A comparison of several feature selection scores applied to hyperspectral data

Band selection consists in identifying the most relevant band subset for a specific application. It is a way to reduce hyperspectral data huge dimensionality and can be applied to design specific superspectral sensors. Automatic feature selection (FS) algorithms are used to achieve this selection. It aims at selecting an optimal band subset according to a criterion. This study focuses on a comparison of several such FS criteria, ranging from simple classic FS scores (e.g. JM separability measure) to wrapper classification accuracies using state-of-the-art classifiers (e.g. Random Forests) making it possible to cope to some extent with complex multi-aspect classes. All these criteria were tested in a stochastic FS optimization framework. They were evaluated considering classification performances reached using several classifiers.