Content uploaded by Chunyang He
Author content
All content in this area was uploaded by Chunyang He
Content may be subject to copyright.
Land-Use/Land-Cover Change Detection Using
Improved Change-Vector Analysis
Jin Chen, Peng Gong, Chunyang He, Ruiliang Pu, and Peijun Shi
Abstract
Change-vector analysis (CVA) is a valuable technique for land-
use/land-cover change detection. However, how to reasonably
determine thresholds of change magnitude and change di-
rection is a bottleneck to its proper application. In this paper,
a new method is proposed to improve CVA. The method (the
improved CVA) consists of two stages, Double-Window Flexible
Pace Search (DFPS), which aims at determining the threshold
of change magnitude, and direction cosines of change vectors
for determining change direction (category) that combines
single-date image classification with a minimum-distance
categorizing technique. When the improved CVA was applied
to the detection of the land-use/land-cover changes in the
Haidian District, Beijing, China, Kappa coefficients of “change/
no-change” detection and “from-to” types of change detection
were 0.87 and greater than 0.7, respectively, for all kinds of
land-use changes. The experimental results indicate that the
improved CVA has good potential in land-use/land-cover
change detection.
Introduction
Land-use/land-cover change is an important field in global en-
vironmental change research. Inventory and monitoring of
land-use/land-cover changes are indispensable aspects for fur-
ther understanding of change mechanism and modeling the
impact of change on the environment and associated ecosys-
tems at different scales (Turner et al., 1995; William et al.,
1994). Remote sensing is a valuable data source from which
land-use/land-cover change information can be extracted effi-
ciently. In the past two decades, there has been a growing trend
in the development of change detection techniques using re-
mote sensing data. A number of techniques for accomplishing
change detection using satellite imagery have been formu-
lated, applied, and evaluated, which can be broadly grouped
into two general types (Singh, 1989; Jensen, 1996; Coppin and
Bauer, 1996; Ding et al., 1998; Johnson and Kasischke, 1998):
(1) those based on spectral classification of the input data such
as post-classification comparison (e.g., Mas, 1999) and direct
two-date classification (e.g., Li and Yeh, 1998); and (2) those
based on radiometric change between different acquisition
dates, including (a) image algebra methods such as band differ-
encing (Weismiller et al., 1977), ratioing (Howarth and Wick-
ware, 1981), and vegetation indices (Nelson, 1983); (b) regres-
sion analysis (Singh, 1986); (c) principal component analysis
(Byrne et al., 1980; Gong, 1993); and (d) change-vector analysis
(CVA) (Malila, 1980). Based on a mixture of categorical and ra-
diometric change information, hybrid approaches have also
been proposed and evaluated (Colwell and Weber, 1981). Selec-
tion of an appropriate change-detection technique, in practice,
often depends on the requirement of information, data avail-
ability and quality, time and cost constraints, and analysis skill
and experience (Johnson and Kasischke, 1998). Among those
radiometric change-based approaches, change-vector analysis
is a useful method for land-use/land-cover change detection be-
cause it not only can avoid the shortcomings of those type 1
approaches, such as cumulative error in image classification of
an individual date, but also it can find changed pixels using
more, even all, the bands and provide “from-to” type of change
information. In the past few years, its advantages and potential
have been demonstrated in some case studies (Michalek et al.,
1993; Lambin and Strahler, 1994a; Lambin and Strahler, 1994b;
Sohl, 1999). Thus, it has recently been adopted as the basis
for the initial 250-meter Land-Cover Change Product using
MODIS Data (Zhan et al., 2000).
However, like other radiometric change approaches, CVA
also has several drawbacks that limit its use. These include
●A strict requirement for reliable image radiometry. Because
CVA
is based on pixel-wise radiometric comparison, the accuracy
of image radiometric correction for alleviating the impacts
caused by disturbing factors such as different atmospheric con-
ditions, solar angle, soil moisture and vegetation phenology,
etc., is more critical for
CVA
than for spectral classification
approaches. However, there exists no valid radiometric correc-
tion method that can be used to reduce the effects of all dis-
turbing factors efficiently, especially for vegetation phenology.
Similar acquisition dates in different years are therefore chosen
to reduce this type of disturbance in
CVA
. This strict requirement
for data acquisition limits the broad application of
CVA
.
●A lack of automatic or semiautomatic methods to effectively
determine the threshold of change magnitude between change
and no-change pixels. Although determination of the optimal
threshold between change and non-change pixels is considered
as the most important task as well as the greatest challenge of
CVA
(Ding et al, 1998; Johnson and Kasischke, 1998; Smits and
Annoni, 2000), the threshold in a specific
CVA
analysis is often
determined according to empirical strategies (Fung and Le-
Drew, 1988) or from manual trial-and-error procedures. This-
usually requires a more experienced image analyst and a long
trial time (Bruzzone and Prieto, 2000).
●Discrimination of different phenomenological types of change
is problematic when the number of bands involved is large.
The methods of discriminating change type in existing literature
can be grouped into three classes: (1) trigonometric functions
of vector angle in two spectral dimensions (Malila, 1980), (2)
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2003 369
Photogrammetric Engineering & Remote Sensing
Vol. 69, No. 4, April 2003, pp. 369–379.
0099-1112/03/6904–369$3.00/0
© 2003 American Society for Photogrammetry
and Remote Sensing
J. Chen, C. He, and P. Shi are with the Key Laboratory of Envi-
ronmental Change and Natural Disaster, Beijing Normal Uni-
versity, Ministry of Education of China, Beijing 100875, China.
