Real-time maize crown detection and robotic autonomous navigation for precision fertilization via lightweight-improved deep learning

Small intelligent robots provide an effective solution for mid-growth stage maize fertilization in hilly regions. However, accurate crown recognition and real-time trajectory planning remain critical challenges for autonomous fertilization. In this study, a lightweight deep learning model was developed for real-time maize crown detection and autonomous trajectory planning. An improved lightweight network based on YOLOv8n was constructed through structural optimization combined with model pruning, and deployed on edge devices. A multi-node parallel software architecture was developed for robot control. A dedicated fertilization mechanism was designed and validated via simulation analysis. Three pruning methods—DepGraph, LAMP, and Taylor—were applied and systematically compared. Experimental results showed that the Taylor-pruned C2f-Star+HSAFPN model achieved strong detection performance. The mean average precision at an IoU threshold of 0.5 (mAP@0.5) was 92.8%. The model size was 2.7 MB. Compared with the original YOLOv8n, the parameters, GFLOPs, and model size were reduced by 56.8%, 59.3%, and 56.5%, respectively. Integrating the lightweight model with k-means clustering enabled real-time crown detection and trajectory planning during fertilization. This study offers a viable technical framework for intelligent field fertilization robots.