вход по аккаунту



код для вставкиСкачать
Remote Sensing Letters
ISSN: 2150-704X (Print) 2150-7058 (Online) Journal homepage:
Improving large-scale moso bamboo mapping
based on dense Landsat time series and auxiliary
data: a case study in Fujian Province, China
Chong Liu, Tianwei Xiong, Peng Gong & Shuhua Qi
To cite this article: Chong Liu, Tianwei Xiong, Peng Gong & Shuhua Qi (2018) Improving
large-scale moso bamboo mapping based on dense Landsat time series and auxiliary
data: a case study in Fujian Province, China, Remote Sensing Letters, 9:1, 1-10, DOI:
To link to this article:
Published online: 25 Oct 2017.
Submit your article to this journal
View related articles
View Crossmark data
Full Terms & Conditions of access and use can be found at
Download by: [UNSW Library]
Date: 26 October 2017, At: 06:22
VOL. 9, NO. 1, 1–10
Improving large-scale moso bamboo mapping based on
dense Landsat time series and auxiliary data: a case study in
Fujian Province, China
Chong Liu
, Tianwei Xiongb, Peng Gongc and Shuhua Qia,b
Downloaded by [UNSW Library] at 06:22 26 October 2017
Key Laboratory of Poyang Lake Wetland and Watershed Research (Ministry of Education), Jiangxi Normal
University, Nanchang, China; bSchool of Geography and Environment, Jiangxi Normal University,
Nanchang, China; cMinistry of Education Key Laboratory for Earth System Modeling, Center for Earth
System Science, Tsinghua University, Beijing, China
Bamboo forest, especially moso bamboo forest, is very important
to human society. However, our ability to detect large-scale moso
bamboo with optical remote sensing is still limited due to the
spectral similarity with other forest species and the influence of
cloud occurrence. In this study, we examined the capability of
dense Landsat time series for moso bamboo forest mapping by
comparing three different interpretation schemes . For each
scheme, two experimental groups were further conducted to
investigate the usefulness of gray-level co-occurrence matrix
(GLCM) textures. Considering classification accuracy, the full-season compositing strategy was regarded as the most efficient. It
was generally beneficial to include GLCM textures as input features, although their usefulness would be partially offset due to
noise/correlation issues. We also investigated the roles of 15 types
of auxiliary covariates in extracting moso bamboo and found some
of them could enhance the classification performance significantly.
With the full-season compositing scheme and crucial auxiliary
covariates, an improved moso bamboo mapping performance
(93.21% in overall accuracy and 73.97% in minimum accuracy)
was observed within the study area. Our evaluation results are
promising to provide robust guidelines for remote mapping of
moso bamboo forest over large areas.
Received 6 June 2017
Accepted 5 September 2017
1. Introduction
Bamboos (Family Poaceae: Sub-family Bambusoideae) are woody grasses broadly distributed in the tropical, subtropical and mild temperate regions. Although these plants
represent a relatively small fraction of the total forest area on earth, they play critical
roles in the integrity of the global ecosystem. China is one of the richest countries in
terms of bamboo coverage and diversity. Of the over 500 bamboo species found in
China, moso bamboo (Phyllostachys pubescens) is the most abundant, occupying more
than half of the nationwide bamboo area (Li et al. 2015). The advent of optical remote
Key Laboratory of Poyang Lake Wetland and Watershed Research
(Ministry of Education), Jiangxi Normal University, Nanchang, China
© 2017 Informa UK Limited, trading as Taylor & Francis Group
Downloaded by [UNSW Library] at 06:22 26 October 2017
sensing has revolutionized our ability to identify moso bamboo distribution in a spatially
explicit manner. During past decades, extensive efforts have been made on remote
detection of moso bamboo in various ecosystems. However, reliable moso bamboo
mapping to date still remains challenging because of two major issues. First, the spectral
characteristic of moso bamboo makes it easily confused with other forest species.
Second, moso bamboo is mostly distributed in subtropical East Asia where cloudy and
rainy climates dominate, so it is difficult to select satellite observations that are free of
cloud impact for a specified time period.
