Performance Evaluation of YOLO Models for Automatic Crop and Weed Detection

Accurate weed mapping is essential for the implementation of weed management in order to maintain high crop yields and less chemical pollution. This paper focuses on the performance evaluation of YOLO models for automatic crop and weed detection. The research aims to develop weed detection models using one-stage deep learning-based object detectors by transferring pre-trained object detection models to a selected weed dataset. The models are trained on annotated images from a public dataset and evaluated based on detection accuracy. Results demonstrate promising training performance, with approximately 70% mean average precision achieved in 100 training epochs. YOLOv5x exhibits the best performance among the studied models, achieving the highest precision, recall, and mean average precision (0.72, 0.75, and 0.74, respectively). The findings indicate that models with deeper or larger backbone networks tend to provide higher detection accuracy.