Weed Detection Based on Fine-Tuned YOLOv5

Conventional weed control often ignores weed locations, potentially causing land damage. AI-based weed detection offers a modern solution. Unfortunately, little attention has been paid to how an augmentation technique improves performance of AI-based weed detection. To address this research gap, we proposed a weed detection model based on fine-tuned YOLOv5 to detect weeds in corn fields aged 1 to 3 weeks. YOLOv5 offers computational benefits, includes auto-augmentation capabilities, and presents a more organized visualization of the outcomes from the training data. The primary dataset used in this study comprised 3,437 images with a total of 18,267 annotated objects, including corn plants and weeds. We observed the effect of mosaic augmentation on the dataset images based on auto-augmentation in YOLOv5. We found that mosaic augmentation with a 0.75 probability outperformed 0.5 and 1.0. The best mAP performance was at the level @0.5 with score of 0.776 for all classes. Meanwhile, the [email protected]:0.95 was 0.573 for all classes, indicating that the model’s performance requires further improvement. Our study is promising as a guidance for future development of AI-based weed detection.