Potential and limitations of Bayesian networks for understanding shoreline mobility: an example in La Réunion island
Abstract
Coastal erosion is a global growing threat as more and more human activities and settlements concentrate on the coastal fringe. Today's climate change and induced sea level rise also contribute to change the risk of erosion. Understanding the current shoreline evolution is a necessary step to predict future changes and better manage this risk. The shoreline mobility results from numerous factors and complex mechanisms acting at different spatial and temporal scales. Here, a data mining approach based on a Bayesian network (BN) is tested enabling one to use readily available data to analyse some causes of decadal-scale shoreline evolution in La Réunion (a volcanic tropical island in the Indian Ocean) and to reproduce the observed evolution trends. The BN is built to define causal relationships between 5 variables describing the state of a given coastal segment: geomorphic settings, exposure to energetic waves, presence of an estuary (importance of continental sediment loads), presence of human works in the vicinity of the segment and current shoreline mobility (accretion, stability or erosion, representative of about 30 years). The retrospective predictions are correct in 79% of the cases. Evaluation of the model performance using log likelihood ratio scores indicates that the BN provides shoreline mobility predictions that are better than the prior probability. By evaluating the model behaviour using from one to four variables, the geomorphic settings are identified as the most important model parameter determining coastal evolution trends. Incorrect predictions of the BN are analysed in details and experts' know-how is used to assess the local causes of the observed mobility and to point out limits of the BN. Among the multiple causes of mis-prediction, the lack of sediment budget information (alongshore transport and interactions between adjacent coastal segments) is the most common.