Projects per year
Species distribution models (SDMs) frequently project substantial declines in the spatial extent of climatically suitable habitat in response to scenarios of future climate change. Such projections are highly disconcerting. Yet, considerable variation can occur in the direction and magnitude of range changes projected by different SDM methods, even when predictive performance is similar. In this study, we assessed whether particular methods have a tendency to predict substantial loss or gain of suitable habitat. In particular, we asked, “are 14 SDM methods equally likely to predict extreme changes to the future extent of suitable habitat for 220 Australian mammal species?”. We defined five non-mutually exclusive categories of ‘extreme’ change, based on stability or loss of current habitat, or the dislocation of current and future habitat: a) no future habitat (range extinction); b) low stability of current habitat (≤10% remains); c) no gain of habitat in new locations; d) all future habitat is in new locations (i.e. completely displaced from current habitat); and e) substantial increase in size of habitat (future habitat is ≥100% larger than current). We found that some SDM methods were significantly more likely than others to predict extreme changes. In particular, distance-based models were significantly less likely than other methods to predict substantial increases in habitat size; Random Forest models and Surface Range Envelopes were significantly more likely to predict a complete loss of current habitat, and future range extinction. Generalised Additive Models and Generalised Linear Models rarely predicted range extinction; future habitat completely disjunct from current habitat was predicted more frequently than expected by Classification Tree Analysis and less frequently by Maxent. Random Forest generally predicted extreme range changes more frequently than other SDM methods. Our results identify trends among different methods with respect to tendency to predict extreme range changes. These are of significance for climate-impact assessments, with implications for transferability of models to novel environments. Our findings emphasise the need to explore and justify the use of different models and their parameterisations, and to develop approaches to assist with optimisation of models.