S⁴DL: Shift-Sensitive Spatial-Spectral Disentangling Learning for Hyperspectral Image Unsupervised Domain Adaptation.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Unsupervised domain adaptation (UDA) is crucial for hyperspectral image (HSI) classification, leveraging labeled source data and unlabeled target data.
• HSIs offer rich spectral information but present significant domain shift challenges, often degrading performance in existing UDA methods.
• Current UDA techniques often overlook spectral domain information, focusing solely on feature distribution alignment.

Purpose of the Study:

• To propose a novel unsupervised domain adaptation method, shift-sensitive spatial-spectral disentangling learning (S4DL), for hyperspectral image classification.
• To address the limitations of existing methods in handling significant spectral domain shifts and varying domain shifts across datasets.
• To improve the robustness and accuracy of cross-scene HSI classification by learning domain-invariant features sensitive to spectral variations.

Main Methods:

• Developed gradient-guided spatial-spectral decomposition (GSSD) to separate domain-specific and domain-invariant representations using domain classification gradients.
• Introduced a shift-sensitive adaptive monitor to dynamically adjust disentangling intensity based on the magnitude of domain shift.
• Utilized a reversible neural network to preserve both semantic and shallow-level spectral domain information.

Main Results:

• S4DL demonstrated superior performance compared to state-of-the-art UDA methods on multiple cross-scene HSI datasets.
• The method showed consistent improvements, validating its effectiveness in handling significant domain shifts.
• Experimental results confirmed the advantage of disentangling spatial-spectral information and adapting to domain shift magnitudes.

Conclusions:

• The proposed S4DL approach effectively addresses the challenges posed by domain shifts in HSI classification.
• S4DL achieves state-of-the-art results by intelligently disentangling and adapting to spatial-spectral domain variations.
• The method offers a robust solution for cross-scene HSI classification, outperforming existing UDA techniques.