Near infrared calibration models for the estimation of wood density in Pinus taeda using repeated sample measurements

Near infrared [NIR) diffuse reflectance was used for the estimation of air-dry density and basic density in wood radial strip samples obtained at breast height (1.4m) from 60 Pinus taeda trees established in three progeny tests in the south-eastern United States. NIR calibration models were fitted using raw spectra and pre-processed spectra with second derivative, multiplicative scatter correction and orthogonal signal correction. Successful calibrations were obtained for both wood properties using data collected in consecutive 10 mm sections from the samples. Data pre-processing did not result in model improvements compared to the models fitted using raw data. The effects of using repeated measures were evaluated by incorporating serial correlation into the partial least squares regression algorithm. The empirical autocorrelation of the normalised residuals showed that serial dependence among residuals was successfully removed by using an autoregressive correlation structure of second order. However, because the initial dependence among observations was not strong, the predictions were similar using the modified algorithm to those obtained with the traditional approach. These results indicate that the use of repeated measurements does not represent a serious problem for the development of NIR calibration models for the prediction of wood properties using radial samples measured in 10mm sections and that the specification of the correlation structure may not be required when the models are used only for predictive purposes.