Integrative Multivariate Analysis and YOLOv5-Based Deep Learning for Quality Evaluation of Underutilized Hyacinth Bean (Lablab purpureus) Landraces

Abstract

The present study investigates the comprehensive quality and compositional profiling of twelve underutilized Hyacinth bean (Lablab purpureus) landrace accessions through a multidisciplinary approach integrating traditional analytical techniques with modern deep learning tools. Initially, we conducted an extensive evaluation of engineering properties, proximate composition, antinutritional factors, and mineral content (including essential minerals and heavy metals) for each accession. To extract meaningful insights from the multivariate dataset, principal component analysis (PCA) was applied separately on proximate composition, polyphenols, antioxidant capacity, antinutritional compounds, essential minerals, and heavy metals. The PCA revealed trait contributions to the primary components (PC1/PC2), optimal clustering structures, and the positioning of accessions across clusters. Hierarchical clustering dendrograms and contribution plots of variables to major dimensions enabled the identification of the top 10 discriminating traits. Radar plots of trait mean per cluster and linear discriminant analysis based on PCA dimensions further strengthened trait-accession linkages and cluster separability. Parallelly, we implemented an automated, real-time quality evaluation framework using a deep learning-based YOLOv5 object detection model. The system was trained to classify hyacinth bean seeds into six quality-related categories: broken, discolored, infested, shriveled, sprouted, and stained. The model achieved high detection performance with a mean average precision of 95.3%, precision of 92.8%, recall of 95.1%, and F1-score of 0.950, enabling accurate, non-destructive, and rapid classification of defective beans. This AI-assisted quality assessment offers a scalable solution for grading large seed batches in supply chains. The integration of multivariate statistics and deep learning in this study provides a holistic framework for accession characterization, trait selection, and real-time quality monitoring. The findings have practical applications in crop improvement, gene bank management, seed grading, and precision agriculture. Future research will focus on expanding the model to other legumes, integrating hyperspectral imaging for chemical quality prediction, and deploying the system at industrial scale with real-time hardware support.