Deep Learning for Tomato Leaf Disease Classification: Comparative Benchmarking of CNN and Vision Transformer Architectures

Abstract

Tomato production is highly vulnerable to foliar diseases that can reduce yield, increase management costs, and complicate timely intervention. Automated image-based diagnosis has therefore become an important research direction for precision agriculture. In this study, we present a comparative evaluation of four representative deep learning backbones for tomato leaf disease classification on the Plant Village dataset: EfficientNetV2-S, ConvNeXt-Base, DeiT3-Base, and Swin-Base. The dataset comprised 18,160 images from ten classes, including nine disease categories and one healthy class, and was divided into training, validation, and test sets using a 70:15:15 split. All models were trained under a standardized transfer learning pipeline with identical preprocessing, augmentation, and optimization settings to enable a fair comparison across architectures. Performance was assessed using accuracy, precision, recall, F1-score, parameter count, and GFLOPs. All evaluated models achieved very high classification performance, with test accuracies of at least 0.9985. Among them, Swin-Base yielded the best overall predictive performance, reaching an accuracy of 0.9989 and an F1-score of 0.9987. In contrast, EfficientNetV2-S provided the most favorable efficiency profile, achieving 0.9985 accuracy with only 20.19 million parameters and 5.4193 GFLOPs. These findings indicate that both convolutional and transformer-based models can deliver highly reliable tomato leaf disease classification under controlled benchmark conditions, while the final model choice should be guided by the application scenario. Swin-Base is preferable when maximum predictive performance is prioritized, whereas EfficientNetV2-S offers a more practical option for computationally constrained deployments.