The spectra of larch wood with different surface roughness were analyzed, and the NIR models suitable for different surface roughness were established, which provided a theoretical basis for improving the accuracy and universality of NIR models in predicting wood density.

Taking larch wood from Xinghuo Forest Farm in Heilongjiang Province as the research object, the near-infrared spectra of unpolished (M0), 150-mesh (M1) and 320-mesh (M2) samples were analyzed and studied. 11-point moving average smoothing, baseline correction and SG smoothing were used for spectral preprocessing to remove redundant spectral information. Manual selection, backward interval partial least squares (BiPLS) and synergy interval partial least squares (SiPLS) were used to complete band optimization. A single prediction model for different surface roughness and a mixed near-infrared model with three surface roughness samples were constructed.

The M0 sample contained more spectral information than the other two samples. Among the three pretreatment methods, the comprehensive evaluation of the modeling effect of SG smoothing pretreatment showed the best. The basic density prediction models of M0, M1 and M2 were established based on the three band optimization methods, and the band selection method of SiPLS had the best effect. For the three surface roughness samples of M0, M1 and M2, the validation set correlation coefficients R and RMSEP were 0.865 9 and 0.022 7, 0.766 0 and 0.021 4, 0.725 6 and 0.027 4, respectively. The prediction ability of the SIPLS mixed prediction model based on the mixture of three different roughness samples was better than that of the single model based on each roughness sample. For the three surface roughness samples of M0, M1 and M2, the RMSEP of the model decreased by 11%, 25% and 5%, respectively.

The NIR models based on the three kinds of surface roughness samples can achieve effective prediction of wood density, and the prediction accuracy of the model is M0 > M1 > M2. The SiPLS band selection method can be used to optimize the influence of surface roughness on the prediction model, and the mixed model established on this basis makes the NIR prediction model more universal. It provides a theoretical basis and technical support for the classification, optimization and fine utilization of wood.