€ 
au = © 
— UC 
N 
e 
IS 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
along two mentioned directions, these lengths are calculated as 
follows. 
Diff, -|postL, — preL,| (2) 
DiffL, = |postL, = preL,| (3) 
where postL,, postL, = lengths of shadows from post-event 
image 
preL;, preL, = lengths of shadows from pre-event 
image 
| | = function to take absolute value 
Li: length of shadow along 
dominant direction 
Lz: length of shadow along 
the perpendicular direction 
to the dominant direction L2 
AP 
    
post-event shadow 
pre-event shadow 
Figure 4. Illustration of length computation along dominant 
directions 
Step5: Thresholding the differences in length. When one 
among two values of length difference is bigger than a 
threshold, it is probably due to the damage. 
Step 6: Merge all extracted shadows from all portions with 
assigned label whether it is damaged or non-damaged. 
4. TEST RESULT 
The proposed methodology was employed for processing two 
pan-sharpened Quickbird scenes acquired over Boumerdes city, 
Algeria. Boumerdes was one of the most heavily-damaged 
areas due to the earthquake of magnitude 6.8 on May 21, 2003. 
The first scene was about one year before the event (April 22, 
2002) and the second one was two days after the event (May 
23, 2003). These scenes were in Standard format, which were 
terrain corrected. It means the ground features were correctly 
mapped, but the scene has not been ortho-rectified. When 
comparing two scenes, the roofs of the same building were not 
in the same location but their shadows were overlapped. 
Furthermore, there is a slight difference between these scenes 
due to difference acquisition condition. Prior to employing the 
proposed extraction method, these two scenes were co- 
registered and extracted into the same region of interest. 
Shadows were successfully extracted by K-mean unsupervised 
classification. The dimension of the test area was 3800 pixels x 
2900 pixels (approximately 2200 m x 1700 m for pan- 
sharpened scene) (Figure 5). 
Detected shadows are shown in Figure 6. However, visual 
checking both scenes, there was just about 5096 of shadows 
really caused by buildings; others were by trees or even clouds 
(Figure 7). Further studies will consider the clearly 
discriminating between those kinds of shadow. If trees and 
buildings could be successfully classified, their shadows would 
be well discriminated. 
609 
  
Figure 5. Pan-sharpening Quickbird scene acquired on May 21, 
2003 in true colour composite of the study area 
uA 
= 4 x 
f t = , Eze? 
- > - -—77 
f a nan : 
rol — - = 
- - 
r - —" - — 
- "zu v 
7 m € 
/ re tt 
* e + 
7 + 
et - ' 
À - 
- y 
-l - - - 
= - T ; e — 
T t: Te LÀ > / 
M er : _ # f 
€ m -— 
_ e: € * 
t n — - x - ve ; x 
KT m i » 
- — 
= un 4 
. e J 
-— & - — t -— 
ot -r ptt 
Figure 6. Detected shadows (post-event scene) 
Detected results 
M MM. 
Number of shadow areas 
on 
© 
  
   
n 
© 
i 
  
Building Tree Cloud 
Class 
Figure 7. Distribution of detected shadows 
Examining the scatterogram between two length differences of 
extracted shadows (Figure 8), the threshold of 10 m was chosen 
to classify damaged and non-damaged buildings. This result 
was compared to visual interpretation of Quickbird carried out 
by Kouchi et al (2004). As illustrated in Figure 9, there is 
highest probability to detect heavily damaged buildings by 
using only shadow analysis. Only 10 % - 20 % of slightly or 
moderate damaged buildings could be detected. A shorter 
threshold could increase the percentage of successful detection