Band selection in Hyperspectral Images using information similarity ranking

Band selection in hyperspectral images is crucial for reducing redundant features and improving subsequent tasks like classification and segmentation. This article presents a novel band ranking method based on information similarity measures from the concepts of information theory. We introduce the Normalized Mutual Information for similarity and Variation of Information for dissimilarity to create a ranking scheme based on the affinity of a band to become a cluster center. We explore hybrid schemes combining structural and informational attributes. Supervised approaches using label-wise similarities using ground truth supervision are also investigated. Evaluation against state-of-the-art band selection algorithms on popular datasets, using both traditional Machine Learning and Deep Learning models, demonstrates superior results. This study underscores the significance of informational and structural attributes in unsupervised band selection and the potential for further enhancement through supervised similarity with ground truth information.