Clustering and classification of soybean leaves based on angle features

Abstract

Introduction Leaf shape is a genetically determined crop phenotype, and its accurate classification underpins soybean germplasm assessment and genetic improvement. Manual classification is highly subjective and struggles to distinguish morphologically similar leaves, while mainstream supervised classification demands large labeled datasets and incurs high development costs. Efficient feature frameworks for soybean leaf categorization are still insufficient. Methods In this study, 581 biologically replicated terminal leaflets sampled from 194 soybean varieties were analyzed at the single-leaflet level using traditional morphological indices and novel leaf contour angular features. Unsupervised K-means clustering was used to classify soybean leaflet morphological phenotypes; t-SNE was applied exclusively for dimensional reduction visualization, while Welch's ANOVA combined with Games-Howell post-hoc tests was adopted to detect inter-cluster phenotypic differences. Clustering stability and external consistency against manual visual labeling were further quantified via Adjusted Rand Index to comprehensively verify the reliability of grouping outputs. Results The results revealed no significant difference in leaflet edge complexity (p = 0.41) between two manually divided leaf groups distinguished by overall leaf outline similarity; these two morphologically similar leaf clusters failed to be fully separated even though the first two principal components accounted for 90.2% of total variance. For K-means clustering, k = 3 achieved better overall performance with a Calinski–Harabasz (CH) index of 395.55, Davies–Bouldin (DB) index of 1.03, and silhouette coefficient (SC) of 0.38, compared with k = 4. Nevertheless, the angular feature attained an F-value of 951.62 in driving sample reallocation across clusters, serving as the core indicator for fine subdivision at k = 4. Under k = 4 clustering, all six morphological indices differed significantly among the four groups (p < 0.05). Additionally, the number of cross-clustered samples increased from 66 to 119 as k rose from 3 to 4, with 96.6% of cross-clustering attributed to the leaflet contour angular feature. Discussion This research provides a novel reference and technical support for the automated identification and fine classification of soybean leaf morphology.

Plant phenotyping relevance

大豆小葉の形態表現型を角度特徴量とクラスタリングで自動分類する手法が研究の中心であり、検証指標も明示されているため。

Code and data availability

The supplied blocks describe 581 scanned soybean leaflet images and a Python/OpenCV clustering pipeline, but contain no public deposit of the image dataset, phenotype data, or analysis code. The Data availability statement is not included in the supplied text, and the only link is the generic Frontiers supplementary材料

No evidence-backed public reproduction asset is currently recorded.