During the classification phase, a region-based co-occurrence 
matrix is computed on each region of the image respectively. 
Figure 8 shows the results of the neural network classification. 
They are more realistic and noiseless compared with a conven- 
tional Bayesian method (Fig. 9). Table 10 and 11 show the 
classification accuracy achieved by a maximum-likelihood 
classification and the neural network approach. The results 
indicate that a neural network approach based on region-based 
co-occurrence matrices can outperform a conventional maxi- 
mum-likelihood method, especially when land-use maps instead 
of land-cover maps are generated. 
4. CONCLUSION 
The classification tests show that region-based co-occurrence 
matrices combined with an ATL network have potential for 
discriminating several intra-urban land-use classes with high 
accuracy. The proposed method produces more realistic and 
noiseless land-use classes compared with a conventional Baye- 
sian classifier. The neural network approach exceeds the overall 
classification accuracy achieved by the maximum-likelihood 
method by 14%. Because of its fast convergence during training 
and its ability to approximate arbitrarily complicated decision 
regions, the ATL algorithm used in this study is an appropriate 
alternative to backpropagation. 
0 ] 2 km 
  
  
N 
LT DEED 
0 l 2 km 
Legend 
Industrial Area 
[ 1 Docks 
EUN Residential Area 
RE Other Areas / 
Unclassified 
  
  
  
Figure 8: Result of ATL classification based on modified co-occurrence matrices 
74 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996 
m m — M 
EC...) pet