Using the Non-Parametric Classifier CART to Model Forest Tree Mortality

Abstract

Abstract A binary classification tree (CART) was used to predict forest tree mortality for two conifer species. CART models were fitted using binary recursive splitting of the data set into increasingly homogeneous subsets. Models were compared in terms of improvement of prediction accuracy, representativeness in average size of the predicted mortality trees, and interpretability of the results. For shade intolerant ponderosa pine, crown ratio, diameter increment prediction, or variables indicating the relative position of a tree in a stand were used for splitting. For shade tolerant white fir, height increment prediction and stand density were selected for splitting. The prediction accuracies for mortality trees of the best CART models were between 28-36% for ponderosa pine and between 11-17% for white fir. CART was also compared with logistic regression using a stochastic and a deterministic assignment of mortality. Efficiencies similar to those achieved with CART were reached with deterministic logistic models using thresholds probabilities. However, CART and the logistic model tended to utilize different predictor variables, especially for white fir. CART uncovered additional factors for white fir important for predicting mortality not identified by the logistic regression. For. Sci. 44(4):507-516.