APRENDIZADO DE MÁQUINAS POR “RANDOM FOREST” PARA A MODELAGEM DA ALTURA DE ÁRVORES DE SERINGUEIRA

Abstract

Total tree height (H) is a fundamental variable, but one that is difficult to measure in forest inventories. It is usually measured in just a few samples, while the rest of the trees are estimated using regression techniques. Recently, machine learning algorithms, such as Random Forest (RF), have shown promise in generating accurate estimates of H. Therefore, this study aimed to evaluate the accuracy of H estimates in rubber trees using RF and Linear Regression (LR) models, trained from different functional relationships. The database came from the measurement of 28 permanent plots distributed over four rubber tree plots (clones RRIM600 and RRIM937), located in the municipality of Paraíso das Águas/MS. In each plot, the following variables were measured: H, in meters; diameter at 1.3 m from ground level (D), in centimeters; and dominant height (Hd), in meters; at five to seven ages (I), between 4.3 and 16.5 years after planting. The database was randomly subdivided into training (80%) and validation (20%). The RF and RL models were trained considering four functional relationships: 1) H = f(D, Hd, I); 2) H = f(D, Hd); 3) H = f(D, I) and; 4) H = f (D), and were then used to generate estimates of H in the validation data. The accuracy of the estimates was assessed. Functional relationships 1 and 2 provide accurate estimates of tree height and are superior to functional relationship 4. The RF models generally perform better in predicting H in rubber trees, indicating that they are a good alternative for modeling the hypsometric relationship.