Regional calibration models for predicting loblolly pine tracheid properties using near-infrared spectroscopy

Abstract

This study developed regional calibration models for the prediction of loblolly pine (Pinus taeda) tracheid properties using near-infrared (NIR) spectroscopy. A total of 1842 pith-to-bark radial strips, aged 19–31 years, were acquired from 268 trees from 109 stands across the southeastern USA. Diffuse reflectance NIR spectra were collected at 10-mm increments from the radial face of pith-to-bark strips using a FOSS 5000 scanning spectrometer. A subset of the 10-mm samples was selected based on the NIR spectra uniqueness and placed in calibration (N = 1020) and validation (N = 998) test sets. The samples were macerated and the tracheid properties measured using an optical analyzer (Techpap MorFi Fiber and Shive Analyzer). Models were created using partial least squares (PLS) regression using the calibration samples, and their performance checked using the validation set. Prediction PLS models for tracheid length were strong when checked with the validation set [R p 2 = 0.87, standard error of prediction (SEP) = 0.23 mm, ratio of performance to deviation (RPD) = 2.8]. Prediction models for tracheid width had moderate fit statistics (R2 = 0.61, SEP = 1.6 μm, RPD = 1.6), but the SEP was similar to the measurement error of the camera (± 2 μm). The NIR wavelengths of importance were largely attributed to cellulose. The 1542 nm wavelength explained 77% of the variation in tracheid length (RMSE = 0.31 mm). These results demonstrate that NIR spectroscopy models for tracheid length and width can be developed from a diverse set of samples and still maintain prediction accuracy.