Cross-Scene Hyperspectral Image Classification via Bidirectional Mamba and Domain Mixing Network.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Domain shift presents a significant challenge in hyperspectral image (HSI) classification.
• Existing domain adaptation (DA) methods struggle with large domain shifts by focusing on feature space alignment.
• Mapping disparate source and target domains into a shared feature space remains difficult.

Purpose of the Study:

• To develop an effective method for hyperspectral image classification under domain shift.
• To propose a novel network architecture that efficiently captures both local and global features.
• To introduce a domain mixing strategy to bridge the gap between source and target domains.

Main Methods:

• A bidirectional mamba module (BMM) is proposed for efficient long-range dependency capture, addressing limitations of CNNs and Transformers.
• A self-distillation strategy is employed using a stable teacher model for reliable target domain predictions.
• A domain mixing supervised learning (DMSL) module creates a mixed domain to reduce the inter-domain gap in the data space.

Main Results:

• The proposed BMDMnet demonstrates superior performance compared to state-of-the-art algorithms.
• Experiments were conducted across three cross-scene datasets, validating the method's effectiveness.
• The integration of BMM and DMSL significantly improves HSI classification accuracy under domain shift.

Conclusions:

• The BMDMnet offers an efficient and effective solution for hyperspectral image classification with domain shift.
• The proposed BMM and DMSL modules successfully address the limitations of existing domain adaptation techniques.
• This work advances the field of HSI classification by providing a robust method for handling domain variability.