A comparative analysis of mutual information based feature selection for hyperspectral image classification

Feature selection is an important task for hyperspectral imagery classification and becomes more critical for the emerging big data analysis. Selection criteria based on mutual information theory have the advantages in terms of distribution free, nonlinearity and low computational load for multiclass cases. However several have been developed and are available to use. In this study, we conduct a comparative analysis on four defined criteria and their performances are evaluated using two hyperspectral data sets with two levels of sample sizes.