Reference-Free Classification of Terahertz Reflection Spectra with Few Selected Frequencies

Terahertz (THz) spectroscopy enables non-invasive material classification but traditionally requires reference spectra, limiting practical applications. This study applies feature selection to achieve accurate, reference-free classification using Logistic Regression (LR), Naïve Bayes (NB), and Support Vector Machine (SVM). Our results show that SVM outperforms other classifiers, achieving 99.3% accuracy with just eight selected features (~1% of the original data). While NB struggles due to its independence assumption, SVM and LR demonstrate superior robustness. These findings suggest that THz classification can be optimized using a minimal number of spectral features, potentially enabling more efficient hardware designs utilizing single-frequency electronic THz sources.