MoSViT: a lightweight vision transformer framework for efficient disease detection via precision attention mechanism

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Maize is a vital global crop facing substantial yield reductions from diseases.
• Current diagnostic methods for maize diseases are often slow, subjective, and hinder effective management.

Purpose of the Study:

• To develop an advanced machine learning model, MoSViT, for accurate and efficient classification of maize diseases.
• To improve upon existing diagnostic tools by leveraging computer vision and deep learning.

Main Methods:

• Developed MoSViT, a classification model based on the MobileViT V2 framework, incorporating CLA focus, DRB module, MoSViT Block, and LeakyRelu6 activation.
• Trained MoSViT on 3,850 images of maize diseases (Blight, Common Rust, Gray Leaf Spot) and healthy samples.
• Evaluated model performance using metrics including accuracy, precision, recall, and F1 score, and performed heatmap analysis for interpretability.

Main Results:

• MoSViT achieved high performance metrics: 98.75% accuracy, 98.73% precision, 98.72% recall, and 98.72% F1 score.
• Outperformed established models like Swin Transformer V2, DenseNet121, and EfficientNet V2 in accuracy and parameter efficiency.
• Demonstrated strong generalization capabilities on small sample datasets, indicating potential for real-world applications.

Conclusions:

• MoSViT offers a highly accurate and efficient solution for maize disease diagnosis.
• The model's interpretability through heatmap analysis aids in understanding its diagnostic process.
• MoSViT shows promise for small-sample detection and timely crop management strategies.