APPLICATION OF NEAR-INFRARED SPECTROSCOPY HYPERSPECTRAL IMAGING USING MACHINE LEARNING MODELS TO PREDICT WOOD CHEMICAL PROPERTIES

Carbon concentration varies within a loblolly pine tree with the proportion of extractives, lignin and cellulose. Because traditional chemical methods are laborious, this research aims to utilize near-infrared spectroscopy (NIR) hyperspectral imaging to better understand the drivers of within-tree variation in carbon concentration. The extractives, lignin and cellulose concentrations were measured from radial pith-to-bark samples to calibrate NIR models. NIR models were built using partial least squares (PLS) regression and Light Gradient-Boosting Machine (LGBM) to predict proportions of chemical properties. The LGBM models, having higher R2 and lower RMSE than PLS, were slightly better at predicting extractives and lignin content in the independent prediction set. Carbon concentration in extractive-free wood averaged 46.8% but accounting for both carbon in wood and extractives resulted in a mean carbon concentration of 47.8%. A positive correlation was found between lignin and carbon concentration and a negative with cellulose concentration. The findings highlight the application of NIR to explain variations in carbon.