Real-time jute leaf disease classification using an explainable lightweight CNN via a supervised and semi-supervised...

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Accurate jute leaf disease detection is crucial for crop health and farmer income.
• Traditional deep learning methods require extensive labeled datasets, which are expensive and time-consuming to create.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) with a semi-supervised learning (SSL) framework for efficient jute leaf disease classification.
• To reduce the dependency on large labeled datasets while maintaining high classification accuracy.

Main Methods:

• A novel lightweight CNN architecture was designed, incorporating modified depthwise separable convolutions, an enhanced squeeze-and-excite block, and a modified mobile inverted bottleneck convolution block.
• A semi-supervised learning (SSL) self-training framework was integrated to leverage unlabeled data.
• Explainable AI (Grad-CAM) and a Flask-based web application were used for interpretability and practical deployment.

Main Results:

• The proposed model achieved 98.95% accuracy in a supervised setting with an 80:10:10 data split.
• In a semi-supervised setting with only 10% labeled data, the model attained 97.89% accuracy, demonstrating near-supervised performance.
• The model has only 2.24M parameters (8.54 MB), making it suitable for resource-constrained environments.

Conclusions:

• Combining SSL with a lightweight CNN offers a novel approach for accurate jute leaf disease detection with significantly reduced labeled data requirements.
• The model provides interpretable disease region visualizations and practical real-time usability.
• This approach enhances agricultural sustainability by improving disease management efficiency and reducing costs associated with data labeling.