Is in band 5 
'asn't homogenous. If I 
network so, I get much 
g three neuron layers. 
and 4 neurons (SSE = 
oesn’t cause increased 
6, 4 neurons) or in the 
ere trained by the mean 
11 happen if l’Il use the 
ctors! 
7 pixels. I’ve selected 
ion for two reasons: 
vith the pixels, but the 
ant. 
, about which I know 
they haven’t taken part 
use these pixels for 
N. 
SSE was 0.0001. This 
is interesting how does 
twork with the same 
ing in model ne42 (12, 
43 (24, 4 neurons) (— 
wed network error was 
ork 
1 radial basis transfer 
from the customary 
on purpose to get exact 
orithm defines also the 
al basis network had 2 
vith a transfer function 
> were 4 linear neurons. 
accelerated: while a 
ochs (nearly 21 million 
aining material, a radial 
38000 flops) (Demuth, 
1a 1996 
  
Forest 
3. COMPARATIVE ANALYSIS 
  
mini 40.98 4.17 13.32 | 41.53 
At first in comparisons I studied how different were the maxi 26.87 2.74 6.80 | 63.60 
testpixels classified by the different methods. Testpixels are the ne2 ' 55.24 10.59 12.88 | 21.29 
pixels of the training areas. (Important to know the models till 
ne3 3 have been trained with the means and covariances ne2 2| 27.49 8.3] 7.01| 62.19 
calculated from these pixels; afterwards the training material nes 60.87 1.65 6.26 | 25.22 
was every tenth pixels of them!) 
3.2 | 29.88 2.82 38.19 | 29.11 
In tableform the methods and their accuracy is the following ne^. 
(Table 1): ne3 3| 62.99 3.71 4.58 | 28.72 
ne4 39.29 4.56 14.55 | 41.59 
  
  
  
  
  
  
  
  
  
Model Wrongly Classification ne42 35.05 3.47 11.47 50.01 
chssified | accuracyn es ne43 | 37.12| 538|  1L63| 45.87 
pixels 
: ne5 37.51 3.21 6.81 | 52.47 
mini 61 2.21 
maxi 8 0.29 
Table 2. Classification results for the whole image 
ne2 244 8.85 
ne2 2* 41 2.56 The described experiment was an analysis of a single research 
= ? area. In the future I would like to test the methods on other 
nel * 95 5.93 areas, too. l’d like to insert further information into the neural 
networks, so to get more accurate and efficient thematic 
ne3 2* 221 13.79 classification. I want to expand my study on multitemporal 
ne3 3* 115 717 images at the end. 
ne4 26 0.95 
ne42 11 0.40 REFERENCES: 
ne43 11 0.40 
nes 9 0.33 Barsi, À. 1994 Thematic Mapping of the Naivasha Region 
  
  
  
  
(Kenya) from LANDSAT Images (in Hungarian) 
Thesis work, Budapest 
  
Table 1. Accuracy of the methods (The * signals that the Barsi, À. 1995 Thematic Classification of Satellite Images by 
testfield had 1603 pixels, in all other models 2757 pixels.) Neural Networks (in Hungarian) 
Essay, Budapest 
Colwell, R. M. 1983 Manual of Remote Sensing 
Sheridan Press, Fall Church 
  
Classification accuracy in differentmethods 
Demuth, H. 1995 Neural Network Toolbox User's Guide 
  
" | 1 Mathworks Inc., Natick 
10 | | Rojas, R. 1993 Theorie der neuronalen Netze 
M 8| | Eine systematische Einführung 
| Bs Springer-Verlag, Berlin 
| 4 1 
I. d 
| 0 
  
  
  
  
  
mini ne2 ne3 ne3_3 ne42 ne5 
Methods | 
  
Figure 8. Barchart of the accuracy of the methods 
Taken the whole image (301 x 460 pixels) in sight, it’s different 
how the methods the pixels classified (Table 2.). 
51 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996