Tree-Based Methods: Concepts, Uses and Limitations under the Framework of Resource Selection Models

Abstract

The use of empirical models to predict species distribution is recognized as an important tool in wildlife management. Tree-based methods gained considerable attention in the last years mostly due to their flexibility and robustness. Here, we provide an overview of tree-based methods by addressing some of their concepts, uses and limitations. For illustrative purposes, we modelled the distribution of a red deer (Cervus elaphus) population using fine-scale predictors while applying four modelling methods: three treebased methods (classification trees, random forests and boosted trees) and the generalized linear model by stepwise regression. In order to explore alternative trees and achieve the best model performance, a series of classifiers were run with different tuning parameters. The random forests and boosted trees models were the most accurate classifiers followed by classification trees and generalized linear model by stepwise regression. Despite differences in the predictive accuracy, the results of the four models were consistent with the species ecological requirements. Red deer occurred further away from disturbed areas (e.g. villages and other human settlements), agricultural fields and near shrubs and forest patches. Furthermore, the species often occurred in areas with gentle slopes, preferentially with a southern exposure. We observed that classification trees are easy to interpret but may produce unstable decision trees and unwieldy results in the presence of sharp discontinuities. We state that ensemble methods such as random forests and boosted trees are valuable tools in predicting species distributions. This study provides the necessary background for the understanding of tree-based methods, which will be of great help in further studies in ecological modelling, as it will shed light in the most appropriate technique to be used.

Keywords: boosted trees, classification trees, ecological modelling, fine-scale predictors, random forests, red deer