Modelagem da densidade básica da madeira de eucalipto utilizando redes neurais artificiais

• Ano de defesa: 2018
• Autor(a) principal: Boa, Ana Carolina
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Ciências Florestais; Centro de Ciências Agrárias e Engenharias; UFES; Programa de Pós-Graduação em Ciências Florestais
• Palavras-chave em Português: Madeira; Qualidade; Eucalipto; Polpa; Redes neurais (Computação); Redes neurais_x000D_ (Computação); Recursos Florestais e Engenharia Florestal; 630
• Link de acesso: http://repositorio.ufes.br/handle/10/10568

Brazilian forestry industries are world leaders in productivity and quality, and use a large volume of raw material to meet the demand of their processes. In addition to quantity, the industry requires materials that meet the quality of their processes and match the required quality of their final products. Among the properties that characterize wood, the basic density is highlighted as an important parameter of quality, since it is related to many different technological and economic aspects. Therefore, the objective of this study was to apply artificial neural network modeling to estimate the basic density of eucalyptus wood for pulpwood. A total of 352 trees from 18 clones of Eucalyptus grandis x Eucalyptus urophylla hybrid, aged between two and eight years old, originated from plantations in the states of Espírito Santo and Bahia. The quantitative variables used in the density estimates were age, diameter at breast height, volume, accumulated precipitation, temperature and relative humidity, added to qualitative variables clone and region. The density of the wood was estimated using artificial neural networks (ANNs) and, for better estimation performance, the estimates were made from seven different combinations of the variables: COMP (complete data), REG-1 (Aracruz-ES), REG-2 (São Mateus-ES), REG-3 (Bahia), CLAS-1 (trees from 2 to 4 years), CLAS-2 (trees from 4 to 6 years) e CLAS-3 (trees from 6 to 8 years). The performances of the hyperbolic logarithmic and hyperbolic tangent activation functions and the Levenberg-Marquardt and Resilient Propagation (RPROP +) training algorithms of the hidden layer of the RNAs were also tested. Based on the hidden layer configurations, the Levenberg-Marquardart algorithm presented better performance and both logarithmic and hyperbolic activation functions performed satisfactorily. In general, artificial neural networks performed well in estimating the basic density of eucalyptus wood, and all combinations of variables used in the estimation were efficient. However, there was a tendency to overestimate the estimated values. Specifying regions and age classes allowed better results to be achieved, with more accurate results being observed in the region Aracruz (REG-1) and trees from 6 to 8 years (CLAS-2).