Remote Sensing of Environment 124 (2012) 689–704

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Monitoring land cover change in urban and peri-urban areas using dense time stacks
of Landsat satellite data and a data mining approach
Annemarie Schneider ⁎
Center for Sustainability and the Global Environment, Nelson Institute for Environmental Studies, University of Wisconsin-Madison, 1710 University Avenue, Madison, WI 53726, United States

a r t i c l e i n f o a b s t r a c t

Article history: Given the pace and scale of urban expansion in many parts of the globe, urban environments are playing an in-
Received 6 August 2011 creasingly important role in daily quality-of-life issues, ecological processes, climate, material flows, and land
Received in revised form 28 May 2012 transformations. Remote sensing has emerged as a powerful tool to monitor rates and patterns of urban expan-
Accepted 3 June 2012
sion, but many early challenges – such as distinguishing new urban land from bare ground – remain unsolved.
Available online 20 July 2012
To deal with the high temporal and spatial variability as well as complex, multi-signature classes within settle-
Keywords:
ments, this paper presents a new approach that exploits multi-seasonal information in dense time stacks of
Urban areas Landsat imagery using a multi-date composite change detection technique. The central premise of the approach
Urbanization is that lands within/near urban areas have distinct temporal trajectories both before and after change occurs, and
Cities that these lead to characteristic temporal signatures in several spectral regions. The method relies on a supervised
Peri-urban classification that exploits training data of stable/changed areas interpreted from Google Earth images, and a
Land cover ‘brute force’ approach of providing all available Landsat data as input, including scenes with data gaps due to
Environment the Scan Line Corrector (SLC) problem. Three classification algorithms (maximum likelihood, boosted decision
Change detection
trees, and support vector machines) were tested for their ability to monitor expansion across five time periods
Classification
(1988–1995, 1996–2000, 2001–2003, 2004–2006, 2007–2009) in three study areas that differ in size, eco-
Machine learning
Decision trees climatic conditions, and rates/patterns of development. Both the decision trees and support vector machines out-
Support vector machines performed the maximum likelihood classifier (overall accuracy of 90–93%, compared to 65%), but the decision
Random forests trees were superior at handling missing data. Adding transformed features such as band metrics to the Landsat
data stack increased accuracy 1–4%, while experiments with a reduced number of features (designed to mimic
noisy or missing data) led to a drop in accuracy of 1–9%. The methodology also proved particularly effective for
monitoring peri-urbanization outside the urban core, capturing >98% of village settlements.
© 2012 Elsevier Inc. All rights reserved.

1. Introduction
During the last two decades, we have made important strides to-
ward developing remote sensing methods that allow for the accurate
characterization of land cover change (Rogan & Chen, 2004), including
urban expansion (Chan et al., 2001). Mapping urban areas remains a
complex challenge, however, because of the many combinations of ma-
terials present and the variations in size/shape of urban features that
can lead to different ‘mixtures’ within pixels (Small & Lu, 2006). Partic-
ularly troublesome is the fact that newly developed urban areas typical-
ly appear identical to fallow farmland at any given time, since both
exhibit high reflectance in the visible-infrared wavelengths. These is-
sues are further compounded in developing countries such as China
and India, since new development is often small, patchy in nature,
and located in peri-urban areas up to 100 km from the urban core
(Long et al., 2009; Webster, 2002).
⁎ Tel.: + 1 608 890 0557; fax: + 1 608 265 4113.
E-mail address: aschneider4@wisc.edu.
The new wave of very high spatial resolution (VHR) data (1–4 m)
holds tremendous promise for resolving these issues, and methods
have emerged to characterize urban features with increased spatial de-
tail (Ban et al., 2010; Del Frate et al., 2007). However, the sparse cover-
age, limited scene availability, and lack of data prior to 2000 make
routine use of VHR data to map change impractical, and in some loca-
tions, impossible. Currently, medium resolution (20–30 m) datasets
such as Landsat and SPOT remain the best option for balancing the
trade-offs involving spatial detail, areal coverage, and availability of his-
torical data. The dense archives as well as routine collection of these
data (as opposed to VHR ‘on-demand’ collection) are also advantageous
when the rate of change is particularly rapid; in cities in China, for in-
stance, the scale and pace of urbanization must be monitored on the
order of years rather than on decades (Ma, 2004). Moreover, studies
that have moved beyond mapping to link social and economic processes
to land use have shown that monitoring change for multiple periods
(i.e. three or more) is pivotal to understand the complex drivers of
urban morphology through space and time, and to forecast future
land use trends (Seto & Kaufmann, 2003).

0034-4257/$ – see front matter © 2012 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2012.06.006
690 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704
With the opening of the Landsat archive and new data sources com-
ing online, it is now possible to take advantage of the temporal dimen-
sion of satellite data to map urban change. The central premise of the
approach presented here is that the confusion between new urban
land and other land cover types can be resolved by including images
(a) from multiple seasons, as well as (b) from multiple years. While
there is likely to be confusion between bare ground and urban areas
during the course of one year, there is often a high probability that near-
by fields or open areas will be vegetated during at least one season of a
given year, and thus be ‘separable’ from built-up areas that are predom-
inantly non-vegetated year round. The temporal information from mul-
tiple years (case b) is also beneficial: expansion of built-up lands is often
unidirectional (i.e. once land is converted to urban uses, it is unlikely to
be converted back to farmland or forest), and thus images that follow
the date of change actually ‘confirm’ that an area has been developed.
Despite the clear advantage of seasonal, multi-year information, few
studies have exploited a multi-temporal approach to resolve urban re-
mote sensing issues.
The primary goal of this research was to develop, test, and validate a
multi-date composite change detection technique that is effective in
complex, heterogeneous urban and peri-urban environments. The
multi-date composite approach was specifically chosen because of its
superior accuracy for difficult change detection problems (Coppin et
al., 2004; Rogan & Chen, 2004), and because it provided a means to ex-
ploit dense time stacks of Landsat data now freely available from the
U.S. Geological Survey (USGS, 2011). Because of the complexity and
size of the datasets for a given study area (~35–50 Landsat scenes),
three supervised classification algorithms were tested: a traditional
maximum likelihood (ML) classifier, and two machine learning algo-
rithms, boosted decision trees (DT) (Quinlan, 1993), and support vector
machines (SVM) (Chang & Lin, 2001). The methods were tested for
their ability to isolate and correctly classify urban expansion for five
time periods spanning 1988 to 2009. To prevent the methodology
from being applicable in only one study area, the algorithms were test-
ed on three cities (Fig. 1) with different city sizes, diverse ecosystem
characteristics, as well as differential rates/patterns of urban develop-
ment. Specifically, the following questions guided this research:
(1) Which supervised classification algorithm performs best for
change detection in urban environments given dense temporal
data stacks?
(2) Does the addition of transformed features increase overall
accuracy?
(3) What is the impact of data quality and quantity on classifier
performance? and
(4) How well do multi-date change detection approaches work in
peri-urban environments given the small size/scale of settle-
ments (e.g. built-up areas ~ 1800 km 2)?
To address these questions, the three algorithms were evaluated
using multiple criteria through a series of experiments that provided
different combinations of data features as input to the classifiers. Recent
applications that exploit dense, multi-temporal datasets have benefit-
ted from the inclusion of transformed data such as band maxima, min-
ima and means (Friedl et al., 2010; Hansen et al., 2008), thus question 2
was designed to test whether the addition of these features was benefi-
cial for monitoring urbanization. Question 3 was included to specifically
test feature selection. While machine learning algorithms can now han-
dle any number of input features, feature selection remains an impor-
tant concern if reduced-quality data are used as input (e.g. data with
clouds or missing observations), or if insufficient imagery is available
to create a dense stack. In the context of these issues, it is important
to understand which features may or may not be necessary to achieve
high accuracy results. Finally, question 4 was designed to test the ability
of this approach to characterize small, piece-meal land development
and village settlements outside the city core.
2. Background: remote sensing of urban change, 1970s to today
Characterizing cities and towns with remotely sensed data has
been challenging since the field of Earth observation began nearly
50 years ago. As medium resolution satellite data (Landsat MSS) be-
came available in the 1970s, early applications relied on simple
band ratios, image thresholding, and image differencing to discern
broad-scale changes at the urban–rural fringe (Friedman & Angelici,
1979; Howarth & Boasson, 1983; Jensen & Toll, 1982; Todd, 1977).
Despite the apparent success of early approaches, the potential user
community – urban planners, developers, land managers, and social
scientists – did not immediately embrace the new technology. Rela-
tively few applications appear in the urban planning literature, due in
large part to the lack of spatial detail and thus inferior information content
of satellite data relative to aerial photos or ground surveys (Michalak,
1993; Ryznar & Wagner, 2001). While GIS technology has been widely
adopted in urban planning, the lack of reliable, easy-to-use methods
and dearth of remote sensing data with sufficient spatial resolution con-
tinue to impede widespread use of satellite-based maps of urban change.
A second user community of remote sensing-based maps of urban
change emerged in the 1990s–2000s, however. Disciplines such as cli-
matology (Romero et al., 1999), hydrology (Carlson & Arthur, 2000),
ecology (Robinson et al., 2005), and public health (Tatem & Hay,
2004) have embraced satellite data to understand the impacts of
urban expansion on environmental systems, as well as human health
and well-being. Moreover, there is a growing body of work looking at
urbanization and its effects from a regional to global perspective
(Mills, 2007; Pataki et al., 2006) which requires medium to coarse res-
olution large-area maps of urban extent and urban change (Schneider
et al., 2009, 2010). With these needs in mind, it is critical that the re-
mote sensing community continues to develop efficient methods and
to explore data sources for mapping urban growth and sprawl.
Although the user community has varied widely, the methods to
generate maps of urban growth have not deviated significantly from
early approaches that exploited spectral profiles of built-up areas and
newly developed land (Ehlers et al., 1990; Jensen & Toll, 1982;
Ulbricht & Heckendorff, 1998; Yang & Lo, 2002). Multi-date composite
approaches – those using images of two dates that are combined during
processing to produce a map of change – began to be used for urban ap-
plications in the 1990s (Ridd & Liu, 1998). Early multi-date techniques
included stacked principal component analysis (Deng et al., 2008; Li &
Yeh, 1998), change vector analysis (Chen et al., 2003), or stacked
multi-date composite classification (Schneider & Woodcock, 2008). To
handle the complexity of the urban environment, machine learning ap-
proaches were adopted in the late 1990s, including neural networks
(Dai & Khorram, 1999; Liu & Lathrop, 2002), boosted/bagged decision
trees (Rogan et al., 2003; Schneider et al., 2003, 2005), and support vec-
tor machines (Griffiths et al., 2010; Nemmour & Chibani, 2006). While
these algorithms provided increased class accuracies, isolating distinct
spectral signatures from the inherently mixed pixels in urban environ-
ments has remained problematic.
It is becoming increasingly clear that resolving class confusion in
urban change detection applications requires taking advantage of ‘do-
mains’ of remote sensing beyond spectral information, such as temporal,
spatial, or polarimetric domains. In this regard, data fusion approaches –
those that combine spectral profiles with either spatial information or
radar responses – have shown great potential. Spatial information has
been exploited through contextual classification and object-oriented pro-
cessing (An et al., 2007; Li et al., 2009), which tackle the urban problem by
using patches to reduce the variability in the urban spectral response. Tex-
tural information has also been tested widely for urban change detection,
but with minimal improvement in detection of built-up areas (Gluch,
2002; Moller-Jensen, 1990). Data fusion has also included merging multi-
spectral and radar data: visible to near-infrared wavelengths are used to
reveal the composition of the land, while the high backscatter of human-
made objects in radar data is used to discern settlement structure
 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704 691

