Low-Data Metric-Learning Phenomic Framework for Interpretable Cultivar Identification and Similarity Analysis in Panax ginseng

Abstract

Image-based cultivar identification remains challenging in perennial medicinal crops because cultivar-specific datasets are often small and morphological differences can be subtle. This study investigated whether a Siamese network-based hybrid learning framework could support cultivar classification and image-derived phenomic similarity analysis in Panax ginseng under low-data conditions. A total of 347 images representing 22 cultivars across four above-ground image acquisition categories were evaluated using stratified five-fold cross-validation. The framework jointly optimized class-weighted cross-entropy and contrastive losses, and five ImageNet-pretrained backbone architectures were assessed across contrastive margins. ConvNeXt-Tiny with a margin of 0.50 achieved the highest five-fold mean performance, with an accuracy of 64.31 ± 7.23% and a macro-F1 score of 61.66 ± 6.63%. UMAP visualization indicated qualitative reorganization of the embedding space after fine-tuning, while Grad-CAM++ localized model responses mainly to plant structures, including leaf, stem, and fruit regions, rather than broad background areas. Hierarchical clustering of cultivar embeddings further suggested structured phenomic relationships, with cosine distance and average linkage yielding a cophenetic correlation of 0.81 ± 0.08 and Kendall’s τ-b of 0.61 ± 0.06. These findings support the potential of hybrid classification and metric learning as a complementary tool for extracting interpretable phenomic representations from limited ginseng image datasets. The framework may provide supporting image-based evidence alongside conventional morphological cultivar assessment. However, the observed cultivar relationships should be considered exploratory and require validation across environments, developmental stages, independent datasets, and genetic information before broader biological interpretation.