359 / 2024-02-29 20:43:31

Quantifying the influence of disaster-inducing factors on flash floods in the Hengduan Mountains, China, using GNN-based flash flood regionalization

flash floods,regionalization,disaster-inducing factors,the Hengduan Mountains,graph neural network (GNN)

Flash floods cause the most casualties among all kinds of floods in China, and the Hengduan Mountains region in Southwest China is one of the most severely affected areas by flash flood disasters. The Hengduan Mountains exhibit pronounced spatial heterogeneity in geomorphology, meteorology, and underlying surface conditions and the environmental factors from these three aspects together act as inducing factors for flash floods in this region. However, the influence of the inducing factors for flash floods in the Hengduan Mountains has not been quantitatively investigated yet. This knowledge gap hinders the better understanding towards the forming mechanisms of flash floods and the implementation of effective disaster prevention and reduction measures accordingly. Flash flood regionalization is a powerful tool to visualize the spatial heterogeneity of flash flood disasters for a target area and it quantifies the influence of input factors, which are normally the inducing factors for flash floods, in the meantime. Therefore, this study aims to quantify the impact of major disaster-inducing factors on flash floods in the Hengduan Mountains by using various combinations of these factors as inputs for flash flood regionalization.

Based on the significant physical processes among the formation of flash floods in the Hengduan Mountains, key disaster-inducing factors from geomorphology, meteorology, and underlying surface conditions were selected and combined into multiple combinations as input datasets. Each input dataset was transformed into a graph structure, with each node representing a grid and its edges connecting the eight adjacent nodes. The Dink-Net, a state-of-the-art unsupervised graph neural network (GNN), was applied to establish the raster-scale regionalization of flash floods, considering both the attributes and spatial topology structure of the nodes. The optimal regionalization result was chosen based on the Clustering Quality Index, Davies-Bouldin Index and the spatial distribution of historical flash flood disasters. Other regionalization results were then compared with the optimal regionalization, and the differences in input data were analyzed. Our results indicate that slope, topographic relief, vegetation type, the gravel content of surface soil, temperature and precipitation exert the most significant influence on flash flood regionalization in the Hengduan Mountains among all the tested inducing factors. Since most of these factors are closely related to the runoff generation mode of subsurface stormflow, our work highlights the significant role of the subsurface stormflow mechanism in the development of flash floods in the Hengduan Mountains.