Cotton Leaf Disease Detection Using Vision Transformers: A Deep Learning Approach

Abstract

Cotton is one of the world's most significant positions as a key cash crop used to make textiles. It also produces cottonseed, which is used as oil and animal feed. Despite being economically significant, cotton crops are prone to various diseases, which can lower both yield and quality. Timely detection of these diseases is essential for reducing losses and saving crops, which helps to grow the economy of a country. The widespread occurrence of cotton leaf disease presents a significant challenge for farmers in India, China, and Pakistan. This paper provides a detailed overview of the methods used to collect both healthy and disease-affected cotton leaves, manually annotate (binary and multiclass), preprocess, and analyze a large dataset of cotton leaf images. The main objective of this dataset was to facilitate automated disease detection systems using advanced artificial intelligent techniques. We explored the data collection process, distribution of the dataset, preprocessing steps, feature extraction techniques, and potential applications. In addition, we present the initial results of our analyses and highlight the importance of such datasets in advancing agricultural technology. It not only helps to protect cotton crops, but also supports economic improvement and enhances the success of cultivation. To achieve this objective, we used five pre-trained deep learning and two advanced transformer-based techniques to predict the best solution for cotton leaf disease. Our proposed methodology (ViT) achieved the highest achievable performance with an accuracy rate of 96.72% in the binary class and 93.39% in the multi-class.