Identification of the key driving factors of flash flood based on different feature selection techniques coupled with random forest method

Abstract

Study regionThe mountainous area of Guangdong Province, China. Study focusA random forest coupling with feature selection (RF-FS) method framework, containing random forest coupling with shadow variable search, random forest coupling with recursive feature elimination, and random forest coupling with genetic algorithm, is introduced to identify the key driving factors (KDFs) of flash flood that can be used to explain the flash flood process and generate the flash flood risk map efficiently.New hydrological insights for the region: Among the KDFs identified by the RF-FS method framework, meteorological factors (maximum 1-day precipitation, total daily precipitation exceeding 95 % threshold, and maximum 5-day consecutive precipitation) are dominant, followed by hydrological factors (topographic wetness index and stream power index). Flash flood risk modeling was then developed by using the KDFs identified by the random forest coupling with genetic algorithm which is the best in the RF-FS method framework in predicting flash flood risks. Furthermore, the modeling results show that about 45 % of the mountainous areas in Guangdong Province are high flash flood-prone areas (high and extremely high-risk areas), and most of them are located in the central, eastern, and southwestern mountainous areas of Guangdong Province. The areas with flash flood-prone occurrences are generally consistent with the areas with large precipitation, implying that these areas should be prioritized for flash flood prevention and early warnings.