Professional Documents
Culture Documents
Median Change Vector Analysis Algorithm For Land Use Land Cover Change Detection From Remote Sensing Data
Median Change Vector Analysis Algorithm For Land Use Land Cover Change Detection From Remote Sensing Data
Improved change vector analysis (ICVA) has recently been promoted as an effec-
tive algorithm for multi-class change detection. Unlike the conventional change
vector analysis (CVA) that works on two-dimensional data, the ICVA works on
multidimensional data. However, ICVA has limitations when the change vector is
fraught with similar direction cosine values. In this article, a new algorithm, named
median change vector analysis (MCVA) has been proposed for multi-class change
detection. The algorithm is based on an enhanced 2n-dimensional feature space
comprising direction cosine values of both the change vector and the median vec-
tor, which allows for more accurate detection of change classes than those obtained
from ICVA. As a case study, the proposed algorithm has been implemented on
Landsat-7 Enhanced Thematic Mapper Plus (ETM+) images of a typical Indian
city and surrounding areas for land-cover change detection.
1. Introduction
Digital image change detection is based on the analysis of difference in the spectral
response of a pixel between two images acquired at different points in time on the
same geographical location. If the difference is more than the statistically determined
threshold value, one land-cover class is assumed to be changed to the other land-
cover class (Singh 1989). Based on this concept, a number of algorithms have been
developed. These include image differencing (ID), image ratioing (IR), principal com-
ponent analysis (PCA), combination of image differencing and principal component
analysis (IDPCA) and so on. These algorithms suffer from the selection of appropriate
thresholds, which at times is quite subjective. Further, these also portray binary change
detection.
Alternatively, a number of other algorithms, namely, artificial neural network
(ANN) change detection (Gopal and Woodcock 1996, Dai and Khorram 1999),
a curve-theorem based on change detection algorithm (Yue et al. 2002), improved
change vector analysis (ICVA) (Chen et al. 2003), independent component analysis
(Zhong and Wang 2006) and unsupervised change vector analysis (CVA) (Kontoes
2008) may also be used. Some of these algorithms produce binary change whereas
others produce multi-class change (i.e. from class 1- to class 2-type changes).
Conventionally, post-classification change detection (Hall et al. 1991) has been used
for the detection of multi-class change but has limitation in that its accuracy of change
detection depends on the quality of the classification of the images at the two different
points of time. Chen et al. (2003) have proposed a novel multi-class change detection
algorithm, named ICVA, wherein classification of only one of the images at two points
in time may yield multi-class changes. The ICVA algorithm states that the change
vector can be represented as a unique point in the feature space z, containing the
mean direction cosine values of change vectors. The mean direction cosine values are
determined from the average of spectral responses of pixels belonging to a set of land-
cover classes mapped in one of the images acquired at two points of time. However,
as demonstrated in this article, the change vector may only be represented as a unique
point in the feature space z, if it contains the direction cosine values of median vector
also in addition to those in change vector. This new feature space has been denoted
as p, containing the direction cosine values of both change and median vectors. The
resulting algorithm based on this feature space has been captioned as median change
vector analysis (MCVA).
2. Concept of MCVA
In a typical CVA problem, the change in land-cover classes is determined based on
both the magnitude and the direction of change (Malila 1980). Let there be two images
consisting of two bands of data: one taken at time 1 and denoted as i1 = (x11 , x12 ) and
the other taken at time 2 and denoted as i2 = (x21 , x22 ), where x1j and x2j are the
spectral responses of pixels in band j (j = 1, 2). The magnitude |m| and the direction θ
of change for two-dimensional data can then be defined as follows:
|m| = (x11 − x21 )2 + (x12 − x22 )2 , (1)
x11 − x21
cos θ = . (2)
|m|
Chen et al. (2003) have extended the conventional CVA to an n-dimensional data
set and called the algorithm as ICVA. Let i1 = (x11 , x12 , . . . , x1n ) and i2 = (x21 ,
x22 , . . . , x2n ) be n-dimensional images acquired at times 1 and 2. The change is
represented as a vector m = (d 1 , d 2 , . . ., dn ) and the magnitude of m is then
n 2
|m| = dj , (3)
j=1
where dj represent the difference in spectral responses of a pixel in band j of the images
taken at two different points in time,
b2
r
m1
v q
m2
w b1
O
b3
dj
cos θj = . (5)
|m|
The direction cosine space of the unknown pixel and the known direction cosine space
of various land-cover change classes as obtained in the training stage are then input
to a supervised classification process to classify the remote-sensing image into those
land-cover change classes (Chen et al. 2003).
