Kernel regression methods for the prediction of wood properties of Pinus taeda using near infrared spectroscopy

Near infrared diffuse reflectance spectra collected in 10-mm sections were used for the estimation of air-dry density (AD), microfibril angle (MFA), stiffness (MOE), tracheid coarseness (COARS), and tracheid wall thickness (WTHICK) in wood radial strip samples obtained at breast height (1.4 m) from 60 Pinus taeda trees. Calibration models were developed using traditional partial least squares (PLS) and kernel regression. The kernel methods included radial basis functions-partial least squares (RBF-PLS) and least-squares support vector machines (LS-SVM). RBF-PLS and LS-SVM models outperformed PLS-CV calibrations in terms of fit statistics. MFA and MOE, two properties that exhibited nonlinearity, showed the most significant improvements compared to PLS. In terms of predictive ability RBF-PLS performed better than PLS for the prediction of MFA, MOE, and COARS. LS-SVM showed better prediction statistics in all cases, except for WTHICK that gave similar statistics compared to PLS and was superior to RBF-PLS. By adding statistically significant factors to the PLS regressions, it was possible to capture some of the nonlinear features of the data and improve the predictive ability of the PLS models.