A Comparison of Column Subset Selection Methods for Unsupervised Band Subset Selection in Hyperspectral Imagery

Abstract

This paper explores the use of column subset selection (CSS) for unsupervised band subset selection (BSS) in hyperspectral imaging. CSS is the problem of selecting the most independent columns of a matrix. Many deterministic and randomized algorithms have been proposed in the literature for CSS. This paper presents a comparison between different algorithms for CSS for BSS. The cosine of the angle between the range space spanned by the selected bands and the corresponding left singular vectors is used to evaluate the quality of the selected bands to represent the image. Numerical experiments are conducted using multispectral and hyperspectral data. Results show that SVDSS outperforms other deterministic algorithms while producing comparable results to a 2-stage randomized CSS in small images and in centered data. However, the randomized algorithm significantly outperforms deterministic approaches in large images.