Diffusion model based downscaling of extreme precipitation in southern Europe

Climate models produce enormous amounts of atmospheric data. However, these models often have very large spatial resolution, making hazard-scale, e.g. an individual city or catchment, forecasts based on future climate data impossible. Diffusion models (DMs) are a class of deep-learning generative models that can rapidly produce ensemble-like realisations of high-resolution weather states, allowing for uncertainty quantification. Numerous studies have demonstrated the efficacy of these models in faithfully downscaling weather variables from both observational datasets and from global climate models to regional climate models. However, little is known about how well DMs can perform when trained and evaluated on heterogeneous and multi-source datasets, and even less regarding their ability to faithfully emulate high-resolution extreme rainfall events. To evaluate this, we train a DM to emulate 0.1° by 0.1° hourly precipitation data from IMERG (satellite-based), using hourly 1° by 1° atmospheric fields from ERA5 (reanalysis) as the model’s input. We are also performing an out-of-distribution experiment in which extreme events are excluded from the DM’s training data in order to investigate to what extent it can accurately extrapolate to severe weather. Our domain is centred in southern Europe and was chosen to cover many diverse regions, including the Alps, Mediterranean Ocean, and northern Africa. According to continuous rank probability score, power spectral density, histograms and many other metrics, after training on balanced data our DM accurately downscales precipitation across all rainfall intensity levels, preserves fine-scale spatial structures, learns regional precipitation dynamics, and captures extreme events in the tails of the distribution. Our DM also outperforms a strong climatological baseline, and it is superior to other commonly used models such as a deterministic deep convolutional network, which tends to over-smooth and underestimate extreme events. Our results affirm the ability of diffusion models to generate robust, hazard-relevant rainfall realisations using coarse atmospheric data.