Panagiotis Athanasiou, Ap van Dongeren, Alessio Giardino, Michalis Vousdoukas, Jose A. A. Antolinez and Roshanka Ranasinghe
Frontiers in Marine Science
Work Package 4
Link: https://doi.org/10.3389/fmars.2021.747754
Highlights
Storms can have severe impacts on coastlines, eroding sandy beaches and dunes, and damaging ecosystems and coastal communities. Modelling the possible impact on a coastline from an incoming storm is a valuable tool in reducing its impact, however this is very computationally expensive when applied over a large area. This study uses a new machine learning-driven clustering method to select a representative subset of 1,430 elevation profiles covering the entire Dutch coast. The team were able to maintain good prediction skill (based on modelled dune erosion from historical storms) using this subset, while reducing the input to the model by 93%. This makes such modelling over a long coastline much more feasible, allowing it to be used by communities and policymakers to prepare for storm events on the coast and mitigate their impacts.
Background
The impacts that storms can have on sandy beaches and dunes have large implications for coastal communities. Their erosion can disturb coastal ecosystems, impact the local economy and leave infrastructure vulnerable to damage. Predictions of coastal change, and more specifically dune impact, using numerical computer modelling have been previously employed for informed coastal zone management. However, these models are computationally expensive and impractical when a large area of coast needs to be covered or there is an urgent need for the results (e.g. early warning). This study attempts to address this issue by reducing a high number of coastal profiles observed along the Dutch coast down to a representative subset. This subset is much easier for the computer models to handle and could significantly improve the efficiency of storm impact prediction systems, helping communities prepare for these events.
Results
The team developed a new clustering method using machine learning techniques to find a representative subset of coastal profiles. Using the Dutch coast as a case study, they began with 1,430 elevation profiles located along the coastline. The new method reduced this to only 100, termed typological coastal profiles (TCPs). This represents a 93% reduction in input to the numerical impact models, substantially improving the computational efficiency of an impact analysis over this area. Predictions of dune erosion based on these TCPs still showed good prediction skill despite their reduction in numbers, proving they are representative of the Dutch coast as a whole. Indeed, 47 of these profiles already represent around 90% of the total, with large stretches of the Wadden and Holland coasts being adequately represented by a small number of TCPs. These findings prove the viability of dune erosion predictions over large stretches of coastline, which can play a key role in adapting to and mitigating the impact of damaging storm events.
Methods
This study used a novel clustering method to find a representative subset of coastal profiles. These profiles were characterised based on their local morphology and hydrodynamics. Pre-storm beach geometries, nearshore slopes, local tidal amplitude, and profile orientations were found to be the most important drivers of dune erosion variability. Input reduction techniques like this are common in coastal applications, however they have so far focussed only on ocean forcings such as wave conditions. This is the first application of this technique on measured beach elevation profiles. This subset was then used to produce dune erosion simulations based on observed conditions during various historical storms. These were compared with a benchmark set of simulations using the full set of profiles to assess the representativeness of the subset.
Policy relevance
These results prove the validity of the authors’ new methodology, allowing dune erosion forecasts of long stretches of coastline to be made ahead of the forecast arrival of a storm or other damaging event. This helps communities and policymakers prepare for the impact of such an event, including putting pre-emptive mitigation measures in place.
Abstract
Dune erosion driven by extreme marine storms can damage local infrastructure or ecosystems and affect the long-term flood safety of the hinterland. These storms typically affect long stretches (∼100 km) of sandy coastlines with variable topo-bathymetries. The large spatial scale makes it computationally challenging for process-based morphological models to be used for predicting dune erosion in early warning systems or probabilistic assessments. To alleviate this, we take a first step to enable efficient estimation of dune erosion using the Dutch coast as a case study, due to the availability of a large topo-bathymetric dataset. Using clustering techniques, we reduce 1,430 elevation profiles in this dataset to a set of typological coastal profiles (TCPs), that can be employed to represent dune erosion dynamics along the whole coast. To do so, we use the topo-bathymetric profiles and historic offshore wave and water level conditions, along with simulations of dune erosion for a number of representative storms to characterize each profile. First, we identify the most important drivers of dune erosion variability at the Dutch coast, which are identified as the pre-storm beach geometry, nearshore slope, tidal level and profile orientation. Then using clustering methods, we produce various sets of TCPs, and we test how well they represent dune morphodynamics by cross-validation on the basis of a benchmark set of dune erosion simulations. We find good prediction skill (0.83) with 100 TCPs, representing a 93% input and associated computational costs reduction. These TCPs can be used in a probabilistic model forced with a range of offshore storm conditions, enabling national scale coastal risk assessments. Additionally, the presented techniques could be used in a global context, utilizing elevation data from diverse sandy coastlines to obtain a first order prediction of dune erosion around the world.
Funded by the European Union under Horizon 2020.