/ area 
  
  
Yosuke Ito 
  
Input layer 
  
Coherence set 
Figure 7: Structure of competitive neural network 
The LVQ2.1 behaves effectively when the PDFs are overlapped each other. In both procedures, the initial learning rate 
is set to 0.03 and the weight vectors are repeatedly updated 40000 times. The extraction method exposes the coherence 
images with unknown category to the trained NN and classifies it as the category to which the winner neuron belongs. 
To find out the effectiveness of the multi-source coherence images for extracting the damaged regions, we defined two 
types of coherence sets: C, = {EC1-3, EC1-4, EC2-3, EC2-4}; C, = Cy + {JC1-2}. The number of coherence images 
corresponds to the number of input neurons N. 
In the experiments, the training and test data (table 3) were chosen from the area covered by either the burned or the 
completely collapsed structures based on the hazard surveying map in figure 4. The training data of w; and v» are 5 x 5 
and 10 x 10 resampled scenes from the test data, respectively. The LVQ method was compared with the maximum 
likelihood (ML) method. All classification methods employed the same training and test data for a fair comparison. 
5 RESULTS AND DISCUSSION 
Extraction results of damaged regions were assessed with regard to an kappa coefficient (Richards, 1993) and distribution 
of the extracted regions in the classified image. Confusion matrix A — [a4], $4, = 1,..., L is produced for each 
classification result where a;; denotes the number of pixels classified w; into w j. The kappa coefficient (x) is defined by 
  
L L 
A++ D Okk — 3, Gk Ok 
iran k=1 k1 
K = ; 7 (4) 
ay — = AR+O+k 
LI L L 
where ayy = N° N^ aij, Qi = > jj and aij = S aij. 
113-1 jg i=1 
Figure 8 shows the comparisons between C, and C» for the LVQ and ML methods in term of & where the number of 
neurons M in the competitive layer varies from 2 to 20 to find out an optimum NN. For C, and C» when Ms are equal 
to 4 and 6, the NNs produce the maximum accuracy. The changes of & with the number of neurons M are small for Ci 
since the distributions of w; and w» in the coherence images of C, are similar as shown in figure 6. However, & for C» is 
sensitive to the number of neurons M since the distributions of JC1-2 and the other ERS coherence images are different. 
The & of the LVQ using the optimum NN is higher than that of the ML for both coherence sets. 
To assess the classification accuracy for all combinations about the classifiers and the coherence sets, the kappa coefficient 
(k) is shown in table 4. The x is improved 23% (0.035) in the LVQ and 34% (0.039) in the ML methods by adding JC1-2. 
Itis also improved 3196 (0.035) in C, and 2096 (0.031) in C; by applying the LVQ method. Table 5 shows the confusion 
matrices using C5. The LVQ method produces better results than the ML method from the view point of balance of the 
number of the correctly classified pixels, that is, a4; of the LVQ decreases 418 and a»» increases 2198 compared with the 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 161