Self-Supervised Masked Graph Autoencoder for Hyperspectral Anomaly Detection

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Hyperspectral anomaly detection is challenging due to difficulties in annotating targets.
• Autoencoder (AE)-based methods excel at image reconstruction but struggle with long-range dependencies and non-Euclidean data.
• Traditional grid-based methods fail to capture complex spatial-spectral relationships in hyperspectral images.

Purpose of the Study:

• To propose a novel self-supervised method for hyperspectral anomaly detection.
• To address limitations of existing methods in capturing long-range dependencies and non-Euclidean structures.
• To improve the accuracy and robustness of anomaly detection in hyperspectral imagery.

Main Methods:

• A Masked Graph AutoEncoder (MGAE) framework is proposed, utilizing a Graph Attention Network (GAT) autoencoder.
• A topological graph structure is constructed for hyperspectral images, processed by the GAT autoencoder with a multi-head attention mechanism.
• A re-masking strategy and a novel loss function (Twice Loss) with graph Laplacian regularization are introduced to enhance reconstruction and prevent trivial solutions.

Main Results:

• The MGAE model effectively reconstructs the background of hyperspectral images.
• The method demonstrates superior performance in identifying anomalous targets compared to existing techniques.
• Experimental results on multiple real-world hyperspectral datasets validate the effectiveness of MGAE.

Conclusions:

• MGAE offers a robust and effective solution for hyperspectral anomaly detection.
• The proposed self-supervised approach overcomes key challenges in hyperspectral image analysis.
• The integration of graph attention networks and masking strategies significantly advances the field.