Work

Wetl-SAM2YNet: A foundation model driven framework for fine-scale wetland mapping

Year: 2026

Type: article

Abstract: Fine-scale wetland mapping is challenged by fragmented patch shapes, spectral confusion between wet and dry surfaces, and the scarcity of annotated high-resolution training samples. This study proposes Wetl-SAM2YNet, a framework based on Segment Anything Model 2 (SAM2, a vision foundation model), designed for community level wetland classification in arid regions. The model couples a frozen SAM2 Hiera backbone with a lightweight auxiliary encoder that takes multispectral bands and spectral indices sensitive to wetlands (NDWI, SAVI, DVI, RVI) as input to resolve spectral ambiguities in the visual of arid region wetlands. An Adaptive Cross-Gated Fusion (ACGF) module reconciles generic visual priors with non-visual wetland features, combining the two feature types at multiple scales as the model processes wetland landscapes. Gaussian weighted sliding window inference preserves the continuity of narrow riverside corridors and small fragmented patches. Evaluated on a GF-1/2 dataset comprising only 226 training patches, Wetl-SAM2YNet achieves an overall accuracy of 84.89%, an IoU of 74.40%, and a Kappa coefficient of 84.37%, substantially outperforming CNN and Transformer baselines. The resulting 2 m resolution map resolves seven vegetation communities along hydrological and salinity gradients, providing finer thematic and spatial detail than the global wetland products evaluated in this study. These results demonstrate effective, data efficient, community-level mapping of fragmented wetlands under limited training samples, supporting ecological assessment and conservation planning along hydrological and salinity gradients.

Source: Ecological Informatics

DOI: 10.1016/j.ecoinf.2026.104119