Comparing Predicted Historical Distributions of Tree Species Using Two Tree-based Ensemble Classification Methods

Fine scale spatial mapping of historical tree records over large extents is important for determining historical species distributions. We compared performance of two ensemble methods based on classification trees, random forests, and boosted classification, for mapping continuous historical distributions of tree species. We used a combination of soil and terrain predictor variables to predict species distributions for 21 tree species, or species groups, from historical tree surveys in the Missouri Ozarks. Mean true positive rates and AUC values of all species combined for random forests and boosted classification, at a modeling prevalence and threshold of 0.5, were similar and ranged from 0.80 to 0.84. Although prediction probabilities were correlated (mean r = 0.93), predicted probabilities from random forests generated maps with more variation within subsections, whereas boosted classification was better able to differentiate the restricted range of shortleaf pine. Both random forests and boosted classification performed well at predicting species distributions over large extents. Comparison of species distributions from two or more statistical methods permits selection of the most appropriate models. Because ensemble classification trees incorporate environmental predictors, they should improve current methods used for mapping historical trees species distributions and increase the understanding of historical distributions of species.