Redes neurais artificiais e regressão não linear na estimativa do volume de oco em troncos de árvores na flona de Saracá-Taquera, Estado do Pará.

Abstract

In tropical forests, among the many challenges for forest measurement is the estimation of hollow volume in tree trunks. In the Brazilian Amazon, in transactions between forestry entrepreneurs, the volumes of this defect are generally discounted using the Francon volume. As the contracts do not provide for the discount of existing holes in the logs, this represents a loss for concessionaire companies in public forests, as they end up paying for a volume above what they actually exploit. Therefore, its estimate in standing trees or logs is important for both the private and public sectors if it is to be adopted in the future, especially in concession contracts where this volume is not discounted yet. Hollow volumes can be estimated using regression techniques, as is done for trunk volumes. An alternative to classical linear regression techniques is the use of machine learning techniques, a subfield of artificial intelligence. This is a potential indirect method to find accurate predictive models for estimating hollow volume in tropical forest tree trunks. In this context, the objective of this research was to evaluate the efficiency of artificial neural networks against regression models to estimate the hollow volume in standing and felled tree trunks in a natural forest in the-Amazon region, specifically in the National Forest of Saracá Taquera. The following variables were measured in 213 sample trunks, namely: length of the log, diameters at the base and top of the log, diameters at the base and top of the hollow, length of the hollow, DBHT welve artificial neural networks (ANN) were trained to estimate the hollow volumes through machine learning were trained and six non-linear regression models were adjusted. Each network had from 2 to 16 neurons in the input layer and the number of intermediate layers depended for each neural network. For predictive modeling purposes, the original dataset (n = 213) was divided into training data (70%) and test data (30%), using the Leave-One-Out Cross Validation (LOOCV) method to evaluate the performance of the models during the learning process. ANNs of the Multilayer Perceptron (MLP) type were used, with a Resilient Propagation algorithm for weight learning, and a linear activation function. Stopping criteria were established based on mean error and number of cycles (1% error and 30,000 cycles). Regression models were fitted using the Levenberg-Marquardt algorithm. It was observed that the neural network R9, which used the input variables, diameter of the base and top of the hollow and length of the hollow, presented the lowest estimation error (4.97%) among the networks tested. While the regression equation of the Schumacher-Hall model for M1 that used the variables of the average of the hollow diameter ((D_0 ) ̅), and hollow length (Co) presented the smallest estimation error (2.60%) . It is concluded that the R9 network with the input variables hollow length, diameter of the base and top of the hollow was able to accurately describe and estimate the hollow volume of logs from Flona de Saracá - Taquera. However, non-linear regression modeling was even more accurate, using mean hollow diameter and hollow length as independent variables. Both methods were especially useful for estimating hollow volume the felled logs, which were reconstituted to verify both the estimated hollow volume in standing trees and in logs from felled trees.