MODELLING SHORELINE DYNAMICS IN COMPLEX MACROTIDAL ENVIRONMENTS USING NEURAL NETWORKS

Overview & Synopsis

The aim of this work is to test an artificial intelligence approach with basic hydrodynamic and morphological variables, in order to assess the effectiveness of such methods in modelling complex beach dynamics. A simple feedforward neural network is used to evaluate the impact of selected variables on the prediction of the dynamics of specific shoreline isocontour proxies extracted from beach profiles, and to build a predictive model that could simulate the position of the proxies. The model was trained on datasets from two sites from the French coastal monitoring program DYNALIT, Porsmilin and Vougot beaches, and their profile measurements, water levels and wave measurements over 20 years. These two sites were selected due to differences in their morphology and hydrodynamics, as a means to assess the performance of neural networks over a larger variety of situations. A range of temporalities encompassing 3 days, 7 days, 14 days, and 30 days of selected hydrodynamic and morphological variables were used to study the impact time scales can have on modelled shoreline positions. The shoreline proxies used for Porsmilin and Vougot beaches correspond respectively to the berm and the contact between the dune toe and the upper beach, which can be assessed and followed along each beach profile. The shallow feedforward network included 1 hidden layer and 5 nodes, and was ran 50 times in order to assess the models' performance. The models were generally successful, with a blind shoreline prediction R of 0.88 in Porsmilin and 0.72 in Vougot. This artificial neural network (ANN) approach showed all-around better performance than previous beach equilibrium models, which is very encouraging regarding the prediction of future beach morphodynamics and the use of Machine Learning algorithms therein.