International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
    
Interne 
  
C-Y eg P) Az» {P,(k)} (12) 
du 
E--S P,)-glr (). M- Seo) 
k=0 k=0 
where &= Shift (distance) from the target point 
k = Brightness difference compared to target point 
P,(k) = Occurrence probability of the pixel where 
k is brightness difference and & shift 
These indexes are combined linearly for the evaluation function 
and the coefficients for the indexes are determined by 
supervised learning in other experiments. 
  
Z-direction polygon matching 
* 
Change detection by 
z difference 
ur 
Change detection by 
texture comparison 
  
  
  
  
  
  
  
Figure 7. Process flow of 3D image matching method 
     
     
Stereo model 
  
Raster image matching 
in 2D space 
e 
: Roofs plane in 
: i assuming altitude 
AH ; 
2 1° <+— Target building 
Figure 8. Process flow of indirect comparison method 
5. EXPERIMENTS 
Evaluation tests were performed with actual aerial photos that 
have been taken right after of the earthquake (1995) and 8 years 
later (2003). The scale of photo is 1/4000 and the image size is 
20000 by 20000 respectively. 
Fig.10 shows an imaginary stereo model by perspective 
projection with automatically detected matching points. Fig.11 
shows the result of adaptive nonlinear mapping in 2D image 
matching method in the case of mapping from Fig.10 (b) to 
Fig.10 (a) and the result of change detection. It was ascertained 
that correct deformation was achieved if increase the iteration. 
The photo interpretation result by human operator for 
evaluating change detection ability is shown in Fig.12. Fig.13 
is a change detection result by 3D image matching method. For 
quantitative estimation ROC chart was applied, which plots the 
sequential probability of detection against the probability of 
false alarm. Fig.14 (a) shows the chart in which x axis is the 0.5-1 
change of building’s height and y axis is the change detection 1.0 for 
ratio or false alarm in the case of 3D image matching method. Anoth 
Similarly, Fig.14 (b) shows ROC chart. As a result, 80 % of photos 
right change detection has been achieved when false alarms are can be 
about 30 % in 2D and 18 % in 3D image matching method matchi 
respectively. compa 
passag 
clouds 
    
(a) Right after the earthquake (b) 8 years later the earthquake 
Figure 9. Aerial images for experiment 
  
  
(a) Perspective projection of 
Fig.9 (a) 
Figure 10. Imaginary stereo model (greyscale image) 
  
1 
0.8 
0.6 
PF/PD 
  
  
   
(a) Mapping process (b) Mapping process Figi 
(iteration = 1) (iteration = 5) 
N 
0.8} 
| 
C oo 
A ' 
e 1 
| 04H 
| \ 
02h" 
| 8 
(c) Mapping process (d) Change detection result | Fig 
(iteration = 20) (binary image) 
Figure 11. Change detection by 2D image matching method | 
As regards texture analysis, change detection showed the best | 
result when coefficients of the indexes are 3.0 - 9.0 for contrast, 
434