Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Abstract. Detecting weeds with precision in variable field conditions is an ongoing challenge for precise weed management. Recent advances in computer vision provide promising solutions that, if adopted, can offer valuable applications in precision weeding. This research presented a robust and improved weed detection method by leveraging advanced object detection models using two open-source datasets. Dataset A is a multi-modal weed image dataset with 3,975 RGB images depicting four weed species commonly found in North Dakota, USA. Dataset B includes 1,118 RGB images featuring six food crops and eight weed species captured in field conditions in Latvia. The research applied four object detection algorithms: the You Only Look Once-v9 (YOLOv9), Real-Time Detection Transformer (RT-DETR), YOLO World model and YOLO-v8. Results indicate that the YOLO-v9 outperformed other deep learning models, achieving mean Average Precision (mAP) values of 0.94 and 0.85 for Dataset A and Dataset B, respectively. It also surpassed YOLOv8 with a 2.15% improvement in performance. RT-DETR had the closest performance to YOLO-v9; however, it lagged significantly in speed, with YOLO-v9's inference speed being 6 ms and RT-DETR's speed being 18 ms. While advanced object detection algorithms excel in speed and precision, further research is crucial to assess their effectiveness on diverse agricultural weed datasets.