Controls on river flood generation: implications for design floods

Flood generation arises from complex, scale-dependent processes that vary across global river catchments. Understanding and predicting the controls on flood generation is key in obtaining robust estimations of flood frequency, particularly for ungauged basins. Indeed, reliable design-flood estimates are fundamental for flood risk assessment, infrastructure design as well as understanding riverine geomorphology and ecology. While recent advances have improved global flood estimation, driven largely by improved hydrological understanding and expanded data availability, key drivers of extreme floods, including compound processes and scale-dependent effects, are still insufficiently represented, and model performance has rarely been evaluated systematically across hydro-climatic regions and flood magnitudes. Here, we leverage recent advances in global river gauge records, high-resolution river hydrography, and comprehensive catchment attribute datasets in order to develop a machine learning model for estimating design floods in ungauged rivers worldwide. We generate a global design-flood dataset and conduct a systematic evaluation of model performance across hydro-climatic regions and flood magnitudes, benchmarking against existing global design-flood products. Using interpretable machine-learning techniques, we identify dominant controls on flood generation and demonstrate how basin classification can inform flood estimation from moderate to extreme events. Our results reveal strong region-specific controls on flood extremes and provide new insights for improving design-flood estimation frameworks worldwide.