Feature Selection Techniques for Classification of Satellite Images with Fuzzy Artmap Neural Networks

The selection of most relevant features can significantly reduce the resource requirement associated with fuzzy ARTMAP neural networks for the classification of high-dimensional data extracted from satellite imagery. This paper introduces a variant of a wrapper-type feature ranking technique that was previously proposed for ARTMAP neural networks by Parsons and Carpenter. As with the originally proposed technique, it evaluates the relevance of features based solely on between-class scatter from the geometry of internal ARTMAP categories. This paper also explores the inclusion into these feature ranking techniques of a within-class scatter measurement. Comparative simulations, performed on the Landsat multi-spectral imagery benchmark ('Satimage' from the StatLog repository), indicate that a significantly lower generalization error and fewer resources may be achieved by learning subsets produced by techniques that evaluate the relevance of features using both betweenand within-class scatter.