Métodos de aprendizagem de máquina para predição de umidade e densidade básica da madeira a partir de espectros no NIR

Abstract

Near-infrared (NIR) spectroscopy is a fast and accurate technique that can be applied to a large number of samples and, when associated with partial least squares regression analysis and artificial neural networks, has been shown to be an efficient tool in prediction. of wood properties. The objective of this study was to verify the performance of least squares regression analysis (PLS-R) and artificial neural networks (ANN) in estimating moisture and basic density of solid wood and chips of Eucalyptus spp. from spectral signatures in the NIR. The NIR spectra and the masses were measured in the 110 chips and in the 110 prismatic samples of wood, at every 10% of mass loss, from the saturated condition to the anhydrous condition, for later determination of the basic density and moisture in the different phases. Thus, wood moisture and basic density, obtained by the conventional method, were correlated with the corresponding spectra in the NIR by means of least squares regression analysis (PLS-R) and Artificial Neural Networks (ANN). For the development of the ANNs, the backpropagation learning algorithm was used with Multilayer perceptron networks. The determination of wood chip moisture, from the NIR spectra, presented the best estimates by the RNA method from the deciles with R2 of 0.97, RMSE of 8.83% and RPD of 5.58. The estimate of the basic density of wood chips, from the spectra in the NIR, was reached with the ANN using deciles, presenting R2 of 0.17, RMSE of 0.02% and RPD of 1.00. Moisture determination of Eucalyptus sp. achieved the best performance by RNA using deciles with R2 of 0.92, RMSE of 10.38% and RPD of 3.34. The best estimates for the determination of density, regardless of prism moisture, were obtained by the PLS-R model with R2 of 0.80, RMSE of 0.06 in the test and RPD of 2.00. Therefore, the models developed by PLS-R and RNA, starting and spectra in the NIR, proved to be a useful tool for fast and accurate prediction of the humidity of chips and prisms of Eucalyptus wood. The determination of basic density, independent of moisture, showed satisfactory results only in wooden prisms through PLS-R and RNA and using the deciles of the spectra.