Several studies have demonstrated the advantages of using time series remote
sensing data for terrestrial environment monitoring (Wilson and Sader 2002; Diao and
Wang 2016). Compared to single-date remotely sensed data, multi-temporal image
stacks are more likely to include essential phenological characteristics and overcome
the cloud influence. With the opening of the Landsat archive from 2008, it is now
possible to take advantage of dense time series satellite data to identify land cover
distributions. Nevertheless, compared to other remote sensing analyses, the utilization
of satellite time series for moso bamboo mapping is still in its infancy, and knowledge is
particularly scarce on how the rich information carried by the dense satellite time series
can be efficiently employed.
In addition to the interpretation of time series images, auxiliary data is another key
factor that impacts on moso bamboo mapping accuracy. Recent studies indicated that
some geospatial covariates play an important role in the distribution of land cover
classes, especially those characterized by climate and topography (Zhu et al. 2016;
Dannenberg, Hakkenberg, and Song 2016). However, it is also noteworthy that the
inclusion of irrelevant auxiliary data probably results in nothing beneficial but increased
computation burden (Li et al. 2016). Given that numerous auxiliary data are currently
becoming available, there is an urgent need to evaluate their usefulness for moso
bamboo identification.
The object of this study was to develop an improved framework to map large-scale
moso bamboo with dense Landsat time series and auxiliary data. We examined the
performances of different image time series interpretation schemes, and then further
investigated the importance of 15 types of auxiliary data which may facilitate remotely
detecting moso bamboo. Based on acquired findings, a definitive moso bamboo mapping procedure was adopted and the result was compared to the reference data.
2. Material and methods
2.1. Study area and tools
The study area covers the whole territory of Fujian Province, which is located on the
southeast coast of mainland China with elevation gradients increasing from its southeast
to northwest (Figure 1). Owing to the subtropical humid climate, Fujian has suitable
temperature, rainfall and sunlight conditions for moso bamboo growth. As one of the
national ‘hot spots’ of moso bamboo industry development, Fujian occupies approximately 1% of China’s land area, but contributes nearly one quarter of China’s moso
bamboo yield (Li et al. 2015). We used the Google Earth Engine (GEE) to process
remotely sensed data and generate final mapping results. As a planetary-scale platform
Downloaded by [UNSW Library] at 06:22 26 October 2017
Figure 1. Geographic location of Fujian Province and the identification of Landsat 8 image footprints
included in this study.
for environmental data and analysis, GEE provides trillions of scientific measurements
and cloud computing resources. These advancements allow users to manipulate the
archives of remote sensing data online and make large area research more feasible. In
addition, in this study ArcGIS 10.3 was used to derive reference samples, and R software
was used for classification and statistical analysis.
2.2. Dataset and preprocessing
Landsat 8 imagery was selected as our primary data source. Hosted on GEE, all available
Landsat 8 surface reflectance data from 2014 to 2016 covering the study area were
gathered. These images were calibrated and atmospherically corrected using the L8SR
algorithm (Vermote et al. 2016). In all, we processed 781 Landsat 8 scenes with a range
of seasonality and fractional cloud cover. To better discriminate moso bamboo from
other forest plants, two spectral indices: the normalized difference vegetation index
(NDVI, Tucker 1979) and the normalized difference moisture index (NDMI, Wilson and
Sader 2002) were also calculated for each image scene.
Downloaded by [UNSW Library] at 06:22 26 October 2017
Auxiliary data were collected and loosely grouped into four categories: topographic,
vegetative, climatic and soil. Topographic covariates included elevation, slope and aspect,
all of which were derived from the Shuttle Radar Topography Mission digital elevation
model data (SRTM DEM). Vegetative covariates included the maximum, minimum, mean
and standard deviation of NDVI and NDMI during the study period. Climatic covariates
included precipitation, temperature and cloud frequency. To eliminate the inter-annual
uncertainty, multi-year average values from 1980 to 2015 were used for precipitation and
temperature data, while the cloud frequency was calculated based on CFmask results from
all historical Landsat 4–8 imagery (Zhu and Woodcock 2012). Additionally, we included a
single soil type layer as an auxiliary input. Among all auxiliary variables, topographic,
vegetative and cloud frequency layers were directly obtained from GEE with 30 m resolution, while precipitation, temperature and soil data were downloaded from the Institute of
Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences
(RESDC, with 500 m resolution. Therefore, these three datasets
were first resampled to 30 m using a bilinear method before their further utilization.