However, the change vector in ICVA may not represent a unique point in the direc-
tion cosine space as can be seen from the spectral plot of two-date three band images
shown in figure 1. In this figure, axes b1 , b2 and b3 represent the spectral response of
pixels in band 1, band 2 and band 3, respectively. Considering two pixels, let points q
and w denote the spectral responses of land-cover classes at these pixels in date 1 image
and points r and v represent spectral responses of land-cover classes of same pixels in
date 2 image. It can be seen that both the change vectors m1 and m2 are parallel to
each other. This indicates that both the vectors have the same direction cosine values
in each band. Therefore, the change vectors m1 and m2 represent the same point in the
direction cosine space z, which may not be unique.
On the other hand, in the proposed MCVA, the change vector based on direction
cosine values has been augmented with a median vector (figure 2). In this figure, Oe
(k1 ) and Os (k2 ) denote the medians of Oqr and Owv, respectively. Clearly, the
median vectors k1 and k2 are not parallel and indicate different directions. Thus,
modification of the change vector by including medians also, the change vector of
an unknown pixel, can be represented as a unique point in the measurement space p.
b2
r
e m1
v k2 q
k1
m2 s
w b1
O
b3
Figure 2. Representation of change vectors (m1 and m2 ) with median vectors (k1 and k2 ) in
different bands.
(a) 77° 57′ 0″ E 78° 0′ 0″ E 78° 3′ 0″ E 78° 6′ 0″ E (b) 77° 57′ 0″E 78° 0′ 0″ E 78° 3′ 0″ E 78° 6′ 0″ E
30° 24′ 0″ N
30° 24′ 0″ N
30° 24′ 0″ N
30° 24′ 0″ N
N N
30° 21′ 0″ N
30° 21′ 0″ N
30° 21′ 0″ N
30° 21′ 0″ N
30° 18′ 0″ N
30° 18′ 0″ N
30° 18′ 0″ N
30° 18′ 0″ N
30° 15′ 0″ N
30° 15′ 0″ N
30° 15′ 0″ N
Figure 3. FCC (R: NIR, G: Red and B: Green) of Landsat 7 ETM + images of study area:
(a) year 2000 image; (b) year 2005 image.
15 m spatial resolution for the years 2000 and 2005 have been used to record land-
cover changes through visual interpretation of the two images. These changes have
been considered as actual changes and form the basis of assessing the accuracy of
changes detected from the proposed algorithm.
Change detection
Extract mean direction cosines of change and Find direction cosine values of
median vectors of all possible change classes the change and median vector
x1j + x2j
aj = , (6)
2
where x1j and y2j are the spectral reflectance of pixels in band j (= 1, 2, . . . , n)
of dates 1 and 2.
2. Determine magnitude of change at each pixel according to the equation (3).
3. Determine the magnitude of n-dimensional median vector k = (a1 , a2 , . . . , an )
as
n 2
|k| = aj . (7)
j=1
610 A. Varshney et al.
4.2 Thresholding algorithm for the separation of change and no-change region
Kontoes (2008) has suggested that a 3 × 3 pixel kernel-based thresholding algorithm
is able to discriminate no-change and change classes more precisely than by setting
thresholding limits directly on the basis of mean (μ) and standard deviation (σ ).
Therefore, in this study also, the same thresholding algorithm with the following
implementation steps has been used:
1. If the value of magnitude (m) of a pixel in the change vector is less than or
equal to μ + (γ −1)σ (γ is a user-defined constant value ranging for confidence
interval at mean to 95% of the magnitude data set), then difference values (dj )
are replaced with 0 in all the bands.
2. If the value of magnitude of a pixel in the change vector is greater than or equal
to μ + γ σ , then difference value is retained.
3. For other values of magnitude greater than μ + (γ −1)σ and less than μ + γ σ ,
the number of pixels in the kernel that have less than or equal to μ + (γ −1)σ
are determined.
If the number of pixels in the kernel are more than 5 (a threshold value used in this
study), then the probability that a pixel lies in the no-change region is more than the
probability that the pixel lies in the change region. Hence, the pixel is assigned to the
no-change class.
aj
cos αn+j = . (8)
|k|
The direction cosine values of pixels in the change vector as well as in median vector
are combined to produce a new 2n-dimensional change vector measurement space p,
given as η = (cos θ1 , cos θ2 , . . . , cos θn , cos αn+1 , . . . , cos α2n ). The values of pixels in
this vector will lie between –1 and 1.
