Modeling of eucalyptus productivity with artificial neural networks

Although it is easy to list the several factors that influence forest productivity, it is almost impossible to isolate or measure all biotic and abiotic components that influence crop growth and development on a field-scale. Artificial neural networks (ANN) have been widely used to model forest productivity, since habitual modeling is hampered by the inclusion of categorical variables as well as by the large number of independent variables and its complex relationships with the dependent variable. This study aimed to obtain ANN to estimate eucalyptus productivity as a function of environmental variables, genotype and silvicultural practices, and infer about those more important in forest productivity. It was used data from continuous forest inventory, climate, soil analysis and fertilization carried out on 507 eucalyptus stands, composed of different genotypes and spacing. Multiple-layer Perceptron networks were trained to estimate the mean annual increment of eucalyptus stands at six years of age (MAI6), testing different combinations of input variables, number of neurons in the hidden layer, training algorithms, data percentage in training and validation subsets, and activation functions. In the validation, it was obtained ANN with correlation between the estimated and observed MAI6 higher than 85 % and root mean square error less than 15 %. Despite data complexity, ANN made it possible to estimate MAI6 with good precision and to include easily numerous variables, even categorical. Genotype, spacing, edaphic characteristics (clay, organic matter and Cation exchange capacity- CEC), climatic characteristics (rainfall, temperature and water deficit) and fertilization were the predictive variables that most influenced eucalyptus productivity at the end of rotation.