of satellite microwave radiometry
and geophysical parameter retrieval and the problem statement
Chapter-2 describes the theory and principles of basic passive microwave radiom­
etry, pertaining to radiative transfer models applied to earth atmospheric system
Emission based microwave radiative transfer m atmosphere is discussed Measure­
ments of the brightness temperature with a brief summary of satellite microwave
radiometers have been presented This chapter also briefly discusses various retrieval
techniques like physical, statistical and neural network
Chapter-3 discusses the overview of neural network technique applied to microwave
radiometers
It mainly deals with the development of neural network for retrieval
purpose specifically the Back Propagation Neural Network (BPNN) technique
Chapter-4 discusses the results of BPNN studies carried out using simulated
brightness temperature for various satelhte sensors configuration. The simulation
dataset are obtained through RT model describe m chapter 2 The BPNN techmque
is specifically discussed for MSMR and TMI configuration. Various NN models have
been developed, like multi-layer configuration with single or multi-parameter output
Microwave radiometric data from MSMR and TMI has been used along with NN tech­
niques for deriving various geophysical parameters The use of data from different
satellites has been made to affirm the effectiveness of NN for retrieval purpose.
Chapter-5 deals with the validation and inter-comparison of derived geophysical
parameters with actual satellite data for MSMR and TMI with m-situ and other
similar satelhte products The details of the validation of various ocean parameters
are discussed
Chapter-6 presents the conclusions and future scope
C hapter 1
Introduction
C hapter 1
Introduction
1.1
Introduction
During the past two decades radiometry from space has developed into a powerful
technique for remote sensing of the earth’s atmosphere and surface. Microwave ra­
diometers on board earth orbiting satellites are now making significant contribution
towards the field of operational meteorology and oceanography in terms of measure­
ments of geophysical parameters viz. Sea Surface Temperature (SST), Sea Surface
Wind speed (SSW), atmospheric Water Vapour Content (WVC) and Cloud Liquid
Water content (CLW) and many more specifically over the oceans These parameters
are of great importance m a large variety of applications, such as meteorological fore­
casting (by means of data assimilation), studies of chmate processes, hydrological bal­
ance and radio communications. Space observations provide synoptic and repetitive
coverage of the ocean m contrast to the sparse and isolated in-situ ship observations,
therefore, greater emphasis is being placed on global geophysical monitoring of the
ocean-atmosphere system by means of microwave observations
Space based observation of earth atmospheric system is realized through exploiting
the interaction mechanism between the electromagnetic radiation and the geophysical
state of the earth and the atmosphere Physical properties of the earth atmospheric
1
1.1 Introduction
system affect these observations The crucial elements of inferring these parameters
from space based observation are the radiative transfer model as a forward model
and various retrieval techmques as inverse models. The retrieval model relates the
satellite radiance observations to the relevant Geophysical Parameter (GP)
The mathematical models used to infer these parameters make use of either sim­
ple techmques like statistical approaches or complex approaches like variational tech­
mques, and nonlinear iterative algorithms
These techmques are computationally
expensive Mostly the inverse problem involve mapping from limited known obser­
vations to many unknown geophysical variables (ill posed problem)
The inverse
modeling based on statistical approach need a prior mapping function which may or
may not be optimum due to non-linearity of radiative transfer model. In the Artificial
Neural Network (ANN) m shore (NN), the non-linearity is included through many
neurons and are found to be better than statistical approach. Neural network on the
other hand, when presented with the known input, and output patterns, inherently
form nonlinear models of the data used to tram them, unlike m nonlinear regression
technique where a particular type of nonlinear function must be specifed in advance.
Thus, we are forced to implement a particular type of nonlinearity known m ad­
vance Because the NN technique is a generic technique for nonlinear mapping, it can
be used beneficially for modeling transfer functions for forward and inverse models
(Krasnopolsky and Chevallier, 2003).
Since the 1970!s, satellites have been making a global sampling of our Earth’s
system which are inherently unobtainable by ground-based systems These remote
sensmg measurements from space axe rapidly increasing as satellites are launched with
more sophisticated, hyper-spectral sensors with more spatial (both vertical and hori­
zontal) dimensions This trend will continue resulting the high-volume of information
2
1.2 Historical Perspective
collected from these satellite For most retrieval algorithms, adding new information
to improve the retrievals is not a simple task because of the nonlinear nature of the
problem as well as the computational difficulties (Vann and Yong-X., 2002) ANN’s
are well adapted to solve nonlinear problems and are specially designed to capitalize
on the inherent statistical relationships (Aires et al., 2001). Here, adaptive means
that the method is able to process a large amount of data or deal with new relevant
variables
The performance of NN approach compared to the multiple regression
approach (Vann and Yong-X , 2002) showed major improvement over other retrieval
methods because it has an obvious advantage m computational speed when applied
to inversion techniques Unlike the iterative technique, requiring to solve radiative
transfer equation hundreds of tames for each sample of satellite measurements, the
neural networks only needs a few simple operations and the inversion can be done m
near real-time
1.2
Historical Perspective
In late 1960’s, microwave remote sensing was first recognized as a powerful technique
for atmospheric studies due to its abihty to measure atmospheric WVC and CLW
(Barrett and Chung, 1962; Staelin, 1966)
They have discussed the possibility of
determining the abundance of high-altitude water vapor by making use of 1 35cm
resonant water-vapor line m emission spectra
Staelin et al. (1976) used passive
microwave spectrometer on the Nimbus- 5 satelhte (Nimbus E) Satellite Microwave
Spectrometer (NEMS) with two channels at 22 235 and 31 4 GHz, for measuring
atmospheric WVC and CLW and observed WVC in the range up to 6 g/am2 with
RMS differences of 0 4 g/crn2 from radiosonde, and CLW m the range 0.01 g/crn2 to
0 2 g/crn2 with an RMS error of 0 01 g/cm 2 (Grody, 1976) used regression technique
3
1.2 Historical Perspective
along with radiation transfer theory and measurements at frequencies near the 22 235
GHz to retrieve WVC and CLW. In these retrievals the effect of wind at the ocean
surface was ignored
Stogryn (1967) developed a theory to account for the wind-
induced roughness, which was further tested by Hollmger (1971) using tower based
radiometric measurement The most obvious foam effects were from the data and
it was found that the roughness effect was somewhat less than the Stogryn theory.
Using airborne data, Nordberg et al (1971) characterized the combined effects if
foam and roughness at 19.35 GHz Geometric optics theory by Stogryn (1967) was
extended to include diffraction effects, multiple scattering, and two-scale partitioning
by Wu and Fung (1972) and Wentz (1975)
Retrieval of SSW, WVC, and CLW
based on airborne data from joint US-USSR Bering Sea Experiment (BESEX), in
wavelengths ranging from 0 8 to 2.8 cm at altitudes from 0 16 to 11 km Wilheit
and Fowler (1977) is well documented for various meteorological conditions over the
Bering Sea. Later Chang and Wilheit (1979) combined two NIMBUS-5 instruments,
the ESMR and the NEMS for SSW, WVC, and CLW retrieval Wilheit (1979a) used
the 37-GHz dual polarized data from the Electrically Scanned Microwave Radiomter
(ESMR) to explore the wind-induced roughness m ocean surface retrieval algorithm
This was later combined with other data to generate a semi-empirical model for
the ocean surface emissivity Wilheit (1979b) to be used for Scanning Multichannel
Microwave Radiometer (SMMR) on the Nimbus-7 and Sea-Sat satellites A theory
for the retrieval of WVC, CLW, SSW and SST ocean parameters was published by
Wilheit and Chang (1980) The launch of the SeaSat and Nimbus-7 SMMR’s spurred
much investigation on SMMR retrieval algorithms and model functions (Wentz, 1983;
Njoku and Swanson, 1983; Ahshouse, 1983, Chang et al., 1984, Gloersen et a l , 1984),
with the state-of-the-art oceanic microwave radiometry
4
It became clear that the
1.2 Historical Perspective
water vapor retrievals were highly accurate. First Indian remote sensing satellite
Bhaskara-I launched m 1979 carried a microwave radiometer (SAMIR) operating at 19
and 22 GHz, followed by Bhaskaxa-II satellite carrying similar microwave radiometer
operating at 19, 22 and 31 GHz Atmospheric WVC and CLW were derived using
SAMIR-I and II (Pandey et al., 1981, 1984; Gohil et a l , 1982)
A major improvement m the wmd retrieval was made when Wentz et al (1986)
combined the SearSat SMMR TB’s and the Sea-Sat scatterometer wind retrievals
to develop an accurate, semi-empirical relationship for the wind-induced emissivity.
The measurement of SST requires relatively low microwave frequencies (4-10 GHz)
The SMMR was the first satellite sensors with the appropriate frequencies to retrieve
SST However, the SMMR suffered from a poor calibration design, and the reported
SST retrievals (Njoku and Swanson, 1983, Milman and Wilheit, 1985) were useful for
httle more than a demonstration of the possibility of SST retrievals for future better
calibrated radiometers. With the launch of the Special Sensor Microwave Imager
(SSM/I) m 1987 many retrieval algorithm were developed. In contrast to SMMR,
SSM/I has an external calibration system that provides stable observations Unfor­
tunately, the lowest SSM/I frequency is 19 3 GHz, and hence SST retrievals are not
possible. Single parameter algorithms such as the Goodberlet et al. (1989) wmd algo­
rithm derived with near simultaneous and collocated measurements made by offshore
ocean buoys, based on the D-matrix approach, but has less expected accuracy, it was
further improvement by Goodberlet et al (1990) and showed retrieval accuracy of
the order ±2 m/s. A physical based approach was used by Schluessel and Luthardt
(1991) for WVC retrieval using SSMI data. Guissard (1998) used simplified physical
based retrieval algorithm apphed to SSM/I data to retrieve WVC and CLW with
5
1.2 Historical Perspective
RMS differences of 0 037g/cm2 and 0 004 g /a n 2 when compared to ECMWF (Euro­
pean Center for Medium Range Weather Forecast) and Gerard’s algorithm (Gerard,
1996) respectively Wentz (1992) used SSM/I measurements with collocated buoy
(NDBC) data and developed physically based algorithm to retrieve the SSW The
RMS difference between SSM/I and buoy SSW -were about 1.6 m /s Wentz (1997)
algorithm for retrieving geophysical parameters over the ocean using SSM/I provided
RMS accuracy between 0 5 and 1 K for SST Similar physical approach is also used
m for AMSR ocean algorithm by Wentz and Meissner (2000)
In November 1997,
the first microwave radiometer capable of accurately measuring SST through clouds
was launched on the Tropical Rainfall Measuring Mission (TRMM) spacecraft The
TRMM Microwave Imager (TMI) is providing an unprecedented view of the oceans
Its lowest frequency channel (10 7 GHz) penetrates non-raining clouds with little at­
tenuation, giving a clear view of the sea surface under all weather conditions except
rain Furthermore at this frequency, atmospheric aerosols have no effect, making it
possible to produce a very reliable SST Connor and Chang (2000) used standard re­
gression techniques with m-situ meteorological buoy measurements to retrieve SSW
A Multi-frequency Scanning Microwave Radiometer (MSMR) similar to SMMR was
launched onboard Indian satellite IRS-P4 in 1999 MSMR operated m channels 6,
10, 18 and 21 GHz with dual polarization providing global measurements of WVC,
CLW, SSW and SST (Gohil et a l , 2000)
The above analyses have shown the use of classical statistical and physical based
retrieval techniques Among other techniques m the artificial intelligent community
like Bayesian algorithm, Neural network and hybrid algorithms (e g. neural network
with genetic algorithm) have also been used Bayesian algorithms have been used
for few retrievals, like cloud liquid water properties from microwave radiometer and
6
1.3 Neural Network based Analysis
millimeter radar data (McFarlane et a l , 2002), Bayesian estimation for land surface
temperature retrieval by Morgan (2005) and Bayesian monte carlo retrieval algorithm
is demonstrated by L’Ecuyer and Stephens (2002), but such algorithms, m addition
need a prior knowledge in the form of conditional probability or probability density
function, which makes it cumbersome to be used However, one of the disadvantages
of Bayesian network is its inability to handle continuous probability distributions
directly They are often approximated by taking the histograms by fitting various
analytical distributions to the data (Frus-Hansen, 2002). The problem is that the
computational complexity becomes intractably large because of fine discretisation of
probability distribution In general Bayesian network has many advantages to offer m
retrieval models However, the complexity of the models may very well grow beyond
traceability. Neural network offer much simple yet powerful modeling tool, extensive
analysis has been discussed below.
1.3
Neural Network based Analysis
Most of the operational retrieval algorithms are based on regression and other statis­
tical approaches During the last decade a new generation retrieval model based on
atrificial neural network technique has found applications m a large variety of different
fields; specifically, m satellite remote sensing, for fast and accurate approximation of
model physics, for solving forward and inverse problem (Krasnopolsky and Chevalher,
2001, 2003; Krasnopolsky et a l , 1999)
A number of publications using NN have been devoted to particular applications,
most of them are single parameter algorithm
NN algorithm as an alternative re­
gression techmque to nonlinear transfer function of radar measurements to retrieve
7
1.3 Neural Network based Analysis
the wind vector from ERS-1 scatterometer data (Thiria et a l, 1993) and later ap­
plied by (Cornford et a l , 2001)
A few review papers related to NN based have
been published m Image processing, classification, and prediction areas by (Atkinson
and Tatnall, 1997, Gardner and Dorlmg, 1998; Hsieh and Tang, 1998) serves as good
service m introducing NN technique to remote sensing community. NN techmque
also applied for inversion of a multiple scattering model to estimate snow parameters
from passive microwave measurements Tsang et al (1992) Smith (1993) used NN for
inversion of a simple two-stream radiative transfer model to derive the leaf area index
from Moderate Resolution Imaging Spectrometer data (Pierce et a l , 1994) developed
an inversion algorithm for radar scattering from vegetation canopies Ozone profile
retrieval from GOME data using a NN was developed by (Mueller, 2003; Abdelgadir,
1998) applied NN for forward and inverse modeling of canopy directional reflectance.
Schiller and Doerffer (1999) used NN techmque for inverting a radiative transfer for­
ward model to estimate the concentration of phytoplankton pigment from Medium
Resolution Imaging Spectrometer (MERIS) Feed-Forward NN was developed for hu­
midity retrievals (Cabrera-Mercader and Staekn, 1995) Davis (1995) applied NN for
inversion of a forward model to estimate soil moisture, surface air temperature and
vegetation moisture from SMMR data
The first NN algorithm for retrieving SSM/I wind speed was developed by Stogryn
et al. (1994) for retrieving single parameter, SSW, from SSM/I X^s Two separate
NN algorithm was developed with the retrieval accuracy of 1 41 m / s and 2.39 m / s for
clear and under cloudy conditions respectively. Krasnopolsky et al. (1995a) showed
with similar single NN algorithm, could generate the same retrieval accuracy as the
two NNs developed by Stogryn et al
for both clear and cloudy conditions with
RMS difference of 1 05 m / s were obtained for with the absence of large absorption
8
1.3 Neural Network based Analysis
Ocean surface air temperature from multiple data sets using artificial neural network
has been derived by (Catherine et a l , 1998)
The performance of ANN approach
compared to the multiple regression approach has been performed by (Jung et al.,
1998) and CLW retrieve with RMS error of 0 009 K g / m 2 over the ocean from SSM/I
data. Krasnopolsky et al (1999) introduces a empirical multi parameter retrieval
algorithm using SSM/I data based on NN approach
The NN approach has also
been used for approximations of model physics m ocean numerical models (Krasopolsky et a l , 2002) Krasnopolsky and Schiller (2003) showing that NNs can be used
for global inversion to explicitly invert a forward model Methodology for the for­
mulation of multi parameter and multi-instrument retrieval from TRMM have been
presented Obligis et al (2001). A parallel retrieval algorithms using BPNN for wa­
ter vapour, cloud liquid water path, surface temperature and emissivity over land
from satellite microwave observation have been developed by Aires et al (2001). NN
based retrieval algorithm has been developed for atmospheric temperature profile us­
ing AMSU-A measurements by Lei (2001). NN based retrieval of atmospheric and
surface temperatures was developed by Aires et al (2002) NN based estimation has
been developed for surface specific humidity and air temperature using MSMR Singh
et al. (2004)
The above literature has shown ANN as a promising technique m retrievals. The
ANN algorithms have three salient and attractive features. Specifically, ANN algo­
rithms are not only computationally efficient, but they are also very useful m repre­
senting nonlinear relationships among a set of parameters, and also they can handle
noisy or incomplete and massive amount of data All these previous studies form a
strong basis for the present attempt to evaluate the performance of neural network
approach for MSMR radiometric channel simulations and retrievals.
9
1.4 Problem Statem ent
1.4
Problem Statement
The thesis focuses on the development of ANN based retrieval algorithms for various
geophysical parameters using passive microwave radiometer data Specific objectives
of the study are• Study the Radiative Transfer Model (RTM) and carry out simulations for de­
veloping retrieval algorithms
• Develop an ANN scheme for geophysical parameter retrieval
• Retrieve various geophysical parameters over oceans from IRS P-4/MSMR and
TRMM/TMI
• Validation and inter comparison of derived geophysical parameters with in-situ
and similar products from other satellites
The research work carried out is presented m the following chapters
10
C hapter 2
B asic P assive M icrow ave
R ad io m e try an d R etriev al
C hapter 2
B asic Passive M icrowave
R adiom etry and R etrieval
2.1
Introduction
This section presents the fundamental principles of radiometry, a brief summary of
radiative transfer theory, radiometric measurement technology, microwave radiometer
systems and various retrieval methods,
2.2
Thermal Radiation and Radiative Transfer
Passive microwave remote sensing is based on the measurement of thermal radiation
emitted by the earth atmospheric system m the microwave portion of the electro­
magnetic spectrum The emitted radiation is largely governed by its physical state
and the radiative properties The microwave region spans in the range 1-1000 mm of
wavelength Any object with temperature above 0° K radiates or absorbs radiation
through transitions of energies during vibration/rotation of molecules. The average
kinetic energy of the vibrating molecules is represented by the absolute temperature
(K) Electromagnetic radiation from the sun is the principal source of energy that
11
2.2 T herm al R ad iation and R ad iative Transfer
drives circulations in both the atmosphere and ocean. Energy received by a radiome­
ter depends upon various contributions from the atmosphere and earth surface within
its field of view as depicted in figure (2.1).
Space
O c e a n S u rfa c e
Figure 2.1: The major components of the radiation received by a radiometer
Radiative Transfer (RT) theory describes the interactions of matter with elec­
tromagnetic radiation. Radiative transfer serves as a mechanism for explaining the
basis for remote sensing. The radiation leaving the earth-atmosphere system sensed
by a satellite borne radiometer is the sum of radiation emitted by the earth surface
and various atmospheric layers up to the top of atmosphere. One useful quantity
describing the amount of thermal radiation emitted by a body is spectral brightness.
Some part of sections are as explained in Chen (2004). Spectral brightness is a mea­
sure of how much energy a body radiates at a specified frequency per unit receiving
area, per transmitting solid angle and per unit frequency. The spectral brightness of
a blackbody {Wstr~lm,~2Hz~l ) is a function of its physical temperature T(K) and
12
2.2 T h e rm a l R a d ia tio n a n d R a d ia tiv e T ra n sfe r
frequency f (Hz) and is given by Planck’s law:
2h f
=
c 2[e hf / k.T _ ^
(^ ^ -
1)
where h is Planck’s constant (J.s), c is the speed of light (m/s), and k is Boltz­
mann’s constant (J/K ). The function defined by eqn. (2.2.1) is graphically portrayed
in figure (2.2) at different temperatures. In the microwave region, hf/kT is much less
than unity, this assumed the familiar Rayleigh-Jeans approximation yielding.
O IT f 2
B(f, T) « -—j —
(2.2.2)
Planck Spectral Exitance for V arious B lackbody Tem peratures
1.E+10
1.E+09
1.E»08
1.E *0 7
1.E+06
~
1 .E t0 5
*
1.E *04
Ti
^
1.E*03
P
1.E+02
^
1.E*01
9E 1.E*00
1.E-01
1.E-02
1.E-03
1.E-04
1.E-05
1E-02
IE-01
1E*00
1E*01
W avelength ip rn)
1E*02
1E*03
IE * 04
ftltlHl* «■<«
r.»i m «
Figure 2.2: Planck’s black body radiation at different temperatures
Thus in microwave region, the black body radiation is proportional to its physical
13
2.2 Thermal Radiation and Radiative Transfer
temperature For black bodies, given an observation of spectral brightness, one can
calculate the physical temperature of a black body-
rp
2kf 2
where T is the physical temperature of the ideal black body.
2.2.1
Brightness Temperature
In microwave remote sensing, it is common to work with brightness temperature (Tb )
rather than the brightness it self Equation (2.2.2) refers the brightness of a black
body at physical temperature T Unlike black bodies, grey bodies reflect some of the
energy incident upon them, so the intrinsic spectral brightness of a grey body is not
equal to that of a black body For a grey body,
B
(
f
cz
,
T
)
=
e
(2.2 4)
where e is the emissivity of the grey body. This equation can be rewritten as
follows2k f 2
R (/,T) = - ^ - £ T
(2.2 5)
The quantity eT is called the brightness temperature of an umlluminated grey
body, l e the temperature of a black body radiating the same brightness under the
conditions of equilibrium, the ability of a body to radiate is closely related to its
ability to absorb radiation Thus the brightness temperature (TB) and the physical
temperature, T, are related through emissivity, e{6, <f>), as follows
14
2.2 Thermal Radiation and Radiative Transfer
TB =e(0,<t>)T
(2 2 6)
The mathematical formulation of this statement is known as Kirchoff’s Law. Thus
the brightness temperature of a grey body is always smaller than or equal to its phys­
ical temperature T Another useful property of a grey body is known as reflectivity
r, and sum of emissivity e, and reflectivity r, is unity (r + £ = 1.)
2.2.2
Radiometric Brightness Temperature of a Scene
In an atmosphere without hydrometeors, the brightness temperature observed by an
earth-observmg spaceborne radiometer can be divided into four components.
1. T b u p , the upwelhng brightness temperature emitted by the atmosphere
2 Tb s , the brightness temperature emitted by the surface
3 T Bd n j the downwelhng brightness temperature emitted by the atmosphere and
reflected by the earth’s surface
4. TBc , the cosmic background brightness temperature attenuated by the at­
mosphere and reflected by the earth’s surface
The observed brightness temperature is a function of several variables such as
atmospheric temperature profile, water vapor profile, surface temperature, surface
roughness and salinity, radiometer operating frequency and observational angle The
radiative transfer equation for a non-scattermg atmosphere under local thermody­
namic equilibrium at a frequency is given by
T b = T b s + Tb u p + TB d n + T b c
15
(2 2.7)
2.2 Thermal Radiation and Radiative Transfer
where
fH
7 W = sec(0) /
T { z ) a f (z)Tf{z - f H,9)dz
(2.2.8)
Jo
TBS = £f (9,p)TsTf ( Q ^ H , 9 )
(2.2 9)
poo
TBDN = ( l ~ s f (9,p))Tf (Q-^H,e)sec(e) /
z,0)<fe
(2 2.10)
Jo
Tec = (1 - e/(0,
- ff, 0)
(2 2.11)
with
TS{zi Where t/ =
(0 1
02, 0) = e“ seoW# “/(z)dz
(2 2 12)
—» 0 2,$) is the atmospheric transmittance between altitudes zi
and z2, W1th cnf(z) (Np per unit length) is total atmospheric absorption coefficient,
e is emissivity of the surface, p is polarization, H is satellite altitude, Tscosmic^) is
the cosmic background brightness temperature and 9 is satellite zemth angle
The emissivity of the ocean surface is dependent on variables such as ocean surface
roughness, salinity, and sea surface temperature
The atmospheric absorption not
only depends upon the concentration of atmospheric gases like oxygen and water
vapour which is governed by temperature and humidity profile but also depends upon
the concentration of water drops present in clouds and precipitation. Thus the total
atmospheric absorption depends on vertical profiles of temperature, humidity cloud
and precipitation In addition, the radiation emitted by earth atmospheric system
also varies with radiometer frequency, polarization and the viewing angle. These
processes are subsequently discussed m detail
16
2.3 Passive Microwave Remote Sensing of the Ocean and Atmosphere
2.3
Passive Microwave Remote Sensing of the Ocean
and Atmosphere
Brightness temperatures observed by a downward looking satellite-borne radiometer
depends on many variables in a complex and nonlinear manner as summarized m
previous section It is sensitive to vertical temperature profile, water vapour content,
cloud liquid water content and the surface parameters Retrieval of geophysical para­
meters by direct inversion (analytically) is very difficult However, the physics of the
atmosphere still allows one to extract useful information about the atmosphere from
microwave frequency bands
2.3.1
Sea Surface Emissivity
The ocean surface plays an important role in modifying the radiation received by a mi­
crowave radiometer firstly, through surface contribution arising from surface temperar
ture and emissivity, and secondly, through surface reflection affecting the atmospheric
downwellmg radiation. The ocean surface emissivity or reflectivity (e + r = 1) is a
function of SST, salinity, surface wind speed and foam. It is also a function of
frequency, incidence angle and polarization. A typical variation of ocean surface
emissivity with varying angle is depicted m figure (2.3) for horizontal and vertical
polarization It is seen that the surface emissivity increases m vertical polarization
while it decreases m horizontal polarization with increase in incidence angle.
