Supervised and Unsupervised Data Engineering for Multimedia Data

10Assessment of Adjusted and Normalized Mutual Information Variants for Band Selection in Hyperspectral Imagery

Bhagyashree Chopade, Vikas Gupta and Divyesh Varade^*

Technocrats Institute of Technology, Bhopal, India and Indian Institute of Technology Jammu, Jammu, India

Abstract

Similarity measures from information theory are widely used in feature selection, particularly based on mutual information. Although several methods exist for feature selection in hyperspectral data utilizing information theory-based approaches, seldom are all aspects of information utilized in the developed techniques. Various normalization, adjustment and weighing schemes have been proposed in the literature for mutual information. However, a detailed investigation of these schemes for feature selection is lacking. In this study, we investigate the normalization scheme for different variants of mutual information for unsupervised feature selection in hyperspectral data. We also examine the potential of adjusted mutual information for the selection of best bands in hyperspectral data. We define a novel scheme for the computation of the expectation of mutual information based on the concept that the mutual information between the mean filtered hyperspectral bands at different neighbourhood sizes yield different mutual information with respect to a reference image. Using this concept to determine the expectation of mutual information satisfies the constant baseline criteria used in the adjusted mutual ...