Beyond static species distribution models: temporally dynamic, spatially nested models reduce extrapolation for future projections

Camila Rocabado , Derek Corcoran, Trine Bilde, Samuel Gascoigne , Signe Normand, Marjo Saastamoinen, Santiago Jose Elias Velazco, Damaris Zurell, Roslyn Henry, Greta Bocedi

Species distribution models are widely used to project biodiversity responses to climate change, but geographically restricted training data can truncate species’ climatic niches, increasing reliance on extrapolation. Spatially nested models can reduce this truncation by incorporating climatic information from species’ wider distributions. However, how accounting for interannual environmental variation rather than temporal averages affects extrapolation remains unclear. Using occurrence records for 28 butterfly species in Denmark, we compared four approaches combining regional versus spatially nested training data with static (long-term average) versus dynamic (annual) predictors. Regional models used climatic information for Denmark, whereas nested models incorporated European-scale climatic suitability. Static models used temporally averaged predictors, while dynamic models matched observations to year-specific conditions. We assessed climatic niche coverage, model’s predictive performance under spatial block cross-validation and a common forward temporal holdout, and future climatic conditions outside the training domain. Regional training excluded ~ 30% of the climatic space occupied across species’ European distributions. Temporally averaged predictors captured nearly all climatic space represented by annual conditions across Europe, but only ~ 37% within Denmark. In the forward evaluation for 2010–2018, dynamic models outperformed static models. Under spatial cross-validation, static regional models outperformed dynamic regional models, whereas static and dynamic nested models did not differ significantly. Spatial nesting and temporal matching substantially reduced future extrapolation, with their combination consistently requiring the least extrapolation. Model choice should therefore reflect the intended application. Static models may be adequate for mapping contemporary suitability within well-represented environmental conditions when supported by spatial validation. For projections under environmental change, broadening training domain across space and time reduces extrapolation, even without improving spatial discrimination. Predictive discrimination and environmental support for transfer should be assessed jointly, with temporal validation and extrapolation risk receiving particular attention for climate-change projections.