We adopted the third-time forest management inventory (3rd FMI) data as a basis to
generate training and testing samples. Implemented by the provincial forestry department, the 3rd FMI results offered comprehensive forest inventory metrics including
moso bamboo distribution, area and culms quantity. The 3rd FMI data of Fujian was
originally produced in 2007, and is updated annually. In this study, the latest 3rd FMI
version from 2015 was used. Here we employed a stratified sampling procedure and
collected a total of 8949 points.
To avoid unnecessary computation and to limit our efforts to forests, we created a
non-forest mask following a similar procedure to Zhu, Woodcock, and Olofsson
(2012). Based on the fact that forests tend to show high NDVI values but low
reflectance in shortwave infrared (SWIR) bands, we masked an image pixel if it
violated at least one of the criteria listed below: (1) NDVI mean value greater than
0.6; and (2) SWIR band 7 reflectance mean value less than 0.1. Here the mean value
refers to the arithmetic average of all cloud-free observations acquired during the
study period (2014–2016).
2.3. Initial classification with different interpretation schemes
We tested three Landsat time series interpretation schemes in the study area (S1-S3 in
Table 1). For S1, we simply selected the Landsat scene with the least cloud coverage from
2015. For S2, we adopted the same full-season strategy as Li et al. (2016), which implemented the S1 method for each season from 2015, and obtained four image scenes as
classification model inputs. For S3, we employed a pixel-level compositing strategy following previous research by White et al. (2014). For each season during 2014–2016, a quality
score band based on two criteria (day of year score and distance to cloud score) were
calculated, and the pixels with the largest quality scores were included in the compositing
results. Here the purpose of using Landsat data from three years rather than one was to
examine whether a sufficiently large number of scenes is helpful to improve performance.
For each scheme, we further designed two experimental groups to investigate the
usefulness of gray-level co-occurrence matrix (GLCM) textures. To alleviate the potential
correlation issue, a principal component analysis was conducted on Landsat bands, then
Table 1. Landsat time series interpretation scheme designed in this study.
Scheme Group
Image time series utilization strategy
Only one image scene with the least cloud coverage during 2015 was used
Only one image scene with the least cloud coverage during 2015 was used, GLCMa textures were
also added
Four image scenes with the least cloud coverage from spring, summer, autumn and winter during
2015 were used
Four image scenes with the least cloud coverage from spring, summer, autumn and winter during
2015 were used, GLCM textures were also added
Composited image based on four scenes with the greatest quality scoreb from spring, summer,
autumn and winter during 2014–2016 were used
Composited image based on four scenes with the greatest quality score from spring, summer,
autumn and winter during 2014–2016 were used, GLCM textures were also added
Downloaded by [UNSW Library] at 06:22 26 October 2017
a. GLCM represents gray-level co-occurrence matrix. The derivation of GLCM textures in this study was based on Lu
et al. (2014) and Li et al. (2016).
b. The generation of a quality score was based on White et al. (2014). Sensor and opacity scores were not
calculated because only Landsat 8 data were used.
the first component was used to generate four types of GLCM metrics (homogeneity,
variance, contrast and entropy) with a moving square window size of 9 9, a direction
of 45° and a grey-level of 64 (Lu et al. 2014; Li et al. 2016).
The random forest classifier (RFC) was chosen to create land cover maps and evaluate
the influence of different schemes on map accuracy. RFC is an ensemble learning
method that produces a series of classification trees using bootstrap samples with
placement from training data (Breiman 2001). Due to reasonable prediction accuracy
and ability of handling high dimensional feature spaces, this classifier has been widely
used in large-area mapping and proven to be particularly effective when a large number
of features is involved (Gong et al. 2013). Here we set RFC with 500 trees and the
number of prediction variables equal to the square root of the number of total input
variables. A 50-fold cross validation was performed with the reference sample set. For
each classification, 80% of the reference samples were randomly selected for model
training, while the withheld 20% were used for accuracy assessment. This procedure was
repeated 50 times and the statistical results of the accuracy measurements were
adopted for the mapping result evaluation.
