Abstract
High-fidelity numerical models can accurately simulate coastal flooding, but their high computational cost makes them unsuitable for real-time emergency response. This paper details a metamodeling approach that utilizes machine learning to mimic complex hydrodynamic processes, providing local flood forecasts for Gâvres, France, in minutes rather than days. The system tracks 14 user-defined impact indicators, ensuring that the forecasts are actionable for local authorities.
Introduction
Coastal flooding, particularly driven by wave overtopping, presents a significant challenge for predictive modeling. While advanced phase-resolving hydrodynamic models like SWASH offer high precision, they are computationally expensive; simulating a 6-hour storm can take several days on a supercomputer. This latency renders them ineffective for short-term forecasting windows. Traditional warning systems often rely on simplistic tide-height thresholds, which fail to account for the complex interaction between wave conditions and local coastal defenses. This research aims to bridge this gap by developing a fast, robust, and user-centered local warning system (SPAS) for the town of Gâvres.
Methodology
The core of the SPAS system is a "metamodel"—a mathematical surrogate trained to learn the relationship between offshore ocean conditions and onshore flooding from a pre-calculated database of high-resolution simulations.
User-Centered Design
A key innovation of this approach is the involvement of local stakeholders (e.g., town halls, emergency services) to identify 14 specific indicators ($I_j$) that directly inform emergency response. These include water depths at critical locations like the town hall or gymnasium, and the "trafficability" of key evacuation routes.
Metamodeling Techniques
The researchers employed three types of metamodels to handle different data structures:
- Scalar-to-Scalar: Quantile Random Forests were used to account for the stochastic nature of waves, providing a range of possible outcomes rather than a single mean.
- Functional-to-Scalar: Gaussian Processes were utilized to translate time-series data (e.g., a 6-hour storm profile) into discrete values like total flooded area.
- Functional-to-Functional: Advanced Gaussian Processes were implemented to predict entire spatial maps of maximum water heights.
To improve accuracy, a "two-step" prediction process was adopted: one metamodel classifies whether a flood will occur, and a second quantifies the magnitude of the event.
Data and Training
The metamodels were trained on synthetic datasets generated by a high-resolution numerical chain (WW3 for offshore waves and SWASH for nearshore flooding). The training sets included 144 scenarios for scalar models and 174 for functional models. Validation was performed using a historical database of 48 flood events in Gâvres between 1900 and 2010. Operationally, the system integrates real-time tide, surge, and wave forecasts from the MARC (Ifremer) and DATASHOM platforms.
Results and Validation
The metamodeling approach achieved a massive computational speedup, delivering predictions in minutes compared to the 72 hours required for full numerical simulations. The models demonstrated high predictive accuracy ($Q^2$) and low root-mean-square error (RMSE) relative to the full physics simulations.
During real-world testing between 2019 and 2021, the system successfully predicted "no flood" or "limit of flooding" for several storm events, resulting in zero false alarms. Calibration against historical data identified an optimal flood detection threshold of 62 m³ of water entering the town within a 15-minute window.
Conclusion
The SPAS system demonstrates that metamodeling can effectively bring high-fidelity coastal flood modeling into the operational timeframe required for emergency management. By focusing on user-defined indicators rather than raw physical data, the system provides actionable intelligence for local authorities. Future work will focus on the operationalization of the system and addressing the challenges of site-specific training and evolving coastal defenses.
Summary updated