Integrating terahertz time-domain spectroscopy with XGBoost for rapid and interpretable species-level wood identification of Pterocarpus

Overexploitation of global forest resources has led to severe survival pressure on endangered Pterocarpus species, and the illegal trade of their wood poses a serious threat to biodiversity and market order. To achieve accurate identification of Pterocarpus wood, this study integrated terahertz time-domain spectroscopy (THz-TDS) with gradient boosting algorithms to develop classification models for seven Pterocarpus wood species. The results showed that the XGBoost model performed best, achieving 100% accuracy in binary classification (P. santalinus and P. tinctorius) and 98.63% in three-class classification (P. indicus, P. macrocarpus, and P. soyauxii). After screening the THz frequency bands and performing feature selection on THz refractive indices using the Uninformative Variable Elimination (UVE) method, the seven-class classification accuracy of the constructed UVE-XGBoost model was improved to 88.64%, confirming that the 0.1-0.3 THz band is the most important frequency range for Pterocarpus wood classification models. SHAP interpretability analysis further revealed that 0.106 and 0.107 THz are the key characteristic frequencies for identifying the seven Pterocarpus wood species. This study demonstrates that THz-TDS combined with the XGBoost algorithm can achieve rapid and accurate identification of Pterocarpus wood, providing effective technical support for endangered wood protection and market supervision.