Beruflich Dokumente
Kultur Dokumente
To cite this article: Haitao Wang, Yang Shao & Lisa M. Kennedy (2014) Temporal generalization
of sub-pixel vegetation mapping with multiple machine learning and atmospheric
correction algorithms, International Journal of Remote Sensing, 35:20, 7118-7135, DOI:
10.1080/01431161.2014.965288
Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
International Journal of Remote Sensing, 2014
Vol. 35, No. 20, 7118–7135, http://dx.doi.org/10.1080/01431161.2014.965288
1. Introduction
Remote sensing has long been used to characterize urbanization (e.g. Masek, Lindsay, and
Goward 2000; Weng 2001; Yuan et al. 2005) and revegetation (e.g. Richards 1984;
Runnström 2003) at local and regional scales. Urban land-cover mapping is challenging
because land-cover types and spatial structures in urban settings are complex and highly
dynamic. For example, vegetation and impervious surfaces are often mixed, particularly
when observed with medium-spatial-resolution (e.g. Landsat, 30 m) remote-sensing data
(Ridd 1995; Song 2005). The use of traditional per-pixel classification for vegetation/
impervious surface characterization thus can be problematic. Many researchers have applied
linear spectral mixture analysis (SMA) to estimate sub-pixel land-cover fraction for urban/
suburban areas (e.g. Smith et al. 1990; Elmore et al. 2000; Phinn et al. 2002; Powell et al.
2007). The sub-pixel vegetation–impervious surface–soil (VIS) model has been intensively
used to characterize biophysical composition of urban environments (Ridd 1995; Wu and
Murray 2003; Lu and Weng 2004). One of the main challenges for VIS models is the
spectral confusion between impervious cover and bare soil, and many researchers have
addressed uncertainties associated with impervious surfaces and bare agricultural lands (e.g.
Ridd 1995; Small 2003; Wu and Murray 2003; Small and Lu 2006). An alternate approach
is to focus on sub-pixel vegetation cover mapping only and use vegetation dynamics as a
surrogate to study urban change (Small and Lu 2006). This approach is appealing since
there is a high inverse relationship between vegetation and impervious surfaces and
vegetation surface could be mapped with a relatively higher accuracy.
More recently, machine learning algorithms have become increasingly popular in sub-
pixel mapping. These approaches typically use high-resolution imagery (e.g. QuickBird,
aerial photography) to build proportional sub-pixel vegetation cover for training data, then
apply machine learning algorithms to establish relationships between Landsat signals and
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
vegetation surfaces. The most commonly used machine learning algorithms for sub-pixel
mapping include artificial neural networks (Shao and Lunetta 2011; Shao, Taff, and Walsh
2011), regression trees (Yang et al. 2003), random forests (Walton 2008; Gessner et al. 2013),
and support vector machines (Brown, Gunn, and Lewis 1999; Esch et al. 2009). For example,
the 30-m resolution percentage of tree canopy map, a part of the US National Land Cover
Database product, was developed using a regression tree algorithm (Homer et al. 2004). For a
given sub-pixel mapping task, the choice of specific machine learning algorithm is largely
dependent on a user’s preference and knowledge. To obtain optimal sub-pixel map products,
multiple machine learning algorithms often need to be implemented and compared for best
generalization performance (e.g. Shao and Lunetta 2012).
For fast-changing urban/suburban areas, an image temporal frequency of approxi-
mately 3–5 years may be used to adequately capture vegetation dynamics (Lunetta et al.
2004). However, sub-pixel mapping of multiple Landsat images across time has not been
routinely implemented. One of the main reasons includes lack of training/validation data
(e.g. high-resolution aerial photographs) for multiple time periods. A potential solution is
to develop a sub-pixel classification algorithm that could be generalized across time. For
example, an image classification algorithm developed for year i could be directly applied
to year j to obtain a land-cover map. In per-pixel classification, temporal generalization is
always an intriguing (but challenging) topic since it allows monitoring land cover at
frequent time intervals without using the traditional ‘image-by-image’ analytical approach
(Pax-Lenney et al. 2001; Woodcock et al. 2001). With free access to the entire Landsat
archive, temporal generalization of classification algorithms will become even more
valuable for operational land-cover mapping and monitoring (Hansen and Loveland
2012). To date, few studies have thoroughly examined the temporal generalization
potential for selected sub-pixel mapping algorithms.
Atmospheric correction or image normalization plays a critical role for the purpose of
temporal generalization since consistent radiometric measurement levels are required
across time (Pax-Lenney et al. 2001; Song et al. 2001; Woodcock et al. 2001). For
Landsat data, advanced atmospheric correction algorithms have been integrated by the
Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) (Vermote and
Saleous 2007) to provide ready-to-use surface reflectance images. LEDAPS uses dense
vegetation targets as dart objects within a given Landsat scene to extract aerosol optical
thickness as input to the 6S radiative transfer model correction (Masek et al. 2006).
