A Smart Agriculture Framework for Apple Leaf Disease Detection Using YOLOv8 and Swin Transformer Networks

Abstract

Detecting diseases in apple leaves plays a vital role in maintaining agricultural productivity, output quantity, and product standards. Enhancing the quality of cultivated crops and increasing their yield can be achieved by reducing losses and adopting better disease management strategies through effective disease diagnosis. Currently, the primary approach to diagnosing plant diseases involves conventional techniques, including the examination of leaves and branches by an experienced professional. This process is not only lengthy and expensive but also prone to inaccuracies due to human disruptions. Over the years, learning methods have developed and shown notable progress in enhancing techniques to combat plant diseases. This research differs primarily in utilizing the YOLOv8s method and a hybrid ResNet50 plus Swin Transformer model with attention features to target the classification of apple leaf diseases. The method enhances performance in object decomposition and disease classification by leveraging Transformer models for global context and CNNs for extracting local features. With a total accuracy of 98.56%, the hybrid system also achieved a weighted average F1-score of 98.56% and a macro average F1-score of 98.59%. Since disease category A achieved an F1-score of 99.66%, it suggests that hybrid plant disease monitoring is highly effective in managing Cedar Apple Rust specifically. The confusion matrix and the precision-recall analysis offered additional evidence supporting this highly uniform classification, which exhibited almost no errors across the different disease categories. The proposed hybrid framework achieved improved classification accuracy, enhanced feature extraction, and superior prediction performance relative to the current YOLOv8s model. The suggested hybrid framework obtained strong classification performance by capturing both basic spatial features and more complex spatial patterns with contextual details, despite YOLOv8s being trained as an end-to-end model that performed well in localization and detection tasks.