Inferência Multimodelos na predição de multiprodutos em povoamentos de Eucalyptus sp.

Abstract

When selecting a single model to estimate a variable, this estimate is conditioned to any and all characteristics of this model, which may be subject to underestimation or overestimation, depending on the model in question and its possible biases. Recent studies have suggested the so-called Multimodel Inference, which consists in obtaining the result through the contribution of two or more mathematical models. The aim of this study was to propose a new methodology for Multimodel Inference, based on the modification of the ARM (Adaptive Regression by Mixing) algorithm, applied to forest inventory to estimating wood volume using volumetric and taper equations. To estimate wood volume with volumetric equations, the modified ARM algorithm was used and to estimate tree taper was used the proposed algorithm, Mixed Taper Equations, or MTE Algorithm. For the two clonal Eucalyptus sp. databases used in this study, the modified ARM algorithm was superior to model selection estimating tree volume more accurately, presenting RMSE of 3.53% and 3.86% while the selected model obtained 3.99% and 3.97 %, respectively. To estimate trees taper, with a database composed of three clonal varieties of Eucalyptus sp. the MTE algorithm was superior to model selection with RMSE of 4.11%, 8.96% and 4.91%, while selected models for each case presented RMSE of 4.49%, 8.88% and 5, 02%, respectively, for the three genetic materials tested. In addition, both modified ARM and MTE showed better dispersion of residues compared to the selected models and, in the case of the MTE algorithm, better results per tree. The use of MTE algorithm within the software R demonstrated that the technique is viable and managed to calculate different weights for the models tested over the measurement heights. In addition, the proposed method for estimating volumes using weights and taper equations was able to perform the tree volume reconstitution with more accurate results per tree than the tested models, individually.