Hyperspectral anomaly detection with self-supervised anomaly prior.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral anomaly detection (HAD) is crucial for Earth observation and military applications.
• Existing methods often rely on low-rank representation (LRR) with handcrafted sparse priors, which can overlook spatial structures and manual sparsity settings.
• This limits the accuracy and interpretability of anomaly detection in hyperspectral imagery.

Purpose of the Study:

• To develop a more accurate and interpretable hyperspectral anomaly detection method.
• To overcome the limitations of handcrafted priors in LRR-based HAD.
• To introduce a self-supervised approach for learning anomaly characteristics.

Main Methods:

• A self-supervised network, self-supervised anomaly prior (SAP), is proposed to redefine the anomaly optimization criterion within the LRR model.
• A novel pretext task involving a classification to distinguish original and pseudo-anomalous hyperspectral images (HSIs) is used for learning.
• A dual-purified strategy is employed to refine background representation using an enriched background dictionary.

Main Results:

• The proposed SAP method demonstrates superior performance compared to existing advanced HAD techniques.
• Experiments on diverse hyperspectral datasets validate the effectiveness and interpretability of the SAP approach.
• The method successfully identifies and locates targets without prior information, even in complex backgrounds.

Conclusions:

• The self-supervised anomaly prior (SAP) offers a significant advancement in hyperspectral anomaly detection.
• The approach provides a more robust and data-driven alternative to handcrafted priors.
• SAP enhances the accuracy and interpretability of anomaly detection in hyperspectral imaging.