Training YOLOv5s under Field-survey Conditions to Detect The Infections of Maize Plants in Real-time | ||||
| Egyptian Journal of Agronomy | ||||
| Volume 46, Issue 1, April 2024, Page 51-60 PDF (1.3 MB) | ||||
| Document Type: Original Article | ||||
| DOI: 10.21608/agro.2024.281521.1422 | ||||
 View on SCiNiTO | ||||
| Authors | ||||
ELSAYED A. E. Ali   1; Ahmed G. Aboelyousr 2; Hassan H. Tarabye 3
| ||||
| 1agriculture engineering research Institute, Dokki, Giza , Egypt | ||||
| 2Department of Agricultural Engineering, Faculty of Agriculture, Aswan University, Egypt | ||||
| 3Department of Agricultural Engineering, Faculty of Agriculture, Aswan University, Egypt | ||||
| Abstract | ||||
| Real-time detection of plant infections by YOLOv5s is important in smart agriculture. Detecting infections in maize poses significant challenges due to the field's complexity. Therefore, YOLOv5s requires multiple training in this aspect. This study aims to explore these obstacles through procedures for training YOLOv5s based on images from field surveying in the real field and evaluate them. It investigated the wide range of infections in maize plants that occurred at the same time, including insect infestations, diseases, and physiological symptoms. A dataset of 938 images was collected from 197 cases (14 infections). YOLOv5s curves were generated using loss and accuracy functions, which rely on metrics such as precision (P), recall (R), mAP@0.5, and mAP@0.5:0.95 to capture detailed model accuracy information. The curves indicate gradual improvement in the model, albeit with some fluctuations attributed to data noise. This fluctuation may be attributed to increased classifications within the dataset. The model shows good R for most object classes, with values over 0.8, indicating accurate identification even for small or difficult-to-see objects. However, it suffers from lower R rates, like corn stunts and phosphorus deficiency, due to its difficulty distinguishing images. The model has strong mAP@0.5:0.95 scores, suggesting its ability to generalize successfully across confidence levels. It works well for most object classes, but its performance for corn stunts and phosphorus deficiency is lower due to visual similarities. To enhance performance, there is a need for further refinement of the detection system, possibly through additional training data or improved algorithms. | ||||
| Keywords | ||||
| YOLOv5; Plant infections; Real-time; machine learning; maize | ||||
|
Statistics Article View: 197 PDF Download: 97 |
||||
