Comparison of discriminant methods and deep learning analysis in plant taxonomy: a case study of Elatine

Abstract

Elatine is a genus in which, flower and seed characteristics are the most important diagnostic features; i.e. seed shape and the structure of its cover found to be the most reliable identification character. We used a combination of classic discriminant methods by combining with deep learning techniques to analyze seed morphometric data within 28 populations of six Elatine species from 11 countries throughout the Northern Hemisphere to compare the obtained results and then check their taxonomic classification. Our findings indicate that among the discriminant methods, Quadratic Discriminant Analysis (QDA) had the highest percentage of correct matching (mean fit—91.23%); only the deep machine learning method based on Convolutional Neural Network (CNN) was characterized by a higher match (mean fit—93.40%). The QDA method recognized the seeds of E. brochonii and E. orthosperma with 99% accuracy, and the CNN method with 100%. Other taxa, such as E. alsinastrum, E. trianda, E. californica and E. hungarica were matched with an accuracy of at least 95% (CNN). Our results indicate that the CNN obtains remarkably more accurate classifications than classic discriminant methods, and better recognizes the entire taxa pool analyzed. The least recognized species are E. macropoda and E. hexandra (88% and 78% match).