On the use of Machine learning approach for assessing the cascading flood risks in estuarine environments

With the global context of climate change and the increasing human pressure on ecosystems, coasts and estuaries are more vulnerable to environmental hazards and are currently facing an intensification of natural hazards. These hazards have become more serious in the last decade due to the overexpansion of urbanization and infrastructures in these areas, along with climate change effects including sea-level rise and storminess increases. Recently, compound flood events have been addressed in some works, such as those induced by the combination of river discharge and surges, rainfall and surges, and rainfall, surge, and waves. In estuarine contexts, controlled by the combined effects of different continental and marine drivers, compound flood events are frequently produced and cause severe consequences. For such systems, the main approach used considers the predominance of high-energy storms, extreme discharge, and wave-induced events responsible for flooding. The effects of hydrogeological forces on superficial flooding are neglected. The present work investigates the dynamics of cascading flood risks in a river-tide environment, the case of the Seine estuary, controlled by different drivers that are produced cascading since they are induced by the same climate forcing. For example, the combined effects of low atmospheric pressure and strong winds associated with synoptic-scale storms can induce storm surges along the coast. This environment is considered an excellent natural laboratory to analyze river-surge interaction because of its time-varying flow and the available water-level records provided by tide gauges along the estuary. A series of cascading events in the lower Seine estuary, extracted from historical records, has been examined by integrating a comprehensive array of environmental data that gathers hydrological (discharge), meteorological (precipitation, wind, temperature, evapotranspiration), marine (swell, tide), and hydrogeological (groundwater levels) factors. To model these events, accurate supervised machine learning techniques, including gradient boosting and XGBoost, were applied at five critical tidal gauge stations. The results obtained highlight good accuracy in predicting water levels, with a margin of error of 30-40 cm during training phases and 40-50 cm during the periods of flooding. This accuracy is more pronounced downstream. Tides, considered as the predominant factor and accounting for more than 90% of the estuary's hydrodynamic dynamics, contribute to a range of between 2 - 7% of the reconstructed signal. Other parameters, such as meteorological and hydrogeological conditions, although in the minority, play an essential supporting role in the overall flood dynamics. In summary, this research demonstrates the effectiveness of machine learning methods for assessing flood risks. It reveals that the dynamics of flooding in estuaries are mainly controlled by the effects of tides and discharges as well as the non-linear interaction between both of them. The findings offer a new perspective on the underlying mechanisms of flooding in a tide-river environment, by integrating a variety of parameters in an AI-based approach whose full consideration would be complex in numerical models.