Deep learning approach for real-time tsunami impact forecasting in near field context – application to the French Mediterranean coastline

Tsunami warning systems currently focus on the first parameters of the earthquake, based on a 24-hour monitoring of earthquakes, seismic data processing (Magnitude, location), and tsunami risk modelling at basin scale.

The French Tsunami Warning Center (CENALT) runs actually two tsunami modelling tools where the water height at the coast is not calculated (i.e., Cassiopee based on a pre-computed database, and Calypso based on real time simulations at basin scale). A complete calculation up to the coastal impact all along the French Mediterranean or coastline is incompatible with real time near field or regional forecast, as nonlinear models require fine topo-bathymetric data nearshore and indeed a considerable computation time (> 45 min). Predicting coastal flooding in real time is then a major challenge in such context. To overcome these limitations, non conventional approches such as machine learning methods are being explored. Among the huge number of actual models, deep learning techniques are becoming increasingly popular. Severals studies have shown the interest of using MLPs (Multilayer perceptrons) and CNNs (Convolutional neural networks) to quickly transform a deep ocean simulation result into a coastal flooding model. Once trained on a specific output area with a large dataset, the networks are able to predict in seconds the tsunami inundation map from any earthquake scenario drawn from a seismic source database representative of the seismotectonic context of the region of interest.

A first study training neural networks to predict the maximum water height maps was performed on three specific French cities (Nice, Antibes and Cannes) to evaluate the capacity of the models to reproduce the ground truth. The objective here is to extend the method to predict, in addition to maximum wave heights and runups, maximum retreats and currents along the entire French Mediterranean coastline. The spatial resolution of the finer bathymetric grids is set to 25 meters. To be representative of reality, the training dataset is fed with seismic scenarios derived from the CENALT fault database and taking into account a stochastic slip distribution. The method provides promising early results.