Estimation of Eucalyptus productivity using efficient artificial neural network

Prediction of plantation productivity from environmental data is challenged by the complex relationships defining the growth and yield processes. Artificial neural network (ANN) can be used to model complex nonlinear interrelations but are challenged by the choice of the training algorithm and the structure and size of the network. The objective of the present study is to find an efficient ANN to estimate the eucalyptus productivity from biotic, abiotic and silvicultural data. We defined the efficiency of an ANN as the processing time to supply an accurate solution. We increase the efficiency of the network in two steps: one outside the ANN, through principal component analysis (PCA), and one inside the ANN, through dedicated pruning methods. To test different network configurations, we used data from 507 Eucalyptus plantations, 3 to 7 years old, located in Minas Gerais, Brazil. An ANN configuration is a combination of the number of neurons in the hidden layers, training algorithm and pruning method. Each ANN was trained five times to assess the impact of the initial weights on the results, which led to a factorial experiment with 9000 combinations. The most accurate result was supplied in 38.81 s by an ANN using the data trimmed with PCA, trained with the Scaled Conjugate Gradient algorithm with four neurons and Magnitude-Based Pruning method. However, an accuracy loss of less than 1% (i.e., the second most accurate result) was obtained in 1.7 s using the same ANN configuration, but no pruning. Our results indicate that an efficient prediction of Eucalyptus productivity does not use all the data or the most complex training algorithms. The proposed efficient ANN supplies the most precise results compared with the existing models. Furthermore, the efficient ANN produces more accurate results from environmental variables than some models that contain dimensional variables. Our study suggests that modeling the yield of Eucalyptus plantations should be executed using a large set of environmental variables, as no clear edaphic or climatic variables supplied accurate results.