Comparative and explainable deep learning for cherry (Prunus avium L.) cultivar classification using convolutional neural networks (CNN) and transformer architectures

Accurate identification of cherry ( Prunus avium L.) cultivars is essential for automated sorting, postharvest management, and cultivar-specific quality control in modern horticultural systems. However, fine-grained classification remains highly challenging due to strong phenotypic similarity among cultivars and the influence of intra-class variability under imaging conditions. In this study, a comparative and explainable deep learning framework was developed to classify eight cherry cultivars, including both registered cultivars and selection cultivars, using three representative architectures: ResNet50, EfficientNet-B0, and Swin Transformer V2. A stratified five-fold cross-validation strategy was employed to provide a consistent fold-wise evaluation while preserving class balance. Model performance was assessed using accuracy, confusion matrices, receiver operating characteristic (ROC) curves, area under the curve (AUC), and Gradient-weighted Class Activation Mapping (Grad-CAM) for interpretability. Among the evaluated models, EfficientNet-B0 achieved the highest classification accuracy (0.970), followed by Swin Transformer V2 (0.955) and ResNet50 (0.895). EfficientNet-B0 achieved an AUC of 1.00 for all evaluated classes in the present dataset and exhibited the most stable training convergence among the evaluated architectures. Misclassification was primarily observed between phenotypically similar cultivars, particularly 10–22 and Lapins, highlighting the intrinsic complexity of fine-grained classification. Grad-CAM analysis revealed that EfficientNet-B0 produced the most spatially focused and fruit-centered attribution maps, whereas ResNet50 showed diffuse activations and Swin Transformer V2 captured broader contextual features. Overall, the results demonstrate the feasibility of deep learning-based fine-grained cherry cultivar classification under controlled imaging conditions. EfficientNet-B0 achieved the highest mean accuracy among the evaluated architectures in the present dataset; however, this result should be interpreted as dataset- and condition-specific rather than as evidence of general architectural superiority. The proposed framework provides an interpretable basis for future validation using independently acquired and more diverse datasets. The proposed framework provides a reproducible and interpretable comparative approach for cultivar-level image recognition and may support future intelligent sorting and precision-agriculture applications.