Dictionary trained attention constrained low rank and sparse autoencoder for hyperspectral anomaly detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral anomaly detection benefits from dictionary representations and autoencoders (AE).
• Dictionary methods offer interpretability but struggle with complex scenes.
• Autoencoders excel in complex scenes but lack self-explanation and spatial detail integration.

Purpose of the Study:

• To develop a hyperspectral anomaly detection method that integrates the strengths of dictionary representations and autoencoders.
• To enhance the utilization of spatial information in hyperspectral anomaly detection.
• To improve the accuracy and robustness of anomaly detection in complex scenarios.

Main Methods:

• Proposed an attention-constrained low-rank and sparse autoencoder (AE) model.
• Incorporated a Global Self-Attention Module (GAM) for global spatial context in the low-rank AE.
• Integrated a Local Self-Attention Module (LAM) for local spatial details in the sparse AE.
• Employed nonlinear fusion to combine detection results from both AE components.

Main Results:

• The proposed AE model demonstrated superior performance in hyperspectral anomaly detection.
• The attention mechanisms effectively captured and utilized both global and local spatial information.
• The method showed significant improvements in background and anomaly reconstruction.
• Experiments on diverse datasets confirmed the model's effectiveness and feasibility.

Conclusions:

• The attention-constrained low-rank and sparse AE is a powerful approach for hyperspectral anomaly detection.
• Integrating self-attention mechanisms significantly enhances the model's ability to leverage spatial information.
• The proposed method offers a promising solution for detecting anomalies in complex hyperspectral imagery.