Fig. 1. The location of each study area within China. The maps of land cover change illustrate the full extent of the study area (the Landsat scene), as well as the 40 km buffer applied
to the central business district to standardize the study areas for comparison.

(Henderson & Xia, 1997). Although multi-source data fusion has been
touted as one of the most promising developments for urban remote
sensing (Dell'Acqua, 2009), it has not seen widespread use because of
the often difficult and time-consuming task of correcting and synthesizing
two or more disparate data sources (Dobson et al., 1996; Sanli et al.,
2009).
In sum, it is clear that the methods to detect urban development are
diverse and often location-dependent, and there is a significant lack of
consensus on the ‘best practices’ for monitoring urban expansion.
While a range of methods for exploiting spectral and spatial information
has been explored, researchers have not taken advantage of the tempo-
ral domain (e.g. seasonal information or dense time series) to isolate
urban change. The reason is likely due to a lack of available data and
methods to efficiently process large, complex datasets. Our initial
work using eight Landsat scenes to map five time periods (Schneider
et al., 2005), and Yuan et al.'s (2005) use of spring–summer imagery
to map urban areas are two of the only studies that rely on multiple sea-
sons/years to map urban land use change. These studies, as well as those
using dense time stacks to monitor forest and cropland change
(Kennedy et al., 2007; Kuemmerle et al., 2009) serve as a point of depar-
ture for the new work described in this research.
3. Study areas
Three Chinese cities (Fig. 1) were selected for the case study
examples because China has experienced rapid urbanization and
692 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704

Fig. 2. Multispectral observations of built-up land and cropland in a peri-urban area southwest of Chengdu, China, from circa 2000–2001 30-m resolution Landsat ETM+ data (the near-
infrared, short-wave infrared and red wavelengths are set to R-G-B) acquired in: (a) winter, (b) spring, (c) summer, and (d) fall. For reference, the same area is shown in (e) using very
high resolution true-color Quickbird imagery (Google Earth, 2011).

industrialization since the 1980s, with significant variability in rates 4. Defining urban and peri-urban land
of change, patterns of development, and land cover types across the
country (Lin & Ho, 2005). The heterogeneity of the landscape, the Before outlining the approach, it is important to clearly define
small-scale, piecemeal nature of development, and growth in peri- ‘urban land’. In physical terms, ‘urban land’ refers to places dominated
urban areas make these locations some of the most difficult areas in by the built environment. The ‘built environment’ includes all non-
the world in which to accurately characterize change. Thus, these cit- vegetative, human-constructed elements, such as roads, buildings,
ies serve as important case studies for testing algorithms to monitor runways, etc., and ‘dominated’ implies coverage greater than 50% of
urban expansion rapidly and efficiently. Also critical to this work a given landscape unit (the pixel). When vegetation (e.g. a park)
was the need to move away from one-time, one-place method devel- dominates a pixel, these areas are not considered urban, although
opment (Woodcock & Ozdogan, 2004), or methodological testing they may function as urban space. Finally, the definition also includes
using simulated datasets. a minimum mapping unit: urban areas are contiguous patches of
The three cities – Chengdu, Xi'An, and Kunming – were selected built-up land greater than 1800 m 2 (two pixels).
based on differences in physical geography, land cover types, climate, In addition, this work focused on characterizing development in
and urban development trajectories. Both Chengdu and Xi'An have a peri-urban areas. Here, peri-urbanization refers to new urban forms
wealth of natural resources and rich agricultural land that have facil-
itated their rapid development since the early 1990s. Chengdu sits in
the Sichuan Basin, where a triple-cropping system of rice, wheat, and
vegetables predominates, while Xi'An is located in Guangzhong Plain,
where a rotation of maize and wheat is typical. Kunming, in contrast,
has been resource-poor, partly due to its location in the rugged land-
scape of the Yungui Plateau. Kunming has a small amount of agricul-
ture near the edge of Lake Dian, but is surrounded primarily by forest.
Climatic conditions also vary across the study areas, further influenc-
ing vegetation growth and patterns: Chengdu is largely mild due to its
humid subtropical climate, and Kunming has similarly mild seasons
typical of its subtropical highland climate. Xi'An, in contrast, has the
greatest seasonal fluctuation, with a temperate continental semi-
arid climate of hot, humid summers and cold, dry winters.
Finally, there are important differences in policy, economic
growth, and urban development across the three cities. Chengdu
was designated early on as a western ‘Gateway City’, receiving invest-
ment and incentives for growth from the central government that led
to rapid urban expansion beginning in the 1990s, and which has con-
tinued to the present day (Webster, 2002). Xi'An, in contrast, did not
receive the same investment, but instead emerged as a leader in tele-
communications and aerospace industry after 2000. Its growth repre-
sents part of the ‘second wave’ of development that occurred in
Western cities (Lin, 2009). Finally, Kunming witnessed a different
type of growth: a spurt of rapid development in the year prior to
the World Horticultural Expo, and later, designation as an ecotourist
destination within China (Feiner et al., 2002). As a result, Kunming
has been part of a ‘third wave’ of growth (after 2006), with recent
government investment in infrastructure and industrial zones, and
planned residential and commercial corridors. These regional differ-
ences have significantly impacted both the rates and patterns of de- Fig. 3. Flow chart illustrating the methods to produce maps of land cover change using
velopment, which are important considerations when evaluating the multi-date composite change detection technique and dense time stacks of Landsat
change detection methods. data.
Table 1
The time series of Landsat 5 and Landsat 7 imagery used to map land cover change in (a) Chengdu, (b) Xi'An, and (c) Kunming, China. The images for each date were stacked and used as input to a multi-date composite change detection
technique. Although the study period ends in 2009, images from 2010 were included to reinforce trends in earlier periods.

Chengdu 1980s 1990s 2000s

p129r039 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

January 01-21 01-27

February 02-15 02-07
March 03-26 03-29 04-03 03-05 03-08 03-16 03-19
a
April 04-24 04-14 04-19 04-06 04-30
a a
May 05-01 05-05 05-10 05-21 05-19 05-06 05-03 05-22
June 06-23
July 07-30
August 08-16 08-12 08-01 08-07
September 09-19

A. Schneider / Remote Sensing of Environment 124 (2012) 689–704

October 10-07
November 11-02 11-03
December 12-09 12-23 12-29

Xi’An 1980s 1990s 2000s

p127r039 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

January 01-10 01-24 01-13 01-16

February 02-06 02-28
March 03-12 03-31 03-04 03-07 03-29
April 04-21 04-16
May 05-07 05-12 05-18 05-24
June 06-29 06-03 06-16 06-27 06-03 06-30
July 06-24 07-13 07-24 07-05 07-24
August 08-23 08-22 08-09 08-22
September 09-01 09-23 09-21
October 10-28
November 11-20 11-08
December 12-20 11-24 12-28

Kunming 1980s 1990s 2000s

p129r043 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

January 01-21 01-30 01-03 01-09 01-30

February 02-20 02-23 02-09 02-28 02-09 02-20 02-23 02-12 02-15
March 03-09 03-23 03-02 02-25 03-08 03-11 03-29 03-19
April 03-25 04-03a 04-24 04-30 04-06 04-01 04-17
May 05-21 05-19 05-06
June 06-14 06-04
July
August 08-16
September 09-15 09-13 09-19
October 10-07
November 11-02 11-21 11-03 11-08 11-11
December 12-23 11-27 12-10 12-29

Table key: Landsat 5 TM Landsat 7 ETM Landsat 7 ETM scan line corrector-off Scenes used to select training data

a
Landsat 5 data acquired from the Chinese Ground Receiving Station (CEODE, 2010). All other data were acquired from the U.S. Geological Survey (USGS GLOVIS, 2011).