Atmospheric conditions such as ozone and water vapour are also considered in the radiative
transfer model. LEDAPS surface reflectance products are now increasingly used and show
high potential for large area mapping and change detection applications (e.g. Masek et al.
2008; Hansen and Loveland 2012).
7120 H. Wang et al.
The impact of LEDAPS products on sub-pixel vegetation mapping, however, has not
been evaluated, especially for arid and semi-arid regions where vegetation canopies are
sparse and an automated atmospheric correction algorithm may not be robust. In areas
where LEDAPS surface reflectance products do not provide consistent temporal general-
ization of sub-pixel mapping, additional atmospheric correction methods may be needed.
Furthermore, the validation of sub-pixel vegetation mapping products for multiple years
presents a significant challenge, especially for study areas where historical aerial photo-
graph coverage is unavailable. Even with historical aerial photographs, it is often time-
consuming and expensive to generate reliable vegetation maps as reference for validation
(e.g. Yang et al. 2003). An alternate approach clearly warrants exploration.
The overall goal of this study was to examine temporal generalization of sub-pixel
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
vegetation mapping with multiple atmospheric correction and machine learning algo-
rithms. We examined two atmospheric correction methods (dark object subtraction
(DOS) and LEDAPS algorithm), three machine learning algorithms (random forests,
neural networks, and classification and regression trees), and their varying combinations
for sub-pixel vegetation mapping across time. These analytical strategies are potentially
important to researchers attempting to monitor vegetation dynamics using a rich archive
of Landsat imagery. We used Zhongwei, northwest China, as an experiment site to test the
accuracy of sub-pixel vegetation mapping methods. Zhongwei is representative of small-
and medium-sized cities in the semi-arid region of northwest China where urban expan-
sion and revegetation processes interact to shape landscape patterns and functions.
2. Study area
Zhongwei City is located in Ningxia Hui Autonomous Region (Ningxia), northwestern
China (Figure 1). The city lies at the edge of an oasis and abuts the southeastern frontier
of the Tengger Desert, the fourth largest desert in China. The Yellow River traverses the
irrigated plain of the city and provides a critical water source for agriculture production.
Zhongwei City was a linchpin of the Silk Road in ancient times. Now it is a hub of
national transportation networks. For example, Lanzhou–Baotou Railway, the nation’s
first desert railway, passes through the city. Zhongwei covers a total area of 16,986 km2.
Over 90% of areas are mountains, loess hilly region, and desert (Wu 2013). Centuries of
overgrazing and forest destruction have caused severe land degradation in this area. In
recent decades, many rehabilitation projects have been implemented in the periphery of
city to restrain ecological degradation (Feng, Fan, and Lu 1998; Wang et al. 2007).
Our study focused on the city municipal district. During the study period, the
population of Zhongwei increased from 289,527 to 335,374 between 1990 and 2000
and to 383,014 by 2012 (Statistical Bureau of Ningxia 2012).
3. Data
We searched Landsat archives including USGS GLOVIS (http://glovis.usgs.gov/) and
EarthExplorer (http://earthexplorer.usgs.gov) for cloud-free images. Four Landsat 5 TM
images (1990, 1996, 2004, and 2010; path 130, row 34) were obtained. The images were
acquired 6–8 years apart (due to availability) over the 20-year period. All images were
acquired in summer (late June to early August), thus phenological variations among the
four images are small for our study area. For each Landsat scene, we created an image
subset (1036 km2) to cover the urban core and its surrounding deserts. All the images
were registered to the 2004 master image through image-to-image registration.
International Journal of Remote Sensing 7121
37° 30′ N
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
Zhongwei City
Study Area
Yellow River
0 5 10 20
km
Figure 1. Zhongwei City (study area) in northwestern China. A Landsat 5 image is used as the
background in the top panel.
We obtained two high-spatial resolution OrbView-3 satellite images from the USGS
EROS Data Center. One OrbView-3 multispectral image was acquired on 20 June 2004 at
4 m spatial resolution. It has four spectral bands (visible and infrared) that are of
comparable spectral resolution to Landsat imagery (OrbImage, Inc. 2005). The other
panchromatic image was acquired on 4 March 2004 at 1 m spatial resolution. The spatial
extents of the multispectral and panchromatic images are approximately 8 km × 25 km
and 8 km × 35 km, respectively. The coverages of these two images run through
Zhongwei City from north to south and overlap at the urban core area (Figure 2). Both
images are high quality and cloud-free.