P. Gong and R. Pu are with the International Institute for Earth
System Science, Nanjing University, China; and the Center for
Assessment and Monitoring of Forest and Environmental Re-
sources, 145 Mulford Hall, University of California, Berkeley,
CA 94720-3114 (gong@nature.berkeley.edu).
sector coding in more than two spectral dimensions (Virag and
is used. Based on the above introduction of CVA and its draw-
Colwell, 1987), and (3) principal component analysis in a multi-
backs, an improved CVA for land-use/land-cover change detec-
temporal space (Lambin and Strahler, 1994a). In most
CVA
ap-
tion was developed (Figure 1).
plications, the change category is mainly distinguished and
assigned by a combination of “”or“” symbols (for increase,
Threshold Search for Identifying Change Pixels
(for decrease) of each computational band and image interpre-
Traditionally, the threshold of change magnitudeis empirically
tation (Virag and Colwell, 1987; Michalek et al., 1993; Johnson
and Kasischke, 1998; Sohl, 1999). When
CVA
is applied in this
determined. This is subjective and varies from person to per-
manner (called “sector coding”), there exist two problems.
son. To overcome this problem, we developed a Double-Win-
First, sector coding can discriminate 2
n
sectors when the number
dow Flexible Pace Search (DFPS) algorithm. This method is
of computational bands is n(Virag and Colwell, 1987). If there
based on selecting a threshold from training samples con-
are nine land-cover types found at two different dates, respec-
taining all possible kinds of changes. The assumption is that
tively, and all kinds of changes between those land covers
the training samples are representative to the entire study area.
probably occur in the period, the number of change types is 9
Thus, a threshold leading to the maximum accuracy of change
872. This means that one sector code certainly represents
detection within the training samples is considered optimal for
more than one change type even if six bands of
TM
are used (2
6
the entire study area. The flowchart of DFPS is shown in Figure 2
64), which may lead to an assignment error of change cate-
gory. Second, it is a strenuous and time-consuming work to
and its main steps are described in the following section.
discriminate and interpret change categories represented by
sector codes with the increase of computational bands (n) be-
Selection of Typical Sample Areas of Land-Use/Land-Cover
cause the number of sector codes increases geometrically.
Change
In light of the abovementioned drawbacks, especially the In the process of optimal threshold search, it is important to
second and third ones, this paper presents an improved CVA for make sure that training samples are as representative as possi-
land-use/land-cover change detection, which includes (1) a ble for all change classes. First, the change magnitude (㥋⌬G㥋)
semiautomatic method, named Double-Window Flexible Pace image is calculated from the two original images of different
Search (DFPS), which aims at determining efficiently the thresh- dates, following precise image-to-image registration and radio-
old of change magnitude, and (2) a new method of determining metric normalization. Then, some typical change areas are cho-
change direction (change category) which combines a single sen as training samples from the change magnitude image. The
image classification and a minimum-distance categorization two original images can be displayed along with the change
based upon the direction cosines of the change vector. The im- magnitude image to assist in the identification of training sam-
proved CVA approach is applied and validated by a case study ples for land-use changes. The criteria for selecting sample ar-
of land-use change detection in the Haidian district of Beijing, eas are (1) training samples should cover as many as possible
China, using multi-temporal TM data. The rest of this paper change types (note that a change type is identified only by com-
outlines the proposed method and gives a detailed description paring the color and context between the two original images
of a case study applying the improved CVA. The conclusions and because the real change type has not been categorized at this
remarks are summarized. stage), (2) training samples should include only change pixels,
Method
Change-Vector Analysis (CVA)
Malila (1980) gave a general idea of change-vector analysis
(CVA). A change vector can be described by an angle of change
(vector direction) and a magnitude of change from date 1 to date
2 (Jensen, 1996). If a pixel’s gray-level valuesin two images on
dates t
1
,t
2
are given by G(g
1
,g
2
,,g
n
)
T
and H(h
1
,h
2
, ...,
h
n
)
T
, respectively, and nis the number of bands, a change vec-
tor is defined as
⌬GHG
冢
h
1
g
1
h
2
g
2
h
n
g
n
冣
(1)
where ⌬Gincludes all the change information between the two
dates for a given pixel, and the change magnitude 㥋⌬G㥋is com-
puted with
얍⌬G얍
冪
(h
1
g
1
)
2
(h
2
g
2
)
2
(h
n
g
n
)
2
. (2)
It represents the total gray-level difference between two dates.
The greater 㥋⌬G㥋is, the higher is the possibility of change. A
decision on change is made based on whether the change mag-
nitude exceeds a specific threshold. Once a pixel is identified
as change, the direction of ⌬Gcan be examined further to deter-
mine the type of change. The type of change is often identified
using the angle of the vector in two spectral dimensions, or sec- Figure 1. Flowchart of land-use/land-cover change detec-
tor codes if more than two spectral dimensions are involved. tion based on improved change vector analysis.
The geometric concept of CVA is applicable to any number of
spectral bands, no matter what measurement scale of radiance
370
April 2003
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Figure 3. An example of selecting a training patch in the
change magnitude image by comparing the two original
images.
Figure 2. Flowchart of the Double-windows Flexible Pace
Search method.
and (3) training samples should be encircled by no-change pix-
els as “islands.” Figure 3 shows an example of selecting one
training sample in the change magnitude image by comparing Figure 4. An example showing the relationship between ac-
the two original images. curacy of change detection and threshold decreasing in one
training sample. The shadow area corresponds to an optimal
The threshold for identifying change and no-change pixels
threshold value.
can be determined by searching an optimal value of change
magnitude to obtain the maximum accuracy of change detec-
tion within the training samples. Obviously, as the threshold
of change magnitude decreases, the number of change pixels in-
side training samples will increase and the accuracy of change
detection will be improved. However, it should be noted that
the possibility of no-change pixels outside the training sam-
ples identified as change pixels would also increase, leading to
higher commission errors. There exists such a threshold, with
which all the pixels inside the training samples are detected
correctly as change pixels and the highest accuracy is obtained
in the training samples while many no-change pixels outside
the training samples are also identified incorrectly as change
pixels. Figure 4 gives an example illustrating the relationship Figure 5. Example of Double-window.
between accuracy of change detection and threshold decrease
in one training patch. In consideration of this situation, an out-
side boundary is created through buffering in a GIS for each
training patch, forming a double area called a double-window magnitude in the first search process. The first search pace (in-
(Figure 5). The outside boundary (outside window) is used to crement) P
1
may be calculated according to the following
prevent the threshold from being too low. formula:
Determination of the Search Range and Pace P
1
(ba)/ m(3)
The threshold search range can be set as a difference between
the minimum value (a) and the maximum value (b) of change where mis a positive integer, which determines the number of
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2003
371
potential thresholds in a search process and can be set manu-
Change Type Discrimination
Discrimination of change types plays an important role inally. The potential thresholds to detect change pixels from the
training samples in the search process are given within the change detection. Problems were noted with the existing
CVA
(Cohen and Fiorella, 1998). Cohen and Fiorella (1998) pointedrange of [a,b]asbP
1
,b2P
1
, ... . It should be noted that the
size of the manually set mdoes not affect the search efficiency out the possibility of other angle measurements for the change
vector in three or more spectral dimensions and noticed the im-and the final results. A large mincreases the number of poten-
tial thresholds during one search, but it decreases the number portance of a “reference image” in change-typediscrimination.