693
694 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704

resulting from region-based urbanization, where the city influence
has expanded to a wider region (beyond suburbs) (Webster, 2002).
This trend is important because the line between urban and rural
areas has become diffuse, and accurate depiction of land cover change
in these extended urban regions requires isolating the large number
of settlements that dominate the Chinese landscape, as well as isolat-
ing the expansion of these areas from surrounding land cover types
that may be spectrally similar.
Four seasonal views southwest of Chengdu illustrate the difficulty
in monitoring peri-urban settlements: the villages are distinct from
surrounding cropland during winter and summer (Fig. 2a, c), but
nearly impossible to discern during spring and fall (Fig. 2b, d). Be-
cause of the similarity in materials between peri-urban settlements
and urban areas (and thus similarity in spectral and temporal signa-
tures), this work followed the convention used in the geophysical sci-
ences by including these areas in the urban class during image
processing. Within the field of urban studies or planning, however,
these areas do not function as urban space nor do they have the min-
imum population size to be designated a city or town.
5. Methods
5.1. Overview of data and methods
The following sections describe the steps involved in this re-
search: (a) acquisition, preprocessing, and stacking of all satellite
data; (b) collecting representative training exemplars for land cover
types in stable and changed areas, (c) classification using the ML, DT,
and SVM algorithms, (d) evaluating algorithm performance, and, final-
ly, (e) accuracy assessment using a stratified-random sample of refer-
ence data. These steps are outlined in Fig. 3.
5.2. Satellite data and pre-processing
This research relied on terrain-corrected Level 1T Landsat 30-m res-
olution data (USGS, 2011). Given the quantity of observations necessary
to monitor seasonal and yearly differences in three regions (Table 1),
this project would not have been feasible without the recent release
of this valuable archive and the continued collection of Landsat 5 and
from before a given period helps establish land cover classes, while
information from after the period helps confirm that change occurred.
5.3. Training data selection and classification scheme
Selection of training sites was performed in-lab for each case
study area through visual interpretation of Landsat data and Google
Earth VHR imagery (multiple dates, 2000–2011) and through on-
the-ground visits to each location. On-site visits (multiple visits,
2009–2011) confirmed that the only land lost to urbanization was ag-
ricultural land, and other land cover transitions (e.g. forest to urban
land) were negligible. Thus, this work focused specifically on the
transition from cropland to urban land in each region, mapped for
five time periods: 1988–1995, 1996–2000, 2001–2003, 2004–2006,
and 2007–2009. In addition, unchanged or ‘stable’ classes such as nat-
ural vegetation and water bodies were mapped, as well as agricultural
and urbanized lands that remained unchanged throughout the period
(see Table 2 for class definitions).
When choosing training sites for changed areas, care was taken to
use Landsat or Google Earth imagery captured from the last year of a
given period (e.g. imagery from fall–winter 2006 was used to define
training data for the 2003–2006 period, Table 1). Images from multi-
ple seasons during that year were used to supplement photo-
interpretation as needed.
5.4. Classification algorithms for supervised multi-date composite change
detection
To monitor urban change across multiple time periods, a ML clas-
sifier, boosted DT, and SVM were tested for their ability to exploit the
spectral and temporal dimensions of the data. While numerous ma-
chine learning algorithms are now available (Rogan et al., 2008), DT
and SVM were selected based on past performance with large, com-
plex datasets (Friedl et al., 2010; Huang et al., 2009). Detailed com-
parisons of the algorithms have been published elsewhere (Huang
Table 2
The classification scheme used in each study area, and the distribution of training ex-
emplars for each class.

7 during the last decade. Unfortunately, Landsat 7 encountered difficul- Classification Definition Training sites per classa
ties in May, 2003 as a result of the malfunction of the scan-line corrector scheme
Chengdu Xi'An Kunming
(SLC). This failure causes wedge-like data gaps in the ETM+ scenes. Be-
Unchanged, stable classes
cause the radiometric and geometric qualities are not impacted by the
Urban, built-up land Built environment 711 725 750
SLC, Landsat 7 ETM+ scenes acquired after May, 2003, (hereafter, > 50%, including
SLC-off data) have been archived (USGS, 2011). While there have been non-vegetated,
a variety of efforts to ‘gap-fill’ the SLC-off scenes (Masek, 2007; human-constructed
Maxwell, 2004; Zhang et al., 2007), the premise of this research was elements, with
minimum area
that gap-filling may not be necessary given a sufficiently large number > 1800 m2
of scenes (this idea is tested during algorithm evaluation, as described Cropland Crops comprise 1641 3549 1151
below). Given any three Landsat 7 SLC-off images, there is a 90% proba- >60% of total land
bility that >90% scene coverage can be achieved (Storey et al., 2005). cover
Forest, natural vegetation Woody vegetation 413 1301 2678
The dense time stacks method essentially uses a ‘brute force’ approach
with height > 2 m
in which all relatively cloud-free Landsat scenes are provided to the comprises >60%
classification algorithm. Gap areas and clouds were not masked or of total land cover
corrected in any way during processing. Water Rivers, seas, lakes, 137 190 192
For each study area, 37–50 images were selected for the period reservoirs

1988 to 2010 (Table 1). For all scenes, pre-processing included check- Changed classes
ing geometric accuracy of all scenes (minor modifications were nec- Cropland to urban 1988–1995 Productive or fallow 192 323 128
essary for one scene), as well as data stacking to place all scenes in Cropland to urban 1996–2000 cropland converted 301 244 324
a single composite dataset. Radiometric normalization was not per- Cropland to urban 2001–2003 to built-up land 285 200 113
Cropland to urban 2004–2006 comprised of >50% 309 343 235
formed, since this step is not required when images are analyzed si-
Cropland to urban 2007–2009 human-constructed 437 234 402
multaneously rather than individually (Song et al., 2001). In elements with
addition, data transformations for testing input features were also minimum area
performed at this stage (discussed in Section 5.5). Finally, all scenes > 1800 m2
were stacked and used as input to the classifier to detect changed Total 4426 7109 5973

areas for the five periods of interest. The idea here is that information a
Each training site is one pixel in size, or 900 m2.
 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704 695

Table 3 Huang et al., 2009). The decision tree algorithm selected for comparison
The number of input features (i.e. bands) for each combination of Landsat imagery, is C4.5, a mature and extensively tested algorithm (Friedl et al., 2010;
NDVI, and band metrics tested in each study area.
Quinlan, 1993). Decision-tree construction is intuitive: the training
Case Input features Chengdu Xi'An Kunming data is split in a recursive manner into increasingly homogeneous sub-
1 All Landsat data 222 246 300 sets based on statistical tests applied to the feature values (the satellite
2 All Landsat data, NDVI, metricsa 343 371 434 data). In C4.5, the statistical test at each split is based on reducing entro-
3 All Landsat data, metrics 306 330 384 py in descendent nodes by maximizing the information gain ratio
4 All Landsat data, NDVI 259 287 350
(Quinlan, 1996). After the tree has been estimated, tested, and ‘pruned’
5 All NDVI, all metrics 121 125 134
6 All NDVI 37 41 50 to eliminate over-fitting to the training data, the decision rules are ap-
7 All Landsat data, no SLC-off data 120 114 150 plied to the entire image to produce a classified map. The efficacy of
8 Top 50% of all featuresb 172 186 217 DT increases significantly with the addition of boosting, an ensemble
9 Top 25% of all featuresb 86 93 109 classification technique that improves class discrimination by estimat-
10 Top 10% of all featuresb 34 37 43
ing multiple classifiers while forcing the classifier to focus on difficult
Abbreviations: NDVI — normalized difference vegetation index, SLC-off — scan line classes (Quinlan, 1996). The final classification is produced by an
corrector-off.
a accuracy-weighted vote across the classifications (Quinlan, 1996).
Band metrics included yearly mean, maxima and minima for each spectral band for
years 2000, 2003, 2006, and 2009. Boosting has been shown to be an equivalent of additive logistic re-
b
Top features were determined for the case 2 inputs for each study area using variable gression (Collins et al., 2002; Friedman et al., 2000), and as a result,
importance estimated by the Random Forests algorithm (Breiman, 2001). probabilities of class membership can be assigned for each class at
every pixel. While a newer version of C4.5 is now available (C5.0),
testing showed that the differences between the algorithms were
et al., 2009; Lippitt et al., 2008), hence the purpose of this work was to in processing time and tree size; error rates were roughly equivalent.
determine if one method was superior at handling (1) the particular com- SVM employ optimization algorithms to determine the location of
plexity of heterogeneous cities and settlements, and (2) multi-temporal decision boundaries that produce the optimal separation of classes in
class signatures. The following section provides a brief description of the training data (Vapnik, 1995). The decision surface between clas-
each classifier (please see Tso & Mather, 2001; Quinlan, 1993, and ses is often called the optimal hyperplane, and the support vectors
Chang & Lin, 2001, for additional information). are those data points closest to the hyperplane. In a two-class pattern
ML is a widely-used parametric classifier that relies on the second- recognition problem, the SVM selects the linear decision boundary
order statistics of a Gaussian probability density function model for that leaves the greatest margin between the two classes (the margin
each class. Training data are used to estimate the mean vector and co- is defined as the sum of the distances to the hyperplane from the clos-
variance matrix, and a discriminant function is then calculated for est points of the two classes). SVM have been adapted to handle non-
each class. The class with the highest value is selected for the final linear boundaries between classes through the use of kernal func-
classified map. ML is often selected for algorithm comparison be- tions; the non-linear kernals transform the training data into higher
cause, if the probability distribution of each class is indeed Gaussian, dimensional feature space where linear class separation is then possi-
it is a robust, optimal classifier (Tso & Mather, 2001). ble (Huang et al., 2002). This work followed the recommendations of
DT and SVM algorithms, on the other hand, offer several advantages Huang et al. (2009) by training the SVM with a Gaussian radial basis
over traditional classifiers: they are both derived from statistical learn- function (for a comprehensive review, see Burges, 1998; Huang et
ing theory, are non-parametric, and have been shown to be superior to al., 2002; Foody & Mathur, 2004). While the SVM is a binary classifier
traditional Bayesian methods for discriminating complex, hierarchical in its simplest form, it can handle multi-class problem by combining
patterns in satellite data (Friedl & Brodley, 1997; Hansen et al., 1996; several binary SVM classifiers using by creating a binary classifier

