Please use this identifier to cite or link to this item: http://repositorio.ufla.br/jspui/handle/1/42144
Title: Funções de afilamento ajustadas por regressão e por redes neurais artificiais para espécies madeireiras da Amazônia
Other Titles: Taper function adjusted by regression and artificial neural networks for Amazonian timber species
Authors: Calegario, Natalino
Gomide, Lucas
Silva, José Antônio Aleixo da
Keywords: Regressão não linear
Inteligência artificial
Espécies nativas
Funções de afilamento
Nonlinear regression
Artificial intelligence
Native species
Issue Date: 30-Jul-2020
Publisher: Universidade Federal de Lavras
Citation: NASCIMENTO, M. Y. M. do. Funções de afilamento ajustadas por regressão e por redes neurais artificiais para espécies madeireiras da Amazônia. 2020. 48 p. Dissertação (Mestrado em Engenharia Florestal) – Universidade Federal de Lavras, Lavras, 2020.
Abstract: This study aimed to evaluate the taper functions efficiency adjusted by classical, generalized, mixed-effect regression and by neural network in stems of 7 species of commercial interest in the Amazon. The data were collected in the area of the sustainable management project of the company Mil Madeiras Preciosas, at Itacoatiara city, in the Amazonas state. 214 stems were measured, divided into 10 equal folds for cross validation of the equations. Data were collected by rigorous cubing, measuring the diameters of each stem at heights (h) relative to the stem base of 0.3m, 1.3m, 2.3m, and so on, up to commercial height (H). Eight traditional taper models were adjusted to estimate diameters at various heights, using linear and non-linear regression techniques, and 20 artificial neural networks. The best performance model using the Akaike information criterion was selected for the adjustment using the mixed effect technique, where the species factor was introduced as a random component. The best performance model by the Akaike information criterion (AIC) was selected for the adjustment using the mixed effect technique, where the species factor was introduced as a random component. The volumes of each stem were calculated using the Smalian formula. In order to assess the accuracy of volume estimates, two neural networks were trained and the Schumacher-Hall model was adjusted. The volume was also estimated by integrating the best performing taper function, using its fixed and mixed structures. Finally, the volume was estimated by the Schumacher-Hall model, in its fixed and mixed structures, including DBH and Height as fixed variables and the species as random. The evaluations of the accuracy of the estimates were made using the correlation coefficient, R² (%) and with the root mean square error, RMSE (%). The results confirm that MLB RNAs, trained as a taper function, had a better performance in estimating the diameter, and the architecture with 10 inputs, including species, and with 10 neurons in the intermediate layer (RNA 20) was the best performance, with R² (%) of 96.07. For the taper functions adjusted by regression, the model proposed by Kozak (2004) was the best performance, with R² (%) values of 95.65 and 94.41 for the mixed and fixed adjustments, respectively. However, in the validation the performance of the Kozak model was better than that of RNA 20. In the evaluation of the volume estimates, the results confirmed the superiority of the Schumacher-Hall model adjusted by the nlme technique, with R² (%) of 96.05, followed by the neural network with R² (%) of 95.37, both including the species as an independent variable. However, the integrated Kozak (2004) model using the parameters of the mixed adjustment obtained the best performance for unknown data, with R²% of 95.31. It is worth mentioning that an integrated tapering function has the flexibility to estimate partial volumes in any position of the stem.
URI: http://repositorio.ufla.br/jspui/handle/1/42144
Appears in Collections:Engenharia Florestal - Mestrado (Dissertações)



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.