Change detection methodology based on region classification fusion

Abstract

In this paper, several classification methods are presented and a fusion structure is included to improve the final classification performance. The definition of "layer" and the method to create it are then introduced. Based on "layer", a multiple level change detection algorithm is proposed, which gives the details of the changes in each region and is demonstrated to be an easy, effective and reliable method. Experimental results are provided using RADARSAT images, which have been registered with the automated registration algorithm of A.U.G. Signals that is currently available under the distributed processing system www.signalfusion.com.

Full text

Page 1

CHANGE DETECTION METHODOLOGY BASED ON
REGION CLASSIFICATION FUSION

Ting Liu, George Gigli, and George A. Lampropoulos
A.U.G. Signals, 73 Richmond St. West
Toronto, ON, Canada
tliu,gigli,lampro@augsignals.com

Abstract - In this paper, several classification methods
are presented and a fusion structure is included to
improve the final classification performance. The
definition of “layer” and the method to create it are then
introduced. Based on “layer”, a multiple level change
detection algorithm is proposed, which gives the details
of the changes in each region and is demonstrated to be
an easy, effective and reliable method. Experimental
results are provided using RADARSAT images, which
have been registered with the automated registration
algorithm of A.U.G. Signals that is currently available
under the distributed processing system
www.signalfusion.com.

Keywords: Distributed Processing, Change Detection, Fusion,
Classification

1 Introduction
Change detection is the process of identifying differences
in the state of an object or phenomenon by observing it at
different times. It is useful in such diverse applications as
land use change analysis, monitoring of shifting
cultivation, assessment of deforestation, crop stress
detection and so on. It is essential for studying changes on
the earth’s surface. Such changes may determine the rate
of change for disaster management (e.g. flooding), ice
monitoring, earthquake prediction and monitoring, urban
planning etc.

Remotely sensed data are now able to estimate changes
with very high accuracy. The accuracy is proportional to
the image resolution, i.e. the higher the resolution of the
images used, the higher the accuracy of the change
detection. There are several sensors used for change
detection. SAR sensors offer the advantage of providing
additional phase information that may be used for change
detection. This is due to the fact that the pixels are
complex numbers. When the pixel-to-pixel phase
information is being used we say that this change
detection process is based on interferometry. When only
the amplitude of the images is used this process is called
photogrammetric change detection.

Change detection may be applied directly on images by
using only the pixel amplitude or both the magnitude and
phase, or transformed pixel values. The well-known
change detection techniques are image differencing,
image ratioing, image regression, Principal Component
Analysis (PCA), wavelet decomposition, change vector
analysis and so on. In topographic change detection, for
example if we want to study changes in a region where the
water level changes, we are interested in studying only the
changes between the two regions (land or water) [1, 2].
Hence, all land pixels may be assigned one value and all
the water pixels another value. In this case, study of
changes is much easier and all unnecessary image land or
water information has been eliminated through an image
segmentation transformation.

To detect the changes for each region, classification
should be performed first. There exist many classification
methods. In this paper, we used three methods, which are
thresholding, fuzzy C-mean and decision tree. To improve
the overall performance, the decisions resulted from
different classification algorithms could be fused before
performing the change detection.

The remainder of the paper is organized as follows. A
detailed topographic change detection method based on
region classification is described in Section 2. Section 3
outlines the fusion structure. The definition of “layer” is
introduced in Section 4. Section 5 discusses the
distributed computing technique. Some simulations are
given in Section 6. In Section 7, the conclusions of the
paper are drawn.

2 Region Calssification
Region classification is a widely used method for
extracting information on surface land cover from
remotely sensed images. The resulting cartography is
helping decision makers in different research fields. There
exist a lot of image classification methods. The change
detection approach that will be proposed in section III is a
kind of post-classification method. In this paper, the
classification methods we used are: thresholding, fuzzy C-
means (FCM) and decision trees.

