Uncertainty-aware methods for enhancing rainfall prediction with deep-learning based post-processing segmentation

Precipitation forecast is critical in flood management, agricultural planning, water resource allocation, and weather warnings. Despite significant advancements in Numerical Weather Prediction (NWP) models, these systems often exhibit substantial biases and errors, particularly at high spatial and temporal resolutions. To address these limitations, we develop and evaluate uncertainty-aware deep learning ensemble architectures, focusing on characterizing forecast uncertainties while achieving high accuracy and an optimal balance between sharpness and reliability. This study presents SDE U-Net, a novel adaptation of SDE-Net designed specifically for segmentation tasks in precipitation forecasting. We conduct a comprehensive evaluation of state-of-the-art ensemble architectures, including SDE U-Net, and compare their forecast uncertainty against that of a Poor Man’s Ensemble (PME, i.e. NWPs forecast average) across diverse meteorological conditions, ranging from non-intense precipitation patterns to intense weather events. As an example case, we focus on predicting daily cumulative precipitation in northwest Italy, though our approach is broadly generalizable. Our findings demonstrate that all the evaluated probabilistic deep learning models outperform the PME benchmark in terms of median RMSE for both non-intense and intense precipitation events. Among them, SDE U-Net achieves the best overall performance, delivering the lowest RMSE for intense events ($2.637\times 10^{-2}$) and demonstrating a more stable error distribution compared to other models. For non-intense events, SDE U-Net perform comparably to other deep learning models, still notably surpassing the baselines. Moreover, SDE U-Net effectively balances sharpness and reliability, making it particularly suitable for operational forecasting of extreme weather. Integrating uncertainty-aware models like SDE U-Net into forecasting workflows can enhance decision-making and preparedness for weather-related hazards.