the human brain. The 
periment I made the 
stance and maximum 
ctors. It’s possible to 
cy can be increased. 
  
ction 
(1) 
(2) 
tion and other curves 
' they are symmetric 
e.g. between 0 and 1, 
development of new 
ch is similar to the 
radial basis transfer 
Font EN 
a Tr 
  
Figure 3. The radial basis transfer function 
Equation of the function: 
flex)=e=" G) 
The new transfer function and method caused revolutionary 
changes in the use of neural networks (see 2.3.) (Demuth, 
1995). 
Let's see how the network operates. The computations of a 
neural network can be solved as vector- and matrix operations. 
If we have S1 neurons in the first layer, S2 in the second one, 
and have R inputs, then the output is the following (Barsi, 
1995) 
O «logsig( W,logsig( WI-B,j]- B.) | (9 
where 
O the output vector (S2 x 1) 
I the input vector (R x 1) 
W, the first weight matrix (S1 x R) 
B, the first bias vector (S1 x 1) 
W, the second weight matrix (S2 x S1) 
B, the second bias vector (S2 x 1) 
logsig logistic sigmoid as transfer function. 
For the previously described operation the network parameters 
(WI, B1, W2, B2) have to be defined. The training vectors 
contain the inputs and that, what the network should answer. 
The differences between the computed and the given answers 
are the errors, which are to be divided on the neurons, and the 
weights and biases are to be modified in order to decrease these 
errors. This phase is the backpropagation. The geometric 
meaning of the method is finding the minimum of the error 
hypersurface. This optimization could be computed in many 
steps, but today using the Levenberg-Marquardt method it takes 
only few epochs: 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Sum-Squared Metwork Error for 12 Epochs 
Figure 4. Network error during the training with Levenberg- 
Marquardt method 
At the end of the training we get the weights and biases which 
we can see in Hinton-diagram: 
  
Figure 5. Hinton-diagram of the weights and biases of a layer 
2. THEMATIC CLASSIFICATION 
2.1. Traditional methods 
Thematic map is a map on which we show the different 
thematics'. Thematics’ are water, forest, meadow, street etc. At 
landuse classification we often use thematic maps. 
One of the most efficient way of thematic mapmaking is the 
process of satellite images. 
I used in my experiment six bands of a LANDSAT TM image 
taken in 1989 (Figure 9.). 
I differentiated four thematics' in my research area: 
e forest 
e meadow 
© water 
e town.