Domain-Adaptive Mamba for Cross-Scene Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cross-scene hyperspectral image classification faces challenges with limited target domain samples.
• Existing methods often align global features, neglecting fine-grained details and efficient modeling.
• Transformer architectures, while effective for dependencies, suffer from quadratic complexity issues.

Purpose of the Study:

• To propose a novel domain-adaptive Mamba (DAMamba) for improved cross-scene hyperspectral image classification.
• To address the limitations of existing methods in fine-grained feature alignment and computational efficiency.
• To enhance classification accuracy and reduce processing time in unsupervised domain adaptation scenarios.

Main Methods:

• Developed a spectral-spatial Mamba for high-order semantic feature extraction.
• Introduced a domain-invariant prototype alignment method (intra-domain, inter-domain, mini-batch) for pseudo-label generation and spectral shift mitigation.
• Utilized a fully connected layer for final classification on aligned target domain features.

Main Results:

• DAMamba demonstrated superior classification accuracy across diverse cross-scene datasets.
• The proposed method significantly improved computational efficiency compared to existing approaches.
• Effective mitigation of spectral shift and reliable pseudo-label generation were achieved.

Conclusions:

• DAMamba offers a more efficient and accurate solution for cross-scene hyperspectral image classification.
• The fine-grained alignment and spectral shift mitigation strategies are crucial for domain adaptation.
• The proposed method advances the state-of-the-art in unsupervised domain adaptation for hyperspectral imaging.