Non-destructive ripeness classification of apricot (Prunus armeniaca L.) using a physically informed deep learning approach

Accurate ripeness assessment is essential for optimizing harvest timing, storage management and commercial grading in climacteric fruit supply chains. However, conventional methods rely on destructive physicochemical measurements that limit scalability and real-time application. Here we develop a physically informed deep learning framework for non-destructive apricot (Prunus armeniaca L.) ripeness classification by integrating physicochemical clustering with image-based analysis. Ripeness labels were defined using combined hardness, °Brix and color parameters, with an independent firmness-based subset used for external validation. A ResNet18 convolutional neural network trained on segmented and augmented images achieved 88.6% accuracy in cross-validation (Macro F1 = 0.888). Performance declined to 63.08% accuracy (Macro F1 = 0.459) on the independent subset, reflecting the continuous nature of ripening transitions. CNN prediction scores were significantly associated with physicochemical ripeness indicators (R² = 0.442, p < 0.001), demonstrating that learned visual features capture biologically meaningful ripening signals and enabling more reliable AI-assisted fruit grading.