Deep Learning-Based Precipitation Nowcasting for Operational and Flash-Flood Applications

Accurate short-term precipitation nowcasting is crucial for disaster risk reduction, flash-flood early warning, and water resource management. Conventional nowcasting approaches, such as extrapolation-based radar methods or numerical weather prediction models, often struggle to capture the nonlinear evolution of convective systems and are computationally demanding for rapid updates at high spatial and temporal resolution. The ability to provide reliable high-resolution forecasts at lead times of minutes to hours is particularly important for mitigating the societal and economic impacts of intense rainfall events. Recent developments in deep learning (DL), in combination with high-resolution radar observations, represent a compelling alternative for improving short-term precipitation forecasting. Radar-based precipitation data are particularly well suited for nowcasting applications due to their fine spatio-temporal resolution and ability to capture the dynamic structure and movement of precipitation systems. In this study, we develop and evaluate an operationally oriented DL framework for precipitation nowcasting that integrates multi-source data including high-resolution radar and satellite observations and automatic weather station measurements via the qPrec system over Slovakia. By incorporating satellite-derived forcing, the framework accounts for convection initiation and cloud development stage, providing a physical advantage over both classical extrapolation and radar-only deep learning methods. The framework leverages modern DL architectures, including convolutional encoder-decoder models such as U-Net and spatio-temporal transformer-based models (e.g., Earthformer), to learn the temporal evolution of precipitation fields inputs. The use of transformer-based models allows the network to capture long-range spatial dependencies and complex motion patterns that traditional CNNs may miss.

The proposed models generate precipitation forecasts at a spatial resolution of 1 km and a temporal resolution of 5 minutes, with lead times of up to 60 minutes. In addition to instantaneous precipitation estimates, the framework produces 15-minute accumulated precipitation for horizons up to 120 minutes. Unlike traditional methods where predictability skill remains static across resolutions, our DL approach leverages varied spatial representations to enhance predictability at these coarser temporal scales, optimizing the forecast for different hydrological requirements. These accumulated fields can be directly applied to flash-flood hazard assessment, enabling estimation of flood likelihood as a function of rainfall intensity and duration. Model performance is evaluated using standard verification metrics such as the Fractions Skill Score, and continuous ranked probability score (reducing to MAE on deterministic outputs), showing improvement over conventional radar extrapolation methods. This study demonstrates that modern DL approaches, particularly when combined with high-resolution radar observations, offer a promising path toward next-generation operational nowcasting.