4. Methods
4.1. Atmospheric correction and LEDAPS surface reflectance products
For image classification involving algorithm generalization across time, atmospheric
correction is required to normalize radiometric measures for images acquired from
different years (Song et al. 2001). A number of atmospheric correction procedures have
been developed: DOS approaches (Chavez 1996), the histogram matching using pseudo-
7122 H. Wang et al.
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
Figure 2. (a) Overlapping spatial regions of the study area and OrbView-3 images, (b) a spatial
subset of the Landsat image, (c) OrbView-3 multispectral image, and (d) OrbView-3 panchromatic
image.
invariant features (Collins and Woodcock 1996), the linear regression or Ridge method
(Song et al. 2001), and advanced radiative transfer model corrections (Gemmell, Varjo,
and Strandstrom 2001).
In per-pixel image classification and change detection studies, Song et al. (2001)
found that simple atmospheric correction methods such as DOS outperformed more
advanced atmospheric correction algorithms (e.g. radiative transfer corrections). Similar
results have also been reported by other researchers (e.g. Gemmell, Varjo, and
Strandstrom 2001; Pax-Lenney et al. 2001). The main argument is that the success of
classification generalization depends on an equivalence of radiometric measurement levels
among time-series remote-sensing data, rather than correction to standard reflectance units
for each image.
Based on these previous findings, we implemented a simple DOS correction method
to normalize Landsat radiometric measurements across time. Song et al. (2001) summar-
ized four specific DOS approaches involving different assumptions of atmospheric trans-
mittance and diffuse irradiance. We followed their DOS1 approach (Equation (1)) to
retrieve surface reflectance (ρ) (Moran et al. 1992; Song et al. 2001):
π Lsat Lp
ρ¼ ; (1)
ðE0 cosðθz ÞÞ
where Lsat is the at-satellite radiance, Lp is the path radiance, E0 is the exoatmospheric
solar constant, and θz is the solar zenith angle. DOS1 assumes no atmospheric transmit-
tance loss and no downwelling diffuse irradiance at the surface. Path radiance (Lp) was
estimated using a minimum digital number (DN with at least 1000 pixels) for each image
International Journal of Remote Sensing 7123
(Teillet and Fedosejevs 1995; Chavez 1996). DOS1 corrections were applied to all four
Landsat images.
In addition to the simple DOS1 approach, we also evaluated LEDAPS-derived surface
reflectance values for sub-pixel vegetation mapping and temporal generalization.
LEDAPS is considered a state-of-the-art atmospheric correction method that is already
operational for Landsat 5 and 7 data (Masek et al. 2006). We downloaded LEDAPS-
derived surface reflectance images for 2004 and 2010 from USGS EarthExplorer. The
other two images were not available in LEDAPS-corrected form, thus we locally imple-
mented the LEDAPS code to derive surface reflectance values.
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
and regression training) package in R (Kuhn 2008). The caret package incorporated
numerous machine learning algorithms such as random Forests (RF), Classification and
Regression Trees (CART), Neural Networks (NN), and Support Vector Machines (SVM).
Data preparation, parameter adjusting, and model training processes are streamlined and
easy to use. We selected RF, CART, and NN for our sub-pixel mapping task due to their
popularity in the remote-sensing research community. For each algorithm, caret automa-
tically adjusts multiple tuning parameters in searching for optimal prediction. For exam-
ple, using RF, the number of classification trees (e.g. ntree = 500, 1000, and 1500) and the
number of random split variables used in each node (e.g. mtry = 2, 3, and 5) need to be
adjusted. Cross-validation is used to select the best set of parameters for final prediction.
To evaluate the impacts of training sample size on the classification performance, we
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
generated a number of training data sets with varying sample sizes: 100, 200, 400, 600, 800,
1000, 2000, 3000, 4000, and 5000. At each sample size, 20 repeated samplings were
conducted to evaluate the variability. For each data set, a fivefold cross-validation was used
for model training and validation. The accuracies of sub-pixel models were assessed using a
root mean square error (RMSE) statistical measure. RMSE is calculated as Equation (2):
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
i¼1 ðEi Ai Þ
RMSE ¼ ; (2)
n
where Ei is the predicted vegetation proportion from the RF model, Ai is the actual
vegetation proportion from the reference vegetation map generated from the OrbView-3
image, and n is the number of sample points used for validation. The accuracy levels of
caret models were compared against training/validation sample sizes.
period. The patterns of vegetation changes inside/outside of the city boundary were
quantified separately for different time periods.
