Answers are grounded in this paper and cite their sources.
Reliable cultivar identification is central to managing heritage apple collections, but routine genetic verification is resource-intensive and expert pomological assessment is difficult to scale. We compared non-destructive cultivar-recognition approaches using controlled multi-view FruitPhenoBox images from a multi-season, multi-location collection. Expert-designed descriptors of fruit color, texture, shape, and size were evaluated against deep-feature classifiers and transfer-learning image encoders under leakage-safe apple-level splitting, tree-holdout evaluation, class imbalance, increasing label-space size, and targeted cross-year/cross-location transfer. In the 38-class Baseline Subset, the strongest image encoder reached 94.9% Top-1 accuracy under tree-holdout evaluation. When the task was expanded within the primary 2021 site-year to 346 accession labels with limited and imbalanced support, the best image models reached approximately 82–83% Top-1 accuracy, whereas the expert-feature model dropped below 40%. In transfer tests from this 2021 site-year, models performed worse on later seasons at the same site than on the second site in the same season. Multi-season training and limited target-year tree calibration partly recovered this loss, indicating that practical use requires target-domain validation rather than reliance on single-season training. Feature-family analyses, binary-mask controls, feature-importance rankings, and qualitative saliency visualizations were consistent with peel color, color variability, and fine-scale surface texture carrying much of the discriminative signal; fruit outline alone was insufficient. Residual errors involved visually similar accessions, including synonym groups and a sport-mutant pair, highlighting the difficulty of strict label-level scoring in germplasm collections. These results support controlled fruit imaging as a decision-support approach for genebank screening, accession triage, and targeted molecular verification.
No takes yet. Share an insight, caveat, or question.
Wróbel et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: