[8] [15]) have used BP for land-cover 
classification, and Lee et al. (1990) used 
BP to classify cloud segmentation. But, 
there are some inhibit problems when 
use BP network for  land-cover 
classification. For examples, given a 
large training data sets required to form 
an adequate representation of the input 
vector space in land-cover classification, 
the BP training algorithms appear 
infeasible using exist computer 
technology because its convergence is 
time consuming. Also, changes in the 
band or scene selection means that a 
classifier under one condition is 
inappropriate for use in another 
condition. Under these circumstances 
(both for handling large training data 
sets and changing  band/scene 
conditions), it may appropriate to work 
with unsupervised learning networks. 
In unsupervised learning, a 
network is expected to self-organize to a 
state the reflects the distribution of input 
patterns while a desired response is not 
given in an explicit form. Therefore, 
unsupervised learning has a possibility 
of discovering unknown relationships 
among input patterns and can be a model 
of rule finding or concept formation. In 
this paper a Competitive Learning (CL) 
network was applied to land-cover 
classification of TM image on a Winner- 
Take-All (WTA) basis. 
2. Methods 
Competitive Learning (CL) 
Networks 
The basic competitive learning 
network is a two-layer, fully connected, 
feedforward network. 
776 
Output Layer ® © eo ® 
Input Loyer ® © 
Figure 1. A Competitive Learning Network 
| 
yn 
A 
b 
Input Vector yl 
The two layers of competitive 
learning network such as the first layer 
consists of n input neurons, and the 
second layer includes k output neurons ( 
Fig. 1). The weight (synaptic) 
connection between the input and output 
neurons is denoted by Wi (i-1,2 ...k). 
The learning of the network is carried 
out by changing the weight (synaptic) 
connection Wi between the neurons in 
the input and output layers. For details, 
further references and applications see 
0101] [17]. 
Competitive Learning with 
Winner-Take-All(WTA) Activation 
Competitive learning can take a 
variety of forms depending on the 
precise update rule used and the method 
for implementing competitive activation 
mechanisms. There exist in the literature 
numerous suggestions that Winner-Take- 
All (WTA) property is based on 
technical as well as biological principles 
([6]). Therefore, one of the most 
common methods is to activate the most 
"excited" neuron and then allow that 
neuron to modify its weight vector using 
a standard linear update rule. The WTA 
activation rules are usually used instead 
of fixed threshold tests, and will activate 
the single neuron whose excitation level 
is the greatest. In practice, unsupervised 
CL network amount to centroid 
estimation and nearest-neighbor 
classification when using WTA property. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996