5. Results
The accuracy of OrbView-3 image classification can have large impacts on the development
of a 30 m vegetation proportional map and subsequent sub-pixel vegetation mapping using
Landsat images. Per-pixel (4 m spatial resolution) classification of an Orbview-3 image needs
to have sufficient accuracy to represent the vegetation distribution. The overall accuracy of
classification is about 71%, and Cohen’s kappa is 0.67 (Table 2), suggesting moderate
classification accuracy. Among all eight land-cover classes, the high turbidity water surface
(the Yellow River) has the highest accuracy (99%). Meanwhile, there were strong spectral
confusions between desert, barren, and urban surfaces, as reflected by the column data and
producer’s accuracy of the desert (61%), barren (59%), and low albedo impervious surface
(43%) categories. The low albedo impervious surface category also has very low user’s
accuracy (30%). Among the 100 randomly selected low albedo impervious surface pixels,
only 30 pixels were correctly labelled. Another 30 and 32 pixels actually belong to the desert
and barren class, respectively. These spectral confusions indicate a potential challenge in
extracting urban impervious surfaces.
When the eight land-cover classes were grouped into two broad cover types (vegetation
and non-vegetation), the overall accuracy increased to 93.8% (kappa = 0.84). The producer’s
and user’s accuracies for vegetation cover were 81.0% and 98.8%. This level of accuracy was
sufficient to generate a proportional vegetation map at a coarser spatial scale (i.e. 30 m) to
serve as training and validation data for Landsat-based sub-pixel vegetation mapping.
7126
Reference
Figure 3. Classification accuracies for three sub-pixel mapping algorithms. RMSE values are
compared at a range of training data sample sizes. RMSE unit is per cent in decimal form.
NN and RF outperformed the CART algorithm at almost all training sample sizes.
Between the NN and RF algorithms, it appeared that RF performed better when relatively
small numbers of training sample points were used (e.g. <600).
In a comparison of the fractional vegetation map generated from Orbview-3 and the
Landsat image, scatterplots were generated for 30 m (Figure 4(a)) and 90 m (Figure 4(b))
spatial resolutions. We focused on the vegetation fraction map derived from the NN
algorithm since it provided the best overall performance. A large degree of scatter of the
data points is apparent at 30 m pixel comparison. If the Orbview-3 derived fractional
vegetation map is considered as the true vegetation distribution at the ground, the Landsat-
estimated sub-pixel vegetation map may contain as much as ±30% error for pixels with
intermediate level (e.g. 50%) of vegetation fractions, although the overall accuracy
(RMSE = 0.099) is still acceptable due to good classification performance for many
pixels with extremely high/low (e.g. 100% or 0%) vegetation cover. One possible reason
for the large uncertainty is the possible error of co-registration and the direct comparison
at 30 m spatial resolution would then not be recommended (Song 2005). When the
fractional vegetation maps were aggregated and compared at 90 m spatial resolution,
the level of point scattering was largely reduced and the RMSE value decreased to 0.051.
We note that the scatter plots fall in the vicinity of a 1:1 line, which is appealing for
unbiased estimation of total vegetation area for the study region.
of the 2010 image classification. The NN sub-pixel classifier, combined with LEDAPS
atmospheric correction, resulted in the lowest RMSE value (0.093) (Figure 5). The NN
classifier combined with DOS1 approach achieved similar accuracy level (RMSE = 0.096;
R2 = 0.94). The scatter plot and intercept of the regression line suggested that LEDAPS-
based atmospheric correction may lead to slight overestimation of sub-pixel vegetation
proportions, especially for areas with lower (i.e. <20%) vegetation fraction. Overall,
accuracy statistics and scatter plots indicated consistent predictions of vegetation fractions
for these invariant pixels, suggesting a very good generalizability of NN sub-pixel
classification algorithm across time.
The RF algorithm showed acceptable but relatively worse generalization performance
compared to the NN algorithm. Relatively larger point scattering leads to higher RMSE
values (~0.120) for both DOS1 and LEDAPS atmospherically corrected data. R2 values
(~0.90) were also lower than the values obtained from the NN classifier (~0.94). Similar
to the NN classifier, LEDAPS-based atmospheric correction resulted in more overestima-
tion of vegetation cover compared to the simple DOS1 approach. A visual interpretation
of the resultant vegetation maps and summary statistics confirmed this finding. For the
2010 sub-pixel vegetation maps, the total vegetation area derived from LEDAPS products
was about 7% higher than DOS1-based map product.
CART performed worst in terms of temporal generalization. RMSE values were 0.152 and
0.193 for DOS1 and LEDAPS atmospherically corrected images, respectively. Corresponding
R2 values decreased to 0.87 and 0.77. The scatter plots, accuracy statistics, and regression lines
suggested that CART did not generate accurate vegetation map products across time, thus
cross-time comparison of vegetation coverage could lead to unreliable results.
