Chocolate Sample Classification by Principal Component Analysis of Preprocessed Terahertz Transmission Spectra

Abstract

We demonstrate the efficiency of the chocolate sample classification by type and manufacturer using the “spectral print” method using THz transmission spectra. To suppress the noise and the Fabry–Perot effect, spectra baselines are determined using the adaptive iteratively reweighted penalized least squares (airPLS) method. The classification was carried out by constructing a low-dimensional space of the principal components of the baselines and applying the methods of cluster analysis in this space. The precision and recall values of the classification of chocolate samples by the k-means, classification and regression tree and hierarchical cluster analysis are 0.85 and 0.83, 0.91 and 0.90, 0.94 and 0.93, respectively. The support vector machine is successfully applied to consider two cases where pairwise classification is most problematic.