A lightweight hybrid CNN and transformer model for medicinal leaf disease classification with explainable AI

Area of Science:

• Agricultural Science
• Computer Science
• Biotechnology

Background:

• Medicinal plants like Tulsi, Neem, and Patharkuchi are vital for bioactive compounds.
• Leaf diseases pose a significant threat to the yield and phytochemical quality of these plants.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for identifying medicinal plant leaf diseases.
• To establish a benchmark for precision phytopathology and sustainable cultivation practices.

Main Methods:

• Proposed LSeTNet, a lightweight hybrid Convolutional Neural Network (CNN) Transformer architecture with Squeeze-and-Excitation (SE) blocks.
• Utilized Explainable Artificial Intelligence (XAI) techniques like Grad-CAM, LIME, and t-SNE for model interpretability.
• Conducted rigorous validation including five-fold cross-validation and external testing on the BD-MediLeaves dataset.

Main Results:

• Achieved 99.72% accuracy and a 1.00 macro F1-score on 12 disease classes with minimal parameters (9.38 M) and computations (2.50 GFLOPs).
• Demonstrated robust performance with 99.74% accuracy in cross-validation and 99.42% on an independent dataset.
• XAI methods confirmed the model's focus on pathological regions, and low inference latency (6.98 ms/image) enables real-time edge deployment.

Conclusions:

• LSeTNet significantly outperforms existing models, offering a transparent, efficient, and generalizable solution for plant disease detection.
• The model supports real-time, edge-based applications, crucial for sustainable medicinal plant cultivation and precision phytopathology.