The variation of emissivity with other parameters are discussed below
17
O c e a n S u rfa c e E m is s iv ity
2.3 Passive Microwave Remote Sensing of the Ocean and Atmosphere
0
10
20
30
40
50
SO
70
SO
90
In cid e n c e A n g le (D e g )
Figure 2,3 A typical variation of Ocean surface emissivity with varying angle
SST and Salinity
Study of dependency of emissivity on SST, salinity, and frequency at nadir angle
for typical ocean salinity and frequency reported m the literature indicates that the
brightness temperature at 1 4 GHz is independent of the SST and emissivity at 1.4
GHz varies inversely with salinity For frequencies above 4 GHz the emissivity m
general first decreases and then increases with SST around 290 K For frequencies
less than about 4 GHz, the emissivity varies inversely with salinity, whereas above
18
2.3 Passive Microwave Remote Sensing of the Ocean and Atmosphere
4 GHz, the emissivity is independent of salinity. For low frequency microwaves, the
brightness temperature depends upon both SST and salinity A typical variations
of emissivity at 6 6 GHz with SST and SSW are shown in figure (2 4) indicating
increasing trend of emissivity with SSW while a decreasing trend with increasing
1»
<b} a s -
Ocean Surface E m s s lv lty
SST
if
as-
S e a S u rfa c e T e m p e ra tu re ■ 3 0 0 K
S u rfa c e W in d S p e e d • 1 0 m /s
In c id e n c e A n g le ** 5 0 D e g .
Fequency® 66G H z
S e a S u rfa c e T e m p e ra tu r e * 3 0 0 K
S a lin ity ® 3 5 p p t
-- — S OH*V-POl
ta a tte W O L
In c id e n c e A n g le “ 5 0 D e g
S a li n it y » 3 5 p p t
...
6GH*H-P0t
-
ip
stutf-poi
272 276 280 284 268 232 286 200304
Sea S urlaceTenrperature(K)
0
4
8
12
IS
20
Sea Surface Wind Speed (m /s )
24
Figure 2.4 A typical variations of emissivity at 6.6 GHz (a) SST and (b) SSW
Ocean W ind Speed
When the sea surface is perturbed by the drag force of wind, the emissivity (or the
reflectivity) deviates significantly from that of a specular surface. The change in
emissivity depends on the amplitude of small scale ocean waves and the surface RMS
slope measured m terms of free space wavelength (A) of the particular radiometer
frequency of interest. The effect of surface roughness is more pronounced as the
wavelength decreases (increasing frequency) The ocean surface roughness has non­
linear relationship with wind speed; and beyond a threshold value of about 6 m/s,
19
2.3 P assiv e M icrow ave R e m o te S en sin g o f th e O c ean a n d A tm o s p h e re
Figure 2.5: Effect of wind speed (surface roughness) on ocean brightness temperature,
SSW W3 > W2 > W1
waves begin to break and sea foam is generated which significantly effects the enrissivity. Wind speed is selected as the appropriate parameter because the small scale
ocean surface roughness is believed to be directly proportional to the energy input
from the ocean surface wind speed (frictional wind speed). Because of the difficulty in
measuring small scale ocean waves (ocean wavelengths < 1 m), the neutral stability
surface wind speed has been used as the oceanographic or meteorological geophysical
parameter to correlate with changes in T b ■ figure (2.5) illustrates the dependency of
T b on surface wind speed for different incidence angles for a typical frequency. These
radiometric signatures are qualitatively representative of majority of microwave fre­
quencies. For horizontal polarization, T b increases monotonically wind speed at all
incidence angles while for vertical polarization, there is a ’’hinge point'’ between 50
and 60 degrees incidence where the brightness is insensitive to wind speed. Below this
hinge point, T b increases with wind speed: and above this point, it decreases. This
figure suggests that a dual polarized T b measurement near 53 degrees can be used to
infer wind speed by determining the differential T b between the two polarizations.
20
2.3 Passive Microwave Remote Sensing of the Ocean and Atmosphere
Empirical Foam Model
As mentioned earlier, sea foam has a significant impact on the ocean surface emissivity The foam covered area depends upon the intensity of winds which generates
whitecaps
Experiments have been performed to determine the Tg for 100% foam
covered surfaces These data were collected using tank experiments where the foam
was artificially generated as well as using field experiments where it was naturally
produced In general the emissrvity is a function of foam thickness until it becomes
thick (i.e , t > A/4), after which, it becomes constant at a value of > 0 9 Further
for thick foam, the emissivity is independent of incidence angle and polarization. For
a partially foam-covered sea surface, the emissivity is a linear combination of the
emissivity due to wind and the foam coverage and is given as:
e{0, A,p) = (1 - C/)£™(0, A,p, SSW) + Cf sf (e, A,p)
where Cj = the fractional foam coverage erw— the emissivity of the wind rough­
ened surface s j = the emissivity of the surface when covered with foam.
The foam emissivity within the microwave domain is very high, close to one
From radiometric measurements below 40 GHz, Stogryn (1972) proposed an empirical
model, depending on incidence angle, frequency and polarization conditions. An
alternative may be to use a physical approach Droppleman (1970), Hosenkranz and
Staelm (1972) But, m addition to the frequency, the incidence angle, the polarization
and the SST, these theoretical models are sensitive to the depth of the foam and to
its density At high frequencies, given the uncertainty related to the foam depth and
density, an emissivity of one seems to be an acceptable approximation for the foam
(weak sensitivity to the SST, to the incidence angle (up to 45) and to the polarization
at 89 and 157 GHz).
21
2.3 Passive Microwave Remote Sensing of the Ocean and Atmosphere
A tm o sp h er e A b so rb tio n
The atmosphere interacts with electromagnetic radiation through the transition of en­
ergies of various rotatmg/vibratmg molecules of the gases present m the atmosphere
Due to certain allowed transitions of energies, the absorption/emission spectra of var­
ious atmospheric gases have many resonance peaks at specific frequencies as shown m
figure (2 6) At these peaks the absorption is higher as compared to the neighboring
frequencies The frequencies regions between two successive absorption peaks are also
known as window region The absorption bands/peaks are suitable for sounding of
the atmosphere while the window region is useful for surface sensing More over, the
hydrometeors present m the atmosphere m the form of clouds and rain interacts with
radiation through their dielectric properties causing absorption, scattering and extinc­
tion The absorption due to hydrometeors increases with increasing frequencies as the
wavelength becomes smaller than the size of hydrometeors However the scattering
also increases with frequency (Chahrne, 1983, Ulaby et a l , 1982). At low frequen­
cies or larger wavelengths the scattering is either negligible or small due to smaller
size parameter (27rr/A, r is hydrometeor radius) Due to scattering by hydrometeors
higher frequencies are useful for cloud and ram detection and monitoring
For MW frequencies in the window regions where atmospheric absorption is less
as compared to those of gaseous absorption resonance frequencies, the atmospheric
transmittance is high and the upwellmg and downwelhng radiations are low, the major
contributions to the total radiation reaching at the top of the atmosphere comes from
the surface which mainly depends upon the surface emissivity This makes window
channels more suitable for the surface sensing Microwave channels m the window
regions are used for remote sensing of surface parameters like surface temperature,
sahmty and wind intensity over oceans These channels are also suitable for oceans
22
ZaumOpacity[dBl
%
2.3 Passive Microwave Remote Sensing of the Ocean and Atmosphere
200
FrequencyJGlL'}
Figure 2 6 Transmission function of the cloud free atmosphere m the microwave
spectral range
due to low sea surface emissivity and also due to well modeled emissivity m terms of
ocean surface parameters In combination with low absorbing water vapour resonance
frequency, these channels are very useful for remote sensmg of total atmospheric water
vapour and cloud liquid water contents
For channels operating in the gaseous absorption resonance bands where at­
mospheric absorption is very high and following the formal solution of RT equation,
the total atmospheric transmittance is almost negligible prohibiting the surface and
the reflected downwellmg radiation but allowing upwelbng atmospheric radiation to
reach up to the top of the atmosphere
This makes channels m the resonance re­
gion useful for atmospheric sensing or the sounding of atmospheric temperature and
humidity. For temperature sounding, channels operating m the oxygen absorption
band are used, while for humidity sounding, channels operating m the water vapour
absorption band are used
23
2.3 P assive M icrowave R em ote Sensing o f th e O cean and A tm osphere
The atmospheric absorption model described by Liebe et al (1992, 1993) is dom­
inated m the microwave region by two Van Vleck-Weisskopf broadened water vapor
lines, at 22 and 183 GHz, together with an oxygen absorption complex of lines taken
from Rosenkranz (1993), as well as a water vapor continuum term Numerous para­
meters of the Liebe’s model have been empirically fit to various observational data
sets
The atmospheric absorption model described m Cruz Pol et al. (1998) is a
modification to Liebe’s that is based on a refined set of observations of atmospheric
downwelhng brightness temperature by a radiometer operating m the near vicinity
of the 22 GHz water vapor line. A 1 3% increase m the line strength, together with
a 6 6% increase m the line width, of the 22 GHz absorption line are determined to
be statistically significant corrections to the Liebe’s model within the range of 18-37
GHz
2.3.2
Overview of the Sea Surface Emissivity Modelling
The specular emissivity of the ocean is a function of the frequency of operation and
the dielectric properties of the sea water Klein and Swift (1977) re-exammed Stogryn
(1972) regression fit to take into account the new sets of measurements of Ho and
Hall (1973) and Ho et al (1974), Klem’s uses a simple Debye expression for sea water
dielectric over a limited frequency range (under 10 GHz), and polynomial fits for the
static dielectric coefficient, the ionic conductivity and the relaxation time as a function
of temperature and salinity. Laboratory measurements made by Ellison et al. (1998)
confirm that the single Debye model can be used at low frequency (under 30 GHz) and
lead to new coefficients of the Debye model for sea water m natural conditions which
was validated using both spaceborne and airborne radiometer data Guillou et al.
(1998) Commonly three mam type of sea surface emissivity are used, as stressed by
24
2.3 Passive Microwave Rem ote Sensing of the Ocean and Atmosphere
Guillou et al. (1996), the simple model, geometry optics, and two-scale approaches.
This section is as describe in Eymard et al. (2000).
Simple model:
Most of the empirical models have been determined from
Hollinger (1971) measurements and from SMMR observations. All these measure­
ments are limited to the low frequency range, up to 37GHz. They are highly depen­
dent on the instrumental calibration and scanning geometry. The input parameters
are usually limited to the sea surface wind speed, the frequency and angle of obser­
vation, and the sea surface temperature. To the flat sea surface specular emissivity,
a corrected term is added to account for the surface roughness. A fractional foam
coverage is often considered. Derived from low-frequency data, some of these models
can potentially be used for other frequencies, at any sounding angle. This is the case
of Francis et al. (1983) algorithm among other. Rosenkranz (1992) extrapolated the
Wentz’s semi-empirical emissivity model Wentz (1983) to the SSM/I channel charac­
teristics.
Geometric optics approach:
The ocean surface is described by a series of
reflecting plane facets characterized by an anisotropic slope distribution. Each indi­
vidual facet is assumed to be infinite in front of the wavelength, and irregularities
small or comparable to the wavelength are ignored. The elementary contribution
of each facet to the upwelling brightness temperature is computed from the Fres­
nel reflection relations. One of the first calculations using this approach is due to
Stogrvn (1967). Wilheit (1979b) model is quite similar, although different in its for­
mulation for the emissivity calculation. Prigent and Abba (1990), model described is
derived from this one. Petty and Katsaros (1994) also propose a parametrization of
a geometric optics model for the specific SSM/1 configuration.
25
2.3 Passive M icrowave R em ote Sensing of th e O cean a n d A tm osphere
Tw o-scale approach: Attempts to enhance the agreement between theoretical
predictions and low frequency observations led to the development of compositesurface models Following previous work to interpret the mechanisms in the radar
backscattenng from the ocean at various incidence angles, Wu and Fung (1972),
Wentz (1975); Guissard and Sobieski (1987), developed two-scale scattering models,
by superimposing small structures (capillary waves, small gravity waves) on the large
undulations (gravity waves). The scattering coefficients are expressed as the sum
of two contributions. The first term concerns the large scale and is given by the
geometric optics solution, slightly modified by the presence of the ripples, which
impose a modification of the Fresnel reflection coefficients. The second term results
from the average of the scattering coefficients due to the small irregularities over
the large scale slope distribution, and these scattering coefficients are derived from
the small perturbation theory The two-scale model by Guissard’s, was improved by
Lemaire (1998) At low frequency, the two-scale models have been shown to provide
better overall agreement with Holhnger’s observations, when fine timing the input
parameters (Wu and Fung, 1972, Wentz, 1975) The separation of the wave spectrum
m two scales is based on a more or less arbitrary choice of the cutoff frequency or
of the small-scale rms height as a function of the wind velocity and of the sounding
wavelength
The radiative transfer model used for simulation m this research, involve computa­
tion of atmospheric absorption Liebe et al. (1992) at MSMR frequencies and incident
angle for known atmospheric physical state comprising of atmospheric constituents
hke dominant gases and hydrometeor. It also involve the computation of ocean sur­
face emissivity Holhnger (1971) Stogryn (1972) at a given frequency, incident angle
and polarization from the known ocean surface physical conditions like sea surface
26
2.3 Passive M icrowave R em ote Sensing of th e O cean a n d A tm osphere
temperature, salmity, wind speeds and foam Wilheit (1979b).
Table (2 1) shows sensitivity of SMMR channels on various geophysical parameters
as reported by Prabhakara et al (1982) It is seen that 22 GHz channel has maximum
sensitivity on WVC while 18 GHz has maximum sensitivity on CLW It is also seen
that 6 and 10 GHz channels have maximum sensitivity with surface parameters like
SSW and SST respectively More over, the radiation observed by the radiometer at
a given frequency/channel is also affected by parameters other than the parameter
having maximum sensitivity implying that multiple channels should be used for better
retrieval
Table 2 1 Sensitivity of SMMR brightness temperature Tg, to geophysical parameter
Prabhakara et al. (1982)
Surface C hannels
Frequency (GHz)
66
A tm ospheric C hannels
10 7
18 0
210
Polarization
V
H
V
H
V
H
V
H
§ k & /K )
0.5
03
05
0.3
0.3
0.2
02
01
0.5
1.0
05
10
0.5
1 0 05
1.0
0.3
05
08
1.3
4.5
7.5
11.9
19.6
10 2 03
05
07
11
1.6
27
15
25
M vc(K cm 2/g )
;
^ j f c ( K c m 2/ g ) *
27
2.4 Microwave Radiometer Instruments and Satellite Missions
2.4
Microwave Radiometer Instruments and Satel­
lite Missions
2.4.1
Microwave Radiometer
Radiometers are highly sensitive receivers used to measure electromagnetic radiation
emitted by earth atmospheric systems The measurement by the radiometer is termed
as antenna temperature (Ta) which incorporates the intensity of radiation mcident
upon the antenna as well as self emission by the antenna structure itself The term
radiometric brightness temperature is used to characterized the emission by a material
through the expression (Ulaby et a l , 1982)
T'b ■
kpG
(2 4 1
)
where k, Boltzmann’s constant 1 38 x 10~23 J/'K, p system bandwidth (Hz), G,
system power gam and P, is the power emitted by the material over the bandwidth
P Corresponding to the power P received by a radiometer antenna, a radiometric
antenna temperature Ta is defined by
Ta
kpG
(2.4 2)
If the scene observed by the antenna beam is characterized by a uniform brightness
temperature Tg, then I \ = Tg
2.4.2
Description of Satellite Microwave Radiometers
A spaceborne imaging microwave radiometer essentially consists of a rotating par­
abolic antenna reflector, which achieves across track scanning either by mechanical
28
2.4 Microwave Radiometer Instruments and Satellite Missions
or electrical scanning techniques The scanning is performed at a constant incidence
angle to obtain a wider data swath as well as for better remote sensing of earth
atmospheric system
A few satellite radiometer instruments have been flown m past and present namely,
Sensor Microwave Imager (SSM/I) on DMSP satellite, Scanning Multi-channel Mi­
crowave Radiometer (SMMR) on Seasat and Nimbus satellites, TRMM Microwave
Imager (TMI) of Tropical Rainfall Measuring Mission (TRMM) satellite, Multi­
frequency Scanning Microwave Radiometer (MSMR), on IRS-P4 satellite, Advanced
Microwave Scanning Radiometer (AMSR) of ADEOS-II satellite. Table (2 2) gives
the details of past, present and future microwave radiometers Some of the radiome­
ters used m the present thesis are described below
Multi-frequency Scanning Microwave Radiometer (MSMR)
India has launched several remote-sensing satellites under the IRS series of satelhtes
IRS-P4 satellite is the fourth satellite launched by the Indian launch vehicle PSLV
from Shriharikota, India
This satellite is also known as Oceansat-1
IRS-P4 was
launched on 26, May, 1999 and is a sun-synchronous satellite m a near polar and
near circular orbit at an altitude of 720 Km with local time of equator crossing
in descending node at 1200 Hrs This satellite carries two payloads, an Ocean Color
Monitor (OCM) and Multi-frequency Scanning Microwave Radiometer (MSMR). The
IRS P4/MSMR with a two days repetivity operated at 6 6, 10 6, 18.0, and 21 0 GHz
channels m vertical and horizontal polarizations It had data swath of 1400 km and
has been, configured with conical scan system with an incident angle of 49° at the
earth surface. Figure (2.7) depicts the observational geometry of MSMR Brightness
temperature from MSMR is available at three-spatial resolutions of 150, 75 and 50
29
2.4 Microwave Radiometer Instruments and Satellite Missions
Table 2 2 Major characteristic of the different microwave satellite sensors used
Radiometer
Incident
Center frequency (GHz)
Angle
Polarization (V,H)
Foot print (km x km)
IRS-P4 MSMR
TRMM TMI
DMSP SSM/I
AD EOS-2 AMSR
49 0°
52.8°
53 0°
55 0°
66
10 65
18 0
21.0
V,H
V,H
V,H
V,H
150x136 75x36
50x46
50x34
10 65
19.35
V,H
-
-
-
-
21.3
37.0
85 5
V,H
V
V,H
V,H
59x36
30x18
23x16 16x10
7x4
-
19 4
22.2
37.0
85 5
V,H
V
V,H
V,H
69x43
50x40 37x29
15x13
69
10 7
18 7
23 8
36 5
89.0
V,H
V,H
V,H
V,H
V,H
V,H
76x44
49x28’
28x16
31x18 14x8
6x4
km’s termed as Grid 1, 2 and 3 respectively. In Grid-1, T bs for all the channels,
i.e. 6 6, 10 65, 18 and 21 GHz in horizontal and vertical polarization are available
In Grid-2, Tg at 10 65, 18 and 21 GHz m both polarization and m Grid-3,
Tb
at
only 18 and 21 GHz m both polarizations are available The operational geophysical
products available from MSMR are ocean surface wmd speed, sea surface temperature,
integrated water vapour content and cloud liquid water content over global ocean
(Gohil et a l , 2000)
30
2.4 Microwave Radiometer Instruments and Satellite Missions
MSMR Viewing Geometry
Figure 2.7. MSMR Viewing Geometry
TEMM Microwave Imager (TMI)
TRMM satellite orbits m a circular orbit at an altitude of about 350 Km, with an
inclination of 35 , and local incidence angle of 52 8 at the earth surface The TMI
instrument measures brightness temperature at 10 7, 19 4, 21.3, 37 0 and 85 5 GHz m
dual polarization except 21 3 GHz, with a swath of 760 km Kummerow et al (1998)
DMSP: Special Sensor Microwave / Imager (SSM/I)
The SSM/I is a seven-channel, linearly polarized passive microwave radiometer aboard
Defense Meteorological Satellite Program (DMSP) satellites
The DMSP design,
builds, launches, and maintains these near-polar orbitmg, sun-synchronous satellites
to monitor the meteorological, oceanographic, and solar-terrestrial physics environ­
ments DMSP satellite orbits at an altitude of approximately 830 km above the earth
surface with SSM/I havmg a swath of 1400 km (Hollmger et a l , 1987) SSM/I data
31
2.5 Retrieval Methods
are used to derive geophysical parameters such as, ocean surface wind speed, area
covered by ice, age of ice, ice edge, precipitation, cloud liquid water, integrated water
vapor, soil moisture, land surface temperature, area covered by snow, snow water
equivalent, and sea surface temperature
2.5
R etrieval M ethods
The radiation emitted by the earth atmospheric system as measured by satellite
sensor depends upon various geophysical variables which can be modeled by radiative
transfer theory RT models are used to simulate radiation using the known state of
the ocean and atmosphere while it is required to derive the physical state of the ocean
and atmosphere using the limited observations from spaceborne radiometer, and this
process is known as inversion or retrieval of geophysical parameters.
In general retrieval methods involve two steps, first, forward models, which emu­
late satellite measurements from given geophysical parameters, and second retrieval
algorithms, which transform satellite measurements into geophysical parameters. A
standard retrieval (conventional methods) approach consists of solving an inverse (or
retrieval) problem for deriving a Transfer Function (TF), / ; which relates a geo­
physical parameter of interest (e g
WVC, CLW, SSW, SST, etc.) with satellite
measurement, brightness temperatures as given below
GP = flTB)
(2 5 1)
where both GP and TB are vectors. Normally, there are two approaches used
for retrieval of geophysical parameters from measured brightness temperature known
32
2.5 Retrieval Methods
as physical and statistical retrieval techniques which can be either linear or nonlin­
ear The hybrid retrieval is the combination of these two. Other useful techniques
developed by artificial intelligent community, like Bayesian network, neural network,
genetic algorithm etc representmg a class of flexible nonlinear regression and dis­
criminating models, are also widely used m remote sensing community
2.5.1
Statistical methods
One of the simple statistical methods is known as empirical retrieval approach m
which GPs coincident and collocated with satellite observations are used to derive a
relationship between them through a suitable statistical technique. However, when
simulated GPs and corresponding Tg simulated through RT model are used to es­
tablish a relationship between them is known as statistical algorithm
The most
well-known purely statistical algorithms are those of Ahshouse et a!. (1990a,b) for
water vapour, which is a linear combination of 19, 22 and 37 GHz channels, with
addition of a quadratic term in Tg22 to better account for the non-lmeanty, and for
CLW, based on ground based microwave radiometer CLW estimates (linear function
of brightness temperatures) Application of inversion techniques to actual measure­
ments generally requires some adjustments due to inherent errors m radiative transfer
model and radiometer calibration Among this category, Wilheit and Chang (1980)
developed retrieval algorithms for WVC and CLW using a multiple linear regression
approach They showed that the nonlinearity of the transmittance (exponential form)
can be reduced by using logarithmic functions of brightness temperatures. Similar
functions of the brightness temperatures have been used by Karstens et al. (1994) and
Gerard and L. (1998), using ECMWF predicted fields for building the learning data­
base. Based on radiosonde profiles with a cloud model, other statistical algorithms
33
2.5 Retrieval Methods
(using linear and nonlinear predictors) were proposed by Bauer and Schlussel (1993).
2.5.2
Physical / Physio-statistical methods
In physico-statistical method, the initial estimate of the retrieved variable is adjusted
iteratively using a RT model The iterative process, starting from a first guess, is con­
tinued until the difference between simulated and observed brightness temperatures
are lower than a given threshold Radiative transfer models are generally simplified.
