TomatoGrow-3: A Composite Dataset and Benchmark for Tomato Growth-Stage Classification using Transfer Learning and Fine-Tuning

Tomatoes are one of the most widely cultivated and economically significant crops worldwide, making their growth-stage classification critical for optimizing irrigation, harvesting, and yield management. Accurate identification of tomato growth stages such as flowering, fructification, and maturity is essential for enhancing productivity and supporting precision agriculture. However, most existing tomato growth-stage image classification datasets focus on the final phase of development primarily ripened, semi-ripened, and unripened samples while neglecting earlier stages such as flowering. Real-time monitoring during early developmental phases, including leafing and flowering, is vital for preventing pest or pathogen-induced disruptions that can affect subsequent stages like fruit set and ripening. To address this gap, this study developed a new tomato growth-stage dataset consisting of flowering, fructification, and mature classes. The dataset was evaluated and benchmarked using multiple popular pretrained deep convolutional neural networks (CNNs) under transfer learning (TL) and fine-tuning (FT) strategies. A baseline Vanilla CNN was first implemented, achieving 97.24% accuracy, 97.30% precision, 97.24% recall, and a 97.21% F1-score, which established a strong foundation for comparison. Further experiments using TL and FT techniques across five pretrained architectures demonstrated significant performance improvements. Specifically, FT models with a data-augmentation setting consistently outperformed TL counterparts, with ResNet50, ResNet101, and VGG16 each achieving an accuracy of 98.74%. The results obtained indicate the critical role of FT in adapting pretrained models to domain-specific agricultural imagery. The high accuracies obtained across architectures also confirm the strong generalization capability of the AI models, even when trained on the proposed dataset derived from multiple sources.