Flood mapping based on the combination of support vector regression and Heun’s scheme

Abstract

In response to the increasing threats of flash flooding, how to generate flood maps in real time has drawn great attention. In this study, a hybrid ML model was developed for flood map generation with multiple lead times based on the combination of the support vector regression (SVR) method and Heun’s scheme for numerical correction, in which the hydraulic simulation results are used for model training while rainfall and flood sensor data are used as input variables (namely, the SVR-RFN model). To evaluate the performance of the hybrid model, two additional SVR models were developed and used as references for comparison—the first one used rainfall as input variables without sensor data (namely, the SVR-R model), and the second one added flood sensor data as input variables without numerical correction (namely, the SVR-RF model). The three models are tested and compared with a hydraulic model for a historical event in Sanyei River Basin, Taiwan. Using the hydraulic model as the benchmark, the hybrid model performed the best by reducing 63% of flood depth error and had the least deterioration of accuracy as lead time increased. The SVR-RF model performed the second by reducing the flood depth errors by 46% whereas the hydraulic model was more accurate than the SVR-R model. In a comparison of flood extents, the hydraulic model showed overestimation, the SVR-R model showed significant underestimation, whereas the two models with sensor data had balanced predictions. The flood maps generated by the hybrid model were applied for dike safety evaluation and showed a good capability in forecasting the risk of dike breach.