Mulberry leaf disease detection by CNN-ViT with XAI integration

Abstract

Mulberry leaf disease detection is vital for maintaining the health and productivity of mulberry crops. In this paper, a novel approach was proposed by integrating explainable artificial intelligence (XAI) techniques with a convolutional neural network (CNN) and vision transformer (ViT) for effective mulberry leaf disease classification with three disease classes. Initially, in this proposed CNN-ViT model, features are extracted using a customized CNN architecture, and then the extracted features are fed into ViT for leaf disease classification in a more streamlined approach. The CNN-ViT model achieved promising results with a projection dimension of 64, utilizing 8 heads and 8 transformer layers, yielding an accuracy of 95.60% with notable precision of 94.75%, recalls of 92.40%, and F1-scores of 93.45%. The proposed method also took 0.0017 seconds to predict an individual image. The accuracy of the proposed method was comparable to that of other state-of-the-art (SOTA) methods reported in the literature. Finally, Grad-CAM was utilized for detecting precise region of interest for diseased leaves, leaf spots, and leaf rust, providing interpretability and insights into the model’s decision-making process. This comprehensive approach demonstrates the effectiveness of explainable artificial intelligence (XAI) integration in the CNN-ViT model for mulberry leaf disease detection, paving the way for improved agricultural disease management strategies.