Integrating Bayesian Networks to Forecast Sea-Level Rise Impacts on Barrier Island Characteristics and Habitat Availability

Evaluation of sea-level rise (SLR) impacts on coastal landforms and habitats is a persistent need for informing coastal planning and management, including policy decisions, particularly those that balance human interests and habitat protection throughout the coastal zone. Bayesian networks (BNs) are used to model barrier island change under different SLR scenarios that are relevant to management and policy decisions. BNs utilized here include a shoreline change model and two models of barrier island biogeomorphological evolution at different scales (50 and 5 m). These BNs were then linked to another BN to predict habitat availability for piping plovers (Charadrius melodus), a threatened shorebird reliant on beach habitats. We evaluated the performance of the two linked geomorphology BNs and further examined error rates by generating hindcasts of barrier island geomorphology and habitat availability for 2014 conditions. Geomorphology hindcasts revealed that model error declined with a greater number of known inputs, with error rates reaching 55% when multiple outputs were hindcast simultaneously. We also found that, although error in predictions of piping plover nest presence/absence increased when outputs from the geomorphology BNs were used as inputs in the piping plover habitat BN, the maximum error rate for piping plover habitat suitability in the fully-linked BNs was only 30%. Our findings suggest this approach may be useful for guiding scenario-based evaluations where known inputs can be used to constrain variables that produce higher uncertainty for morphological predictions. Overall, the approach demonstrates a way to assimilate data and model structures with uncertainty to produce forecasts to inform coastal planning and management. Plain Language Summary Bayesian networks provide a means for estimating the probability of various outcomes that depend on input conditions that also have a range of probabilities. They can, for example, predict the probability that a specific location is a suitable nesting site for shorebirds if provided information about the site. The information required can be obtained from mapping on-the-ground habitat conditions, or as is done here, be predicted from another BN. The methods for doing that, and the additional uncertainties introduced by using multiple networks, are discussed here.