Probabilistic Postprocessing of Hourly Precipitation Ensemble Forecasts Using UNet

Title: Probabilistic Postprocessing of Hourly Precipitation Ensemble Forecasts Using UNet 

Authors: Marcos Esquivel-González, Albano González, Juan Carlos Pérez, Juan Pedro Díaz, Pierre Simon Tondreau

Affiliation of authors: Grupo de Observación de la Tierra y la Atmósfera (GOTA), Avenida Astrofísico Francisco Sánchez, s/n, La Laguna, 38200, Canary Islands, Spain

Abstract: Reliable precipitation forecasting is crucial in sectors like public safety, agriculture and water management. Numerical Weather Prediction (NWP) models, which form the backbone of modern forecasting, are prone to errors due to their limitations and the chaotic behavior of equations, requiring postprocessing to improve accuracy and quantify uncertainties. Thus, this study evaluates probabilistic postprocessing models tailored for the Canary Islands, with the aim of enhancing Weather Research and Forecasting (WRF) ensemble forecasting accuracy in hourly precipitation forecast. UNet-based models were explored using two approaches,  one incorporating  the full set of km-scale convection-permitting ensemble forecast simulations (25) and another applying dimensionality reduction via Principal Component Analysis (PCA) and feature selection methods. These models were compared to traditional benchmarks like the Censored Shifted Gamma Distribution (CSGD) with Ensemble Model Output Statistics (EMOS) and the Analog Ensemble method. 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.

In general, UNet models outperformed traditional approaches. The UNet with PCA excelled in probabilistic and deterministic metrics but struggled in regions without weather station data. Conversely, the UNet with feature selection, while slightly less accurate overall station locations, showed better generalization to unseen locations, maintaining consistent performance across the region and reducing computational demand. Additionally, the Integrated Gradients technique, an interpretability method that quantifies the contribution of each input feature to a model’s predictions by analyzing gradients, was employed to evaluate the impact of input variables on model performance. This analysis revealed that the integration of digital terrain elevation data significantly contributed to the UNet's outputs, underscoring the importance of topographic data in rainfall prediction.