A physically-based forward model can be written as-
Tb = F(GP)
(2.5.2)
where F is a simplified Forward Model (FM) which relates a vector GP to a
Tg. Simplified forward models are derived using radiative transfer model and known
geophysical parameters Thus, the forward problem eqn (2 5 2) is a well-posed con­
trasting to the inverse problem eqn
(2 5.1)
In a few specific cases, the forward
model can be inverted explicitly m the complete parameter space and the solution
eqn. (2.5.1) simultaneously provides the complete vector of geophysical parameters.
In real case, the relationship betv/een Tg and GP can not be mverted explicitly. Hence
numerical iterative inversion is applied based on the linearization of forward model
(eqn. 2.5.2). Based on linearization of forward model, numerical iterative inversion
is applied by Wentz (1992, 1997). When the errors (model, observations) are taken
into account, these methods are called optimal estimation algorithms Many retrieval
techniques make use of such a physical method Schluessel and Emery (1990) derived
WVC through a quasi physical retrieval scheme for non scattering situations based
on RT and regression algorithm. Greenwald et al. (1993) used two channels 19 and 22
for WVC, 19 and 37 for CLW. They simplified the radiative transfer model by taking
34
2.5 Retrieval M ethods
chmatologic surface emissivity values This method is similar to that developed by
Wentz (1983), Wentz et al (1986) in which, the wind speed is first estimated with
an empirical model, and then the atmospheric transmittance is calculated at 22 and
37 GHz, in order to separate the contributions of WVC and CLW (Liu and Curry,
1993). Phahppou (1996) used the assimilation scheme of a meteorological model to
retrieve the WVC and CLW. The first guess is the model analysis at the collocated
point, and a 1-D Variational Assimilation procedure is applied to all SSM/I channels.
Other methods are based on a selection of channels or combinations of channels to
reduce the model uncertainties and optimize the retrieval. Petty and Katsaros (1992)
established a model of the depolarization effect of droplets on the sea surface signal m
window channels (37 / 85 GHz). Other methods require additional information like
atmospheric and surface temperature, cloud top height etc. Liu and Curry (1993)
presented physical retrieval schemes based on RT calculation for different atmospheric
conditions like, clear sky, liquid water cloud, cloud with ice, raining water clouds, and
raining cloud with ice It was shown that cloud sensitive schemes especially improve
the CLW retrieval as compared to the previous algorithm (Ahshouse et al., 1990a,
Petty, 1990; Greenwald et a l , 1993) The nonlinear physical retrieval algorithm is
based on the formulation initiated by Rodgers (1976) and further developed by Eyre
(1989a,b), Eyre et al (1993) the algorithm employs a Newtoman iterative method
that finds the maximum probability solution to the nonlinear mversion of the radiative
transfer equation
35
2.5 Retrieval Methods
2.5.3
Neural Network method
Neural networks offer interesting possibilities for solving problems involving transfer
functions (Rumelhart et a l , 1986, Lipmann, 1987) Firstly, NN are adaptive provid­
ing a flexible and easy way of modeling a large variety of physical phenomena Here,
adaptive means the ability of the method to process a large number of data or new
additional relevant variables can be easily incorporated Secondly, even if the learning
phase of the network is computational expensive to finalize the correct NN architec­
ture, the operational phase is very efficient. This phase requires few calculations and
can be carried out on small computers.
Neural networks are capable of modeling a large variety of complex inverse non­
linear processes The NN is a generalization of the well-known multilayer perceptron
(MLP) model (Rumelhart et a l , 1986) and is widely used for transfer function ap­
proximation (Thiria et a l , 1993) Multidimensional transfer function with NN can
be developed by using specific architectures The number of layers and the number
of neurons m each layer can be chosen considering the complexity of the function
Basic theory of NN used in the present work is discussed m detail m Chapter (3).
In principle, NN can be used to emulate FMs (2 5 2) and TFs (2 5 1) because FM
and TF both are continuous mappings. Instead of using a statistical approach for
developing FM and TF, NN can be incorporated with better accuracies due to their
capability of modeling inherent nonlinearity as well as handling multiple parameters
individually or simultaneously (Kerhrzm and Refregier, 1995; Attali and Pages, 1997;
Chen and Chen, 1995a,b, Cybenko, 1989, Funahashi, 1989, Hormk, 1991)
NN is a fast, accurate and robust tool for nonlinear (continuous) mappings and
can be effectively used for modeling multi-parameter retrieval algorithms
For re­
trieval algorithms, NN can be used m several different setups In physical retrieval
36
2.5 Retrieval Methods
algorithms, emulating the complex and slow RT models and its Jacobian, NN can be
used to speed up local inversion process In many cases NN can be used for global
inversion to explicitly invert a FM In such cases, after inversion, the NN provides an
exphcit retrieval algorithm (or TF)
All these previous studies, also discussed m chapter (1) form a strong basis for
the present attempt to evaluate the performance of neural network approach for
MSMR radiometric channel simulations and retrievals Various architecture, model
development and its implementation for specific apphcations have been discussed in
subsequent Chapters
37
C hapter 3
D evelopm ent of A rtificial N eu ral
N etw ork for G eophysical
P a ra m e te r R etriev al
Chapter 3
D evelopm ent o f A rtificial N eural
N etw ork for G eophysical
Param eter R etrieval
3.1
Introduction
This chapter deals with the theory of Artificial Neural Networks (ANNs) specifically
the Back Propagation Neural Network (BPNN), an overview of neural network tech­
niques applied to microwave radiometers and related retrieval algorithm
3.2
Artificial Neural Network
ANNs exploit an analogy to the human brain The idea behind ANN was to transfer
the idea of parallel distributed processing, as found in the brain, to the computer
m order to take advantage of the processing features of the bram ANNs have been
studied almost from the beginning of the computer era ANNs are mostly referred
to as Neural Networks (NNs) The field of artificial neural network has been made
tremendous progress m the past 20 years m terms of theory, algorithms and applica­
tions Notably, the majority of real world neural network applications have involved
38
3.2 Artificial Neural Network
the solution of difficult statistical signal processing problems Compared to the con­
ventional signal processing algorithms that are mainly based on linear models, ANNs
offer attractive training algorithms
The availability of such powerful modeling tools has motivated numerous research
efforts to explore the new signal processing applications The nonlinear nature, ability
of learning from their environment m supervised and/or unsupervised ways, as well as
the universal approximation property make them highly suitable for solving difficult
signal processing problems, like the retrieval and time series prediction
Prom the
retrieval respective, it is imperative to develop a proper understanding of basic neural
network structure and its effectiveness m retrieval algorithms and applications
A
challenge m surveying the field of neural network paradigms is to distinguish those
neural network structures that have been successfully applied to solve real world
problems In addition, it is also important to assess the impact of neural networks on
the performances, robustness of the systems and develop methodologies for integrating
neural network with other retrieval algorithms.
Estimating high quality geophysical parameters (information about physical, chem­
ical, and biological properties of the oceans, atmosphere, and land surface) from re­
mote (satellite, aircraft, etc ) measurements is very important problem m geosciences
such as meteorology, oceanography, climatology and environmental modeling. The
quality of geophysical parameters derived from these measurements varies signifi­
cantly on the strength and uniqueness of the signal from the geophysical processes
and mathematical methods applied to extract these parameters, i.e. to solve forward
and inverse remote sensing problems
The development of artificial neural networks started from the efforts to simulate
biological nervous systems by combining many simple computing elements neurons
39
3.2 Artificial Neural Network
mto a highly interconnected system. A computer can be used to simulate a biological
neural network This computer simulated neural network is called an artificial neural
network A biological neuron cell, as shown in figure (3.1a) is the basic building block
of a human brain. Figure (3 lb} is a computer model of neuron
fhrtprt
A
(a) An human bram-neuron
(b) A artificial neuron
Figure 3 T Neurons
A large NN might have hundreds or thousands of processing units whereas a mam­
malian brain has billions of neurons with a corresponding increase m magnitude of
their overall interaction and emergent behavior In practical terms, artificial neural
networks are essentially very simple computer programs that can automatically find
non-linear relationships/patterns m data without any pre-defined model form or do­
main knowledge They consist of an often large number of neurons, i e simple linear
or nonlinear computing elements, interconnected m often complex ways and often
organized mto layers Artificial neural networks are mainly used as:
• models of biological nervous systems and intelligence
• real-time adaptive signal processors or controllers implemented in hardware for
applications such as robots
40
3.2 Artificial Neural Network
• data analytic methods
The mam feature of neural network is that they can learn the internal character­
istic of a system by analyzing the datasets The network then can reproduce or even
predict new output of the system These networks are good when the data is large,
noisy, and has unknown relationships The learning m a bram is based on synap­
tic modification of the strength of the connection between the neurons. So neural
network resembles brain in two ways, firstly, the knowledge acquired through the
learning process and secondly, the strength of connections between neurons, termed
as synaptic weight used to store the knowledge
3.2.1
Application of Neural Network
In general the mam application of NN is data analysis. A NN requires dataset to be
trained and provides solutions with better accuracies by learning the internal charac­
teristic of the dataset. NN are mainly used m following applications, 1) Classification,
2) Forecasting and 3) Modeling and others
Classification: These applications categorize various characteristics of the dataset
Some of the various applications are pattern recognition/pattern classification, speech
recognition, natural language processing, expert system, data-minmg application, im­
age processing, and Remote Sensing.
Forecasting: These applications identify the time dependent behavior of the
dataset to be used for predicting a certain variable m future, like function fitting and
regression, function approximation, forecastmg, and scheduling.
M odeling: Modehng apphcation focuses on the modeling on certain intelligent
processes like emotions, dialog, robot-control etc
41
3.2 Artificial Neural Network
3.2.2
Advantages and Limitations of Neural Network
Like others techniques, NN also has certain advantages and limitations as mentioned
below.
A dvantages: NN is in general effective for dataset with missing point and noise,
does not require prior information of approximate model or algorithms, suitable for
systems with non-hnear behavior, applicable to large dataset with numerous variables
or parameters and has parallel distributed processing capability with multiple mputs
and outputs There are many practical advantages like computational speed, accuracy
and robustness
In NN, all continuous bounded functions can be learned and the
shape and location of the bounded functions can be adjusted. NNs are capable of self
learmng and generalization
L im itations: NNs are computationally expensive during training phase, but are
efficient m recall phase Despite many advantages NN is considered as a ’’black box”
because of the inability to explain comprehensively how a trained neural network
reaches its output NN does not provide information about the nature of the relar
tionship between predictor and target variables, except a predicted value with some
statistics about goodness of fit
Recently Zwaag et al (2002) described domain-
specific NN analysis methods that utilize domain-specific base functions which are
easy to interpret by the user and can even be used to optimize NN systems This NN
analysis uses two-dimensional vectors applied to some well-known image filters, en­
abling comparison of conventional edge detectors known from literature and the neural
network edge detectors NN sometimes also suffer with the problem of over/under
training which can be solved experimentally
42
3.2 Artificial Neural Network
—
in p u t 1
—
in p u t 2
—
in p u t 3
—
in p u t 4
Figure 3.2: A feed-forward neural network
3.2.3
Building Blocks of Neural Network
Every neural network consists of a number of a building block. In general the network
consists of neurons and connections Neurons are organized m layers. The number
of neurons, number of connections, type of connections, number and type of layers
are all dependants on the classification of NN Figure (3 2) shows one example of a
possible neural network structure
Neuron
A neuron (also called single perceptron)is the smallest unit of NN Figure (3 3) shows
the detail structure of a neuron. The neuron is a simple model of a real brain neuron
The neuron consists of various parts like input and output connections, synaptic
weight, summing junction, activation function and a threshold value. The threshold
value can be modeled as an offset input
A simple perceptron computes a hnear combmation of the inputs (possibly an
intercept or bias term) called the net input Then an activation function is apphed
to the net input to produce the output An activation function, maps any real input
43
3.2 Artificial Neural Network
into a usually bounded range, often 0 to 1 or -1 to 1
Figure 3 3 Single neuron
/
Activation function
The activation or transfer function decides the manner m which weighted inputs
are connected to the output of a neuron. The activation function is necessary to
introduce the non-hnearity m the network This non-linearity makes it possible to
learn non-hnear functions. For hidden layers m Multi Layer Perceptrons (MLP),
usually sigmoidal functions, are preferred yielding best results m most cases. These
functions are easier to change than the threshold functions because threshold functions
usually do not change the output when the weights changes very little. A perceptron
can have one or more outputs Each output has a separate set of weights Usually
the same activation function is used for each output, although it is possible to use
different activation functions Perceptrons are most often trained by least squares,
i e , by attempting to minimize
where the summation is over all outputs and
44
3.2 Artificial Neural Network
over the traimng dataset. A perceptron with a linear activation function is thus a
linear regression model. A perceptron with a threshold activation function is a linear
discriminant function (Hand, 1981, McLachlan, 1992, Weiss and Kukkowski, 1991),
and a perceptron with a sigmoid activation function is thus a non-linear regression
model
Layers
Neurons are organized m layers like, one input layer, one or more hidden layers and one
output layer The input layer consists of neurons serving as input to the network The
neurons in the mput layer transform the input data into information to be processed
by NN The hidden layers consist of neurons and are placed between the input and
output layers One or more layers can be designed with varying number of neurons
m each layer. The hidden layers transform the information from input layer to the
subsequent hidden layer(s) or to the output layer and the output layer consists of
neurons, which transform the information from the last hidden layer to the output of
the network.
Connections
Connection between various layers m a NN can be of four types. One lateral connec­
tions, connects neurons mside a layer, recurrent connections, connects a neuron to
itself, and feed-forward connection, is unidirectional connections between neurons of
consecutive layers
NNs can be modeled into three broad categories, namely, feed-forward, feed-back,
and cellular, which are further classified based on 1) pattern of connections between
the neurons (also called its architecture or model), 2) activation function used m the
45
3.2 Artificial Neural Network
neurons, and 3) learning algorithm (criteria for determining weights)
All NN paradigms involve a training phase and a testing phase. In the training
(learning) phase (usually offline) the NN is trained until it has learned its tasks while
the testmg (recall) phase is used to realize the task Some NN paradigms are named
after their proposer such as Hopfield, Kohonen, etc
Most NNs axe named after
their learning algorithm such as Backpropagation, Competitive learning, Counter
propagation, conjugate gradient, ART, etc and some are named after their model
such as BAM basically a particular NN. A NN classification is shown in Table (3.1)
Table 3 1' Neural network model classification
NN Models
Feed forward
Supervised
Feed back
Least Mean Square
Recurrent backpropagation
Backpropagation
Reinforcement learmng
Cascaded Correlation
Unsupervised
Self-organizing maps (KNN)
Adaptive resonance theory
Competitive learning
Fuzzy ART
Counter propagation
Boltzmann learning
Hopfield network
3.2.4
Learning Law
There are a variety of commonly used learning laws. These laws are mathematical
algorithms used to update the connection weights The process of finding the best set
of weights for the NN is referred to as traimng or learning The approach used most,
46
3.2 Artificial Neural Network
to estimate the weights is backpropagation Each time the network cycles through
the training data', it produces a predicted value for the target variable This value
is compared to the actual value of the target variable and an error is computed for
each observation The errors are fed-back (’’feed back”) through the network and new
weights are computed to reduce the overall error. Despite the neural network termi­
nology, the framing process is actually a statistical optimization procedure Typically,
the procedure minimizes the sum of squared residuals. The human brain basically
learns from experience NNs are sometimes called machine learning algorithms, be­
cause changing of its connection weights (training) causes the network to learn the
solution to a specific problem. The strength of connection between the neurons is
stored as a weight for that specific connection The system learns new knowledge
by adjusting these connection weights The learning abihty of a neural network is
determined by its architecture and by the algorithmic method chosen for training.
Two types of learning prevailed m NNs, supervised an unsupervised learning.
Supervised learning: it is performed using the teacher (target) signals Differ­
ence between the NN output and the target treated as error signal, is then minimized
through contmuous adaptation of the weights to solve the problem through a learning
algorithm When the error found with in the acceptable limits, the NN is assumed to
have learned the task and the training process is stopped In the thesis, for retrieval
purpose the dataset used to tram the NN is the brightness temperature simulated
through radiative transfer model.
U nsupervised learning: In this case, the neurons must find a way to organize
themselves without the help of target signal. Most of these learning laws are a sort of
variation of the best known and oldest learning law, the Hebb’s Rule. Some of these
learning laws are Hebb’s Rule, Hopfield Law, Delta Rule, and Kohonen’s.
47
3.3 The Back-Propagation Neural Network
3.3
The Back-Propagation Neural Network
The back-propagation algorithm is the learning strategy for NNs (Rumelhart et al.,
1986). The back-propagation learning law is a very popular for inversion related
problems, and are often referred to as the Back-Propagation Neural Network (BPNN)
It has also been used in this research In BPNN, the weights are modified to reduce
the difference between the target value and the output of NN. This rule changes the
connection weights m a manner chat minimizes the mean squared error of the network
The error is back propagated into previous layers one layer at a time Thus backpropagation algorithm cycles through two distinct passes, a forward pass followed by
backward pass through these layers of the network. The process of back propagating
the network errors continues unril the first layer is reached The algorithm alternates
between these passes several times for a given input/output vector and modifies the
NN weights till it reaches the minimum desired error, i e the minimum error limit
This process consumed several iterations (this is also termed as local iteration as it
is use in minimizing of single vector) This iterations is part of the global iteration
counter set as maximum number of iteration as the input parameters of NN model.
Iteration counter continues as it scans the training data one by one till the NN scan
the entire training data once, at this point NN is said to complete one epoch NN
iterates through several epochs till it reaches the minimum desire RMS error for the
parameters under investigation
Data can be fed to NN for training in two ways, pattern training mode and batch
training mode In the pattern training mode each input/output vector data point is
fed to NN and error is minimized iteratively through back propagation for this vector
(weight is updated m each iteration) and similarly the entire dataset is processed.
In batch training mode the weight is update only after all the input/output vectors
48
3.3 T he B ack -P rop agation N eural N etw ork
are processed after each epoch. In both the training mode typically the whole set
of training data is scanned several times before the networks learns to make good
learning.
Pattern training algorithm is faster and more effective than the batch
procedure especially for large database (Cichocki and Unbehaue, 1993).
3.3.1
L earning A lg o rith m s for S in gle P e r ce p tro n
To explain the backpropagation algorithm in its basic form, let us first consider the
learning of a single neuron as shown in figure (3.4). The hyperbolic tangent function
as given by equation (3.3.1) is used as an activation function.
5
/
+1
/
X,
*
:
\
: W»
■
"\
\
UJ
(Output)
'--.-I
*2
.
w
-
x.
........
-
s
: v ’f V
'
'
■
'
. ....
^Adaptive
.Algorithm
--------- / I
vl
e
---------(' X
. , (Target Output)
-
il
\
;
i,
J
Figure 3.4: Single perceptron learning
1 _ e -2Uj
Vj
where
iij —
Wjix i
= V\u,i) = tan(uj) =
+
®j
with
Wj0
=
©j
and
(3.3.1)
xq
= +1.
and
&
\ ( d , ~ V j? =
49
(3.3.2)
3.3 The Back-Propagation Neural Network
Where d3 is target value and y3 is NN output The aim of learning is to minimize
the square error (also known as cost function) by modifying the weight W , and need
to determine how to increase or decrease the weights to minimize the local error
function
This can be achieved using a steepest descent gradient rule
=
dt
-V
(3 3 3)
‘dW3i
where r] is a positive learmng parameter determining the speed of convergence, r)
is an important tuning parameter that is chosen by trial and error by repeated runs
on the training data Typical values for r] are m the range 0 1 to 0 9 Lower values
give slow but steady learning while higher values give erratic learning and may lead
to an unstable network Applying chain rule of differentiation to the right hand side
of eqn (3 3 3) yields the following expression
W»
dt
de3
-ye]aw
-r/e3
de3 du3
du3 dW3l
dip(u3)
ye3.... —x,
<9u,
r}53x%
(3 3 4)
Where S3, called the learning signal or local error, is expressed as
^ = e3^ '(u 3) =
(3 3 5)
for the sigmoid activation function to be a hyperbolic tangent function given by
equation (3 31), then the derivation ip'{u3) is given by
50
3.3 The Back-Propagation Neural Network
au3
= i1 - ( ta n h ^ ) 2] = (1 - y2)
(3 3 6)
equation(3 3 4) can be written m the form
~^r = VZ] (1 - y , ) ^ = y f y (1 - z/2)xt
(3 3 7)
with % > 0 that weight update stabilized if y3 approaches -1 or +1 since the
derivative dy3/du3, equals to (1 —y2), reaches its maximum for y3 = 0 and its minimal
for ±1
The weights are usually changed incrementally and the neuron gradually conver­
gence’s to a set of weight, which solves the specific problem. Change m weight W3%
cab be determined by
AWn (k) = A F jt = Wn [{k + l)r] - W3t(kr) = r}3S3x3
(3 3 8)
The weights are usually changed incrementally and the neuron gradually converges
to a set of weight, yielding the solution to a specific problem Change m weight W3X
can be determined by
W3l(k + 1) = Wn {k) = A l ^ ( t )
(3 3 9)
The above describes the approach of adaptive learning of the weight which can be
extended to multi-layer perception (MLP) as shown m figure (3 5)
51
3.3 The Back-Propagation Neural Network
Figure 3 5 NN architecture for the standard backprogation algorithm of a three-layer
perceptron (Cichocki et al 1993)
3.3.2
Function Minimization Evaluation
In pattern training mode the RMS error of the entire dataset as compared to the
target values is momtored and the training process is iterated till the desired RMS
error is achieved The RMS error of the dataset is defined by
Hilf S E rro r —
v~' ((h —tz)2
1
.,,i/
n
(3 3 10)
Where, n is the number of training cases Along with the RMS error, bias between
52
3.4 Development of NN Algorithm for Geophysical Parameter Retrieval
the target and predicted values given by eqn. (3 3 13) is also monitored and the
training process is resumed with bias correction incorporated at appropriate stage.
(3 3 11)
3.4
D evelopm ent of N N A lgorithm for G eophysi­
cal Param eter Retrieval
As discussed m Chapter 2, the satellite observations are related to various geophysical
parameters through the process of radiative transfer hence a relationship exists be­
tween geophysical parameter and satellite observed radiation which is required to be
determined using various inversion techniques. Among various inversion techniques,
the ANN technique has been chosen for developing retrieval algorithm for geophysi­
cal parameters For selecting the appropriate retrieval model it is necessary to have
information about the dominant dependency of satellite observations on various geo­
physical parameters. This is explained by a specific example of sensitivity of SMMR
channels on various geophysical parameters as given m figure (3.6) (Wilheit, 1979a)
As seen in figure (3 6), the sensitivities of sea surface salinity and temperature are
maximum around 1 and 6 GHz, respectively. While sea surface wind has maximum
sensitivity beyond 10 GHz The sensitivity of atmospheric water vapour is maximum
around 22 GHz due to water absorption line at 22.235 GHz The sensitivity of cloud
liquid water increases with frequency
More over, the radiation observed by the
radiometer at a given frequency/channel is also affected by parameters other than
the parameter having maximum sensitivity implying that multiple channels should
be used for better retrieval
53
3.4 Development of NN Algorithm for Geophysical Parameter Retrieval
Apart from this the satellite data is also available at grids with varying spatial
resolutions which also contains different channels for different resolutions, for example,
for IRS-P4 MSMR for 150 km resolution grid all the channels (6, 10, 18 and 21 GHz)
are available while for 50 km resolution grid only 18 and 21 GHz channels are available
which restrict the retrieval of different geophysical parameters under different grid
resolutions like only water vapour and cloud liquid water is possible for 50 km grid
while sea surface temperature, sea surface wmd speed, water vapour and cloud liquid
water are possible for 150 km grid
The sensitivity analysis suggests that different combinations of channels are suit­
able for retrieving different geophysical parameters. Hence retrieval models for various
parameters may have different channels as predictors In view of availability of dif­
ferent channels and varying sensitivity suitable NN models can be designed Due to
Figure 3 6- Sensitivity of
Tb
to geophysical parameters, the arrow indicates the
SMMR frequencies, Wilheit (1979a)
54
3.4 Development of NN Algorithm for Geophysical Parameter Retrieval
the special capability of NN of mapping multiple inputs to single or multiple outputs
it is also possible to develop single or multi-parameter retrieval models. As reported
by (Krasnopolsky et a l , 1999), single-parameter algorithms may have additional sys­
tematic bias and unknown component of RMS errors e g. SSM/I algorithms for wmd
speed (Goodberlet et a l , 1989), for water vapor (Alishouse et a l, 1990a), and for cloud
liquid water (Petty, 1993) The obvious way to improve single-parameter retrievals
is to include the other parameters m the retrieval process. A typical example of NN
models using single and multi-parameter developed for IRS-P4 MSMR is depicted in
Table (3 2). These NN based retrieval models have been classified according to the
single or multi-parameter retrievals termed as class-I and class-II models respectively.
