Pyramid Vision Transformer and Context-Aware CNN for Accurate and Explainable Bean Leaf Disease Classification

Abstract

: Two major problems with bean production that result in yield loss are bean rust and angular leaf spot. Though the older procedures require specialist knowledge, timely disease detection promotes production. Combining Pyramid Vision Transformer (PVT) and Group Context Aware Depthwise Shuffle Network (GCADSN) demonstrates an explainable deep learning model. The GCADSN focuses on the acquisition of small, context-specific characteristics, while the PVT efficiently models long-range relationships, identifying disease trends across larger leaf areas. By combining these methods, the input picture is presented richer, which allows for better categorization of illnesses. Grad-CAM visuals help to explain the model by focusing on the regions of the picture that are important for the model's predictions and enable class-specific understandings. A dataset called IBean, which consists of a number of photos of various behaviors of bean leaves including common rust, angular leaf spot, and healthy leaves, was used to comprehensively test the model's ability to deliver accurate results. With a high accuracy of 97.66%, our proposed network outperformed some of the most advanced deep learning models available today. With a 97.67% F1 score, 97.83% accuracy, and 97.67% recall, it also performed well in terms of other important criteria. In the study by B. Yang(2024), Y. Wang(2024) and J. Wang(2024), the computational effectiveness of the model and its ability to assess the early, mid, and late stages of illness development ensures its use for agricultural field deployment and timely intervention in the real world.