Spatial prediction of flash flood susceptible areas using novel ensemble of bivariate statistics and machine learning techniques for ungauged region

Flash floods are considered one of the most devastating natural hazards due to a short time scale. Ensemble-based approaches have recently become popular in flash flood susceptibility modeling due to their strength and flexibility with data. This study aimed to incorporate new ensemble approaches to bivariate statistical model, such as the quantitative approach of weight of evidence (WOE) with multivariate statistical models, such as artificial neural networks (ANN), support vector machine (SVM), and the K nearest neighbor (KNN) model. The Uttarakhand state of India was selected as a study area. A flash flood and geospatial database were developed in this regard. In the historical database, a total of 122 flash flood points were identified. A geospatial dataset was created with aspect, plan curvature, elevation, normalized difference vegetation index (NDVI), slope, stream power index (SPI), topographic wetness index (TWI), annual rainfall, distance from river, distance from road, land use/cover (LULC), and sediment transport index (STI) in GIS. Weights were assigned to each influencing factor based on correlation using WOE in R open-source software, then ensembled with ANN, SVM, and KNN. Finally, all models were validated with different statistical indices, and subsequently, their performances were compared. All of the built models performed well, according to the results. However, WOE-ANN outperformed all machine learning models. The results of the study can help local governments and researchers with flash flood management.