A surrogate machine learning model using random forests for real-time flood inundation simulations

Abstract

Real-time simulation of flood inundation helps to mitigate the catastrophic effects on human lives by facilitating emergency evacuations. Traditional two-dimensional (2D) physics-based hydrodynamic models, though accurate, require significant computational time, thereby rendering them unsuitable for such real-time applications. To address this limitation, we developed Random Forest (RF) models as surrogate hydrodynamic models for predicting maximum flood depth and velocity under complex fluvial conditions with backwater effects. These models integrate hydrological parameters, such as upstream discharge, physical catchment characteristics, to enhance predictive accuracy and generalizability. A comprehensive assessment revealed that the inclusion of physical characteristics increased the prediction accuracy of RF models by 1.72 times and 2.60 times for depth and velocity models with root mean square error of 0.494 m and 0.148 m/s respectively, compared to baseline models. Furthermore, the RF models required only 1.5 %–4 % (for minor flood event and major flood event respectively) of the computational time needed by hydrodynamic models. With its ability to understand complex flooding scenarios with high prediction accuracy and computing efficiency, the proposed RF models have demonstrated great potential for real-time flood inundation modelling. Efforts in this direction to improve the real-time flood inundation predictions may greatly aid the decision makers for undertaking emergency evacuations during catastrophic flood events.