Integrating social media data and machine learning methods for flash flood susceptibility mapping in China

Abstract

Flash floods represent one of the most hazardous natural disasters globally, with China ranking among the nations most severely impacted by such events. Assessment of flash flood susceptibility and risk provides critical information for relevant authorities, and is essential for disaster prevention and mitigation in mountainous regions. This study compiled historical flash flood data from social media platforms, to construct a consistent dataset for the spatiotemporal analysis of flash floods across China. Five machine learning algorithms were employed to model and spatially map flash flood susceptibility nationwide. The results reveal distinct spatiotemporal patterns: between 2012 and 2023, the distribution of flash floods shifted from an initial concentration in central China to progressive northeast and southwest expansion. Temporal analysis indicates a statistically significant upward trend in disaster frequency over the study period. Model validation metrics demonstrated superior predictive performance by XGBoost (Accuracy: 0.931; AUC: 0.993), followed by SVM, RF, NB, and ANN. Key determinants of flash flood susceptibility include road network density, daily maximum precipitation, sand ratio, and average typhoon frequency. Western Sichuan, Yunnan-Guizhou Plateau and Zhejiang’s Hilly Terrain are found with highest flash flood susceptibility. This study demonstrates the reliability of social media data, offering novel approaches for flash flood risk assessment. Based on the findings of this study, it is recommended to implement ecological restoration in western mountainous areas prone to flash floods and establish a portfolio of preventive measures against typhoon-triggered flash floods in the southeastern coastal regions.