Species distribution models (SDMs) used to make spatial predictions of habitat suitability or species occurrence are widely applied in conservation assessment, impact analysis, and projecting the impacts of climate change and invasive species.  Predictors represent environmental gradients and landscape variables that are distally or proximally related to habitat suitability or species’ realized niches.  Considering the increasingly widespread use of these models, there is a need to establish a framework to guide the operational use of these methods for biodiversity assessment and landscape management.  Therefore, the objective of our study was to explore the relative effects of species’ biogeographical, ecological, and functional traits on prediction accuracy.  We hypothesized that species’ traits would determine the accuracy of SDMs to a greater extent than the modeling method used.

Results/Conclusions

For 45 plant species in the southwestern ecoregion of California, we found that the accuracy (AUC) of SDMs varied to a much greater degree among species (61% of variance in model accuracy explained by species) than among types of models (comparing logistic regression, classification trees, generalized additive models, Random Forests) (18% of variance explained by model type).  While the average predictive accuracy of the models was similar for Random Forests, generalized additive models, and logistic regression (AUC ranging from 0.78 – 0.81), it was substantially lower for classification trees (average AUC of 0.69).The effect of species on model accuracy primarily reflected differences in rarity, life form and disturbance response. When species were grouped into rarity classes, life forms and disturbance functional types, regression models showed that both of these variables explained significant variation in model accuracy.  Accuracy was significantly higher for species classified as either rare or habitat specialists than for species classified as common or widespread sparse.  These differences show that it is more difficult to discriminate suitable from unsuitable habitat for habitat generalists.  Nevertheless, developing distribution models for rare species is critical for conservation planning and environmental impact analysis. Models also performed better for shrubs and succulents than for sub-shrubs and perennial herbs.  The functional type with the best accuracy was obligate seeder chaparral shrubs that regenerate via fire-cued seed germination from a dormant seed bank.  Obligate seeder shrubs have low dispersal distances and poor recruitment between fires, and this site fidelity may explain higher distribution model accuracy.  Our framework for interpreting SDM results with respect to species traits and model types will improve their usefulness in landscape-scale conservation planning and large-scale impact analysis.