[A-HW29] Climate, Rivers, and Floods: Exploring Hydro-Geomorphological Interactions

[AHW29-P02] A graph-based hybrid DL model for flash flood susceptibility simulation

*Jun Liu^1,2, Gang Zhao¹ (1.Institute of Science Tokyo, 2.Sun Yat-sen University)

Keywords:Flash flood susceptibility, Spatiotemporal simulation, Deep learning models, Graph neural network

Flash floods are sudden flood events triggered by intense rainfall and rapid runoff generation in mountainous regions. To support disaster mitigation and prevention, deep learning (DL) models have been widely applied to flash flood susceptibility (FFS) simulation. However, conventional DL models fail to account for inter-catchment interactions and the temporal dynamics in FFS modeling. To address this limitation, this study proposes a graph-based hybrid model (named LTG model) for spatiotemporal FFS modelling at a daily scale in China. A graph structure was introduced to reconstruct the topological connectivity between catchments, serving as the embedding of physical mechanisms. Based on this graph structure, the proposed LTG model integrates long short-term memory (LSTM), temporal graph convolutional network (TGCN), and graph convolutional network (GCN) to enable spatiotemporal dynamic convolution. This enables the LTG model to accurately simulate FFS while also incorporating catchment topology information. We demonstrated the proposed model in China and found that the proposed LTG model outperforms three baseline DL models. The FFS simulated by the proposed model accurately identifies locations prone to flash flood occurrence on a daily scale. Overall, the proposed hybrid model not only enhances the accuracy of spatiotemporal FFS simulation but also effectively captures the spatial dependencies of flash floods. This provides a novel insight for incorporating physical mechanisms and temporal dynamics into DL models for FFS modelling.