4.4 Extraction of mean direction cosine values of change and median vectors of all
possible change classes from reference image
The following are the steps to extract mean directional cosine values of pixels
belonging to all possible change classes:
(8) for each possible change class. For example, if there are 5 classes in the ref-
erence image, then there will be a maximum of 25 possible combinations with
5 combinations of no-change classes. Thus, there will be a total of 20 possible
change classes and one no-change class.
4.5 Detection of multi-class change using minimum distance to mean decision rule
The classical minimum distance to mean decision rule has been applied to detect the
multi-class changes based on the following steps:
1. Compute the Euclidean distances between direction cosine values of a pixel and
the mean direction cosine values of possible change classes using equation (9).
n 2
n
2
l= cos θj − cos θj + cos αn+j − cos αn+j , (9)
j=1 j=1
Table 1. Mean direction cosine values of change and median vectors for some change classes.
Thick vegetation to −0.05 −0.12 −0.1 −0.82 −0.5 −0.22 0.3 0.24 0.17 0.79 0.41 0.17
thin vegetation
Thin vegetation to 0.05 0.12 0.1 0.82 0.5 0.22 0.3 0.24 0.17 0.79 0.41 0.17
thick vegetation
Built-up to sand −0.29 −0.4 −0.48 −0.43 −0.45 −0.36 0.39 0.36 0.36 0.46 0.48 0.38
Sand to thick 0.31 0.39 0.5 0.02 0.47 0.53 0.39 0.34 0.32 0.59 0.45 0.31
vegetation
30° 24′ 0″ N
Legend
No change
Thick vegetation–thin vegetation
Thick vegetation–built-up
Thick vegetation–sand/soil
30° 21′ 0″ N
30° 21′ 0″ N
30° 18′ 0″ N
Built-up–sand/soil
Sand/soil–thick vegetation
Sand/soil–thin vegetation
Sand/soil–built-up
30° 15′ 0″ N
30° 15′ 0″ N
Figure 5. Multi-class change map from MCVA with 0.3 γ constant value (the class before the
hyphen is the ‘from’ class, and the class after the hyphen is the ‘to’ class).
Note: MCVA, median change vector analysis.
(Congalton 1991). From the change error matrix, three measures, namely, the overall
accuracy of change, the quantity disagreement and the allocation disagreement have
been determined to portray the quality of multi-class change (Pontius and Millones
2011). A total of 529 testing pixels (207 belong to no-change class and 322 belong to
various change classes) have been selected to generate the change error matrix.
The overall accuracies of change from the post-classification, ICVA and MCVA
algorithms have been obtained as 56%, 60% and 64%, respectively. The values of quan-
tity disagreement and allocation disagreement for the corresponding algorithms have
been found to be (25%, 19%), (29%, 11%) and (25%, 11%). These values indicate that
the MCVA algorithm has yielded the highest overall accuracy of change and the lowest
quantity disagreement and allocation disagreement amongst all the algorithms for the
Median change vector analysis 613
Post-classification 56 25 19
ICVA 60 29 11
MCVA 64 25 11
Note: ICVA, improved change vector analysis; MCVA, median change vector
analysis.
Post-
classification ICVA MCVA
No change 69 74 77 78 77 78
Thick vegetation–thin vegetation 79 68 54 50 58 74
Thick vegetation–built-up 50 67 35 35 40 44
Thick vegetation–soil/sand 31 53 44 48 44 50
Thin vegetation–thick vegetation 50 58 67 61 67 63
Thin vegetation–built-up 68 45 45 56 48 58
Thin vegetation–soil/sand 56 48 56 54 62 57
Built-up−thick vegetation – 0 – 0 – 0
Built-up−thin vegetation – 0 – 0 – 0
Built-up−soil/sand – 0 – 0 – 0
Soil/sand−thick vegetation 29 65 45 61 57 78
Soil/sand−thin vegetation 35 78 54 79 66 87
Soil/sand−built-up 66 79 34 55 41 57
Note: UA, user accuracy; PA, producer accuracy; ICVA, improved change vector analysis;
MCVA, median change vector analysis.
present experiment (table 2). This clearly demonstrates the superlative performance of
the proposed change detection algorithm over the other change detection algorithms
used here.
