Embedded Feature Selection and Machine Learning Methods for Flash Flood Susceptibility-Mapping in the Mainstream Songhua River Basin, China

Abstract

Mapping flash flood susceptibility is effective for mitigating the negative impacts of flash floods. However, a variety of conditioning factors have been used to generate susceptibility maps in various studies. In this study, we proposed combining logistic regression (LR) and random forest (RF) models with embedded feature selection (EFS) to filter specific feature sets for the two models and map flash flood susceptibility in the mainstream basin of the Songhua River. According to the EFS results, the optimized feature sets included 32 and 28 features for the LR and RF models, respectively, and the composition of the two optimal feature sets was similar and distinct. Overall, the relevant vegetation cover and river features exhibit relatively high effects overall for flash floods in the study area. The LR and RF models provided accurate and reliable flash flood susceptibility maps (FFSMs). The RF model (accuracy = 0.8834, area under the curve (AUC) = 0.9486) provided a better prediction capacity than the LR model (accuracy = 0.8634, AUC = 0.9277). Flash flood-prone areas are mainly distributed in the south and southwest and areas close to rivers. The results obtained in this study is useful for flash flood prevention and control projects.

Keywords

• flash flood
• susceptibility mapping
• feature selection
• random forest
• logistic regression