Classification of THz pulse signals using two-dimensional cross-correlation feature extraction and non-linear...

Area of Science:

• Spectroscopy and Spectroscopic Imaging
• Machine Learning Applications
• Terahertz (THz) Science and Technology

Background:

• Terahertz (THz) spectroscopy is sensitive to material properties like extinction coefficient, refractive index, and scattering.
• Similar optical properties of substances can make classification challenging based solely on complex insertion loss.
• Scattering effects in spectroscopic experiments can be difficult to quantify, limiting direct classification accuracy.

Purpose of the Study:

• To compare the performance of four machine learning classifiers (MLR, KNN, SVM, NB) for analyzing THz transient time domain sequences.
• To evaluate the effectiveness of 2-D cross-correlations in noise suppression and feature extraction for hyperspectral datasets.
• To establish a general methodology for assessing hyperspectral dataset classifiers using 2-D cross-correlations.

Main Methods:

• Utilized terahertz (THz) transient time domain spectroscopy on pixelated images of powder samples with varying thicknesses (2mm, 3mm, 4mm) and mixtures.
• Applied two-dimensional (2-D) cross-correlations between background and sample interferograms to generate statistical features.
• Performed cross-validation by classifying samples and mixtures using multinomial logistic regression (MLR), k-nearest neighbours (KNN), support vector machine (SVM), and naïve Bayes (NB) classifiers.

Main Results:

• MLR (least accuracy 88.24%) and KNN (least accuracy 90.19%) consistently outperformed SVM (least accuracy 74.51%) and NB (least accuracy 56.86%) in classification accuracy and robustness.
• The use of 2-D cross-correlations provided effective noise suppression and generated robust features for classification.
• Classification accuracy was assessed across different sample thicknesses and complex mixtures, confirming classifier consistency.

Conclusions:

• MLR and KNN are highly effective and robust classifiers for hyperspectral THz imaging data, particularly when enhanced by 2-D cross-correlation analysis.
• The developed methodology using 2-D cross-correlations offers a reliable approach for evaluating other hyperspectral dataset classifiers.
• This work supports the advancement of automated THz imaging systems for applications in biomedical imaging, industrial processing, and quality control.