“Before”:09/30/2003 
2.4m multispectral 
1094X886 pixels 
SAR and optical images fusion 
based on wavelet transform 
Registration to “ before ” using 60 
ground control points with first 
order polynomial transform 
“After”:03/01/2004 
2.4m multispectral 
1166X843 pixels 
Table 1. High-resolution QuickBird imagery of Bam 
to t> 
CO to 
CM 
CO 
o 
CM CM 
iH 05 
CO 05 
Track=120 S'" 
CM 
to © C- 
CM M © 
• CO • ✓—n 
cm 3 h co 
H O' OCR 
eo -p 
8 S S' 
cm « cm 
g 
•H O 
CM 
Dt=175/Bp=495 
Dt=35/Bp=581 
Track=385 
trr or 
CM CM^ 
5 8 
1-H to 
1-1 IO 
. 05 
CO 00 
1-1 00 CM 
OlO © 
1 
<5- S'" S' 
CM CM CM 
Figure 1. ENVISAT ASAR data in Bam 
3. METHODOLOGY 
Object-based change detection method avoids the disadvantage 
of pixel-based approaches and is more suitable for high 
resolution satellite images.The technique workflow as 
follows(Figure 2). 
First of all, images from different and same sensors are 
registered to each other. SAR and optical images fusion based 
on wavelet transform is executed on CASMImageinfo3.0.And 
then multiresolution segmentation divides the processed images 
into homogeneous regions ,r'espectively.Make sure that the 
number of before event image regions is more than that of after 
event image regions in order for the characteristic consistency 
of regions.After that,region features which represent elements 
accurately and exclusively are extracted,including(l)spectral 
feature:mean,(2)shape featurexentral geographic coordinate and 
mininmal ex-rectangle of regions,(3)texture feature:gray-level 
co-occurance matrix. In order to eliminate the influence of 
shadow on final results,those are labeled by 
human-assissted.According to the segmented before-event 
image regions(exclusive of shadow regions) centriods,we use 
ARs searching method proposed in this paper to look up ARs on 
after-event ones.Repetition is executed for those which haven’t 
been corresponded on after-event images.Thus all regions 
between before-event and after-event images have 
one-against-one relationship.Afterwards the feature space 
distance is calculated between the ARs. Finally, threshold 
defined by users determines whether regions change or not. 
4. MAIN TECHNIQUES 
4.1 Multiresolution Segmentation 
Segmentation is an important method for acquiring 
objects.Multiresolution segmentation is a bottom-up 
region-merging technique starting with one-pixel objects.In 
numerous subsequent steps,image objects are merged into 
bigger ones.Throughout this pairwise clustering process,the 
heterogeneity criterion is considered,defined by scale 
parameter.If the smallest growth exceeds the scale parameter,the 
merging process stops.This heterogeneity criterion is a weighted 
by the mean between spectral heterogeneity (hc 0lor )and spatial 
heterogeneity(h shape ). The definition as follows: 
h = W * Kotor + 0 — w) • Khope (1) 
where w=color weight(against shapejwith 0 < W < 1 
4.2 ARs Searching Based On Minimum Ex-rectangle 
ARs searching is the premises of region feature comparison.If 
not,many false alarms will generate so as to reduce the detection 
accuracy. 
The arithmetic workflow(Figure 3) is delineated as follows:We 
establish correspondence relationship for all regions between 
before and after event images according to region centroids. 
There are three kinds of result:one against none(l-a-0),one 
against one(l-a-l) and one against rest(l-a-r).As for the case of 
l-a-r,the offsets between correspondence regions centroids are 
calculated.In the images pre-processing step,we have the 
pro-knowledge of the accuracy of image registration between 
before and after event images.Therefore,we can estimate the 
association confidence and establish the correspondence 
regions. 
4.3 Region Features Similarity Comparison Measured By 
Feature Space Distance 
The similarity of feature space between ARs is a key to 
determine whether change or not.The similarity measurement is 
not only consistent with human vision,easy to compute but also 
robust for the accuracy of segmentation techniques.The distance 
of feature space between ARs is a weighted mean between 
spectral(brightness) and texture(gray-level co-occurence matrix) 
factors. The distance is compared with threshold defined by 
users,which is larger considered as changed,no change 
contrarily. 
4.4 Accuracy Assessment Using Simple Confusion Matrix 
Accuracy assessment is a way to estimate the effectiveness of 
change detection methods.There are two aspects: (1) attribute 
accuracy,detecting change/no change and change type,(2)