Image-based seed phenotyping for characterization and discrimination of basil (Ocimum basilicum L.) breeding materials derived from local genetic resources
Abstract
Reliable phenotypic characterization of plant genetic resources is essential for germplasm evaluation and breeding use. Here, 2750 single-seed images from 11 genotype-designated basil breeding materials (TRB2–TRB12; 250 seeds per material), produced in 2021 and 2022 at a single site, were evaluated using a spatially calibrated 1200-dpi flatbed-scanner workflow. From 288 image-derived descriptors, 44 SD/CV-derived variables were excluded, leaving 244 hand-crafted predictors. Four classical machine-learning models, four deep-learning architectures, and two fixed 50:50 decision-fusion strategies were evaluated using five locked maternal-seed-lot-grouped outer folds. The highest mean outer-fold Macro-F1 was obtained by the ML-family mean + DL-family mean fusion (0.831 ± 0.012); ConvNeXtV2-Tiny and SwinV2-Tiny each achieved 0.798, and CatBoost was the strongest classical model (0.770 ± 0.014). No paired Wilcoxon comparison remained significant after Holm correction. Bidirectional production-year transfer with the classical models yielded lower Macro-F1 values (0.6439–0.6983), indicating reduced transferability across production years. Maternal-lot cluster bootstrap analyses quantified uncertainty without treating individual seeds as independent replicates. Grouped permutation importance, held-out SHAP, out-of-fold t-SNE, and architecture-specific Grad-CAM were used as interpretive diagnostics rather than causal evidence. The framework supports non-destructive germplasm documentation, breeding-material comparison, and preliminary seed-lot screening, but broader cultivar-authentication or genetic-identity claims require validation across additional environments, germplasm panels, imaging systems, and laboratories.