Domain-Adaptive Segment Anything Model for Cross-Domain Water Body Segmentation in Satellite Imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Environmental Monitoring

Background:

• Accurate surface water body monitoring is vital for environmental protection and resource management.
• Current satellite image segmentation techniques often lack generalization across different satellite domains.
• Domain adaptation is crucial for robust water body segmentation.

Purpose of the Study:

• To introduce DASAM, a domain-adaptive Segment Anything Model (SAM) for cross-domain water body segmentation.
• To enhance the generalization capabilities of segmentation models for diverse satellite imagery.
• To improve the accuracy and robustness of water body detection in environmental analysis.

Main Methods:

• DASAM utilizes a contrastive learning module to align image features, enabling domain generalization without target-domain annotations.
• A prompt-enhanced module and encoder adapter are integrated to capture fine-grained spatial details and global context.
• The model is evaluated using experiments on the China GF-2 dataset and cross-domain evaluations on GLH-water and Sentinel-2 datasets.

Main Results:

• DASAM demonstrated superior performance compared to existing methods on the China GF-2 dataset.
• Cross-domain evaluations confirmed DASAM's strong generalization and robustness on GLH-water and Sentinel-2 datasets.
• The proposed method effectively addresses the challenge of limited generalization in cross-domain water body segmentation.

Conclusions:

• DASAM offers a robust solution for cross-domain water body segmentation in satellite imagery.
• The model's domain-adaptive approach enhances its applicability to large-scale and diverse environmental monitoring tasks.
• DASAM shows significant potential for improving the accuracy of environmental analysis through satellite data.