Multi-Scale Autoencoder Suppression Strategy for Hyperspectral Image Anomaly Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Autoencoders (AEs) are widely used for hyperspectral anomaly detection (HAD).
• Existing AE methods struggle with spatial information and may reconstruct anomalies, reducing detection accuracy.
• There is a need for improved AE strategies that prioritize background reconstruction.

Purpose of the Study:

• To propose a novel Multi-scale Autoencoder Suppression Strategy (MASS) for enhanced hyperspectral anomaly detection.
• To improve the accurate reconstruction of background information while suppressing anomalies.
• To enhance the overall performance of hyperspectral anomaly detection.

Main Methods:

• Developed a Multi-scale Autoencoder Suppression Strategy (MASS).
• Integrated Local Feature Extractor (Convolution and ODConv) and Global Feature Extractor (Transformer) for multi-scale feature extraction.
• Introduced a Self-Attention Suppression (SAS) module to reduce anomaly influence and focus on background reconstruction.
• Incorporated an iterative mask into the loss function to guide background learning.

Main Results:

• The proposed MASS method demonstrated superior performance compared to traditional and deep learning methods.
• Experiments on eight datasets confirmed the effectiveness of the MASS strategy.
• The method successfully prioritized background reconstruction over anomaly reconstruction.

Conclusions:

• The MASS strategy significantly improves hyperspectral anomaly detection accuracy.
• The integration of multi-scale feature extraction and self-attention suppression is effective.
• MASS offers a promising approach for robust hyperspectral anomaly detection.