Leveraging pre-trained models within a semi-supervised and explainable AI RealTime framework: A pioneering paradigm for betel leaf disease detection
Betel leaf is a significant crop considering its popularity and economic benefits. Although its betel vines are prone to several infections that are generally specified as betel leaf disease. Fungal and bacterial infections are simultaneously responsible for the disease, triggering the early decay o...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | Journal of Agriculture and Food Research |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2666154325005137 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Betel leaf is a significant crop considering its popularity and economic benefits. Although its betel vines are prone to several infections that are generally specified as betel leaf disease. Fungal and bacterial infections are simultaneously responsible for the disease, triggering the early decay of leaves and resulting in diminished production. Notably, artificial intelligence can deliver a significant imprint in the agricultural sector of betel leaf, as this assists in the production increment by predicting disease. The deep learning approach for disease classification can reduce the extensive nature of human-led examination for the primary symptoms, indicating its importance. After leveraging multiple pre-trained models, the study has selected DenseNet-201 as the best pre-trained model, which scored the highest accuracy of 99.23%. Later, a semi-supervised learning approach was integrated with the pre-trained model DenseNet-201, where FixMatch top-performed with 98% accuracy, manifesting a lightweight deep learning framework with superior performance. Next in line is MixMatch 91.27% and lastly 81% by MeanTeacher. The semi-supervised methods used only 30% of the labeled data for classification, applying 1000 original and 2589 preprocessed images from the Mendeley website. Furthermore, to enable effective feature extraction and facilitate real-time decision-making in practical applications, the experiment has integrated a pre-trained model, DenseNet-201, with XAI methods such as Smooth Grad, Vanilla Saliency, GradCAM++, and Faster ScoreCAM that help to visualize the affected lesions. Finally, this process has been fused with a real-time application providing transparency for the end users to identify the unhealthy betel leaves more easily. The incorporation of semi-supervised and XAI with the best pre-trained model insinuates that the classification process of the healthy leaves from diseased ones can have meticulous consequences without over-fitting issues with less amount of labeled data. |
|---|---|
| ISSN: | 2666-1543 |