Configuration of artificial neural networks for estimation of total height of eucalyptus trees

Abstract

The aim of this study was to define appropriate configurations of Artificial Neural Networks (ANN) to obtain the total height of eucalyptus trees. The data used came from continuous forest inventories in stands aged 21-137 months located in southern Bahia. The ANN configurations tested varied according to the number of neurons in the hidden layer, activation function, number of cycles and learning algorithms with their parameters. The tests were performed in Neuroforest system and the estimates were evaluated using the correlation coefficient, the root mean square error (RMSE%), and graphical analysis of residues. The estimation of the height of trees may be made by various ANN configurations using the learning algorithms Resilient Propagation, Quick Propagation and Scaled Conjugate Gradient, with number of hidden neurons varying between 03 and 08 for the Quick Propagation algorithm and 13 and 20 to Scaled Conjugate Gradient algorithm. The activation functions hyperbolic tangent, sigmoid, log and sine are suitable for the hidden and output layers, and functions linear and identity proved suitable only for the output layer. Two thousand cycles are sufficient for the training of ANN.