Combining Numerical and Statistical Models to Predict Storm‐Induced Dune Erosion

Abstract

Dune erosion is an important aspect to consider when assessing coastal flood risk, as dune elevation loss makes the protected areas more susceptible to flooding. However, most advanced dune erosion numerical models are computationally expensive, which hinders their application in early-warning systems. Based on a combination of probabilistic and process-based numerical modeling, we develop an efficient statistical tool to predict dune erosion during storms. The analysis focuses on Dauphin Island, AL, in the northern Gulf of Mexico, where we combine synthetic sea storms with a calibrated and validated XBeach model to develop and test a range of different surrogate models for their ability to predict barrier island geometric parameters under storm conditions. Surrogate models are developed by combining the oceanographic forcing from 100 optimally sampled sea storm events covering the entire multivariate parameter space (used as XBeach input) and associated changes in the dune system (XBeach output). We test four surrogate models using a k-fold approach for validation. All models perform well in predicting changes in dune elevation, barrier island area, and width but are less accurate in predicting alterations in the cross-shore locations of dune morphological features. Multivariate adaptive regression splines is identified as the best surrogate model based on its fast development and good performance, attaining a modified Mielke index of 0.81 for dune crest height. As demonstrated at Dauphin Island, our approach shows potential to be used in an operational framework to predict dune response (in particular crest elevation change) when water level and wave forecasts are available. Key Points: • Surrogate models of XBeach are developed and tested, showing that multivariate adaptive regression splines exhibits the best performance • Carefully trained statistical and machine-learning models predict changes in barrier island geometric parameters at low computational cost • The highest performance is attained for the most extreme dune impact regimes, allowing fast and accurate dune height change predictions Correspondence to:Here, surrogate models are designed to predict coastal dune erosion by addressing issues related to computational performance and absence of sufficient forcing data. Our main goal is to understand complex morphological interactions through simulation of a wide range of realistic scenarios. To accomplish this, we develop and test surrogate models for XBeach that are trained with simulated, physically consistent oceanographic variables and associated dune response. This includes the following steps: (1) select a subset of synthetic multivariate sea storm events from Wahl et al. (2016) and consider time-dependent evolution of the oceanographic variables, (2) use XBeach to predict morphological response parameters for the selected sea storms, and (3) develop and test a range of different surrogate models to mimic XBeach at low computational cost. The different analysis steps are summarized in Figure 1. In...

A Clustering Approach for Predicting Dune Morphodynamic Response to Storms Using Typological Coastal Profiles: A Case Study at the Dutch Coast

“…A skill value of 1 indicates a perfect prediction. The bias, RMSE and the modified index of Mielke (λ) (Duveiller et al, 2016;Santos et al, 2019) were also explored as performance metrics. The predictive capabilities of the model were assessed using the previously described performance metrics (Sections "Clustering Sensitivity Analysis" and "Applying the Typological Coastal Profiles").…”

“…With respect to the clustering of the profiles and defining the TCPs, various sensitivities were tested in the study herein, mainly focusing on (1) the parameters used, (2) their weights, (3) the clustering algorithm and (4) the matching technique. With respect to the parameters used, the extraction of parameters per profile was based on data availability and previous studies that explored their importance in driving dune erosion (Beuzen et al, 2019;Cohn et al, 2019;Santos et al, 2019). Undoubtedly, there are parameters that were not studied here and could be of importance for simulating the dune erosion more accurately and obtaining a better grouping of the profiles and thus more representative TCPs.…”

“…Additionally, in some other cases dune toe detection was not consistent in the prestorm and post-storm profiles (i.e., erroneous detection of dune toe in post-storm profile) leading to inconsistencies in calculating the relative change, which could potentially introduce artificial inhomogeneities during the clustering phase. It would be expected tough, that for locations with barriers islands and lower dune features, other indicators like dune crest lowering or overwash volumes might be of more importance (Santos et al, 2019).…”

“…In recent years, predictive tools or meta-models that employ large number of observations or usually simulations along with statistical techniques (e.g., Bayesian Networks), have shown potential for producing efficient predictions of coastal hazard indicators (Gutierrez et al, 2011;Poelhekke et al, 2016;Beuzen et al, 2017;Pearson et al, 2017;Giardino et al, 2019;Rueda et al, 2019;Santos et al, 2019;Sanuy and Jiménez, 2021). The concept underlying these meta-models is the use of a large number of combinations of forcing conditions and/or coastal elevation profiles to predict coastal response.…”

Machine Learning Predicts Reach‐Scale Channel Types From Coarse‐Scale Geospatial Data in a Large River Basin

“…However, with increasing volume of data, ease of access to computing resources and availability of ML toolboxes, this situation is rapidly evolving. In geomorphology in particular, novel ML applications include delineating landforms (Bugnicourt et al, 2018), predicting geomorphic disturbance (Perry & Dickson, 2018) or dune erosion (Santos et al, 2019), mapping susceptibility to landslide and gully erosion (Lee et al, 2018;Pham et al, 2018;Rahmati et al, 2017), inferring ecohydrological parameters (Bassiouni et al, 2018), analyzing model residuals (Hassan et al, 2018), clustering river profiles (Clubb et al, 2019), classifying and predicting sediment-discharge relationships (Hamshaw et al, 2018;Vaughan et al, 2017), and assessing stream diversity with large-scale top-down approaches (Beechie & Imaki, 2014;McManamay et al, 2018). In the face of such rising popularity, the interpretability and assessment of uncertainty in ML models remain key issues (e.g., Reichstein et al, 2019).…”