Wetland mapping from sparse annotations with satellite image time series and temporal-aware segment anything model

Accurate wetland mapping is critical for ecosystem monitoring and management, yet acquiring dense pixel-level annotations is prohibitively costly. In practice, only sparse point labels are typically available. Existing deep learning-based models face significant challenges in capturing accurate wetland extents under such weak supervision, particularly when coupled with the strong seasonal dynamics of wetlands, which, meanwhile, makes single-date imagery insufficient and causing substantial omission and commission errors when mapping. Although powerful foundation models like the Segment Anything Model (SAM) provide promising generalization from point prompts, it is intrinsically designed for static natural images, resulting in spatially fragmented masks in heterogeneous wetland environments and cannot exploit satellite image time series. To address these challenges, we propose WetSAM, a novel SAM-based framework that effectively leverages satellite image time series to enhance wetland mapping from sparse point annotations. WetSAM adopts a dual-branch design: (1) The temporal branch extends SAM to learn temporal contexts from satellite image time series via hierarchical adapters and a dynamic temporal aggregation module. This branch equips SAM with the ability to capture and model temporal features of wetlands, allowing it to learn complex temporal patterns and phenological changes; (2) The spatial branch reconstructs distinct boundaries via a temporal-constrained region-growing strategy, iteratively expanding sparse points into reliable dense pseudo-labels; (3) A bidirectional consistency regularization enforces minimizing the discrepancy of the predictions from two segmentation heads of two branches. We validate the effectiveness of WetSAM across eight diverse global locations, each covering an area of around 5000 k m 2 and with various wetland types and geographical features. WetSAM reaches an average F1-score of 85.58%, considerably outperforming other state-of-the-art algorithms. Results demonstrate that WetSAM achieves accurate, structurally consistent segmentation from sparse labels. With minimal labeling effort, our framework shows strong generalization ability and holds promise for scalable, low-cost wetland mapping at high spatial resolutions.