Page 2

2.1 Thresholding
Considering a grayscale image, it is possible to do the
classification by applying the thresholding technique
using the map histogram. Thresholding permits the
distinction of relevant topographic information, such as
the lakes, rivers, wetlands, wooded areas, eskers, roads,
etc., from contours and grid lines. The map thresholding
classification technique is based on the fact that different
textures have different mean gray values on the map. This
technique is defined as follows. If a pixel represents the
texture of interest, we set its value to “1” in the new
classified image, and all the other pixels are set to “0”,
such as
),(1
),(

where
),(
yxf
is the pixel value in the new classified
image, and
),(
yxr
is the original pixel value.
are gray-scale values used as thresholds. Normally, we are
interested in more than one region. In this case, different
values will be assigned to
(
f
distinguish them. The most appropriate threshold values
have to be determined by the operator, since these values
may vary according to the printing and scanning specifics.

Take a look at Figure 1, in which there are two
RADARSAT images taken in May and August 1997.
These images were provided by the Defence Research and
Development Canada (DRDC)-Ottawa. These images
were registered by the automatic registration algorithm of
A.U.G. Signals Ltd that is available through the
distributed computing at www.signalfusion.com. Roughly
there are three regions in these images: deep water,
shallow water and land. We can easily see the differences
of water levels due to flooding of the river in May. We
extract the regions of interest from Figure 1 and display
them in Figure 2. To apply the thresholding method to
find the exact water and land regions, we have to
determine the threshold first. Pick up some small regions
with known classes (water or land) from the two images.
The pixels in these regions are used as the training data.
The histogram of these training data will then be plotted.
Suppose there are N regions needed to be classified, the
histogram should have N peaks. The thresholds should be
set as the lowest points between every two consecutive
amplitude peaks in the histogram. Figure 4 gives the
classification results of these two images using this
thresholding method, where the black regions represent
the deep water, grey ones are the shallow water, and the
white regions stand for the lands.

,
),(),(0


><
≤≤
g
=
ji
ji
gyxrandyxrfor
gyxrgfor
yxf

ig and
j
g
),
yx
for different regions to
2.2 Fuzzy C-Means
Fuzzy clustering is very well suited for the imprecise
nature of geographical information including remote
sensing data. According to the fuzzy clustering
framework, each cluster is a fuzzy set and each pixel in
the image has a membership value associated to each
cluster, ranging between 0 and 1, measuring how much
the pixel belongs to that particular cluster [13]. There have
been many different families of fuzzy clustering
algorithms proposed in the last decade. The one used in
this work is the Fuzzy C-Means algorithm (FCM), which
is an iterative technique based on the minimization of a
generalized group sum of squared error objective
functions [14], [15].
∑∑
==
ik
where the real number m is a weighting exponent on each
fuzzy membership with
1
<≤ m
clusters and n is the total number of pixels in the image
being classified.
,,(
21
vvv
?
=
prototypes.
}{
,k
U =
is a
,)();,(
2
11
−=
cn
ik
m
ikm
vxuxvUJ

.
∞
c is the total number of
),
cv
are geometric cluster
c× matrix, where the element
iu
n
of
kiuU
,
,
satisfies
] 1 , 0 [
,∈
kiu
and ∑
=
=
c
i
kiu
1
,
1for all k.

Figure 1: Two registered RADARSAT images


AugustMay




Figure 2: sub-images of the images in Figure 1.

Jis based on the suitable selection of U
and v using an iterative process using the following steps.

1. Determining values for c, M, error (e) and loop
counter t=1.
2. Creating a random
c× membership matrix U.
3. Computing cluster centers.
∑
=
=
n
mt
ik
u
1

Minimization of
m
n
∑
=
k
n
k
k
mt
ik
t
i
xu
v
)(
1
)(
)(
)(
)(

Page 3

where s represent sub-band images, acquired from
stationary wavelet transform.

