SAMWater: a prompt-guided multimodal framework for transferable water mapping with the segment anything model

Accurate and transferable surface water mapping is critical for water resource assessment, flood monitoring, and aquatic ecosystem management, but remains difficult because SAR speckle, optical contamination, heterogeneous backgrounds, and scarce high-quality pixel-level annotations limit robust large-scale mapping. We propose SAMWater, a prompt-guided multimodal framework that integrates Sentinel-1 SAR, Sentinel-2 optical imagery, the Segment Anything Model (SAM), and Segment Anything Model 3 (SAM 3). Instead of using SAR and optical data as independent classifiers, SAMWater transforms their complementary information into structured prompts for coarse-to-refined water segmentation. A physically interpretable multisensor coarse water prior is first generated from Sentinel observations using an automatic threshold-based strategy. This prior is then refined through prompt-based boundary refinement, a multi-prompt excavation strategy that derives point, box, and Gaussian-style mask prompts, and Water-IoU fine-tuning for water-specific candidate ranking. Using globally distributed samples and manually interpreted PlanetScope reference labels, SAMWater achieved average Kappa coefficients of 90.71%, 85.63%, and 89.05% in evaluations stratified by water body type, land cover background, and climatic zone, respectively. It generally outperformed traditional thresholding and clustering methods, public deep learning models, and existing global water or land cover products, particularly in structurally complex river, reservoir, urban, agricultural, and mountainous scenes. Ablation experiments confirmed that multimodal prior fusion, structured prompt construction, and water-specific ranking jointly drive the performance gains. Moreover, SAMWater-generated refined labels improved the training of U-Net, DeepLabV3+, and SegFormer, indicating their utility for weakly supervised segmentation. SAMWater therefore demonstrates that integrating physically interpretable remote sensing priors with promptable foundation models provides a practical pathway for transferable surface water mapping, large-scale multisensor weak-label generation, and more consistent monitoring of surface water extent.