Hyperspectral Anomaly Detection via Hybrid Convolutional and Transformer-Based U-Net With Error Attention Mechanism

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Hyperspectral anomaly detection identifies pixels with unique spectral signatures.
• Existing methods struggle to simultaneously utilize spectral and spatial information.
• This limitation hinders the performance of traditional hyperspectral anomaly detection techniques.

Purpose of the Study:

• To propose a novel framework, HCT-Unet, for enhanced hyperspectral anomaly detection.
• To improve the simultaneous leveraging of spectral and spatial information in hyperspectral images (HSIs).
• To overcome the limitations of traditional anomaly detection methods.

Main Methods:

• Developed a hybrid convolution and Transformer-based U-Net (HCT-Unet) architecture.
• Integrated convolutional layers for local feature extraction and Transformers for long-range dependency modeling.
• Introduced an error attention mechanism for adaptive multiscale feature fusion and proposed a new anomaly score combining reconstruction error and SSIM.

Main Results:

• The HCT-Unet architecture effectively captures both local and global spatial-spectral interactions.
• The error attention mechanism significantly enhances feature representation capacity.
• Experimental results on seven datasets demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed HCT-Unet framework offers a significant advancement in hyperspectral anomaly detection.
• The hybrid approach effectively addresses the limitations of traditional methods.
• The novel anomaly scoring method provides a robust measure for pixel anomaly identification.