Flash flood disaster risk simulation technology based on classified recursive feature elimination-random forest optimization algorithm

Abstract

[Objective] Flash flood disasters cause severe economic losses and casualties to human society, making the scientific identification and assessment of flash flood disaster risk an urgent issue to be addressed. The aim of the study is to improve the accuracy of flash flood risk prediction by coupling feature selection with the Random Forest algorithm, thereby providing a scientific basis for disaster early warning. [Methods] Seventeen feature factors associated with the occurrence of flash flood disasters were selected. A feature selection approach that integrated Classified Recursive Feature Elimination(RFE-class) with a Random Forest optimization algorithm was proposed to identify the optimal feature combination for flash flood risk simulation. [Results] The result showed that the optimal feature combination obtained using the RFE-class method significantly improved the predictive performance of the Random Forest model, achieving a Receiver Operating Characteristic(ROC) curve value of 94.7%, representing an approximately 5% improvement in accuracy compared to using the Random Forest algorithm alone. [Conclusion] In Fujian Province, the high-risk areas for flash flood disasters are primarily distributed in the Wuyi Mountains, Daiyun Mountains, and Daimao Mountain regions, covering an area of approximately 49 000 km² and affecting 27 million people.

Keywords

flash flood disaster risk / Random Forest / recursive feature elimination method / small watershed scale / influencing factors