Optimizing hyperspectral classification: spectral similarity-based band selection with chord k-means

Abstract

Band selection is crucial for achieving high classification accuracy in hyperspectral image (HSI) analysis, especially when ground truth data are limited. While unsupervised algorithms are preferred in such scenarios, the effectiveness of k-means clustering depends heavily on the choice of similarity measure. This article presents a novel two-level clustering approach for band selection. In the first level, bands are clustered using k-means with various similarity measures such as Euclidean distance, spectral angle mapper (SAM), and spectral information divergence (SID). Subsequently, the second level leverages the chord metric k-means clustering to form clusters of HSI scenes upon optimal band clusters from the first level. This initial band selection reduces dimensionality and guides subsequent k-means clustering. The proposed chord-based clustering method, utilizing the chord metric, outperforms standard k-means variants, demonstrating significant improvements in accuracy. Experimental results on publicly available hyperspectral datasets confirm the effectiveness of the proposed approach as an alternative to traditional k-means algorithms, showcasing significant improvements in accuracy.