Figures (3 7, 3 8) depict various models developed for IRS-P4 MSMR
Table 3 2. Different NN Models
Various Neural Network Models
Class-I (Single Output)
M o d e ls
In p u ts
O u tp u ts
4Tb - 1G P
I' b ISV, I' b LSII, T b 21V, T b 21H
WVC / CLW
T bo&v i T B q&h , T b w v , T B w h , T b i s v ,T b w h
SSW / SST
T b 06V, T boSH, T b i OV,T b i OH,TB 1BVi T b \SH, T B21V, T B21H
WVC / CLW / SSW / SST
6T b
- 1G P
8T b - 1G P
Class-II (Simultaneous Output)
4T b - 2G P
T b ISV, T b i BH, T B21V, T B21H
WVC, CLW
6T b
- 2G P
T bo bv , T b o sh , T b io v , T B io h , T b w v , T b w h
SSW, SST
8T b
- 2G P
T bo bv , T b o sh , T b io v , T b io h , T b i b v , T b w h , ?B2i v , T B21h
WVC, CLW / SSW, SST
8T b - 4.GP
T b o s v , T bo bh , T b i o v , T b io h , T b i s v , T b i s h ,T b 2i v , T B2i h
WVC, CLW, SSW, SST
In general, the radiance measurement involves responses of detector system and
electronics of sensor system as well as scanning systems (involving integration times).
Therefore, the observations are laced with ’’measurement errors” . These measurement
55
3.4 D e v e lo p m e n t o f N N A lg o r ith m fo r G e o p h y s ic a l P a r a m e t e r R e tr ie v a l
Input Layer
Hidden Layers
Output Layer
Input
»
T b 18V
T b 18U
T b 21V
T , 2111
(a) (47s-lG P)
Input Layer
Hidden Layers
Output Layer
(b) (8Tb -1GP)
Figure 3.7: Class-I neural network model configurations
56
3.4 D e v elo p m e n t o f N N A lg o rith m for G eo p h y sical P a ra m e te r R e trie v a l
In p u t L a y e r
H id d e n Lay ers
O u tp u t L a y e r
(a) (47V2GP)
In p u t L a y e r
H id d e n L a y e rs
O u tp u t L a y e r
T r 06H
(b) (8Tb -4GP)
Figure 3.8: Class-II neural network model configurations
57
3.4 D evelopm ent of N N A lgorithm for G eophysical P a ra m e te r R etrieval
errors specify the accuracy of possible retrievals as well as utility of particular spectral
channel(s) for a particular geophysical parameter
Hence, all retrieval schemes are
subjected to noise (A T B ) considerations statistically. The solution to optimization
problem performed better when sensor specific anticipated noise is introduced. For
each sensors channels, Gaussian noise with certain standard deviation m their respec­
tive channels and with zero mean is added by means of a random noise generator
While developing retrieval model such noise are incorporated m their respective chan­
nels and this database is actually used m model development. This added noise will
ensure the robustness of the retrieval model to perform in the real world scenario
For MSMR NN model development the simulated
Tb
have been individually per­
turbed with Gaussian noise for different channels with certain standard deviation
The noise figures are preliminary m nature, which has been evaluated from the lim­
ited MSMR data based on simulated data using the minimized approach used m
MSMR operational algorithm (Gohil et al., 2000).
Radiative transfer based simulations have been used for NN modeling, and statis­
tical analysis for selecting the final NN configuration and the results of this analysis
are presented m Chapter 4
58
C hapter 4
N eu ral N etw ork A p p ro ach and
R etriev al of G eophysical
P a ra m e te rs from S atellite D a ta
C hapter 4
N eural N etw ork A pproach and
R etrieval o f G eophysical
Param eters from Satellite D ata
4.1
Introduction
This Chapter deals with the radiative transfer based simulations of microwave radia­
tion and the use of NN approach m retrievals For simulating the radiation database,
the required surface and atmospheric parameters have also been simulated using ther­
modynamical and other relevant relationships. The radiation data has been simulated
at radiometer sensor specification, observational geometry and anticipated noise sce­
nario. The dataset of simulated radiation and corresponding geophysical parameter
has been used m developing NN based retrieval approaches for IRS-P4/MSMR and
TRMM/TMI sensors for evaluating the retrieval performance of various NN models.
Figure (4 1) depicts the overall flow of algorithm development, testing, implemen­
tation to satellite data and initial in orbit data quahty evaluation
Moreover, a
preliminary comparison with standard products has also been attempted
59
4.1 Introduction
Figure 4 1 Block diagram of retrieval algorithm development, testing and data qual­
ity evolution
60
4.2 Simulation of Brightness Temperature over Oceans
4.2
Simulation of Brightness Temperature over Oceans
4.2.1
Simulation of Geophysical Parameters
In order to simulate the brightness temperature, vertical profiles of atmospheric tem­
perature, pressure, humidity and cloud liquid water density are required The tem­
perature profile is simulated using region (like tropical, mid-latitude and polar) de­
pendent average temperature lapse rate. The temperature lapse rate of the standard
atmosphere given by Silby et al. (1978) has been used For this purpose, sea surface
temperature values in the appropriate range for different regions obtained through
climate data have been considered. Using appropriate values at the surface, the pres­
sure profiles have been simulated using standard hydrostatic equation Water vapour
density profile has been simulaued by making use of profiles of simulated tempera­
ture and relative humidity m which relative humidity has been varied linearly from
surface to top of the atmosphere SST dependent surface relative humidity variation
has been considered in which for low SST values, the surface relative humidity vari­
ations are allowed in narrow range while for higher SST, wide range of humidity has
been considered This is a general trend seen m the global observations Clouds of
different thickness varying with altitude have been introduced. In this model, cloud
liquid water density diminishes linearly from freezing level to the boundaries of cloud
as given by Pans (1971) Moreover, the relative humidity is saturated at the cloud
levels Earning clouds have not been considered m the simulations, due to reason
that under precipitating conditions, retrieval of SST and SSW are erroneous The
total atmospheric water vapour and cloud hqmd water contents have been estimated
by integrating the density over the whole atmosphere. The surface wind speed val­
ues have been chosen to vary with Rayleigh distribution with the peak at 7.4 m/s
61
4.2 Simulation of Brightness Temperature over Oceans
Figure 4 2 Frequency distribution of simulated GP’s m training dataset (MJ3-I,
29500 points), (a) WVC g/cm2. (b) CLW g/cm2, (c) SSW m /s , (d) SST K
as given by Wentz et al (1984). The typical discrete values of environmental para­
meters have been used for simulating brightness temperature. Typical distributions
of water vapour, cloud liquid water, wind speed and surface temperature are shown
m figure (4 2) The distribution of these environmental parameters is a part of the
global simulated dataset, same distribution patterns are also seen m global dataset
and this distribution m general is also observed m climatological data The bright­
ness temperature thus simulated from the known atmospheric and surface geophysical
variables are used to establish suitable statistical or NN relationship between them
for the retrieval purpose In order to develop MSMR global retrieval model, a global
database of surface and atmospheric conditions and corresponding simulated radia­
tion are required, where as, m case of TMI a tropical region database is required. For
MSMR a general dataset have been generated for three regions viz. tropical (from 0
to 34° N/S), mid-latitude (from 30° N/S to 64° N/S) and polar (from 60° N/S to 90°
N/S). The regional and global statistics of simulated geophysical dataset are given m
Tables (4 1 and 4.2) respectively For TMI a dataset has been generated for tropical
region between 40° N and 40° S, discussed more m section (4 7)
62
4.2 Simulation of Brightness Temperature over Oceans
Table 4.1: Regional statistics of simulated geophysical parameters
Region
Parameter
Mm
SST (K)
281
SSW (ra/s)
00
Tropics NP=37560
WVC (g/cm2) 03
CLW (g/cm2) 00
SST (K)
273
SSW (m/s)
0.0
Mid-Lat NP=75000
WVC (g/cm2) 00
CLW (g/cm2) 0.0
s s t ( in
273
SSW (m/s)
0.0
Polar NP=49650
WVC (g/cm2) 0 08
CLW (g/cm2) 00
NP is number of points
Max.
305
24
818
0 132
303
24
7.70
0 09
291
24
3.14
0 09
Mean
294 7
8.67
3.32
0.013
287.2
8 67
2 15
0.009
282 1
8 67
119
0 009
Std Dev
716
5 31
181
0 028
8 64
5 31
162
0.020
5.49
5 31
750
0 020
Table 4 2: Global statistics of simulated geophysical parameters
Parameter
(points 162210)
WVC (g/cm2)
CLW (g/cm2)
SSW (m/s)
SST (K)
Mm
Max
Mean
0 084
00
00
273 0
8 1777
01315
24 0
305 00
2.1248 1.6565
0 0105 0 0227
8 6670 5 3084
287 36 8 7428
63
Std Dev
4.2 Simulation of Brightness Temperature over Oceans
4.2.2
Simulation of Brightness Temperature
Using the radiative transfer model as discussed m Chapter 2, the brightness tempera­
tures have been simulated using die surface and atmospheric simulated data discussed
above at sensors’ specifications of MSMR and TMI and their observational geometry
The channels of MSMR used are 6 6,10 6, 18.0, 21 0 GHz with dual polarization while
TMI channels at 10 6,19 3, 21 3 and 37 0 GHz are used The local incidence angles of
MSMR and TMI at 49° and 52 8° at the earth surface have been used, respectively.
The simulated brightness temperatures have been perturbed with respective noise
figures of MSMR and TMI in development of retrieval models.
MSMR operational retrieval algorithms are based on the statistical approach used
m generating operational product from MSMR (Gohil et a l , 2000) assume one of the
following formM
GP = Co(CSST, 0) +
J2^ ( C
s s t , ff)
f{TBl(M, R, 9))
(4 2.1)
l— l
where C sst —
GP is desired geophysical parameters viz SST, SSW,
WVC, CLW. C'sst is season and region specific monthly mean climate SST over 2° x
2° longitude-latitude grid, C q and C\ are retrieval coefficients for a parameter GP for
%
th channel, T b i is brightness temperature data of MSMR for ith channel, N is the total
number of channels used, 9 is local mcident angle of MSMR observation, M is month
of MSMR observation, R is geographical region (viz. tropical, mid-latitude or polar)
of MSMR observation and f is the function of TB The retrieval coefficients have been
derived using simulated TB through RTM and simulated environment parameters
as described earlier The SST dependent retrieval algorithms have been developed
for deriving geophysical parameters from MSMR TB data Separate coefficients are
derived for different geographical regions for WVC and CLW. Table (4.3) depicts the
64
4.2 Simulation of Brightness Temperature over Oceans
Table 4 3 Range of SST (K) used in MSMR operational algorithms
Tropics
Min Max
281
286
291
296
302
<
<
<
<
<
286
291
296
302
305
281 305
Mid-latitude
Mm Max
Split Range
273 < 277
277 < 281
281 < 286
286 < 291
291 < 297
297 < 303
SmgleRange
273 303
Polar
Min Max
273
277
281
286
<
<
<
<
277
281
286
291
273 291
Table 4 4' Theoretical RMS retrieval accuracies of geophysical parameters by MSMR
operational SR algorithm
Parameters
WVC (g/cm2)
CLW (g/crri2)
SSW (m/s)
SST (K )
Tropical
0 20
0 013
1.63
1.52
Mid-latitude
018
0 011
159
1.92
Polar
015
0 009
151
190
Average
0.17
0 011
157
178
range of SST used m these models for the three geographical regions. Retrieval was
performed separately in these geographical regions while m the overlapping regions,
average of retrieved values were used. Theoretical retrieval accuracy for WVC, CLW,
SSW and SST using Statistical Regression (SR) technique is shown in Table (4 4),
(Gohil et ah, 2000). In SR, only the high frequency channels (18V, 18H, 21V and
21H GHz) have been used to derive WVC and CLW for MSMR data for Grid-1, 2
and 3
The radiative transfer model used for simulation in this research, involves compu­
tation of atmospheric absorption given by Liebe et al (1992) at MSMR frequencies
and incidence angle for known atmospheric physical state comprising of atmospheric
65
4.3 Preparation of Training database
constituents like dominant gases and hydrometeor It also involves the computation
of ocean surface emissivity (Holhnger, 1971, Stogryn, 1972) at a given frequency, in­
cidence angle and polarization from the known ocean surface physical conditions like
sea surface temperature, salinity, wind speed and foam cover (Wilheit, 1979b)
4.3
Preparation of Training database
For developing MSMR NN based global retrieval algorithm, a global training dataset
has been prepared from the simulated database discussed above by randomly picking
up approximately 40%, 30% and 30% of the data from tropical, mid-latitude and
polar regions respectively, amounting to about 18% (29500 points) of the total global
dataset termed as MJS-I The statistics of selected geophysical and brightness tem­
perature datasets are given m Table (4 5) The remaining data is used for testing the
NN performance and the dataset is termed as M_S-II, which consists of 82% (132710)
data points. Since satellite measurements are subjected to noise, so there is always a
slight inconsistency between simulated and measured brightness temperature Hence,
for a robust retrieval algorithm, anticipated sensor noise m respective channels have
been incorporated into these dataset, which is then used for model development and
evaluation
The noise figures of MSMR have been incorporated m MSMR NN model develop­
ment algorithm The same noise figure was also used m MSMR operational algorithm
(Gohil et a l , 2000) The datasets M.S-I and M_S-II have been individually perturbed
with Gaussian noise for different channels with standard deviation values as given m
Table (4 6) Similarly, for developing TMI NN based tropical retrieval algorithm, a
training dataset has been prepared from the simulated database by randomly picking
up approximately 30%, of the data from tropical region, amounting to about (24000
66
4.3 Preparation of Training database
Table 4 5 Statistics of simulated geophysical parameters and brightness temperatures
used m MSMR NN training (M_S-I dataset)
Parameter
(points 29500)
WVC (g/cm*)
CLW (g/cm2)
SSW (m/s)
SST (K)
Tb 06V (K)
T b 06H (K)
T b 10V (K )
Tb 10H (K )
Tb 18V (K)
T b 18H (K )
T b 21V (K)
T b 21H (K )
Mm
Max
Mean
Std. Dev
0.084
00
00
273 0
143 64
77.290
150 71
82.080
163 89
92 34
170 52
98 66
81777
10315
24 0
305 0
1701
110.93
184.24
131 98
227 72
195 91
264 97
252 32
2 5607
0.0161
8 7217
289 04
153 397
87 9210
160 405
95.9027
181.527
123.118
206.180
160.329
1.8566
0 0334
54310
9 4755
5.60
5.40
6.42
8 31
13 07
19 58
23.29
35.72
Table 4 6- Initial noise figures in MSMR channel based on limited data and simulation
Channels
Noise (K)
6V
1.0
6H
11
10V
13
10H
10
18V 18H 21V
18
1.4
16
21H
1.7
points) of the total global dataset used for training and remaining dataset is used for
testing. The noise for different TMI channel are describe m section (4 7) has been
used. The statistics of selected training dataset is gven m Table (4.7)
67
4.4 Analysis of Different Neural Network Models
Table 4.7' Statistics of simulated training dataset used m TMI NN model
Parameter (K)
(pts 24000)
WVC (s/cro2)
CLW (g/cm2)
SSW (m j s )
SST (K)
Tb IQV (K )
Tb 10H (K )
TBI W {K)
Tb 19H (K )
Tb 22V (K )
T b 37F (K)
Tb 37H (K)
4.4
Mm
Max.
0 29
0 00
0 00
281 00
158 48
78 85
173.67
92.31
179 42
200.93
12147
817
2 97
0.130
0 0149
24.00
8.66
305 00 293 43
189.47 168 31
130 56 93 66
247 74 199 30
221 91 139 25
270.53 220 40
269.75 220.85
258 03 166.31
Mean
Std. Dev
164
0 0283
5.35
6 78
5 52
7.85
14 86
23 70
20 74
14 24
27 01
Analysis of Different Neural Network Models
As mentioned m Chapter 3, various NN models were planned based on varying sen­
sitivity of channels with respect to different geophysical parameter as well as based
on the availability of the channels for a given sensor, e g MSMR Here the detailed
analysis of NN models is presented for the selection of optimum NN model for all the
geophysical parameters.
4.4.1
Selection of Optimum NN Configuration
In general neural network applications, neural networks, are trained using finite mput
samples It is important to know, firstly, how many hidden neurons are needed to
learn the input dataset, and secondly, how many different input data can be learned
by a neural network with a pre defined complexity. It is known that N arbitrary
distinct samples can be learned precisely by standard single hidden layer feedfor­
ward network with JV hidden neurons with almost any activation function found m
68
4,4 Analysis of Different Neural Network Models
applications (Huang and Babri, 1998)
However, in most applications large num­
bers of input samples often need to be dealt with Tamura and Tateishi (1997) m
their research showed that Two Hidden Layer Feedforward Networks (THLFNs) with
(JV/2) + 3 hidden neurons and sigmoid activation function can represent N distinct
samples (*„ dt) with negligibly small error However, in neural network designed by
Tamura and Tateishi (1997) the layers and neurons can still be very large m many
applications Further, Huang (2003) m his research has proved that the significant
results of THLFNs with 2^/(m + 2)N hidden neurons can learn N distinct sample
(x„dt) with negligibly small error, where m is number of output neurons
That
means, the upper bound of the required hidden neurons for a THLFN can be reduced
significantly and markedly, and thus, the complexity of the required THLFN can also
be reduced sharply For example, one such THLFN with only 340 hidden neurons
can learn 10000 distinct (®„dj) data with small error The above research has been
theoretically proved without any experiment
In the earlier study (Vasudevan et a l , 2004), retrieval of WVC and CLW from
MSMR using NN technique The best NN configuration selected was based on NN
model analysis with minimum number of channels required to retrieve WVC or CLW.
Here four Tg (18V, 18H 21V and 21H) as input parameters and WVC as one output
parameter (class-I model) was analyzed. In this study, the best NN configuration
yielding minimum RMS error found was 3-hidden layers (HL), containing 15-neurons
(N) m each layer termed as 3HL-15N or (4-15-15-15-1) NN configuration The varia­
tion of RMS error with the global iterations for such a case is shown in figure (4 3).
This experiment was carried out to test which NN configuration is to be selected
for the final retrieval As a pilot model, single-parameter NN model, with four input
vectors and one output vector were tested with varying numbers from 1 to 3 hidden
69
4.4 A n alysis o f D ifferent N eural N etw ork M od els
Figure 4.3: Evolution of RMS error minimization for the training dataset for 47 b —
IG P NN, configuration for WVC
layers and by varying 5 to 20 neurons in each layer has been considered. In figure (4.3),
it is seen that different NN configurations comprising of different number of hidden
layers with varying number of neurons lead to minimization of errors at different
number of iterations and it was found that the 3HL-15N configuration (black dash
curve) yields minimum RMS error with less number of iterations (fast convergence).
The research work by Huang (2003) has proved that NN configuration with more
than 2 hidden layers makes NN capable of learning large amount of samples with
very small output error taking care of the non-linearity in the dataset. The above
analysis provides the experimental evidence that three hidden layer can handle large
sample size with small error, as seen in figure (4.3). The same NN architecture used
for CLW also shows similar convergence. The minimization process was carried out
with normalized error limit of 0.003 with a learning rate of 0.7. Further optimization
70
4.4 Analysis of Different Neural Network Models
of error limit and learning rate is discussed later.
4.4.2
Impact of Noise on N N M odels
Single output case:
Subsequent to configuration optimization, the impact of noise m training and test
data on the performance of NN retrieval has also been studied. For this purpose, the
trammg dataset (input as well as output parameters) is perturbed with different noise
values and the EMS error of the optimized NN configuration is evaluated This study
is required anticipating noise m the actual satellite data. The study was carried out
for class-I, 3HL-15N configuration for WVC NN model from MSMR. For various noise
scenario starting with 0.0 K to 2.0 K noise m input Tb
s
and an error of 0 0 g/cm 2 to
1 0 g/cm 2 in WVC at the output m training dataset (M_S-I, 29500 points) and test
dataset (MJ3-II, 132710 points) were analyzed In general, it has been observed that
with different noise scenario the performance of NN models does not vary significantly
during traimng as well as testing phases The results are depicted m Table (4.8)
Table 4 8. Impact of Noise on class-I, 3HL-15N configuration for WVC
D a ta Type
M.S-I (29500)
MJS-II (132710)
N oise in T b ( K )
00
1.0
2.0
00
10
2.0
N oise
0.0
0 0857
0.0874
0 0867
0.0921
0 0868
0.0869
in W V C (,g / c m 2)
0.5
1.0
0 0877 0 0886
0.0877 0 0886
0 0869 0 0875
0.0924 0 0932
0 0871 0.0879
0 0871 0.0878
More over a negligible variation m performances with noise is noticed which could
be due to the use of different number of iterations in the minimization purpose
71
4.4 A nalysis of D ifferent N eu ral N etw ork M odels
Multi-parameter output case:
For testing the noise impact for simultaneous output ease, a second test is per­
formed on 8-15-15-15-4, NN configuration The performance of NN model with Eight
Tb 'b (06V, 06H, 10V, 10H, 18V, 18H 21V and 21H) as input parameters and WVC,
CLW, SSW, and SST as four simultaneous output parameters (class-II model) has
been evaluated by varying the noise in the input to account for Gaussian random
noise of AT b ’s = 0 K to 2 K (1<t) added to input T& channels and also a case assum­
ing error m geophysical parameters of 0 5 g/cm 2, 0 001 g/cm 2, 0 5 m /s and 0.5 K m
WVC, CLW, SSW and SST respectively have been considered as the noisiest MSMR
observation on M.S-I dataset. The same noise conditions have been introduced in
test dataset (M_S-II) and tested the performances on this dataset Table (4.9) shows
impact of noise in (Class-II) NN model m terms of RMS errors tested on MJ3-II
dataset.
Table 4.9 Impact of noise on (8Tn —4GP) NN Model, performed on test dataset
(MJ3-II)
Epochs
1134
Iteration s
5611488
9434609
10060918
11883139
N oise Im p act on (8Tb —4GP) N N M odel
Noise =»
W VC
CLW
SSW
(g/cm2)
(g/cm2)
(m /s)
Ts(curue) (J. 0.0
0.0
0.0
OK (red)
0.1797
0.0495
0.0018
IK (green)
0 0489
0.0018
0.1417
2K (blue)
0 0519
0 0018
01448
0.5
0.001
0.5
2K (black)
0 0967
0 0023
0.2310
SST
(K)
0.0
0 2387
0 2866
0 2708
0.5
0 3318
The variation of RMS error vs epochs for such a case is shown in figures (4 4) (a)
WVC, (b) CLW, (c)SSW, and (d) SST, and each of these plots has four curves. The
first three curves (red, green and blue) are the case when the input,
72
T b ’s , is
perturbed
4.4 A n aly sis o f D ifferen t N e u ra l N e tw o rk M o d els
RUS (grvon*)
Figure 4.4: Impact of noise in (87# —4GP) NN Model. RMS Vs Epochs
If
if
Figure 4.5: Impact of noise in (87# —A G P ) NN Model, RMS Vs Global Iterations
73
4.4 Analysis of Different Neural Network Models
with noise of OK (no noise), IK, and 2K (lc) respectively m all channels, but with
no perturbation at output, GP's NN started converging at about 1134 epochs m all
the case The red curve (with no perturbation both at input as well as at the output)
converges faster It took about 5 x 10® iterations, compared to the case where the
noisy input TVs (ATb = 1 K and 2 K) for the same number of epochs (epochs and
global iteration are described m section (3 3)), NN took more iterations hence more
time to converge In all the three cases, model achieved almost the same retrieval
RMS error but took more iterations to converge as seen m figure (4.5). It is found
that when no noise is added to TVs the global iteration required for minimization is
less compared to when noise is added. Table (4 9) shows the RMS values at the end
of 1134 epochs Here the number of global iterations taken by NN model for different
noise scenario is different when iterated through same number of epochs For a case
of A T b = 2 K noise m all the input channels and with no noise at output shows
the RMS error of 0 0519 g/cm 2, 0 0018 g/cm 2, 0.1448 m /2, and 0 2708 K for WVC,
CLW, SSW and SST respectively after about 1134 epochs
The fourth curve (black), shows a very complex and noisiest scenario where, noise
of A Tb = 2K is added m all the mput channels and with noise m the output as well,
a sample case of noise added at the output with 0 05 g/cm 2, 0 001 g/cm 2, 0 5 m /s
and 0 5 K m WVC, CLW, SSW and SST respectively. This is the case where noise
m the training dataset is added at both mput and output vectors The RMS errors
of this curve almost remained nearly within the acceptable error limit, 0 0967 g/cm 2,
0 0023 g/cm 2, 0.2310 m /2, and 0 3318 K for WVC, CLW, SSW and SST respectively
after about 1134 epochs.
