Flash flood susceptibility prediction mapping for a road network using hybrid machine learning models

Flash flood is one of the most common natural hazards affecting many mountainous areas. Previous studies explored flash flood susceptibility models; however, there is still a lack of case studies in the transport sector. This paper aimed to develop advanced hybrid machine learning (ML) algorithms for flash flood susceptibility modeling and mapping using data from the road network National Highway 6 in Hoa Binh province, Vietnam. A single ML model of reduced error pruning trees (REPT) and four hybrid ML models of Decorate-REPT, AdaBoostM1-REPT, Bagging-REPT, and MultiBoostAB-REPT were applied to develop flash flood susceptibility maps. Field surveys were conducted about the flash flood locations on the 115-km route length of the National Highway 6 in 2017, 2018, and 2019 flood events. This study used 88 flash flood locations and 14 flood conditioning factors to construct and validate the proposed models. Statistical metrics, including sensitivity, specificity, accuracy, root mean square error, and area under the receiver operating characteristic curve, were applied to evaluate the models’ performance and accuracy. The DCREPT model showed the best performance (AUC = 0.988) among the training models and had the highest prediction accuracy (AUC = 0.991) among the testing models. We found that 12,572 ha (Decorate-REPT), 9564 ha (AdaBoostM1-REPT), 11,954 ha (Bagging-REPT), 14,432 ha (MultiBoostAB-REPT), and 17,660 ha (REPT) of the 3-km buffer area of the highway are in the high- and very high-flash-flood-susceptibility areas. The proposed methodology could be potentially generalized to other transportation routes in mountainous areas to generate flash flood susceptibility prediction maps.