A Superpixel and Mutual Information Based Band Selection Method for Hyperspectral Images

Abstract

Dimension reduction is an important task in hyperspectral image (HSI) processing to minimize computational complexity. This work presents an efficient band selection method to select a set of highly relevant yet non-redundant bands using class-band mutual information (MI) and clustering technique.In HSI, the bands comprise continuous reflectance values, while the class labels in the ground truth are discrete. Thus before calculating class-band MI, the continuous reflectance values are converted to several discrete values using a novel graph-based superpixel generation method based on the spectral angle divergence (SAD). To ensure non-redundancy of the selected bands, clustering is performed to using correlation distance and then from each cluster only one representative band is selected using the calculated class-band MI. Experiments are carried out over three hyperspectral images to evaluate the efficacy of the proposed band selection technique. For each image, classification is performed using Support Vector Machine (SVM) classifier over both the compressed image with only the selected bands and the original image. The comparison results reveal the capability of the proposed method to maintain the classification accuracy with a substantially compact set of selected bands.