Optimal Band Selection in Hyperspectral Images Using Improved K-Means Clustering with Spectral Similarity Measures

Hyperspectral imaging provides abundant spectral information with hundreds of bands, but selecting an optimal number of bands is crucial for efficient and accurate data analysis. The k-means clustering method is widely used for band selection, but the quality of clustering and efficiency of band selection depends on the similarity measure used. In this paper, we propose an improved version of k-means clustering using spectral similarity measures (SSM) such as Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), and the hybrid measure of SID and SAM (SIDSAM) for optimal band selection. Empirically it is proved that the proposed k-means with spectral similarity measures will identify the best bands and thereby improve the classification accuracy.