Learning From Vision Foundation Models for Cross-Domain Remote Sensing Image Segmentation.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cross-domain image segmentation is vital for remote sensing.
• Existing mean-teacher models struggle with domain discrepancy and class overlap, limiting performance.
• There is a need for robust domain adaptation methods in remote sensing image segmentation.

Purpose of the Study:

• To introduce LFMDA, a novel domain adaptation method for cross-domain semantic segmentation in remote sensing.
• To leverage vision foundation models (VFMs) to enhance segmentation performance across different domains.
• To address the limitations of current methods by improving feature invariance and discriminability.

Main Methods:

• Proposing a prototypical contrastive knowledge distillation (PCD) loss to distill knowledge from a domain-generalized VFM teacher.
• Implementing a local region homogenization (LRH) strategy using the Segment Anything Model (SAM) for high-quality pseudo-label generation.
• Developing a robust domain adaptation framework (LFMDA) for remote sensing image segmentation.

Main Results:

• LFMDA significantly outperforms existing approaches in cross-domain remote sensing image segmentation.
• The method achieves state-of-the-art (SOTA) performance by producing domain-invariant and category-discriminative features.
• PCD loss and LRH strategy effectively enhance segmentation accuracy and robustness across domains.

Conclusions:

• LFMDA represents a significant advancement in domain-adaptive remote sensing image segmentation.
• Leveraging VFMs with novel distillation and pseudo-labeling strategies offers a powerful approach to overcome domain shift challenges.
• The proposed method sets a new benchmark for cross-domain segmentation tasks in remote sensing.