Work

KM-CV-SAM: A Weighted K-Means Band Selection Method for Hyperspectral Image Classification

Year: 2026
Type: preprint
Abstract: Hyperspectral image (HSI) classification is an active research area with various applications, including forestry, land cover analysis, defense, and environmental monitoring. Along with numerous applications, HSI analysis is challenging due to the large number of spectral bands with high redundancy, which demands high computational cost. Band selection plays an important role in identifying compact and informative spectral band subsets. To this end, this article presents a weighted clustering-based band selection framework termed KM-CV-SAM. The proposed approach follows a two-step process. First, a weight is assigned to each band. In the second step, weighted bands are grouped into a predefined number of clusters, and representative bands are selected from each cluster. The weight for each band is computed using the Coefficient of Variation (COV) and Spectral Angle Mapper (SAM), thereby emphasizing discriminative and information-rich bands and yielding more representative and diverse subsets. To establish the effectiveness of the proposed approach, a rigorous experimental study has been conducted on publicly available benchmark HSI datasets, including Indian Pines, Pavia University, Salinas, and WHU-Hi-LongKou, covering a wide range of land-cover classes and imaging conditions. Experimental results show that KM-CV-SAM consistently outperforms several state-of-the-art band selection techniques. The model achieves improved classification accuracy with only 5–30 selected bands with a little compromise in classifcation accuracy. The proposed framework is well suited for practical hyperspectral applications, particularly in scenarios with limited labeled samples and constrained computing resources.
Language: English
Open Access status: green
OpenAlex ID: W7235454627
Cites: 23
Cited by:
Related to: 10