A hybrid deep learning framework using convolutional and transformer models for robust plant disease classification

Abstract

Abstract Plant diseases continue to pose a significant threat to worldwide food security, resulting in notable yield reductions and economic consequences. Automated disease diagnosis through machine learning has arisen as a potential solution; nevertheless, current methods frequently have difficulty in capturing both detailed local attributes and overarching contextual patterns found in plant leaf images. This study presents a thorough comparative examination of conventional and deep learning methods—such as Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), YOLO, Support Vector Machines (SVMs), and Random Forests—for the classification of multi-class plant diseases. To overcome the constraints of individual CNN and transformer models, a new hybrid framework that integrates EfficientNet-B7 for strong spatial feature extraction with a Vision Transformer (ViT-B16) for comprehensive contextual modeling is suggested. The system is assessed on an extensive dataset consisting of 21,534 images covering 38 classes of plant diseases and healthy specimens. Experimental findings show that the suggested hybrid model reaches an accuracy of 98.13%, surpassing standalone CNN baselines and other rival models, while consistently achieving high precision, recall, and F1-scores for all classes. The results emphasize the success of combining convolutional and transformer-based models for scalable and precise plant disease detection, aiding the creation of smart decision-support systems for precision farming.

Plant phenotyping relevance

植物葉画像から病害状態を分類する新規ハイブリッド画像解析手法を提案し、複数手法との比較評価と大規模データセットでの検証を行っており、フェノタイピング手法が中心である。

Code and data availability

The paper's plant disease image dataset (New Plant Diseases Dataset on Kaggle) and the authors' complete hybrid CNN–ViT implementation (GitHub repository with Zenodo DOI) are both publicly and explicitly available.