Maximum relevance, minimum redundancy feature extraction for hyperspectral images

In this paper we propose a new feature extraction scheme for hyperspectral images based on mutual information. Relevance of extracted feature set to class label has been measured by average of mutual information between each of them and class label and Redundancy of them is measured by average of mutual information between each pair of them. Based on relevance of features and redundancy between them, we propose a cost function that maximize relevance of extracted features and simultaneously minimize redundancy between them. This cost function has been already used for feature selection. In this paper we will find the parameters of an optimal linear mapping by optimizing the proposed cost function with respect them. Linear methods are attractive due to their simplicity. Because of nonlinear and nonconvex relation between proposed cost function and the parameters, we use genetic algorithm for optimization. Mutual information accounts for higher order statistics, not just for second order as PCA and LDA do. Hence mutual information is a better criterion for hyperspectral images because they have higher order statistics than two. Our classification results for AVARIS data shows proposed method has better performance over PCA and LDA.