Fig. 4. Multi-temporal trajectories of urban areas, forests/natural vegetation, and one ‘change class’ (areas that were converted from cropland to built-up land circa 2003–2006) in
Xi'An, derived from Landsat NDVI images spanning August, 1999 to January, 2010. Years 2001 and 2003 were omitted due to insufficient data. To illustrate these trends, the Landsat
data were corrected to surface reflectance using LEDAPS (Masek, 2007). Similar temporal trajectories were observed in Chengdu and Kunming.
696 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704

for each possible pair of classes. The SVM classification was per-
formed using LIBSVM developed by Chang & Lin (2001).
5.5. Evaluation criteria and experimental design
Criteria for evaluating the performance of classification algorithms
include assessing (1) the impact of feature selection, (2) the effects of
noise, and (3) the impact of training set size on classification accuracy,
as well as (4) the interpretability/reliability of the results (Rogan et al.,
2008). This study focused on three of these criteria; the impact of train-
ing set size was not assessed directly because of evidence showing that
DT and SVM performance improves when additional training data are
provided (Huang et al., 2009; Lippitt et al., 2008; McIver & Friedl,
2001). The complexity and inter-/intra-annual variability of the real-
world Chinese landscape necessitates a large number of exemplars,
thus any sites collected should therefore be used during processing.
The impact of training set variability was minimized, however, through
the use of randomized subsets during cross-validation.
To assess the remaining criteria, traditional accuracy measures were
estimated using a ten-fold cross-validation of the sites collected for each
city. A random stratification procedure was used to produce ten unique
datasets for training (80%) and testing (20%) (note that all sites were col-
lected as one-pixel samples to prevent spatial autocorrelation effects dur-
ing cross-validation) (Friedl et al., 2000). The accuracy measures for each
classification were then averaged to calculate mean accuracy rates across
the ten classifications. The same training and test data were used to train
and test the three classifiers, thus making it possible to assess the relative
performance and consistency of the algorithms across the study areas.
To evaluate the impact of feature selection (both the addition of
transformed features and reduction of inputs), a series of experiments
was devised in which different combinations of bands were tested as
input to each classifier. The first experiment used all available Landsat
scenes as input (Table 3, case 1); this experiment was treated as the
‘control’ since the majority of change detection methods use raw,
untransformed data as input. In the second set of experiments, the
Landsat data were augmented with transformed features to produce
an expanded dataset for each study area (cases 2–4). The foundation
for this approach comes from recent studies with multi-temporal data
(e.g. MODIS) that exploit band metrics to increase classification accuracy
in global land cover mapping (Friedl et al., 2002; Hansen et al., 2005;
Schneider et al., 2010). In this research, the normalized difference vege-
tation index (NDVI) for each date and annual band metrics were both
tested as additional inputs. While it may seem counter-intuitive to use
NDVI to discriminate urban features, the temporal differences in NDVI
are particularly effective for distinguishing the removal of vegetation
that accompanies urban expansion (Stefanov et al., 2001). Moreover,
NDVI signatures for natural vegetation and urban areas (2000–2010) re-
veal that the cover types are separable through time by differences in
the amplitude, timing, and frequency of peak greenness (Fig. 4). Fig. 4
also illustrates the unique temporal trend for areas that became urban-
ized during the 2003–2006 period (red dashed line). Prior to conversion,
these areas followed the bi-modal vegetation cycle characteristic of

Fig. 5. The top 25% of the features for each study area as determined by the variable importance measure from random forests (Breiman, 2001). The selection frequency is illustrated
as both raw counts (left graph) and percentages (right graph); for the latter, the frequency was normalized by the number of features available for each category. The plots display
the selection frequency stratified by Landsat band (top row), time period (middle row), and season (bottom row).
 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704 697

cropland, but after 2003, the NDVI signal drops, and follows the pattern
within established urban areas. While the classification algorithms may
be adept at handling complex class signatures, it was hypothesized that
the addition of NDVI would help ‘nudge’ the classifier to detect these
temporal differences (Friedl et al., 2002).
In addition to NDVI, annual band metrics for the years 2000, 2003,
2006, and 2009 were also included. To do this, >4 Landsat scenes cap-
tured during one year were transformed into annual means, maxima,
and minima for each spectral band and NDVI. Annual metrics provide
a more generalized feature space to the classification algorithm by cap-
turing the salient features of phenological variation without reference to
time of year, and have been shown to perform as well as time-sequential
image composites in mapping land cover (Hansen et al., 2005).
In the third set of experiments, the impact of missing or reduced-
quality data was tested by using subsets of the Landsat bands, NDVI
and band metrics (cases 5–7). To provide an objective selection
process, random forests (RF) were then used to define a subset of var-
iables considered critical for classification (cases 8–10). RF – a variant
of ensemble decision trees (Breiman, 2001) – have seen increased use
in remote sensing for their ability to highlight variables considered
important in a given classification (Guo et al., 2011; Pal, 2005). RF
have been favored over principal components analysis due to their
explicit estimation of variable importance, as well as their increased
flexibility with large datasets, speed in processing, and lack of as-
sumptions on data distribution (Menze et al., 2009). The RF algorithm
draws a series of bootstrap samples from the training data to con-
struct a large number of trees (a ‘forest’ of trees). Each tree is grown
with a randomized set of predictors, a portion of which have been
randomly permuted. This follows the idea that a random permutation
of a feature mimics the absence of that feature from the model. The
measure of feature importance, then, is the difference between pre-
diction accuracy (i.e., the number of observations correctly classified)

Fig. 6. Results of the cross-validation assessment for each combination of Landsat imagery and transformed features (NDVI, band metrics) using the maximum likelihood classifier
(top row), boosted decision trees (middle row), and support vector machines (bottom row). For these tests, the training exemplars used to create the maps were randomly split ten
times to produce training (80%) and testing (20%) datasets; the results from the ten classifications were then averaged to estimate mean overall accuracy and class accuracy.
698 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704

before and after permuting the feature, averaged over all trees
(Breiman, 2001). Thus, a high decrease in accuracy denotes the im-
portance of that feature.
An ensemble of 1000 trees was grown by the RF algorithm
(implemented in Matlab, Jaiantilal, 2009) using the ‘all Landsat,
NDVI, metrics’ subset (case 2) for each study area. The features were
then ranked in order of importance according to the overall decrease
in accuracy offered by that feature across all classes. From these rank-
ings, the top 10, top 25 and top 50% of features were selected for each
city to create cases 8–10 (Table 3).
Finally, three additional analyses were used to test map accuracy
and to supplement the cross-validation assessment. First, the land
cover change maps (following final edits, spatial filtering, etc.) were
assessed for accuracy using an independent set of reference data. A
stratified-random sample of >200 points was selected for each city,
and Landsat and Google Earth VHR times series imagery were
assessed to label the points. Second, the effect of missing data (here,
the SLC-off data gaps) was further examined by comparing the per-
pixel confidence estimates from the classification output against the
number of high-quality values in the Landsat band stack for each
city. Third, the maps were analyzed to determine the ability of the
classifier to accurately map peri-urban settlements.
6. Results
6.1. Feature importance determined by random forests
While the primary goal of the RF analysis was to identify feature
subsets for classification, several interesting conclusions can be
drawn from the RF results. The results in Fig. 5 are presented as the
frequency with which an input feature appears ranked in the top
25% ranking, stratified according to Landsat band, time period, and
season (note that the patterns in the top 25% of features were repre-
sentative of those in the top 10 and top 50% subsets as well).
The results show that certain Landsat bands play a more important
role than others in classification (Fig. 5, top row). For example, the
near-infrared wavelength (Landsat band 4) is highlighted as important
more often than other bands in Chengdu and Xi'An, while the visible
wavelengths (Landsat bands 1–3) are selected as important in Kun-
ming. This result is likely due to the dominant land cover type in each
region: Chengdu and Xi'An are both primarily agricultural outside the
core city, while Kunming is dominated by evergreen forest. The near-
infrared appears to play an important role for inter- and intra-year
crop variations, but less so for areas where the spectral response does
not change. Interestingly, the NDVI and band metrics are important

Fig. 7. A series of panels showing the performance of the maximum likelihood classifier, boosted decision trees, and support vector machines for three hard-to-classify areas of
change: multi-period urban expansion in near Kunming's core (top row), development of a town nucleus outside Chengdu (middle row), and urban expansion into villages and
settlements in Xi'An (bottom row). These maps were produced using the ‘all Landsat, NDVI, metrics’ (case 2) as input to the DT and SVM, while the ML was run using ‘all Landsat,
no SLC-off’ data (case 7). For reference, Google Earth imagery (~ 4 m resolution, far left) and Landsat data (30 m resolution, the short-wave infrared, near-infrared, and red wave-
lengths are set to R-G-B) are shown.
 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704 699

features across all study areas, often weighing in as heavily as the other algorithms outperform the ML classifier, but the difference between
top bands for each city. the DT and SVM is again negligible.
With respect to time period, features from all years appear in the top While the cross-validation revealed that, in general, the SVM per-
ranking of important inputs for all study areas (Fig. 5, middle row). A formed as well or slightly better than the DT algorithm, examination
slightly greater number of features from the two early periods of the maps in Fig. 7 provides a different story. A final map was pro-
(1988–1995; 1996–2000) are considered critical for classification, de- duced for each city using each of the three classifiers and the full set
spite the small number of inputs available in the 1980s–90s. However, of training data as input. The results indicate that DT provide the best re-
the raw count of features also reveals that the number of bands consid- sults for each city: the size and shape of urban features are in good
ered important for classification is influenced by the number of inputs agreement with the expected morphology of each region, and there
for those categories. Each study area had a greater number of features are no noticeable artifacts from the input data. In contrast, the SVM con-
in the 2001–2003 and 2007–2010 periods, and these periods appear sistently produced a striping effect as a result of misclassification of the
more often in the top rankings when judged by frequency alone. SLC-off gap areas. In Kunming and Chengdu, the striping is noticeable
Finally, the results for season indicate that images during certain across all of the study area, while in Xi'An, only the eastern and western
times of year appear in the list of important features more often than scene edges (not pictured) suffered from the deleterious effects of the
others (Fig. 5, bottom row). The results suggest that some seasons SLC-off data inputs. Fig. 7 also reveals that the ML classifier does not
played a fairly important role during classification despite a small num- achieve useable map results. The ML either does not classify areas of
ber of inputs. In Kunming, for example, images from the summer change (error of omission, top two), or conversely, or it over-classifies
weighed in more heavily than other seasons (Fig. 5i), although only a one of the change classes (significant errors of commission).
few inputs were available from June, July and August. Finally, the results of the accuracy assessment using independent
reference data confirm that the DT map results are indeed superior
6.2. Which supervised classification algorithm performs best given dense to those from the SVM. The overall accuracy rates for the final DT
temporal data stacks? maps were 90.3, 89.9, and 92.7% for Chengdu, Xi'An, and Kunming, re-
spectively, while the maps produced with SVM had overall accuracies
The pooled cross-validation results reveal two major trends re- of 58.3, 82.6, and 78.2%.
garding classifier performance (Fig. 6). First, both the DT and SVM
outperformed the ML in terms of overall accuracy (Fig. 6a–c); the re- 6.3. Does the addition of transformed features increase overall accuracy?
sults for the two machine learning approaches (83–92%) are marked-
ly higher than for the ML classifier (63–82%), regardless of the input Relative to the original stack of Landsat imagery, the use of band
band combination. These results confirm previous studies that metrics and NDVI provided modest improvement in both overall ac-
showed higher rates of accuracy can be achieved with machine learn- curacy and class accuracy when used in conjunction with the DT or
ing approaches (Lippitt et al., 2008). Given the high degree of spatial SVM algorithms. The increase in accuracy was as high as 0.5–2.0% in
and temporal variability, the large, noisy datasets, and the complexity the pooled results; however the results at the individual city level
of the change detection problem in each study area, the low accuracy showed increases of 1–4% across both the DT and SVM algorithms
rates for the ML algorithm are to be expected. (Fig. 6). This result indicates that the addition of transformed data
Second, the results reveal only minor differences in overall accuracy can be advantageous in producing high accuracy maps.
between the DT and SVM algorithms. For the band combinations that With respect to the ML classifier, the added information from the
include all Landsat data (cases 1–4), both classifiers achieve a mean transformed bands was not helpful. Rather, the highest overall accuracy
overall accuracy of 92% across the three city sample. Looking more for each city (89.4, 74.6, and 76.8% for Chengdu, Xi'An and Kunming,
closely at the individual city results, the overall accuracy from the
SVM is higher than those achieved by the DT, but the difference is
never more than 0.7 percentage points across the case 1–4 band inputs.
The cross-validation results were assessed further to understand
how each algorithm performed at the class level. For ease of compar-
ison, the class accuracies for the three cities were pooled into ‘stable’
classes (Fig. 6d–f) and ‘change’ classes (Fig. 6g–i). The assumption
here is that the change classes provided added variability that
makes these areas more difficult to classify. Evaluating the class-
level results therefore offers a more nuanced view of classifier perfor-
mance than overall accuracy.
Not surprisingly, the results at the class level corroborate the overall
accuracy results: the DT and SVM performed better than the ML classi-
fier. The mean accuracies for the stable classes (including agriculture,
forest, urban areas and water) are nearly identical to the overall accura-
cy results (91–92% for cases 1–4), with some added variability as indi-
cated by the increase in the error bars for all feature combinations.
This variability can be traced to the class results for Chengdu: the exem-
plars for agriculture in this region included a wide range of spectral sig-
natures due to the different cropping systems and cycles in the region,
and as a result, the accuracy across the cross-validation runs varied sub-
stantially for this class (68–95%). The remaining stable classes had rela-
tively low variability across validation runs for all cities.
All three classifiers had more difficulty with the change classes than
the stable classes, as indicated in the accuracy rates (79–82% for cases
1–4). Note that several of the cross-validation runs could not be run
for the ML, due to the small number of training sites in the change clas- Fig. 8. Visualization of the location and number of missing features in the full Landsat
ses relative to the number of band inputs. Both machine learning image stack (300 bands) for Anning, a satellite city near Kunming, China.
 700
A. Schneider / Remote Sensing of Environment 124 (2012) 689–704
Fig. 9. Box plots showing the range, median, minimum and maximum accuracies per pixel (y-axis) according to the number of Landsat observations (x-axis, binned in 5-band increments) used as input to the classifier for that pixel. Here,
‘accuracy’ is determined using the per-pixel confidence values from the decision tree classification. The band stacks for Chengdu, Xi'An and Kunming differed in size (222, 246, and 300 features, respectively). A low number on the x-axis is
indicative of a pixel with a large amount of missing data due to the Scan Line Corrector (SLC)-off problem in Landsat 7 data acquired after 2003.
 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704
respectively) was achieved using a reduced number of bands, such as
the ‘top 10 percent of all features’ or ‘all NDVI’ options (cases 10 and 6).
It is not surprising that the ML performed best with only a fraction of
the number of inputs, since it is well-established that this algorithm
handles the variability of fewer inputs better (Tso & Mather, 2001).
The gaps in the SLC-off data also appear to have caused problems for
the ML; using only transformed inputs or excluding these scenes pro-
duced higher overall accuracy with the ML.
6.4. What is the impact of data quality and quantity on classifier
performance?
The impact of data quality and quantity was assessed by systemati-
cally reducing the input features to those considered critical to classifi-
cation (cases 5–10). Three trends are worth noting in the results. First,
reducing the band inputs to NDVI or band metrics only (cases 5 and
6) led to reduced accuracy across the pooled and individual city results
for both the DT and SVM. For example, the case 6 ‘All NDVI’ subset pro-
duced the lowest accuracies of any of the band combinations, averaging
83–84% for overall accuracy, and 65–66% for the pooled change classes.
Second, removing the SLC-off scenes (case 7) also caused accuracy to
drop for the DT and SVM. Because the SLC-off scenes make up 40–50%
of the Landsat data, removing these features resulted in a reduction in
accuracy of 6–14%, with a significant impact on the change classes in
particular. This result confirms that additional data, even SLC-off scenes
with missing observations, are indeed helpful for change detection
when a machine learning algorithm is used.
Finally, reducing each expanded data stack to a small number of inputs
using the RF importance measure (cases 8–10) produced mixed results.
Again, the reduction in number of features caused overall accuracy for
the ML to climb (e.g. from 67% for case 1 to 82% for case 8), but fall consid-
erably for the DT and SVM algorithms (e.g. dropping from 92% for case 2
to 87% for case 8). However, a more modest reduction in inputs – such as
that provided by the ‘top 25’ or ‘top 50 percent of all features’ subsets (cases
8–9) produced accuracies roughly similar to those from the ‘all Landsat,
NDVI, metrics’ experiment (case 2). This result suggests that the full
range of features may not be necessary to produce a high quality change
detection result. However, close inspection of the results in Fig. 5 provides
strong support for including features from different parts of the spectrum
(visible, near-infrared), from all time periods (including periods with
Fig. 10. The final maps of land cover change produced in this research for (a) Chengdu, (b) Xi'An, and (c) Kunming, People's Republic of China.
701
missing observations), and from all seasons when monitoring change
across multiple periods and classes.
To explore the effect of data quality and quantity further, this
study compared per-pixel accuracy to the number of high-quality
Landsat observations used to classify a pixel. Because the number of
missing observations varies across the SLC-off scenes, these data pro-
vide a natural experiment for testing the impact of noisy or missing
data on class accuracy (each SLC-off had ~ 22% missing data in gap
areas located at the scene edges, Fig. 8). The working hypothesis
was that a greater number of high-quality looks (i.e. fewer missing
observations) results in higher overall accuracy. Per-pixel accuracy
was estimated using the confidence measure from the DT output for
the ‘all Landsat’ case. The results are shown as box plots in Fig. 9,
where the y-axis indicates per-pixel accuracy, and the number of in-
puts is binned in 5-band increments along the x-axis.
Although the band stacks for Chengdu, Xi'An and Kunming differed in
size (222, 246, and 300 features, respectively), the overall trend is similar
across the cities and across different classes: fewer missing features re-
sults in greater per-pixel accuracy, confirming the original hypothesis.
For example, cropland pixels in Xi'An (Fig. 9d), have accuracies >90%
when a large number of observations were available (>165 inputs of a
possible 246). The accuracy drops to b50% as the number of observations
drops from 160 to 125. This trend is most pronounced for Xi'An and
Chengdu, but less so for Kunming, regardless of the class. The number of
inputs was greatest for Kunming, so it is possible that a drop in accuracy
is not as apparent when extra observations are available.
6.5. How well do multi-date change detection approaches work in
peri-urban environments?
Based on the results above, the final maps of land cover change
produced by the DT were chosen (Fig. 10), and used to assess the ef-
ficacy of the multi-date composite approach to characterize small
peri-urban settlements. The large number of settlements (several
thousand) outside each city makes collecting ground truth data im-
practical, and the horizontal positional accuracy of Google Earth
VHR imagery (±39.7 m, Potere, 2008) makes pixel-based assessment
of these areas unreliable. Instead, the maps were tested for the
presence/absence of settlements against Google Earth VHR data.
Using a stratified-random sample of 60–100 9-km 2 sites for each
702 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704
city, the settlements in the maps and reference data were indepen-
dently counted using a double-blind procedure.
In the scatter plots in Fig. 11, the samples that fit to the 1:1 line indi-
cate cases where the number of settlements in the DT map matches that
in the VHR data. Samples below the line represent errors of omission
(where the map did not characterize a settlement) and those above
the line indicate errors of commission (where the map falsely depicts
a settlement due to misclassification). The results for all three cities
show a tight fit to the 1:1 line and high R2 values (R2 > 0.96), indicating
that settlements were characterized correctly in the majority of sample
sites. Although the total number of villages in each study area sample
varied from 762 to 2579, the error rates are low for all three cities. Errors
of omission across the total number of villages ranged from 1.0 to 1.8%,
and errors of commission from 0.3 to 2.7%.
7. Discussion and conclusions
With the release of the Landsat archive, new data sources coming
online, and the increasing availability of multi-temporal VHR data
from Google Earth, we have entered a new era of remote sensing. Spe-
cifically, it is now possible to exploit multi-date observations to map
land cover change, and to resolve confusion between classes with sim-
ilar spectral characteristics. The research in this paper presented a new
approach leveraging dense time stacks of Landsat data to monitor urban
change across multiple periods. The method was tested in some of the
most complex urban environments in the world, namely, three metro-
politan regions in China. There are four main results from this work.
First, the cross-validation accuracies from the final maps averaged
90–94% using DT and SVM, although the DT outperformed both the
SVM and ML algorithms based on visual interpretation of the resulting
maps against VHR reference data. Second, the inclusion of annual
band metrics and NDVI as input to the classifier provided a modest in-
crease in overall accuracy of 1–4%, depending on the algorithm and fea-
tures used. Third, the results from the experiments with a significantly
reduced number of inputs (cases 7 and 10) to mimic low quality or
missing data caused a decrease in accuracy of 4–7%. Clearly, the infor-
mation from additional seasons and additional years is beneficial for ac-
curate mapping of both change and stable classes in urban areas when a
multi-date composite change detection approach is used. Finally, the re-
sults showed that >98% of villages >1800 m 2 in size can be reliably
mapped with the new methodology, providing a significant advance-
ment over previous methods that attempted to map peri-urban areas.
Fig. 11. A comparison of the number of villages in very high resolution Google Earth imagery (reference data, x-axis) to those present in the final land cover change maps (y-axis).
Here, the presence/absence of villages was determined using a sample of 60–100 9-km2 sites (shown as n above).
The results provide several insights for monitoring urban change
which are applicable to future studies:
The use of decision trees and support vector machines: First, it is clear
from the results that advanced machine learning algorithms must be
used when dealing with high dimensional data such as dense time sta-
cks, large amounts of noisy/missing data, or the complexity of a rapidly
urbanizing landscape. While the cross-validation results showed that
both SVM and DT achieve high overall accuracy given this type of prob-
lem, the map results showed that the SVM could not effectively deal
with the gaps in the Landsat SLC-off data. Although SVM have been
shown to be extremely successful at handling difficult classification
problems, this work revealed a weakness in the SVM approach. SVM
are designed to use only the training data that that lie on the edge of
the class distribution to define class boundaries; all other training data
provide no contribution to the classification analysis (Foody & Mathur,
2004). The striping effect may be a result of the algorithm observing a
large numerical difference between gap areas and non-gap areas, and
then using this difference to define the support vectors, and ultimately,
the final class labels. The boosted DT out-performed the SVM since they
were able to effectively ‘ignore’ the striping in favor of high quality ob-
servations. Because the striping issue showed up in the maps and not in
the cross-validation results, this result also points to a potential weak-
ness in relying exclusively on cross-validation techniques to test algo-
rithm efficacy.
The importance of seasonal information: This study revealed the need
to consider seasonality when attempting to discern urban change. For
many areas of China with high inter- and intra-annual variability in
land cover types, it is nearly impossible to use medium-resolution
data sources to separate land cover classes at just one point in time.
Areas of bare ground, fallow or post-harvest agriculture, and new con-
struction are easily confused with one another at any given time of
year, while vegetated cover in small settlements is also easily confused
with surrounding land cover types if assessed at only one time during
the year.
While the use of dense time stacks and seasonal information is clear-
ly advantageous, there are a few drawbacks to this approach. First, such
large quantities of data can require extra processing time or special com-
puting routines such as parallel processing during classification. Second,
and more critically, it can be extremely difficult to find cloud-free data in
the seasons of interest, or during certain decades (i.e. the lack of Landsat
coverage in the 1980s–90s). In some areas, a dense time series may sim-
ply not be feasible. This issue points to the need to develop methods that
can exploit multi-sensor data from Landsat-like sensors such as SPOT,
 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704
ASTER, the Indian Remote Sensing (IRS) satellites, and the Chinese–
Brazilian Remote Sensing (CBERS) satellites. In addition, multi-
resolution methods may also be appropriate, such as new work to
synthesize Landsat and MODIS observations (Gao et al., 2006, 2010).
The need for high-quality training data: One additional finding from
this work is that successful mapping of urban environments requires
high-quality exemplars to be used as input to the classifier. The availabil-
ity of Google Earth data has revolutionized training site collection; using
VHR data to define sites worked particularly well in this research. The
‘bottom-up’ approach of using data collected in situ or from sources
with increased detail to define training sites is not new (Friedl et al.,
2002), but the vast quantities of VHR Google Earth data now available
for cities means that information at the characteristic scale of urban fea-
tures (e.g. trees, buildings) can finally be exploited. Given that supervised
classification approaches have come to dominate remote sensing studies
in recent years, the next step will be to develop semi-automated ways to