The initial classification performance was evaluated using two complementary
metrics: overall accuracy (OA) and minimum accuracy (MA). OA refers to the proportion
of correctly classified samples to the total sample set, while MA refers to the minimum of
user’s accuracy and producer’s accuracy (Zhu et al. 2016). Our primary aim was to find
the best interpretation scheme that could mostly elevate MA without sacrificing OA.
2.4. Auxiliary data selection and final classification
On the basis of the optimal interpretation scheme (i.e., highest OA and MA), the
importance of 15 auxiliary covariates were quantified by a mean increase in the classification error using RFC. We assumed an auxiliary covariate to be important and
necessary if its absence could elevate the error rate by over 1% for both classes (moso
bamboo/other forest plants) together or could elevate the error rate by over 1% for the
moso bamboo class alone. With the optimal Landsat time series interpretation scheme
and auxiliary covariates labeled as important, a final classification procedure was then
conducted with GEE platform to map moso bamboo in the study area.
3. Results and discussion
Downloaded by [UNSW Library] at 06:22 26 October 2017
3.1. Impact of image time series interpretation scheme
We found substantial differences in the initial classification results (Figure 2). This
disagreement revealed that different Landsat time series interpretation strategies can
have a large impact on moso bamboo estimation. Relative to the single-date image
selection (S1), the use of full-season images (S2) led considerable improvement in both
OA and MA. These patterns are consistent with research of Diao and Wang (2016), and
support the idea that including images from different months is more capable of
recovering from misclassifications and characterizing phenological trajectory (Li et al.
2016). Analysis results also showed further accuracy increases by S3 groups. We believe
these enhancements were mainly results of alleviating the influence of cloud as previous
studies suggested. Based on the same ‘best available pixel paradigm’ used in this study,
Hermosilla et al. (2016) reported the superiority of composited Landsat images for
broad-scale forest monitoring. Considering the severe cloud impact in our study area,
it is almost impossible to find clear Landsat scenes for all seasons, thus the compositing
strategy turned out to be a more desirable option.
According to our experiments, the inclusion of GLCM textures generally led to
increases in accuracy. However, the degree of accuracy improvement varied across
different schemes. For instance, in areas severely influenced by cloud, adding GLCM
textures as model inputs may cause greater uncertainty due to the missing information
of original imagery. This possibly explains the relatively small variations displayed by
groups of S2, as they were more prone to cloud influence.
3.2. Impact of auxiliary data
We also provided a quantitative evaluation of auxiliary data importance measured by
the mean increase in classification error (Figure 3). The exclusion of a certain covariate
caused a 0–2% error rate increase for both classes together, and a 0–7% error rate
Figure 2. Initial classification performance for minimum accuracy (upper row) and overall accuracy
(lower row) with different interpretation schemes.
Downloaded by [UNSW Library] at 06:22 26 October 2017
Figure 3. Auxiliary data importance measured by the mean increase in classification error. The black
lines suggest thresholds of 1% mean increase in error for the moso bamboo class alone (upper row)
and for both classes together (lower row). G_p: maximum NDVI; G_m: mean NDVI; G_v: minimum
NDVI; G_s: standard deviation of NDVI; W_p: maximum NDMI; W_m: mean NDMI; W_v: minimum
NDMI; W_s: standard deviation of NDMI; asp: aspect; ele: elevation; slp: slope; pre: precipitation; soil:
soil type; ta: temperature; cld_f: cloud frequency.
increase for the moso bamboo class alone. This importance inconsistence highlights the
complex suite of biotic and abiotic stresses that can affect moso bamboo distribution.
According to the evaluation result, five covariates (maximum NDVI, mean NDMI, minimum NDMI, elevation and temperature) met the importance criteria and were therefore
included for the final classification. Among all vegetative covariates, the mean and
minimum NDMI layers were the most helpful, which can be possibly explained by the
lower leaf water content of moso bamboo than other forest plants (Li et al. 2016).
Interestingly, it was the maximum NDVI rather than other NDVI metrics that passed the
importance test, revealing the most useful time window should be within the growing
seasons. Along with these biotic covariates, some abiotic covariates also bear special
mention. For example, the elevation layer was regarded as a necessary input in identifying moso bamboos, probably because they tend to occur in distinctly different elevation
landscape positions compared to other forest types (Li et al. 2016). Zhu et al. (2016)
reported a crucial role of cloud frequency in enhancing land cover classification.
