Files
Abstract
This dissertation has four chapters focused on using near-infrared (NIR) hyperspectral imaging as a rapid, nondestructive evaluation tool for precise estimation of various properties in longleaf pine and loblolly pine wood (chapters 1-3), and loblolly pine forest residues (chapter 4). The first chapter compared 31,314 ring-level extractives content predictions from a deep learning model and a gradient boosting machine learning model (LGBM) with those from a partial least squares regression (PLSR) model using NIR spectral data as the input. Results showed that LGBM provided biologically realistic values for extractives content. In the second chapter, the ring-level predictions from LGBM were modeled in a non-linear mixed effects framework to explain variability in extractives content as a function of stand type, cambial age, and disk height. Results showed that extractives content was highest at the pith and then decreased with cambial age. At the pith, extractives content decreased with height within tree with naturally regenerated longleaf pine trees having the least variation in height and old agricultural field planted longleaf pine having the most variation in height. The third chapter aimed to explain the within-tree variation in cellulose and lignin content in longleaf pine and loblolly pine trees. Results indicated that cellulose content increased whereas lignin content decreased with an increase in cambial age for any given disk height. At any given cambial age, cellulose content increased with height up to a certain point, this point was different for trees from different stand origins, then decreased. Lignin content increased nominally with increasing height in naturally regenerated longleaf pine trees whereas it decreased with increasing height for trees from other stand types. In the last chapter, calibration models capable of rapid and precise estimation of chemical, proximate, and ultimate composition were developed from samples collected from different physiographic regions and age groups. PLSR models showed good fit statistics for most of the properties measured, with 9 of 18 properties having Rcv2 > 0.9, thus demonstrating that NIR hyperspectral imaging is a rapid and reliable tool for estimation of properties which strongly correlate to pyrolysis bio-oil yield.