CDAFormer: Hybrid Transformer-based contrastive domain adaptation framework for unsupervised hyperspectral change...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) change detection is crucial for monitoring environmental and urban changes.
• Current deep learning methods struggle with generalization to new HSI datasets due to distribution shifts.
• Acquiring labeled HSI data for training is costly and time-consuming.

Purpose of the Study:

• To develop an unsupervised method for HSI change detection that overcomes generalization issues.
• To reduce the reliance on expensive, manually annotated HSI datasets.
• To improve the performance of HSI change detection by leveraging domain adaptation techniques.

Main Methods:

• A hybrid Transformer-based contrastive domain adaptation (CDAFormer) framework is proposed.
• The method aligns changed and unchanged difference features between source and target domains.
• Contrastive learning at the feature level refines domain alignment, ensuring feature separability.

Main Results:

• CDAFormer demonstrates superior performance on widely used HSI datasets.
• The method effectively bridges domain discrepancies without requiring labeled training samples.
• Achieved improved detection accuracy compared to existing state-of-the-art techniques.

Conclusions:

• The proposed CDAFormer framework offers a robust solution for unsupervised HSI change detection.
• Effective domain adaptation significantly enhances model generalization across diverse HSI data.
• This approach reduces the burden of data annotation while maintaining high detection performance.