MangoLeafNet-XAI: an attention-enhanced deep learning architecture for accurate and interpretable mango leaf disease...

Area of Science:

• Agricultural Automation
• Deep Learning
• Plant Pathology

Background:

• Mango leaf diseases significantly impact crop yield and quality.
• Existing deep learning models are often too computationally intensive for resource-constrained agricultural settings.
• Precise and efficient disease detection is crucial for automated agricultural systems.

Purpose of the Study:

• To develop a lightweight deep learning model for accurate mango leaf disease detection.
• To address the limitations of computational complexity in current agricultural AI models.
• To create a model suitable for deployment on edge devices in smart agriculture.

Main Methods:

• Proposed MangoLeafNet-XAI, a lightweight architecture combining Efficient Channel Attention (ECA) with a DenseNet-121 backbone.
• Evaluated the model using 5-fold cross-validation and soft-voting ensemble on three diverse public datasets (MLDID, Mango Leaf Disease, Harumanis).
• Employed explainable AI (XAI) techniques like Grad-CAM and LIME for model interpretability.

Main Results:

• Achieved state-of-the-art accuracies: 98.83% (MLDID), 98.09% (Mango Leaf Disease Dataset), 98.76% (Harumanis).
• The model utilizes only 6.9 million parameters, demonstrating high computational efficiency.
• XAI methods validated the model's focus on clinically relevant pathological features.

Conclusions:

• MangoLeafNet-XAI offers an optimal balance between performance and computational efficiency for mango disease diagnosis.
• The model's lightweight design and high accuracy make it suitable for edge deployment in smart agriculture.
• MangoLeafNet-XAI sets a new benchmark for reliable, interpretable, and generalizable AI-driven disease detection systems.