A Multi-Considered Seed Coat Pattern Classification of Allium L. Using Unsupervised Machine Learning

Abstract

The seed coat sculpture is one of the most important taxonomic distinguishing features. The objective of this study is to classify coat patterns of Allium L. seeds into new groups using scanning electron microscopy unsupervised machine learning. Selected images of seed coat patterns from more than 100 Allium species described in literature and data from our samples were classified into seven types of anticlinal (irregular curved, irregular curved to nearly straight, straight, S, U, U to Ω, and Ω) and five types of periclinal walls (granule, small verrucae, large verrucae, marginal verrucae, and verrucate verrucae). We used five unsupervised machine learning approaches: K-means, K-means++, Minibatch K-means, Spectral, and Birch. The elbow and silhouette approaches were then used to determine the number of clusters required. Thereafter, we compared human- and machine-based results and proposed a new clustering. We then separated the data into six target clusters: SI, SS, SM, NS, PS, and PD. The proposed strongly identical grouping is distinct from the other groups in that the results are exactly the same, but PD is unrelated to the others. Thus, unsupervised machine learning has been shown to support the development of new groups in the Allium seed coat pattern.