Zero-shot segmentation meets EfficientNetB7-MHA: an explainable deep learning framework for real-time plant disease detection

The detection of plant leaf diseases is essential for ensuring crop health and productivity. This study uses a comprehensive merged dataset, including the Mendeley Plant Leaf Dataset, which includes 22 classes, while the Jute and Mulberry datasets provide 2 classes each, representing healthy and diseased categories from various species. The final dataset, consisting of 26 classes, was augmented to ensure 500 training samples per class. In this study, the Segment Anything Model (SAM) was employed for zero-shot segmentation, enabling the automatic extraction of precise regions of interest (ROIs) from leaf images without the need for task-specific training for region-focused analysis. An EfficientNetB7- Multi-Head Attention (MHA) model that combines EfficientNetB7 with MHA was used to improve plant disease classification across 26 distinct classes. The proposed model is designed to handle the variety and diversity of leaf diseases, achieving a high classification accuracy of 98.01%, with precision, recall, and F1-scores all exceeding 97.9%. The integration of MHA allows the model to focus on disease-specific features, significantly enhancing its ability to generalize across diverse agricultural settings while maintaining scalability. Experimental results show that the EfficientNetB7-MHA model consistently outperforms several state-of-the-art (SOTA) models in terms of accuracy and robustness, making it a promising tool for precision agriculture. Explainable AI (XAI) including the use of Grad-CAM was utilized to detect the precise regions of interest for each class providing interpretability and insights into the model’s decision-making process. Finally, to demonstrate the model’s performance, a web application was developed that displays the predicted outputs when given different images of healthy and diseased leaves from the dataset. This comprehensive approach demonstrates significant potential for improving agricultural disease management through accurate, scalable, interpretable, and efficient disease detection.