Tree species identification based on ATR-FTIR spectra in case of extensive intraspecies spectral variations

Algorithmic identification of tree species based on FTIR spectra without sample pretreatment is challenging due to extensive intraspecies spectral variance, driven by spatial tissue heterogeneity, encompassing heartwood and sapwood distinctions, and environmental modulation of xylogenesis. This study presents a robust, hierarchical machine learning methodology for direct on-site species classification, evaluated on a dataset of 6443 spectra derived from 112 trees across 22 hemiboreal species. Analysis was restricted to the 1800 to 800 cm⁻¹ fingerprint region, where the linear separability of species was established and Linear Support Vector classifier achieved the highest predictive accuracy. Primary binary discrimination accuracy between softwood and hardwood species was 100% within the narrow 1300 to 1200 cm⁻¹ range, demonstrating absolute zero-shot generalization for taxonomically unseen species. Subsequent pairwise classification was empirically optimized within discrete, highly informative spectral windows. Within these tailored ranges, 91% of hardwood and 60% of softwood species pairs achieved greater than 99% identification accuracy, with the spruce and pine pair exhibiting the lowest discrimination accuracy (91.3%). The finalized algorithmic pipeline integrates the primary softwood/hardwood discriminator with these targeted pairwise verification models. Across 20 independent test iterations, the presented method correctly identified 97.5% of the evaluated trees, yielding unidentified and misclassification rates of 1.6% and 0.9%, respectively.