Dynamic semantic-geometric guidance and structure transfer network for cross-scene hyperspectral image...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cross-scene hyperspectral image classification (HSIC) using domain adaptation is a growing research area.
• Existing methods often fail to fully mine source domain information or use noise-sensitive characterization, leading to negative transfer and performance degradation.

Purpose of the Study:

• To propose a novel Dynamic Semantic-Geometric Guidance and Structure Transfer (DSGG-ST) network for enhanced cross-scene HSIC.
• To overcome limitations of existing domain adaptation methods in HSIC by improving information mining and structure transfer.

Main Methods:

• Introduced a Dynamic Semantic-Geometric Guidance (DSGG) module for domain-invariant learning from semantic and geometric perspectives.
• Developed a Graph Attention Learning-Matching (GALM) module utilizing Graph Attention Networks and SeedGNN for effective structure information transfer and alignment.

Main Results:

• The proposed DSGG-ST network achieved new state-of-the-art (SOTA) performance on three common cross-scene hyperspectral image datasets.
• Experimental results validated the effectiveness of the DSGG module in guiding domain alignment and the GALM module in structure transfer.

Conclusions:

• The DSGG-ST network offers a robust solution for cross-scene hyperspectral image classification by effectively handling domain shift.
• The proposed approach demonstrates significant improvements over existing methods, highlighting the importance of dynamic semantic-geometric guidance and advanced structure transfer.