312 
TERRAIN CLASSIFICATION BY ARTIFICIAL NEURAL NETWORKS 
Joji Iisaka 
Canada Centre for Remote Sensing 
1547 Merivale Rd. Nepean, Ontario, Canada 
Wendy Russell 
Intera Technologies Ltd. 
Ottawa, Ontario, Canada 
ABSTRACT 
This paper will report the results of an investigation into the applicability of neural networks to the 
classification of remotely sensed images, especially polarimetrie SAR images. Traditionalclassification methods, 
such as maximum likelihood classification for terrain classification, assume simple mathematical models such 
as Gaussian data distribution. Very often, the exact distribution is not well known and a priori probabilities 
must be assumed. However, this behaviour of data is not reflected in the real world and is very difficult to 
model in simple mathematical terms. 
In recent years, SAR technology has advanced to allow the collection of fully polarimetrie data, which includes 
the HH, HV, VH, and W components. The behaviour of polarimetrie data has not been well examined yet. 
Although polarimetrie imagery is becoming more popular because of new capabilities which cannot be provided 
by other sensors, there are few powerful information extraction methods available for it. 
Neural network computing does not need to assume the shapes of data distribution. It has the capability to 
learn when given examples to learn from. Better classification results are expected with the advantages of 
neural network computing. Using a microcomputer-based neural network simulator, this paper will report on 
terrain classification based on the back propagation paradigm.