The above investigations show high tolerances of NN to additional noise during
the training and testing, and suggest that the 3-layers with 15-neurons m each layer
74
4.4 Analysis of Different Neural Network Models
is an optimally configured NN to derive the transfer function for GPs either single
or multiple output, discussed above. Based on the inferences of the configuration op­
timization and the noise impact studies, NN models for retrieving other geophysical
parameters like WVC, CLW, SSW and SST individually as well as simultaneously
have been developed using different input and output combinations of MSMR chan­
nels Results of the theoretical (using simulations only) RMS errors of different NN
models for these geophysical parameters (developed on training dataset and tested
on test dataset) are depicted m Table (4 10)
Table 410 Theoretical RMS error of different NN Models
Data type
M.S-I (29500)
M.S-II (132710)
N N Models
4Tb - 1GP
4Tb - 2GP
6Tb - 1GP
6Tb - 2GP
8Tb - 1 GP
8Tb - 2GP
8Tb - 4GF
4Tb - 1GP
4Tb - 2GP
6Te - 1G P
6Tb - 2GP
8T b - 1GP
8Tb - 2GP
8Tb - 4GP
WVC
(g/cm2)
0 1034
01002
CLW
(g/cm2)
0 0031
0.0022
-
~
-
-
-
0.0432
01089
0 0996
0 0022
0.0029
0 0027
-
-
-
-
-
0.0413
0 0020
SSW
(m/s)
SST
(K)
-
-
-
0 3324
0 1500
0.3609
0 2352
0.3431
0 3473
0.2931
0 3573
0 3407
0 4598
-
-
-
-
0 3224
0.1358
0 3549
0.2324
0.3418
0.3364
0.2753
0 3643
0 3137
0 4525
As shown m Table (4 10), the NN models (class-I as well class-II) for different
geophysical parameters were developed with 3HL-15N configuration. The noise figures
given m Table (4.11) have been used to perturb for the models given m Table (3 2)
Here NN minimum normalized error limits of 0 0029 and learning rate of 0 7 have
been used
75
4.4 Analysis of Different Neural Network Models
Moreover the distribution of retrieval errors of elass-II models for all the pa­
rameters on test dataset has also been studied, is as shown in figures (4 6)
The
error distribution for all the geophysical parameters is found to be Gaussian The
(8Tj?..4GP) NN models yielding minimum error with a Gaussian error distribution
is considered for geophysical parameter retrieval using actual satellite data and their
comparison and validation with other datasets
-0 8
-0 4
-0 0
04
Figure 4.6- Error distribution of class-II NN model (8T b — AGP) on test dataset
(M_S-II) for (a) WVC {9 /cm 2), (b) CLW (g/cm2), (c) SSW (m/s), (d) SST(K)
76
M SMR (8T#-4GP)NN C LW (gnv'cm2)
MSMR (8Tb-4GP)NN WVC (gm /orf)
4.4 Analysis of Different Neural Network Models
2
4
6
0.00
MSMR W VC Simu (gnvcm2)
0.02
0.04
0.06
0.08
0,10
0.12
0.14
M SMR (8T0-4GP)NN SSW tnVs)
MSMR C lW Simu (gnVcm2)
3
6
9
12
15
18
21
273
24
276
279
282
285
288
291
294
297
300
MSMR SST Simu(K)
MSMR SSW Simu(m/$)
Figure 4.7: Scatter plot of simulated and retrieved parameter using (87B —AGP) XX
model, trained with M_S-I (29500 data points) and applied on test dataset (M.S-II.
132710 data points)
77
4.4 Analysis of Different Neural Network Models
The 8Tj5-4GP NN model has minimum RMS error because not only more number
of channels are used but also the correlations among brightness temperature and the
geophysical parameters themselves is considered by NN technique. This fact is often
ignored in the design of simpler statistical regression retrieval method (Jung et a l ,
1998).
Another interesting point to note about the evolution of transfer function by NN
model is that the NN is trained with only 18% (M_S-I) of the total dataset and is then
applied to test dataset (M_S-II) containing 82% of remaining dataset The scatter
plots between the simulated values and the NN derived WVC, CLW, SSW and SST
applied on the test dataset (MJ3-II) are shown m figure (4 7) depicting correlation
coefficients of 0 9998, 0.9975, 0 9995 and 0.9999, with RMS errors of 0 0413 (g / c m 2),
0 0020
( g / c m 2),
0 3418 (m/s) and 0.4525 (K) respectively In spite of training the
NN model with small number of traimng dataset (this data contains all range of
the parameters at discrete interval, distribution as shown in figure (4.2)) is able to
represent the test dataset which is about 82% of the dataset.
4.5 Retrieval of Geophysical Parameters from MSMR
4.5
Retrieval of Geophysical Parameters from MSMR
The NN algorithms have been applied to MSMR brightness temperature data and
the geophysical parameters have been derived for different NN models (single or multi
parameter). Global distribution of WVC, CLW, SSW and SST for a typical period
during Jul 15-16, 1999 is shown in figures (4 8 to 4.11) respectively Also shown in
these figures are the MSMR operational products and similar available products from
TMI and SSM/I radiometers (http://ssm i com)
These plots have been generated
using 2° x 2° spatially averaged values over global oceans
Global distribution of
SSW and SST from MSMR as depicted m respective figures shows stnpiness
This stripiness could be firstly due to the presence of noise as well as bias in
Tb
data exceeding the parameter sensitivity specifically at the extremes of the swath, and
secondly due to possible polarization coupling in MSMR T b data caused by scanning
mechanism with fixed feed and rotating reflector To confirm the presences of possible
stripiness m
Tb
data itself a global distribution of MSMR T b data for all the channels
has also been analyzed as shown m figure (4 12). These plots are monthly averaged
over 2° x 2° spatial box over global oceans. Among all the channels pronounced
striping is seen m horizontal polarization distribution specifically at 6 and 10 GHz
which are the prime channels for surface parameter like SSW and SST. However mild
striping is also seen in vertical polarization specifically at 6 and 10 GHz
79
4.5 R etrieval o f G eophysical P aram eters from M S M R
(a) MSMR, NN (4TB - 1GP) WVC (g/cm2)
(b) MSMR, SR (4TB - 1GP) WVC (g/cm2)
(c) MSMR, NN (8TB - AGP) WVC (g/cm2)
(d) SSM/I, Wentz WVC (g/cm2)
(f) TMI. Wentz WVC (g/cm2)
Figure 4.8: Global distribution of MSMR retrieved WVC (g/cm2) using various NN
models, MSMR operational products and Wentz products for SSM/I and TMI
80
4.5 R e trie v a l o f G e o p h y sic al P a r a m e te r s fro m M S M R
(a) MSMR, NN (4TB - 1GP) CLW (g/cm2)
(b) MSMR, SR (4TB - IG F) WVC (g/cm2)
(c) MSMR. NN (87 b - 4GP) CLW (g/cm2)
(d) SSM/I, Wentz CLW (g/cm2)
(f) TMI, Wentz CLW (g/cm2)
Figure 4.9: Global distribution of MSMR retrieved CLW ( g / c m 2) using various NN
models, MSMR operational products and Wentz products for SSM/I and TMI
81
4.5 R e trie v a l o f G e o p h y sic al P a r a m e te r s fro m M S M R
(a) MSMR, NN (8TB - 1GP) SSW (m / s )
(b) MSMR, SR (STB - 1GP) SSW (m/s)
(c) MSMR, NN (6TB - 1GP) SSW (m/s)
(d) SSM/I, Wentz SSW (m/s)
(e) MSMR, NN (8TB - 4GP) SSW (m/s)
(f) TMI, Wentz SSW (m/s)
Figure 4.10: Global distribution of MSMR retrieved SSW (m/s) using various NN
models, MSMR operational products and Wentz products for SSM/I and TMI
82
4.5 R e trie v a l o f G e o p h y sic al P a r a m e te r s fro m M S M R
(a)
MSMR, NN (8TB - 1GP) SST (K )
(c)
MSMR, NN (67 b - 1GP) SST (.K )
(b)
(e) MSMR, NN (87 b - 4GP) SSW (K)
MSMR, SR (87 b - 1GP) SST (K )
(f) TMI, Wentz SST (K)
Figure 4.11: Global distribution of MSMR retrieved SST (K) using various NN mod­
els, MSMR operational products and Wentz products for SSM/I and TMI
83
4.5 R e trie v a l o f G e o p h y sic al P a r a m e te r s fro m M S M R
Figure 4.12: Monthly average of MSMR T b at 2° x 2° grid for Jul-1999 (a) TB06V
(b) TB06H (c) TB10V (d) TB 10H (e) TS 18V (f) TS 18H (g) 7’S 21V (h) r fl21H
84
4.6 Analysis of Across Track Distribution of Brightness Temperature
4.6
Analysis of Across Track Distribution of Bright­
ness Temperature
The errors in brightness temperatures are most likely related to instrument calibration
process. An increase of noise and bias m brightness temperature may come from im­
proper along-track averagmg, maccurate spill-over correction, ignorance m non-hnear
correction, degradation of on-board calibration targets and polarization rotation due
to cross-track scanning. Spill-over effects from antenna sidelobes, where the antenna
reflector does not fully subtend the view of the antenna feed, introducing a reduction
to the earth scene brightness temperature that enters the feed. In addition, stray
energy that does not come from the earth scene also enters the feed, and introduces
errors into the antenna feed temperature. In such cases, the extraneous energy must
be corrected from the antenna temperature, TA, to get the earth scene component of
the antenna temperature.
In principle, the measurement characteristics of the sensor should be completely
independent of scan position. It is assumed that the temperature field does not vary
significantly over one scan. However m practice, problems such as spacecraft attitude
errors, obstructions m field of view, and side lobes seeing the spacecraft may result
m systematic errors that are a function of scan position. Given enough observations,
we might determine the along-sean error by making the assumption that the effect
of weather and surface features will, on average, be the same at all scan positions
(Wentz et a l , 2001) If this is the case, then the along-scan errors can be detected
by taking a simple across track histogram of Tb
To further examine the reasons for striping in SSW and SST, long period of
six month (Jun - Oct, 1999) across track distribution of normalized histogram of
T b for all the MSME channels has been studied and compared with the similar
85
4.6 Analysis of Across Track Distribution of Brightness Temperature
distribution for TMI and SSM/I Tb ’s, the distribution are shown in figure (4.13) It
is observed that histogram peak bends irregularly across the swath which other wise
should be invariant across the swath. Similar distribution for TMI and SSM/I TB’s
data as shown by figures (4 14 and 4 15 respectively) do not show such irregularity as
anticipating due to the simultaneous scanning of feed and reflector. The comparison of
across track histogram distribution of MSMR with those of TMI and SSM/I indicates
the presence of relatively large errors at the extremes scan position as compared to
the central scan position.
Across track bias corrections are needed for a cross-track scanning instrument
when the reflector normal angle and the polarization angle depart from the nominal
designs
Unlike the conical scanning sensor that views at a fix angle and whose
reflector rotates with the feed horn and the energy captured by the radiometer is
associated with the single polarization Across track instrument whose reflector scans
receives the electric components at both horizontal and vertical polarization. MSMR
on the other hand though has conical scanmng mechanism but with fixed feed and only
scanning reflector Hence, the energy contributing factor to the polarization mismatch
is the rotation of the antenna pattern polarization coordinates caused by the scanning
of the antenna reflector relative to the fixed antenna feed. The polarization rotation
due to antenna,scan motion relative to fixed feed for correcting antenna temperature
measurements and retrieving the true brightness temperatures have been developed
for the SMMR (Njoku, 1980) Other empirical algorithms have been also developed
and produced corrections for radiances from polarization misalignment and other
sources of radiation (Weng et a! , 2003). For a conical scanning system, some crossscan biases are also detected and corrected (Colton and Poe, 1994).
For MSMR cross-track bias correction, the results of across track Tb ’s histograms
86
4.6 A n a ly sis o f A cross T rack D is trib u tio n o f B rig h tn e ss T e m p e ra tu re
(h)
(g) MSMR 21H T b
Figure 4.13: Across swath MSMR
of all channels
Tb
MSMR 21V TB
, six month (Jun - Oct, 1999) data histogram
87
4.6 A n a ly sis o f A cro ss T rack D is trib u tio n o f B rig h tn e ss T e m p e ra tu re
(b) TMI 10V T b
(a) TMI 10H T b
(f) TMI 22V Tb
(g) TMI 37H T b
(h) TMI 37V TB
Figure 4.14: Across swath TMI Tg, six month (Jun - Oct, 1999) data histogram of
low resolution channels
88
4.6 A nalysis o f A cross T rack D is trib u tio n of B rig h tn e ss T e m p e ra tu re
(b) SSM /I 19V T b
(a) SSM/I 19H T b
S S M / I F - 1 3 T „ - 2 2 V | J u n - O c t , 1 8 ff» |
1
5
s
£
I-
i.'
fig *
■
a.
I B B
T „ |K |
(d) SSM /I 22V T b
T „ |K |
T „(K |
(f) SSM /I 37V T b
(c) SSM/I 37H T b
Figure 4.15: Across swath SSM/I
low resolution channels
TB,six month (Jun - Oct, 1999) data histogram of
89
4.6 Analysis of Across Track Distribution of Brightness Temperature
are used to correct the bias across the swath with references to the values near the
center of scan position of the swath. In order to do so, the difference of TB value
corresponding to histogram peak at any scan position from the value near center
of scan location is used as an additional correction to TB data
An adjustment is
calculated for each scan position with respect to the center of the swath But this
activity does not help much m removing the stnpmess m the Tg data as it has
been found that these errors are highly irregular, and are different for ascending and
descending pass and they rarely correspond to a real statistical bias However, the
correction has been applied to only those extreme scan points which have differences
more than 2 K from that of average value of three points around center This irregular
Tb error at the extremes of the scan could be due to the polarization mixing arising
from geometry with fixed feed and rotating reflector.
Figure (4 16) shows a sample of the instant across track SSW retrieved using 8Tb —
AGP NN model, before and after the across track bias correction. It clearly shows
the discrepancies at the extremes of the swath The difference between brightness
temperatures at center and at the swath edge is as high as 14 i f among all the
brightness temperatures. Instead of plottmg individual brightness temperature the
most affected geophysical parameter SSW, has been used in the analysis and shown
m figure (4 16)
Figure (4 17c) shows the differences m SSW retrieval before and after the across
track bias correction It is found that the difference in retrieved SSW is more than 3
m /s at the extreme edges of the scan across the track The SSW distribution here
for 2 days average on 2° x 2° grids is shown. Figures (4.17a) and (4 17b) shows the
retrieval before and after across track bias correction, respectively.
90
4.6 A n a ly sis o f A cro ss T ra ck D is trib u tio n o f B rig h tn e ss T e m p e ra tu re
Figure 4.16: Instant across track SSW, (a) before and (b) after across track bias
correction
91
4.6 A n a ly sis o f A cro ss T ra ck D is trib u tio n o f B rig h tn e s s T e m p e ra tu re
Figure 4.17: SSW retrieval using (8T b —4G P ) NN model, (a) before and (b) after
across track bias correction and (c) the differences
92
4.6 A n alysis o f A cross Track D istrib u tion o f B righ tn ess T em perature
4.6.1
I n - O r b it D a ta Q u a lity E v a lu a tio n
Having identified various errors in MSMR GP and also in
Tb
it was required to
evaluate the in-orbit quality of T b data and apply correction as far as possible prior
to validation and fine tuning of retrieval algorithms.
T b 'ms_dev(i)
T b bias (i)
(a) Iterative minimization process
(b) Bias minimization
(c) Channel bias
Figure 4.18: MSMR in-orbit data quality evaluation
93
4.6 Analysis of Across Track Distribution of Brightness Temperature
In-orbit data quality evaluation is performed by minimizing the differences of
m-orbit Tb , measurement with T§, simulated for a large database of geophysical
conditions. Global Tb data for a limited period is processed using equation (4 6 1)
and database is generated.
ch= 8
mint
(E
4=1
■rp
Bt • ■ T iJ
Nd
)+ {Cssr —SSTs)2
(4 61)
where C s s t , is SST climatology at l°xl° grid, S S T s,ls the sea surface temperature
from the simulated database and N ^ , is the number of MSMR channels. Further these
dataset is used to estimate the bias (Tsbias) and RMS deviations (TBrrns^dev) for
each channel, as shown by equations (4 6.2) and (4.6 3). Here NP, is number of points
m this database
k*=NP
TBTmsjdev(i) =
*=1
\ E
k—N P /r p
^ a s (z ) a
(?W -r|
NP
rp S
a) 2
(4.6.2)
'i
(4 6 3)
k=1
The bias m the data is used iteratively so that new bias dimmish to a minimum
yielding final RMS deviation as shown m figure (4.18) In this process, the iterations
were truncated on the bases of mean of bias value of all the channels, is minimized till
consecutive changes is zero as shown m figure (4.18 b) The RMS deviation resulted
from quality evaluation is mcorporated m retrieval algorithm while the bias term is
used along with actual respective T b’s data for retrieval of geophysical parameters.
The results of this analysis correspond to initial quality of the Tb data in terms of
Tgbtaslt) and TBrms„dev(i) with respect to radiative transfer model based on which
all the retrieval algorithms are developed. For this purpose MSMR data for several
94
4.6 A nalysis of A cross Track D istrib u tio n of B rightness T e m p era tu re
Table 4 11: MSME bias RMS deviation with respect to RT model for all the channels
Channels
6V
6H
10V
10H
18V
18H
21V
21H
Term s Aev
12133
16212
14167
1 9098
2.1003
2.1501
21
2 0032
Tsbias
-0 2525 -1 4438
2 8599 0 9721
2 0994 -1 9709 -0 9260
0 9400
months is used to evaluate the Tebias(t) and Term s-dev^) as mentioned above with
respect to radiative transfer model for all the channels The new biases are given m
Table (4 11) which are used in die present analysis. These new offsets are adjusted
from the observed 2Vs, and remove the overall bias between the model T § ’s and the
observations
Across track bias correction along with m-orbit bias correction obtained for long
period MSMR data have been applied for updating the retrieval of geophysical pa­
rameters The results of updated retrieval and comparison with other datasets are
discussed m Chapter 5
95
4.7 Retrieval Algorithm and Analysis of Derived Parameters: TMI case
4.7
Retrieval Algorithm and Analysis of Derived
Parameters: TMI case
In the case of TEMM/TMI retrieval, a tropical training dataset has been prepared
from the simulated database (discussed xn section 4 3) by randomly picking up ap­
proximately 30% of total datasets The various procedures like optimization of NN
models with single and multi-parameter retrievals, m orbit data quality evolution etc
are similar to those performed for MSMR The m-orbit data quality evaluation results
are depicted in Table (4.12)
Table 4 12 TMI bias and EMS deviation with respect to RT model for low resolution
channels
channels
10V
10H
19V
19H
22V
37V
37H
SST
SSW
Tgrras-dev
12615
12030
1 3336
1 2299
2 3193
18343
17156
12619
1 5759
T sh a s
-0 8705
0 5964
-1 0256
-0 1102
16780
0 4354
-0 7410
0 7593
11369
The final retrieval model developed for TMI is a multi-parameter (class-II) model
utilizing 7 channels as input with 3 hidden layers containing 15 neurons m each layer,
and 4 parameters like WVC, CLW, SSW and SST as output The theoretical RMS
errors of various parameters are depicted in Table (4.13)
This model has been applied to TMI data for the period Jul 15-16, 1999 for
retrieving various geophysical parameters The global distribution of derived para­
meters and the corresponding Wentz products are shown in figures (4.19) and figures
(4.20) Results of comparison of derived products with Wentz’s products and m-situ
data are discussed in Chapter 5
96
97
(d) TMI, NN (77 b - 4CP) SST ( K)
(b) TMI, NN (77b - 4GP) CLW ( j / m 2)
Figure 4.19: Global distribution TMI (77# —4 0 P) NN derived parameters for 15-16 Jul, 1999
(c.) TMI, NN (77 b - 4GP) SSW (m /s)
(a) TMI, NN (7TB - 4GP) WVC (g/cm2)
4.7 R e trie v a l A lg o rith m a n d A n aly sis o f D e riv ed P a ra m e te rs : T M I case
98
Figure 4.20: Global distribution Wentz finished product for 15-16 Jul, 1999
(c) TMI, Wentz SSW (rn/s)
-frit
SH
(b) TMI, Wentz CLW (g/cm2)
(>/) .LSS Z1U3A\ 'IWX (P)
(a) TMI, Wentz WVC (g/cm2
4.7 R e trie v a l A lg o rith m a n d A n a ly sis o f D e riv ed P a ra m e te rs : T M I case
4.7 Retrieval Algorithm and Analysis of Derived Parameters: TMI case
Table 4 13’ Theoretical RMS error for TMI {7Tb —4GP) retrieval on testing data
points
Simulated
Error limit 0.0029
N N (7Tb - AGP)
(pts 80000)
Mm
Max
Mean
Mm
Max
Mean
RMS
Bias
RMS*
Corrl
SST(jRr)
2810
305 0
293 4
280 67
305 61
293 5
0 4112
0.0842
0 4025
0 947
SSW (m/s)
0 000
24 00
8 665
0 1266
23 908
8 664
0 2018
-0 0014
0 2018
0 999
WVC (g/cm2)
0 296
8177
2 975
0 3962
7 8626
2 983
0 0384
0 0076
0 0376
0 999
a M (g /cm ? )
0 000
0 130
0 014
0 0004
01325
0 015
0 0014
0 0001
0 0014
0 999
RMS* is RMS error after bias removal
The purpose of retrieval using TMI data was to examine the striping seen m the
MSMR retrieved parameters In order to do so, an interim NN based retrieval algo­
rithm was developed for TMI and the parameters were derived from TMI data. The
interim algorithm based derived parameters for TMI did not indicate such stripmess
Hence, the problems were in MSMR data The possible reason could be due to polar­
ization mixing which is due to fixed feed and rotating reflector Retrieved products
from MSMR shows striping m surface parameters while that from TMI do not show
such stripmess. Global MSMR 'TVs analysis and along-track analysis shows errors
in the m-orbit Tb s measurements These errors m brightness temperatures are most
likely related to the available MSMR Tb data.
99
C hapter 5
C om parison a n d V alid atio n of
D erived G eophysical P a ra m e te rs
C hapter 5
Com parison and V alidation of
D erived G eophysical Param eters
5.1
Introduction
This Chapter deals with the validation and comparison of various derived geophysical
parameters with m-sttu and other satellite products. Validation of satellite retrievals
is complicated by several important differences between satellite and m-situ measure­
ments. It needs to be recognized that there is significant spatial/temporal mhomogeneity between m-situ and satellite measurements In-situ measurements are time
averages at a smgle point, while satellite measurements are instantaneous measure­
ments averaged over a large spatial footprint. The comparison of derived parameters
has been performed for a limited period of two days while validation has been carried
for longer period for SST and SSW, where as for WVC small number of available
collocated datasets have been used. Data for comparison and validation purpose has
been obtained from various sources like ships, buoys, radiosonde and finished prod­
ucts from different satellites
Data sources and results of comparison of different
parameters have been described.
100
5.2 D escrip tion o f D ata set
5.2
D escription o f D ata set
Two types of data sets used for validation and comparison are described below.
5.2.1
V a lid ation D a ta
The datasets have been procured from various sources, namely, ICOADS (Interna­
tional Comprehensive Ocean-Atmosphere Data Set) containing surface meteorolog­
ical observations from various ships (commercial, research, and fishing) and buoys
(moored and drifting). NIOT (National Institute of Ocean Technology, India) buoy
dataset containing surface metrological data in deep Sea region (Premkumar et al.,
2000) and Vaisala radiosonde data over oceans containing vertical profile of temper­
ature humidity and winds.
