Estimation of the physical wood properties of Pinus taeda L. radial strips using least squares support vector machines

Near infrared (NIR) spectroscopy offers a rapid method for estimating many important wood properties, including air-dry density, microfibril angle (MFA) and SilviScan estimated stiffness (E-L(SS)). Wood property calibrations may be improved by using non-linear calibration methods. In this study, we compare calibrations developed using partial least squares (PLS) regression and least-squares support vector machine (LS-SVM) regression, a relatively new technique for modelling multivariate, non-linear systems. LS-SVM regression provided the strongest calibration statistics for all wood properties. For an equivalent number of latent variables, the predictive performance of the MFA LS-SVM calibrations were superior to those of the corresponding PLS calibration, while predictive results for air-dry density and E-L(SS) were similar for both calibration methods.