On the role of hydrodynamic and morphologic variables on neural network prediction of shoreline dynamics

Abstract

Predicting shoreline change is a key issue in coastal research. Predictors, process-based or data-driven, tend to be developed and tested on high-frequency and high-quality data sets. Combining hydrodynamic and morphological variables extracted from video images and artificial neural network allows us to evaluate if sparse data could still provide physically-sound shoreline change predictions. The data set covered a 3-year period with shoreline position data (with an accuracy of ±5 m) available 73 % of the time and 66 % for the morphological parameters (beach state or bar location). The best configuration of the trained shallow (one hidden layer) Feedforward Artificial Neural Network (ANN), includes 10 input variables and 10 nodes allowing to capture the shoreline dynamic at different time scales, from the storm-event to the seasonal scale, and to predict the shoreline position on a 1-year period with a RMSE of about 6.7 m. Increasing the complexity of the architecture of the ANN by increasing the number of hidden layers did not improve the predictions. By modifying the number of input variables in the algorithm, the ANN also allows us to highlight the mitigation effect of the bar during the storm event and its role as sediment buffer during seasonal accretion.