Abstract
Purpose Weed management in vegetable production systems is increasingly constrained by labor shortages, rising input costs, and the need to reduce herbicide use while avoiding crop injury. Precision, site-specific spraying offers a promising alternative to broadcast application; however, its effectiveness under real field conditions is often limited by unreliable weed detection, sprayer resolution, and timing inaccuracies. This study evaluates an integrated high-precision smart spraying system combining real-time weed detection, plant tracking, and micro-jet spray actuation for selective weed control in vegetable fields. Methods The system employed a YOLOv10-small model trained on a five-season crop-weed dataset (14,186 images and 103,266 annotated plant instances), coupled with a ByteTrack algorithm for spray timing. A micro-jet sprayer equipped with 12 independently controlled nozzles spaced at 1-cm intervals was mounted on a ground-based, robotic platform to target early-stage weeds, and an optimized multithreaded software architecture was implemented for system integration and real-time performance. Following an initial dataset-based crop-weed detection evaluation, the system was tested in a lettuce field plot, a 15-m crop row containing 74 lettuce plants and 174 weeds, to further evaluate plant detection and spraying performance. Blue dye-based fluid was used in the spraying testing of the system at a forward speed of 0.91 km/h. Results Video-based evaluation yielded a detection performance of 80.0% mAP@50. Field spraying tests achieved a weed hit rate of 84.5% and a crop hit rate of 17.6%, representing a substantial improvement over previous system configurations. Conclusion This study demonstrates the benefits of integrating artificial intelligence (AI)-driven detection and tracking with a high-precision sprayer, advancing the practical deployment of intelligent sprayer systems for precision weed management. Effective site-specific weed control still requires further reducing crop contact while improving weeding accuracy, which could be possible by developing more robust weed detection models and incorporating a buffer zone around crops.