Flood susceptibility assessment and mapping in a monsoon-dominated tropical river basin using GIS-based data-driven bivariate and multivariate statistical models and their ensemble techniques

Floods constitute the most frequent and damaging natural hazard in monsoon dominated Indian subcontinent. Therefore, flood susceptibility modeling and mapping at basin scale is necessary for implementing management projects to reduce the adverse impact of flood hazards. This study evaluates the efficiency of bivariate frequency ratio (FR), and Dempster-Shafer based evidential belief function (EBF) models in flood susceptibility mapping of the tropical lower Dwarakeswar river basin of India, both individually and in combination with the multivariate logistic regression (LR) model. For this purpose, a flood inventory map (FIM), containing 162 past flood locations was prepared. After multicollinearity testing, the spatial data layers were created for 11 flood-related conditioning factors (FCFs): elevation, slope, sediment transport index (STI), topographic wetness index (TWI), stream power index (SPI), distance from streams, plan curvature, terrain ruggedness index (TRI), lithology, Land use/ cover (LU/LC), and normalized difference vegetation index (NDVI). A total of 113 flood locations (70%) were used for training the models and preparing flood susceptibility maps. The remaining 49 flood points (30%) were used for validating the models and comparing their efficiency using the area under the curve (RoC-AUC) and modified seed cell area index (mSCAI) methods. The validation result indicated that the FR model had the highest predictability (72.26%), and the FR-LR ensemble model had the lowest efficiency (58.65%). All the models showed that the southeastern parts of the study area are most susceptible due to lower elevation and gentle slopes.