Lightweight dual-stage feature refinement for black gram leaf disease classification using ConViTSE.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Black gram (urad bean) is vital to Indian agriculture but suffers significant losses from leaf diseases.
• Manual disease identification is time-consuming and inaccurate, hindering effective crop management.
• Automated disease detection systems are crucial for improving agricultural productivity and farmer livelihoods.

Purpose of the Study:

• To develop a lightweight, hybrid deep learning model for accurate black gram leaf disease classification.
• To enhance feature extraction and representation for improved disease detection.
• To evaluate the model's performance and generalization capabilities across different crops.

Main Methods:

• Proposed ConViTSE, a hybrid architecture integrating ConvMixer, Vision Transformer (ViT), and Squeeze and Excitation (SE) blocks.
• Incorporated Local Channel Attention Refinement (LCAR) and Global Channel Attention Refinement (GCAR) modules.
• Trained and evaluated the model on a black gram leaf disease dataset and cross-domain datasets (rice, maize, wheat).

Main Results:

• ConViTSE achieved a classification accuracy of 99.30% on the black gram dataset.
• Demonstrated strong cross-domain generalization with accuracies of 98.75% (rice), 98.20% (maize), and 95% (wheat).
• Outperformed traditional deep learning models in disease classification accuracy.

Conclusions:

• ConViTSE is a highly accurate and computationally efficient model for black gram leaf disease detection.
• The model's robust generalization potential supports its application in precision agriculture for various crops.
• ConViTSE offers a practical solution for real-time disease management in diverse agricultural settings.