4. Updating the membership matrix U.



k
vx
,
1
1
2
−
1
) 1
+
(
ik
−
=













−
−
=∑
m
c
j
i
ik
t
vx
U

Stop if
to step 3.
eUU
t
ik
t
ik
<−
+
)( ) 1(
, otherwise increase t and go

The FCM algorithm is very well suited to remote sensing
image segmentation. But at the same time it exhibits
sensitivity to the initial settings with regard to both speed
and stability and also shows sensitivity to noise. Figure 5
is the three-region classification results for the images in
Figure 2 using the FCM method.

2.3 Decision Trees
Another common approach to classification is to use
decision trees. The decision tree itself is a set of decision
rules that describe each group's patterns learned from
these given examples. The decision tree algorithm used
here is the "Quick, Unbiased, Efficient Statistical Trees"
(QUEST). The algorithm is described in [16] and the
performance of this algorithm compared with other
classification methods can be found in [17]. Applying the
QUEST to the original images in Figure 2 to discriminate
regions of land and water, Figure 6 gives the classification
results. The three-region results are plotted in Figure 8.

We must note before applying the above classification
techniques, denoising method should be applied to the
original images. In this paper, we use the wavelet
denoising method combined with simple nonlinear
speckle reduction filters (i.e. median filters). At first we
apply median filtering to the original images. Median
filtering is a widely used nonlinear process useful in
reducing impulses, or salt-and-pepper noise. It is also
useful in preserving edges in an image while reducing
random noise. The wavelet denoising method is then
applied. Wavelet transform is a useful tool for the time-
frequency analysis of signals. From the viewpoint of
signal processing, wavelet analysis represents a signal by
its components in a series of independent frequency
channels (scales). By analyzing the behavior of the signal
in each scale, we can find the features of the signal or
discriminate different parts (such as the noise and the
useful signal) of the combined signal. Mallat’s [11]
research indicated that the local maximums of the wavelet
transform of noise and signal have different variation rules
with the change of the scale. So denoising by wavelet
method can be realized by observing these local
maximums at each scale. A commonly used wavelet
denoising method proposed by Donoho [12] regards the
wavelet coefficient below a threshold as noise and set
them to zero.

Comparing the classification results using these three
different techniques, it is easy to find that the classified
images in Figure 4 using thresholding method are the
clearest. The FCM algorithm is very noise sensitive. The
images in Figure 5 present a lot of salt-and-pepper noise.
Since in this example the images are single band, the
decision tree method is very similar to the thresholding
method. By analyzing the training data, a tree is structured
with the pixel value being the only split variable for each
node. It is like using the sample data to find the threshold
and then do the thresholding classification. The
performance of the decision tree method depends on the
accurateness of the sample data and is more sensitive to
the additive noise than the thresholding technique. Among
these three methods, the FCM algorithm is the most
automatic one, which doesn’t need the training data, but at
the same time, gives the worst results.

3 Fusion Archetecture
The aim of fusing the decision result from different
classification methods is to increase the overall
performance. Hierarchically there are three levels of
fusion: data fusion, feature fusion and decision fusion
[17]. Decision level fusion was chosen over the other
forms of fusion for a variety of reasons. The most
important of these is that, unlike the others, it is always a
feasible approach, though not necessarily the most
desirable. The obvious practical reason aside, decision
level fusion does have many desirable qualities in a
change detection application. First it is the most tolerant
to individual errors in the data Secondly it has a lower
computational complexity than feature level fusion.
Thirdly, because of its low coupling of information, it is
more robust to the removal or addition of individual data
sources. The main disadvantage of this form of fusion is
that information lost from a lower level of fusion cannot
be recovered from a higher level, although some of this
can be compensated for by providing information about
the quality of the decision reached.

