The effect of pseudo-absence selection method on transferability of species distribution models in the context of non-adaptive niche shift

Abstract

Transferability of species distribution models (SDMs) is key to predicting invasion patterns and can be challenged if niche shift occurs in the invaded range. When using native occurrences to estimate potential invasions with presence-only modeling methods, it is important to constrain the pseudo-absence (PA) sampling to the species’ native range. However, some studies including highly cited ones, do not follow this approach to selecting PA samples. In this research, we addressed two questions using an invasive species in the United States (U.S.), kudzu bug (Megacopta cribraria): 1) is model transferability challenged by a non-adaptive niche shift? and 2) is model performance affected by use of PA samples from outside the native range of the species? Kudzu bug is native to Asia, with recently observed non-adaptive niche shift in the U.S. To answer the first question, we quantified the environmental space anisotropy and non-adaptive niche change, and then evaluated the performances of seven SDMs. To answer the second question, we further compared the interpolation and transferability of seven SDMs trained with PAs from the native range and from both native and invaded ranges. We confirmed that the environmental space anisotropy (P = 0.01) and non-adaptive niche change (P = 0.01) are both statistically significant. Of the seven SDMs used, four models had transferability indices higher than 0.9. Boosted regression tree and random forests both had good interpolation and transferability (AUC>0.80 and kappa>0.60), whereas three other models showed good interpolation and fair transferability (AUC>0.70 and kappa>0.40). Inclusion of pseudo-absences from the invaded range significantly increased the interpolation (P < 0.001) but decreased the transferability (P < 0.01) of almost all models. Our findings suggest that SDMs can show good transferability with non-adaptive niche shift, thus native occurrence information should be used in similar situation. We confirmed that it is crucial to constrain the PAs to the same spatial range as presences to accurately model potential invasions.