Vision transformer meets convolutional neural network for plant disease classification

Plant diseases are the primary cause of crop losses globally, which have an impact on the world economy. To deal with these issues, new agriculture solutions are evolving that combine the Internet of Things and machine learning for early disease detection and control. Most of these systems use machine learning and computer vision for real-time disease identification and diagnosis. With advancements in deep learning techniques, various methods have emerged that employ convolutional neural networks for plant disease detection and identification. More recently, vision transformers have attracted the attention of researchers due to their strikingly better performance in classification problems in different vision-based applications. Accordingly, researchers have begun to explore vision transformers for plant pathology applications as well. In the present work, a hybrid model is proposed that combines the strength of a vision transformer with the inherent feature extraction capability of convolutional neural networks for disease identification using plant leaves. It can efficiently identify a large number of plant diseases for several crops. The proposed model has a lightweight structure with only 0.85 million trainable parameters, which makes it suitable for IoT-based agriculture systems. The performance of the proposed model is compared against nine state-of-the-art techniques on five publicly available datasets. The model is shown to outperform all nine methods even under challenging background conditions. On 'PlantVillage' dataset it achieves 98.86% accuracy and 98.9% precision, while on 'Embrapa', it shows 89.24% accuracy and 91.17% precision. On small datasets too its performance is better than other competing methods. Explainability of the proposed model is evaluated using gradient-weighted class activation maps and local interpretable model agnostic explanations.