Each of the subsystems, D(DSi), operating on a data
source, DSi, can be viewed as a Feature In - Decision Out
configuration while the final fusion of all the subsystems
is a Decision In - Decision Out configuration. In the
general case we will have L Data source measures of an
underlying object, o, that can belong to one of c classes in
Ω = {ω1, ω2, ... ,ωc}. Then di,j(o) is the support that the
subsystem working on the i’th data source, DSi, gives to
the hypothesis that object, o, comes from class ωj. This
information can be summarized in a Fusion Matrix,
FM(o), as
Figure 3. Here the i’th row of the matrix corresponds to
the output of the subsystem operating on the i’th data
source, while the j’th column represents the support for
class ωj, from all of the subsystems [18].

given in

Page 4

There are two general approaches to using the information
present in the Fusion Matrix, FM(o). The first takes into
account that the columns of FM(o) are the support for a
class, and so a level of support can be developed using all
the subsystems. Examples include taking the average,
product, maximum, minimum, etc and then choosing the
class that has the largest support. The second approach is
to treat all of the outputs of the Fusion Matrix as an
intermediate feature space regardless of the context and
then train a new classifier on this intermediate feature
space.


d1,1(o) d1,j(o) d1,N(o)

…

…

…

Output of
subsystem
D (DSi)
di,1(o) … di,j(o) … di,N(o)


…

…

…



dL,l(o) dL,j(o)
Support
for class
j
dL,N(o)




Figure 3: Fusion Matrix, FM(o)

Using the first general approach, the decisions resulted
from thresholding, FCM and decision tree could be fused
by majority rule. Figure 7 presents the fused three-region
classification results.

4 Creation of Layers
“Layers” can be defined as images containing part of the
information of the original image. For example, for a
multi-band image, each band can be viewed as a layer.
The mean of all the bands could also be viewed as a layer.
Applying the Principle Component Analysis (PCA) to the
multi-band image, the images generated by the principle
components are also layers of the original image. Another
example of layers is
decomposition to the original image, the resulting
orthogonal components are the layers of the image. A set
of layers is “complete” if original image can be fully
generated using the set of layers.

The layers are generated based on a user’s need with each
layer containing only information of interested. Normally,
compared with the original image, each layer contains less
information, so it’s easier to perform the calculations, and
transformations based on layers. Furthermore, in some
cases, only parts of the layers are useful such as in image
fusion by PCA.

For the topographic change detection, we are interested in
the region changes at different times, so the layers we
used in this paper are based on the region classifications.
applying the orthogonal
Each layer contains only one region from the original
image. In Figure 5, each image contains three regions that
are land, shallow water and deep water. These regions
should be extracted one by one to generate the layers.
Figure 8 shows the corresponding layers of both images.
The images in red are the layers of the image taken in
May, and the layers taken in August are plotted in green.
(a) and (d) are the layer-of-land with land represented in
red/green. In (b) and (e), except the regions of shallow
water, all the others are in black. So they are the layer-of-
shallow water. Similarly, (c) and (f) are the layers-of-deep
water.

5 Topographic Change Detection
5.1 Change Detection Based on Region
Classification
Topographic change detection detects changes on the
surface of the earth. Satellite images are used to perform
topographic detection at very high accuracy. In this paper
we present a topographic change detection method that
applies the automatic update algorithm presented in [1].
Namely, for a two level classification problem we
consider an image Ι = S1 + S2, where Si, i=1,2 are
compact regions of the image represented by contours.
The contours enclose pixels that correspond to the same
region. When a change occurs, two groups of pixels are
changing region. Those that move from region S1 to
region S2 are named as “additions” (A), while the others
that change from region S2 to region S1 are called
“deletions” (D). The total change C in the image I is
expressed as the summation of additions and deletions,
C=A+D. Namely,



) 1(
21
SSSD
∩−=
(
1
)

(1)
where the subscript “i” and “i-1” are the time index,
which represent the current and previous time,
respectively. The advantage of this method is that details
of the changes are provided. In many applications, we are
not only interested in where the changes happen, but also
how they change.

In the following, we will extend this concept to multiple
regions and automatic update of information. In a
distributed processing system this mechanism may be
programmed to keep updates of changes of classification
regions or other features over time.

