[PLSR model based on near-infrared spectroscopy for the detection of wood fiber anatomy of Schima superba.]

To rapidly acquire fiber phenotypic data for wood quality assessment, we used a portable NIR spectro-meter to collect spectral data in 100 individuals of Schima superba at 18-year-old of 20 different provenances, and simultaneously collected wood cores. Wood basic density and the anatomical structure of wood fiber were measured. The standard normal variate (SNV), orthogonal signal correction (OSC), and multiplicative scatter correction (MSC) methods were used for spectral preprocessing, the competitive adaptive reweighted sampling (CARS) method were used for wavelength selection, and the partial least squares regression (PLSR) model were established. The results showed a significant difference for the absolute reflectance data between forest and indoor environments, and the spectral data of which were relatively independent. SNV, OSC and MSC showed significant differences for predictive performance of the model. OSC had the excellent preprocessing capability in multiple cha-racteristics of wood fiber ether in forest and indoor environments. The predictive accuracy of the models with R² was 0.47-0.78 in forest (average=0.63), and R² was 0.54-0.82 in indoor environment (average=0.71). However, the SNV and MSC methods could not establish the models, except the fiber wall-cavity ratio from forest data. After wavelength selection through the CARS method, the predictive accuracy of the models was significantly improved using both forest and indoor data (R²=0.58 and 0.72, respectively). When performed OSC before and after CARS, the predictive accuracy of the models was improved to 0.68 and 0.84 respectively using forest and indoor data. The OSC and CARS could significantly improve the accuracy of the models for wood fiber anatomical structures. First OSC, then CARS, and finally OSC methods could be used to establish the PLSR model for fiber length, fiber cell wall thickness, fiber lumen diameter, wood basic density, fiber cavity-width ratio, and fiber wall-cavity ratio, and the R² ranged from 0.80 to 0.95. These models had effective predictive ability and accuracy to assess the physical properties of wood fibers of S. superba.