Figure 5. Comparison (2004 vs. 2010) of vegetation fraction for pseudo-invariant pixels: (a)
DOS1 atmospheric correction combined with NN algorithm, (b) DOS1 atmospheric correction
combined with RF algorithm, (c) DOS1 atmospheric correction combined with CART algorithm,
(d) LEDAPS atmospheric correction combined with NN algorithm, (e) LEDAPS atmospheric
correction combined with RF algorithm, and (f) LEDAPS atmospheric correction combined with
CART algorithm. The solid line and dash line represent 1:1 and regression line, respectively.
this time provided a clear boundary differentiating urban/human settlement and deserts.
We applied a threshold value (e.g. >50% vegetation cover per pixel) to the 1990 vegeta-
tion map to generate an arbitrary urban/human settlement boundary (Figure 7). The
vegetation changes within and beyond the boundary resulted from two different human
activities: urbanization and revegetation.
For each sub-pixel vegetation map, we summarized vegetation cover fractions for
areas inside and outside the boundary and assessed trends of vegetation change patterns.
Within the urban boundary, vegetation cover decreased over the study period (1990–2010)
by 15.3% (217.15 km2 in 1990 and 183.83 km2 in 2010) (Table 3). Outside the boundary,
vegetation cover increased by 78.5% (45.61 km2 in 1990 and 81.43 km2 in 2010) over the
same period. The revegetation process outside the city mainly started in the mid-1990s.
From 1990 to 1996, 16.75 km2 of barren desert was converted into vegetation cover
around the city fringe. From 2004 to 2010, an area of 16.93 km2 was revegetated, mostly
in the northwestern industrial zone of the city.
While the changes in land cover varied spatially over the study period, the total
vegetated areas in the entire study region increased initially (1990–1996), and then
slightly decreased from 1996 to 2010. We derived three maps illustrating vegetation
cover change for three time periods (1990–1996, 1996–2004, and 2004–2010) at sub-
pixel level. The contrasting colors indicate vegetation increase/decrease within each time
period (Figure 8). Within the city boundary, the trajectory of vegetation change shows a
clear tendency of urban development. The most dramatic vegetation change occurred
7130 H. Wang et al.
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
Figure 6. Predicted vegetation fraction for 1990, 1996, 2004, and 2010 Landsat images.
during the period 2004–2010, with northward urban expansion toward the Yellow River
and the emergence of new road networks across the city and along the Yellow River.
6. Discussion
We observed some important advantages of the caret package for sub-pixel mapping or
remote-sensing image classification in general. The caret package provides easy access to
a large number of machine learning and statistical data analysis methods. More impor-
tantly, it streamlines data pre-processing, parameter adjusting, and model selection (Kuhn
2008). This provides an efficient implementation and comparison mechanism for multiple
sub-pixel mapping algorithms. Comparison of multiple mapping algorithms for a given
remote-sensing classification problem should be routinely conducted since there is no
universally optimal algorithm.
Among the three selected sub-pixel mapping algorithms, we found that both the NN
and RF algorithms performed better than CART for 2004 sub-pixel mapping. In per-pixel
applications, previous studies showed superior performance from NN if sufficient training
sample points are available (Shao and Lunetta 2012). The RF algorithm is now widely
used in the remote-sensing community due to its ease of use and good generalization
ability (Walton 2008). The superior performance of RF compared to CART was expected
since it takes an ensemble approach to reduce variance in learning (Breiman 2001).
International Journal of Remote Sensing 7131
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
Figure 8. Change of sub-pixel vegetation fraction for three time periods. Areas in red suggest
increase of vegetation cover. Areas in blue indicate decrease of vegetation cover.
7132 H. Wang et al.
For the NN and RF algorithms, a relatively small number (~3000) of training samples
were needed to achieve strong classification performance. For urban core and suburban
areas, vegetation and impervious cover are often mixed within the Landsat pixels. In the
urban–desert interface in Zhongwei City, revegetation programmes have been implemen-
ted, but newly developed vegetation cover (trees or artificial grasslands) could be sparse
and mixed with the desert background. We found fractional vegetation cover could be
estimated with high accuracy using the NN and RF algorithms even though it mixed with
different cover types (e.g. impervious surfaces and sand in desert).
The choice of specific sub-pixel classification algorithm appeared to be important for
temporal generalization. Combined with atmospheric correction methods such as DOS1
and LEDAPS, the NN algorithm showed impressive generalization capability across time.
