Utilizing convolutional neural networks and vision transformers for precise corn leaf disease identification

Corn is not only widely used in industry but also a crucial staple food. Early detection of corn leaf diseases is vital to prevent crop loss. Farmers and agricultural engineers often rely on computer-aided systems for early diagnosis of plant diseases. Among the various methods, deep learning stands out as the most popular and effective approach for detecting corn leaf diseases. In this study, we utilized cutting-edge Vision Transformer (ViT) models like MaxViT, DeiT3, MobileViT, and MViTv2, which have recently gained more traction compared to Convolutional Neural Networks (CNNs). Additionally, we incorporated well-known CNN architectures such as VGG, ResNet, DenseNet, and Xception to accurately diagnose corn leaf diseases. To enhance the models’ effectiveness, we employed image preprocessing, data augmentation techniques, transfer learning, and optimized parameters. Furthermore, we implemented a soft voting ensemble technique with an adaptive thresholding method to dynamically boost performance, leading to higher accuracy and balanced metrics in detecting corn diseases. Our approach was trained and evaluated on both the well-known PlantVillage dataset and the novel CD&S dataset. The results showed that four models from the MaxViT architecture, along with other deep learning models, achieved a high accuracy of 100% on the CD&S dataset’s test data, the highest performance recorded in the literature. On the PlantVillage dataset, the approach attained an impressive 99.83% accuracy, surpassing other studies. This proposed method offers an early and autonomous solution for diagnosing corn plant diseases in the agricultural field with high accuracy. This innovation highlights the potential of advanced ViT models to outperform traditional CNNs and improve crop disease detection.