Latent diffusion model for quantitative precipitation estimation and forecast at km scale

Abstract

Accurate high-resolution precipitation estimation remains a significant challenge in weather prediction due to computational limitations and sub-grid process parameterization difficulties. We present a latent diffusion modeling (LDM) framework that estimates 4 km resolution precipitation using 25 km resolution atmospheric and topographic inputs. The LDM transforms precipitation data into a compact Quasi-Gaussian latent space and progressively refines predictions through neural network-guided diffusion, effectively avoiding common deep learning issues such as mode collapse and blurry artifacts. Compared to traditional numerical models and other deep learning approaches, LDM achieves superior performance with over 30 % reduction in root mean squared error and 40 % improvement in critical success index for extreme events. For the extreme precipitation event (>300 mm/d) in California on October 25, 2021, LDM maintained effective 7-day forecast skill using circulation predictions from a data-driven weather forecasting model. The framework demonstrates significant potential for operational weather prediction applications.