The Use of Feature Selection Techniques in the Context of Artificial Neural Networks

Feature selection is an important issue, especially for classification problems where artificial neural networks are involved. It is known that using large number of inputs can make the network overspecific and require significantly longer time to learn the characteristics of the training data. Such over-specificity also reduces the generalisation capabilities of a neural network, so the network may fail to classify new data outside the range of the training data. Although feature selection methods have been used in remote sensing studies for many years, their use in the context of artificial neural networks has not been fully investigated. This paper sets out some results of an investigation of feature selection techniques, specifically the separability indices, in the problem of determining the optimum network structure in terms of achieved accuracy. For this purpose, separability indices, including divergence, transformed divergence, Bhattacharyya distance and Jeffries-Matusita distance, and the Mahalanobis distance classifier (MDC) based on two accuracy measures are employed to determine the best eight-band combination out of a 24 band multitemporal dataset. Two search procedures, sequential forward selection and the genetic algorithm, have been used to search for the best band combinations using separability measures as evaluation functions.