This study evaluated the use of stacked machine learning for wood species identification based on Laser Induced Breakdown Spectroscopy data. The approach combined three types of Support Vector Machine classifiers (linear, polynomial, and radial kernels) using a One-vs-All strategy. The outputs were combined through a Partial Least Squares Discriminant Analysis meta-learner. Variable selection was performed using Principal Component Analysis loadings to reduce dimensionality and improve model performance.
The experimental dataset comprised wood samples from 18 distinct species, with species verification performed through wood anatomical analysis or Direct Analysis in Real Time-Time of Flight Mass Spectrometery (DART TOFMS). Samples were sourced from three collections: the U.S. Forest Service International Programs and Trade (USFS IPT) Wood Identification & Screening Center (WISC), the U.S. National Fish and Wildlife Forensics Laboratory (USNFWFL), both located in Ashland, Oregon, and the U.S. Geological Survey (USGS) Reston Stable Isotope Laboratory (RSIL) on loan to the University of Delaware. This dual-source approach enabled the development of classification models incorporating samples from both highly controlled (WISC and USNFWFL) and standard laboratory (University of Delaware) environments, thereby enhancing the robustness of the model by accounting for natural variations in sample conditions.
The method significantly improved classification accuracy compared to flat models, demonstrating its potential for field-deployable, non-invasive wood identification systems supporting regulatory enforcement and conservation efforts.
