100d 
| the 
er to 
hose 
Irya 
tion 
the 
ility 
6% 
ples 
els, 
] as 
  
time was 60 seconds, while the classification time was 320 
seconds. 
A standard back-propagation neural network used in our 
researchs was composed of an input layer, one hidden layer and 
an output layer. The input layer consisted of six processing 
elements, one for each of the Thematic Mapper nonthermal 
channels. The output layer consisted of fifteen processing 
elements, corresponding to fifteen land-cover classes. The input 
image pixel data was linearly scaled to a value between 0.0 and 
1.0 for input to the neural network.. At the moment when the 
neural network was first created; the weights for the interunit 
connections were initialized to random values in the range of -0.5 
to +0.5. An output of 0.9 (the processing element with the 
highest activation) represented the correct class, and an output of 
0.1 represented all other classes. In all the experiments the initial 
learning rate and momentum were 0.1 and 0.9, respectively. The 
learning and momentum rates affects how quickly the neural 
network stabilizes. High learning rate (0.9) would converge 
quickly, but may exit prematurely. On the contrary, low learning 
rate (0.1) would take more iterations to train, without a danger of 
prematurely exit. Usage of a momentum term helps in reducing 
oscillation between iterations, and allows a higher learning rate 
to be specified without the risk of non-convergence In our 
experiments we tested networks with from 2 to 31 processing 
elements in a hidden layer. Percentages of correctly classified 
pixels for all these networks are shown in the Table 5. The 
values are obtained after 500 iterations. The obtained results are 
very similar to the networks with a number of processing 
elements between 7 and 31 for the training set of data. The same 
holds for the networks with a number of processing elements in 
a hidden layer between 11 and 31 for the testing data This 
indicates an increase of neural network classifier performance for 
large number of processing elements in a hidden layer. 
We used 19 processing elements in the hidden layer. The 
leaming rate and momentum rate were both 0.01. The 
normalized total error (Figure 3.) shows that the neural network 
converges and stabilizes by the 10000th training iteration. The 
average accuracy and the overall accuracy of the train pixels for 
the back-propagation algorithm were 98.29% and 98.24% 
respectively. The average accuracy and the overall accuracy of 
the test pixels for the back-propagation algorithm were 97.04% 
and 96.93% respectively. Comparing these results to those 
obtained by the network of the same architecture with a 500 
iterations, we can not see an important increase of network 
performance. As a consequence there is a significant reduction 
of the training time for the network with a less number of 
Table 4. Confusion matrix of the test side results for maximum likelihood classifier. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Areas Percent Pixels Classified by Code 
Code Name Pixels 0 10 20 30 40 50 60 76 
10 New Forest 230 17 94.3 2.6 0.0 0.0 0.0 0.0 0.0 
20 Forest 235 17 17 94.5 0.0 0.0 0.0 0.0 0.0 
30 Wet siol 293 3.1 0.0 0.0 92.5 0.0 0.0 0.0 0.0 
40 Arid soil 344 4.9 0.0 0.0 0.0 86.9 0.0 0.0 0.0 
50 Maize-I 243 3,8 0.0 0.0 0.0 0.0 93.8 0.0 0.0 
60 Maize-II 241 8.7 0.0 0.0 0.0 0.0 0.0 91.3 0.0 
70 Dry vegetat. 156 6.4 0.0 0.0 0.0 0.0 0.0 0.0 93.6 
80 Stubble-field 227 $3 0.0 0.0 0.0 0.0 0.0 00 [ 00 
90 Indiv. parcels 255 2.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
100 Barley I 297 9.4 0.0 0.0 0.0 2.4 0.0 0.0 0.0 
110 Barley I 275 73 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
120 Oil rape I 213 1.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
130 Oil rape I 210 3.8 0.5 2.9 0.0 0.0 0.0 0.0 0.0 
140 Wet vegetat. 260 9.6 0.0 0.0 0.0 0.0 0.8 0.0 0.0 
150 Sunflower 164 5.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
Code Name Pixels 80 90 100 110 120 130 140 150 
10 New Forest 230 0.0 0.0 0.0 0.0 0.0 13 0.0 0.0 
20 Forest 235 0.0 0.0 0.0 0.0 0.0 2.1 0.0 0.0 
30 Wet siol 293 0.0 0.0 0.3 0.0 4.1 0.0 0.0 0.0 
40 Arid soil 344 0.0 0.0 7.8 0.0 0.3 0.0 0.0 0.0 
50 Maize-I 243 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.0 
60 Maize-II 241 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
70 Dry vegetat. 156 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
80 Stubble-field 227 92.5 0.0 1.3 0.0 0.0 0.9 0.0 0.0 
90 Indiv. parcels 255 0.0 93.7 0.0 0.4 3.9 0.0 0.0 0.0 
100 Barley I 297 0.0 0.0 88.2 0.0 0.0 0.0 0.0 0.0 
110 Barley II 275 0.0 0.4 0.0 92.4 0.0 0.0 0.0 0.0 
120 Oil rape I 213 0.0 3.3 0.0 0.0 95.3 0.0 0.0 0.0 
130 Oil rape II 210 1.4 0.0 0.0 0.0 0.0 91.4 0.0 0.0 
140 Wet vegetat. 260 0.0 0.0 0.0 0.0 0.0 0.0 89.2 0.4 
150 Sunflower 164 0.0 0.0 0.0 0.0 0.0 0.0 0.0 94.5 
  
  
  
  
  
  
  
  
  
  
  
  
  
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
379