Flash flood prediction in St. Lucia island through a surrogate hydraulic model

Abstract

ABSTRACT Recent flood disasters caused by extreme meteorological events highlight the need of fast and reliable tools for flooding forecast. For our purposes, the danger associated with floods is embodied in a single risk-level flag which considers both local water depth and velocity. The methodology here derived is applied and validated for the case study of the St. Lucia island in the eastern Caribbean Sea that experiences flash flooding as a result of combined intense rainfall and steep slopes, difficult to predict with traditional early-warning systems. A multi-layer perceptron neural network is trained on a high-fidelity dataset generated through full two-dimensional shallow water simulations of real and synthetic events. The dataset is validated against social markers obtained from real events. The predictive capabilities of the neural network model are tested on the out-of-box case of the Dean and Tomas hurricanes and compared with the solutions of the shallow water solver. The surrogate solver allows a significant speed-up in the prediction time with respect to traditional CFD (seconds vs hours), showing a high precision and accuracy, with accuracy, precision and F1-score above 0.99.