Multiple crop disease detection with optimized Vision Transformers

The field of agriculture has witnessed a growing need for efficient and accurate methods for plant disease diagnosis to ensure crop health and maximize yield. This paper introduces an optimized technique for detecting plant disease using the vision transformer (ViT) model. The proposed system leverages the power of ViTs which are modern deep learning architecture that has demonstrated recognizable success in various computer vision real time problems. The proposed Vision Transformer combines local feature extraction strengths of convolutional neural networks (CNNs) with local and global feature extraction of transformers, offering a robust and scalable solution for plant disease detection. Experimentation carried out on multiple crops images to evaluate performance based on number of parameters and training time. The Vision Transformers are optimized with hyperparameters like input size, number of layers, batch size, patch size etc. The paper presents promising results for early-stage diagnosis and correct detection of plant diseases in comparison with traditional convolutional neural networks. The proposed model will help in farming by timely intervention and resource optimization for sustainable agriculture development. The paper presents research work for fifteen disease classes of three crops. The proposed model is achieving better accuracy of 0.97 whereas popular convolutional neural network VGG16 is of 0.87 and ResNet is 0.90.