Author(s): Yifan Li; Chendi Zhang; Shunyu Yao; Guotao Zhang; Yang Zhao

Linked Author(s): Chendi Zhang

Keywords: Disaster-inducing factors; Flash floods; Graph neural network (GNN); Regionalization; The Hengduan Mountains region (HMR)

Abstract: The Hengduan Mountains region (HMR) in China is heavily threatened by flash flood disasters, which cause severe human casualties and economic losses. Flash floods regionalization divides geographical space into homogeneous regions based on the factors that trigger these disasters and hence, offers scientific insights for flash flood prevention and mitigation efforts, especially in areas with limited monitoring activities. This research aims to explore an innovative flash flood regionalization method to understand and reveal the spatial distribution of flash flood disasters in the HMR. We gathered comprehensive input data on geomorphology, climate, meteorology, hydrology, and the underlying surface conditions for both the Graph Neural Network (GNN) and traditional clustering techniques as regionalization methods. The results exhibited the superior clustering performance of the GNN model, which effectively captured the spatial heterogeneity of flash flood disasters. The Shapley Additive exPlanations (SHAP) model was employed to quantify the impact of the inducing factors, showing that extreme rainfall, temperature, soil moisture, and elevation to be the key factors. The GNN-based regionalization method highlighted in this study has shown great potential in supporting disaster prevention and mitigation practices for flash flood disasters in the HMR.

DOI:

Year: 2024