However, in our study, this improvement was not significant. On the contrary, temperature was labelled as an important variable, indicating its major role of explaining the
spatial variation of moso bamboo compared to other climatic factors.
3.3. Final moso bamboo mapping with the optimal strategy
A definitive map of moso bamboo was generated by adopting the optimal strategy
(i.e., full-season compositing scheme and important auxiliary covariates). Pixel-level
statistical results showed that the final result yielded a higher prediction accuracy
Downloaded by [UNSW Library] at 06:22 26 October 2017
performance than any initial classification trials (93.21% in OA and 73.97% in MA),
implying the usefulness of combining dense Landsat time series and crucial auxiliary
data for moso bamboo identification. Based on the resultant map, we found moso
bamboos are generally found in most regions of Fujian. Spatially, the highest moso
bamboo densities exist in mountainous zones, and the moso bamboo density
decreases significantly from the northwest to the southeast coastline areas
(Figure 4).
Figure 4. Final moso bamboo mapping result of the study area. The scatter plot at the upper right
corner displays the comparison between reference and estimated results of moso bamboo area at
county-level (ha is an abbreviation for hectare).
At provincial-level, we noted a slight omission trend of our moso bamboo area prediction
when comparing it with the 3rd FMI data (0.98 million hectares vs. 1.08 million hectares).
This is maybe because the ‘threshold-based’ masking procedure in our study not only
filtered out non-forest regions, but also removed some isolated moso bamboo forests. To
examine the omission phenomenon on mapping performance, we further selected 44
counties with moso bamboo areas from 300 to 90,000 hectares and conducted a scatter
plot analysis. It can be found the linear regression relationship between our estimates and
3rd FMI data had a near 1:1 relationship, supported by its slope (0.97 close to 1) and a
coefficient of determination (R2) of 0.89. Therefore, we believe the final mapping result is
generally reasonable and can be used to identify moso bamboo distribution.
Downloaded by [UNSW Library] at 06:22 26 October 2017
4. Conclusions
This study comprehensively investigated different interpretation schemes for mapping
large-scale moso bamboo using dense Landsat time series, and provided a quantitative
importance evaluation of auxiliary data. Compared to the commonly used single-date or
sparse multi-date image selection approaches, it was found the adoption of full-season
image compositing strategy significantly enhanced the classification accuracy by incorporating phenological information and alleviating the influence of cloud. Additionally,
including GLCM textures as model inputs was generally beneficial, yet the degree of
precision increase varied with specific interpretation schemes. We also highlight the
contribution variability of auxiliary data, indicating that care should be given to the
selection of appropriate auxiliary covariates for moso bamboo identification. With the
full-season image compositing scheme and selected auxiliary covariates, the final classification exhibited a reasonable and improved performance (93.21% in OA and 73.97% in
MA). Our conclusions contribute to a better understanding of the roles of dense satellite
image time series and auxiliary data in identifying moso bamboo, and potentially serve
as a reference for other large-scale plant distribution mapping applications.
We thank the International Network for Bamboo and Rattan (INBAR) for offering valuable support.
This work was supported by National Natural Science Foundation of China [41601453], Natural
Science Foundation of Jiangxi Province of China [20161BAB213078] and Open Research
Foundation of Key Laboratory of Watershed Ecology and Geographical Environment Monitoring,
State Bureau of Surveying and Mapping of China [WE2016007].
Chong Liu
Breiman, L. 2001. “Random Forests.” Machine Learning 45 (1): 5–32. doi:10.1023/A:1010933404324.
Dannenberg, M. P., C. R. Hakkenberg, and C. Song. 2016. “Consistent Classification of Landsat Time
Series with an Improved Automatic Adaptive Signature Generalization Algorithm.” Remote
Sensing 8: 8. doi:10.3390/rs8080691.
Downloaded by [UNSW Library] at 06:22 26 October 2017
Diao, C., and L. Wang. 2016. “Incorporating Plant Phenological Trajectory in Exotic Saltcedar
Detection with Monthly Time Series of Landsat Imagery.” Remote Sensing of Environment 182
(September): 60–71. doi:10.1016/j.rse.2016.04.029.