5.2.2
In ter-com p arison D a ta
Finished products from TMI and SSM/I produced by Wentz products available from
Global Hydrology Resource Centre (GHRC). NASA, USA. (http://ssmi.com) have
been used for inter comparison purpose. The finished products from TMI contain
daily averages of WVC, CLW, SSW and SST. over 0.25° x 0.25° grid over tropical
region only while the finished products from SSM/I contain global daily averages of
WVC, CLW and SSW, over 0.25° x 0.25° grid. Moreover, finished products from
MSMR contain swath data of WVC, CLW. SSW and SST at 1.5 degree spatial res­
olution. Retrieval of SST and SSW products from operational MSMR makes use
of all the channels (6.6, 10.6, 18.0 and 21.0 GHz channels with dual polarization)
while WVC and CLW products makes use of 18.0 and 21.0 GHz channels with dual
polarization. The MSMR finished products data is also referred as MSMR SR.
101
5.3 Validation and Comparison of MSMR Derived Parameters
5.3
Validation and Comparison of MSMR Derived
Parameters
Various geophysical parameters derived from MSMR data using class-II multi-parameter
NN model (8Tb —4GP), many time also referred as MSMR NN, have been validated
with m-situ data. Validation of different parameters is discussed separately m view
of differences in data type, validation data source and validation period.
5.3.1
Water Vapour Content
Water vapour derived from MSMR data for the periods from Jul 15 to Aug 28 1999
(BOBMEX-1999 campaign) and during Mar 2001 has been validated with Vaisala
radiosonde water vapour content
The collocation of the MSMR and m-stiu data
were earned out withm ±1 hour duration and within 0 5° x 0 5° grid yielding 25
collocated data points The comparison of this dataset indicates bias corrected RMS
difference of 0 35 g/cm 2, as shown m Table (5 1) and its scatter plot is shown m
figure (5.1). The corresponding comparison of MSMR finished product of WVC with
m-situ data shows bias corrected RMS difference of 0 38 g/cm 2.
Table 5.1 Statistics of validation of MSMR derived water vapour with Vaisala
Parameters
MSMR NN
Vaisala (Ship)
(points 25)
Mm
Max
Mean
Mm
Max
Mean
RMS
Bias
RMS*
Corrl
WVC(gm/cm2)
2 308
6 872
5 388
2 820
6 660
5 231
0 381
-0 156
0.347
0 957
0 484
-0 298
0.381
0 948
MSMR SR
WVC(gm/cm2)
2 710
7100
5 529
Vaisala (Ship)
2 820
RMS * is RMS error after bias removal.
102
6 660
5 231
l
■
l
■
I
.
[
.
I
.
I
W V C (8 T B-4 G P ) N N (g /cm 2)
5.3 V alidation and C om parison of M S M R D erived P a ram e te rs
V a is a la W V C (g /cm 2)
V a is a ia W V C (g /c m 2)
Figure 5 1- Comparison of MSMR derived WVC with Vaisala (after bias removal)
(a)NN and (b)SR
The comparison of these RMS deviations indicates retrieval by neural network is
relatively better compared to finished products. A slight overestimation at the higher
WVC (> 5
g / c m 2)
by SR is observed which has improved m the neural network
inversion. Moreover, the WVC values are in higher range above 2 3 g / c m 2 , this may
be because the collocated datasets obtained are corresponding to monsoon period
over the Indian Ocean As the simulation based global non-linear neural network
algorithms are developed for ram free situations. The derived parameters under
rainy condition may have large errors Hence, during the monsoon period the derived
water vapour may overestimate for any rainy situation
In addition to the validation, inter-comparison of MSMR derived WVC with those
of TMI and SSMI is also carried out for the period Oct 9-10, 1999 The statistic of
inter-comparison and respective scatter plots are shown m Table (5.2) and figure
103
5.3 Validation and Comparison of MSMR Derived Parameters
Table 5 2. Inter-comparisons of water vapour of MSMR derived and MSMR finish
products with TMI and SSM/I
W YC (g / c m 2) Num . of Points: 4447
Mm
Max
MSMR-NN
0 530
7.154 3193
TMI
0 713 6 204
3.241
MSMR-SR
0 862
3 465
TMI
0 713 6 204 3 241
MSMR-NN
0 530
7154
3.193
SSM/I
0 778
6.443
3 273
MSMR-SR
0 862
6 710
3 465
SSM/I
0 778
6 443 3 273
TMI
0 713
6 204 3.241
SSM/I
6.710
0 778 6 443
Mean
Corrl.
RMS
Bias
0 971
0 3739
0.0477
0.964
0 3925 -0 2239
0 965
0.3852
0 961
0 3944 -0.1922
0 977
0 2878 -0 0317
0.0793
3.273
(5 2) respectively The mter-comparison results indicate that NN derived MSMR
parameters are closer to other satellite products as compared to MSMR finished
products
104
CD
J.
i
■ t
«
•
r
M
i ■ i
-
*
o
f
( ■UJ3/B) 3 A M NN U D * - B L B ) - h V 4 S \ N
M S M R -(8 T B -4 G P ) N N W V C ( g /c m 2)
5.3 V a lid atio n a n d C o m p a riso n o f M S M R D e riv ed P a ra m e te rs
SSM /I WVC (g/em *)
T M I W V C ( g /c m : )
l - x ....L -i-.-L ...
u 3
K
s 2
Ul
5
1
0
0
1
2
3
4
S
6
T M I W VC t a lc m 2)
7
9
0
1
2
3
4
5
6
7
SSM /I WVC (g /cm 2)
Figure 5.2: Inter-comparison of 2° x 2° averaged MSMR WVC with TMI and SSM/I
finished products, (a) TMI versus MSMR NN, (b) SSM/I versus MSMR NN, (c) TMI
versus MSMR SR, (d) SSM/I versus MSMR SR
105
5.3 Validation and Comparison of MSMR Derived Parameters
5.3.2
Cloud Liquid Water Content
The absolute validation of cloud liquid water is difficult to perform as m-situ observa­
tions firstly do not exist, and even if available through aircraft observations, datasets
are very limited m space and time and m its dynamic range As mention by Wentz and
Meissner (2000), apart from using upward-looking radiometers to calibrate downwardlooking radiometers (or vise versa), there are no other calibration sources for cloud
liquid water Several organizations maintain upward looking radiometers or carry out
routine aircraft flights to measure cloud liquid water to support of their meteorolog­
ical data requirements. The comparison of cloud liquid water inferred from upward
looking radiometers with those inferred from downward looking satelhte radiometers
has limited utility due to large spatial and temporal variability of clouds rendering
the comparisons relative to each other. Because of the non-availability of ground
truth pertaining to cloud liquid water, the result can only be inter-compared with
other satellite products. In view of such problems, only comparison of histogram of
cloud liquid water obtained from various satelhte sensors are possible (Wentz, 1997),
also proposed for AMSU by Wentz and Meissner (2000). The same method is applied
for MSMR NN CLW comparison with finished product (Wentz, 1997, Moreau et a l ,
2002) It is assumed that the probability density function (PDF) for the derived CLW
should have a peak at zero and should decrease exponentially as cloud liquid water
increases Any deviation of peak from zero value indicates error m either retrieval
process or in the Tb measurements. In a similar manner comparison of histogram of
CLW from MSMR NN and MSMR finished products has been performed as shown
m figure (5 3) for the period Oct 9-10, 1999 It is observed that histogram of both
the CLW data peaks at zero value, however, the shape seems to be more realistic
than those exhibited by SR algorithm indicating gradual decrease m NN CLW, while
106
O
o
CM
% Occurrence
5.3 Validation and Comparison of MSMR Derived Parameters
0
0.0
0.2
0.4
0.6
0.8
(8T b-4GP) NN CLW (kg/m2)
0.0
0.2
04
0.6
0.8
Operational SR CLW (kg/m2)
Figure 5.3- Comparison of histogram of CLW from MSMR (a) MSMR NN and (b)
MSMR finished product
abrupt decrease in finished products histogram, indicating the effectiveness of NN for
CLW retrieval
In addition to the comparison of histogram, global fields of MSMR cloud liquid
water derived from NN and the finished products from TMI and SSM/I are mtercompared for period Oct 9-10, 1999 The scatter plots of inter-comparison are shown
in figure (5 4), while the statistics of inter-comparison is shown in Table (5.3)
107
\\
M S M R -N N (B T B -4G P ) C L W (m m )
5.3 V a lid atio n a n d C o m p ariso n o f M S M R D e riv ed P a ra m e te rs
E
jE
s
Figure 5.4: Inter-comparison of 2° x 2° averages of MSMR cloud liquid water with
TMI and SSM/I finished products, (a) TMI versus MSMR NN (b) SSM/I versus
MSMR NN (c) TMI versus MSMR SR (d) SSM/I versus MSMR SR
108
5.3 V alidation a n d C om parison o f M S M R D erived P a ra m e te rs
Table 5 3' Inter-comparisons of cloud liquid water of MSMR derived and MSMR
finish products with TMI and SSM/I
CLW (mm) N u m b er of P oints: 4409
Mm.
Max
Mean
MSMR-NN
0000
8282
.0543
TMI
.0076
.6467
0863
MSMR-SR
0000
.4445
0423
TMI
6467
.0863
MSMR-NN
0000
8282
0543
SSM/I
.0075
9587
.0955
MSMR-SR
.0000
4445
0423
SSM/I
0075
9587
0955
TMI
0076
6467
.0863
SSM/I
0075
9587
0955
109
Corr
RMS
Bias
.5825
0 0710
0.0320
4387
0.0894
0 0441
.5502
0 0916
0 0412
4237
0.0944
0.0532
6756
0 0617 -0.0091
5.3 Validation and Comparison of MSMR Derived Parameters
The scatter plots for TMI show an overestimation and underestimation for cloud
liquid water by NN and SR algorithm, respectively. Where as the NN derived CLW
is in reasonable comparison with that of SSM/I but, the MSMR finished product
still underestimates as compared to SSM/I. W hat is more important, however, is
that the NN derived CLW show's a relatively large variability compared to finished
products, with RMS error of 0.0710 mm and 0 0894 m m compared to TMI, and
RMS error of 0 0916 m m and 0 0944 mm compared to SSM/I, respectively The
inter-comparison results indicate that NN derived MSMR parameters are closer to
other satellite products as compared to MSMR finished products
Further research efforts are needed to validate the oceanic CLW over long period
of time at various sites especially m the tropics, with ground truth datasets. These
ground based and satellite datasets will be valuable tools for testing the general
circulation models that currently predict cloud liquid water and will play vital role
m better understanding of the significance of cloud m climate and climate change
(Greenwald et a l , 1993)
110
5.3 V a lid a tio n a n d C o m p a riso n o f M S M R D e riv e d P a r a m e te r s
5 .3 .3
Sea Su rface T em p era tu re and W in d S p ee d
MSMR derived SST and SSW are validated with in-situ data from ICOADS ships and
buoys as well as from NIOT buoys for the period from Jan to Dec 2000. Validation
is carried out within ±1 hour and 0.5° x 0.5° spatial grids. The plots of collocated
positions of MSMR with ICOADS ships and buoys are shown in figures (5.5 a and
5.5 b). respectively.
O'
40'
80'
120'
160'
2 00'
2 40'
280'
0‘
40 '
80 '
120'
160 '
200 '
240 '
260 '
320
320 '
3 60'
360'
Figure 5.5: Positions of collocated MSMR and ICOADS datasets (a) Ships (10,911
points) (b) buoys (2569 points)
The scatter plots of MSMR derived parameters and ICOADS ships data are shown
in figure (5.6). The statistics of validation of MSMR NN derived SST and SSW with
in-situ data is given in Table (5.4) indicating bias corrected RMS deviations of 1.87
(K) and 3.02 (m/s) in SST and SSW respectively. The bias corrected RMS deviations
111
5.3 Validation and Comparison of MS MR Derived Parameters
Table 5 4 Validation of MSMR NN and SR derived SST and SSW with m-situ
(ICOADS Ships, » 10900, collocated data points)
Parameters
MSMR NN
ICOADS (Ship)
(pts 10900)
Mm
Max
Mean
Min
Max
Mean
RMS
Bias
RMS’
Corrl
SSW(m/s)
0 827
24 106
8 383
0 500
24 000
7 721
3 098
-0 663
3.027
0 612
SST(K)
278 29
304 58
296 75
274 90
305 10
297 09
1899
0 336
1.869
0 921
MSMR SR
ICOADS (Ship)
SSW(m/s)
0 000
28 800
9 019
0 500
24 000
7 721
3 462
-1 299
3.209
0 554
SST(iC)
280 60
304 99
297 03
274 90
305 10
297 09
2 10
0 059
2.099
0 891
RMS * is EMS error after bias removal
of SST and SSW values from MSMR finished products are found to be 2 09 (K) and
3 20 (m/s), respectively, as seen m Table (5.4)
The vahdation results shown for SSW and SST are based on the 8Tjg — AGP
multi-parameter neural network global retrieval model. The comparison of these
RMS deviations indicates retrieval by neural network is relatively better compared
to finished products In both SSW and SST the spread is less and wider ranges are
covered compared to finished products as seen m figure (5 6)
112
NN (8TB-4GP) SST (K)
5.3 V a lid atio n a n d C o m p a riso n of M S M R D e riv ed P a ra m e te rs
n—'-T” "'.. r ”1...r" !—r
276 280 284 288 292 296 300 304
Insitu (Ship) SST(K)
4
8
12
16
20
Insitu (Ship) SSW (m/s)
4
8
12
16
20
Insitu (Ship) SSW (m /s)
SR (8TB-1GPJSST (K)
SR (8TB-1GP) SSW (m/s)
<d|
0
276 280 284 288 292 296 300 304
Insitu (Ship) SST (K)
Figure 5.6: Scatter plots of MSMR derived parameters with ICOADS ships (a) NN
SSW (b) SR SSW (c) NN SST (d) SR SST
113
5.3 V alidation and C om parison o f M S M R D erived P aram eters
0 .0 0 0
(a )
0 .0 0 6
0 .0 1 2
0 .0 1 8
0 .0 2 4 0 .0 3 0
M S M R N N C L W (of'cm*)
0 .0 3 6
u.ovu u.uuo u.un u.uio u . u < u .ujo u .ujo
(b )
M S M R S R C L W (g /cm *)
0
(C )
<d >
1
2
3
4
5
6
M S M R N N W V C <g/cm*}
-
-
7
0
(e )
-
M S M R S R W V C (g fc m *)
2
4
6
8
10
12
14
16
M S M R NN S S W {rrVs)
~
(U
M S M R S R S S W (nVa)
Figure 5.7: The residual A S S T , MSMR (NN/SR) SST - ship SST are plotted against
(a/b) MSMR NN/SR CLW, (c/d) MSMR NN/SR WVC, (e/f) MSMR NN/SR SSW.
The solid lines (blue) on each figure indicate the average residual, while the dashed
lines (red) are ± one standard deviation from the mean.
Dependency of SST retrieval from MSMR using NN and SR algorithms is analyzed
under varying cloud liquid water, columnar water vapor, and surface wind. The
analysis is verified by examining the residual A S S T (MSMR - Ship), as a function of
simultaneously derived various other environmental parameters. Here ship data are
from ICOADS dataset. Figure 5.7(a/b) shows ASST* plotted against MSMR NN/SR
derived cloud liquid water. It demonstrates the ability to retrieve SST through cloud,
with no dependence on cloud amount. The NN based SST bias is almost independent
on CLW, while SR based SST bias has weak dependency on CLW. Figure 5.7(cf) indicate a weak dependency on NN based SST bias on WVC and SSW, while
SR based SST bias indicates relatively more dependency on WVC and SSW. This
indicates more robustness of NN as compared to SR algorithm.
114
18
5.3 Validation and Comparison of MSMR Derived Parameters
Validation of MSMR derived SST and SSW with that of ICOADS buoys is also
carried out for the period from Jan to Apr 2000 within ±1 hour and 0 5° x 0 5° spatial
grids The moored-buoy m Tropical Ocean Global Atmosphere/Tropical Atmospheric
Ocean (TOGA/TAO) array covers the tropical Pacific ocean These buoys are placed
at approximately 10 to 15 degree longitude intervals and 2° to 3° degree latitude
intervals The scatter plots of MSMR derived parameters and ICOADS buoys are
shown m figure (5 8) The statistics of validation of MSMR derived SST and SSW
with ICOADS buoys is given m Table (5 5) indicating NN derived bias corrected
RMS deviations of 1 TO (m/s) and 1.55 (K) m SSW and SST respectively. The
bias corrected RMS deviations of MSMR finished products of SSW and SST from
ICOADS buoys are found to be 1 95 (m/s) and 1 33 (K) respectively, as seen m
Table (5 5) The reason for differences found m comparisons using ICOADS ships
and buoys could possibly be due to difference in reference heights as normalization
with respect to height is not performed (Mears et a l , 2001). In addition some bias
as well as scatter specifically for low wind speed are possible due to neutral stability
conditioning required.
115
5.3 Validation and Comparison of M SM R Derived Parameters
. I
■ i
i
'
■ '
■ '
cd
o
r
o o
d
x
In-situ (Buoy) SSW(rrVs)
.
i
co
co
o
o
SR SST(K)
i
ro
.
i
ro
,
C
£>
o
CN
c
o c
o
o
o
i
C\J
o
CO
o
o
■ '
) 2 4 6 8 10 12 14 1618
j __ t
o
CO
CO
■ i
5) O) O
2 4 6 8 10 12 14 16 18
to
CO
o>
■ i
RMSE = 1.951 F
Bias = -2.584
Corrl = 0.691
2H
In-situ (Buoy) SSW(nVs)
MSMR NN SST(K)
■ i
tO
I, I. I. I , I, I
i
A
SR SSWfnv's)
MSMR NN SSW(nrVs)
18
16
t
296 298 300 302 304
296 298 300 302 304
In-situ (Buoy) SST(K)
In-situ (Buoy) SST(K)
Figure 5.8: Scatter plots of MSMR NN/SR derived parameters with ICOADS buoys,
(a) NN SSW (b) SR SSW (c) NN SST (d) SR SST
116
5.3 Validation and Comparison of MSMR Derived Parameters
Table 5.5 Validation of MSMR NN/SR derived SST and SSW, with in-situ (ICOADS
Buoys, number of collocated pomts » 2569)
Parameters
MSMR NN
ICOADS (Buoy)
(pts 2569)
Mm
Max
Mean
Mm
Max
Mean
RMS
Bias
RMS*
Corrl
SSW (m /s)
0 079
16 504
6 727
0 000
12 400
5 893
2 122
-0 833
1.704
0 694
SST(R')
294 97
304 49
300 59
294 10
303 40
300 04
1 591
-0 551
1.555
0 699
MSMR SR
ICOADS (Buoy)
SSW (m /s)
2.400
16 400
8 478
0 000
12 400
5 893
3 096
-2 584
1.951
0 619
SST(i7)
286 39
304 99
300 39
29410
303 40
300 04
1 366
-0 307
1.331
0 682
RMS* is RMS error after bias removal
Study of variation of NN/SR algorithm RMS errors as well as biases of SSW and
SST with their magnitude has been studied For this purpose, the magnitude of SSW
and SST have been binned at 1 m / s and 1°K, respectively, and the corresponding
RMS error and bias value have been evaluated (as depicted m figure (5.9) indicated
by respective legends). The number of points m each bm is shown at the top of the
each figure, and the combined RMS error and bias value are shown at the right side
of vertical dash line It has been observed that RMS errors and biases for NN derived
parameters (SSW and SST) before bias correction for all the bins are less compared
to SR derived parameters.
In general, the accuracy of MSMR NN and SR shows the similar trends, but RMS
error of NN derived SSW is relatively less compared to those of SR, while SR derived
SST error is better than the NN derived SST
117
5.3 V a lid a tio n a n d C o m p a riso n o f M S M R D e riv e d P a r a m e te r s
Figure 5.9: RMS error and bias, binned at every 1 m / s and 1 K for SSW and SST
with refrences to ICOADS buoys respectively, for MSMR NN derived and finished
products, (a) SSW (b)SST
Similar to validation with ICOADS data, MSMR derived parameters are also
validated with NIOT deep Sea buoys at different locations in Arabian Sea and Bay of
Bengal, Three buoy ( DS1 (15.5°IV, 69.3°E), DS2 (10.PW, 72.5°£) and DS3 (13.o°V,
90.8°E)) data for different period in year 2000 have been considered. Validation is
carried out within ±1 hour and 0.5° x 0.5° spatial grids. The combined statistics of
the validation is shown in Table (5.6). The bias corrected RMS differences of NN and
finished product for SST are 1.61 K and 1.89 K respectively while those for SSW
are 2.17 m / s and 2.50 m / s respectively. The NN derived parameters are found to be
closer to in-situ data as compared to those of finished products. The time series for
SSW from NIOT buoys along with MSMR NN derived and finished product values is
shown in figure (5.10) while time series for SST is shown in figure (5.11), also shown
location of the buoys.
In general, the trends of derived and finished products are found to be similar to
118
5.3 Validation and Comparison of MSMR Derived Parameters
OS-2 SSW(nV£)2000
-D S -3 —
1-Jah
31*Jan
i- h
31-Mar.
30^Apr^ 3QiMay-v 28*Mun'
29tJui
28 Aug ^27-Sep
27-Oot
- S R - — NN.
28 Nov
28 Deo
Dale
Figure 5.10: Time series of SSW with different NIOT buoys, MSMR NN derived and
finished products
119
5.3 Validation and Comparison of MSMR Derived Parameters
"VdiWi.,
-31i*lar
3 0 -A fir,
,2%'Juft
2& 0ul
28* iu 3
2 ?-$ ep
2 t-O rtj
26N oy
2^D *o
Figure 5.11- (a) Time series of SST with different NIOT buoys, MSMR NN derived
and finished products, (b) Location of buoys
120
5.3 Validation and Comparison of MSMR Derived Parameters
Table 5 6 Validation of MSMR NN/SR derived SST and SSW, with tn-sttu (NIOT
buoys, number of collocated points « 250).
parameters
Mm
Max.
Mean
MSMR NN
Mm
Max
Mean
DS (1,2,3) (buoys)
RMS
Bias
RMS* Corrl
Number of points 250
SSW(m/s)
0 068
14 617
6.747
0 100
13 670
5154
2 688
-1 593
2.165
0 701
SST(A)
29712
305 97
30180
29813
302 77
300 85
1870
-0 952
1.609
0 569
MSMR SR
DS (1,2,3) (buoys)
SSW(m/s)
2 090
18 290
9157
0100
13 670
5154
4 718
-4 003
2.498
0 574
SST(Jf)
286 26
304 82
300 98
298 13
302 77
300 85
1898
-0138
1.894
0 430
RMS* is RMS error after bias remora!
those of buoys, except during monsoon period where the retrievals are known to be
more erroneous under heavy cloudy and rainy conditions. Apart from this, during
monsoon period the finished products are seen to be more fluctuating as compared
to NN derived values
Increased variability of SSW and SST mainly for DS1 and
DS2 buoy measurements after May 30 is typical of Asian monsoon over Arabian Sea
The MSMR NN as well SR wmd speed faithfully captures the major variability. The
observed variability of MSMR derived SST could be due to the fact that the MSMR
SST is the skm temperature, whereas the buoy measures the bulk temperature Bulk
SST is representative of the temperature of the upper few centimeter of the ocean
surface Skin SST is the temperature of the surface skm of the ocean, which is less
than a m m thick. No correction has been applied to the NIOT buoy SSW data to
a standard height for comparison with MSMR NN/SR data. This may lead to some
bias as well as scatter, specifically for low wmd speed
121
5.3 Validation and Comparison of MSMR Derived Parameters
In addition to the validation, comparison of 2° x 2° averages of MSMR derived
parameters with MSMR finished products and TMI and SSM/I finished products has
also been carried out for the period Oct 9-10, 1999. Comparison is carried out within
±1 hour The statistics of comparison and the respective scatter plots are shown m
Table (5 7) and figures (5.12 and 5 13) respectively The comparison results indicate
that NN derived MSMR parameters are closer to other satellite products as compared
to MSMR finished products.
122
5.3 Validation and Comparison of MSMR, Derived P aram eters
Table 5 7 Inter-eompansons of SSW and SST for MSMR NN/SR derived with TMI
and SSM/I finished products
Parameter
SSW (m
/s
) N um ber of Points: 4652
Model 4
Mm
Max.
