Testing the skill of a species distribution model using a 21st century virtual ecosystem

Plankton communities play an important role in marine food webs, in biogeochemical cycling, and in Earth's climate; yet observations are sparse, and predictions of how they might respond to climate change vary. Correlative species distribution models (SDM's) have been applied to predicting biogeography based on relationships to observed environmental variables. To investigate sources of uncertainty, we use a correlative SDM to predict the plankton biogeography of a 21st century marine ecosystem model (Darwin). Darwin output is sampled to mimic historical ocean observations, and the SDM is trained using generalized additive models. We find that predictive skill varies across test cases, and between functional groups, with errors that are more attributable to spatiotemporal sampling bias than sample size. End-of-century predictions are poor, limited by changes in target-predictor relationships over time. Our findings illustrate the fundamental challenges faced by empirical models in using limited observational data to predict complex, dynamic systems.