Seeding is the first and one of the most critical operations in crop production, as it directly affects all subsequent field processes. This study proposes a real-time seeding monitoring and mapping system for a sowing-type garlic planter that detects...
Seeding is the first and one of the most critical operations in crop production, as it directly affects all subsequent field processes. This study proposes a real-time seeding monitoring and mapping system for a sowing-type garlic planter that detects seeding events—single seeding, missing seeding, and double seeding—during operation and visualizes their spatial distribution by integrating GNSS-based position information. A hardware platform consisting of an RGB camera, a GNSS receiver, and an embedded PC was developed, and a lightweight YOLO11n-based object detection model was implemented to perform real-time, row-wise detection of garlic seeding events. For seeding position estimation, an algorithm was proposed that integrates the longitudinal front offset between the camera detection point and the actual seed discharge location, row-wise lever arms derived from the planter geometry, and GNSS-derived position and heading information to estimate seeding locations in a local ENU coordinate frame. Detected seeding events and estimated positions were discretized into grid cells of 1 m × 0.14 m, enabling intuitive visualization of missing and double seeding patterns along the travel direction and across planter rows. The developed GUI displays real-time cumulative counts of single, missing, and double seeding events for each row during operation, and automatically generates CSV logs and seeding maps after completion, supporting both real-time monitoring and post-operation analysis. System performance was evaluated using field video data acquired during actual garlic seeding operations, including quantitative assessment of seeding event detection accuracy and validation of position estimation accuracy based on ground-truth measurements obtained after removing soil cover over an approximately 16 m test section of a single row. Among multiple YOLO families (YOLOv8/10/11/12) and model sizes evaluated, YOLO11n was selected as the final model, achieving real-time performance of 25.6 FPS on a Jetson TX2 while maintaining high precision and recall. In a camera-based validation test, the estimated seeding positions exhibited a mean deviation of −0.25 cm from the theoretical 100 cm spacing. The estimated in-row seeding spacing showed an MAE of 4.11 cm, an RMSE of 5.12 cm, and a correlation coefficient of r = 0.9819 compared with tape-measured ground truth. Furthermore, field evaluation over 384 bucket slots yielded F1-scores of 0.973 for single seeding and 0.954 for missing seeding, with an overall accuracy of 96.6%. In conclusion, the proposed real-time seeding monitoring and mapping system provides a practical framework for spatially quantifying garlic seeding quality by integrating vision-based event detection with GNSS-based position estimation, and demonstrates strong potential for future applications in planter setup optimization, operation history management, and precision agriculture decision support.