A new method of determining change type (category) is devel-of searches.
oped in this study, which combines single image classification
(as a reference image) with minimum-distance categorizing
Definition and Calculation of Test Parameter based on direction cosines of change vectors.
A success rate of change detection is defined to evaluate the
The direction of a vector can be described by a series of co-
performance of each potential threshold during one search
sine functions in a multi-dimensional space. This series is
process for identifying change/no-change pixels. The success
called direction cosines (Hoffmann, 1975). The direction of the
rate (L
k
) is calculated for a potential threshold of k: i.e.,
change vector containing change type information can also be
defined by the cosine function of angles between the vector and
L
k
(A
k1
A
k2
)100
A% (4) each spectral axis. The change vector direction of one change
pixel corresponds to one and only one point in a multi-dimen-
sional space constituted by direction cosines. If some typical
where A
k1
is the number of change pixels detected inside all feature points (hereafter called seed points) and their corres-
training patches (in the inner windows), A
k2
is the number of ponding change types in the space of direction cosines are
change pixels which are detected incorrectly in the outside known, the change type of a change pixel can be determined by
boundary of all training patches (in the outside windows), and a supervised classification on the basis of its proximity to those
Ais the total number of pixels within all training patches (in the seed points. Obviously, to obtain seed points and their corres-
inner windows). In order to keep L
k
from becoming negative, ponding change types becomes the key to change-type dis-
the outside border (outside window) should be set to one or two crimination for all change pixels. For two images acquired on
pixels for each training sample. From this definition, it can be different dates t
1
and t
2
, after rigorous radiometric normaliza-
seen that the double-window concept is useful in controlling tion, we may assume that the spectral feature difference be-
the commission error caused by low thresholds, because low tween any two kinds of land-use/land-cover types on either
thresholds increases A
k2
and reduces L
k
.date are similar to their spectral change features from t
1
to t
2
.
After all L
k
for all mthresholds in one search are obtained, With this assumption, we can first carry out a land-use/land-
the maximum and minimum values of L
k
can be found and des- cover classification from the image of a selected date, either t
1
ignated as L
max
and L
min
during this search process. If the two or t
2
(the reference image). This reference image should be the
parameters do not satisfy the conditions described in the next one with more ancillary information available for accurate im-
section, a new search begins, the new search range is set in the age classification. The spectral difference vectors between any
range [k
max
P
1
,k
max
P
1
], and a new smaller search pace is two land-use/land-cover types in the reference image can then
set based on the modified search range with Equation 3. Here, be calculated based on the classified land-use/land-cover
k
max
is the potential threshold value corresponding to L
max
in types and transplanted into the direction cosine space. These
the search process. points can be considered as the seed points (mean vectors) in a
minimum-distance classification for land-use/land-cover
Condition to Exit the Iteration change-type discrimination because their features are typical
The steps described in the previous two sections represent an and their coordinates and change types are known.
iterating process, which will terminate when the following for-
mula is satisfied: Definition of Direction Cosines of Change Vector
Suppose X(x
1
,x
2
,x
3
, ..., x
n
)isanndimensional vector, its mag-
L
max
L
min
␦
(5) nitude can be calculated as
where L
max
and L
min
are the maximum and minimum values of 앚X앚
冪
x
12
x
22
x
32
x
42
x
n2
. (6)
the success rate in one search process, and ␦is an acceptable
error constant. The condition indicates that the change of If the angles between Xand each axis are (
1
,
2
,
3
,
4
, ,
n
),
search pace has little influence on the result of change/no- respectively, then its direction can be described by cosine
change pixel detection. The threshold corresponding to L
max
is functions of these angles as (Hoffmann, 1975)
considered as an optimal threshold for change detection.
The method of Double-Window Flexible Pace Search
(
DFPS
) is a semiautomatic method for determining the optimal cos
1
x
1
앚X앚, cos
2
x
2
앚X앚, ..., cos
n
x
n
앚X앚(7)
threshold of change detection that only requires the involve-
ment of the image analyst during the selection of typical train-
ing samples of land-use/land-cover change. The advantages of Using Equation 7, the direction of a change vector can be
represented as one and only one point, defined as a new vectorthis method can be summarized as (1) the optimal threshold
can be obtained automatically after the selection of typical Z(cos
1
, cos
2
, cos
3
, ..., cos
n
) in the direction cosine space. All
of the change pixels have their corresponding points in thistraining samples of land-use/land-cover change, (2) the com-
mission error caused by over decreasing the threshold can be space. According to this definition, the determination of
change types is turned into a classification problem of pointscontrolled effectively through the double-window technique,
and (3) search efficiency is improved with the flexible and var- in the direction cosine space. Moreover, using the direction co-
sines instead of angle measurements can avoid the difficulty ofied search pace. Although the new method may perform better
compared with some previous empirical methods, it should be “baseline” establishment for angle measurement such as using
Tasseled-cap transformation to build the Plane of Soil (definednoted that the new method is dependent on training samples,
which to a certain degree relies on the experience and skills of by brightness and wetness axes) and Vegetation (defined by
brightness and greenness axes) (Crist and Cicone, 1984).the image analyst.
