Postprocessing of convection permitting precipitation forecast using UNets

Reliable precipitation forecasting is crucial in sectors like public safety, agriculture and water management. Although Numerical Weather Prediction (NWP) models form the backbone of modern forecasting, their inherent limitations and the chaotic behavior of atmospheric equations often result in errors, requiring postprocessing to improve accuracy and quantify uncertainties. This study assesses hourly precipitation probabilistic postprocessing models tailored for the Canary Islands, aiming to improve ensemble forecasting accuracy. UNet-based models were explored using two strategies: one incorporating all 25 km-scale convection-permitting ensemble forecast simulations as input, and another applying dimensionality reduction techniques to reduce input complexity. These models were compared to traditional benchmarks like the Censored Shifted Gamma Distribution (CSGD) with Ensemble Model Output Statistics (EMOS), the Analog Ensemble method and deep learning strategies through MLP architectures. In the analysis of the results, not only the reliability of the predictions for the set of available meteorological stations was considered, but also the generalization capacity of the UNet models to obtain precipitation predictions for the whole region. UNet models generally improved the raw WRF ensemble and performed comparably or better than traditional approaches. Probabilistic (CRPSS, BSS) and deterministic (MAESS, RMSESS, Relative Bias) skill scores demonstrated improvements, particularly in forecasting light-to-moderate rainfall. The Integrated Gradients technique revealed that incorporating a height terrain map significantly influenced UNet’s outputs, emphasizing the critical role of topographic data in rainfall forecasting. Furthermore, UNet models demonstrated strong spatial generalization, showcasing their potential for operational forecasting in areas with limited observational networks. • UNet postprocessing enhance probabilistic and deterministic precipitation forecasts. • Spatial generalization is achieved, leveraging convolutional encoder–decoder structure. • Feature selection reduces computational costs while maintaining predictive skill. • High sensitivity of UNet models to the terrain height.