Probabilistic Localized Radar-Based Nowcasting of Flood-Inducing Rainfall Events

Accurate short-term rainfall forecasts, also known as nowcasts, are essential for building effective flood early warning systems, especially for convective events in urban areas with rapid hydrologic response. In addition, developing and using probabilistic ensemble forecasts can provide decision makers and stakeholders in highly populated areas with a clearer understanding of forecast uncertainty, supporting better flood risk management and response planning.Traditional rainfall nowcasting techniques, such as extrapolation with Lagrangian persistence, are not able to predict the growth and decay of precipitation. To address these limitations, deterministic methods like RadVIL, which uses mass balance equations of the Vertically Integrated Liquid (VIL), and Spectral Prognosis (SPROG), which performs a spectral decomposition of rainfall fields and an autoregressive (AR) model, have been developed and improved over time. Methods, like SPROG-Localized (SPROG-LOC) and Autoregressive Nowcasting using the VIL (ANVIL), improve the internal evolution of rainfall fields by adding localization in the AR model and the accuracy for intense rainfall by using an autoregressive integrated (ARI) model on the VIL, respectively. Additionally, the probabilistic version of SPROG, called the Short-Term Ensemble Prediction System (STEPS), adds stochastic noise to include uncertainties in the precipitation providing an ensemble nowcast.Building on these methods, we propose Short-Term Autoregressive Nowcasting (STAN), a novel integrated approach designed to leverage the strengths of existing techniques. STAN enables the seamless combination and reproduction of previous methods, incorporating improvements such as adaptive localization, which extends the lifetime of small convective cells. This approach aims to improve nowcasting performance, particularly in cases with large, non-uniformly distributed precipitation areas and isolated convective features.While we are currently optimizing and evaluating the STAN configuration to achieve the best possible performance, we will present results with nowcasting lead times up to 2 hours of our deterministic and probabilistic approaches for different precipitation events that caused flood in Belgium. Preliminary results show that STAN in the deterministic version, which is using the VIL, an AR model and adaptive localization, shows a better Fraction Skill Score (FSS) with high precipitation thresholds (> 5 mm/hr) and Root Mean Squared Error (RMSE) than SPROG. In the probabilistic version, when stochastic noise is included, STAN shows better performance in FSS, RMSE, Equitable Threat Score (ETS) and False Alarm Ratio (FAR) than STEPS for high thresholds.