Fewer interpretable bands via self-supervision for hyperspectral anomaly detection

Abstract

Hyperspectral band selection (BS) algorithms aim to identify the most important bands relevant to model decisions. While existing methods can effectively assign importance to different bands, they face challenges in unsupervised hyperspectral anomaly detection (AD) due to the lack of predefined labels to guide the selection process. This paper addresses these challenges by proposing a self-supervised approach for identifying fewer interpretable bands that best separate background and anomalous data. By formulating band selection as a subset selection problem using pseudo-anomalies, bands are chosen based on high confidence, rich feature diversity, and inter-collaboration. Specifically, the pseudo-anomalies and the corresponding map are utilized to identify bands with high confidence through a self-supervised strategy. To further enhance feature diversity and collaboration among bands, we impose a feature diversity constraint on the selection of subsets and assess the collaboration ability of various subsets. A novel evaluation function is designed to discover more useful bands for AD. Experiments demonstrate the effectiveness of each module and show that the proposed method selects bands that are beneficial for AD across four different datasets. The code is released at https://github.com/rk-rkk/Fewer-Interpretable-Bands-for-Hyperspectral-Anomaly-Detection.