Comparação entre redes neurais artificiais e comitê de redes na estimação do volume individual de árvores de Eucalyptus

Abstract

With the advancement of computer programs and the diffusion of Computational Intelligence, artificial neural networks (ANNs) have been used as an alternative for the modeling and prognosis of forest production. The objective of this work was to evaluate the performance of an artificial neural network and a committee of artificial neural networks in the estimation of the individual Eucalyptus trees volume. Data from Eucalyptus spp. stands were used, from nine clones distributed in 28 plots. ANNs were obtained through ten training sessions and in each session was retained the best one. The networks were of the Multilayer Perceptron type, with 1 hidden layer. The number of neurons in the hidden layer was defined in an automated way by the Statistica software and the back propagation algorithm was used. The activation function used was logistics; learning rate of 0.001 and the term momentum 0.5. As a result, it was observed that the use of the committee managed to reduce the average relative error of the ten ANNs, which changed from an average of 4.30% to 1.49%. The volume estimates obtained by artificial neural networks committee presented an average relative error lower than that presented by the volume estimates for the best isolated network, which was 1.73%. The ANNs committee is shown to be applicable to the processes of estimation of individual volume of shafts, presenting good performance.