Classification of Horsetails Using Predictive Modelling on NIR Spectra

Le résumé fourni par la source

ABSTRACT Common horsetail (Equisetum arvense L., syn.: field horsetail) holds a long tradition in the supportive treatment of numerous diseases. A frequently observed problem is the risk of confusing Equisetum arvense plants with another closely related species Equisetum palustre (syn.: marsh horsetail) due to its morphological similarities. The distinction between the two species during collection/harvest is further complicated by the fact that both species share similar habitats. This, however, is of particular importance because E. palustre contains toxic alkaloids (palustrine and palustridiene) while this is not the case for E. arvense used for medicinal purposes (Equiseti herba). The aim of this study was the classification of horsetails using near infrared spectroscopy (NIR). Therefore, over 370 E. arvense and E. palustre samples originating from all over Germany, consisting of 2 years of harvest, were analysed using two different devices from different manufacturers: (a) a miniature (portable) NIR device and (b) a benchtop NIR device. Initial unsupervised machine learning techniques (PCA and t‐SNE) provided insightful visualizations for the distribution of both species within the data space. After applying variable screening to the spectral data, a variety of supervised machine learning models based on different algorithms were trained to predict the species from an individual spectrum. In a repeated cross‐validation (CV) approach, it could be shown that the spectra from both spectrometers are sufficient to achieve classification accuracies around 90%. Additionally, the data allowed for discriminating between harvesting seasons as well. The success of the complete workflow is further emphasized by assessing its reliability through posterior probabilities, which were high for the predicted class labels, implying a satisfying model certainty.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.