29 Mar 2022
29 Mar 2022
Status: this preprint is currently under review for the journal NHESS.

Estimating dune erosion at the regional scale using a meta-model based on Neural Networks

Panagiotis Athanasiou1,2, Ap van Dongeren1,3, Alessio Giardino4, Michalis Vousdoukas5, Jose A. A. Antolinez6, and Roshanka Ranasinghe1,2,3 Panagiotis Athanasiou et al.
  • 1Deltares, Delft, Netherlands
  • 2Water Engineering and Management, Faculty of Engineering Technology, University of Twente, Enschede, Netherlands
  • 3Department of Coastal and Urban Risk & Resilience, IHE Delft Institute for Water Education, Delft, Netherlands
  • 4Water Sector Group, Sustainable Development and Climate Change Department, Asian Development Bank, Manila, Philippines
  • 5Joint Research Centre (JRC), European Commission, Ispra, Italy
  • 6Department of Hydraulic Engineering, Faculty of Civil Engineering and Geosciences, Delft University of Technology, Delft, Netherlands

Abstract. Sandy beaches and dune systems have high recreational and ecological value, while they offer protection against flooding during storms. At the same time, these systems are very vulnerable to storm impacts. Process-based numerical models are presently used to assess the morphological changes of dune and beach systems during storms. However, such models come with high computational costs, hindering their use in real-life applications which demand many simulations and/or involve a large spatial/temporal domain. Here we design a novel meta-model to predict dune erosion volume (DEV) at the Dutch coast, based on Artificial Neural Networks (ANN), trained with cases from process-based modelling. First, we reduce an initial database of ~1,400 observed sandy profiles along the Dutch coastline to 100 representative Typological Coastal Profiles (TCPs). Next, we synthesize a set of plausible extreme storm events, which reproduces the probability distributions and statistical dependencies of offshore wave and water level records. We choose 100 of these events to simulate the dune response of the 100 TCPs using the process-based model XBeach, resulting in 10,000 cases. Using these cases as training data, we design a 2-phase meta model, comprised of a classifying ANN (which predicts the occurrence (or not) of erosion) and a regression ANN (which gives a DEV prediction). Validation against a benchmark dataset created with XBeach and a sparse set of available dune erosion observations, shows high prediction skill with a skill score of 0.82. The meta-model can predict post-storm DEV 103–104 times faster (depending on the duration of the storm) than running XBeach. Hence, this model may be integrated in early-warning systems or allow coastal engineers and managers to upscale storm forcing-dune response investigations to large coastal areas with relative ease.

Panagiotis Athanasiou et al.

Status: final response (author comments only)

Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor | : Report abuse
  • RC1: 'Comment on nhess-2022-106', Víctor Malagón-Santos, 16 May 2022
  • RC2: 'Comment on nhess-2022-106', Anonymous Referee #2, 22 Aug 2022

Panagiotis Athanasiou et al.

Panagiotis Athanasiou et al.


Total article views: 663 (including HTML, PDF, and XML)
HTML PDF XML Total BibTeX EndNote
490 154 19 663 11 9
  • HTML: 490
  • PDF: 154
  • XML: 19
  • Total: 663
  • BibTeX: 11
  • EndNote: 9
Views and downloads (calculated since 29 Mar 2022)
Cumulative views and downloads (calculated since 29 Mar 2022)

Viewed (geographical distribution)

Total article views: 622 (including HTML, PDF, and XML) Thereof 622 with geography defined and 0 with unknown origin.
Country # Views %
  • 1
Latest update: 27 Sep 2022
Short summary
Sandy dunes protect the hinterland from coastal flooding during storms. Thus, models that can efficiently predict dune erosion are critical for coastal zone management and early warning systems. Here we develop such a model for the Dutch coast based on machine learning techniques, allowing for dune erosion estimations in a matter of seconds relative to available computationally expensive models. Validation of the model against benchmark data and observations shows good agreement.