Neural Networks Applied To The Classification Of Remotely Sensed Data

Abstract

A Neural network with topology 2-8-8 is evaluated against the standard of supervised non-parametric maximum likelihood classification. The purpose of the evaluation is to compare the performance in terms of training speed and quality of classifica-tion. Classification is done on multispectral data from the Thematic Mapper(TM3,TM4) in combination with a ground reference class map. This type of data is familiar to professionals in the field of remote sensing. This means that the position of clusters in feature space is well known and under-stood, and that the spatial pattern is equally well known. As a spin-off, the application of a neural net to a classical task of statistical pat-tern recognition helps to demystify neurai networks. neural nets, k-nearest neighbours, remote sensing, classification INTKODUCTION After the sobering up, out of speculations about neural networks as prevalent in the behaviouristic school of artificial intelligence, in the late nineteen-seventies, as a iesult of the publication on Perceptrons by Minsky (Minsky, 1969) (2), a new wave of speculations started off in tne early nineteen-eighties. Neural networks are of technical interest because of the parallel nature of the calculations. Several manufacturers are attempting to build and market hardware neural necwork devices now. In case these parallel processing devices become available at the right pricr/performance figure, image processing of remote sensing data, with its massive data flows, could profit from these devices. For this reason it is worthwhile to investigate the performance of simulated (neural) nets, when applied to e.g. computation intensive classifica-tion tasks. Duda & Hart (Duda, 1973) (1) evaluated the performance of perceptrons for speech recog-nition. Tne authors decided to evaluate the performance of