PMA-SAM2: Prior-Memory Adapted Framework for Automatic Remote Sensing Water Body Extraction

SAM2 has great potential for remote sensing water body extraction due to its strong feature representation capability. However, when directly applied to the water body segmentation task, SAM2 still faces several challenges, including domain discrepancy, spectral confusion between water bodies and shadow-like objects, significant intra-class variations, and dependence on manual prompts. To address these challenges, this paper proposes the PMA-SAM2, which integrates domain prior injection, prototype memory calibration, and automatic prompt guidance for robust water body extraction. First, a Multi-level Prior Adaptation Module is designed to inject spatial geometric information and texture features into the SAM2 encoder, enabling effective alignment between general visual representations and remote sensing domain features. Then, a Prototype Memory Calibration Module is introduced to dynamically accumulate water feature prototypes under different geographical environments and imaging conditions, improving the model’s robustness to intra-class variations and inter-class spectral ambiguity. Furthermore, an Automatic Prompt Decoder is proposed to transform prototype-retrieved similarity maps into dense prompts, replacing manual prompts and enabling fully automatic water body segmentation. Experimental results on four datasets demonstrate that the proposed method achieves superior water body extraction accuracy compared with mainstream segmentation methods, verifying its strong generalization capability and scene adaptability.