Deep Learning Approaches for Jute Leaf Disease Detection Using CNN Models Final Year Design Project

Abstract

One of the major cash crops in Bangladesh is jute, which is crucial for the economy and supports millions of farmers. However, jute production often faces the risk of various leaf diseases, resulting in decreased yield and quality. Thus, the development of the ability to timely identify the disease is crucial. Currently, such methods are based on manual leaf image detection and characterize the disease through expert principles. This thesis suggests developing a deep learning-based system that would automatically identify and classify jute leaf diseases using convolutional neural networks. The dataset consisted of approximately 3,500 leaf images; a total of five classes could be identified: bug attack, spot disease, twisted leaf, yellow leaf, and healthy leaf. The research included training and analyzing four popular CNN models. The best results were achieved by DenseNet169 which reached nearly 88% accuracy. MobileNetV2 was slightly worse, reaching nearly 86%, and Xception achieved an 85% result. ResNet152 showed the lowest rate with approximately 74% accuracy. Beyond the immediate contribution for the possibility to characterize and detect the presence of diseases in jute leaves, the research can be used for further development for the mobile or web application that will be highly useful in Bangladesh.