Predicting coastal hazards for sandy coasts with a Bayesian Network Laurens Poelhekke a,b,c, ⁎, Wiebke S. Jäger b , Ap van Dongeren a , Theocharis A. Plomaritis d , Robert McCall a , Óscar Ferreira d a Deltares, Department of Applied Morphodynamics, Delft, The Netherlands b Delft University of Technology, Faculty of Civil Engineering and Geosciences, Department of Hydraulic Engineering, Delft, The Netherlands c CDR International, Amersfoort, The Netherlands d University of the Algarve, CIMA, Faro, Portugal abstract article info Article history: Received 11 April 2016 Received in revised form 16 August 2016 Accepted 23 August 2016 Available online xxxx Low frequency, high impact storm events can have large impacts on sandy coasts. The physical processes governing these impacts are complex because of the feedback between the hydrodynamics of surges and waves, sediment transport and morphological change. Predicting these coastal changes using a numerical model requires a large amount of computational time, which in the case of an operational prediction for the pur- pose of Early Warning is not available. For this reason morphodynamic predictions are not commonly included in Early Warning Systems (EWSs). However, omitting these physical processes in an EWS may lead to potential under or over estimation of the impact of a storm event. To solve this problem, a method has been developed to construct a probabilistic Bayesian Network (BN). This BN connects three elements: offshore hydraulic boundary conditions, characteristics of the coastal zone, and onshore hazards, such as erosion and overwash depths and velocities. The hydraulic boundary conditions are derived at a water depth of approximately 20 m from a statistical analysis of observed data using copulas, and site character- istics are obtained from measurements. This BN is trained using output data from many pre-computed process- based model simulations, which connect the three elements. Once trained, the response of the BN is instanta- neous and can be used as a surrogate for a process-based model in an EWS in which the BN can be updated with an observation of the hydraulic boundary conditions to give a prediction for onshore hazards. The method was applied to Praia de Faro, Portugal, a low-lying urbanised barrier island, which is subject to fre- quent flooding. Using a copula-based statistical analysis, which preserves the natural variability of the observa- tions, a synthetic dataset containing 100 events was created, based on 20 years of observations, but extended to return periods of significant wave height of up to 50 years. These events were transformed from offshore to onshore using a 2D XBeach (Roelvink et al., 2009) model. Three BN configurations were constructed, of which the best performing one was able to predict onshore hazards as computed by the model with an accuracy ranging from 81% to 88% and predict events with no significant onshore hazards with an accuracy ranging from 90% to 95%. Two examples are presented on the use of a BN in operational predictions or as an analysis tool. The added value of this method is that it can be applied to many coastal sites: (1) limited observations of offshore hydrodynamic parameters can be extended using the copula method which retains the original observations' natural variability, (2) the transformation from offshore observations to onshore hazards can be computed with any preferred coastal model and (3) a BN can be adjusted to fit any relevant connections between offshore hydraulic boundary conditions and onshore hazards. Furthermore, a BN can be continuously updated with new information and expanded to include different morphological conditions or risk reduction measures. As such, it is a promising extension of existing EWSs and as a planning tool for coastal managers. © 2016 Elsevier B.V. All rights reserved. Keywords: Early Warning System Bayesian Network Sandy coasts XBeach Probabilistic Hazards 1. Introduction Over the past decades a number of large storm events have demon- strated the vulnerability of the coastal zone in Europe, such as to the North Sea Flood of 1953 in the Netherlands, Belgium and the United Kingdom (Gerritsen, 2005), Xynthia (2010) affecting the entire coast of south-western Europe (Bertin et al., 2012) and Hercules (2014) caus- ing severe coastal erosion and flooding in parts of France and the United Kingdom (Masselink et al., 2015a) among many others. Larger and more extreme events such as the hurricanes in the USA, e.g. Katrina in 2005 (Knabb et al., 2006) and Sandy in 2012 (Blake et al., 2013), and Typhoons in Asia, e.g. Haiyan in 2013 and Nargis in 2008, have also shown the devastating effects of these low-frequency, high impact flood events. Coastal Engineering 118 (2016) 21–34 ⁎ Corresponding author at: CDR International, Amersfoort, The Netherlands. E-mail address: l.poelhekke@cdr-international.nl (L. Poelhekke). http://dx.doi.org/10.1016/j.coastaleng.2016.08.011 0378-3839/© 2016 Elsevier B.V. All rights reserved. Contents lists available at ScienceDirect Coastal Engineering journal homepage: www.elsevier.com/locate/coastaleng