Using a deep learning model to predict flood arrival times after a dike breach

When a dike is at the point of breaching during a flood, quick decisions needs to be made to minimize the number of casualties. To do so, safety region managers rely on either pre-run inundation models for dike breaches at a specific location with general characteristics, or run real-time a model with situation specific parameters. The problem is however that the first option does not represent the actual situation, while the second option takes 1-2 days for the model to compute the results. 

There is therefore a need for a quick though accurate flood prediction model. The aim of this research is therefore to explore the possibilities of a deep learning model to predict flood arrival times after a dike breach. The great advantage of deep learning models is its capability of reducing the computation time in orders of magnitudes while maintaining high accuracies. 

This can only be achieved when the model is properly trained, which requires many flood simulations. As retrieving a sufficient amount of actual flood information data is cumbersome and labor-intensive work, synthetic landscapes are created in Python and thereafter flooded by making use of the numerical flooding model Delft3D FM Suite 1D2D. The flood arrival time of these simulations is then used to evaluate the performance of the deep learning model. 

Once the deep learning model is sufficiently trained on the synthetic cases, it will be applied to a real case scenario of Dijkring 48. The waterboard of Rijn and Ijssel is therefore included in the research as well, which is supervised by Deltares and the TU Delft.