372
April 2003
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Extraction of all Possible Change Types and Their reference image without strenuous and time-consuming train-
ing work. Thus, the minimum-distance classifier was selectedCorresponding Seed Points
For two images acquired on different dates t
1
and t
2
, afterrigor- for this study to identify land-use/land-cover change types
based on its effectiveness and its simple requirement of onlyous radiometric normalization,the spectral feature differences
between any two kinds of land-use/land-cover types on either the estimation of the mean vector of each spectral class (Rich-
ards and Jia, 1999). With this classifier, an unknown pixel isdate are similar to their spectral change features from t
1
to t
2
.
This assumption can be denoted as assigned to a certain class or unclassified class based on a mini-
mum distance to means of all candidate classes when the dis-
tance is within a certain threshold. The minimum-distance
⌬W
ij
P
j
Q
i
⇔⌬T
ij
H
j
G
i
(8) classifier was employed in this study to categorize change
types in three steps: (1) calculating direction cosines of the
where P
j
and Q
i
are gray value vectors of land-use/land-cover spectral change vector for each change pixel according to
types jand iin either image of date t
1
or t
2
,⌬W
ij
is their spectral Equation 7; (2) calculating Euclidean distances of the changed
difference vector, G
i
and H
j
are gray-level vectors of land-use/ pixels to seed points corresponding to all possible change
land-cover types iand jat different dates t
1
and t
2
, and ⌬T
ij
is types in the direction cosine space, and those seed points are
spectral change vector from date t
1
to t
2
(the same as the change obtained through land-use/land-cover classification in a refer-
vector as described above). Based on the land-use/land-cover ence image; and (3) determining change types of land-use/land-
classification from the reference image, the spectral difference cover by applying the minimum-distance rule. Like a conven-
vectors between any two kinds of land-use/land-cover types in tional minimum-distance classifier, the unclassified pixels are
the reference image are calculated. Using Equation 8, these those falling outside the threshold range based on the mean
spectral difference vectors can be thought of as equivalents of and standard deviation of every class. Those pixels may stand
spectral change vectors of those land-use/land-cover changes for new change types, which are not included in all possible
from date t
1
to t
2
. The mean of spectral difference vectors repre- change types obtained through land-use/land-cover classifica-
sents a typical feature of various changes between any two tion in the reference image.
kinds of land-use/land-cover types, and direction cosine val-
ues of the mean of spectral difference vectors can be specified
Case Study— Land-Use Change Detection in the Haidian District,
as seed points in the direction cosine space. Moreover, the
standard deviation of spectral difference vectors that belong to
Beijing, China
the same type of land-use/land-cover change can also be used
Study Area and Data
to determine the threshold of an unclassified class in a mini- The Haidian district (one district of the city of Beijing) is lo-
mum-distance classifier. cated in the west and northwest part of Beijing with a popula-
The mean of spectral difference vectors and their standard tion of 1.25 million and an area of 426 km
2
(Figure 6). Three-
deviations can be calculated based on change samples in a par- quarters of the district lie in a plain with an elevation of less than
ticular “from ito j” change class with an assumption that the 50 meters, while hills cover the remaining quarter of the dis-
probability distributions of spectral differences are normal. trict. It is an educational and cultural center, a vegetable pro-
This is usually possible because each spectral class (land-use/ duction area for Beijing, and a famous Information Technology
land-cover class) is usually normally distributed in the spec- (
IT
) industrial area of China. With a rapid growth in the econ-
tral space (this is the same assumption used in a maximum- omy during the past 20 years, tremendous land-use/land-cover
likelihood classification; Richards and Jia, 1999). We used EX
i
,
DX
i
,EX
j
, and DX
j
to denote the mean and standard deviation
vectors of land-use/land-cover types iand jin multispectral
space (n bands), respectively: i.e.,
EX
i
(EX
i1
,EX
i2
, ..., EX
in
)
T
(9)
DX
i
(DX
i
,DX
i2
, ..., DX
in
)
T
(10)
EX
j
(EX
j1
,EX
j2
, ..., EX
jn
)
T
(11)
DX
j
(DX
j
,DX
j2
, ..., DX
jn
)
T
(12)
Based on stochastic theory, if pand qare mutually inde-
pendent and subject to a normal distribution, then zpqis
also subject to a normal distribution. Thus, the mean and the
standard deviation of spectral difference vector of land-use/
land-cover types jand ican be simply deduced from EX
i
,DX
i
,
EX
j
, and DX
j
as follows:
EX
ij
(EX
j1
EX
i1
,EX
j2
EX
i2
, ..., EX
jn
EX
in
)
T
(13)
DX
ij
(DX
j1
DX
i1
,DX
j2
DX
i2
, ..., DX
jn
DX
in
)
T
(14)
Change-Type Discrimination Based on Minimum-Distance
Classification
Types of change pixels can be discriminated by classifying di-
rection cosines of their change vectors. As described in the pre- Figure 6. The study area (shadow area) in Beijing.
vious section, the mean and the standard deviation vectors of
each land-use/land-cover change type can be obtained from the
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2003
373
changes have taken place during this period. The area was se- and 2, change magnitudes between 1991 and 1997 were calcu-
lated and shown in Figure 7. The change magnitudes rangelected as a study area to evaluate the performance of the im-
proved
CVA
approach. In order to eliminate the effects of dis- from 0 to 177 (Figure 7b) and most change magnitudes fall un-
der 100. From Figure 7a, it is obvious that greater values occurturbance factors, especially vegetation phenological
differences, two scenes of Landsat
TM
image data (path/row: in the north part of the image.
