Configuration of artificial neural networks for height-diameter relationship of Eucalyptus spp.

Abstract

This study explores algorithms and functions of activation of artificial neural networks (ANNs) to predict the total height of Eucalyptus spp. The objective was to recommend the best RNA configurations for this variable. The data came from 2,888 trees. The trained ANNs presented DBH, clone, age, age class and diametric class as input variables. The total height was the output variable. Five algorithms and six activation functions were combined in the hidden and output layers, totaling 18,125 trained ANNs. ANNs were evaluated using linear correlation (yyr), square root of the average error (RMSE%), bias and histograms of yyr and RMSE%. The trained ANNs obtained RMSE% ranging from 0.07% to 396.3% and yyr of -0.7130 to 0.9998. The ANNs was performed using the Neuro 4.0.6 software. With the exception of ANN with the Manhattan Update Rule algorithm, the best ANN selected in the validation showed a correlation above 0.97, and RMSE% and bias close to zero. The Backpropagation, Resilient Propagation, Scaled Conjugate Gradient and Quick Propagation algorithms presented satisfactory results in height modeling. The logistic and log activation functions are efficient for the hidden and output layers, respectively. In validation, the 12-10-1 network architecture with a Resilient Propagation algorithm showed the highest precision, with RMSE of 0.067 m. On the other hand, the architecture 12-14-1 with the Manhattan Update Rule algorithm resulted in the lowest precision, with RMSE of 3.13 m. The 12-10-1 network architecture, with Resilient Propagation algorithm and logistical activation function, can be used in the training for the prediction of the total height of Eucalyptus spp.