Feature Selection Based on Mutual Information and Its Application in Hyperspectral Image Classification

Abstract

This paper investigates mutual information-based feature selection for high dimensional hyperspectral imagery, which accounts for both the relevance of features on classes and the redundancy among features. A representative method shortly known as min-redundancy and max-relevance (mRMR) was adopted and compared with a baseline method called Max- Relevance (MR) in experiments with AVIRIS hyperspectral data. Supervised classifications were also carried out to identify classification accuracies obtainable with hyperspectral data of reduced dimensionality through five different classifiers. The results confirm that mRMR is more discrimination- informative than MR in feature selection due to the additional redundancy analysis. Different classifiers with different accuracies manifest that a more impact but more informative subset may exist. However, the intrinsic dimensionality which indicates the optimal performance of a classifier remains an issue for further investigation.