Improved K-Means Clustering Algorithm for Band Selection in Hyperspectral Images

Band selection plays a major role in hyperspectral pixel classification. Clustering-based approaches gained more attention in recent years for optimal band selection. Despite some limitations in identifying non-convex-shaped clusters, the k-means clustering-based band selection is widely used because of its simplicity to implement. The quality of clustering and the efficiency of band selection depends on the similarity measure. This article presents an improved k-means clustering algorithm for band selection using the spectral similarity measures like Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), and the hybrid similarity measure of SID and SAM (SID-SAM), Normalised Cross-correlation (NCC), Jeffery-Matusita (JM), Jeffery-Matusita With Spectral Angle Mapper (JM-SAM). Further, the efficiency of these spectral similarity measures for clustering based band selection are analyzed on various benchmark datasets utilizing the state-of-the-art machine learning classifiers like K-Nearest Neighbours(KNN), Decision Tree(DT) and Support Vector Machines(SVM) for pixel classification of HSI.