ISPRS, Vol.34, Part 2W2, "Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001 
279 
oyu Road, Wuhan, 
t regional or global 
problem in remote 
/arious geo-spatial 
is the central part 
ture-based image 
change accuracy 
9 detection, which 
s separately and 
-by-pixel basis to 
Direct Multi-date 
mages of the two 
ation between two 
the whole change 
lassification. The 
tntages over the 
itly recognize the 
:urred .robustness 
Jitions at the two 
poral and/or multi- 
include greater 
evere difficulty in 
/ and difficulties 
' assessment on 
s generation of an 
usually a difficult 
Consequently ,the 
ction methods is 
a ground truth is 
upervised change 
ice image”. These 
ge ratio, change 
nalysis(PCA) and 
k .morphological 
mathematics. They process the two multi-spectral images 
acquired at two different dates in order to generate the further 
image. The computed difference image is such that the values 
of the pixels associated with land cover changes present 
values significantly different from those of the pixels 
associated unchanged areas. The changed areas are then 
identified by analyzing the difference image. For example, the 
image difference technique generates the difference image by 
subtracting pixel by pixel, a single spectral band of the two 
multi-spectral images under analysis. The choice of the 
spectral band depends on the specific type of change to be 
detected. An analogous concept is applied by the widely used 
change vector analysis (CVA) technique. In this case, each 
pairs of corresponding pixels is represented by two vectors in 
feature space called change vectors .The change vector takes 
the difference between the feature vectors at the two times. 
The magnitude of the change vector represents the degree of 
change, while the direction of the change vector indicates the 
types of change with the help of supervision on the change 
types. In spite of their relative simplicity and widespread use, 
the aforementioned change detection techniques exhibit a 
major drawback : a lack of automatic and non-heuristic 
techniques for the analysis of the difference image. In fact, in 
classical techniques, such an analysis is performed by 
thresholding the difference image according to the empirical 
strategies or manual trial-and-error procedures, which 
significantly affect the reliability and accuracy of the final 
change detection map. Although many analysts proposed a lot 
of automatic threshold selection methods 
[Rosin,1999;Bruzzone 1999],they are only suitable for some 
specific situations not for common use. 
2.Problems & Requirements 
There are eleven major problems associated with the 
current change detection techniques. 
i) Lack of theoretical basis for change detection is the 
key problem. Many change detection techniques can detect 
some change information in some specific situations, but when 
the situations changed the results changed. In fact, because of 
complexity of image problems, it is difficult to illustrate one kind 
of universal truths. 
ii) This is the further step for the first problem. Even if we 
have no universal theory for change detection, it is practical 
that we have some criterion for selecting different change 
detection techniques according to different situations. But this 
point is still not achieved. 
iii) Most change detection techniques are based on pixel 
level. But general speaking, a mere thresholding of the 
difference signal obtained from two corresponding pixels was 
insufficient to distinguish between changes of interest. So 
some feature-based algorithms should be developed for 
improving the reliability and accuracy of detection. 
iv) We have too little information about spectral 
characteristics of ground objects. This affects our 
understanding to images. Of course this task is very time- 
consuming and expensive. 
v) Often we have no good methods for processing the 
bad effects on image such as uncertain atmosphere 
conditions, sensor noises, radiometric differences and so on. 
These factors causes the low accuracy of change detection. 
vi) Considering how humans detect changes from 
images, it is obvious that very limited information and/or 
knowiedge (about sensors, images, spatial relations and so 
on) is utilized in current change detection techniques. 
vii) Finding change (i.e. the amount of change detected) 
is one of the most important objectives in change detection 
applications, but most of the current change detection 
techniques need a user-specified threshold which is often set 
empirically and subjectively since there is theoretical guidance 
to this problem. 
viii) In most change detection techniques, the 
dependency information between the two images is ignored. 
ix) Only very limited or no information at all about the 
direction and characteristics of actual changes occurred on the 
ground can be induced using most current change detection 
techniques [Xiaolong Dai, 1998] 
x) One practical problem with difference image is that the 
images are not in perfect spatial registration before analyzing 
so the difference image will contain artifacts caused by 
incomplete cancellation of the unchanged background objects. 
This registration noise causes problems for most change 
detection algorithms [Xiaolong Dai, 1998]. 
xi) Most techniques are not fully automated and some 
are even non-quantitative. 
In summary, many interactive change detection 
techniques are in practice today. However, the majority of the 
techniques themselves can only provide the binary change 
mask and classification procedure must be applied to the 
individual multi-temporal images in order to obtain the 
categorical information of multi-date land covers. Besides,