Species distribution models are commonly used to project the ranges of species of concern into the future, but unlike within-era projections to the same region, their accuracy cannot be assessed because of lack of data from the future. One method for testing projections across time is to project ranges based on data collected from the past to the present and assess model performance using contemporary records. Here, we assess the ability of SDMs to project ranges accurately by comparing projections trained on historical museum records of vertebrates collected by Joseph Grinnell and his colleagues between 1900 and 1940 and resurveys of these same sites since 2003.
Results/Conclusions
We compare several common modeling techniques that have performed well when used for modeling the same era and region from which data are drawn, and we assess whether high within-era performance predicts cross-era performance and under what circumstances models perform poorly. Although within-era model performance can be high and in general is a good predictor of cross-era performance, good performance can be due to overfitting in some cases. Thus, projections to different eras can suffer vis-à-vis models that perform poorly within-era but generate smoother response functions.