123/32, 06 May 1991 and 16 May 1997) covering the whole
study area with good image quality were collected and used. Threshold Determination Using the Double-Window Flexible
Moreover, some auxiliary data were also collected, including Pace Search Method
1:50,000-scale topographic maps from 1972, a 1:100,000-scale Based on a preliminary comparison of the two images and inter-
land-use map from 1991,
GPS
survey data from 1999, and a
SPOT
pretation, typical training samples of changes were selected
panchromatic image from May 1997. with their 1-pixel outer buffer boundaries (some of the bigger
ones shown in Plate 1a). Then the search range was set to [0,
Image Prepreprocessing
180] and the pace as 20 initially. The Double-Window Flexible
Pace Search method was used to determine the threshold of
Geometric Registration change magnitude. The threshold search process iterated until
High precision geometric registration of the multi-temporal im- the success rate difference between the maximum and the min-
age data is a basic requirement for change detection (Gong et imum value was less than 0.1 percent. As a result, the threshold
al., 1992; Dai and Khorram, 1998; Morisette and Khorram, of change magnitude was obtained as 33.4 with a success rate
2000). First, the 1991
TM
image was transformed to the
UTM
of 63.78 percent. The search process was recorded in Table 2,
projection at a 30- by 30-m resolution, using second-order poly- and the search range changed five times with the paces of 20, 5,
nomial and bilinear interpolation. Twenty-one ground control 1, 0.5, and 0.1. The number of thresholds tested totaled 36. The
points were collected from the 1:50,000-scale topographic change pixels in the study area at threshold 33.4 were ex-
maps. The root-mean-square error (
RMSE
) was less than 1 pixel. tracted and shown in Plate 1b.
Then the 1997
TM
image was registered to the 1991 image by
image-to-image registration with an
RMSE
of registration of less
Change Type Discrimination
than 0.5 pixels using 29 tie points. Finally, the boundary of the
Haidian district, digitized into a
GIS
database, was used to clip Land-Use Classification Using the 1991 TM Image
the study area from the images. Referring to the land-use map of 1991 and the results of the un-
supervised classification of the 1991
TM
image, all land classes
Image Radiometric Normalization in the study area were grouped into nine classes: urban land,
A problem associated with the use of multi-temporal remote barren land (mostly are transitional areas), water, paddy field,
sensing data for change detection is that the multi-temporal wheat land, vegetable land, other agricultural land, shrub and
data are usually acquired under different sun angle, and atmos- grassland, and forest land. Based on this classification system,
pheric and soil moisture conditions. Ideally, such data should a land-use/land-cover classification of the 1991
TM
image (ref-
be radiometrically normalized so that the effects of those unde- erence image) was carried out using the maximum-likelihood
sirable conditions can be minimized or eliminated, which is classifier and, after postpreprocessing (Plate 1a), an overall ac-
critical to change detection, especially to those methods based curacy of 89 percent and a Kappa coefficient of 0.82 were
on radiometric change (Hall et al., 1991; Jensen et al., 1995). In reached.
this study, the 1991 image was chosen as a primary reference
image because the detailed ground reference information was Change-Type Discrimination
available in that year. Then a Scattergram Controlled Regres- Based on the results of the land-use/land-coverclassification of
sion (
SCR
) method was used to develop the radiometric nor- the 1991
TM
image, the mean and standard deviation of each
malization equations ( y
k
a
k
x
k
b
k
) (Table 1) (Elvidge et al.,land-use/land-cover type were first extracted. Then the means
1995). Here the independent variable x is the pixel values of of spectral difference vectors between any two kinds of land-
the 1997
TM
image while the dependent variable y is the nor- use/land-cover types and their deviations were also calculated.
malized pixel values of the 1997
TM
image. The R
2
of equations There were 72 (9 8) possible land-use/land-cover change
for most bands were greater than 0.95 except for band 2 with an types in the study area. The program of change-type discrimi-
R
2
of 0.88. Such equations were applied to normalize the 1997 nation was performed nine times, respectively, using the
image to assure that the images from 1991 and 1997 were com- change pixel data belonging to the nine land-use/land-cover
parable in terms of radiometric characteristics. types in 1991, and the change types of all change pixels were
labeled, including a new type of unclassified class. By impos-
Change/No-Change Pixel Detection
ing the land-use/land-cover change image, including change
pixels and change types, on the land-use image classified from
Change Magnitude Calculation the 1991
TM
image, the land-use/land-cover image of 1997 was
Bands 1, 2, 3, 4, 5, and 7 of the
TM
1991 image and TM 1997 obtained and is shown in Plate 1d. Only 4 percent of the change
image were used in the
CVA
analysis. According to Equations 1 pixels labeled as unclassified fell outside the threshold range of
mean 2 times the standard deviation for each class. This sug-
gests that a few new change types had occurred during this pe-
T
ABLE
1. P
ARAMETERS
U
SED FOR
R
ADIOMETRIC
N
ORMALIZATION FOR
1997
riod, which were not included in all possible change types ob-
TM I
MAGE
tained through land-use/land-cover classification in 1991.
Comparing the land classification images in 1991 with
Band Slope (a) Intercept (b) R
2
1997 (Plates 1c and 1d), the changes in the study area during
1 0.780 4.705 0.953
this period may be characterized by two aspects. (1) Land-use
2 0.786 2.380 0.882
changes within agricultural sectors took place notably in the
3 0.759 1.40 0.959
north and northwest area, especially from other agricultural
4 0.791 2.491 0.959
land changed to water pools, which were used mostly as fish-
5 0.648 11.669 0.981
ponds. These changes resulted from increasing urban popula-
7 0.808 2.325 0.991
tion and the rising living standard of city dwellers. (2) Urban
374
April 2003
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Figure 7. Change magnitude between 1991 and 1997 in the Haidian District, Beijing. (a) Change magnitude distribution. (b)
Histogram of change magnitudes.
land expanded quickly in the urban-rural fringe area and the In order to show the superiority of the improved
CVA
pro-
posed in this paper, it is necessary to compare this methodnorthern study area by encroaching upon agricultural land, es-
pecially paddy field and vegetable land and other agricultural with other change-detection methods. Because the traditional
CVA
is computation intensive and is dependent on the user’slands, as a result of economic development.
