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

TL;DR

WetSAM addresses the challenge of accurate wetland mapping under sparse point supervision by extending the Segment Anything Model (SAM) to satellite image time series. It introduces a temporal branch with hierarchical adapters and dynamic aggregation to capture phenology and hydrological events, a spatial branch that generates dense pseudo-labels via temporal-constrained region growing, and a bidirectional consistency regularization to align their predictions. Across eight global regions (~5,000 km² each), WetSAM achieves an average F1-score of 85.58% while requiring minimal labeling, demonstrating strong generalization and boundary coherence in dynamic wetlands. The proposed framework offers a scalable, low-cost solution for high-resolution wetland mapping and motivates future work on multi-modal data fusion and parameter-efficient fine-tuning.

Abstract

Accurate wetland mapping is essential for ecosystem monitoring, yet dense pixel-level annotation is prohibitively expensive and practical applications usually rely on sparse point labels, under which existing deep learning models perform poorly, while strong seasonal and inter-annual wetland dynamics further render single-date imagery inadequate and lead to significant mapping errors; although foundation models such as SAM show promising generalization from point prompts, they are inherently designed for static images and fail to model temporal information, resulting in fragmented masks in heterogeneous wetlands. To overcome these limitations, we propose WetSAM, a SAM-based framework that integrates satellite image time series for wetland mapping from sparse point supervision through a dual-branch design, where a temporally prompted branch extends SAM with hierarchical adapters and dynamic temporal aggregation to disentangle wetland characteristics from phenological variability, and a spatial branch employs a temporally constrained region-growing strategy to generate reliable dense pseudo-labels, while a bidirectional consistency regularization jointly optimizes both branches. Extensive experiments across eight global regions of approximately 5,000 km2 each demonstrate that WetSAM substantially outperforms state-of-the-art methods, achieving an average F1-score of 85.58%, and delivering accurate and structurally consistent wetland segmentation with minimal labeling effort, highlighting its strong generalization capability and potential for scalable, low-cost, high-resolution wetland mapping.

Figures (10)

• Figure 1: Overview. We propose an end-to-end, single- stage model for wetland mapping from satellite image time series under sparse point supervision. Note the difficulty of semantic segmentation of wetlands from a single image, highlighting the need for modeling temporal dynamics and spatial contexts.
• Figure 2: The WetSAM framework. (a) is the overview of the framework. (b) is the detailed architecture of the Encoder.
• Figure 3: The temporal aggregation module.
• Figure 4: The study areas in our study, including 1) Poyang Lake, China (PL); 2) Mississippi River, United States (MR); 3) Sundarbans, India & Bangladesh (SD); 4) Sudd Wetland, South Sudan (SW); 5) Amazon, Brazil (AM); 6) Biesbosch, Netherlands (BS); 7) Pantanal, Brazil (PT); 8) Bayanbulak, China (BA). The sub-figure on the top right denotes the distributions of wetland classes in each region based on GWD30 yuan2025gwd30.
• Figure 5: The number of wetland samples of each study area.