Predictive Assessment of Neural Network Classifiers For Applications In GIS

Gordon German, Mark Gahegan and Geoff West

Department of GIS, School Of Computing,
Curtin University, P.O. Box U 1987, Perth 6001,
Western Australia

Presented at the second annual conference of GeoComputation ‘97 & SIRC ‘97, University of Otago, New Zealand, 26-29 August 1997

Abstract

Artificial Neural Networks (ANNs) are well suited to implementing supervised classification tools for GIS data. They make no assumptions about the statistical nature of the data, can be used with ordinal and nominal data types together and can be trained with comparatively few training points, as they do not have to choose a data distribution model, unlike techniques such as Maximum Likelihood Classification.

However, training these neural network classifiers can be a time-consuming process, with no guarantee of the outcome. In this paper, the author presents a methodology for determining whether learning is practical for a given network on a given data-set, prior to commencement of the training phase. This is achieved by examining the error scores at the initial class boundaries and checking for redundancy in the network hyperplanes. This redundancy indicates how much flexibility is available in the network to learn complex boundaries.