A deep learning-based probabilistic approach to flash flood warnings in mountainous catchments

Advanced warning of flash flooding plays a vital role in mitigating risk. Established warning models focus on deterministic indicators and do not consider the probability of flash flood occurrences. This study proposed a probabilistic framework based on CNN-LSTM-MultiHead-Attention (CLMA) by considering the threshold optimization for the flash flooding warnings, and compared it with a deterministic method based on hydrological modeling. In addition to the pivotal factor of rainfall, the proposed model also incorporated antecedent precipitation index (API) and rainfall pattern (RP) as input factors to explore the hidden patterns of flash flood occurrences. The results show that the proposed method gave more informative results than the deterministic method by additionally generating hourly probability values. This greatly reduced false and missing alarms and enabled warnings to be issued average 1-3 h in advance. Meanwhile, incorporating API and RP data into the CLMA model enhanced its ability to forecast flash flooding, with API having a greater effect than RP. Moreover, probabilistic threshold optimization using the Best_DIFF criterion improved the CLMA model, decreasing the Euclidean Distance (ED) from 0.12 to 0.05 and increasing the Critical Successful Index (CSI) from 0.54 to 0.83. This study demonstrates that deep learning has the potential to be applied within a probabilistic framework for the flash flood forecasting, which can effectively assess risk information and support early decision-making to issue warnings and flood control actions.