Explainable deep learning-based comparative study for guava fruit and leaf disease classification: advancing...

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Early detection of guava diseases is crucial for preventing significant yield losses and ensuring agricultural sustainability.
• Deep learning models offer potential for disease identification but face challenges in accuracy and interpretability for practical use.

Purpose of the Study:

• To develop an explainable deep learning framework for classifying guava fruit and leaf diseases.
• To enhance the accuracy and interpretability of AI models for agricultural disease diagnosis.

Main Methods:

• A dataset of 527 annotated guava images across five classes was used.
• Six hybrid deep learning models were created by combining transfer learning backbones (VGG16, MobileNetV2, InceptionV3, ResNet50) with custom CNN classifiers.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was used for model interpretability.

Main Results:

• The VGG16 + MobileNetV2 hybrid model achieved the highest performance with 96% accuracy and an F1-score of 0.96.
• The model demonstrated strong generalization across all disease classes.
• Comparative analyses confirmed the superior performance of the proposed hybrid model.

Conclusions:

• Combining deep feature extractors with lightweight architectures improves classification accuracy and efficiency.
• Explainable AI techniques like Grad-CAM increase trust and interpretability in AI-driven disease diagnosis.
• The framework shows potential for real-time smart farming and mobile diagnostic applications, especially in resource-limited settings.