Multi-Scale Semantic Selection and Spatial Constraint-Guided Network for Cross-Scene Hyperspectral Image...

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Cross-scene classification of hyperspectral images (HSI) faces challenges due to significant distribution discrepancies between different scenes.
• Existing methods often fail to fully utilize spatial semantics during sample expansion and lose spatial information by compressing features into 1D vectors.

Purpose of the Study:

• To propose a novel network, MSCGnet, that addresses the limitations of existing methods in cross-scene HSI classification.
• To improve the exploration of spatial semantics and preserve spatial location information during feature alignment and sample augmentation.

Main Methods:

• The proposed multi-scale semantic selection generator optimizes Mamba's tokenization with a spatial diffusion scanning strategy, arranging pixels from center to periphery to maintain spatial continuity.
• Multi-scale spectral-spatial features are extracted, and a semantic selection matrix guides Mamba for diverse augmented sample generation.
• A spatial constraint-guided discriminator uses class activation map projection for explicit spatial constraints on feature distributions.

Main Results:

• MSCGnet achieves superior classification accuracy on multiple cross-scene HSI datasets.
• The method demonstrates enhanced generalization performance compared to existing approaches.
• The proposed network maintains low model complexity while delivering high performance.

Conclusions:

• MSCGnet effectively mitigates domain shift in cross-scene HSI classification by preserving and leveraging spatial semantics.
• The spatial constraint-guided approach enhances the reliability of augmented samples and improves classification outcomes.
• The method offers a promising direction for robust hyperspectral image analysis across diverse scenes.