5 - Hyperspectral image classifications and feature selection

Abstract

This chapter discusses the use of a modified radial basis function (RBF) network for the classification of hyperspectral data. The use of RBF networks has already been reported in remote sensing literature with different approaches to determine the weights between the hidden and output layers. The RBF network proposed in this chapter uses Cholesky decomposition and the least mean square error method to determine the weights between the hidden layer and the output layers. The availability of the large number of features in hyperspectral data represents a challenge due to data redundancy for classification analysis, and the use of too many features requires a larger training set size. A collection of a large amount training data is difficult and costly, therefore, a feature selection process has been adopted to reduce the requirement for ground data for hyperspectral data classification. The Bayesian framework for feature selection, a filter-based feature selection algorithm based on Bayesian theory and the receiver operating characteristic, is used in this chapter.