A comparative study of vision transformers and convolutional neural networks: sugarcane leaf diseases identification

Abstract

Diseases in agricultural products cause significant decrease on harvest efficiency and economic values of the products, early detection of diseases can prevent this loss. The development of artificial intelligence has brought its use in the field of agriculture. These practices have facilitated the work of farmers and increased productivity. Sugarcane is one of the agricultural crops with high economic value, and in this study, diseases in sugarcane leaves were classified by using deep learning methods. The dataset we use contains a total of 2521 images and there are 5 classes; healthy, mosaic disease, redrot disease, rust disease and yellow leaf disease. DenseNet121, one of the convolutional neural network (CNN) models, is applied to this dataset first, followed by the Vision Transformers (ViT) model, and finally the ViT + CNN combination is applied, and the results are compared. As a result of the observations, it is understood that the precisions of 92.87%, 93.34%, and 87.37%, respectively, were obtained.