Propagating uncertainty from physical and biogeochemical drivers through to top predators in dynamic Bayesian ecosystem models improves predictions

With the global rapid expansion of offshore renewable energies, there is an urgent need to assess and predict effects on marine species, habitats, and ecosystem functioning. Doing so will require dynamic, multitrophic, ecosystem-centric approaches coupled with oceanographic models that can allow for physical and/or biogeochemical indicators of marine ecosystem change to be included. However, in such coupled approaches, indicators carry uncertainties that can propagate and affect species higher up the trophic chain. Dynamic Bayesian networks (DBNs) are pragmatic approaches that probabilistically represent ecosystem-level interactions. They allow for uncertainties to be better estimated than mechanistic models that only account for expected values. In this study, we calculated variance as a measure of uncertainty from selected indicators and used them to build DBN models. A hidden variable was incorporated to model functional ecosystem change, where the underlying interactions dramatically change, following a disturbance. We wanted to assess whether propagating uncertainty into the modelling process affects the predictive accuracy of the models in the context of reconstructing the time series of the ecosystem dynamics. Model accuracy was improved for 60 % of the species once variance was added. The models were better in capturing the temporal inter-annual variability, once variance was calculated with a rolling window approach. The hidden variable successfully modelled previously identified ecosystem changes, however, now with the added uncertainty, the changes that implicated the ecosystem state were identified earlier in the time series. The results indicate that using DBNs is highly valuable as it gains accuracy with the addition of uncertainty.