Mean
MSMR-NN
0.790
21 417
7.748
TMI
0 998
22.155
7137
MSMR-SR
3 600
22 200
9 512
TMI
0.998
22 155
7136
MSMR-NN
0 790
21417
7.746
SSM/I
1283
21 150
7 207
MSMR-SR
3 600
22 200
9 483
SSM/I
1283
21.150
7.207
Corr
RMS
Bias
0.5201 2 4279 01999
0 6255 3 0380 2 3722
0 5202 2 6958 0 1340
0 5723 3 1198 2 2908
Parameter
SST (K
=>
)
N um ber of Points: 4674
Max.
Mean
Model #
Mm.
MSMR-NN
280.839 305.928 296 745
TMI
282 395 304 578 296 906
MSMR-SR
280.950 304 850 296.473
TMI
282.395 304 578 296 906
123
Corr
RMS
0 9672
1 3604 01696
0.9642
1 3546 0 4340
Bias
M SM R (8TB -4G P ) NN SST
A
N
O
O)
M SM R (8TB -4G P ) NN SSW
5.3 V a lid atio n a n d C o m p ariso n of M S M R D e riv ed P a ra m e te rs
288
294
300
TM I (W e n tz) SST
306
288
294
300
TMI (W entz) SST
306
M SM R O p r (SR) SST
M S M R O pr (SR) SSW
o r o * - c r > c o o t O i f » . o >
282
Figure 5.12: Inter-comparison of MSMR NN/SR derived SST (K) and SSW (m/s)
with TMI finished products (a) TMI versus MSMR NN, SSW (b) TMI versus MSMR
NN, SST (c) TMI versus MSMR SR, SSW (d) TMI versus MSMR SR, SST
3 _______ - - l i - -----------r
Co r r l - 0
2
4
5
6
7
8
2
3
~
10 12 14 16 18
SSM I (W e n tz ) SS W
Figure 5.13: Inter-comparison of MSMR NN/SR derived SSW (m/s) with SSM/I
(Wentz) products (a) SSM/I versus MSMR NN, SSW (b) SSM/I versus MSMR SR.
SSW
124
5.3 V alidation and C om parison of M S M R D erived P a ram e te rs
5.3.4
Comparison of MSMR NN derived parameters with
Finished Products
So far, validation and comparison of various geophysical parameters derived from
MSMR with m-situ and other satellite datasets have been carried out as discussed
above It is worthwhile to carry out the comparison for NN technique based derived
parameters using MSMR with similar parameters available as finished products from
MSMR which are based on statistical techniques
This comparison should bring
out relative performances of these two techniques Hence, neural network technique
based derived geophysical parameters from MSMR have been compared with those
of finished products for the period Oct, 9-10 1999
The global distribution of MSMR derived, WVC, CLW, SSW, and SST using NN
and SR techniques and the difference of NN and SR derived parameters are shown in
figures (5.14 to 5.17) respectively The comparison of these distributions reveal many
features. The retrieval of WVC using NN shows improvement over those using SR
technique Despite the discrepancies seen in across track histogram for horizontal and
vertical channel of 21 GHz, which is sensitive for atmospheric parameters, m general
the global distribution of WVC as well as CLW retrieval appears to be better in both
the NN and SR derived values This is due to observed compensatmg distribution of
21 GHz Tb across the swath (figures 4.13) as well as large sensitivity of T b to WVC
(« 11 K / gcm~2)
Significant difference in distribution of CLW derived using these two techniques is
observed, specifically over Arabian Sea, South Fiji Island and along band of convection
located in Pacific mid-tropical oceanic region A qualitative comparison of cloudiness
usmg Meteosat-7 IR, satellite imagery from EUMETSAT (European Organisation
for the Exploitation of Meteorological Satelhtes) and that of CLW from NN and SR
125
5.3 V alidation and C om parison o f M S M R D erived P aram eters
a
1
t o
as
ss a
a
a a
*
■* a
ea
a
aa
Figure 5.14: Global distribution of MSMR derived WVC (g / c m 2) for Oct 9-10 1999.
2 day. averaged 2° x 2° box. (a) NN (b) SR. and (c) NN-SR
5.3 V a lid a tio n a n d C o m p a riso n o f M S M R D e riv ed P a r a m e te r s
M a tto a a t XR ia ia g a ,
» O ct 1
M a ta o a a t XR i u g i ,
j—
„ „.
„
.Iff
10 Occ 1*»«
L. U
Figure 5.15: Global distribution of MSMR derived CLW (g /c m 2) for Oct 9-10 1999.
2 day, averaged 2° x 2° box, (a) NN (b) SR and (e) NN-SR
127
5.3 V a lid a tio n a n d C o m p a riso n o f M S M R D e riv e d P a r a m e te r s
Figure 5.16: Global distribution of MSMR derived SSW (m/s) for Oct 9-10 1999, 2
day, averaged 2° x 2° box, (a) NN (b) SR and (c) NN-SR
128
5.3 V a lid a tio n a n d C o m p a riso n o f M S M R D e riv ed P a r a m e te r s
Figure 5.17: Global distribution of MSMR derived SST (K) for Oct 9-10 1999, 2 day,
averaged 2° x 2° box, (a) NN (b) SR and (c) NN-SR
129
5.3 Validation and Comparison of MSMR Derived Parameters
derived product was carried out. The finished product CLW is underestimated as
compared to NN values. These regions axe clearly seen m the global distribution of
difference between NN and SR derived values as shown in figure (5 15 c)
Nearly
all CLW values m this region extends beyond 0 03 g/cm 2 indicatmg possible heavy
cloudiness In other areas a good correspondences can be seen between other smaller
individual cloud systems
Among all the possible combinations examined (4Tg —1GP, 4Tb —2GP, 8Tb
—
1GP, and 8Tg—4GP) NN algorithm, the best performance obtained is for 8Tg —4GP
NN algorithm The best performance is obtained usmg multi-frequency measurements
and is with simultaneous output configuration The eight MSMR channels allow an
estimation of the integrated cloud liquid water in a better way. The performance of
NN algorithm in the presence or absence of clouds shows that the clouds are taken
into account well in the simulation database. As a part of mter-comparison analysis
it is found that the CLW from a multi-parameter NN model is much better compared
to other configuration
Another difference observed in SSW is that low and high wind speed regions are
well picked up by NN while these features are found to be shallow m finished products
In order to bring out the differences m parameters generated by these two techniques,
scatter plots and comparison statistics of these parameters are shown in figure (5.18)
and Table (5 8) respectively
130
5.3 V a lid atio n a n d C o m p a riso n o f M S M R D e riv ed P a ra m e te rs
MSMR (8TB-4GP) NN CLW
c j
m
co
w
—
MSMR (8TB-4GP) NN WVC
0.12
0-,
0
0.1
0.08
0.06
0.04
0.02
0
1
2
3
4
5
(b) CLW (g/cn P )
C o rrl = 0.6022
RMS = 0.0083
Bias = -0.0009
6
0
0.02 0.04 0 06 0.08
o!l
0.12
MSMR (Opr.) SR CLW
MSMR (Opr.) SR WVC
CD
o
o
MSMR (8TB-4GP) NN SST
CO
ro
ro
CO
O
CO
ro
03
CO
CO
CO
CD
C\2
r-
Figure 5.18: Scatter plots of MSMR NN derived geophysical parameter with that of
MSMR finish products, (a) WVC, (b) CLW, (c) SSW and (d) SST
131
5.3 Validation and Comparison of MSMR Derived Parameters
Table 5.8- Comparisons of MSMR derived geophysical parameters with MSMR finish
products
Parameters
(8T b - 4GP) N N
finish p ro d u cts (SR )
(Pts 6950)
Mm
Max
Mean
Mm
Max
Mean
RMS
Bias
Corrl
WVC (g/cm2)
0 1669
6 7850
2 5202
0 2660
6 8006
2 7806
0 3497
0 2604
0 9900
c m (g/cm2)
0 0000
0 0981
0 0075
0 0000
0 0797
0 0066
0 0083
-0 0009
0 6022
SSW(m/s)
0 6448
21 4176
8 7602
3 6000
25 4000
10 3055
2 1454
15454
0 8379
SST(K)
271 370
305 928
291 454
273 070
304 850
291 795
2 0374
0 3415
0 9797
Operational S R S S W (mis)
(8TB-4G P) NN S S W (m/s)
Figure 5.19: Comparison of histogram of SSW from MSMR, (a) NN and (b) finished
product
132
5.3 Validation and Comparison of MSMR Derived Parameters
The scatter plots show that finished products CLW is under estimated while the
finished product SSW has relatively smaller dynamic range as compared to NN, as
also confirmed by histogram shown m figure(5 19) The comparison of SSW histogram
indicates shifting of peak toward higher values and has narrower width m finished
product
133
5.4 T M I V alidation
5.4
5.4.1
T M I V alid atio n
S ea Su rface T em p era tu re and W in d S p eed
TMI derived SST and SSW are validated with in-situ data from ICOADS ships and
buoys for 8 days during Jan to Dec 2000. Validation is carried out within 1 hour and
0.25° x 0.25° spatial grids. The plots of collocated positions of TMI with ICOADS
ships and buoys are shown in figures (5.20a and 5.20b), respectively.
Figure 5.20: Positions of collocated TMI and ICOADS datasets (a) Ships (b) buoys
The scatter plots of TMI derived parameters and ICOADS ships data are shown
in figure (5.21). The statistics of validation of TMI derived SST and SSW with in-situ
data is given in Table (5.9) indicating bias corrected RMS deviations of 1.41 (K) and
1.68 (m/s) in SST and SSW respectively. The bias corrected RMS deviations of SST
134
5.4 TMI Validation
Table 5.9 Comparison of TMI (7TB-4GP) NN derived SSW and SST with ICOADS
ships data with m ±1 hour with m 0.25° x 0 25° grid
ICOADSships
Num of Pt
N N (7Tb - 4GP)
(pts 443)
Mm
Max
Mean
Mm
Max
Mean
RMS
Bias
RMS*
Corrln
SSW(m/s)
0 000
19 60
8 289
0 047
2 580
8 329
1 677
-0 040
1.676
0 939
SST(K)
284 10
303 80
295.49
283 27
305 98
295 49
1415
0 059
1.415
0 949
ICOADSships
W entz products
SSW(m/s)
0 00
19 60
8 289
0 950
21 528
8 258
1616
-0 030
1.616
0 945
SST(K)
28410
303 80
295.49
283 22
305 50
295 54
1305
0 054
1.304
0 958
RMS* is RMS error after bias removal
and SSW values from TMI finished products axe found to be 1 30 (K) and 1 61 (m/s),
respectively, as seen m Table (5 9)
The number of collocated observations obtained for TMI T’a and m-situ ship, is
about 443 points for 8 days of available TMI data. These 8 days are for different
month (Jan, Jul, Apr, and Oct, 2000), with 2 days m each month Validation of TMI
retrieved parameters is performed with limited number of collocated TMl/vn-situ
observations. The comparison of these RMS deviations indicates retrieval by NN is
close to the Wentz finished products
135
NN (71B -4G P) SST (K)
5.4 TMI Validation
W
k>
O
O
O
A
W
O
D
I C
O
W
^ O
W entz SSW (m /s)
”T" 4r*’ '*■" t~
M S S ( d o r a u ) NN
!NOCO(D'f(NOCO(D'fCO
ro
(N CN
(S /U J )
T ~
2
2
4 6 8 1 0 1 2 14 1 6 18 2 0 22
In s ltu (Ship) SSW (m/s)
4 6 8 10 1 2 14 1 6 18 2 0
In s ltu (S hip) SSW (m /s)
Figure 5 21- Scatter plots of TMI NN derived parameters with ICO ADS ships (a)
NN SSW (b) SR SSW (c) NN SST (d) SR SST
136
5.4 TMI Validation
Table 5.10' Comparison of TMI (7TB-4 G P ) NN derived SSW and SST with ICOADS
buoys data with m 1 hour in 0.25° x 0 25° grid
In-situ
(pts 448)
N N (7Tb - AGP)
(348)
Mm
Max
Mean
Mm
Max
Mean
RMS
Bias
RMS *
Corrl
SSW(m/s)
0 900
1110
6181
0154
12 029
6 234
1402
0 053
1.401
0 787
SST(K)
294 0
303 0
299 39
290 30
305 51
299 29
1547
-0 104
1.510
0 757
In-situ
W entz products
SSW(m/s)
0 900
11 10
6 181
1 154
12 18
6 204
1165
0 022
1.165
0 840
SST(K)
294 0
303 0
299 39
292 44
306 10
299 44
1467
0 045
1.423
0 796
RMS' is RMS error after bias removal
Validation of TMI derived SST and SSW with that ICOADS buoys is also carried
out. The scatter plots of TMI derived parameters and ICOADS buoys are shown in
figure (5 22). For m-situ buoy, collocated observations obtained are about 448 pomts
for 8 days of available TMI Tb The 8 days data are from different month (Jan, Apr,
Jul and Oct, 2000), with 2 consecutive days for each month
The statistics of validation with limited dataset of TMI NN derived SST and SSW
with ICOADS buoys, is given m Table (5 10) indicating bias corrected RMS deviations
of 1 51 (K) and 1.40 (m/s) m SST and SSW respectively. The bias corrected RMS
deviations of TMI Wentz finished products of SST and SSW from ICOADS buoys
are found to be 1 42 (K) and 1 16 (m/s), respectively, as seen in Table (5.10). The
comparison results of these RMS deviations indicates that retrieval by NN are close
to the Wentz finished products
137
5.4 TMI Validation
306
306
303
303
300 H
297
294
300
297
294
291
291
T------- 1------- 1------- 1------- 1------- 1------- 1------- 1------
291
NN (7TB -4G P ) SS W (m /s)
W entz S S T (K )
<b>
294
297
300
303
Insltu (To ga Buoy) S S T (K)
294
297
300
303
Insltu (To ga B uoy) S S T (K)
291
14 -
14 -
12
-
12 -
rm s - 1 1664
corrfn = 0 8405
10
-
10 -
(«
8
-
6
-
W e n tz S S W (m /s)
NN (7 T B 4 G P ! S S T (K )
r m s » 1 4679
corrJn * 0 7961
4 -
2
-
0
-
6 -
*♦
SjfB?
*55****
v#
.
4 -
0 12
++
& ***
8 -
2 -
2
4
6
8
10
Insltu (To ga B uoy) S S W (m/s)
+
306
•*%
— r-*r—
i— i— r— 1— i— 1— i— r-
2
4
6
8
10
12
Insltu (To ga B uoy) S S W (m /s)
14
Figure 5 22. Scatter plots of TMI derived parameters with ICOADS buoys (a) NN
SSW (b) SR SSW (c) NN SST (d) SR SST
138
5.4 TMI Validation
Table 5 11 Comparisons of TMI NN derived geophysical parameters with TMI finish
products for 2 days 15-16, Jan 2000, 0 25° x 0.25° (latitude-longitude)average
Parameters
Class-II (7TB -AGP)
W entz (Single O utput)
(pts ~ 30000)
Mm
Max
Mean
Mm
Max
Mean
RMS
bias -
corrl.
SST {K)
280 06
305 91
298 64
280 575
307 05
297 38
3 2713
-1 2174
0 9030
SSW (m/s)
0 0010
24 001
6 7697
01500
23 400
5 2607
2 5903
-1 5091
0 7324
WVC (g/cm2)
0 2060
8 4900
3 3008
0 3300
7 3500
3 3285
0 3107
0 0277
0 9766
CLW(g/em2)
0 0001
0 1412
0 0105
0 0010
Q 1730
0 0094
0 0109
-0 0011
0 7389
In addition to the validation, comparison of TMI derived parameters with TMI
Wentz finished products has also been carried out for 2 days (15-16, Jan 2000),
over 0 25° x 0 25° (latitude-longitude grids)
The statistics shown in Table (5 11)
depicts comparison of geophysical parameters from Wentz’s finished products with
those derived using NN from TMI measurements and a scatter plot is shown in figure
(5 23) For high WVC (6 > g/crii2) and CLW (0 04 > mm) situations the NN derived
values overestimated compared to Wentz’s finished products WVC and CLW In case
of NN derived SSW there is slight overestimation around SSW « 6m /s as compared to
Wentz’s SSW The comparison results indicate that NN derived TMI parameters are
close to TMI Wentz’s finished products. However, validation using more collocated
observations may be needed for stable comparison
The purpose of retrieval using TMI data was to examine the stnpmg seen in the
MSMR retrieved parameters In order to do so, an interim NN based retrieval algo­
rithm was developed for TMI and the parameters were derived from TMI data. The
interim algorithm based derived parameters for TMI did not indicate such stripmess
Hence, the problems were m MSMR data The possible reason could be due to polar­
ization mixing which is due to fixed feed and rotating reflector Retrieved products
139
NN WVC
5.4 T M I V a lid atio n
Wentz WVC
Wentz CLW
W entz SSW
W entz SST
Figure 5.23: Scatter plots of TMI NN derived geophysical parameter with that of
TMI finish products, a) WVC, b) CLW, c) SSW and d) SST
140
5.4 TMI Validation
from MSME shows striping m surface parameters while that from TMI do not show
such stripmess. Global MSMR 7V s analysis and along-track analysis shows errors
m the m-orbit TVs measurements These errors m brightness temperatures are most
likely related to the available MSMR Tb data.
In this chapter, validation of various geophysical parameters derived using MSMR
and TMI data has been carried out with m-siiu and similar products from other
sources In general it has been found that validation errors using NN are relatively
less compared to those of finished products specifically in the case of MSMR How­
ever, inter-comparison of the MSMR NN derived parameters with TMI and SSM/I
derived products are also found to be m better agreement compared to MSMR finished
products.
141
C hapter 6
C onclusions a n d F u tu re Scope
C hapter 6
C onclusions and Future Scope
The research work carried m this thesis is mainly related to the development of a new
multi-parameter retrieval algorithm based on the back-propagation neural network
approach for geophysical parameters like water vapour, cloud liquid water content,
sea surface wind speed and sea surface temperature from MSMR and TMI The
algorithms have been developed using simulated database through radiative transfer
models employing simulated atmospheric and surface conditions
NN is trained to estabhsh the relationship between WVC, CLW, SSW and SST,
and brightness temperature for MSMR and TMI radiometers The neural network
architecture has been optimized through error analysis which revealed that the si­
multaneous multiple parameter retrieval model, having the configuration of 3 hidden
layers with 15 neurons m each layer, as the best model Separate optimum models
have been developed for MSMR (8 input - 4 output) and TMI (7 input - 4 output)
respectively Study of impact of noise on NN based algorithms has been earned out
and it is found that retrieval errors are mostly in variant with noise, indicating its
effectiveness for retrieval.
The theoretical RMS errors of MSMR NN model for WVC, CLW, SSW and SST
are 0 0415 (g /c m 2), 0 0020 (g /c m 2), 0 3148 (m/s) and 0 4525 (K), respectively, while
142
the theoretical RMS errors of TMI NN model for WVC, CLW, SSW and SST are
0 0376 (g/cm2), 0.0014 (g/cm2). 0 2018 (m/s) and 0 4025(K), respectively The the­
oretical RMS errors of MSMR NN model are found to be much less than those of
quoted errors for MSMR operational model.
Validation of NN based retrievals using MSMR data with those from m-situ has
been carried out reveahng RMS deviations of 0 35 (g/cm2), 3 04 (m/s) and 1 87 (K) m
WVC, SSW and SST, respectively, which are found to be better than corresponding
RMS deviations of 0 38 (g/cm2), 3 23 (m/s) and 2 09 (K) of MSMR operational
products.
More so, validation of NN based retrieval using TMI data with those from in-situ
has also been carried out reveahng RMS deviations of of 1.67 (m/s) and 1 41 (K) m
SSW and SST, respectively, and found to be closer to those of TMI finished products
(Wentz products) having RMS differences of 1.61 (m/s) and 1 30 (K), respectively.
Analysis of TMI data has been carried out m view of striping seen in MSMR retrieved
parameters which were not observed m TMI retrievals.
Like-wise, inter comparison of MSMR retrievals with those of TMI and SSM/I
have also been carried out leading to conclusions similar to those of validation results
Based on the present analysis, it can be, m general, concluded that NN performs
better than statistical technique for retrieval specifically in the case of MSMR
Some of the future studies envisaged are, the development of regional algorithms
for improving the retrieval performances by limiting the dynamic range of geophys­
ical parameters, further improvements m the implementation of error minimization
scheme and the development of NN architecture with hidden layer with dynamically
varying neurons
143
B ib lio g ra p h y
Bibliography
A Abdelgadir
Forward and inverse Modeling of Canopy Directional Reflectance
using a Neural Network. International Journal of Remote Sensing, 19 453-471,
1998
F. Aires, C. Prigent, W. B. Rossow, and M. Rothstem
A New Neural Network
Approach Including first Guess for Retrieval of Atmospheric Water Vapor,Cloud
Liquid Water Path, Surface Temperature, and Emissivities over Land from Satellite
Microwave Observations. Journal of Geophysical Research, 106 14887-14907, 2001
F Aires, A Chedm, N. Scott, and W. B. Rossow. A Regularized Neural Net Approach
for Retrieval of Atmospheric and Surface Temperatures with the IASI Instrument
Journal of Applied Meteoroloy, 4T144-159, 2002
J C Ahshouse
Total precipitable water and rainfall determinations from the
SEASAT Scanning Multichannel Microwave Radiometer
Journal of Geophysical
Research, 88 1929-1935, 1983
J C Ahshouse, S Synder, J Vongsathorn, and R R Ferraro
Determination of
Oceanic total precipitable water from the SSM/I. IEEE Transactions on Geoscience
and Remote Sensing, 28.811-816, 1990a
144
BIBLIOGRAPHY
J. C. Ahshouse, J. B. Synder, E. R. Westwater, C. L. Swift, G. S. Ruf, S A. Synder,
J. Vongsathorn, and R R Ferraro Determination of cloud liquid water content
using the SSM/I. IEEE Transations on Geoscience and Remote Sensing, 28-817822, 1990b
P. M Atkinson and A. R L. Tatnall Neural Networks in Remote Sensing, Introduc­
tion International Journal of Remote Sensing, 18(4).699-709, 1997
J G. Attali and G Pages Approximation of Functions by a Multilayer Perceptrona New Approach Neural Networks, 10 1069-1081, 1997
A.H Barrett and V K Chung A Method for the Determination of High-Altitude
Water-vapor Abundance from Ground-based Microwave Observations, Journal of
Geophysical Research, 67-4259-4266, 1962.
P Bauer and P. Schlussel Rainfall, total Water, and Water Vapor over Sea from Po­
larized Microwave Simulations and Special Sensor Microwave/Imager Data Journal
of Geophysical Research, 98.20737-20759, 1993.
C R Cabrera-Mercader and D. H Staelin
Passive Microwave Relative Humidity
Retrievals using Feed-Forward Neural Networks IEEE Transactions on Geoscience
and Remote Sensing, 33 1324-1328, 1995.
G Catherine, P Pete, and J Charles Ocean Surface Air Temperature Derived from
Multiple Data sets and Artificial Neural Networks Geophysical Research Letters,
25(22):4217-4220, 1998.
A T C Chang and T T Wilheit Remote Sensing of Atmospheric Water Vapor, Liq­
uid Water and Wind Speed at the Ocean Surface by Passive Microwave Techniques
from the Nimbus-5 Satellite. Radio Science, 14 793-802, 1979
145
BIBLIO GRAPHY
H. D. Chang, P H. Hwang, T T Wilheit, A T. C Chang, D. H Staehn, and P W
Rosenkranz. Monthly Distributions of Precipitable Water from the NIMBUS 7
SMMR Data Journal of Geophysical Research, 89 5328-5334, 1984.
F. W Chen Global estimation of precipitation using opaque microwave bands. De­
partment of Electrical Engineering and Computer Science, M IT, USA, 2004.
T Chen and H Chen Approximation Capability to Functions of Several Variables,
nonlinear Functionals, and Operators by Radial Basis Function Neural Networks
Neural Networks, 6 904-910, 1995a.
T. Chen and H. Chen Universal Approximation to nonlinear Operators by Neural
Networks with Arbitrary Activation Function and its Application to Dynamical
Systems Neural Networks, 6 911-917, 1995b
A Cichocki and R. Unbehaue Neural Network for Optimization and Signal Processing
Application. John Wiley and Sons 122-133, New York and Brisbane and Toronto
and Singapore, 1993.
