Large-scale flash flood warning in China using deep learning

Flash flood warning (FFW) systems play a fundamental role in flood hazard prevention and mitigation. In this study, we propose the first deep learning-based approach for large-scale FFW and demonstrate the application of this approach to mountainous and hilly areas of China. Specifically, the time series of precipitation before flash floods and three spatial features (maximum daily precipitation, curve number, and slope) are selected as predictors. A long short-term memory (LSTM)-based approach is adopted to predict the occurrence of flash floods, and we compare this approach with two widely used FFW methods, namely the rainfall triggering index (RTI) and flash flood guidance (FFG). The results demonstrate the following: (1) The LSTM-based approach provided a reliable FFW 1 day ahead with a hit rate (HR) of 0.84 and false alarm rate (FAR) of 0.09. It demonstrated moderate warning performance 2 days before flash floods, with an HR of 0.66 and FAR of 0.21. (2) The LSTM-based approach outperformed the benchmark RTI and FFG methods, achieving the highest critical success index (CSI) of 0.77. The FFG also provided satisfactory performance, with a CSI of 0.71, and the RTI demonstrated the lowest performance (CSI = 0.68). (3) The LSTM-based approach provides better results (CSI = 0.75) than RTI (CSI = 0.68) when only the time series of precipitation is used for prediction. The performance of the LSTM-based approach can be improved by considering the spatial features and a long time series of precipitation during model development. (4) The proposed approach did not exacerbate the effect of precipitation uncertainty on the flash flood warning; and we suggest using ensemble results for FFW to reduce the uncertainty caused by small or unbalanced learning samples. We conclude that the proposed approach is a valid method for large-scale FFW without using commercially sensitive observations, and can improve the capabilities of flood disaster mitigation, particularly in ungauged areas.