Abstract
Accurate mapping of small wetlands remains challenging due to their limited spatial extent, fragmented boundaries, and spectral similarity to surrounding land-cover types. To address these challenges, this study proposes RSF-HG-BDResUNet, a deep learning network based on asymmetric optical–synthetic aperture radar (SAR) feature fusion for object-level small wetland mapping in the Dianchi Lake Basin, China. To account for differences between optical and SAR information, the proposed network introduces a learnable residual gating mechanism that selectively incorporates SAR features into a jointly encoded feature representation. This asymmetric fusion strategy is combined with lightweight multiscale contextual modeling, selective skip connections, and boundary-aware supervision to improve the identification of small and irregular water bodies. Pixel-level segmentation results are subsequently processed using eight-neighborhood connected-component analysis and area filtering to delineate small wetland patches smaller than 8 ha. The proposed model was evaluated against U-Net, ResUNet, DeepLabV3+, PSPNet, and SegFormer-B0 using both pixel-level and object-level metrics. It achieved a pixel-level F1-score of 0.852 ± 0.003, a mean intersection over union of 0.871 ± 0.002, and an object-level F1-score of 0.780 ± 0.002, outperforming the comparison models on these metrics. Ablation experiments further demonstrated the contributions of residual SAR-aware fusion and the associated feature representation modules. The results indicate that asymmetric optical–SAR feature fusion can improve water-body segmentation while enhancing the detection completeness of small wetland patches. The proposed approach provides a practical framework for fine-scale wetland mapping and regional wetland resource inventories.