Gong, P., J. Wang, L. Yu, Y. Zhao, Y. Zhao, L. Liang, Z. Niu, et al. 2013. “Finer Resolution Observation
and Monitoring of Global Land Cover: First Mapping Results with Landsat TM and ETM+ Data.”
International Journal of Remote Sensing 34 (7): 2607–2654. doi:10.1080/01431161.2012.748992.
Hermosilla, T., M. A. Wulder, J. C. White, N. C. Coops, G. W. Hobart, and L. B. Campbell. 2016. “Mass
Data Processing of Time Series Landsat Imagery: Pixels to Data Products for Forest Monitoring.”
International Journal of Digital Earth 9 (11): 1035–1054. doi:10.1080/17538947.2016.1187673.
Li, M., L. Congcong, H. Jiang, C. Fang, J. Yang, Z. Zhu, L. Shi, S. Liu, and P. Gong. 2016. “Tracking
Bamboo Dynamics in Zhejiang, China, Using Time-Series of Landsat Data from 1990 to 2014.”
International Journal of Remote Sensing 37 (7): 1714–1729. doi:10.1080/01431161.2016.1165885.
Li, P., G. Zhou, D. Huaqiang, L. Dengsheng, M. Lufeng, X. Xiaojun, Y. Shi, and Y. Zhou. 2015.
“Current and Potential Carbon Stocks in Moso Bamboo Forests in China.” Journal of
Environmental Management 156 (June): 89–96. doi:10.1016/j.jenvman.2015.03.030.
Lu, D., L. Guiying, E. Moran, L. Dutra, and M. Batistella. 2014. “The Roles of Textural Images in
Improving Land-Cover Classification in the Brazilian Amazon.” International Journal of Remote
Sensing 35 (24): 8188–8207. doi:10.1080/01431161.2014.980920.
Tucker, C. J. 1979. “Red and Photographic Infrared Linear Combinations for Monitoring
Vegetation.” Remote Sensing of Environment 8 (2): 127–150. doi:10.1016/0034-4257(79)90013-0.
Vermote, E., C. Justice, M. Claverie, and B. Franch. 2016. “Preliminary Analysis of the Performance of
the Landsat 8/OLI Land Surface Reflectance Product.” Landsat 8 Science Results 185 (November):
46–56. doi:10.1016/j.rse.2016.04.008.
White, J. C., M. A. Wulder, G. W. Hobart, J. E. Luther, T. Hermosilla, P. Griffiths, N. C. Coops, et al. 2014.
“Pixel-Based Image Compositing for Large-Area Dense Time Series Applications and Science.”
Canadian Journal of Remote Sensing 40 (3): 192–212. doi:10.1080/07038992.2014.945827.
Wilson, E. H., and S. A. Sader. 2002. “Detection of Forest Harvest Type Using Multiple Dates of Landsat
TM Imagery.” Remote Sensing of Environment 80 (3): 385–396. doi:10.1016/S0034-4257(01)00318-2.
Zhu, Z., A. L. Gallant, C. E. Woodcock, B. Pengra, P. Olofsson, T. R. Loveland, S. Jin, D. Dahal, L. Yang,
and R. F. Auch. 2016. “Optimizing Selection of Training and Auxiliary Data for Operational Land
Cover Classification for the LCMAP Initiative.” ISPRS Journal of Photogrammetry and Remote
Sensing 122 (December): 206–221. doi:10.1016/j.isprsjprs.2016.11.004.
Zhu, Z., and C. E. Woodcock. 2012. “Object-Based Cloud and Cloud Shadow Detection in Landsat
Imagery.” Remote Sensing of Environment 118 (March): 83–94. doi:10.1016/j.rse.2011.10.028.
Zhu, Z., C. E. Woodcock, and P. Olofsson. 2012. “Continuous Monitoring of Forest Disturbance Using All
Available Landsat Imagery.” Landsat Legacy Special Issue 122: 75–91. doi:10.1016/j.rse.2011.10.030.
Без категории
Размер файла
1 822 Кб
2017, 1378454, 2150704x
Пожаловаться на содержимое документа