A novel hybrid of meta-optimization approach for flash flood-susceptibility assessment in a monsoon-dominated watershed, Eastern India

The exponential growth in the number of flash flood events is a global threat, and detecting a flood-prone area has also become a top priority. The flash flood-susceptibility mapping can help to mitigate the worst effects of this type of risk phenomenon. However, there is an urgent need to construct precise models for predicting flash flood-susceptibility mapping, which can be useful in developing more effective flood management strategies. In this present research, support vector regression (SVR) was coupled with two meta-heuristic algorithms such as particle swarm optimization (PSO) and grasshopper optimization algorithm (GOA), to construct new GIS-based ensemble models (SVR–PSO and SVR–GOA) for flash flood-susceptibility mapping (FFSM) in the Gandheswari River basin, West Bengal, India. In this regard, 16 topographical and environmental flood causative factors have been identified to run the models using the multicollinearity (MC) test. The entire dataset was divided into 70:30 for training and validating purposes. Statistical measures including specificity, sensitivity, PPV, NPV, AUC–ROC, kappa and Taylor diagram have been employed to validate adopted models. The SVR-based factor importance analysis was employed to choose and prioritize significant factors for the spatial analysis. Among the three modeling approaches used here, the ensemble method of SVR–GOA is the most optimal (specificity 0.97 and 0.87, sensitivity 0.99 and 0.91, PPV 0.97 and 0.86, NPV 0.99 and 0.91, AUC 0.951 and 0.938 in training and validation, respectively), followed by the SVR–PSO (specificity 0.84 and 0.84, sensitivity 0.87 and 0.86, PPV 0.85 and 0.82, NPV 0.87 and 0.87, AUC 0.951 and 0.938 in training and validation, respectively) and SVR (specificity 0.80 and 0.77, sensitivity 0.93 and 0.89, PPV 0.82 and 0.77, NPV 0.91 and 0.89, AUC 0.951 and 0.938 in training and validation, respectively) model. The result shown that 40.10 km² (10.99%) and 25.94 km² (7.11%) areas are under very high and high flood-prone regions, respectively. This produced reliable results that can help policymakers at the local and national levels to implement a concrete strategy with an early warning system to reduce the occurrence of floods in a region.