Hybrid Deep Learning Approach for Sweet Orange Leaf Disease Detection Using CNN and Vision Transformers

Abstract

Sweet orange leaf diseases significantly threaten agriculture, necessitating accurate and timely detection for sustainable farming. This study presents a hybrid deep learning approach combining Vision Transformers (ViT) and Convolutional Neural Networks (CNN) for classifying sweet orange leaf diseases. The methodology includes data preprocessing, such as resizing, normalization, and augmentation, to enhance dataset quality and prepare it for deep learning models. Three models—ViT, ResNet50v2, and the hybrid ViT-CNN—were implemented and evaluated. The hybrid ViT-CNN model achieved the highest test accuracy of 98%, surpassing the individual performances of ViT (90%) and ResNet50v2 (97%), with consistent training and validation accuracies of 97%. The hybrid model integrates the localized feature extraction of CNNs with the global contextual capabilities of ViTs, enabling superior disease classification. This research highlights the scalability and robustness of the hybrid approach, addressing dataset scarcity and computational efficiency challenges. Implemented on Google Colaboratory, the system is optimized for deployment in resource-constrained environments, ensuring accessibility for small-scale farmers. The findings contribute to precision agriculture by reducing crop losses, minimizing pesticide use, and promoting sustainable practices. This study establishes a reliable framework for agricultural disease detection, paving the way for advancements in AIdriven solutions for broader crop management applications.