Hyperspectral Image Classification via Cascaded Spatial Cross-Attention Network.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral images (HSIs) contain rich spectral information, but classifying land cover (LC) using limited bands leads to information loss and poor accuracy.
• Distinguishing between various LC classes in HSIs is challenging due to overlapping spectral signatures and spatial complexities.

Purpose of the Study:

• To propose a novel deep learning method, Cascaded Spatial Cross-Attention Network (CSCANet), for accurate and robust HSI classification.
• To address the information loss and low average accuracy issues in traditional HSI classification methods.

Main Methods:

• Developed a cascaded spatial cross-attention module that integrates local and global spatial features.
• Employed a group cascade structure for sequential propagation of spatial information across channels.
• Designed a two-branch feature separation structure to enhance the separability of spatial-spectral features for different LC classes.

Main Results:

• CSCANet achieved excellent performance in enhancing classification accuracy for HSIs.
• The proposed method demonstrated improved robustness in distinguishing between various land cover classes.
• Experimental results validated the effectiveness of the cascaded spatial cross-attention and feature separation mechanisms.

Conclusions:

• CSCANet effectively leverages spectral-spatial information for improved HSI classification.
• The method offers a robust solution for land cover mapping using hyperspectral data.
• The proposed architecture provides a significant advancement in the field of hyperspectral image analysis.