AI-driven hybrid vision transformer for automated disease classification in tomato (Solanum lycopersicum L.) crop under high altitude covered cultivation

Advanced artificial intelligence has the potential to completely transform crop disease management in agriculture, particularly for high-value crops like tomatoes. Using traditional disease detection techniques in automated, modern agriculture systems is inefficient. In the classification of plant diseases, recent developments in deep learning (DL), especially Vision Transformers (ViT), have shown great promise. However, the optimization of model hyper-parameters to maximize performance has been underexplored. This study presents a novel hybrid ViT model for real-time disease classification in tomato (Solanum lycopersicum L.) crops, focusing on optimizing key hyper-parameters - epochs, batch size, and learning rate, to improve the model’s precision, recall, F1 score and a Matthews Correlation Coefficient (MCC). Using response surface methodology (RSM), the influence of these parameters was systematically analyzed to enhance the model’s classification accuracy. The model’s performance was validated using a dataset of 1690 images of healthy and diseased tomato leaves, achieving a classification accuracy of 91.53%, with statistically significant F1, precision, and recall metrics. The model demonstrated inference latency of 634–681 ms per image, indicating the capabilities of the suggested framework to be used in near-real-time tomato disease classification.The model performance was cross-validated by employing the k-fold method and it was further compared with baseline models such as basic ViT, MobileNet, and VGG16, proving the proposed hybrid approach to be superior. This approach presents a scalable solution for automated disease management in tomato crops, particularly in high-altitude, covered cultivation environments.