d feedfor- 
ncept from 
longing to 
(Restricted 
CE got its 
yous to the 
site electri- 
ork. Input 
he internal 
An output 
re active it 
993). 
out Layer 
len Layer 
it Layer 
work 
IS of attrac- 
ws a simple 
are the at- 
ptic weight 
je ith basin 
ctor, i, falls 
ciated with 
iss B 
nternal neuron 
S not active 
1al neuron 
ive 
th training 
raction; the 
hen training 
' two rules, 
'oduce these 
1. Ifa training vector is applied that does not lie within a basin 
of attraction of the same class as that of the training vector, 
a basin of radius 6,64, is created, centred at that training 
vector. The radius is chosen to be less than the distance to 
the center of the nearest basin of any other class (Fig. 4). 
Class A 
~ 
Ne 
Class B 
Class A 
oy 
5 
Yor 
Class B 
Figure 4: Insertion of a new basin of attraction 
2. If an applied training vector falls in the basin of attraction 
of a different class, the radius of that basin is reduced until 
the training vector lies just outside of the basin (Fig. 5). 
Following these steps, training vectors generate multiple basins 
approximating contours of underlying classes A and B (Fig. 6). 
3. EXPERIMENTS 
The area selected for study is the city of Santos/Brazil; the input 
data used in this study is a geocoded 7-channel Landsat TM 
image acquired on 16 April 1992. A subscene of 384 by 384 
pixels which covers a large portion of the harbour of Santos 
(approximately 9.6 km by 9.6 km) was selected for the classifi- 
cation tests. A black-and-white reproduction of a natural false- 
color composite covering the subscene is shown in Figure 7. 
The objective of the study is the discrimination between the 
intra-urban land-use classes ‘Residential Area’, ‘Industrial 
Area’. ‘Docks’, and ‘Other Areas’. The training stage of the 
classification procedure is given by the following steps: 
73 
Class A 
pois 
    
Class B 
  
Figure 6: Approximated decision regions 
Step 1: An unsupervised ISODATA classification was executed 
on the 7-channel TM-subscene. The resulting land- 
cover map consists of 20 classes. 
Step 2: Representative training samples for each land-use class 
were selected. 
Step 3: Modified co-occurrence matrices based on a buffer 
distance of 100 m were generated for each region of the 
training samples. The elements beneath the main diago- 
nal of the region-based co-occurrence matrix form an 
input vector for the ATL network. The output layer 
consists of four neurons, one neuron for each land-use 
class. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996