In the DataBeach study a machine learning model was developed with the aim to improve the efficiency and sustainability of design of soft coastal defense projects. A morphological model based on machine learning was trained and tested to predict beach volume changes with significantly reduced computational time compared to traditional process-based models. The machine learning model was then applied, combined with a ‘penalty function’ for inclusion of morphological feedback, to predict beach volume changes for the study area of approximately 2 km alongshore and on a 10-year project timescale. In order to run many different scenarios for the 10-year prediction, a probabilistic methodology was developed to take into account the uncertainties in this time period. The machine learning-based model provides great benefits for probabilistic simulations, due to the lower computational time, compared to process-based numerical models such as XBeach, and a flexible way in which it can incorporate measurement data. The performance of the tested machine learning models was comparable to that of the short term volume predictions of XBeach. Comparison of 10-year predicted volume changes using the machine learning model and measured beach topography (LiDAR) showed good agreement between measured and predicted volume changes for the dry beach area (when accounting for nourishments), but overestimation in the beach volume change predictions for the intertidal beach. These differences are partially attributed to the poor performance on XBeach for long term, normal wave conditions, which are an important factor in the intertidal area.

In this paper, a new approach to model wave-driven, cross-shore shoreline change incorporating multiple timescales is introduced. As a base, we use the equilibrium shoreline prediction model ShoreFor that accounts for a single timescale only. High-resolution shoreline data collected at three distinctly different study sites is used to train the new data-driven model. In addition to the direct forcing approach used in most models, here two additional terms are introduced: a time-upscaling and a time-downscaling term. The upscaling term accounts for the persistent effect of short-term events, such as storms, on the shoreline position. The downscaling term accounts for the effect of long-term shoreline modulations, caused by, for example, climate variability, on shorter event impacts. The multi-timescale model shows improvement compared to the original ShoreFor model (a normalized mean square error improvement during validation of 18 to 59%) at the three contrasted sandy beaches. Moreover, it gains insight in the various timescales (storms to inter-annual) and reveals their interactions that cause shoreline change. We find that extreme forcing events have a persistent shoreline impact and cause 57–73% of the shoreline variability at the three sites. Moreover, long-term shoreline trends affect short-term forcing event impacts and determine 20–27% of the shoreline variability.

Implementation of multiple shoreline response factors to consider different timescales and their interplay in equilibrium shoreline models.• Extreme forcing events can have a persistent impact on the longer term state of the beach.• The long term shoreline location can modulate extreme event impacts.

Shoreline change is affected by a multitude of complex processes operating at various spatiotemporal scales. Comprehensive multi-year simulations of shoreline changes and forecasts are feasible with process-based models. However, these detailed and computationally expensive numerical simulations do not always lead to increased predictive skill in comparison to simpler shoreline models (Davidson et al., 2013). ShoreFor (Davidson et al., 2013) employs the concept of (dis-) equilibrium of shoreline location following Wright and Short (1985). In this research, the current ShoreFor model (Splinter et al., 2014) is used as baseline. ShoreFor seeks for an optimum decay factor that best describes the morphological response of a coastal system to the corresponding hydrodynamic forcing. This parameter is measured in days and effectively controls the shoreline response timescale. Currently, the ShoreFor model provides a single value for φ, representing a single dominant shoreline response timescale. As morphological systems can contain multiple dominant timescale responses, a new approach to multi-timescale shoreline change modelling is proposed. Three video-derived datasets are used to improve the model towards a generally applicable one which incorporates multiple temporal scales: Narrabeen (Australia), Nha Trang (Vietnam) and Grand Popo (Benin). Each dataset has different hydrodynamic- and morphological characteristics. The storm timescale is a dominant mode of shoreline response for Narrabeen, whereas for Nha Trang and Grand Popo the seasonal timescale is the most dominant. Furthermore, all sites are subjected to more modes of shoreline response such that the application of a single memory decay factor will hamper shoreline modelling. The existing model is improved using 3 steps.In the first step, the raw wave- and shoreline signals are filtered to distinguish temporal scales. Then filtered temporal scales in shoreline position are forced with the corresponding scales in the wave signals. For each temporal scale, a distinct memory decay factor φ is found. In the second step, the effect of small temporal scales in wave forcing on larger temporal scales in shoreline position is accounted for. The improved model takes this effect into account using the envelope of the filtered wave signals. The envelope is used to force the model and to calculate shoreline change with the same timescale. In the third and final step, the effect of large temporal scales in shoreline position on smaller scales in shoreline response is accounted for. The efficiency with which waves induce cross-shore sediment transport can be dependent on the large scale shoreline variation. A time varying response factor c is introduced that controls the efficiency with which waves induce cross-shore sediment transport. The dynamic response factor varies over time with the shape of the larger scale shoreline signal: it represents the effect of the large scale shoreline variation on the smaller scale shoreline response.