Abstract
How the temporal structure of the environment should be represented is a central question when describing and predicting dynamic marine ecosystems. We examine this through plankton beta-diversity in an advective marine ecosystem, asking how predictor design, specifically the temporal representation of environmental predictors alongside the formulation of spatial separation, influences the performance of dissimilarity-based models. Using daily simulations for 2019 in the coastal and shelf seas of mid-Norway, we generated communities of diatoms, flagellates, ciliates, bacteria, and heterotrophic nanoflagellates with a coupled hydrodynamic–biogeochemical ocean model (SINMOD), and calculated pairwise Bray–Curtis dissimilarity among sampled sites. Beta-diversity was modelled in a Bayesian spatial generalised dissimilarity mixed modelling framework (spGDMM) to compare alternative predictor formulations within a common probabilistic structure.
We compared temporally aligned environmental predictors with annual-mean summaries, and evaluated spatial terms based on Euclidean distance and a transport-informed connectivity metric derived from Lagrangian particle simulations. Predictive performance was assessed using spatial hold-out validation and probabilistic scoring rules. Time-resolved predictors consistently improved predictive skill and reduced worst-case errors relative to annual means, with the advantage increasing as additional environmental predictors were included. Spatial separation terms produced smaller but consistent gains, and the connectivity metric outperformed Euclidean distance. These results suggest that, in dynamic marine systems, predictive performance depends strongly on representing the temporal structure of environmental forcing, while transport-informed spatial terms provide secondary improvements. In such dynamic systems, predictor design should therefore prioritise temporal representation over further refinement of spatial separation terms.