Fusedvit: A Hybrid Vision Transformer and Convolution Model for Accurate Crop Disease Classification in Digital Agriculture

With the increase in population, raising crop yields to meet the growing demand for food has become a livelihood issue of public concern. However, crop diseases have long been a key factor in jeopardizing crop yields. In the past, crop health could only be monitored manually, which not only required significant time and labor costs but also faced the problem of poor classification accuracy. It also does not conform to the trend of the construction of digital agriculture. To alleviate the above problems, this paper proposes a novel and efficient classification model named Fusedvit utilizing deep learning techniques. In Fusedvit, the preprocessed images are fed into the convolution modules and the self-attention modules for feature extraction, respectively. The model utilizes the self-attention modules to obtain global representations and pipes the pooling strategy to enhance the local relevance of global information. Thereafter, Fusedvit associates different feature representations by effectively combining the global representation learning capability of Vision Transformer and the inductive biases of convolution and optimizes the training using a loss function optimization strategy to obtain accurate classification results. In an effort to illustrate the generalizability and stability of the model, experimental validations are conducted in this paper to validate the model on various crop disease datasets. Fusedvit achieves 99.231%, 95.262% and 97.333% results in accuracy on the validated datasets, outperforming the other comparative models and fully demonstrating the model’s excellent performance.