CC BY
8THz-I-12 _________
_ 22
NONLINEAR AND TERAHERTS PHOTONICS
Classification of chocolates by multivariate methods
in THz spectroscopy
M. Khodasevich1, A. Lyakhnovich1, H. Eriklioglu2
1 - B.I.Stepanov Institute of Physics, National Academy of Sciences of Belarus, Minsk, Nezavisimosti Ave., 68-2
2 - Middle East Technical University, Ankara, Turkiye e-mail address: m.khodasevich@ifanbel.bas-net.by
Non-destructive and non-contact methods of diagnostics of the composition, quality and authenticity of food products are intensively developing in the world. Interest in these methods is due to the need for food safety and protection of the rights of consumers and producers. The THz frequency range has recently been added to the traditional UV, VIS and NIR spectroscopy. We study different types of chocolates (bitter, dessert, milk) from different manufacturers from Belarus, Russia, Turkiye and Ukraine. A description of the THz spectrometer used and measurement conditions is given in [1].
The aim of the work is to demonstrate the possibility of classifying chocolates by type and manufacturer by the "spectralprint" method [2]. Correction of transmission spectra baselines by the method of adaptive iteratively reweight-ed least squares with a penalty [3] is applied as a preprocessing procedure to the THz spectra. It allows to get rid of significant noise caused by Fabry-Perot effects and the presence of water vapor in the measuring path of spectrometer. The baselines of THz transmission spectra are used for multivariate analysis. Score plot in two-dimensional space of principal components [4] is shown in Figure 1 on the left. The symbols indicate the type of chocolates. Two principal components explain 95.6 % of the total variance of the baselines. The application of hierarchical cluster analysis [5] in this space makes it possible to classify 5 studied types of chocolates with precision (the ratio of true positive solutions of the classifier to the sum of true positive and false positive solutions) of 0.94 and recall (the ratio of true positive solutions to the sum of true positive and false negative solutions) of 0.93. The dendrogram of the classifier is shown in Figure 1 on the right where serial numbers correspond to the left figure. This figure is useful when analyzing cluster agglomeration and shows erroneous decisions of the classifier. The obtained precision and recall values are sufficient for practical application for the classification of chocolates. The applied methods are based on the a priori hypothesis of compactness of the location of objects in the space of the principal components of the baselines of THz transmission spectra of various types of chocolates.
2
.3 • 1
•6
• 4 •24
16 O 5 • J #12 23 19 ••
Ol8 9 oil 22
• 21
15 017 8o «10 20
O 13° 014 25 <P o 30 26
270 °29
28
_O
-0.2 -0.1 0 0.1 0.2 0.3 first principal component (88.8 %}
E
I 2 3 6 4 16 13 14 15 18 17 5 7 11 12 8 9 10 25 26 30 23 27 28 19 20 21 22 23 24 serial number of the sample
Fig. 1. Score plot in two-dimensional space of principal components of THz spectra baselines and dendrogram of hierarchical cluster analysis with selection of 5 types of chocolate.
[1] M. Khodasevich, A. Lyakhnovich, H. Eriklioglu, Chocolate sample classification by preprocessed THz transmission spectra analysis, Journal of Applied Spectroscopy, vol. 89, pp 251-255 (2022).
[2] R. Ríos-Reina [et al.], Spectralprint techniques for wine and vinegar characterization, authentication and quality control: Advances and projections, Trends in Analytical Chemistry, vol. 134, pp. 116121:1-21 (2021).
[3] Z. M. Zhang [et al.], Baseline correction using adaptive iteratively reweighted penalized least squares, Analyst,vol.135, pp. 1138-1146 (2010).
[4] R. Bro, A.K. Smilde, Principal component analysis, Analytical Methods, vol. 6, pp. 2812-2831 ( 2014).
[5] T.W. Liao, Clustering of time series data - a survey, Pattern Recognition, vol. 38, pp. 1857-1874 (2005).
