Weed detection in cotton farming by YOLOv5 and YOLOv8 object detectors

Abstract

Weeds are undesirable plants that pose significant challenges to agricultural crops by competing with them below and above ground. Traditional manual methods of identifying and managing weed infestations are time-consuming and labor-intensive, limiting their effectiveness. To address this problem, an experiment was conducted during kharif-2022 at Main Agricultural Research Station, University of Agricultural Sciences, Dharwad, India. The objective was to create a cotton-weed dataset of Indian cotton production system and to evaluate the performance of YOLO (You Only Look Once) object detection models. High-resolution images were captured using a digital camera mounted on a tripod stand, positioned vertically downward at a height of 80 cm. A dataset of 2300 images was created accompanied by 44130 bounding box annotations of two weed classes (grasses and broad-leaf weed) and a crop class i.e. Cotton. 12 state-of-the-art YOLO object detectors of two versions (YOLOv5 and YOLOv8) were evaluated. The algorithms demonstrated promising results, with detection accuracy ([email protected]) ranging from 69.88 % (YOLOv8n) to 76.50 % (YOLOv5s6). YOLOv5n (3.07 ms inference time) was the fastest model. Additionally, it had lower number of model parameters (1.7 million) and GFLOPs (4.1) making it suitable for real-field applications in resource-constraint conditions. Other YOLO models also exhibited significant potential for real-time weed detection. The study underscores the capabilities of YOLO object detectors for real-time weed detection in cotton. The models can be implemented on specialized computing hardware, for integrating into a robotics and sensor platform for real-time weed identification enabling targeted herbicide application, reducing chemical use, and enhancing crop yields.