Downscaling Extreme Precipitation with Wasserstein Regularized Diffusion (2410.00381v2)

Abstract: Understanding the risks posed by extreme rainfall events necessitates both high-resolution products (to assess localized hazards) and extensive historical records (to capture rare occurrences). Radar and mesonet networks provide kilometer-scale precipitation fields, but with limited historical records and geographical coverage. Conversely, global gauge and blended products span decades, yet their coarse 30-50 km grids obscure local extremes. This work introduces Wasserstein Regularized Diffusion (WassDiff), a generative downscaling framework that integrates diffusion modeling with a distribution-matching (Wasserstein) regularizer, suppressing bias throughout the entire generative denoising process. Conditioned on 55 km CPC gauge-based precipitation and the 31 km ERA5 reanalysis, WassDiff generates 1 km precipitation estimates that remain well-calibrated to targets across the full intensity range, including the extremes. Comprehensive evaluations demonstrate that WassDiff outperforms existing state-of-the-art downscaling methods, delivering lower reconstruction error and reduced bias. Case studies further demonstrate its ability to reproduce realistic fine-scale structures and accurate peak intensities from extreme weather phenomena, such as tropical storms and cold fronts. By unlocking decades of high-resolution rainfall information from globally available coarse records, WassDiff offers a practical pathway toward more accurate flood-risk assessments and climate-adaptation planning.