MÉTODOS DE SELEÇÃO DE VARIÁVEIS EXPLICATIVAS PARA A ESTIMATIVA DA ALTURA DE ÁRVORES DE EUCALIPTO.

Abstract

Knowing the height of trees in stands is fundamental for quantifying forest resources. However, obtaining this variable is difficult to operationalize, performed by means of indirect methods, through the use of hypsometers, which are subject to several errors. The objective of this work was to evaluate the performance of different strategies for selection of variables on the accuracy of estimates of the height of eucalyptus trees. In each plot, five rectangular plots were delimited, with an area of approximately 540m², and the diameter (cm) of all trees was measured, and the total height (m) of about 33% of the trees in the plots, also with the identification and measurement the height of the dominant trees. The dataset was randomly divided into two subsets: fit (75%) and validation (25%). The adjustment data were used to apply different strategies for selecting variables to be included in the hypsometric models, and subsequently, in the adjustments of the selected models. Different statistical procedures were used for the selection of explanatory variables: empirical model, Pearson correlation, Stepwise regression (Forward method), path analysis and Random Forest algorithm, being executed, with and without the dominant height variable (Hd), in the set of variables. The following hypotheses were evaluated: H0(1): The methods of selection of explanatory variables provide the selection of variables with greater predictive capacity, which result in improved precision in estimating the height of eucalyptus trees; and H0(2): The variable Hd can be replaced by others that are easier to obtain, without losing the predictive capacity of the models. Path analysis is the most efficient statistical procedure for selecting explanatory variables in the construction of hypsometric models, with superior performance than commonly used reference models. Therefore, the hypothesis H0(1) was accepted. All models obtained without including the variable Hd have a loss in the quality of precision of the estimates, when compared to the models whose dominant height was included. Therefore, hypothesis H0(2) was rejected.