AG-Vision: a dual-module approach for tomato leaf disease diagnosis

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Accurate tomato leaf disease identification is vital for precision agriculture.
• Convolutional Neural Networks (CNNs) excel at local features but struggle with global context, limiting real-world robustness.
• Existing methods often lack the ability to model both local details and global relationships effectively.

Purpose of the Study:

• To develop a hybrid deep learning architecture, AG-Vision, that integrates local and global feature learning for robust tomato leaf disease identification.
• To enhance the model's spatial awareness and ability to capture long-range dependencies.
• To evaluate AG-Vision's performance, efficiency, and interpretability in diverse datasets.

Main Methods:

• Proposed AG-Vision, a dual-module framework combining an EfficientNet-B4 CNN (DeepFolia) for local features and a Transformer encoder (VisiLeaf) for global context.
• Incorporated positional encoding and optimized attention heads to improve spatial understanding.
• Evaluated on PlantVillage (controlled) and PlantDoc (real-world) datasets, with ablation studies and Grad-CAM for interpretability.

Main Results:

• AG-Vision achieved state-of-the-art results: 99.97% accuracy and 99.53% F1-score on PlantVillage; 96.97% accuracy and 94.47% F1-score on PlantDoc.
• The model demonstrated real-time efficiency with ~25 ms inference time per image.
• Ablation studies confirmed the benefits of combining CNN and Transformer modules, positional encoding, and attention mechanisms.

Conclusions:

• Fusing local and global feature learning significantly improves classification accuracy and robustness for plant disease detection.
• AG-Vision provides an efficient and scalable solution for edge deployment in precision agriculture.
• The hybrid approach enhances disease identification capabilities under varied field conditions.