speed up training data collection. To this end, ideas such as training site
generalization across regions (Woodcock et al., 2001) and chain classifi-
cation (Knorn et al., 2009) hold tremendous promise.
The value of Landsat SLC-off data: This research illustrated that the
‘brute force’ approach of providing all available data to the classifier –
even reduced quality data with missing values – worked well. These
additional data inputs provided value to the change detection process
in the form of increased overall map accuracy. Several conditions
should be met, however, to achieve success with this method. First,
a sufficient number of data inputs must be provided. Mapping five pe-
riods spanning 23 years required >150 features (25 scenes) to
achieve high per-pixel accuracy (Fig. 9). Second, the SLC-off data
were never more than 50% of the total Landsat scenes for a given
city, thus ensuring that high quality observations were available. Fi-
nally, there is some advantage to augmenting the Landsat data with
NDVI and band metrics; the band metrics, especially, help overcome
the missing data problem by effectively eliminating gaps through
the use of annual means, maxima or minima.
In summary, several conclusions can be drawn from this study. This
research has illustrated that land cover types in urbanizing regions have
characteristic temporal profiles in several spectral regions, and in par-
ticular, that conversion from agriculture to urban land leads to a reduc-
tion in year-to-year greenness. Most critically, this work has illuminated
the utility of multi-date, multi-season imagery for change detection, es-
pecially in complex urban landscapes such as those in Southeast Asia. As
large area, wall-to-wall monitoring of land cover change with medium
resolution data (Landsat, SPOT, AWiFS, Sentinel-2, etc.) becomes more
common, there is an increased need for approaches that can deal with
the high data volumes, reduced-quality data, and time consuming na-
ture of data integration, interpretation, and processing. Anything that
can speed up or facilitate processing is of critical importance. To this
end, classifier selection and performance are key. This work specifically
revealed that the DT algorithm performed more effectively in heteroge-
neous urban environments than the traditional ML or other machine
learning approaches (SVM) for a large, complex dataset with missing
observations.
References
An, K., Zhang, J. S., & Xiao, Y. (2007). Object-oriented urban dynamic monitoring — A
case study of Haidian district of Beijing. Chinese Geographical Science, 17, 236–242.
Ban, Y. F., Hu, H. T., & Rangel, I. M. (2010). Fusion of Quickbird MS and RADARSAT SAR
data for urban land-cover mapping: Object-based and knowledge-based approach.
International Journal of Remote Sensing, 31, 1391–1410.
Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32.
Burges, C. J. C. (1998). A tutorial on support vector machines for pattern recognition.
Data Mining and Knowledge Discovery, 2, 121–167.
Carlson, T. N., & Arthur, S. T. (2000). The impact of land use — Land cover changes due
to urbanization on surface microclimate and hydrology: A satellite perspective.
Global and Planetary Change, 25, 49–65.
Center for Earth Observation, Digital Earth (CEODE) (2010). Chinese Academy of Sci-
ences (CAS) data catalog (available online at: http://cs.rsgs.ac.cn)
703
Chan, J. C. W., Chan, K. P., & Yeh, A. G. O. (2001). Detecting the nature of change in an
urban environment: A comparison of machine learning algorithms. Photogrammet-
ric Engineering and Remote Sensing, 67, 213–225.
Chang, C. C., & Lin, C. J. (2001). LIBSVM: A library for support vector machines (software
available online at: http://www.csie.ntu.edu.tw/~cjlin/libsvm)
Chen, J., Gong, P., He, C. Y., Pu, R., & Shi, P. (2003). Land-use/land-cover change detec-
tion using improved change-vector analysis. Photogrammetric Engineering and Re-
mote Sensing, 69, 369–379.
Collins, M., Schapire, R. E., & Singer, Y. (2002). Logistic regression, Adaboost, & Bregman
distances. Machine Learning, 48, 253–285.
Coppin, P., Jonckheere, I., Nackaerts, K., Muys, B., & Lambin, E. (2004). Digital change
detection methods in ecosystem monitoring: A review. International Journal of Re-
mote Sensing, 25, 1565–1596.
Dai, X. L., & Khorram, S. (1999). Remotely sensed change detection based on artificial
neural networks. Photogrammetric Engineering and Remote Sensing, 65, 1187–1194.
Del Frate, F., Pacifici, F., Schiavon, G., & Solimini, C. (2007). Use of neural networks for
automatic classification from high resolution images. IEEE Transactions on Geosci-
ence and Remote Sensing, 45, 800–809.
Dell'Acqua, F. (2009). The role of SAR sensors. In P. Gamba, & M. Herold (Eds.), Global mapping
of human settlements: Experiences, datasets, and prospects. Boca Raton, Florida: CRC Press.
Deng, J. S., Wang, K., Deng, Y. H., & Qi, G. J. (2008). PCA-based land-use change detec-
tion and analysis using multitemporal and multisensor satellite data. International
Journal of Remote Sensing, 29, 4823–4838.
Dobson, M. C., Pierce, L. E., & Ulaby, F. T. (1996). Knowledge-based land-cover classifi-
cation using ERS-1/JERS-1 SAR composites. IEEE Transactions on Geoscience and Re-
mote Sensing, 34, 83–99.
Ehlers, M., Jadkowski, M. A., Howard, R. R., & Brostuen, D. E. (1990). Application of SPOT
data for regional growth analysis and local planning. Photogrammetric Engineering
and Remote Sensing, 56, 175–180.
Feiner, J., Salmeron, D., Joos, E., & Schmid, W. (2002). Priming sustainability: The Kun-
ming urban region development project. DISP, 151(38), 59–67.
Foody, G. M., & Mathur, A. (2004). Toward intelligent training of supervised image clas-
sifications: Directing training data acquisition for SVM classification. Remote Sens-
ing of Environment, 93, 107–117.
Friedl, M. A., & Brodley, C. E. (1997). Decision tree classification of land cover from re-
motely sensed data. Remote Sensing of Environment, 61, 399–409.
Friedl, M. A., McIver, D. K., Hodges, J., Zhang, X., Muchoney, D., Strahler, A., Woodcock,
C. E., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao, F., & Schaaf, C. (2002).
Global land cover mapping from MODIS: Algorithms and early results. Remote
Sensing of Environment, 83, 287–302.
Friedl, M. A., Sulla-Menashe, D., Tan, B., Schneider, A., Ramankutty, N., Sibley, A., & Huang,
X. (2010). MODIS Collection 5 global land cover: Algorithm refinements and charac-
terization of new datasets. Remote Sensing of Environment, 114, 168–182.
Friedl, M. A., Woodcock, C., Gopal, S., Muchoney, D., Strahler, A. H., & Barker-Schaaf, C.
(2000). A note on procedures used for accuracy assessment in land cover maps de-
rived from AVHRR data. International Journal of Remote Sensing, 21, 1073–1077.
Friedman, J., Hastie, T., & Tibshirani, R. (2000). Additive logistic regression. The Annals
of Statistics, 28, 337–374.
Friedman, S. Z., & Angelici, G. (1979). The detection of urban expansion from Landsat
imagery. Remote Sensing Quarterly, 1, 58–79.
Gao, F., Masek, J., Schwaller, M., & Hall, F. (2006). On the blending of the Landsat and
MODIS surface reflectance: Predicting daily Landsat surface reflectance. IEEE Trans-
actions on Geoscience and Remote Sensing, 44, 2207–2218.
Gao, F., Masek, J., Wolfe, R., & Huang, C. (2010). Building a consistent medium resolu-
tion satellite data set using moderate resolution imaging spectroradiometer prod-
ucts as a reference. Journal of Applied Remote Sensing, 4, 043526 (article).
Gluch, R. (2002). Urban growth detection using texture analysis on merged Landsat TM
and SPOT-P data. Photogrammetric Engineering and Remote Sensing, 68, 1283–1288.
Google Earth (2011). Satellite imagery for Chengdu (30°39′30.96″ N, 104°3′53.48″ E),
Xi'An (34°15′53.95″ N, 108°56′39.37″ E), and Kunming (25°2′15.80″ N, 102°43′
19.93″ E), China, multiple dates (software available online at: http://earth.google.com)
Griffiths, P., Hostert, P., Gruebner, O., & van der Linden, S. (2010). Mapping megacity
growth with multi-sensor data. Remote Sensing of Environment, 114, 426–439.
Guo, L., Chehata, N., Mallet, C., & Boukir, S. (2011). Relevance of airborne lidar and mul-
tispectral image data for urban scene classification using Random Forests. ISPRS
Journal of Photogrammetry and Remote Sensing, 66, 56–66.
Hansen, M. C., Townshend, J. R. G., Defries, R. S., & Carroll, M. (2005). Estimation of tree
cover using MODIS data at global, continental and regional/local scales. Interna-
tional Journal of Remote Sensing, 26, 4359–4380.
Hansen, M., Dubayah, R., & DeFries, R. (1996). Classification trees: An alternative to tradi-
tional land cover classifiers. International Journal of Remote Sensing, 17, 1075–1081.
Hansen, M., Stehman, S., Potapov, P., Loveland, T., Townshend, J., DeFries, R., Pittman,
K., Arunarwati, B., Stolle, F., Steininger, M., Carroll, M., & DiMiceli, C. (2008).
Humid tropical forest clearing from 2000 to 2005 quantified by using multi-
temporal and multiresolution remotely sensed data. Proceedings of the National
Academy of Sciences of the United States of America, 105, 9439–9444.
Henderson, F. M., & Xia, Z. G. (1997). SAR applications in human settlement detection,
population estimation and urban land use pattern analysis: A status report. IEEE
Transactions on Geoscience and Remote Sensing, 35, 79–85.
Howarth, P. J., & Boasson, E. (1983). Landsat digital enhancements for change detection
in urban environments. Remote Sensing of Environment, 13, 149–160.
Huang, C., Davis, L. S., & Townshend, J. R. G. (2002). An assessment of support vector ma-
chines for land cover classification. International Journal of Remote Sensing, 23, 725–749.
Huang, C., Goward, S., Schleeweis, K., Thomas, N., Masek, J., & Zhu, Z. (2009). Dynamics
of national forests assessed using the Landsat record: Case studies in eastern
United States. Remote Sensing of Environment, 113, 1430–1442.
704 A. Schneider / Remote Sensing of Environment 124 (2012) 689–704
Jaiantilal, A. (2009). Classification and regression by Random Forest — Matlab (soft-
ware available online at: http://code.google.com/p/randomforest-matlab)
Jensen, J. R., & Toll, D. L. (1982). Detecting residential land use development at the
urban fringe. Photogrammetric Engineering and Remote Sensing, 48, 629–643.
Kennedy, R. E., Cohen, W. B., & Schroeder, T. A. (2007). Trajectory-based change detec-