The user’s and producer’s accuracies of change classes have been listed in table 3.
For this case study, it can be seen that the accuracy values of change classes from
thick vegetation to thin vegetation, sand/soil to thick vegetation and sand/soil to thick
vegetation detected in MCVA algorithm are higher than those produced from ICVA
algorithm (table 3).
6. Conclusion
In this study, a new CVA algorithm, named as MCVA, has been proposed and its
performance was assessed with the other two established change detection algorithms,
post-classification and ICVA for multi-class change detection. The MCVA algorithm
has yielded higher accuracy (64%) than that obtained from the ICVA (60%) and post-
classification (56%) for the experimental data set considered for change detection. The
following specific conclusions can be derived from this study:
614 A. Varshney et al.
1. The ICVA and MCVA algorithms use only one image to detect multi-class
change thereby reducing the efforts in classifying two images acquired at
different times.
2. The feature space suggested in the proposed MCVA algorithm appears better
and more effective than that suggested in the ICVA algorithm for multi-class
change detection.
Nevertheless, for the use of the proposed algorithm at operational level, further exper-
iments on a wide array of data sets in different field conditions are recommended.
Moreover, the usage of a non-parametric classifier in place of minimum distance to
mean classifier may result in the increase in accuracy of multi-class change detection
based on a single image.
References
CHEN, J., PENG, G., HE, C., PU, R. and SHI, P., 2003, Land-use/land-cover change detec-
tion using improved change-vector analysis. Photogrammetric Engineering and Remote
Sensing, 69, pp. 369–379.
CONGALTON, R., 1991, A review of assessing the accuracy of classification of remotely sensed
data. Remote Sensing of Environment, 37, pp. 35–46.
DAI, X.L. and KHORRAM, S., 1999, Remotely sensed change detection based on artificial neural
networks. Photogrammetric Engineering and Remote Sensing, 65, pp. 1187–1194.
GOPAL, S. and WOODCOCK, C.E., 1996, Remote sensing of forest change using artificial neural
networks. IEEE Transactions on Geoscience and Remote Sensing, 34, pp. 398–404.
HALL, F.G., BOTKIN, D.B., STREBEL, D.E., WOODS, K.D. and GOETZ, S.J., 1991, Large-scale
patterns of forest succession as determined by remote sensing. Ecology, 72, pp. 628–640.
KONTOES, C.C., 2008, Operational land cover change detection using change vector analysis.
International Journal of Remote Sensing, 29, pp. 4757–4779.
LEONARDO, P., FRANCEISCO, G., SOBRINO, J.A., JAUAN, C., JIMENEZ, M. and HAYDEE,
K., 2006, Radiometric correction effects in Landsat multi-date/multi-sensor change
detection studies. International Journal of Remote Sensing, 27, pp. 685–704.
MALILA, W.A., 1980, Change vector analysis: an approach for detecting forest changes with
Landsat. In Proceedings of the 6th Annual Symposium on Machine Processing of
Remotely Sensed Data, 3–6 June 1980, West Lafayette, IN (West Lafayette: Purdue
University), pp. 326–335.
PONTIUS, R.G.J. and MILLONES, M., 2011, Death to Kappa: birth of quantity disagreement
and allocation disagreement for accuracy assessment. International Journal of Remote
Sensing, 32, pp. 4407–4429.
SINGH, A., 1989, Digital change detection techniques using remotely-sensed data. International
Journal of Remote Sensing, 10, pp. 989–1003.
SUNDERASAN, A., VARSHNEY, P.K. and ARORA, M.K., 2007, Robustness of change detec-
tion algorithms in the presence of registration errors. Photogrammetric Engineering and
Remote Sensing, 73, pp. 375–384.
YUE, T.X., CHEN, S.P., XU, B., LIU, Q.S., LI, H.G., LIU, G.H. and YE, Q.H., 2002, A curve-
theorem based approach for change detection and its application to Yellow River Delta.
International Journal of Remote Sensing, 23, pp. 2283–2292.
ZHONG, J. and WANG, R., 2006, Multi-temporal remote sensing change detection based
on independent component analysis. International Journal of Remote Sensing, 27,
pp. 2055–2061.
Copyright of Remote Sensing Letters is the property of Taylor & Francis Ltd and its content may not be copied
or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission.
However, users may print, download, or email articles for individual use.