Highly-Constrained Unsupervised Hyperspectral Band Selection via Quality-aware Latent Feature Clustering

Abstract

Unsupervised band selection (UBS), which reduces dimensionality without relying on costly labeled data, is pivotal for hyperspectral image (HSI) analysis. However, highly-constrained UBS scenarios, defined by selecting fewer than ten bands, present significant challenges in achieving promising performance for downstream tasks. As such, we propose quality-aware latent feature clustering (QLFC), where a novel band-based adjacency-quality-weighted distance metric is introduced to construct an HSI graph to integrate interband physical correlations with quality priors. This metric facilitates a graph decomposition that maps bands into a latent space where high-quality bands gravitate toward cluster centroids while degraded ones are marginalized. Subsequently, a quality-guided adaptive band selection strategy is proposed. By integrating an adaptive threshold based on the desired band count with a secondary refinement stage focused on band priority, the framework reliably captures the most informative representatives within each cluster. Extensive experiments on HSI anomaly detection and classification across four benchmark datasets demonstrate the superiority of QLFC under varying numbers of selected bands. For anomaly detection, it enhances the detection accuracy while simultaneously reducing the computational overhead by up to 59.49%. For classification, QLFC consistently outperforms six state-of-the-art methods, notably achieving up to 13.55% higher classification accuracy than competitors with only three selected bands. These results highlight QLFC's superior ability to preserve critical spectral information, making it suitable for UBS enabled resource-constrained applications.