Abstract
Although early diagnosis of pomegranate fruit disorders reduces yield losses, we examined pomegranate image datasets that were cross-sectional, with each fruit captured only once. Three stage labels were created for healthy, early and established using weekly photographs of 233 tagged fruits (October 2024 to January 2025) from 10 trees, as well as a persistence rule where three or more consecutive observations of damage were necessary. PomStage-Net was used for split-conformal prediction with a single-image Swin-T classifier for referring ambiguous images to an expert. For fruit-wise cross-validation, the macro-F1 and early-stage recall were 0.733 and 0.737, respectively, on the 173 fruits that had consistent labeling sequences, while for the tree-grouped validation, the macro-F1 was 0.740, and the early-stage recall was 0.782. Fruit-level calibration at α = 0.10 led to 0.904 coverage (0.913 in early class), and 33.5% of the images were referred. The rest (66.5%) were labeled one, with a correct label for 89.6% and a confident wrong label for 6.9% of all images. The results for the 60 fruits that lacked a consistent label were significantly poorer (macro-F1 0.529). The model is able to detect early damage only when it is visible; it cannot predict damage prior to symptoms showing. The results are only valid for one orchard, season and cultivar.