Modelagem da relação hipsométrica e do crescimento e produção utilizando aprendizagem de máquina e modelo de efeito misto

Abstract

We developed this study based on forest inventory data collected from temporary and permanent plots. The objective was to model the hypsometric relationship, the growth, and the production of eucalyptus forests using traditional regression techniques, mixed modeling, and machine learning to obtain models capable of representing the reality of forest stands. We divided the study into three chapters. In the first chapter, we performed a literature review to support the development of the second and third chapters. The second chapter consisted of evaluating four nonlinear models for modeling the hypsometric relationship. We compared the traditional regression technique with the mixed modeling technique with the addition of covariates and heteroscedasticity modeling to obtain more accurate models. The use of nonlinear mixed models allowed us to reduce the standard error by approximately 60%. We decomposed the Gompertz model and included the clone, basal area, site, and age covariates to provide better accuracy, modeling the heteroscedasticity. The heteroscedasticity modeling allowed us to reduce the estimate of the residual standard error and percentage of approximately 55% when compared to the mixed effects model. For the second chapter, we aimed to analyze the performance of machine learning techniques (Artificial Neural Networks - ANN and Support Vector Machine - SVM) in the projection of the basal area and volume, comparing them to the traditional method of regression analysis using the Clutter model and variations thereof. The data derive from 2,550 permanent sample units of clonal stands of the Eucalyptus grandis x E. urophylla hybrid, comprising seven genetic materials. The machine learning methods to project the basal area provided good training and generalization skills. For the volumetric projection, the methods presented low generalization capacity. The estimates produced by the Clutter system and its variations were superior to the machine learning techniques.