FDGNet: Frequency Disentanglement and Data Geometry for Domain Generalization in Cross-Scene Hyperspectral Image...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cross-scene hyperspectral image classification (HSIC) faces challenges due to domain shifts.
• Existing domain adaptation (DA) methods require target data, limiting practical application.
• Current domain generalization (DG) methods often compromise semantic information or generate unrealistic samples.

Purpose of the Study:

• To propose a novel domain generalization network (FDGNet) for cross-scene HSIC.
• To address limitations of existing DG methods, such as semantic compromise and unrealistic sample generation.
• To improve the generalization capability of hyperspectral image classification models to unseen domains.

Main Methods:

• Developed a spectral-spatial encoder with frequency disentanglement (FDSS encoder) to maintain semantic consistency while simulating inter-domain gaps.
• Incorporated data geometry into adversarial training to diversify new domains realistically.
• Proposed an intermediate domain sampling strategy using class-wise perceptual manifolds to synthesize reliable intermediate domains.

Main Results:

• FDGNet demonstrated superior performance in cross-scene HSIC tasks.
• The proposed methods effectively preserved semantic consistency and generated realistic domain variations.
• The class-wise perceptual manifold strategy enhanced the learning of domain-invariant representations.

Conclusions:

• FDGNet offers a robust solution for domain generalization in HSIC.
• The integration of frequency disentanglement and data geometry is effective for handling inter-domain variations.
• The approach successfully generalizes to unseen domains without requiring target data during training.