Data-driven shoreline modelling at timescales of days to years

Abstract

An increased availability of long-term coastal imaging datasets has opened the door to the use of data-driven modelling approaches to predict shoreline evolution in response to wave and water level conditions. In this study an autoregressive neural network approach has been applied to predict shoreline change over daily to yearly timescales. A dataset comprising two embayed beaches (Narrabeen Beach, Australia and Tairua Beach, New Zealand) has been used, spanning 10 years of daily shoreline position observation at each site. The model shows good cross-validation performance, predicting the shoreline position with an average 4.64 m RMSE (0.78 NMSE) at Tairua and 5.73 m RMSE (0.46 NMSE) at Narrabeen over approximately 2-year test periods. The autoregressive component of the model involved the use of the last predicted shoreline position in the prediction of shoreline change over the next timestep. This “memory” of past conditions was found to be crucial to maintaining model stability and prediction accuracy over timescales of weeks to years. Model outputs were interrogated to show the structure of the equilibrium response to previous shoreline position which was more prevalent at Tairua. The model is quite robust to changes in the quantity and temporal resolution of the training data, though training data of more than 2-years was desirable, particularly at Narrabeen. • Auto-Regressive Neural Network’s are capable of reproducing observed shoreline variability at two storm dominated beaches in Australia and New Zealand • Hysteresis in the model is important for long-term stable model predictions • Model architecture learns equilibrium-type behaviour of the shoreline with respect to incident wave forcing • The machine learning model is not overly sensitive to the amount of data required for training beyond 2 years. • Noisy data, as would be found from satellite derived shorelines, reduces overall model skill

Topics & Concepts