Predicting Coastal Hazards with a Bayesian Network

Abstract

Recent and historic events have demonstrated the European vulnerability to coastal floods. Larger and more extreme events in Asia and the Americas have shown the devastating effects that these low-frequency high-impact floods can have. This thesis contributes to the development of an Early Warning System (EWS) for marine coastal hazards. The relevance of an EWS is supported by the UN who have identified it as key in reducing casualties and economic losses due to flood events (UNISDR, 2002). For coastlines with sandy shores (beaches and dunes) the response of the coastline to high impact events is very large (e.g. dune and beach erosion). Current flood hazard models do not include these morphological processes and are therefore not sufficient for these types of coastlines. Process based models such as XBeach (Roelvink et al., 2009) are capable of modeling the coastal response but are not very useful as an EWS due to the long duration of the computations. A solution is found in utilizing a probabilistic model as a surrogate for a process based model. The probabilistic model is fed with data that is created with the process based model. This way the probabilistic model gains the same knowledge about the processes within the bounds that it is trained. More specifically a Bayesian Network (BN) is used as a surrogate for an XBeach model. As a case study site for the development and initial implementation the beach settlement of Faro, Praia de Faro, has been selected. Praia de Faro is pestered by yearly recurring overwash events and has experienced damages to houses and infrastructure due to severe erosion of the coastline. To be able to feed the BN with data a large storm dataset is needed. This dataset is created by using a local storm dataset to create bivariate distributions using copulas. These bivariate distributions are then used to sample a synthetic dataset that mimics the characteristics of the original dataset. The final BN is able to translate offshore storm boundary conditions to onshore hazard intensities. It can, however, only do this with relatively low complexity. For it to be useful in an operational EWS it will either need extra training data or increased complexity which implicitly also means more training data is needed.