TomatoGrow-3: A three-stage tomato phenological benchmark with systematic evaluation of transfer learning and fine-tuning

Tomato is one of the most extensively cultivated vegetable crops worldwide and plays a significant role in global agriculture and food production. Reliable identification of its developmental stages is fundamental for supporting precision agriculture practices, including irrigation scheduling, nutrient management, disease monitoring, harvest planning, and yield optimization. Despite the growing availability of tomato image datasets, existing resources predominantly focus on fruit maturity classification, while earlier developmental stages, particularly flowering, remain comparatively underrepresented. To address this gap, this study presents TomatoGrow-3, a three-stage phenological benchmark comprising Flowering, Fructification, and Mature classes, constructed through a structured annotation, harmonization, and quality-control protocol. Operational class definitions, mixed-stage exclusion rules, fruit-colour criteria, and image-quality controls were applied to establish consistent phenological labels. The benchmark was systematically evaluated using eleven pretrained deep-learning architectures under transfer-learning (TL) and fine-tuning (FT) strategies, together with a custom convolutional neural network (CNN) trained from scratch. Data augmentation (DA) was further evaluated to examine its influence on classification performance under the different training strategies. The custom CNN achieved 97.24% accuracy, 97.30% precision, 97.24% recall, and 97.21% F1-score after DA. Among the pretrained configurations, ShuffleNetV2 with FT and DA achieved the highest numerical held-out accuracy of 98.99%, with precision, recall, and F1-score of 99.00%, 98.99%, and 98.99%, respectively. However, marginal differences among the highest-performing architectures were not interpreted as evidence of definitive model superiority. Overall, the comparative results indicate that effective tomato phenological classification depends not solely on architectural complexity but also on pretrained representation transfer, target-domain adaptation, and augmentation, with lightweight architectures remaining competitive with substantially larger models.