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
7. Conclusion
We examined temporal generalization of sub-pixel vegetation mapping using multiple
machine learning (NN, RF, and CART) and atmospheric correction methods (DOS1 and
LEDAPS). We first used high-resolution OrbView-3 images to generate vegetation frac-
tions to support sub-pixel mapping of a 2004 Landsat image. Cross-validation resulted in
RMSE values of 0.099, 0.100, and 0.123 for NN, RF, and CART, respectively. The trained
International Journal of Remote Sensing 7133
classifiers were then directly applied to Landsat images from 1990, 1996, and 2010 to
examine temporal generalizability. The NN algorithm, combined with simple DOS1 or
LEDAPS surface reflectance products, showed best generalization capability for sub-pixel
vegetation mapping. Predicted sub-pixel vegetation proportions were consistent for
pseudo-invariant pixels (RMSE ~ 0.093). The RF algorithm performed the second best
(RMSE ~ 0.120), while CART generated the worst temporal generalization results (e.g.
RMSE = 0.193 for LEDAPS products). Our overall results suggested that machine
learning algorithms, combined with appropriate atmospheric correction methods, have
high potential for temporal generalization of sub-pixel vegetation mapping in arid and
sub-arid urban environments.
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
References
Breiman, L. 2001. “Random Forests.” Machine Learning 45: 5–32. doi:10.1023/A:1010933404324.
Brown, M., S. R. Gunn, and H. G. Lewis. 1999. “Support Vector Machines for Optimal
Classification and Spectral Unmixing.” Ecological Modelling 120: 167–179. doi:10.1016/
S0304-3800(99)00100-3.
Cao, S. 2008. “Why Large-Scale Afforestation Efforts in China Have Failed to Solve the
Desertification Problem.” Environmental Science & Technology 42: 1826–1831. doi:10.1021/
es0870597.
Chavez, P. S. 1996. “Image-Based Atmospheric Corrections-Revisited and Improved.”
Photogrammetric Engineering and Remote Sensing 62: 1025–1035.
Collins, J. B., and C. E. Woodcock. 1996. “An Assessment of Several Linear Change Detection
Techniques for Mapping Forest Mortality Using Multitemporal Landsat TM Data.” Remote
Sensing of Environment 56: 66–77. doi:10.1016/0034-4257(95)00233-2.
Congalton, R. G. 1991. “A Review of Assessing the Accuracy of Classifications of Remotely
Sensed Data.” Remote Sensing of Environment 37: 35–46. doi:10.1016/0034-4257(91)90048-B.
Elmore, A. J., J. F. Mustard, S. J. Manning, and D. B. Lobell. 2000. “Quantifying Vegetation
Change in Semiarid Environments: Precision and Accuracy of Spectral Mixture Analysis and
the Normalized Difference Vegetation Index.” Remote Sensing of Environment 73: 87–102.
doi:10.1016/S0034-4257(00)00100-0.
Esch, T., V. Himmler, G. Schorcht, M. Thiel, T. Wehrmann, F. Bachofer, C. Conrad, M. Schmidt,
and S. Dech. 2009. “Large-Area Assessment of Impervious Surface Based on Integrated
Analysis of Single-Date Landsat-7 Images and Geospatial Vector Data.” Remote Sensing of
Environment 113: 1678–1690. doi:10.1016/j.rse.2009.03.012.
Feng, W., X. Fan, and T. Lu, ed. 1998. Zhongwei Yearbook 1994-1995 [Chinese Edition]. Yinchuan:
Ningxia People’s.
Gemmell, F., J. Varjo, and M. Strandstrom. 2001. “Estimating Forest Cover in a Boreal Forest Test
Site Using Thematic Mapper Data from Two Dates.” Remote Sensing of Environment 77: 197–
211. doi:10.1016/S0034-4257(01)00206-1.
Gessner, U., M. Machwitz, C. Conrad, and S. Dech. 2013. “Estimating the Fractional Cover of
Growth Forms and Bare Surface in Savannas. A Multi-Resolution Approach Based on
Regression Tree Ensembles.” Remote Sensing of Environment 129: 90–102. doi:10.1016/j.
rse.2012.10.026.
Hansen, M. C., and T. R. Loveland. 2012. “A Review of Large Area Monitoring of Land Cover
Change Using Landsat Data.” Remote Sensing of Environment 122: 66–74. doi:10.1016/j.
rse.2011.08.024.
Homer, C., C. Huang, L. Yang, B. Wylie, and M. Coan. 2004. “Development of a 2001 National
Land-Cover Database for the United States.” Photogrammetric Engineering & Remote Sensing
70: 829–840. doi:10.14358/PERS.70.7.829.
Kuhn, M. 2008. “Building Predictive Models in R Using the Caret Package.” Journal of Statistical
Software 28: 1–26.
Lu, D., and Q. Weng. 2004. “Spectral Mixture Analysis of the Urban Landscape in Indianapolis
with Landsat ETM+ Imagery.” Photogrammetric Engineering and Remote Sensing 70: 1053–
1062. doi:10.14358/PERS.70.9.1053.
