Modelos mistos generalizados e redes neurais artificiais aplicados na estimativa da altura, afilamento e volume total e parcial de árvores de Pinus spp.

Abstract

Concerning planning, ordering, and the use of wood increasingly requires greater precision. Furthermore, the growing need for predicting forest multiproduct has supported the use of various modeling techniques to describe the profile of boles. These techniques include the use of mixed effect models and the use of machine learning, having provided gains in accuracy. Therefore, this work evaluated, in addition to classical local regression models and generic models, the application of alternative methods, such as mixed models and artificial neural networks in different stages of forest measurement (height prediction and stem taper modeling). The data used in the present study were obtained from Pinus spp. plantations, located in the municipality of Nova Ponte, Minas Gerais. Different configurations of the methods were tested and based on statistical measurements of the quality of the predictions, the gain in precision was verified with the use of the proposed techniques. To evaluate the predictions made by the different methods, at each stage, and select the best models, the following criteria were considered: root mean square error (RMSE), mean absolute percentage error (MAPE), model efficiency index (EF) , mean deviation (BIAS), Akaike Information Criterion (AIC), Schwarz Bayesian Criterion (BIC), likelihood ratio test (MLRT), paired t-test, paired Wilcoxon signed rank test and graphical analysis of residual distributions and histograms. All analyses were carried out using the R programming language. Ratkowsky's equation (1990) stood out as the most appropriate to explain the interaction between height and diameter of Pinus caribaea var. hondurensis. The inclusion of dominant diameter and dominant height resulted in significant improvements in predictions of the total height of trees at different sites and with different ages. The introduction of random effects and calibration per plot demonstrated great effectiveness in improving hypsometric models. Furthermore, the most appropriate calibration alternatives were identified, requiring nine trees distributed equally in different diameter classes for the local model in mixed form and, for the model that already had plot-level variables, calibration was not necessary. Artificial neural networks demonstrated poor generalization capacity, resulting in biased predictions for the test database. Considering the modeling of the stem tapering, the modeling with random effects and the continuous first-order autocorrelation function provided significant improvements to the Kozak (2004) model, making the models more coherent with the data, resulting in values significantly equal to the real ones for Pinus caribaea var. caribaea and Pinus oocarpa. Artificial Neural Networks demonstrated efficiency in predicting diameter along the stem and total volume for Pinus caribaea var. caribaea and Pinus caribaea var. hondurensis, generating predictions statistically equal to the observed values simultaneously for both species using dummy variables. The results found in this work demonstrate the ability of mixed modeling and artificial neural networks to provide significant gains in accuracy in predicting forest attributes. Furthermore, it demonstrated the need for more in-depth studies on the optimization of strategies for selecting the most appropriate structure of artificial neural networks and interpreting their adjusted parameters. Also, it is recommended to add other attributes at the plot level, to improve the representation of the tapering of tree stems, if the collection of these variables does not compromise the viability of forest inventory operations.