ries 
d in 
the 
een 
vork 
0 6 
data 
ired 
data 
tion 
ded 
tely 
; nO 
4-1, 
for 
very 
: all 
ion. 
y of 
735. 
ory 
very 
After several runs the first network had an accuracy in the 
training set 2.2 % errors. The structure of this network is 
greater: 10-14-1. The desired error goal was 0.05. There was no 
need for many epochs, the 6 iteration has stopped. The final 
run was successful: 0 % training error at 0.01 desired network 
error, network structure 12-16-1. The modification of neural 
parameters (learning) was repeated 13 times. The test set 
accuracy was 0.6 %. The resulting whole map is presented in 
Figure 11. 
  
Figure 11. Thematic map with 4-neighborhood 
The result map seems like it would have been filtered by edge 
detection filters. The classifier hadn’t reject any pixel. 
Comparing the result to the previous maps, again the 
Urban/Meadow ratio is differing. This time the Urban area was 
34.1 % and Meadow2 26.5 %. 4-neighborhood ordered more 
pixels also into the class Meadow 1 (7.5 %). 
The management of the 8-neigborhood necessitated the most 
patience. The training was terrible slow because of the 63- 
dimensional intensity vector. The phases of the design with the 
important parameter settings are following: 
  
  
  
  
Training | Number | Training | Network | Error 
errors of time [s] | structure | goal 
epochs 
35% 10 ~ 2 hours | 15-20-1 0.05 
0.6 % 16 ~ § hours | 17-22-1 0.01 
0% 11 ~ 5 hours | 17-22-1 0.01 
  
  
  
  
  
  
  
Table 3. Designing the 8-neighborhood network 
The training error is reduced very slowly. There was no need for 
designing too complex network, the final structure is 17-22-1. 
The desired error goal is moderate. The most extreme values are 
the training times. For the management of 8-neighborhood, the 
LM memory reduction was “life-saving”. 
The resulting thematic map is similar to the 4-neighborhood 
classification, but the “edge detecting” effect is even more 
stronger. The statistics about class distribution has proved the 
filtering effect (Table 4). 
It’s worth to compare Table 4 to Table 1. The Urban class is 
almost the half of the original mapping, water is also less, but 
meadow area has grown strongly (3 and 2.5 times more). 
The classifier didn’t reject any pixel. 
Barsi, Arpad 
  
  
  
  
  
  
  
  
Class Amount % 
F1 19577 18.2 
F2 11054 10.3 
M1 18786 17.5 
M2 24531 22.8 
U 27589 25.6 
W 6099 8.7 
  
  
  
  
Table 4. The result of the 8-neighborhood classification 
Figure 12 illustrates the result considering all neighboring 
pixels. 
  
Figure 12. Thematic map with all (8) neighbors 
There're more disturbing wrong pixels in the Water class (river 
Danube). 
3.4. Handling PCA and neighborhood with neural networks 
As it was mentioned in the methods' chapter, the training set for 
the neural network was prepared from the previously PCA- 
transformed data extending with the neighborhood pixels. The 
dimension of the input intensity vector was 15 
(4-neighborhood) and 27 (8-neighborhood). 
Although the data dimensionality is reduced the training were 
longer, about 700-900 s. The first successful realization of 
neural network had a structure 12-16-1. The structure was kept, 
but newer trainings were started. The final solution is found in 
13 epochs, the desired error goal was just 0.01. The accuracy in 
the training set was 0.1 %, in the test set 0.3 %. No rejected 
pixels are found. Figure 13 illustrates the neural network’s 
output. 
The resulting thematic map of neural network trained with PCA 
transformed bands and 8-neighborhood is very close to the 
antecedent map (Figure 14). After expectations the network’s 
training was slow: it took about 2 hours. The desired error goal 
was 0.01. The successful trained network had 20-24-1 structure. 
The training accuracy was 0.1 %, the test’s one 0.3 %. There 
were 36 pixels rejected. The produced thematic map has the 
same high quality as the previously with slightly “edge 
detecting effect". 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 145