Explainable deep learning model for automatic mulberry leaf disease classification.

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Sericulture, responsible for 90% of global raw silk, faces reduced yields due to mulberry leaf diseases.
• Manual identification of these diseases is time-consuming and prone to errors.
• No deep learning models currently exist for detecting mulberry leaf diseases.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for accurate and efficient identification of mulberry leaf diseases.
• To address the limitations of manual disease diagnosis in sericulture.
• To create an explainable AI tool for sericulture specialists.

Main Methods:

• Collected and annotated images of healthy and diseased mulberry leaves (leaf rust, leaf spot).
• Utilized image augmentation to generate 6,000 synthetic images from an initial dataset of 764 images.
• Developed a lightweight Parallel Depth-wise Separable Convolutional Neural Network (PDS-CNN) model.
• Applied SHapley Additive exPlanations (SHAP) to ensure model explainability.

Main Results:

• The PDS-CNN model achieved high accuracy: 95.05% for three-class and 96.06% for binary classification.
• The model is lightweight, featuring only 0.53 million parameters, 8 layers, and a size of 6.3 MB.
• PDS-CNN outperformed established deep transfer learning models in accuracy, parameter count, layer count, and size.
• SHAP visualizations confirmed the model's predictions aligned with expert assessments.

Conclusions:

• The explainable AI-based PDS-CNN offers an effective solution for accurate mulberry leaf disease categorization.
• This technology can significantly aid sericulture specialists in early disease detection and yield improvement.
• The developed model demonstrates the potential of lightweight, explainable AI in agricultural applications.