Impact Assessment of Extreme Storm Events Using a Bayesian Network

Abstract

The most densely populated and economical most valuable areas in The Netherlands lie below mean sea level. These areas are protected against the sea by a coastal dune system. The vital importance of this dune system is reflected in the extensive collection of Dutch legal regulations that ensure the safety level of the dunes. Current safety assessment method for dunes, prescribed by these regulations, is based on conducting large numbers (thousands) of simulations to estimate dune erosion at individual locations along the Dutch coast. With the use of more complex dune-erosion models (i.e. Xbeach; Roelvink et al. (2009)) in the safety assessment, the method gets computationally more intensive. This means that conducting large numbers of simulations for a dune safety assessment are not feasible. Therefore new probabilistic approaches (e.g., Bayesian approach) are needed in order to apply state-of-the-art insights and models for dune erosion in a safety assessment. The aim of this study was to gain an insight in the applicability of the Bayesian network approach for dune safety assessment against extreme storm events on the Dutch coast. First a database was generated that serves as input to the Bayesian network. The content of this database is obtained in a way similar to current assessment method. This means, data-sources (wave conditions, bottom profiles) and the dune-erosion model (duros+; Vellinga (1986), van Gent et al. (2008)) to simulate the dune erosion process were equal. Qualitatively the Bayesian network is represented by nodes (variables) and arrows (relations). Variables and relations were selected such, they were physically related to the dune erosion process and could be obtained out of the database. Quantitatively the Bayesian network is described by a conditional probability table defined by a expectation-maximization training algorithm. For this training process, cases (a case is a record in the database and represents a storm event) out of the dataset were selected. Sensitivity analysis of the required number of training data, as well the determining variable of this Bayesian network were made. Results show the Bayesian network is capable of reproducing the dune erosion process (given this set of data) for 89%. The loss in skill is a consequence of the discretization of the variables in the Bayesian network. The number of training-cases needed to make reliable predictions that are more accurate then predictions made with the use of the prior probability, is approximately 5,000 cases. Results of the sensitivity analysis indicates water level information is a determining term regarding the prediction skill. However, when other hydraulic conditions (wave height, peak period) were known the water level was redundant. Indicating water level information is captured by the variables wave height and/or peak period. Dune erosion volumes predictions for locations at the Holland coast, shows a high amount of uncertainty and results in unreliable predictions (negative log-likelihood ratio) when using a Bayesian network trained on cases representing the Wadden coast. Indicating coastal features between both coastal regions are too diverse. Furthermore, a Bayesian network is a useful tool to improve insight into data. Expected relations between variables can be investigated and visualized in the Bayesian network immediately.