Flash Flood Susceptibility Assessment Based on Binary Statistics and Machine Learning

Flash flood susceptibility assessment is critical for risk identification, early warning system optimization, and disaster mitigation planning. The objective of this study is to investigate the applicability of bivariate statistical and machine learning methods in flash flood susceptibility assessment. This study focuses on Jiangxi Province, a high-frequency flash flood region in southeastern China. Small watersheds are adopted as assessment units. First, 12 risk factors are selected from three aspects: hazard triggers, environmental susceptibility, and vulnerability exposure. Second, we establish three assessment models using two statistical methods (Frequency Ratio and Weight of Evidence) and a machine learning algorithm (Random Forest). The importance of each risk factor is quantitatively analyzed. Ultimately, model performance is systematically evaluated using the area under the curve (AUC), Accuracy, and Precision. Additionally, assessment accuracy is reflected by the capture rate of historical flash flood points. The results show that the Random Forest model outperformed statistical methods, achieving 0.874 AUC, 0.822 Accuracy, and 0.813 Precision. In terms of historical disaster point capture within moderate to high susceptibility zones, the Random Forest, Weight of Evidence, and Frequency Ratio models achieve capture rates of 97.0 %, 83.4 %, and 67.7 %, respectively. These findings indicate that Random Forest and Weight of Evidence demonstrate strong predictive capability, while Frequency Ratio exhibits limited applicability.