Accurate prediction of cross-species wood density by fusing terahertz time-domain spectroscopy and explainable machine learning.

This study employed terahertz time-domain spectroscopy (THz-TDS) to acquire spectral signals of wood samples with different densities and extract their refractive indices. Wood density prediction models were developed using three machine learning algorithms: Elastic Net Regression (ENR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). The Uninformative Variable Elimination (UVE) algorithm was used for the feature selection of refractive index spectra, and the Particle Swarm Optimization (PSO) algorithm was applied to optimize model hyperparameters. Meanwhile, the Shapley Additive exPlanations (SHAP) method was employed to perform interpretive analysis on the optimal model. The results showed a positive correlation between wood density and THz refractive index. In the task of predicting the wood density of fast-growing tree species, the XGBoost model demonstrated excellent performance, with a test set coefficient of determination (R2) of 0.9594. When the dataset was expanded to include fast-growing wood species and high-density Pterocarpus wood species, the test set R2 increased to 0.9846, fully validating the universality and effectiveness of the method in cross-species wood density prediction. After feature selection using the UVE algorithm, the test set R2 was further improved to 0.9862, while significantly reducing computational complexity. SHAP feature importance analysis revealed that the refractive index at 0.216 THz made the greatest contribution to the wood density prediction model, and the top 9 key feature frequencies were all located in the 0.2-0.3 THz low-frequency band, all exhibiting significant positive impacts on the prediction model. This study demonstrates that the integrated application of terahertz time-domain spectroscopy technology and machine learning algorithms provides an innovative solution for rapid and precise detection of wood density.