Reducing the computational cost of process-based flood frequency estimation by extracting precipitation events from a large-ensemble climate dataset

The generation of flood projection ensembles for large areas and at a high resolution has been a long-standing computational challenge. Previous studies focused on improving model efficiency and hardware acceleration. An intriguing question arises; is it possible to reduce the computational demand for obtaining specific flood characteristics (e.g., flood frequency curves, FFCs) through precipitation data preprocessing, while maintaining high accuracy? In this study, we developed an aggregating grid event (AGE) method based on hydrological concepts to extract essential precipitation events from large-ensemble climate-change dataset. A total of 2,966 events were extracted from dynamically downscaled 720-year precipitation data covering all over Japan with the 5-km spatial resolution. By inputting the precipitation data into the Rainfall-Runoff-Inundation model at a 150 m resolution, we computed hourly discharges across all river grids in the study area, Shikoku Island, Japan. Based on the simulated peak discharges, we computed FFCs for return periods exceeding 10 years at each river grid using the peak-over-threshold method. The results demonstrated the effectiveness of the AGE method, with a relative bias (BIAS) of-1.38 % and a root mean square error (RMSE) of 36.45 m3/s for flood peaks across locations. Furthermore, the BIAS of quantiles at 100-year return period was-1.04 % compared to the references, which were estimated from all the valid 25,700 precipitation events. By using 2,966 events instead of 25,700 events, the AGE method significantly reduces the computational burden in estimating FFCs at all river grids while maintaining accuracy. This approach is applicable for any grid-based precipitation dataset, marking a crucial advancement in regional hyper-resolution flood studies based on climate projection ensembles.