Flood hazard assessment using machine learning and hydrodynamic modeling: case study in the Vu Ga–Thu Bon basin in Vietnam

Abstract

The objective of this study was the construction of a theoretical framework to evaluate flood hazards by integrating machine learning and hydrodynamic modeling in Vietnam’s Vu Gia–Thu Bon basin. MIKE FLOOD was used to simulate historical floods in 2017 in order to obtain flood depth and velocity. Support vector machine-stochastic gradient descent (SVM-SGD), decision trees (DT), and dagging (DA) were used to determine flood susceptibility. Flood hazard was constructed by combining the flood depth, velocity, and susceptibility using the analytic hierarchy process technique. The statistical indices Area under the curve-receiver operating characteristic (AUC), Root mean square error (RMSE), mean absolute error (MAE), and NASH were used to evaluate the precision of the hydrodynamic and machine learning models. The results showed that hydrodynamic modeling was highly accurate, with a NASH value of 0.86. The proposed models achieved AUC values of 0.98 for SVM-SGD, 0.93 for DT, and 0.92 for DA. The results showed that 7.59% of the flood zones is located in the very low flood hazard zone, 108.2 km2 in the low flood hazard zone, 24.59% in the moderate flood hazard zone, 22.53% in the high flood hazard zone, and 10.01% in the very high flood hazard zone.