Identification of Cotton Leaf Curl Disease Using CNN and Vision Transformer

Cotton is one of the most widely cultivated crops in the world, with a large proportion grown in developing countries. For better cotton management and yield, deep learning techniques are developed in this work. Therefore, the aim of this paper is twofold: first to create an open-source dataset of healthy and diseased cotton leaves (leaf curl virus-affected). A new custom dataset and the training/validation/testing sets and the raw dataset themselves have been provided in the GitHub repository. Secondly, to develop image classification models based on Convolution Neural Networks (CNNs) through an initial baseline model and Vision Transformer (ViT) of the cotton leaves. It shows how the vanilla model for a vision transformer with the addition of existing algorithms such as shifted patch tokenisation and locality self-attention can be used in this context to give over 80% accuracy on an unseen testing dataset. Facebook Research's ConViT hybrid model with GPSA layers is also evaluated in this context, using the automatic and manual implementation from code, and has shown the “convit-base” model providing nearly 85% accuracy and better generalisation over the epochs of training than the CNN baselines and the ViT model.