Physiological and image-based characterization of Botrytis cinerea infection progression in postharvest 'Haegeum' kiwifruit using interpretable deep learning.

Gray mold caused by Botrytis cinerea severely limits the postharvest storage and marketability of kiwifruit, creating a need for accurate and non-destructive infection assessment. In this study, a symptom based, ROI-aware, and interpretable deep-learning framework was developed to classify infection levels in golden-fleshed 'Haegeum' kiwifruit. Infection severity was categorized into four levels based on external visual symptoms: Level 1: no lesions; Level 2: slight discoloration; Level 3: localized mycelial growth; Level 4: extensive mycelial growth and tissue breakdown. The associated physiological responses were subsequently characterized for each level. Disease progression induced fruit softening with firmness decreasing from 4.82 at Level 1 to 2.79 N at Level 4, while respiration rates increased from 0.05 to 0.24 µmol kg-1 s-1. Ethylene production began at Level 3 and sharply increased at Level 4, reaching 0.015 µmol kg-1 s-1. To quantify lesion severity, a region-of-interest-aware pipeline using RGB blue and LAB L* and a* channels was developed, enabling lesion-focused visualization of infection. Compared with the baseline model, the proposed approach improved classification performance, achieving accuracies of 90.3% with EfficientNet-B0, ResNet-50, and ShuffleNet-V2, and 87.6% with ResNet-18. Gradient-weighted class activation mapping further indicated that the ROI-aware pipeline reduced background-related activation and improved the spatial confinement of model attention within fruit regions. These findings demonstrate that symptom-based RGB classification, supported by post hoc physiological characterization and explainable deep-learning analysis, provides a biologically interpretable framework for monitoring visible fungal infection progression in postharvest kiwifruit.