Classification of the geographical origin of medicinal herbs using multivariate spectral analysis

Abstract

Classification of the geographical origin and manufacturer of medicinal herbs was carried out by multivariate analysis of the optical density spectra of 70 % alcohol tinctures in the wavelength range 230–2600 nm using the example of chamomile from Russia and Belarus. Principal component analysis, classification and regression tree method, and spectral variable selection were used to build the models. The principal component analysis allows one to significantly reduce the dimension of the feature space. Classification and regression trees are being constructed in it. The maximum number of principal components considered is limited to 10, which made it possible to describe more than 0.999 of the total dispersion of the measured spectra. Classification and regression trees with tenfold cross-validation classify the country of origin of samples in a four-dimensional space and the manufacturer in a three-dimensional space of the principal components of broadband optical density spectra with an accuracy of more than 0.93. Ranking the spectral variables in decreasing order of the absolute value of the average deviation of optical density from the average value made it possible to improve the accuracy of classification models. A reliable classification of the geographical origin of chamomile is achieved in the space of principal components of 20 variables out of 2623 available in the broadband spectra. The manufacturer’s classification accuracy was improved to 0.94 by selecting 14 spectral variables.