Print PagePredictive modeling of species-environment relationships can enhance our understanding of climate change effects on biodiversity, assist in assessment of invasion risk by exotic organisms and inform basic research on individual species. However, predictions resulting from models may be sensitive to the initial conditions from which models are trained. Although forecast models of climate change effects on species require that models be trained using data on contemporary climate and many alternative datasets describing contemporary climate are available, we lack systematic comparisons of model performance and predictions across alternative climate datasets used for model training. Here, we seek to fill that gap by exploring how model performance and spatial predictions vary using a standardized suite of occurrence data and predictor climate variables obtained from two global datasets describing contemporary climate. We investigate differences in performance and predictions across six ecologically-distinct terrestrial vertebrate species and four modeling algorithms.
Results/Conclusions
Although variation in performance between climate datasets was minor compared with variation among species and algorithms, the average spatial correlation between predictions made with the two datasets was low considering the equivalent data inputs used in model construction (r = 0.676, SD = 0.288 for contemporary conditions and r = 0.615 ± 0.345 for future conditions). The discrepancy between prediction maps created using the two climate datasets was driven by differences in estimated temperature in grid cells occupied by modeled species. We suggest that the inconsistency in spatial predictions made using alternative climate datasets may contribute heretofore unconsidered uncertainty into predictive models of climate change effects that may be best captured using an ensemble approach (e.g., by averaging observations from multiple datasets) to characterize climate data inputs.