7134 H. Wang et al.
Lunetta, R. S., D. M. Johnson, J. G. Lyon, and J. Crotwell. 2004. “Impacts of Imagery Temporal
Frequency on Land-Cover Change Detection Monitoring.” Remote Sensing of Environment 89:
444–454. doi:10.1016/j.rse.2003.10.022.
Masek, J., F. Lindsay, and S. Goward. 2000. “Dynamics of Urban Growth in the Washington DC
Metropolitan Area, 1973–1996, from Landsat Observations.” International Journal of Remote
Sensing 21: 3473–3486. doi:10.1080/014311600750037507.
Masek, J. G., C. Huang, R. Wolfe, W. Cohen, F. Hall, J. Kutler, and P. Nelson. 2008. “North
American Forest Disturbance Mapped from a Decadal Landsat Record.” Remote Sensing of
Environment 112: 2914–2926. doi:10.1016/j.rse.2008.02.010.
Masek, J. G., E. F. Vermote, N. E. Saleous, R. Wolfe, F. G. Hall, K. F. Huemmrich, F. Gao, J. Kutler,
and T.-K. Lim. 2006. “A Landsat Surface Reflectance Dataset for North America, 1990-2000.”
Geoscience and Remote Sensing Letters, IEEE 3: 68–72.
Moran, M. S., R. D. Jackson, P. N. Slater, and P. M. Teillet. 1992. “Evaluation of Simplified
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
Procedures for Retrieval of Land Surface Reflectance Factors from Satellite Sensor Output.”
Remote Sensing of Environment 41: 169–184. doi:10.1016/0034-4257(92)90076-V.
OrbImage, Inc. 2005. OrbView-3 Commercial Satellite Image Product Catalog. Dulles, VA:
OrbImage, Inc.
Pax-Lenney, M., C. E. Woodcock, S. A. Macomber, S. Gopal, and C. Song. 2001. “Forest Mapping
with a Generalized Classifier and Landsat TM Data.” Remote Sensing of Environment 77: 241–
250. doi:10.1016/S0034-4257(01)00208-5.
Phinn, S., M. Stanford, P. Scarth, A. Murray, and P. Shyy. 2002. “Monitoring the Composition of
Urban Environments Based on the Vegetation-Impervious Surface-Soil (VIS) Model by
Subpixel Analysis Techniques.” International Journal of Remote Sensing 23: 4131–4153.
doi:10.1080/01431160110114998.
Powell, R. L., D. A. Roberts, P. E. Dennison, and L. L. Hess. 2007. “Sub-Pixel Mapping of Urban
Land Cover Using Multiple Endmember Spectral Mixture Analysis: Manaus, Brazil.” Remote
Sensing of Environment 106: 253–267. doi:10.1016/j.rse.2006.09.005.
Richards, J. 1984. “Thematic Mapping from Multitemporal Image Data Using the Principal
Components Transformation.” Remote Sensing of Environment 16: 35–46. doi:10.1016/0034-
4257(84)90025-7.
Ridd, M. K. 1995. “Exploring a VIS (Vegetation-Impervious Surface-Soil) Model for Urban
Ecosystem Analysis through Remote Sensing: Comparative Anatomy for Cities.”
International Journal of Remote Sensing 16: 2165–2185. doi:10.1080/01431169508954549.
Runnström, M. 2003. “Rangeland Development of the Mu Us Sandy Land in Semiarid China: An
Analysis Using Landsat and NOAA Remote Sensing Data.” Land Degradation & Development
14: 189–202. doi:10.1002/ldr.545.
Schneider, A., K. C. Seto, and D. R. Webster. 2005. “Urban Growth in Chengdu, Western China:
Application of Remote Sensing to Assess Planning and Policy Outcomes.” Environment and
Planning B: Planning and Design 32: 323–345. doi:10.1068/b31142.
Shao, Y., and R. S. Lunetta. 2011. “Sub-Pixel Mapping of Tree Canopy, Impervious Surfaces, and
Cropland in the Laurentian Great Lakes Basin Using MODIS Time-Series Data.” IEEE Journal
of Selected Topics in Applied Earth Observations and Remote Sensing 4: 336–347. doi:10.1109/
JSTARS.2010.2062173.
Shao, Y., and R. S. Lunetta. 2012. “Comparison of Support Vector Machine, Neural Network, and
CART Algorithms for the Land-Cover Classification Using Limited Training Data Points.”
ISPRS Journal of Photogrammetry and Remote Sensing 70: 78–87. doi:10.1016/j.
isprsjprs.2012.04.001.
