An Ultra Lightweight Interpretable Convolution-Vision Transformer Fusion Model for Plant Disease Identification:...

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Plant diseases cause significant global crop losses, impacting the agriculture industry.
• Current IoT and AI solutions for plant disease detection face challenges with diverse crops, limited generalizability, and data scarcity.
• Existing deep learning models often lack explainability and struggle with real-world, in-field data.

Purpose of the Study:

• To propose a lightweight and efficient model, ConViTX, for plant disease classification.
• To enhance generalizability and explainability in AI-based plant disease detection systems.
• To develop a solution suitable for resource-constrained precision agriculture setups.

Main Methods:

• Developed 'ConViTX', a compact model fusing convolutional neural networks and vision transformers.
• Utilized Gradient Weighted Class Activation Maps and Locally Interpretable Model-Agnostic Evaluations for model explainability.
• Evaluated the model on four public datasets and a self-collected in-field maize dataset.

Main Results:

• ConViTX outperformed nine state-of-the-art deep learning methods.
• Achieved 98.8% accuracy on a maize dataset and 61.42% on raw drone-captured images.
• The model has only 0.7 million parameters and 0.647 billion operations per second.

Conclusions:

• ConViTX offers improved generalizability and explainability for plant disease classification.
• The model's lightweight nature makes it suitable for deployment in precision agriculture.
• This research contributes to more efficient and accessible AI solutions for crop disease management.