Enhancing detection of common bean diseases using Fast Gradient Sign Method-trained Vision Transformers.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Common bean production in Tanzania faces significant threats from diseases like bean rust and anthracnose.
• Effective disease management hinges on timely and accurate early detection systems.

Purpose of the Study:

• To develop a robust deep learning model for early detection of common bean diseases.
• To enhance model reliability for real-world farm conditions, particularly in resource-constrained environments.

Main Methods:

• A Vision Transformer (ViT)-based deep learning model was developed and enhanced with adversarial training.
• A dataset of 100,000 annotated images was augmented using geometric, color, and FGSM perturbations to simulate field variability.
• The model was fine-tuned using transfer learning and validated via cross-validation.

Main Results:

• The adversarial training significantly improved the model's robustness against image perturbations.
• The fine-tuned ViT model achieved a high accuracy of 99.4% in disease detection.
• The study demonstrated the model's effectiveness for mobile-based plant disease diagnostics.

Conclusions:

• Integrating adversarial robustness is effective for enhancing the reliability of deep learning models for plant disease detection.
• The developed model shows promise for practical application in mobile-based disease diagnostics in resource-limited agricultural settings.
• This approach can aid in mitigating crop losses and improving food security in regions reliant on common bean production.