Banca de QUALIFICAÇÃO: ALEXANDRE CAMPELO DE CARVALHO

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : ALEXANDRE CAMPELO DE CARVALHO
DATE: 16/11/2022
TIME: 09:00
LOCAL: Video conferência
TITLE:

ARTIFICIAL NEURAL NETWORKS AND LINEAR REGRESSION FOR ESTIMATING VOLUME OF TRUNKS IN FLONA SARACÁ-TAQUERA, STATE OF PARÁ

KEY WORDS:

Forest measurement; Mathematical models; R programming language; Artificial neural networks; Amazon.

PAGES: 29
BIG AREA: Ciências Agrárias
AREA: Recursos Florestais e Engenharia Florestal
SUBÁREA: Manejo Florestal
SUMMARY:

A major challenge for forest science is estimating the volume of hollows in the trunks of standing trees. Recent research carried out in the Brazilian Amazon has revealed that more than half of the trees harvested have a hollow in the trunk, with an impact on the volume of commercial timber. In public forests under concession contracts, although the occurrence of this defect is significant, there is no discount for existing hollows in the logs in the payment made to the government for the volume logged. Determining the volume of hollows and their discount is, therefore, of great interest to concessionaires. The great difficulty in estimating the volume of hollows in standing trees is the estimation of the height or length of the hollow, because it is impractical to determine it by direct methods in the field. The use of models based on machine learning techniques, a subfield of artificial intelligence, may constitute a potential indirect method to find more accurate predictive models for estimating the volume of hollows in commercial tree trunks compared to traditional regression models. Therefore, the objective of this research is to study and compare Classical Linear Regression (CRL) and Artificial Neural Networks (ANN) approaches in fitting models to predict hollow volume in commercial tree trunks in the Saracá-Taquera National Forest in the state of Pará. The trunk volumes of 213 sample trees were determined. The volumes of the existing hollows in the logs were calculated using the truncated cone formula, when it was possible to measure the radii at both ends of the logs, and with the cone formula, when the hollow did not reach the other end. The height or length of the hollow in logged trees was measured with the aid of a graduated stick introduced into the hollows of the trunks. To model the volume of the hollows, the classic models of equations of trunk volume presented in the forestry literature will be tested by means of the standardized least squares method and by non-linear regression. To estimate hollow volumes through machine learning, ten artificial neural networks (ANN) will be trained. Each network will have two neurons in the input layer with the variables stem height and diameter at breast height (DBH). For predictive modeling purposes, the original dataset (n = 213) was divided into training (70%) and test (30%) data, using the Leave-One-Out Cross Validation (LOOCV) method to assess the performance of models during the learning process. Multilayer Perceptron (MLP) ANNs will be used, with Resilient Propagation algorithm for weight learning, and Sigmoidal activation function. Stopping criteria will be established based on the average error and number of cycles (1% error and 30,000 cycles). For the training and evaluation of the networks, the R programming language will be used, as well as for the adjustment of the regression models. The statistical criteria for the evaluation of the best regression model and the best ANN will be: higher adjusted coefficient of determination (R²aj), lower standard error of the estimate (Syx %); distribution of standardized residues; Akaike information criterion; for the ANNs, the following indicators will be used: correlation coefficient (Rŷy), bias, root-mean-square-relative error (rRMSE) and Statistical Prediction Residual Error Sum of Squares (PRESS). It is expected, with this study, to select a safe, effective technique with acceptable accuracy to estimate hollow volumes in tropical forests in the Amazon, thus contributing to the advancement of forest science and to improving the management of public forests in the region.

BANKING MEMBERS:
Presidente - 142.060.489-91 - JOSE NATALINO MACEDO SILVA - Oxford
Interno - 4143218 - JOAO OLEGARIO PEREIRA DE CARVALHO
Interno - 2411802 - RODRIGO GERONI MENDES NASCIMENTO
Externa à Instituição - GABRIELA CRISTINA COSTA SILVA - UFPA
Externo à Instituição - EDSON MARCOS LEAL SOARES RAMOS - UFPA