Bridging CNNs and vision transformers for efficient tea leaf phytopathogen diagnosis: GL-MobFormer.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Accurate identification of tea plant (Camellia sinensis) diseases is crucial for sustainable cultivation.
• Balancing diagnostic accuracy with computational efficiency for edge devices is a key challenge in agricultural phytosanitary monitoring.

Purpose of the Study:

• To develop a lightweight, hybrid deep learning framework for efficient and accurate tea plant disease identification.
• To address the challenge of deploying accurate diagnostic tools on resource-constrained agricultural edge devices.

Main Methods:

• Proposed GL-MobFormer, a hybrid deep learning framework integrating MobileNetV3 and Transformer Encoder.
• Utilized CutMix data augmentation for enhanced model robustness in field conditions.
• Evaluated the framework on a dataset of 5,278 tea leaf images across seven phytosanitary categories.

Main Results:

• Achieved 95.13% classification accuracy and a Matthews Correlation Coefficient (MCC) of 0.9417.
• Maintained a low computational footprint of 0.33 G FLOPs.
• Occlusion-based sensitivity analysis confirmed that the model focuses on pathological features, with a 70.61% confidence drop when critical regions were masked.

Conclusions:

• GL-MobFormer offers an optimal balance between diagnostic accuracy and computational efficiency for phytosanitary monitoring.
• The framework provides a practical and efficient solution for real-time, on-site disease detection in precision agriculture.