Selecting predictors to maximize the transferability of species distribution models: lessons from cross-continental plant invasions

Résumé 0

Aim: Niche-based models of species distribution (SDMs) are commonly used to predict impacts of global change on biodiversity but the reliability of these predictions in space and time depends on their transferability. We tested how the strategy to choose predictors impacts the SDMs' transferability at a cross-continental scale. Location: North America, Eurasia and Australia Method: We used a systematic approach including 50 Holarctic plant invaders and 27 initial predictor variables, considering 10 different strategies to variable selection, accounting for predictors' proximality, multicollinearity and climate analogy. We compared the average performance per strategy, some of them using a large number of random predictor combinations. Next, we looked for the single best model for each species across all possible predictor combinations, by pooling models across all strategies. Transferability was considered as the predictive success of SDMs calibrated in native range and projected onto the invaded range. Results: Two strategies showed better SDMs' transferability on average: a set of predictors known for their ecologically-meaningful effects on plant distribution, and the two first axes of a principal component analysis calibrated on all predictor variables (Spc2). From the >2000 combinations of predictors per species across strategies, the best set of predictors yielded SDMs with good transferability for 45 species (90%). These best combinations consisted in a random selection of 8 predictors (45 sp) and in Spc2 (5 sp). We also found that internal cross-validation was not sufficient to fully inform about SDMs' transferability to a distinct range. Main conclusion: Transferring SDMs at the macroclimatic scale, and thus anticipating invasions, is possible for the large majority of invasive plants considered in this study, but the predictions' accuracy relies strongly on the choice of predictors. From our results, we recommend including either the state-of-the-art proximal variables or a reduced and orthogonalised set to obtain robust SDMs' projections.