THEMATIC CLASSIFICATION OF A LANDSAT IMAGE USING NEURAL NETWORKS 
Árpád BARSI 
Technical University of Budapest 
Dept. of Photogrammetry 
1521 Budapest 
abarsi@epito.bme.hu 
Comission III, Working Group 3 
KEY WORDS: Landsat, Thematic Classification, Neural Networks 
ABSTRACT: 
Since the 60s we know the basic operation of artificial neural networks which are similar to the components of the human brain. The 
hardware and software necessary for the computation has become adequate just only nowadays. 
It was proved in my experiment that the thematic classification by neural networks is possible. In the experiment I made the 
classification for a LANDSAT TM image with six bands containing Budapest in it by traditional (minimum distance and maximum 
likelihood) and neural methods. I used 2 and 3 layer neural networks with different number of neurons. 
The classification results show that the operation is highly depending on the network structure and training vectors. It's possible to 
find such a structure which has the accuracy of the traditional methods. Inserting further information the accuracy can be increased. 
1. THEORY OF NEURAL NETWORKS 
There are many types of neural networks; I used the so-called 
feedforward and radial basis networks for thematic 
classification. Neurons build layers at these networks. 
Neural Network 
A ne 
HA D Sri 
LC jr WET > nA ne 
er 7 T LA AE 
ne 
EN Re, X : ; 
Inputs : ST, e Ae ^ = 
AU ES. Yd. A rs neis Outputs 
una c Ci CE AT ; 
i4 x ne LT " ui | 
= ne AN 
Sy ne 7 SZ 
2.layer 
Qr 
1.layer 
Figure 1. A 2 layer neural network 
The operation of the neuron is the following: 
l. The incoming signals are multiplied by the 
corresponding weights and are added. 
2. The sum and bias are added. 
3. The transfer function gives the neuron answer. 
Important requirements of the transfer function to be 
differentiable relative easily. In the beginnings the logistic 
sigmoid was used almost exclusively: 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Logistic Sigmoid Transfer Function 
  
Figure 2. The logistic sigmoid transfer function 
Equation of the function: 
fie (1) 
which has the derivative function: 
df x) 
"s er i 
  
(Rojas, 1993). 
Later the line, the tangenthyperbolic function and other curves 
were used as transfer function. Generally they are symmetric 
curves giving answers in a given interval (e.g. between 0 and 1, 
or between -1 and 1 etc.). Lately with the development of new 
methods a new function appeared, which is similar to the 
Gaussian curve. This is the so-called radial basis transfer 
function. 
      
    
    
  
   
  
    
    
    
    
    
    
   
   
     
    
  
    
   
    
    
    
     
   
   
   
   
    
   
    
  
Figure 3. Th 
Equation of t 
The new tra 
changes in | 
1995). 
Let's see ho 
neural netwo 
If we have S 
and have R 
1995) 
O «logs 
where 
O 
I 
W 
B, 
W 
B 
log 
For the prev 
(WI, BI, W 
contain the 
The differen 
are the error 
weights and 
errors. This 
meaning of 
hypersurface 
steps, but toc 
only few epc