Aprendizado de máquina para modelagem do crescimento e produção em povoamentos de Eucalyptus spp

Abstract

The forestry sector, especially that of planted forests, is one of the most crucial pillars of Brazilian agribusiness. Brazilian planted forests stand out for their high productivity. In the planning of timber production, it is essential to have the ability to estimate production both in the present and, particularly, in the future. Studies in the field of biometrics have developed important approaches and models that accurately estimate production, especially based on linear and non-linear regression techniques. Currently, machine learning algorithms have been tested, and the results have proven to be very promising. Thus, this study aimed to evaluate machine learning algorithms for estimating present and future production in eucalyptus plantations. The following machine learning algorithms were assessed: Artificial Neural Networks (ANN), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). These methods were compared to the Clutter model, a system of linear equations traditionally used by the forestry industry for production prognosis. The models were fitted in four forest analyses, measured in 1280 permanent plots between 2013 and 2019 in the state of Minas Gerais. Machine learning models were adjusted using field biometric data and derivatives of remote sensing, such as vegetation indices. In general, machine learning models had lower estimation errors than the Clutter model, with SVM standing out. Additionally, the use of vegetation indices in the SVM model improved the accuracy of current production estimates. The use of traditional variables together with forest inventory registration variables in the SVM model enhanced the accuracy of future production estimates. Thus, the analyses demonstrated that machine learning models are good production estimators when associated with field biometric data and remote sensing derivatives.