Discriminação dos planos anatômicos em tábuas de Pinus sp. com dispositivo NIR portátil visando à produção de lápis

Abstract

The classification of wood according to anatomical planes is relevant for pencil manufacturing, as it contributes to organizing the raw material into more homogeneous groups. This separation is important because variability along the log may influence wood behavior during processing and negatively affect the performance and final quality of pencils. Therefore, the potential of near-infrared spectroscopy (NIR), combined with multivariate analysis techniques, was evaluated to differentiate Pinus sp. boards into three anatomical plane classes: (1) longitudinal-radial, (2) outer longitudinal-tangential plane (near the bark), and (3) inner longitudinal-tangential plane (near the pith), aiming at sorting and quality control in production. A total of 540 NIR spectra were obtained using a portable trinamiX Palm NIR device from 90 wood samples measuring 190 × 80 × 4 mm, distributed into three anatomical classes, with 30 samples per class and six collection points per sample. Of the total spectra obtained, 70% (n = 378; 126 per class) were used for calibration and 30% (n = 162; 54 per class) for external validation/prediction. PCA was applied as an exploratory analysis, and PLS-DA models were fitted and evaluated by cross-validation (LOOCV) and external validation. Spectral preprocessing methods were compared, including raw data, SNV, Savitzky–Golay derivative, and combinations, and model performance was assessed by specificity (SPE), sensitivity (PRE), and accuracy (ACU); in external validation/prediction, performance was assessed by the percentage of correct classifications (CCV). The models showed high discrimination ability among the planes. Even without preprocessing, consistent predictive performance was observed, with CCV = 95.7% and mean accuracy of 97.0%. The use of SNV alone reduced performance, with CCV = 92.0%. Although the derivative reached maximum classification metrics, it showed lower consistency in prediction, with CCV = 63.0%, suggesting lower generalization ability. In contrast, the combinations of derivative + SNV provided the best performance in external validation, with CCV = 100%. These results indicate that NIR spectroscopy associated with PCA/PLS-DA is an efficient approach for classifying Pinus sp. boards according to orthogonal anatomical cutting planes, with potential for rapid sorting and quality control in pencil production, highlighting the importance of prioritizing external validation metrics when selecting the preprocessing method.