WindHCT: context and local hybrid CNN Transformer for wind speed super-resolution

Abstract

Global climate models operate at coarse spatial resolutions that limit their ability to represent localized atmospheric phenomena, motivating deep learning–based super-resolution (SR) for meteorological fields. However, most existing SR models are designed for natural images and fail to fully capture the characteristics of wind-speed fields. This study proposes WindHCT, a hybrid CNN–Transformer SR model that jointly learns the local variability and global spatial dependencies of wind-speed fields. WindHCT comprises two parallel branches—a convolutional neural network (CNN)-based local branch and a Transformer-based context branch—whose representations are integrated through an adaptive fusion module, with a coordinate attention mechanism enhancing directional information. Experiments were conducted under a 5 × upscaling setting with temporally disjoint splits on two datasets: a WIND Toolkit-based dataset and an ERA5–CERRA-based dataset. WindHCT achieves consistent and, in most comparisons, statistically significant improvements over representative SR models, including EDSR, RCAN, SwinIR, HAT, and PFT, in terms of root mean square error (RMSE), MAE, PSNR, and SSIM. Physically oriented evaluations further suggest that the proposed model improves several aspects of wind-field reconstruction beyond pixel-wise accuracy. The WindHCT implementation is publicly available at https://github.com/Jowonji/WindHCT.