Super-Resolution of NCEP Reanalysis Wind Fields to ERA5 Resolution via Conditional Diffusion Models

High-resolution near-surface wind fields are essential for analyzing atmospheric circulation, diagnosing extreme weather events, and conducting regional climate research. However, the spatial resolution of global reanalysis datasets is often limited by computational constraints, which hinders their ability to accurately represent mesoscale and fine-scale atmospheric structures. Conventional interpolation methods, such as bilinear interpolation, typically fail to restore realistic fine-scale wind variability.

To address this, we develop a conditional diffusion-based super-resolution framework to downscale NCEP wind fields from a 2° resolution to 0.25° over the East Asia region (15°–55°N, 65°–135°E). The model leverages a denoising diffusion probabilistic process with a U-Net backbone, where low-resolution wind fields serve as conditioning information to guide the generation of high-resolution wind structures. The model is trained using daily NCEP 10-m wind components from 1981 to 2015, validated during 2016–2019, and evaluated over an independent test period from 2020 to 2023.

Model performance is assessed against bilinear interpolation using multiple statistical and spectral metrics, including root-mean-square error (RMSE), spatial correlation coefficient, and isotropic power spectrum analysis. Results show that the proposed diffusion model substantially outperforms bilinear interpolation across the entire domain. The spatially averaged correlation coefficient between the reconstructed and ERA5 reference wind fields reaches approximately 0.94, while the mean RMSE is reduced to about 0.86. Furthermore, power spectrum analysis demonstrates that the diffusion-based results closely match the high-resolution ERA5 reference across a wide range of spatial scales, effectively recovering high-frequency energy that is largely absent in interpolated fields.

These results indicate that conditional diffusion models provide a promising data-driven approach for wind field downscaling, offering improved spatial realism and scale-consistent representations. The proposed framework has potential applications in high-resolution reanalysis reconstruction, regional climate analysis, and ensemble-based uncertainty quantification.