Species distribution models (SDMs) have become central tools in ecology and
biogeography. Although they can be fitted with different types of species data (e.g.
presence-absence, abundance), the most common approach, based on data from
large species repositories, is to use simple occurrences (i.e. presence-only) combined
with background points (BP; also called pseudo-absences). But how should we sample
these background points, and how does this choice affect SDMs? In most studies so
far, BP were sampled randomly in geographic space, yet theory rather suggests, if a
species is at equilibrium, that it is better to sample them in a stratified way in
environmental space. However, this potential improvement of SDM predictions has
never been tested. Furthermore, a typical assumption behind SDMs is that the
modelled species are at equilibrium with their environment. But how do these models
perform when species are in disequilibrium, as is the case for most invasive species?
To answer these questions, we selected 30 different species (10 insects, 10 mammals
and 10 plants; for each group 5 were invasive and 5 were considered at equilibrium)
and for each we calibrated SDMs with different types of background selections:
random in environmental space, random-stratified in environmental space, random in
geographic space, and random-stratified in geographic space. For each SDM we
assessed both predictive performance using standard metrics and their stability using a
new approach that compares the model’s habitat suitability projection with those of a
SDM calibrated with virtual occurrence data generated from the most suitable areas.
Finally, we compared the predictive performance of species distribution models of
invasive alien (disequilibrium) species versus native (equilibrium) species by
comparing model stability and performance metrics of the two groups. We found that
sampling BP in a stratified-random way in environmental space yields the highest
performance metrics, and that sampling fully randomly in environmental space yields
the most stable models. This has implications for the use of SDMs in conservation, as
the classical and frequently used fully random in geographic space BP are found to
produce both less accurate and less stable models. Our results indicate that the best
approach is to use stratified random in environmental space BP sampling if accuracy is
essential, and fully random in environmental space BP sampling if model stability is
essential.