Shao, Y., G. N. Taff, and S. J. Walsh. 2011. “Comparison of Early Stopping Criteria for Neural-
Network-Based Subpixel Classification.” IEEE Geoscience and Remote Sensing Letters 8: 113–
117. doi:10.1109/LGRS.2010.2052782.
Small, C. 2003. “High Spatial Resolution Spectral Mixture Analysis of Urban Reflectance.” Remote
Sensing of Environment 88: 170–186. doi:10.1016/j.rse.2003.04.008.
Small, C., and J. W. Lu. 2006. “Estimation and Vicarious Validation of Urban Vegetation
Abundance by Spectral Mixture Analysis.” Remote Sensing of Environment 100: 441–456.
doi:10.1016/j.rse.2005.10.023.
Smith, M. O., S. L. Ustin, J. B. Adams, and A. R. Gillespie. 1990. “Vegetation in Deserts: I. A
Regional Measure of Abundance from Multispectral Images.” Remote Sensing of Environment
31: 1–26. doi:10.1016/0034-4257(90)90074-V.
International Journal of Remote Sensing 7135
Song, C. 2005. “Spectral Mixture Analysis for Subpixel Vegetation Fractions in the Urban
Environment: How to Incorporate Endmember Variability?” Remote Sensing of Environment
95: 248–263. doi:10.1016/j.rse.2005.01.002.
Song, C., C. E. Woodcock, K. C. Seto, M. P. Lenney, and S. A. Macomber. 2001. “Classification
and Change Detection Using Landsat TM Data: When and How to Correct Atmospheric
Effects?” Remote Sensing of Environment 75: 230–244. doi:10.1016/S0034-4257(00)00169-3.
Statistical Bureau of Ningxia. 2012. “Ningxia Statistical Yearbook – Population.” http://www.nxtj.
gov.cn/nxtjjxbww/tjsj/ndsj/2012/indexch.htm.
Teillet, P., and G. Fedosejevs. 1995. “On the Dark Target Approach to Atmospheric Correction of
Remotely Sensed Data.” Canadian Journal of Remote Sensing 21: 374–387. doi:10.1080/
07038992.1995.10855161.
Vermote, E., and N. Saleous. 2007. LEDAPS Surface Reflectance Product Description, 21 p.
College Park, MD: University of Maryland Department of Geography. ledaps.nascom.nasa.
Downloaded by [Virginia Tech Libraries], [Haitao Wang] at 15:59 25 October 2014
gov/ docs/pdf/SR_productdescript_dec06.pdf.
Walton, J. T. 2008. “Subpixel Urban Land Cover Estimation: Comparing Cubist, Random Forests,
and Support Vector Regression.” Photogrammetric Engineering and Remote Sensing 74: 1213–
1222. doi:10.14358/PERS.74.10.1213.
Wang, S., Y. Ma, Q. Hou, and Y. Wang. 2007. “Coping Strategies with Desertification in China.” In
Managing Weather and Climate Risks in Agriculture, edited by M. V. Sivakumar and R. P.
Motha, 317–341. Heidelberg: Springer.
Weng, Q. 2001. “A Remote Sensing-GIS Evaluation of Urban Expansion and Its Impact on Surface
Temperature in the Zhujiang Delta, China.” International Journal of Remote Sensing 22: 1999–
2014.
Woodcock, C. E., S. A. Macomber, M. Pax-Lenney, and W. B. Cohen. 2001. “Monitoring Large
Areas for Forest Change Using Landsat: Generalization across Space, Time and Landsat
Sensors.” Remote Sensing of Environment 78: 194–203. doi:10.1016/S0034-4257(01)00259-0.
Wu, C., and A. T. Murray. 2003. “Estimating Impervious Surface Distribution by Spectral Mixture
Analysis.” Remote Sensing of Environment 84: 493–505. doi:10.1016/S0034-4257(02)00136-0.
Wu, J., ed. 2013. Zhongwei Yearbook 2012 [Chinese Edition]. Yinchuan: Ningxia People’s.
Yang, L., C. Huang, C. G. Homer, B. K. Wylie, and M. J. Coan. 2003. “An Approach for Mapping
Large-Area Impervious Surfaces: Synergistic Use of Landsat-7 ETM+ and High Spatial
Resolution Imagery.” Canadian Journal of Remote Sensing 29: 230–240. doi:10.5589/m02-098.
Yuan, F., K. E. Sawaya, B. C. Loeffelholz, and M. E. Bauer. 2005. “Land Cover Classification and
Change Analysis of the Twin Cities (Minnesota) Metropolitan Area by Multitemporal Landsat
Remote Sensing.” Remote Sensing of Environment 98: 317–328. doi:10.1016/j.rse.2005.08.006.