Spatial-Gating Dual-Branch Feature-Learning Network for Marine Environmental-Parameter Downscaling

High-resolution marine environmental fields are essential for ocean monitoring, forecasting, and coastal applications, but the spatial resolution of reanalysis products may be insufficient for representing localized gradients and fine-scale structures. This study proposes a spatial-gating dual-branch feature downscaling network (SG-DFDS) for fixed-scale marine environmental-parameter downscaling. Parallel large- and small-kernel pathways are designed to capture spatial information at different receptive-field scales, while a spatial gating unit adaptively modulates feature responses in the large-kernel branch. Experiments were conducted using hourly ERA5 fields from January 2020 to December 2021 over the seas surrounding China for wind speed (WS), mean wave direction (MWD), mean wave period (MWP), and significant wave height (SWH) at 2× and 4× scale factors. SG-DFDS achieved the best overall PSNR, SSIM, and MAE among the evaluated interpolation- and learning-based methods. At 4×, its PSNR improvements over the strongest comparison models ranged from approximately 0.20 to 1.09 dB across the four parameter tasks. Ablation experiments further verified the contributions of both the dual-branch feature structure and the spatial gating unit.