Search

The ability to understand and predict storm erosion along sandy coastlines is important for safely and sustainably managing these coastlines now and in the future. This thesis investigates the use of data-driven Bayesian networks to assist coastal scientists, engineers and managers to better understand and predict sandy coastline storm erosion using large observational datasets acquired by coastal remote sensing technologies. A methodological study was undertaken using a 10-year dataset of coastal storm events extracted from Argus coastal imaging technology. It was found that Bayesian networks have key advantages for modelling storm erosion including the illumination of causality, uncertainty quantification, and low computational cost. However, these advantages can be limited by the data requirements of the approach. It was determined that, based on data availability and the modelling objective, Bayesian networks can be applied to descriptive or predictive applications. Descriptive networks can be useful for exploring the physical relationships between variables within a specific dataset. In contrast, predictive networks identify general relationships in a dataset that can be used to predict unseen data. A descriptive Bayesian network was used to investigate controls of spatial variability in storm erosion of the berm and dune based on detailed storm erosion observations spanning a 400 km region of the southeast Australian coastline. It was found that spatial variability in storm erosion was driven by both the antecedent morphology of the coastline and hydrodynamic forcing conditions of the storm event, and that controls of erosion of the berm and dune were different. A predictive Bayesian network was developed using multiple large storm erosion datasets from three different global locations. Results showed that while Bayesian networks developed at one specific location poorly predicted the storm erosion response at differing locations, a Bayesian network developed on all three datasets generalised storm erosion response well, predicting changes to the shoreline, dune toe and dune crest with accuracies between 61 – 76% when tested on unseen data. Importantly, similar variables were important in the Bayesian network for predicting storm erosion at the different study regions and included descriptors of the antecedent beach morphology and hydrodynamics forcing of the storm event.