Feature selection aplicada à biometria florestal

Abstract

Computational advances made possible greater viability of data collection, storage, and algorithms processing with the expansion of big data in the forestry sector. In line with this, computational intelligence techniques have been increasingly applied to support decision-making in several problems. Among their applications, the feature selection process successfully contributes to the task automation of reducing the dimensionality of the data for optimizing a subset of relevant variables in the models building. Given this perspective, the thesis focuses on the genetic algorithms' use in association with the Random Forest (GA-RF) for selecting variables in the modeling of forest machine productivity (Article 1) and the periodic annual diameter increment in a Semideciduous seasonal montane forest in Brazil (Article 2). In article 1, the objective of the work was to test different methodological approaches in the generation of models with good predictive capacity, in addition to investigating the importance of variables arising from soil and climate conditions, operator records, and forest inventory. We selected the GA-RF because it has a high generalization power by reducing the errors' estimates, in addition to maximizing the importance of relevant variables in the machine's productivity. Article 2 aimed to evaluate the incorporation of competition effects in a growth model at individual trees level, based on the investigation of different categories of classical competition indices and an additional methodology proposed in this study, known as metrics of complex networks. The GA-RF methodology was efficient by combining ecological meaning and accuracy improvements. It selected distance-independent indices and complex network metrics for modeling the growth of the species, Xylopia brasiliensis, and Copaifera langsdorffii, respectively.