Abstract
High-resolution sea surface wind fields are essential for marine meteorology and offshore wind energy development. While traditional statistical and dynamical downscaling methods suffer from limitations in accuracy or computational cost, deep learning-based super-resolution methods provide a promising alternative. However, many existing models prioritize global-average accuracy, which limits their performance for extreme wind speeds. Based on paired low- and high-resolution sea surface wind fields in the South China Sea derived from 12-km and 4-km two-way nested Weather Research and Forecasting (WRF) simulations, this study proposes a SwinIR-based downscaling framework in which coarse-resolution zonal and meridional wind components are used as inputs to reconstruct their high-resolution counterparts. The framework leverages shifted-window self-attention to capture long-range spatial dependencies and incorporates a two-stage training strategy to better handle the long-tailed distribution of wind speeds. In Stage 1, the model learns a general low-to-high-resolution mapping using a mean squared error (MSE) loss, whereas in Stage 2, a weighted loss is introduced to improve reconstruction in high-wind regimes. On an independent test set, the proposed method achieves a mean absolute error (MAE) of 0.11 m/s and a root mean square error (RMSE) of 0.21 m/s, outperforming bilinear interpolation, Efficient Sub-Pixel Convolutional Neural Network (ESPCN), and DeepSD. Ablation experiments show that the two-stage strategy reduces RMSE by 4.25% for wind speeds exceeding 25 m/s, and a case study of Tropical Storm Wipha (2019) further demonstrates its capability in reconstructing extreme wind fields.