An attention enhanced CNN ensemble for interpretable and accurate cotton leaf disease classification

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Manual cotton leaf disease identification is inefficient and error-prone.
• Existing automated methods lack dataset diversity, consistent evaluation, explainability, and efficiency.
• There is a need for accurate and robust automated solutions for cotton disease diagnosis.

Purpose of the Study:

• To develop an accurate, robust, and interpretable automated system for cotton leaf disease classification.
• To address limitations of existing automated disease identification methods.
• To propose an attention-enhanced Convolutional Neural Network (CNN) ensemble named CottonLeafNet.

Main Methods:

• Developed an attention-enhanced CNN ensemble (CottonLeafNet) integrating lightweight CNNs.
• Utilized two publicly available cotton leaf disease datasets for training and evaluation.
• Employed Gradient-Weighted Class Activation Mapping (Grad-CAM) for model interpretability.
• Validated real-time feasibility through web-based deployment.

Main Results:

• CottonLeafNet achieved state-of-the-art performance, with accuracies up to 99.43% on separate datasets and 99.08% on a merged dataset.
• Demonstrated high macro F1-scores (e.g., 0.9942), Cohen's kappa (e.g., 0.9924), and low inference times (0.40-0.51s per image).
• Showcased model robustness under class imbalance and moderate generalization across datasets.
• Grad-CAM visualizations confirmed attention to disease-relevant regions, enhancing interpretability.

Conclusions:

• CottonLeafNet provides an accurate, robust, interpretable, and computationally efficient solution for automated cotton leaf disease diagnosis.
• The model's performance and real-time feasibility suggest significant potential for practical agricultural applications.
• This work advances automated crop disease management through explainable AI.