AgriCure: A web application based layered augmentation-enhanced YOLOv8 for disease and nutrient deficiency detection in bitter gourd leaves
Bitter gourd is an important cucurbitaceous vegetable widely grown in India and other tropical and subtropical regions and appreciated for its nutritional, medicinal, and economic values. Traditional way of detecting diseases and nutrient deficiencies in bitter gourd leaves requires significant effo...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-09-01
|
| Series: | Current Plant Biology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2214662825000854 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Bitter gourd is an important cucurbitaceous vegetable widely grown in India and other tropical and subtropical regions and appreciated for its nutritional, medicinal, and economic values. Traditional way of detecting diseases and nutrient deficiencies in bitter gourd leaves requires significant effort and expertise whereas, precision farming and automated disease detection methods can greatly support farmers by facilitating sustainable agriculture To address this challenge a novel web based application AgriCure was developed which incorporated a multilevel approach to detect the plant disease and nutrient deficiency with high level. It uses a hybrid augmentation-based YOLOv8 DL model for image analysis. The study focuses on detecting diseases like Downy Mildew, Leaf Spot, and Jassid, as well as nutrient deficiencies such as Potassium, Magnesium, and Nitrogen Deficiency and their combinations. The initial dataset of 785 images was increased to 2430 images using advanced data augmentation. The results on the augmented dataset after 100 epochs demonstrated high effectiveness with the augmented dataset. The model achieved an impressive mean Average Precision (mAP50) of 92.9 % at an Intersection over Union (IoU) threshold of 0.50 and a mAP50–95 of 91.5 % across IoU thresholds from 0.50 to 0.95. Nearly all predicted positive instances were true positives, with a precision rate of 89.6 % and a recall of 86.6 %, which showed the capacity of the model in identifying true positives. The F1 score of 91.66 % highlighted balanced performance of the model between precision and recall, emphasising its reliability and accuracy. The model shows low losses, with a Box loss of 0.2435, a Class loss of 0.1689, and a Distribution Focal Loss (dfl loss) of 0.9024. This approach offered a valuable tool for early and accurate detection of disease and nutrient deficiency. Detection results indicate that, compared to previous methods, the proposed approach significantly improves overall performance and addresses challenges tied to limited dataset sizes. |
|---|---|
| ISSN: | 2214-6628 |