Results

We found that reasonably parsimonious bioclimatic models are driven by either 3 or 4 climate variables (Table: Statistical output from bioclimatic models). Overall out-of-bag classification errors from fitting of the Random Forests algorithm ranged from 1.39 percent for Macfarlane’s four-o’clock to 2.2 percent for subalpine larch. The models predicted species occurrence where species were known to occur with perfect accuracy (0 errors of omission); the error therefore is due to commission, predicting the presence of a species when, in fact, it was absent. The use of digital elevation model (DEM) and GIS to make predictions on a 1-km grid introduces additional error and uncertainty. The use of Little’s range maps to validate the models’ mapped projects also introduced the possibility of error. Range maps are known to contain errors (see Rehfeldt and others 2006) but perhaps more importantly, they only provide 2-dimensional limits of the species distribution and do not indicate where species actually occur within the species boundaries. Even so, a visual comparison between these range maps and our predictions can be used for general validation of the models.