Real-Time Weed Detection Using Advanced Object Detection Models: A Comparative Study

Detecting weeds automatically and accurately is essential for enhancing agricultural productivity while addressing the challenges of rising labor costs and the environmental impact of conventional herbicide use. Deep learning algorithms have the potential to achieve the level of weed detection performance required for automated weed control systems. This study presents a comparative evaluation of cutting-edge object detection frame-works, including YOLOv8, RT-DETR, and YOLOv10, for real-time weed detection in unstructured field images. We evaluated all the models on a publicly available dataset from Kaggle, correcting approximately 10% of mislabeled annotations to ensure data integrity. Additionally, we applied various data augmentation techniques to enhance data diversity and thereby improve the generalizability and robustness of the models. Furthermore, this study emphasizes the importance of appropriate optimizer selection in improving the performance of deep learning algorithms. Both Adam and Stochastic Gradient Descent (SGD) optimizers were employed to train the YOLO-based frameworks. The experimental results revealed that YOLOv10m, when trained with SGD, outperformed both YOLOv8 and RT-DETR, achieving a precision of 94.0%, a recall of 95.5%, an mAP@50 of 98%, and an inference time of 72 ms. These results demonstrate that YOLOv10m effectively balances accuracy and computational speed for real-time weed detection in complex agricultural environments. By providing a comprehensive analysis, this study potentially delivers a foundation for the development of auto-mated weed control systems, ultimately contributing to enhanced agricultural productivity and sustainability.