Architecture and hyperparameter optimization of YOLOv5 and YOLOv8 for weed detection in garlic fields

Weed adaptability to environmental conditions poses a major challenge in agricultural production, often leading to yield reduction or even complete crop loss. Effective weed detection requires algorithms tailored to the unique visual and biological characteristics of each crop. However, despite the agricultural importance of garlic, no studies have developed optimized state-of-the-art deep learning models for weed identification in garlic fields. Addressing this gap, the present study focuses on optimizing two prominent object detection frameworks-YOLOv5 and YOLOv8-for accurate weed detection under real field conditions. A dataset of 600 RGB images containing garlic plants and five weed species was collected directly from the field. Different versions of YOLOv5 and YOLOv8 were optimized and evaluated based on mAP@0.5 and inference speed. Experimental results showed that YOLOv8n achieved superior performance, yielding an mAP@0.5 of 87.0% with an average inference time of 16 ms, compared to YOLOv5m with an mAP@0.5 of 85.0% and 50 ms. The findings highlight that YOLOv8n, with an input resolution of 320 × 320 pixels and optimized using the SGD optimizer, provides the best trade-off between detection accuracy and computational efficiency. These results demonstrate the potential of YOLOv8n for real-time weed detection and its application in precision agriculture machinery.