experiences to a certain extent (to select the threshold and in-
terpret change types), it was not used for comparison in this
Accuracy Assessment
In order to evaluate the performance of the proposed improved study. A comparison between the improved
CVA
and post-clas-
sification comparison, which is the most widely used method
CVA
method, the accuracy of change detection was estimated at
both “change/no-change” detection and “from-to change” de- in change detection, was carried out. The post-classification
comparison was performed based on a supervised classificationtection levels. This is different from a conventional accuracy
assessment method that applies only to change detection errors using the 1991 and 1997 images, respectively, and the 1991
training data. At the “change/no-change” detection level,(Khorram et al., 1998; Biging et al., 1998). At the “change/no-
change” detection level, a random sampling technique was Table 6 gives the error matrix of post-classificationcomparison
based on the same 2,400 sample pixels together with the auxil-used. Because change pixels only occupy a small proportion of
the image, the number of sampling pixels belonging to no- iary information also used in the improved
CVA
. It can be seen
that the post-classification comparison causes an overestima-change were much more than that of change pixels. Table 3
shows an error matrix of “change/no-change” detection con- tion of change due to the cumulative errors of classification on
each individual date. For example, 118 unchanged pixelsstructed from 2,400 sample pixels together with the auxiliary
information visually interpreted from the 1997 image and
SPOT
(only 32 in the improved
CVA
) have been assigned to change in
post-classification comparison. A comparison of Table 3 andpanchromatic image of May 1997. A Kappa coefficient of 0.87
and an overall accuracy of 96.3 percent were achieved. Table 6 indicates that the improved
CVA
is better than the con-
ventional method at the “change/no-change” detection levelAt the “from-to change” detection level, the accuracy as-
sessment was carried out based on those change pixels belong- (the kappa coefficient is 0.87 for the improved
CVA
and 0.69 for
post-classification comparison). Table 7 gives the accuracy as-ing to different land-use/land-cover types in 1991. The sam-
pling pixels used for accuracy assessment were selected using sessment of the “from-to change” detection for all kinds of
change types using the post-classification comparison. Thethe randomly stratified sampling method, in which the de-
tected change pixels (Figure 7b) were used for stratification. same sample pixels in Table 5 were used to construct error ma-
trices and calculate accuracies. It is evident that the post-clas-More than 60 samples were selected for each land-use/land-
cover type in 1991 according to Congalton (1991). The ground sification comparison is more problematic than the improved
CVA
in almost every change type because a lot of unlikely land-truth wasproduced from visual interpretation of the1997image
and a
SPOT
panchromatic image acquired in May 1997. Table 4 use conversions and overestimation of change were produced
from cumulative classification errors. The above results sug-shows the change-detection error matrix from wheat land in
1991 to other land-use types in 1997 as an example of all other gest that the improved
CVA
is effective and applicable for land-
use/land-cover change detection.change-detection error matrixes. Table 5 gives the accuracy as-
sessment of the “from-to change” detection for all kinds of
change types. It indicates that the proposed method is effective
Conclusion and Discussion
In this paper, we report some improvements made to the tradi-for change-type discrimination because the kappa coefficients
for all kinds of change types exceeded 0.7. tional change vector analysis in two aspects, determination of
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2003
375
Plate 1. The selected change sample areas and results of change/no-change detection. (a) Typical change
sample areas (red) with outer no-change buffer boundary (blue). (b) Change pixels detected by the improved
CVA
method. Land-use maps of the Haidian District, generated (c) by the maximum-likelihood classifier using the
1991
TM
image, and (d) by the improved
CVA
with the 1997
TM
image.
376
April 2003
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
T
ABLE
2. R
ESULTS OF
D
OUBLE
-W
INDOW
F
LEXIBLE
P
ACE
S
EARCH
Range 160-20 Pace 20 Range 60-20 Pace 5 Range 40-30 Pace 1 Range 35-33 Pace 0.5 Range 34-33 Pace 0.1
Threshold Success Rate Threshold Success Rate Threshold Success Rate Threshold Success Rate Threshold Success Rate
160 0.21 55 51.35 39 62.88 34.5 63.73 33.9 63.76
140 0.65 50 55.6 38 63.14 34 63.75 33.8 63.77
120 1.88 45 59.36 37 63.24 33.5 63.77 33.7 63.77
100 8.13 40 62.25 36 63.43 33.6 63.77
80 22.69 35 63.75 35 63.67 33.5 63.77
60 46.07 30 62.76 34 63.75 33.4 63.78
40 62.25 25 60.4 33 63.68 33.3 63.75
20 55.73 32 63.24 33.2 63.72
31 62.65 33.1 63.69
T
ABLE
3. E
RROR
M
ATRIX FOR
“C
HANGE
/N
O
-C
HANGE
”D
ETECTION
U
SING THE
I
MPROVED
CVA
Reference Change
Change No-Change Commission
Pixels Pixels Sum Error
Classified Change Change Pixels 368 32 400 8%
No-Change Pixels 57 1943 2000 2.85%
Sum 425 1975 2400
Omission Error 13.4% 1.6%
Overall Accuracy 96.29% Kappa Coefficient 0.8698
T
ABLE
4. C
HANGE
D
ETECTION
E
RROR
M
ATRIX FROM
W
HEAT
L
AND IN
1991
TO OTHER
T
YPES IN
1997 U
SING THE
I
MPROVED
CVA
Reference Change
Wheat Land
Other Shrub
Urban Barren Paddy Agricultural Vegetable and Forest
1991 1997 Land Land Water Field Land Land Grassland Land Sum
Classified Wheat
Change Land Urban Land 13 114
Barren Land 2 20 123
Water 12 214
Paddy Field 3 20 124
Other Agricultural Land 0
Vegetable Land 4 11 15
Shrub and Grassland 77
Forest Land 11
Sum 19 20 15 22 1 12 8 1 98
Overall Accuracy 85.71% Kappa Coefficient 0.8264
T
ABLE
5. T
HE
A
CCURACY
A
SSESSMENT OF
“F
ROM
-T
O
”C
HANGE
D
ETECTION
U
SING THE
I
MPROVED
CVA
Land Use in 1991 Land Use in 1997 Sampling Pixels Overall Accuracy Kappa Coefficient
Urban Land Other Land-Use Types 60 84.56 0.7156
Barren Land 90 87.33 0.838
Water 80 92.86 0.8401
Paddy Field 90 81.36 0.7796
Other Agricultural Land 90 86.72 0.8108
Wheat Land 98 85.71 0.8264
Vegetable Land 90 89.04 0.8516
Shrub and Grassland 90 92.05 0.8668
Forest Land 90 85.71 0.7456
the threshold of change magnitude and change type discrimi- 0.7, respectively, for all kinds of change types. The results sug-
gest that the improved
CVA
is effective and has potential innation. When the improved method was employed for the Hai-
dian District, Beijing with Landsat
TM
imagery acquired in 1991 land-use/land-cover change detection.
