Artificial intelligence-driven classification method of grapevine major phenological stages using conventional RGB imaging

Accurate monitoring of grapevine phenological stages is essential for optimising vineyard management. This study evaluates the performance of three deep learning architectures (ResNet-34, YOLOv11-Classification and Vision Transformer (ViT)) for automated classification of vineyard canopy images into four key phenological stages: i) Shoot and inflorescence development (E-L 12–18), ii) Flowering (E-L 19–26), iii) Berry formation (E-L 27–33), and iv) Berry ripening (E-L 35–38). These categories correspond to broad developmental periods that may span several E-L stages. A dataset comprising 4,381 images was used to train and validate the models, incorporating data augmentation techniques to improve robustness. Results indicate that all three models achieved high classification accuracy, with ResNet-34 obtaining the highest accuracy (97.4 % validation, 95.6 % test), reinforcing its strong feature extraction capabilities. However, its lower F1-score (95.3 % validation, 91.8 % test) suggests challenges in handling class imbalances. YOLOv11-Classification demonstrated the most balanced classification performance, achieving a high F1-score (93.6 % validation, 91.8 % test) while maintaining the fastest training time, making it particularly suitable for real-time applications. ViT exhibited competitive classification performance but had higher computational demands, limiting its feasibility for real-time vineyard monitoring. A confusion matrix analysis highlighted misclassification trends, particularly between early shoot development and flowering, due to their visual similarities. Despite these challenges, the study confirms that AI models can effectively automate vineyard phenology classification, reducing manual assessment efforts, and contributing to more efficient viticultural decision-making.