ality 
the 
jhts 
as. 
der 
pod 
lich 
tter 
lich 
ose 
| to 
cul- 
are 
lata 
tely 
sks 
ca- 
and 
tur- 
em. 
Sor 
are 
und 
red 
the 
and 
nal 
ern 
Jeo- 
68- 
In- 
nal 
ing, 
jata 
| Si- 
tral 
nfe- 
ulti- 
for 
-irst 
ce, 
fied 
ote 
nul- 
)70- 
Tr 
Classification of TM image Using a Competitive Learning Neural Network 
Jee-cheng Wu 
jewu@magnus.acs.ohio-state.edu 
Department of Civil and Environmental Engineering and Geodetic Science 
The Ohio State University, U.S.A. 
Committee VII, WG/1 
Abstract 
Recent progress in Artificial Neural Networks (ANNs) research has demonstrated the 
usefulness of ANNs in variety of applications. In remote sensing community, a 
supervised learning network paradigm, Back Propagation (BP), has been successful 
applied to land-cover classification of satellite images. Two of the major problems 
associated with BP network for land-cover classification are that its convergence time is 
usually very long, and it does need a suitable training data set. This paper describes 
applying an unsupervised learning network paradigm, Competitive Learning (CL), to 
land-cover classification of TM image on a Winner-Take-All (WTA) basis. The final 
results showed that CL of WTA still needs to improvement for satisfying the 
requirements of land-cover classification. 
1. Introduction 
The renewed interest in Artificial 
Neural Networks (ANNs) is mainly due 
to the development of multi-layer 
learning algorithms ([17]), which enable 
the ANNS to learn using a more complex 
Structure to solve the problems in the 
book  Perceptrons ([16]). Another 
important reason is that more internal 
dynamics of the network is revealed 
through insights into the fundamental 
laws governing the convergence of the 
group behavior of interacting physical 
elements. Parallel ANNs of the type 
described by Hopfield, Kohonen, and 
Grossberg have been shown to be well 
775 
suited to the task of pattern recognition 
and classification ([4] [14]). Generally it 
involves setting up a network 
architecture, and then training the 
network through a set of training data. 
The result network is then found to be 
capable of classify subsequent testing 
data in terms of a set of training data. 
In remote sensing community, a 
number of researchers have 
demonstrated the use of ANNs 
techniques. Back Propagation (BP), a 
supervised learning paradigm of ANNs, 
has been shown to be useful in 
classification of satellite images. For 
examples, a number of researchers ([3] 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996