A Dual-Branch Deep Learning Framework for Wetland Mapping Using Sentinel-1 and Sentinel-2

Wetlands are among the most ecologically significant yet threatened ecosystems, providing critical services such as carbon storage, water filtration, and biodiversity support. However, their mapping remains challenging due to their complex and heterogeneous nature. This study introduces a novel deeplearning (DL) framework for accurate wetland mapping by integrating convolutional neural networks (CNNs) and vision transformers (ViT). The model uses local features extracted through CNN blocks and global features extracted by ViT's self-attention mechanism, enhanced by spatial attention (SA) and channel attention (CA) modules. Using Sentinel-1 (S1) and Sentinel-2 (S2) data, the framework achieves high classification accuracy across diverse wetland types, including bogs, fens, marshes, swamps, and other classes. Experimental results demonstrate strong performance, with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex>-scores of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 9 0}$</tex> for bogs and 0.89 for swamps, highlighting the model's ability to distinguish spectrally similar classes. The gradient-weighted class activation mapping (Grad-CAM) analysis further shows the complementary roles of local and global features in capturing fine-grained details and large-scale contextual information.