A Graph‐Based Deep Learning Approach for Daily Flash Flood Susceptibility Modeling in China

Abstract Flash floods are sudden flood events triggered by intense rainfall, and often exacerbated by mountainous terrain that accelerates surface runoff. To support disaster mitigation and management, deep learning (DL) models have been widely applied to flash flood susceptibility (FFS) modeling. However, traditional deep learning (DL) models overlook both the intra‐annual temporal variations and the spatial interactions between nearby catchments. To address these, this study proposes a graph‐based DL model (named LTG model) for spatiotemporal FFS simulation at a daily scale in China considering catchment topology. The proposed LTG model mainly integrates three components: Long Short‐Term Memory Networks, Temporal Graph Convolutional Networks, and Graph Convolutional Networks aiming to capture temporal dependencies, spatio‐temporal interactions, and spatial dependencies between catchments, respectively. This enables the LTG model to perform spatiotemporal dynamic FFS simulation as well as incorporating catchment topology information. We demonstrated the proposed LTG model in China and found that the proposed model outperforms the baseline models with the highest Area Under the ROC Curve (AUC) of 0.911 and Critical Success Index of 0.719. Compared to yearly‐scale modeling, the daily‐scale simulation generated by LTG model exhibits a higher ability to capture seasonal variations, with a significant intra‐annual standard deviation of 0.263. By conducting a detailed analysis of FFS changes along river networks and establishing biased rainfall scenarios, we found that the proposed model not only considering the spatial clustering along river networks, and upstream–downstream dependence, but also enhances its inferential ability by leveraging information from nearby catchments.