Assessment of Flash Flood Risk Using Random Forest Regression Model Integrated with Binary Statistics

Flash floods are one of natural disasters that result in significant economic losses and human casualties. Previous studies on flash flood assessment generally applied binary or classification models. In this paper, we proposed a regression model based on the integration of Random Forest (RF) algorithm and binary statistics to assess flash flood risk. Jiangxi Province, one of the Chinese provinces most severely impacted by flash floods, was chosen as the case study. The Flood Disaster Index (FDI) in each affected small catchment is calculated based on historical flash flood data in Jiangxi Province. A total of 15 risk factors were collected. The binary statistics algorithm of Certainty Factor (CF) method was used to examine the interval sensitivities of risk factors, which were then used to calculate the integrated weights of the risk factors. Both regression models including RF regression model with FDI and the 15 risk factors as inputs, and RF regression model with FDI and the integrated weights of the risk factors from CF as inputs (CF-RF) were applied to achieve flash flood disaster prediction in Jiangxi Province. The percentage of the historical disaster points captured, high-risk misclassification rate and low-risk error rate were used to evaluate the performances of the regression models. Results indicated that the integrated CF-RF model showed the optimal performance, whose high-risk misclassification rate and low-risk error score rate were 3.18% and 10.34%, respectively. Moreover, the integrated CF-RF regression model captured 96.18% of the historical disaster points. This study indicated the enhanced prediction performance of machine learning regression models integrated with binary statistics.