)(
1
) 1
−
(
1
) 1
−
(
1
)(
)(
1
S
) 1
−
(
1
S
)(
1
)(
) 1
−
(
) 1
−
(
)(
2
+
)(
1
S
)(
iii
i
iii
i
ii
i
ii
i
SA
SI
SS
=
I
∩−=
+=
−

Page 5

,
)(
k
) 1
−
(
k
)() 1
−
(
〉〈
=∩
ii
i
k
i
k
LLRR
?
,
)(
k
) 1
−
(
k
)(
k
)(
k
〉
−
〈
=
iiii
LLLA
?
.2
) 1
−
(
k
)(
k
) 1
−
(
k
)(
k
)(
k
)(
k
)(
k
〉
−+
〈
=+=
iiiiiii
LLLLDAC
?
.
)(
k
) 1
−
(
k
) 1
−
(
k
)(
k
〉
−
〈
=
iiii
LLLD
?
,
)(
k
) 1
−
(
k
)( ) 1
−
(
〉〈
=∩
ii
i
k
i
k
LLRR
?
.'
)()(
ii
CC
=
For the images with multiple level classification, we are
interested in the changes in each region, i.e. addition and
deletion. Assume we have M regions of interest, which
are presented in M “layers”,
,,,,
21
M
LLL
?
where the
region-of-interest in the layer
iL is denoted as
iR. The
pixels in
to zero. The basic idea is to compare the pairs of layers
taken at different times, one by one. Namely, find the
addition and deletion for each
the region-of-interest
discussion, and the other part having zero values is the
In this way, if we use
k
the common region of interest will be:

iRhave values “1” and all the other pixels are set
iL . For each pair of layers,
iRis exactly the
2 S in our previous
1S.
)(i
L to denote the kth layer at time i,
(2)

where the operator “?” represents the element-by-element
multiplication of two matrices, and “ 〉
region which is composed of the pixels whose values are
ones in “*”. In this way, the addition of
〈* ” represents a
k
R will be:
(3)


and the deletions is:

(4)

The total change for the k th region will be:

(5)

However, if we perform frame-to-frame subtraction, we will
obtain:

(6)

and we have:
(7)

From the above we can see in a two level classification
problem the total change may be expressed through the
absolute value of a frame-to-frame differencing. In
practice we are interested in more details of the changes
such as additions and deletions. Our proposed
formulations give these details.

Apply the above procedures to each pair of layers. Step by
step the addition and deletion for every class will be
detected sequentially.

5.2 Change Detection Based on Pixel Level
Characteristics
The ability to preserve the pixel characteristics from frame
to frame when change detection is performed is essential
if multiple classification inferences are derived from the
changes. In this case image classification process is
carried out on the change detection results. Two methods
have been studied for change detection on images with
multiple classification regions, i.e. the principal
component analysis and the wavelet method.

5.3 Matched filtering and change detection.
Change detection may be applied using matched filters.
Matched filters tend to suppress clutter and emphasize the
changes of interest. When matched filters are applied the
change detection performance increases. Matched filtering
for change detection is normally applied to multispectra
and/or multipolarized images [3], [5]-[7].

6 Example
Let’s consider the two images in Figure 2. Their layers are
presented in Figure 8. We apply the proposed multi-level
change detection method to the pair of layers {(a), (d)},
{(b), (e)} and {(c), (f)}, respectively. The result is
displayed in Figure 9, where the red regions represent
deletions, green ones stand for additions, and the yellow
means no change happens. Figure 9 clearly gives the
details of change in each region. It is easy to find from
Figure 9 (c), because of the flooding in May, some regions
of shallow water and land in the image of August become
deep water (the red region in (c)). For the same reason, in
(a), the green regions are the parts that are changed from
shallow and deep water in May to land in August. Using
this method avoids need for strict radiometric calibration,
and it designates the types of changes occurring for each
region of interest. It is simple, reliable and effective.


May
Figure 4: Three level region classification results using
thresholding method.

August

May August