AppleLeafNet: a lightweight and efficient deep learning framework for diagnosing apple leaf diseases.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Accurate apple disease identification is vital for industry sustainability and apple quality.
• Analyzing complex leaf images for disease detection presents significant computational challenges.
• Existing deep learning models can be resource-intensive for practical field applications.

Purpose of the Study:

• To develop a novel, lightweight deep learning model for efficient apple leaf disease identification.
• To implement a two-stage framework for initial condition assessment and subsequent disease subclassification.
• To evaluate the model's performance using a publicly available dataset.

Main Methods:

• A custom 37-layer lightweight deep learning model was designed from scratch.
• The model was first trained to classify leaves as healthy or diseased.
• Transfer learning was applied using the trained model for subclassification of specific diseases (rust, complex, scab, frogeye).

Main Results:

• The two-stage framework achieved 98.25% accuracy in identifying apple leaf conditions.
• The model demonstrated 98.60% accuracy in diagnosing specific apple leaf diseases.
• The developed model is significantly lighter with fewer learnable parameters compared to pre-trained models.

Conclusions:

• The proposed lightweight deep learning model offers an effective and efficient solution for apple disease identification.
• The two-stage approach enhances diagnostic precision for various apple leaf conditions.
• This model presents a practical tool for improving apple production and disease management.