Model complexity affects species distribution projections under climate change

Abstract

Aim: Statistical species distribution models (SDMs) are the most common tool to predict the impact of climate change on biodiversity. They can be tuned to fit relationships at various levels of complexity (defined here as parameterization complexity, number of predictors, and multicollinearity) that may co-determine whether projections to novel climatic conditions are useful or misleading. Here, we assessed how model complexity affects the performance of model extrapolations and influences projections of species ranges under future climate change. Location: Europe.Taxon: 34 European tree species. Methods:We sampled three replicates of predictor sets for all combinations of 10 levels (n = 3-12) of environmental variables (climate, terrain, soil) and 10 levels of multicollinearity. We used these sets for each species to fit four SDM algorithms at three levels of parameterization complexity. The >100,000 resulting SDM fits were then evaluated under environmental block cross-validation and projected to environmental conditions for 2061-2080 considering four climate models and two emission scenarios. Finally, we investigated the relationships of model design with model performance and projected distributional changes. Results: Model complexity affected both model performance and projections of species distributional change. Fits of intermediate parameterization complexity performed best, and more complex parameterizations were associated with higher projected loss of current ranges. Model performance peaked at 10-11 variables but increasing number of variables had no consistent effect on distributional change projections. Multicollinearity had a low impact on model performance but distinctly increased projected loss of current ranges. Main conclusions: SDM-based climate change impact assessments should be based on ensembles of projections, varying SDM algorithms as well as parameterization complexity, besides emission scenarios and climate models. The number of predictor variables should be kept reasonably small and the classical threshold of maximum absolute Pearson correlation of 0.7 restricts collinearity-driven effects in projections of species ranges. | 131 BRUN et al.

Predicting range shifts of African apes under global change scenarios

“…Firstly, we extracted all variables for the extent of the range of each taxon, resampled onto a 5km x 5km equal-area grid and projected them into the WGS 1984 geographic coordinate system. Secondly, we used Spearman rank correlations to select a subset of least correlated variables to minimize multicollinearity (Brun et al, 2019). For this, we used a graphical representation of the correlation values between variables to identify five least correlated variables for each species to avoid overfitting in model predictions ( Fig.…”

“…Modelling species responses to global environmental changes carries many uncertainties (Araújo & New, 2007;Thuiller et al, 2019). Using two algorithm approaches, two future scenarios, two dispersal scenarios, an ensemble forecasting and including only a few but highly important predictors of the distribution of African apes, should have reduced uncertainties in our distribution models (Brun et al, 2019;Thorne et al, 2013). A recent study proposed that SDMs include historical records to produce better predictions of range shifts rather than relying on contemporary records alone (Faurby & Araújo, 2018).…”

On the importance of predictor choice, modelling technique, and number of pseudo‐absences for bioclimatic envelope model performance

“…The selection of pseudo-absence or background data is a third key consideration in SDM building, given that the majority of species observations concern presence-only data (Ponder et al, 2001). The impact of predictor set, modeling technique, and pseudo-absence selection on the performance of bioclimatic envelope models has been studied in isolation (Barbet-Massin et al, 2012;Beaumont et al, 2005;Moreno-Amat et al, 2015;Pearson et al, 2006;Pliscoff et al, 2014) or for two of the three factors (Brun et al, 2019;Bucklin et al, 2015;Dormann et al, 2008;Jarnevich et al, 2017;Petitpierre et al, 2017;Verbruggen et al, 2013;Warren et al, 2019). However, to our knowledge, no study so far has systematically evaluated the combined effect of predictor set, modeling technique, and pseudo-absence selection on the performance of bioclimatic envelope models.…”

SDMtune: An R package to tune and evaluate species distribution models

“…Despite the fact that highly correlated environmental variables are not a problem when the aim of the study is prediction in the same extent of the observed data, reducing collinearity is recommended in order to reduce model complexity and increase the interpretability of the predictors (Dormann et al., 2013; Merow et al., 2013). In addition, although the degree of accepted model complexity varies according to the modeling scope(s) (Halvorsen, 2012; Halvorsen, Mazzoni, Bryn, & Bakkestuen, 2015), it has been pointed out that models might perform best when trained with a reduced number of predictors (Brun et al., 2020; Halvorsen et al., 2015). Even though the selection should be driven by the knowledge of the modeled species, this might be difficult when the user must decide among several a priori ecologically relevant predictors for the species, or if the ecology of the species is poorly known.…”