Good data quality (similar acquisition dates in differentand 1997, the Kappa coefficients of “change/no-change” detec-
tion and “from-to change” detection were 0.87 and greater than years and cloud-free) and image radiometric normalization
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2003
377
T
ABLE
6. E
RROR
M
ATRIX FOR
“C
HANGE
/N
O
-C
HANGE
”D
ETECTION
U
SING
P
OST
-C
LASSIFICATION
C
OMPARISON
Reference Change
Change Pixels No-Change Pixels Sum Commission Error
Classified Change Change Pixels 321 118 439 26.88
No-Change Pixels 104 1857 1961 5.30
Sum 425 1975 2400
Omission Error 24.47 5.97
Overall Accuracy 90.75% Kappa Coefficient 0.6867
T
ABLE
7. T
HE
A
CCURACY
A
SSESSMENT OF
“F
ROM
-T
O
”C
HANGE
D
ETECTION
U
SING
P
OST
-C
LASSIFICATION
C
OMPARISON
1991 1997 Sample Pixels Overall Accuracy Kappa Coefficient
Urban Land Other Land-Use Types 60 62.15 0.4923
Barren Land 90 76.37 0.6421
Water 80 86.91 0.7902
Paddy Field 90 74.24 0.6332
Other Agricultural Land 90 81.23 0.7216
Wheat Land 98 79.24 0.681
Vegetable Land 90 69.32 0.5637
Shrub and Grassland 90 73.47 0.5718
Forest Land 90 73.25 0.5689
ods and Applications (Ross S. Lunetta and Christopher D. Elvidge,
have a strong impact on the final change-detection result be-
editors), Sleeping Bear Press, Inc., New York, N.Y., pp.281–308.
cause the proposed method is based on an assumption of radio-
metric similarity among multi-temporal remotely sensed data.
Bruzzone, L., and D.F. Prieto, 2000. Automatic analysis of the difference
image for unsupervised change detection, IEEE Transactions on
In addition, the selection of typical change sample areas and
Geoscience and Remote Sensing, 38(3):1171–1182.
single-image classification are also critical to the search of an
optimal threshold of change magnitude and identification of
Byrne, G.F., P.F. Crapper, and K.K. Mayo, 1980. Monotoring land-cover
by principal component analysis of multitemporal Landsat data,
change types. In our study, thanks to the good quality of image
Remote Sensing of Environment, 10:175–184.
data, rigorous radiometric and geometric correction, and rich
ancillary materials, relatively satisfactory results were ob-
Cohen, W.B., and M. Fiorella, 1998. Comparison of methods for de-
tecting conifer forest change with Thematic Mapper imagery, Re-
tained. However, desirable imaging conditions are not always
mote Sensing Change Detection: Environmental Monitoring Meth-
guaranteed; therefore, the change-detection accuracy of the im-
ods and Applications (Ross S. Lunetta and Christopher D. Elvidge,
proved CVA approach may decrease.
editors), Sleeping Bear Press, Inc,, New York, N.Y., pp. 89–102.
In most land-use/land-cover change-detection applica-
Colwell, J., and F. Weber, 1981. Forest change detection, Fifteenth
tions, collection of rich ancillary information in all years is usu-
International Symposium on Remote Sensing of Environment, 11-
ally difficult. This may lead to error accumulation due to unre-
15 May, Ann Arbor, Michigan, pp. 65–99.
liable classification of images from all dates. The new method
Congalton, R.G., 1991. A review of assessing the accuracy of classifica-
of determining change type developed in this paper takes ad-
tions of remotely sensed data, Remote Sensing of Environment,
vantage of the spectral differences between images acquired in
37:35–46.
two dates t
1
to t
2
, and the rich ancillary information from one
Coppin, P.R., and M.E. Bauer, 1996. Digital change detection in forest
date for classification for reference. It is based on two assump-
ecosystems with remote sensing imagery, Remote Sensing Re-
tions: (1) the direction cosine of the change vector contains
views, 13:207–234.
change type information, and (2) the spectral feature difference
Crist, J.B., and R.C. Cicone, 1984. A physically-based transformation
between any two kinds of land-use/land-cover types on either
of thematic mapper data - the TM tasseled cap, IEEE Transactions
date are similar to their spectral change features from t
1
to t
2
.If
on Geoscience and Remote Sensing, 22(3):256–263.
these assumptions are reasonable, no ancillary information
Dai, X., and S. Khorram, 1998. The effects of image misregistration on
from another date is needed to perform change type discrimi-
accuracy of remotely sensed change detection, IEEE Transactions
nation. Although we need some ancillary reference data on
on Geoscience and Remote Sensing, 36(5):1566–1577.
change to evaluate the change detection result, in practice, this
requirement is much less than in other change-detection meth-
Ding, Y., C.D. Elvidge, and Ross S. Lunetta., 1998. Survey of multispec-
tral methods for land cover change detection analysis, Remote
ods. The improved CVA is an effective approach, especially
Sensing Change Detection: Environmental Monitoring Methods
when ancillary information for classification is only available
and Applications (Ross S. Lunetta and Christopher D. Elvidge,
on one date.
editors), Sleeping Bear Press, Inc., New York, N.Y., pp. 21–39.
Elvidge, C.D., and Y. Ding, 1995. Relative radiometric normalization
Acknowledgments
of Landsat multispectral scanner (MSS) data using an automatic
This research was partially supported by the National Basic
scattergram-controlled regression, Photogrammetric Engi-
Science Program of China (
2001
CB
309404
), and NASA’s land-
neering & Remote Sensing, 61:1015–1026.
cover and land-use change grants (NAG
5-9231
;NCC
5-492
).
