Spatial Entropy Based Mutual Information in Hyperspectral Band Selection for Supervised Classification

Abstract

Hyperspectral band image selection is a fundamental problem for hyperspectral remote sensing data processing. Accepting its importance, several information-based band selection methods have been proposed, which apply Shannon entropy to measure image information. However, the Shannon entropy is not accurate in measuring image information since it neglects the spatial distribution of pixels and is computed only from a histogram. This paper investigates the potential of spatial entropy in measuring image information and proposes a new mutual information (MI) band selection method based on the spatial entropy. Then selected band images are validated for supervised classification via Support Vector Machine (SVM). Using a hyperspectral AVIRIS 92AV3C dataset, experiment results show that with 20 images selection from 220 bands, the supervised classification accuracy can reach 90.6%. Comparison with a previous Shannon entropy-based band selection method shows that the proposed method selects band images which can achieve more accurate classification results.