M C Colton and G. A. Poe
Shared Processing Program, Defense Meteorological
Satellite Program, Special Sensor Microwave /Imager Algorithm Symposium. Bul­
letin American Meteorological Society, 75.1663-1669, 1994
L N Connor and P S Chang. Ocean Surface Wind Retrievals using the TRMM
Microwave Imager IEEE Transactions on Geoscience and Remote Sensing, 38(4)
2009-2016, 2000
D Comford, I T Nabney, and G Ramage. Improved neural network scatterometer
forward models Journal of Geophysical Research, 106:22331-22338, 2001
146
BIBLIOGRAPHY
S L Cruz Pol, C S Ruf, and S J Keihm. Improved 20-32 GHz Atmospheric
Absorption Model Radio Science, 33(5) 1319-1333, 1998.
G. Cybenko Approximation by Superposition of Sigmoidal Functions Mathematics
of Control Signals and Systems, 2 303-314, 1989.
D T Davis. Solving inverse problems by Bayesian Iterative Inversion of a Forward
Model with Apphcation to Parameter Mapping usmg SMMR Remote Sensing Data.
IEEE Transactions on Geoscience and Remote Sensing, 33.1182-1193, 1995.
J D. Droppleman Apparent microwave emissivity of sea foam. Journal of Geophysical
Research, 75 696-698, 1970
W. J. Ellison, A Balana, G Delbos, K Lamkaouchi, L Eymard, C. Guillou, and
C Prigent New permittivity measurements of sea water. 1998.
L Eymard, S English, P Sobieski, D Lemaire, and E Obligis. Ocean surface emissiv­
ity modeling COST Action 712, C Mtzler - UE GOST and Umv. Bern, Brussels,
Bern, pages 129-148, 2000
J R. Eyre Inversion of Cloudy Satellite Sounding Radiances by nonlinear Optimal
Estimation I Theory and Simulation for TOVS Quarterly Journal of the Royal
Meteorological Society, 115 1001-1026, 1989a
J R. Eyre
Inversion of Cloudy Satelhte Soundmg Radiances by nonlinear Opti­
mal Estimation II Application to TOVS Data. Quarterly Journal of the Royal
Meteorological Society, 115T027-1037, 1989b
J R Eyre, G A Kelly, A P McNally, E. Andersson, and A Persson Assimilation
147
BIBLIOGRAPHY
of TOVS Radiance Information Through One-Dimensional Variational Analysis.
Quarterly Journal of the Royal Meteorological Society, 119, 1993
C R. Francis, D P Thomas, and E P L Windsor. The evaluation of smmr retrieval
algorithms, m satellite microwave remote sensing
edited by T D Allen, Ellis
Horwood, Chichester, England, pages 481^198, 1983
A Frus-Hansen Bayesian networks as a decision support tool in marine applications
PhD Thesis, Department of Naval Architecture and Offshore Engineering, Technical
University of Denmark, 208, December, 2002.
K. Funahashi. On the Approximate Realization of Continuous Mappings by Neural
Networks. Neural Networks, 2:183-192, 1989
M W. Gardner and S R. Dorling Artificial Neural Networks (the Multiplayer Per­
ception) a Review of Applications in the Atmospheric Sciences Atmospheric En­
vironment, 32'2627-2636, 1998
E Gerard
Restitution de leau hquide nuageuse par radiometrie hyprefrequence.
Th‘ese de doctorat, Univ. Paris VI Paris, 1996
E. Gerard and Eymard L Remote Sensing of Integrated Cloud Liquid Water Devel­
opment of Algorithms and Quality Control. Radio Science, 33.433-447, 1998.
P. Gloersen, D. J. Cavalieri, A. T. C. Chang, T T Wilheit, Campbell, O. M. Johannessen, K. B Katsaros, K. F Kunzi, D. B Ross, D Staelin, E P L Windsor, F T.
Barath, P. Gudmandsen, E Langham, and R. 0 . Ramseier. Summary of Results
from the First NIMBUS 7 SMMR Observations Journal of Geophysical Research,
89 5335-5344, 1984
148
BIBLIOGRAPHY
B S Goh.il, T A Hanharan, P C Pandey, and A. K Sharma. Remote sensing of
atmospheric water content from Bhaskara SAMIR data International Journal of
Remote Sensing, 3(3).235-241, 1982
B S Gohil, A K Mathur, and A. K. Varma Geophysical Parameter Retrieval over
Global Oceans from IRS-P4/ MSMR Procedmg of Pacific Ocean Remote Sensing
Conference 5-8 December, NIO, Goa, India 207-211, 2000
M. A. Goodberlet, C. T Swift, and J. C Wilkerson Remote Sensmg of Ocean Surface
Winds with the Special Sensor Microwave Imager. Journal of Geophysical Research,
94:14547-14555, 1989
M A Goodberlet, C T Swift, and J C Wilkerson
Ocean Surface Wind Speed
Measurements of the Special Sensor Microwave/Imager (SSM/I). IEEE Transations
on Geoscience and Remote Sensing, 28'823-828, 1990
T. J. Greenwald, G. L. Stephens, T H. Vonder Haar, and Jackson D L A Phys­
ical Retrieval of Cloud Liquid Water over the Global Ocean using Special Sensor
Microwave/Imager (SSM/I) Observation.
Journal of Geophysical Research, 98
18471-18448, 1993
N C Grody Remote sensing of Atmospheric Water Content from Satellites using
Microwave Radiometry IEEE Transactions on Antennas and Propagation, AP-24:
155-162, 1976
C Guillou, S J Enghsh, C Prigent, and D C Jones. Passive microwave airborne
measurements of the sea surface reponse at 89 and 157 ghz. 1996.
C Guillou, W. Ellison, L. Eymard, K Lamkaouchi, C Prigent, G Delbos, G. Balana,
149
BIBLIOGRAPHY
and S A Boukabara. Impact of new permittivity measurements on sea surface
emissivity modeling m microwaves. Radio Science, 33(3) 649-667, 1998
A. Gmssard The retrieval of atmospheric water vapor and cloud liquid water over the
oceans from a simple radiative transfer model: Application to ssm/i data IEEE
Transations on Geoscience and Remote Sensing, 36(1) 328-331, 1998.
A. Guissard and P Sobieski An approximate model for the microwave brightness
temperature of the sea
International Journal of Remote Sensing, 8 1607-1627,
1987
D. J. Hand. Discrimination and Classification. John Wiley, Chichester, 1981
W W Ho and W F Hall Measurements of the dielectric properties of sea water
and nacl solution at 2 65 ghz. 1973
W W Ho, A W. Love, and M J Van Melle Measurements of the dielectric properties
of sea water at 1 43 ghz NASA Contractor Report, CR-2458, 1974
J P Holhnger. Passive Microwave Measurements of Sea Surface Roughness IEEE
Transactions on Geoscience Electronics, GE-9T65-169, 1971
J P Hollmger, R Lo, G Poe, R Savage, and J Peirce
Special Sensor Mi­
crowave/Imager users Guide Tech Rep , Naval Research Laboratory, Washington,
DC, 1987.
K Hornik. Approximation Capabilities of Multilayer Feedforward Network Neural
Networks, 4-251-257, 1991
W W Hsieh and B Tang Applying Neural Network Models to Prediction and data
150
BIBLIOGRAPHY
Analysis m Meteorology and Oceanography Bulletin of American Meteorological
Society, 79 1855-1870, 1998
Guang-Bm Huang Learning capability and storage capacity of two-hidden-layer feed­
forward networks IEEE Transactions on Neural Networks, 14(2) 274-281, 2003.
Guang-Bm Huang and H. A. Babri Upper bounds on the number of hidden neurons
m feedforward networks with arbitrary bounded non-lmear activation functions
IEEE Transactions on Neural Networks, 9 224-229, Jan 1998
T. Jung, E. Euprecht, and F Wagner Determination of Cloud Liquid Water Path
over the Ocean fron Special Sensor Microwave/Imager (SSM/I) Data Using Neural
Networks International Journal of Applied Meteorology, 37.832-844, 1998.
U Karstens, C Simmer, and E Ruprecht Remote Sensing of Cloud Liquid Water.
Meteorology and Atmospheric Physics, 54-157-171, 1994
P Kerhrzm and P Refregier. Theoretical Investigation of the Robustness of Multilayer
Perceptrons Analysis of the Linear case and Extension to nonlinear Networks.
1995.
L A. Klein and C. T. Swift. An improved model for the dielectric constant of sea water
at microwave frequencies. IEEE Transations On Antennas Propagation, AP-25(1)
104-111, 1977.
V M Krasnopolsky and F Chevallier Some Neural Network Applications m Envi­
ronmental Sciences Advancing Computational Eefficiency of Environmental Nu­
merical Models. 2001
151
BIBLIOGRAPHY
V. M. Krasnopolsky and F. Chevallier. Some Neural Network Applications in Envi­
ronmental Sciences. Part II: Advancing Computational Efficiency of Environmental
Numerical Models. Neural Networks, 16:335-348, 2003.
V. M. Krasnopolsky and H. Schiller. Some Neural Network Applications in Envi­
ronmental Sciences. Part I: Forward and Inverse Problem in Geophysical Remote
Measurements. Neural Networks, 16:321-334. 2003.
V. M. Krasnopolsky. L. C. Breaker, and W. H. Gemmill. A Neural Network as a
Nonlinear Transfer Function Model for Retrieving Surface Wind Speeds from the
Special Sensor Microwave Imager. Journal of Geophysical Research, 100:1103311045, 1995a.
V. M. Krasnopolsky, W. H. Gemmill, and L. C. Breaker. A Multiparameter Empirical
Ocean Algorithm for SSM/I Retrievals. Canadian Journal of Remote Sensing. 25:
486-503, 1999.
V. M. Krasopolsky, D. V. Chalikov, and H. L. Tolman. A Neural Network Tech­
nique to Improve Computational Efficiency of Numerical Oceanic Models. Ocean
Modelling, 4:363-383, 2002.
C. Kummerow, W. Barnes, T. Kozu, J. Shiue, and J. Simpson. The Tropical Rainfall
Measuring Mission (TRMM) Sensor Package. Journal of Atmospheric and Oceanic
Technology, 15:809-817, 1998.
T. S. L’Ecuyer and G. L. Stephens. An Uncertainty Model for Bayesian Monte Carlo
Retrieval Algorithms: Application to the TRMM Observing System. Quarterly
Journal of the Royal Meteorological Society, 128:17131737, 2002.
152
BIBLIOGRAPHY
Shi Lei. Retrieval of Atmospheric Temperature Profiles from AMSU-A Measurement
Using a Neural Network Approach. Journal of Atomsphere and Ocean Technolgy,
18 343-350, 2001.
D Lemaire Non-fully developed sea state characteristics from real aperture radar
remote sensing Ph.D. Thesis, Umversit Cathohque de Louvain, Facult des Sciences
Appliques, page 220, 1998
H. J. Liebe, P. W. Rosenkranz. and G A Hufford. Atmospheric 60-GHz Oxygen
Spectrum’ New Laboratory Measurements and Line Parameters. Journal of Quan­
titative Spectroscopy & Radiative Transfer, 48 629-643, 1992.
H J. Liebe, G. A Hufford, and M G. Cotton
Propagation Modeling of Moist
Air and Suspended Water/Ice Particles at Frequencies Below 1000 GHz. AGARD
Conference Proceedings 542, Atmospheric Propagation Effects through Natural and
Man-Made Obscurants for Visible to MM-Wave Radiation, 1993.
R P Lipmann
An Introduction to Computing with Neural Nets
IEEE ASSP
Magazine, pages 4-21, 1987
G. Liu and J A Curry. Determination of Characteristic Feature of Cloud Liquid
Water from Satellite Microwave Measurements Journal of Geophysical Research,
98 5069-5092, 1993
S. A. McFarlane, K. F. Evans, and A S. Ackerman
A bayesian algorithm
for the retrieval of liquid water cloud properties from microwave radiometer
and millimeter radar data
Journal of Geophysical Research, 107(D16) 4317
doi 10 1029/2001JD001011, 2002.
153
BIBLIOGRAPHY
G. J McLachlan. Discriminant Analysts and Statistical Pattern Recognition John
Wiley and Sons, New York, 1992
C A Mears, D K Smith, and F J Wentz. Comparison of Special Sensor Microwave
Imager and buoy-measured wind speeds from 1987 to 1997. Journal of Geophysical
Research, 106(06) 11719 - 11730, 2001.
A S Milman and T T Wdheit. Sea surface temperatures form the Scanning Multi­
channel Microwave Radiometer on Nimbus-7. Journal of Geophysical Research, 90
11631-11641, 1985
E Moreau, C Mallet, B Maxbboux, F Badran, and C Klapisz Atmospheric Liquid
Water Retrieval Using a Gated Experts Neural Network Journal of Atmosphere
and Ocean Technology, 19:457-467, 2002
J.A Morgan. Bayesian estimation for land surface temperature retrieval The nui­
sance of emissivities. IEEE Transactions on Geoscience and Remote Sensing, 43
(6)T279-1288, 2005
M D Mueller Ozone Profile Retrieval from GOME data using a Neural Network
Approach (NNORSY) Journal of Geophysical Research, 2003.
E. G Njoku Antenna pattern correction procedures for the Scanning Multichannel
Microwave Radiometer (SMMR) Boundary-Layer Meteorology, 18(1) 79-98, 1980
E. G Njoku and L Swanson
Global Measurements of Sea Surface Temperature,
Wind Speed, and Atmospheric Water Content from Satelhte Microwave Radiometry Monthly Weather Remew, 111 1977-1987, 1983
154
, BIBLIOGRAPHY
W Nordberg, J Conaway, D B Ross, and T. T Wilheit. Measurement of Microwave
Emission from a Foam Covered Wmd Driven Sea. Journal of Atomsphenc Sciences,
38, 1971.
E Obligis, L Eymard, and R Gairola. Multiparameter Retrieval Algorithms Part
A
Forward Modeling And Performances of Multi-Lmear Regression
MEGHA-
TROPIQUES 2nd Scientific Workshop, 2-6 July, Pans, France, 2001.
P C. Pandey, B S Gohil, and A K Sharma Remote sensing of atmospheric water
content from satellite microwave radiometer (SAMIR) on Bhaskara. Mausam, 32
(1) 17-22, 1981.
P. C. Pandey, B S Gohil, and T. A. Hariharan A two-frequency logarithmic dif­
ferential technique for retrieving precipitable water from the satellite microwave
radiometer (SAMIR-II) onboard Bhaskara-II
IEEE Transations on Geoscience
and Remote Sensing, GE-22-647-655, 1984
J F. Paris. Transfer of thermal microwaves m the atmosphere Department of Mete­
orology, Texas A and M University, 2, Texas, 1971
G Petty On the response of the Special Sensor Microwave Imager to the marine
environment Implications for Atmospheric parameters Retrievals 1990.
G W Petty A Comparison of SSM/I Algorithms for the Estimation of Surface Wind.
Proceedings Shared Processing Network DMSP SS M /I Algonthm, Symposium, 8-10
(June), 1993
G W Petty and K B Katsaros The response of the ssm/i to the marine environment
part n a parameterization of the effect of the sea surface slope distribution on
155
BIBLIOGRAPHY
emission an reflection. Journal of Atomsphere and Ocean Technolgy, 11(3).617628, 1994.
G W Petty and K B Katsaros. Nimbus 7 SMMR precipitation observations cali­
brated against surface radar during TAMEX Journal of Applied Meteorology, 31
489-505, 1992.
L Phalippou Variational Retrieval of humidity profile, Wind Speed and Cloud liquid
Water path with SSM/I Potential for numerical weather prediction Quart. J Roy.
Meteor. Soc , 122 327-355, 1996
L. Pierce, K Sarabandi, and F T Ulaby. Application of an Artificial Neural Network
m Canopy Scattering Inversion International Journal of Remote Sensing, 15:32633270, 1994
C. Prabhakara, I Wang, Chang A. T. C., and P. Gloersen A statistical Examination
of Nimbus 7 SMMR Data and Remote Sensing of Sea Surface Temperature, Liquid
Water Content m Atmosphere and Surfaces Wind Speed GSFC NASA TM 84927,
1982
K. Premkumar, M Ravichandran, S R. Kalsi, Sengupta D., and S. Gadgil First
Result from a New Observational System Over the Indians Seas. Current Sciences,
78 323-330, 2000.
C. Prigent and P Abba. Sea surface equivalent brightness temperature at millimeter
wavelength Annales Geophysicae, 8'627-634, 1990.
C D Rodgers Retrieval of Atmospheric Temperature and Composition from Remote
Measurements of Thermal Radiation
Review of Geophysical Space Physics, 14,
1976
156
BIBLIOGRAPHY
P W Rosenkranz Rough-sea microwave emissivities measured with the ssm/i. IEEE
Transations on Geoscience and Remote Sensing, 30.1081-1085, 1992.
P W Rosenkranz Absorption of microwaves by atmospheric gasses, Atmospheric
Remote Sensing by Microwave Radiometry
M Janssen (ed), Chapter 2, John
Wiley & Sons, NY, 1993
P W Rosenkranz and D H. Staelm Microwave emissmty of ocean foam and its
effect on nadiral radiometric measurements Journal of Geophysical Research, 77.
6528-6538, 1972.
D. E Rumelhart, G. Hinton, and R. Williams
Learning representations by error
propagation to adaptive learning Algorithms. Foundations, Rumelhart, D. E., and
• McClelland J L , Eds The M IT Press, page 318, 1986.
H. Schiller and R. Doerffer. Neural Network for Emulation of an Inverse Model Op­
erational Derivation of case II Water Properties from MERIS Data. International
Journal of Remote Sensing, 20:1735-1746, 1999.
P Schluessel and W. J. Emery Atmospheric Water Vapour Over the Oceans from
SSM/I Measurements 1990.
P Schluessel and H. Luthardt Surface wind speeds over the North Sea from Special
Sensor Microwave/Imager Observations Journal of Geophysical Research, 96 48454853, 1991
J. E A Silby, F. X Kneizys, J . H Chetwynd, and R. A. McClatchey. Atmospheric
'IVansmittance/Radiance. Computer code Lowtran4 1978.
157
BIBLIOGRAPHY
R. Singh, B G. Yasudevan, P. K Pal, and P C. Joshi Artificial Neural Network
Approach for Estimation of Surface Specific Humidity and Air Temperature using
Multifrequency Scanning Microwave Radiometer Proceedings Indian Academy of
Sciences (Earth Planet Sci ) t 113*89—101, 2004
J. A Smith LAI inversion using a Back-Propagation Neural Network Trained with a
Multiple Scattering Model IEEE Transactions on Geoscience and Remote Sensing,
GE-31T102-1106, 1993
D. H. Staehn. Measurements and Interpretation of the Microwave Spectrum of the
Terrestrial Atmosphere Near 1-cm Wavelengths. Journal of Geophysical Research,
71 2875-2881, 1966
D H. Staehn, K F Kunzi, R L Pettyjohn, R K. L Poon, R. W. Wilcox, and J W
Waters Remote Sensing of Atmospheric Water Vapor and Liquid Water with the
Nimbus-5 Microwave Spectrometer. Journal of Applied Meteoroloy, 15 1204-1214,
1976
A Stogryn The Apparent Temperature of the Sea at Microwave Frequencies IEEE
Transactions on Antennas and Propagation, AP-15 278-286, 1967
A. Stogryn The emissivity of sea foam at microwave frequencies Journal of Geo­
physical Research, 1972
A P Stogryn, C T Butler, and T. J Bartolae, Ocean surface Wind Retrievals
from Special Sensor Microwave imager data with Neural Networks. Journal of
Geophysical Research, 90‘981-984, 1994
158
BIBLIOGRAPHY
S. Tamura and M. Tateishi Capabilities of a Four-Layered Feedforward Neural Net­
work
Fourlayers versus Three
IEEE Transactions on Neural Networks, 8(2)
251-255, Max 1997
S Thiria, C Mejia, F Badran, and M Crepon, A Neural Network Approach for Mod­
eling nonlinear Transfer Functions. Application for Wind Retrieval from Spaceborn
Scatterometer Data Journal of Geophysical Research, 98 22827-22841, 1993.
L Tsang, L Chen, S Oh, R. J Marks, and A T C Chang
Inversion of Snow
Parameters from Passive Microwave Remote Sensing Measurements by a Neural
Network trained with a Multiple Scattering Model IEEE Transactions on Geo­
science and Remote Sensing, GE-30 1015-1024, 1992
F. T. Ulaby, R K. Moore, and A. K. Fung. Microwave Remote Sensing Active
and Passive, Vol. II - Radar Remote Sensing and Surface Scattering and Emission
Theory Addison-Wesley, Advanced Book Program, Reading, Massachusetts, 609
pages, 1982.
L. Vann and Hu. Yong-X A neural network inversion system for atmospheric remotesensmg measurements IEEE Instrumentation and Measurement Technology Con­
ference Proceedings, 2 1613-1615, 2002
B G Vasudevan, B S Gohil, and V. K Agarwal. Backpropagation Neural-NetworkBased Retrieval of Atmospheric Water Vapor and Cloud Liquid Water From IRSP4 MSMR IEEE Transations on Geoscience and Remote Sensing, 42(5).985-990,
2004
S M Weiss and C. A. Kulikowski Computer systems that learn Classification and
159
BIBLIOGRAPHY
prediction methods from statistics. Neural Networks, Machine Learning and Expert
Systems, Morgan Kaufmann, San Mateo, 1991
F. Weng, L Zhao, G Poe, R Ferraro, X. Li, and N. Grody Advanced Microwave
Sounding Unit (AMSU) cloud and precipitation algorithms
Radio Science, 38'
8086-8096, 2003
F J. Wentz A Two-scale Scattering Model for Foam-free Sea Microwave Brightness
Temperatures. Journal of Geophysical Research, 80 3441-3446, 1975.
F. J. Wentz
A Model Function for Ocean Microwave Brightness Temperatures.
Journal of Geophysical Research, 88 1892-1908, 1983.
F J Wentz Measurement of Oceanic Wind Vector using Satellite Microwave Ra­
diometers. IEEE Transactions on Geoscience and Remote Sensing, 30 960-972,
1992.
F
J
Wentz
A Well-Calibrated Ocean Algorithm for Special Sensor Mi-
crowave/Imager Journal of Geophysical Research, 102:8703-8718, 1997
F J Wentz and T. Meissner Algorithm Theoretical Basis Document (ATBD) Version
2 RSS Technical Proposal 121599A-1, Santa Rosa CA, 2000.
F J Wentz, S. Petecheryeh, and L A. Thomas A Model Function for Ocean Radar
Cross Sections at 14 6 GHz Journal of Geophysical Research, 89 3689-3704, 1984.
F J Wentz, L A Mattox, and S Peteherych
New Algorithms for Microwave
Measurements of Ocean Winds: Applications to SeaSat and the Special Sensor Mi­
crowave Imager Journal of Geophysical Research, 91.2289-2307, 1986
160
BIBLIOGRAPHY
F J Wentz, Ashcroft P , and Gentemann C L Post-launch calibration of the TRMM
Microwave Imager IEEE Transactions on Geoscience and Remote Sensing, 39 415
422, 2001
T T Wilheit
The effect of Wind on the Microwave Emission from the Ocean’s
Surface at 37 GHz NASA STI/Recon Technical Report N, 78 30480-+, 1979a
T T Wilheit. A Model for the Microwave emissivity of the Ocean’s surface as a
function of Wind Speed
IEEE Transactions on Geoscience Electronics, GE-17*
244-249, 1979b
T T Wilheit and A T C. Chang
An Algorithm for Retrieval of Ocean surface
Atmospheric Parameters from the Observation of the scanning multichannel Mi­
crowave Radiometer Radio Science, 15:525-544, 1980
T T Wilheit and M G Fowler Microwave Radiometric Determination of Wind speed
at the Surface of the Ocean During BESEX IEEE Transactions on Antennas and
Propagation, AP-25 111-120, 1977
S T Wu and A K. Fung. A Non-Coherent Model for Microwave Emission and
Backscattermg from the Sea Surface Journal of Geophysical Research, 77'59175929, 1972
B J van der Zwaag, C Cees Slump, and L. Spaanenburg. Analysis of neural net­
works m terms of domain functions Procedmg 3rd IEEE Benelux Signal Processing
Symposium (SPS-200&), Leuven, Belgium,, March 2122, 2002
Q u I.U n l. Library
mil
T2854
— 161—
Документ
Категория
Без категории
Просмотров
0
Размер файла
12 897 Кб
Теги
sdewsdweddes
1/--страниц
Пожаловаться на содержимое документа