Hyperspectral Anomaly Detection with Self-Supervised Anomaly Prior

TL;DR

This work tackles hyperspectral anomaly detection (HAD) by replacing handcrafted sparsity priors in low-rank representation (LRR) with a self-supervised anomaly prior (SAP) learned via a dedicated pretext task. The pretext task trains a ResNet34-based classifier to distinguish original HSIs from pseudo-anomalies generated as prism-like prisms with arbitrary polygons and spectral bands, producing a generalizable anomaly prior used in a plug-and-play LRR solver. A dual-purified dictionary construction further refines background modeling, and an ADMM-style optimization unifies the self-supervised prior with background/decomposition terms. Experiments on four HAD benchmarks show SAP achieving state-of-the-art detection accuracy and interpretability, with ablations confirming the benefits of the SSL prior and the enriched dictionary. The approach offers robust HAD performance without manual sparsity tuning and demonstrates strong generalization across sensors and scenes.

Abstract

The majority of existing hyperspectral anomaly detection (HAD) methods use the low-rank representation (LRR) model to separate the background and anomaly components, where the anomaly component is optimized by handcrafted sparse priors (e.g., $\ell_{2,1}$-norm). However, this may not be ideal since they overlook the spatial structure present in anomalies and make the detection result largely dependent on manually set sparsity. To tackle these problems, we redefine the optimization criterion for the anomaly component in the LRR model with a self-supervised network called self-supervised anomaly prior (SAP). This prior is obtained by the pretext task of self-supervised learning, which is customized to learn the characteristics of hyperspectral anomalies. Specifically, this pretext task is a classification task to distinguish the original hyperspectral image (HSI) and the pseudo-anomaly HSI, where the pseudo-anomaly is generated from the original HSI and designed as a prism with arbitrary polygon bases and arbitrary spectral bands. In addition, a dual-purified strategy is proposed to provide a more refined background representation with an enriched background dictionary, facilitating the separation of anomalies from complex backgrounds. Extensive experiments on various hyperspectral datasets demonstrate that the proposed SAP offers a more accurate and interpretable solution than other advanced HAD methods.

Figures (10)

• Figure 1: Schematic of the proposed SAP method.
• Figure 2: The flowchart of the pretext task for self-supervised learning.
• Figure 3: Schematic of the proposed dual-purified strategy. (a) The output of estimator network. (b) and (c) are the purified results. The atoms farther from the class center indicate a lower probability of occurrence.
• Figure 4: Analysis of parameters $\lambda$ over experimental datasets
• Figure 5: Visual detection results with different anomaly priors on four HAD datasets. (a) With self-supervised prior. (b) With $\ell_{2,1}$-norm.