EDISP: a hybrid CNN-ViT framework for robust maize leaf disease detection and classification

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Maize foliar diseases cause significant crop losses, necessitating timely and accurate identification.
• Manual disease inspection is subjective and inefficient.
• Deep learning models like CNNs and ViTs offer automated diagnosis but have limitations: CNNs focus on local features, while ViTs require large datasets and may miss fine details.

Purpose of the Study:

• To develop and evaluate EDISP, a novel hybrid deep learning framework integrating CNNs and ViTs for enhanced maize foliar disease detection.
• To overcome the limitations of standalone CNNs and ViTs by leveraging their complementary strengths for both local feature extraction and global contextual learning.

Main Methods:

• The EDISP framework combines CNNs for local feature extraction and ViTs for global contextual learning.
• A diverse dataset of controlled-environment and real-field maize leaf images was utilized.
• Data preprocessing included normalization, augmentation, and stratified splitting for robust model training and validation.

Main Results:

• EDISP achieved superior performance over standalone CNN and ViT models, with an overall accuracy of 99.40%.
• The model demonstrated high precision (99.43%), recall (99.38%), and F1-score (99.40%) in identifying various maize diseases and healthy leaves.
• External validation confirmed EDISP's robustness and generalizability to real-world conditions with minimal false positives/negatives.

Conclusions:

• The hybrid EDISP architecture offers a powerful solution for accurate and automated maize leaf disease detection.
• EDISP shows promise for precision agriculture, providing a scalable tool for farmers lacking expert diagnostic resources.
• Further research should address potential limitations related to image quality, lighting, and novel disease variations.