Towards Accurate Flood Predictions in Small, Fast-Responding Catchments Using Hourly CAMELS-DE Data

Long Short-Term Memory (LSTM) networks have recently emerged as powerful data-driven approaches for rainfall-runoff modeling, often outperforming traditional hydrological models. However, their application has been predominantly tested on daily time steps and larger catchments (>250 km²). In this study, we push these boundaries by investigating the potential of LSTMs for flash flood prediction in smaller, fast-responding catchments. We leverage a refined version of the CAMELS-DE dataset, processed at hourly resolution, to capture the rapid hydrological dynamics that typify flash flood events. Hourly discharge and water level observations from federal agencies in Germany are combined with meteorological inputs from the German Weather Service (DWD), enabling a detailed assessment of the benefits of refined temporal resolution for LSTM-based modeling. 

Our findings reveal that while LSTMs demonstrate reasonable skill in predicting peak discharges and event timing, performance degrades significantly during summer convective storms, characterized by localized and intense rainfall. We investigate whether this drop in performance is related to limitations in the LSTM architecture and training strategy or is due to increasing uncertainties in the meteorological boundary conditions. We further investigate when, where and how the use of hourly resolution data affects model performance. The study provides critical insights into the challenges and opportunities of using data-driven approaches for flash flood forecasting in small, fast-responding catchments, contributing to the development of more robust hydrological prediction systems. In addition, we present a preliminary version of CAMELS-DE in hourly resolution, opening new possibilities for research in the field of large sample hydrology.