Yosuke Ito 
  
  
0.20 
0.18[ 
0.16[ 
  
0.14 [ 
0.127 
  
  
  
  
0.10 
0 M 
Number of neurons in the input layer 
Figure 8: Comparison with kappa coefficient (x) using LVQ and ML methods 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Table 3: Training and test data Table 4: Assessment of kappa coefficient 
| Category | Training data | Test data | | Coherence set | Classifier ET 
wi 131 2994 C1 LVQ (M = 4) | 0.149 
wo 150 16221 ML 0.114 
Total 281 19215 C» LVQ (M = 6) | 0.184 
(pixels) ML 0.153 
  
  
  
  
  
ML. As a result of using the ML, the test data tend to be classified into w;. These results mean that the non-parametric 
approach is more significant when the PDFs are unknown. 
Finally, we show the classification images using C» in figures 9 (a) and (b) which are generated by the LVQ with M = 6 
and the ML methods, respectively. As the extracted pixels using the ML are too many (table 5 (b)), the extraction regions 
of the LVQ result in figure 9 (a) is more similar to the hazard map (figure 4) than those of the ML result in figure 9 (b). 
It is experimentally shown that the classification results can be improved using the multi-source coherence image and the 
neural classifier. 
6 CONCLUSIONS 
We have produced and assessed classified images for extraction of damaged regions by using multi-source and temporal 
coherence images and classification methods. It is suggested that extraction accuracies can be improved by using the 
multi-source coherence images and the neural classifier on an experimental basis. In comparison with the parametric 
method, the LVQ produces higher classification accuracy in term of the kappa coefficient. For future study, we will 
consider an extraction method which can classify the coherence images into more detailed categories according to degree 
of collapse. To enhance the detection accuracy, it is desirable that the damaged regions are roughly specified based on 
coseismic crustal deformation using differential interferogram. 
Table 5: Confusion matrices using C» 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
(a) LVQ with M — 6 (b) ML 
Result Result 
[aij] Q1 | wo | Total [a;;] wh | wo | Total | 
Test | Gi 1978 1016 | 2994 Test W1 2396 598 2994 
data | wo 5799 | 10422 | 16221 data | wo 7997 | 8224 | 16221 | 
Total | 7777 | 11438 | 19215 Total || 10393 | 8822 | 19215 
(pixels) (pixels) 
  
162 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 
ACK 
The. 
woul 
feror 
Rese 
REF 
Fujis 
inter 
Ito, ' 
Rem 
Koh 
Lee, 
IEEE 
pp. 4 
Rich 
Yone 
tiona 
Zeb 
Rem