LeafSightX: an explainable attention-enhanced CNN fusion model for apple leaf disease identification

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Accurate apple leaf disease identification is vital for precision agriculture to prevent yield loss.
• Existing deep learning models often lack real-world applicability, interpretability, and statistical validation.

Purpose of the Study:

• To develop and validate LeafSightX, a novel deep learning framework for precise and interpretable apple leaf disease identification.
• To enhance model resilience and transparency through specialized preprocessing, limited data augmentation, and explainable AI techniques.

Main Methods:

• LeafSightX employs a dual-backbone architecture combining DenseNet201 and InceptionV3 with Multi-Head Self-Attention (MHSA).
• The framework incorporates specialized preprocessing, limited data augmentation, and Grad-CAM for explainable AI (XAI).
• Rigorous evaluation involved a five-class apple leaf disease dataset and an independent dataset for robustness testing.

Main Results:

• LeafSightX achieved a test accuracy of 99.64% and an F1-score of 0.9962 on the primary dataset.
• AUC and PR-AUC scores reached 1.000, with cross-validated Cohen's Kappa (mean = 0.9917) indicating high consistency.
• The model demonstrated real-time inference capabilities and achieved 99.69% accuracy on an independent dataset, proving robustness and generalization.

Conclusions:

• LeafSightX offers a highly accurate, interpretable, and robust solution for apple leaf disease identification in precision agriculture.
• The framework's real-time performance and generalization capabilities support practical AI application in agriculture.
• This rigorously evaluated system provides a reproducible foundation for AI-driven plant disease management.