Uncertainty-Driven Cyclic Background Purification for Hyperspectral Anomaly Detection

Abstract

Hyperspectral anomaly detection (HAD) aims to identify spectrally distinct and spatially sparse targets from complex backgrounds without prior anomaly labels. In recent years, reconstruction-based deep learning (DL) methods have achieved detection by modeling the background distribution. However, due to the influence of imaging conditions, detection networks inevitably reconstruct anomalies alongside the background. Some recent methods indirectly mitigate anomaly overfitting by applying a series of delicate masking strategies, but they usually rely on predefined occlusion rules that are not fully adaptive to diverse anomaly distributions. To address this issue, we propose an uncertainty-driven cyclic framework for hyperspectral anomaly detection (UC-HAD), where pixel-wise reconstruction uncertainty is introduced as a data-adaptive cue to guide background modeling. Within UC-HAD, uncertainty estimation and background reconstruction operate cyclically. Specifically, UC-HAD estimates pixel-wise reconstruction uncertainty by approximating the posterior distribution over model parameters, thereby characterizing reconstruction reliability across spatial locations. Then, an uncertainty-guided adaptive filling strategy (UAFS) is designed. Pixels with low uncertainty are used to adaptively repair high-uncertainty regions, effectively reducing the anomaly interference in subsequent background modeling. As the loop proceeds, the reconstructed background becomes progressively cleaner, and the separability between anomalies and backgrounds is correspondingly enhanced. Experiments on six public HAD datasets demonstrate that UC-HAD consistently outperforms competitive methods, validating the effectiveness of uncertainty-guided cyclic background purification for HAD. The code is available at https://github.com/tobeaTrueman/UC-HAD