Suitability of five cross validation methods for performance evaluation of nonlinear mixed-effects forest models - A case study

Abstract

Five cross validationmethods, the k-fold, leave one plot out (LOP), leave one tree per plotout (LOT), 0.632and 0.632+ bootstrapmethods, were examined in this study for their suitability for performance evaluation of seven nonlinear mixedmodels basedona height-diameter relationship. Thek-fold, LOP, 0.632 and 0.632+methods usedplot asthe basic unit for data resampling, and applies to situations where predictions are needed for all trees in a new plot not used for model development. All four methods were suitable for evaluating the predictive performance of the selected model(s), and the 0.632 and 0.632+ methods were better than the k-fold and LOP methods. The LOT method used tree as the basic unit for data resampling, and applies to situations where predictions are needed for a portion of trees in a plot not used for model development, while the remaining trees of the plot are used for model development. The LOTmethod was not suitable for performance evaluation of the selected model(s).