Enhanced Maize Leaf Disease Detection and Classification Using an Integrated CNN-ViT Model.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Maize crop productivity is threatened by foliar diseases, necessitating efficient detection methods.
• Traditional classification techniques struggle with the complexity of disease-affected leaf imagery.
• Automated disease detection is crucial for timely intervention and optimal crop management.

Purpose of the Study:

• To develop a robust hybrid deep learning framework for enhanced maize leaf disease classification.
• To overcome the limitations of traditional methods in capturing complex visual patterns.
• To improve the accuracy and reliability of automated maize disease detection.

Main Methods:

• A hybrid deep learning framework integrating Convolutional Neural Networks (CNNs) for local feature extraction and Vision Transformers (ViTs) for contextual dependencies.
• Utilizing self-attention mechanisms within the ViT module to capture long-range relationships.
• Concatenating features from CNN and ViT modules, followed by fully connected layers for classification.

Main Results:

• The proposed hybrid CNN-ViT model achieved a validation accuracy of 99.15% with precision, recall, and F1-score of 99.13%.
• 5-fold cross-validation demonstrated high average accuracies: 99.06% on the combined Kaggle + Mendeley dataset and 95.93% on the Corn Disease and Severity (CD&S) dataset.
• The hybrid model outperformed standalone CNNs, indicating superior performance and generalization capabilities.

Conclusions:

• The hybrid CNN-ViT model offers a reliable and highly accurate solution for maize leaf disease identification.
• The framework's effectiveness across multiple datasets highlights its potential for real-world agricultural applications.
• Improvements in stability and performance were noted due to dropout regularization and the RAdam optimizer.