Evaluation of YOLO-based weed detection models on commercial horseradish fields in Southern Illinois

Horseradish ( Armoracia rusticana ) is a high-value specialty crop whose production in Southern Illinois is constrained by limited herbicide options and labor-intensive weed management, particularly in commercial fields. This study developed a multisource image dataset and evaluated lightweight deep learning models to enable real-time, vision-based weed detection for future robotic weeding systems in commercial horseradish fields of Southern Illinois. Image data collection was conducted during the 2024 growing season at two commercial fields and one research site using a handheld smartphone and an unmanned ground vehicle (Farm-ng Amiga robotic platform) equipped with two stereo cameras (Luxonis OAK-D cameras). The data was first annotated, assigning appropriate labels to horseradish and weed instances, then augmented to improve data diversity. We trained and compared nine YOLO models (v8, v11, v12; nano, small, medium) using standard object detection metrics (precision, recall, F1-score, mAP@50) and computational indicators (inference time, GFLOPs, model size, training time), and selected the best configuration for hyperparameter tuning with an automated search over learning rate, regularization, and optimizer. The tuned YOLOv8-nano model achieved the best balance of detection performance and computational efficiency, and was subsequently benchmarked on multiple desktop and edge-computing platforms to assess real-time feasibility. The results demonstrated that lightweight YOLO architectures can provide accurate, fast horseradish-weed detection suitable for deployment on embedded hardware, offering a key sensing component for future autonomous mechanical weeding in commercial horseradish production. This study makes three key contributions: (i) a multisource image dataset of horseradish and weeds collected from both commercial and research fields using manual imaging and a robotic platform; (ii) an evaluation protocol that combines accuracy metrics with computational indicators to guide model selection for embedded deployment; and (iii) a cross-platform benchmarking workflow that assesses the real-time feasibility of lightweight YOLO models on desktop and edge-computing hardware for robotic weeding applications.