2.2. Environmental Variables and Multicollinearity

A Digital elevation model (DEM) was downloaded from the U.S. geological survey (https://www.usgs.gov) at 30 arc-seconds spatial resolution. The 19 bioclimatic predictors of current and future climates were downloaded from the WorldClim v 2.1 (https://www.worldclim.org/data/worldclim21.html) at a 30-arcsecond resolution [31].

To evaluate the impact of projected climate change on the potential distribution of Juniperus species, we used two global general circulation models (GCMs): BCC-CSM1.1 (Beijing Climate Centre–Climate System Modelling 1.1, http://forecast.bcccsm.ncc-cma.net/web/channel-34.htm) and MIROC5 (Model for Interdisciplinary Research On Climate, http://www.icesfoundation.org/Pages/ScienceItemDetails.aspx?siid=181). BCC-CSM1.1 is widely used for Asian regions and performs well when describing vegetation dynamics compared to other GCMs [32]. Simultaneously, MIROC5 simulates extreme and summer precipitation better than other GCMs for the South Asian region [33]. We used an ensemble average of the two GCMs to reduce the uncertainty arising from a single GCM [34].

Two representative concentration pathways’ (RCP4.5 and RCP8.5) emission scenarios of 2070 (average of 2061–2080) were applied. The RCP4.5 pathway represents a moderate scenario, but RCP8.5 indicates a high scenario. The reason behind choosing these two RCP scenarios is because China is the largest emitting country of carbon dioxide [35].

We used the crop and mask functions of the “raster” package in R 3.5.3 [36] to clip the bioclimatic and elevation layers according to a China shapefile and then resampled the output into 60 arc-second (approximately 2 km) resolution, which is required for AOO calculation as described by [30]. Finally, based on the occurrence coordinates of each species, the values of bioclimatic and elevation variables were extracted for the analysis of multicollinearity.

To reduce overfitting of SDM models, we removed the highly correlated variables based on their variance inflation factor (VIF), which measures how strongly each predictor can be explained by the rest of the predictors [37]. To perform VIF analysis, we used the vifcor and vifstep functions of the package “usdm” [38] in R 3.5.3 to exclude the variables with VIF values more than five and a correlation threshold of 0.75, as recommended by [39]. The relative importance of predictor variables was estimated using the function getVarImp of the “SDM” package in R 3.5.3.