tion for automated characterization of forest disturbance dynamics. Remote Sensing
of Environment, 110, 370–386.
Knorn, J., Janz, A., Radeloff, V. C., Kuemmerle, T., Kozak, J., & Hostert, P. (2009). Land
cover mapping of large areas using chain classification of neighboring satellite im-
ages. Remote Sensing of Environment, 113, 957–964.
Kuemmerle, T., Chaskovskyy, O., Knorn, J., Radeloff, V. C., Kruhlov, I., Keeton, W. S., &
Hostert, P. (2009). Forest cover change and illegal logging in the Ukrainian Car-
pathians in the transition period from 1988 to 2007. Remote Sensing of Environ-
ment, 113, 1194–1207.
Li, X., & Yeh, A. G. O. (1998). Principal component analysis of stacked multi-temporal
images for the monitoring of rapid urban expansion in the Pearl River Delta. Inter-
national Journal of Remote Sensing, 19, 1501–1518.
Li, X., Yeh, A. G. O., Qian, J. P., Ai, B., & Qi, Z. (2009). A matching algorithm for detecting
land use changes using case-based reasoning. Photogrammetric Engineering and Re-
mote Sensing, 75, 1319–1332.
Lin, G. C. S. (2009). Scaling up regional development in globalizing China: Local capital
accumulation, land-centred politics, and reproduction of space. Regional Studies,
43, 429–447.
Lin, G. C. S., & Ho, S. P. S. (2005). The state, land system, and land development processes in
contemporary China. Annals of the Association of American Geographers, 95, 411–436.
Lippitt, C. D., Rogan, J., Li, Z., Eastman, R., & Jones, T. G. (2008). Mapping selective log-
ging in mixed deciduous forest: A comparison of machine learning algorithms.
Photogrammetric Engineering and Remote Sensing, 74, 1201–1211.
Liu, X., & Lathrop, R. G. (2002). Urban change detection based on an artificial neural
network. International Journal of Remote Sensing, 23, 2513–2518.
Long, H., Liua, Y., Wu, X., & Dong, G. (2009). Spatio-temporal dynamic patterns of farm-
land and rural settlements in Su–Xi–Chang region: Implications for building a new
countryside in coastal China. Land Use Policy, 26, 322–333.
Ma, L. J. C. (2004). Economic reforms, urban spatial restructuring, and planning in
China. Progress in Planning, 61, 237–260.
Masek, J. G. (2007). LEDAPS disturbance products: User's guide and algorithm (avail-
able online at: http://ledaps.nascom.nasa.gov/docs/pdf/LEDAPS_DI_ATBD.pdf)
Maxwell, S. (2004). Filling Landsat ETM + SLC-off gaps using a segmentation model ap-
proach. Photogrammetric Engineering and Remote Sensing, 70, 1109–1111.
McIver, D. K., & Friedl, M. A. (2001). Estimating pixel-scale land cover classification
confidence using non-parametric machine learning methods. IEEE Transactions on
Geoscience and Remote Sensing, 39, 1959–1968.
Menze, B. H., Kelm, M., Masuch, R., Himmelreich, U., Bachert, P., Petrich, W., &
Hamprecht, F. (2009). A comparison of random forest and its Gini importance
with standard chemometric methods for the feature selection and classification
of spectral data. BMC Bioinformatics, 10, 213 (article).
Michalak, W. Z. (1993). GIS in land use change analysis — Integration of remotely
sensed data into GIS. Applied Geography, 13, 28–44.
Mills, G. (2007). Cities as agents of global change. International Journal of Climatology,
27, 1849–1857.
Moller-Jensen, L. (1990). Knowledge-based classification of an urban area using tex-
ture and context information in Landsat TM imagery. Photogrammetric Engineering
and Remote Sensing, 56, 899–904.
Nemmour, H., & Chibani, Y. (2006). Multiple support vector machines for land cover
change detection: An application for mapping urban extensions. ISPRS Journal of
Photogrammetry and Remote Sensing, 61, 125–133.
Pal, M. (2005). Random forest classifier for remote sensing classification. International
Journal of Remote Sensing, 26, 217–222.
Pataki, D. E., Alig, R. J., Fung, A. S., Golubiewski, N. E., Kennedy, C. A., McPherson, E. G.,
Nowak, D. J., Pouyat, R. V., & Romero Lankao, P. (2006). Urban ecosystems and the
North American carbon cycle. Global Change Biology, 12, 2092–2102.
Potere, D. (2008). Horizontal positional accuracy of Google Earth's high-resolution im-
agery archive. Sensors, 8, 7973–7981.
Quinlan, J. R. (1993). C4.5: Programs for machine learning. New York: Morgan
Kaufmann Publishers.
Quinlan, J. R. (1996). Bagging, boosting, and C4.5. Proceedings of the 13th National Con-
ference on Artificial Intelligence (AAAI-96), 4–8 August 1996 (pp. 725–730). Portland,
Oregon: AAAI Press.
Ridd, M. K., & Liu, J. J. (1998). A comparison of four algorithms for change detection in
an urban environment. Remote Sensing of Environment, 63, 95–100.
Robinson, L., Newell, J. P., & Marzluff, J. A. (2005). Twenty-five years of sprawl in the
Seattle region: Growth management responses and implications for conservation.
Landscape and Urban Planning, 71, 51–72.
Rogan, J., & Chen, D. M. (2004). Remote sensing technology for mapping and monitor-
ing land-cover and land-use change. Progress in Planning, 61, 301–325.
Rogan, J., Franklin, J., Stow, D., Miller, J., Roberts, D., & Woodcock, C. (2008). Mapping
land cover modifications over large areas: A comparison of machine learning tech-
niques. Remote Sensing of Environment, 112, 2272–2283.
Rogan, J., Miller, J., Stow, D., Franklin, J., Levien, L., & Fischer, C. (2003). Land-cover
change monitoring with classification trees using Landsat TM and ancillary data.
Photogrammetric Engineering and Remote Sensing, 69, 793–804.
Romero, H., Ihl, M., Rivera, A., Zalazar, P., & Azocar, P. (1999). Rapid urban growth,
land-use changes and air pollution in Santiago, Chile. Atmospheric Environment,
33, 4039–4047.
Ryznar, R. M., & Wagner, T. W. (2001). Using remotely sensed imagery to detect urban
change — Viewing Detroit from space. Journal of the American Planning Association,
67, 327–336.
Sanli, F. B., Kurucu, Y., & Esetlili, M. T. (2009). Determining land use changes by
radar-optic fused images and monitoring its environmental impacts in Edremit re-
gion of western Turkey. Environmental Monitoring and Assessment, 151, 45–58.
Schneider, A., Friedl, M. A., Mciver, D. K., & Woodcock, C. E. (2003). Mapping urban
areas by fusing multiple sources of coarse resolution remotely sensed data. Photo-
grammetric Engineering and Remote Sensing, 69, 1377–1386.
Schneider, A., Friedl, M. A., & Potere, D. (2009). A new map of global urban extent from
MODIS satellite data. Environmental Research Letters, 4, 044003 (article).
Schneider, A., Friedl, M. A., & Potere, D. (2010). Mapping urban areas globally using
MODIS 500m data: New methods and datasets based on urban ecoregions. Remote
Sensing of Environment, 114, 1733–1746.
Schneider, A., Seto, K. C., & Webster, D. R. (2005). Urban growth in Chengdu, Western
China: Application of remote sensing to assess planning and policy outcomes. En-
vironment and Planning B, 32, 323–345.
Schneider, A., & Woodcock, C. E. (2008). Compact, dispersed, fragmented, extensive? A
comparison of urban expansion in twenty-five global cities using remotely sensed
data, pattern metrics and census information. Urban Studies, 45, 659–692.
Seto, K. C., & Kaufmann, R. K. (2003). Modeling the drivers of urban land use change in
the Pearl River Delta, China: Integrating remote sensing with socioeconomic data.
Land Economics, 79, 106–121.
Small, C., & Lu, J. (2006). Estimation and vicarious validation of urban vegetation abun-
dance by spectral mixture analysis. Remote Sensing of Environment, 100, 441–456.
Song, C., Woodcock, C. E., Seto, K. C., Pax-Lenney, M., & Macomber, S. A. (2001). Classi-
fication and change detection using Landsat TM data: When and how to correct at-
mospheric effects? Remote Sensing of Environment, 75, 230–244.
Stefanov, W. L., Ramsey, M. S., & Christensen, P. R. (2001). Monitoring urban land cover
change: An expert system approach to land cover classification of semiarid to arid
urban centers. Remote Sensing of Environment, 77, 173–185.
Storey, J., Scaramuzza, P., & Schmidt, G. (2005). Landsat 7 scan line corrector-off
gap-filled product development. Proceedings of the Pecora 16 Conference on Global
Priorities in Land Remote Sensing, 23–27 October 2005, Souix Falls, South Dakota.
Tatem, A. J., & Hay, S. I. (2004). Measuring urbanization pattern and extent for malaria
research: A review of remote sensing approaches. Journal of Urban Health—Bulletin
of the New York Academy of Medicine, 81, 363–376.
Todd, W. J. (1977). Urban and regional land use change detected by using Landsat data.
Journal of Research of the U.S. Geological Survey, 5, 529–534.
Tso, B., & Mather, P. M. (2001). Classification methods for remotely sensed data. New
York: Taylor and Francis (332 pp).
Ulbricht, K. A., & Heckendorff, W. D. (1998). Satellite images for recognition of land-
scape and landuse changes. ISPRS Journal of Photogrammetry and Remote Sensing,
53, 235–243.
United States Geological Survey (USGS) (2011). Global Visualization Viewer (GLOVIS)
(Landsat data archive, available online at: http://glovis.usgs.gov)
Vapnik, V. N. (1995). The nature of statistical learning theory. New York: Springer-Verlag.
Webster, D. R. (2002). On the edge: Shaping the future of peri-urban East Asia, discussion
paper. The urban dynamics of East Asia series. Stanford University, Stanford, California:
Asia Pacific Research Center Publication (68 pp.).
Woodcock, C. E., Macomber, S. A., Pax-Lenney, M., & Cohen, W. B. (2001). Monitoring
large areas for forest change using Landsat: Generalization across space, time
and Landsat sensors. Remote Sensing of Environment, 78, 194–203.
Woodcock, C. E., & Ozdogan, M. (2004). Trends in land cover mapping and monitoring.
In G. Gutman, A. Janetos, & C. Justice (Eds.), Land change science: Observing, moni-
toring, and understanding trajectories of change on the Earth's surface. Dordrecht,
The Netherlands: Kluwer Academic Publishers.
Yang, X., & Lo, C. P. (2002). Using a time series of satellite imagery to detect land use
and land cover changes in the Atlanta, Georgia metropolitan area. International
Journal of Remote Sensing, 23, 1775–1798.
Yuan, F., Sawaya, K. E., Loeffelholz, B. C., & Bauer, M. E. (2005). Land cover classification
and change analysis of the Twin Cities (Minnesota) Metropolitan Area by multi-
temporal Landsat remote sensing. Remote Sensing of Environment, 98, 317–328.
Zhang, C., Li, W., & Travis, D. (2007). Gap-fill of SLC-off Landsat ETM + satellite image
using a geostatistical approach. International Journal of Remote Sensing, 28,
5103–5122.