Detection techniques for tomato diseases under non-stationary climatic conditions

Area of Science:

• Plant Pathology
• Computer Vision
• Machine Learning

Background:

• Tomato diseases pose a significant threat to crop yield, necessitating accurate and timely detection.
• Deep learning models, like YOLO, struggle with domain shifts when applied to new environments, leading to performance degradation.
• Existing methods often fail to address the continuously evolving nature of real-world agricultural data.

Purpose of the Study:

• To develop a continuous test-time domain adaptation framework (CTTA-DisDet) for robust tomato disease detection.
• To enhance the generalization capability of pre-trained models in unseen and evolving environments.
• To improve the practical applicability of automated disease detection systems in agriculture.

Main Methods:

• Proposed CTTA-DisDet framework employing a teacher-student architecture for domain adaptation.
• Introduced dynamic data augmentation, including explicit image corruption and implicit data generation using large language models (LLMs).
• Implemented neuron weight restoration to mitigate catastrophic forgetting and exponential moving average (EMA) for continuous teacher model updates.

Main Results:

• CTTA-DisDet achieved 67.9% performance in continuously changing cross-domain environments.
• Demonstrated significant improvement in generalization across unseen domains compared to baseline methods.
• Validated the effectiveness of dynamic augmentation and neuron restoration techniques.

Conclusions:

• CTTA-DisDet offers a robust solution for tomato disease detection in non-stationary agricultural settings.
• The framework effectively adapts to domain shifts, enhancing model reliability.
• This approach holds significant promise for practical, real-world agricultural applications requiring adaptive AI systems.