Fung, T., and E. LeDrew, 1988. The determination of optimal threshold
levels for change detection using various accuracy indices, Photo-
References
grammetric Engineering & Remote Sensing, 54:1449–1454.
Gong, P., 1993. Change detection using principal component analysisBiging, Gregory S., David R. Colby, and Russell G. Congalton, 1998.
Sampling system for change detection accuracy assessment, Re- and fuzzy set theory, Canadian Journal of Remote Sensing,
19(1):22–29.mote Sensing Change Detection: Environmental Monitoring Meth-
378
April 2003
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Gong, P., E.F. LeDrew, and J.R. Miller, 1992. Registration noise reduc- Morisette, J.T., and S. Khorram, 2000. Accuracy assessment curves for
satellite-based change detection, Photogrammetric Engineering &tion in difference images for change detection, International Jour-
nal of Remote Sensing, 13(4):773–779. Remote Sensing, 66:875–880.
Hall, F., D. Strebel, J. Nickeson, and S. Goetz, 1991. Radiometric rectifi- Nelson, R.F., 1983. Detecting forest canopy change due to insect activity
cation: Toward a common radiometric response among multidate, using Landsat MSS, Photogrammetric Engineering & Remote
multisensor images. Remote Sensing of Environment, 35(1):11–27. Sensing, 49:1303– 1314.
Hoffmann, B., 1975. About Vectors, Dover Publications, Inc., New York, Richards, J.A., and X. Jia, 1999. Remote Sensing Digital Image Analysis,
N.Y., 134 p. Third Edition, Springer-Verlag, Berlin, Heidelberg, Germany,
363 p.
Howarth, P.J., and G.M. Wickware, 1981. Procedures for change detec-
tion using Landsat digital data, International Journal of Remote Singh, A., 1986. Change detection in the tropical forest of northeastern
Sensing, 2:277–291. India using Landsat, Remote Sensing and Tropical Land Manage-
ment (M.J. Eden and J.T. Parry, editors), Chichester Wiley Press,
Jensen, J.R., 1996. Introductory Digital Image Processing: A Remote London, United Kingdom, pp. 237–254.
Sensing Perspective, Second Edition, Prentice Hall, Upper Saddle
River, New Jersey, 316 p. ———, 1989. Digital change detection techniques using remotely-
sensed data, International Journal of Remote Sensing,
Jensen, J.R., K. Rutchey, M.S. Koch, and S. Narumalani, 1995. Inland
10:989–1003.
wetland change detection in the Everglades Water Conservation
Area 2A using a time series of normalized remotely sensed data, Smits, Paul C.; and Alessandro Annoni, 2000, Toward specification-
Photogrammetric Engineering & Remote Sensing, 61:199–209. driven change detection, IEEE Transactions on Geoscience and
Remote Sensing, 38(3):1484–1488.
Johnson, R.D., and E.S. Kasischke, 1998. Change vector analysis: A
technique for the multispectral monitoring for land cover and Sohl, T.L., 1999. Change analysis in the United Arab Emirates: An
condition, International Journal of Remote Sensing, 19:411–426. investigation of techniques, Photogrammetric Engineering & Re-
mote Sensing, 65:475–484.
Lambin, E.F., and A.H. Strahler, 1994a. Change vector analysis in multi-
temporal space: A tool to detect and categorize land-cover change Turner II, B.L, D. Skole, Steven Sanderson, Gunther Fischer, Louise
processes using high temporal-resolution satellite data, Remote Fresco, and Rik Leemans, 1995. Land-Use and Land-Cover Change
Sensing of Environment, 48:231–244. Science/Research Plan, IGBP Report No. 35, URL: http://
———, 1994b. Indicators of land-cover change for change vector analy- www.geo.ucl.ac.be/LUCC/scienceplan/lucc.html.
sis in multitemporal space at coarse spatial scales, International Virag, L.A., and J.E. Colwell, 1987. An improved procedure for analysis
Journal of Remote Sensing, 15:2099–2119. of change in Thematic Mapper image-pairs, Proceedings of the
Li, X., and A.G.O. Yeh, 1998. Principal component analysis of stacked Twenty-First International Symposium on Remote Sensing Envi-
multi-temporal images for the monitoring of rapid urban expan- ronment, 26-30 October, Ann Arbor, Michigan, pp. 1101–1110.
sion in the Pearl River Delta, International Journal of Remote Weismiller, R.A., S.Y. Kristof, D.K. Scholz, P.E. Anuta, and S.A. Momin,
Sensing, 19:1501–1518. 1977. Change detection in coastal zone environments, Photogram-
Malila, W.A., 1980. Change vector analysis: an approach for detecting metric Engineering & Remote Sensing, 43:1533–1539.
forest changes with Landsat, Proceedings of the 6th Annual Sym- William, E.R., B.M. William, and B.L. Turner II, 1994. Modeling land
posium on Machine Processing of Remotely Sensed Data, 03-06 use and land cover as part of global environmental change, Cli-
June, Purdue University, West Lafayette, Indiana, pp. 326–335. matic Change, 28:45–64.
Mas, J.F., 1999. Monitoring land-cover changes: a comparison of change Zhan, X., R. Defries, J.R.G. Townshend, C. Dimiceli, M. Hansen, C.
detection techniques, International Journal of Remote Sensing,Huang, and R. Sohlbert, 2000. The 250m global land cover change
20:139–152. product from the Moderate Resolution Imaging Spectroradiometer
Michalek, J.L., T.W. Wagner, J.J. Luczkovich, and R.W. Stoffle, 1993. of NASA’s Earth Observing System, International Journal of Re-
Multispectral change vector analysis for monitoring coastal ma- mote Sensing, 21:1433 –1460.
rine environments, Photogrammetric Engineering & Remote Sens-
ing, 59:381–384. (Received 07 February 2001; revised and accepted 11 February 2002)
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2003
379
