HiFA-Diff: A High-Frequency-Aware Diffusion Framework for Meteorological Downscaling to 1-km Resolution

Abstract

High-resolution meteorological fields at the kilometer scale are essential for urban forecasting, renewable energy scheduling, and climate resilience planning. However, publicly available datasets such as ERA5 remain limited to coarse resolutions, making it difficult to reconstruct fine-scale atmospheric patterns such as wind convergence zones, coastal gradients, and urban heat islands. Conventional statistical and deep learning-based downscaling approaches often oversmooth these high-frequency components, especially when targeting ultrafine resolutions. To address this limitation, we propose high-frequency-aware diffusion framework (HiFA-diff), a high-frequency-aware deep learning framework for meteorological downscaling to 1-km resolution. HiFA-diff employs a two-stage coarse-to-fine architecture, where a UNet-based network first generates preliminary high-resolution estimates from interpolated ERA5 inputs, followed by a conditional diffusion model that explicitly learns and refines the high-frequency residuals in the spectral domain. This design decouples large-scale background structures from fine-scale fluctuations, leading to sharper and more realistic reconstructions of mesoscale weather features. Moreover, topography and geolocation information are incorporated to enhance terrain-sensitive predictions. Experiments on an ERA5-to-weather research and forecasting (WRF) benchmark over the Yangtze River Delta (YRD) demonstrate that HiFA-diff substantially outperforms conventional baselines. Both quantitative results and visual analyses confirm the ability of the framework to recover intricate wind structures and sharp spatial gradients, underscoring its potential for operational high-resolution weather modeling and forecast enhancement.