Early-Warning System for Cyclone-Induced Wave Overtopping Aided by a Suite of Random Forest Approaches

Abstract

Full process-based hydrodynamic models allows reproducing with high fidelity the complexity of cyclone-induced wave dynamics and their potential impact regarding overtopping-induced inundation. Yet, their major drawback is the high computational time cost (typically of several hours), which prevents them from a direct integration in a real-time early-warning system to support crisis/emergency management. A possible alternative relies on the statistical exploitation of pre-calculated simulation results to build a fast (low computation time cost) prediction model given the offshore conditions (wave and sea level). In the present study, we present a suite of random forest (RF) techniques for fast prediction of key indicators to support forecasting of cyclone-induced wave overtopping, namely: (1) the occurrence likelihood of the overtopping event using a RF-based classification method; (2) the maximum cumulative volume overtopping using a RF-based regression method; (3) the starting time of the overtopping event (referred to as time-to-event) using a multi-output RF-based classification method. We apply the technique at Sainte-Suzanne city at Reunion Island (Indian Ocean basin) using a database of simulation results, which relates offshore conditions (wave and sea level) induced by ~500 synthetic cyclones and their consequences in terms of wave overtopping, namely the time evolution of the cumulative overtopping volume in the vicinity of the emergency center. Through an extensive cross-validation exercise, we show the high performance of the RF models with respect to the prediction of the three indicators. More specifically, the accuracy and the area under the ROC curve (AUC) of the RF-based classifier reaches values above 95%; the R-squared of the RF-based regression model reaches values above 80%; the time-averaged accuracy of the multi-output RF-based classifier reaches values above 80%. As a complementary analysis, the comparison to simulated historical cases (Dumile in 2013, and Dina in 2002) shows error less than 10% on flooding indicators 2 and 3. Finally, we take advantage of the probabilistic information provided by RF models to evaluate some measure of confidence associated to the prediction result.