Coastal erosion is a worldwide societal issue and problems associated with it are expected to worsen as sea levels increase due to global climate change and the impacts of storms become more severe. With the recognition of the hazards facing coastal development, there is growing demand for predictive models that can be used to forecast where coastal erosion hazards are highest. Existing models that forecast coastal cliff response to sea-level rise or high water levels (i.e. storm surge or swell) include geometric models such as a modified Bruun Rule, empirical models based on historical water level data, or more simply, hazard zones can be delineated by the forward projection of historic rates. These methods provide deterministic predictions but often don’t account for the spatial and temporal variability of cliff retreat processes, or for the fact that cliff failure is episodic and does not necessarily respond instantaneously to forcing conditions. Furthermore, the response may depend on the influence of previous events. Incorporating probabilistic methods may be more appropriate to account for the complexity of coastal cliff retreat.
A Bayesian network is constructed for Southern California using long-term (70-yr) cliff retreat rates, cliff slope and height, and material strength and wave impact information. The goal of the study is to use the network as a diagnostic tool to predict the probability and distribution of coastal cliff retreat over a three-year period along the Southern California coast. The performance and accuracy of the Bayesian network can be evaluated for this case using a dataset of measured cliff retreat from the same three year period. The Bayes approach is well-suited because it can take advantage of correlations between key variables that influence the cliff-retreat processes. For the three year test period, the Bayesian network accurately predicts the correct outcome in 80 - 90% of the study area. The results also indicate that information about prior behavior is crucial for accurately predicting cliff response. In this study, without the long-term historical cliff-retreat rate, which represented the prior behavior, the model performed poorly.