Multinomial and Ordinal Logistic Regression Models and Artificial Neural Networks for lumber grading

Abstract

Lumber classification is one of the most subjective activities of the final phase of log sawing process in sawmills. The objective of this research was to propose tools that assist in conifers lumber grading. The research was carried out at the sawmill Combate de Tenerías of Macurije integrated forest company, Pinar del Río, Cuba. The data used comes from 259 lumber pieces of Pinus caribaea var. caribaea classified following the requirements (24 variables) and classes established by the conifers lumber grader used in Cuba. We proceeded to fit a Multinomial and Ordinal Logistic Regression (MOLR) model and train Artificial Neural Networks (ANNs). The parameters of the MOLR model were estimated using the maximum likelihood method optimized with the Newton-Raphson algorithm. ANNs were trained with Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm. The multicollinearity problem, usually present in modeling with numerous predictive variables, was addressed with factor analysis, using the factors retained as inputs to the models. Based on the percentage of correct classification, the ANN RBF 24-8-4 was superior to the ordinal logistic regression equations.