Answers are grounded in this paper and cite their sources.
In precision agriculture, the integration of artificial intelligence with targeted weeding technologies provides critical support for efficient and accurate weed control. However, existing detection and localization methods struggle to accurately identify morphologically similar and diminutive weed seedlings in complex field environments, which limits the precision and reliability of targeted interventions. This study proposes GSP-YOLO, an enhanced lightweight model based on YOLOv8s, designed for real-time weed detection and edge deployment. To balance performance and efficiency, the original backbone was replaced with GhostNetV3, and a synergistic effects between spatial and channel attention (SCSA) module was integrated to improve feature extraction for small targets. Additionally, powerful-IoU (PIoU) loss was employed to refine bounding box regression and multi-scale adaptability. Validation on the NIAM9weeds dataset shows that GSP-YOLO achieves 97.74% precision, surpassing mainstream models like YOLOv5n/s, v6s, v7-tiny, v8n/s, v9s, and v10s. Compared to the baseline YOLOv8s, the precision, recall, and mAP improved by 1.94, 2.76, and 2.63 percentage points, respectively, while parameters, FLOPs, and model size were reduced by 32.88%, 31.93%, and 29.20%. Post-TensorRT acceleration, the inference speed reached 41 FPS on edge terminals. In dynamic greenhouse laser weeding trials, the system achieved a weed recognition accuracy of 93.40% at an operational speed of 0.72 km/h, with an average elimination time of 275 ms per weed and a crop injury rate below 3.40%. This algorithm balances high precision with the lightweight requirements of embedded deployment, providing a core technical foundation for real-time dynamic target detection in laser weeding robots.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: