A Novel Deep Learning Approach for Tomato Leaf Disease Detection Using Optimized CNN Architecture

Abstract

Early disease detection in plants is essential for sustaining agricultural yield and guaranteeing food security. A comparative analysis of transformer-based and convolutional based deep learning models for classifying tomato leaf diseases is presented. Specifically, it examines the performance of a Vision Transformer (ViT), tested both in a form of scratch training setup and through transfer learning, against well-known CNN architectures such as Inception V3, VGG16, ResNet50, and a custom-designed lightweight CNN.  This is one of the few studies to rigorously benchmark ViT against CNNs in the context of agricultural disease detection using the PlantVillage dataset. The fine-tuned ViT model delivered the best results, achieving an accuracy of 95.53%, significantly outperforming all CNN counterparts. The lightweight CNN demonstrated strong performance with 93.12% accuracy, while offering clear benefits in terms of smaller model size and reduced computational cost making it well-suited for on-device or edge-level applications. Conversely, the ViT model trained from scratch underperformed due to dataset constraints, reinforcing the necessity of transfer learning for transformer architectures. Evaluation metrics included recall, accuracy, F1-score, and precision, which collectively illustrated the trade-off between high-capacity models and deployment feasibility. The main contribution of this work lies in introducing transformer-based learning into the plant pathology domain and the presentation of a scalable, low-computation alternative via lightweight CNNs. Future directions involve enlarging the dataset, integrating explainable AI techniques, and enabling real-time applications for precision agriculture.