Band Selection by Divergence Distance Based on Gaussian Mixture Model for Hyperspectral Image Classification

In this work, we investigate a new band selection approach by Divergence distance based on the Gaussian Mixture Model (GMM) for Hyperspectral image classification. The main motivation in modeling the Divergence distance with GMM is due to the fact that GMM is well known to be less sensitive to estimation error problem than non-parametric models and can capture non Gaussian statistics of multivariate data. To estimate the parameters of GMM, the Expectation Maximization (EM) with the Bayesian Information Criterion (BIC) and a Robust Expectation Maximization (REM) algorithm are used. This investigation is inspired by our previous work on the Bhattacharyya distance hence we are particularly interested in using the Divergence distance to find out which one gives better results. The performances of the proposed approach are compared to those of the Bhattacharyya distance in terms of global classification accuracy and numbers of retained bands through two Classifiers; Extreme Learning Machine (ELM) and Support Vector Machine (SVM). The experiments are carried out on three hyperspectral images, the Indiana Pines (92AV3C), the Botswana and the Kennedy Space Center dataset (KSC).