Development and Preliminary Evaluation of a Machine Vision‐Guided Smart Sprayer Prototype Toward Precision Vegetable Weeding

Sustainable weed management necessitates herbicide‐reduced, precision weed control. Recent advancements in machine vision and artificial intelligence (AI) for weed recognition and localization have created new opportunities for high‐precision weeding. Although blanket spraying systems are widely used for weed removal, machine vision‐guided, AI‐powered smart sprayers are becoming increasingly important but remains underexplored for many vegetable cropping systems. This study presents a new engineering effort to develop a ground‐based, fully integrated smart sprayer prototype for precision weeding in vegetable crops. The system incorporates a machine vision system for real‐time weed detection and a custom‐designed weeding actuator equipped with an array of 12 evenly spaced spraying nozzles controlled independently, spanning a row width of approximately 0.66 m. A dedicated graphical user interface was designed and implemented to support system integration and operation. Using a diverse, three‐season weed data set consisting of 1464 still images and 16,215 weed instances for the detection of Lambsquarters and Purslane, a benchmark evaluation of a suite of 12 variants of YOLOv8, YOLOv9, and YOLOv10 models was conducted to enabled informed model selection. Among these models, YOLOv10s provided the most favorable balance between inference speed and accuracy, achieving an overall mAP@50 of 86.6%, and was therefore selected for onboard deployment and field testing. With the model deployed an the edge computer, the sprayer prototype was evaluated at a platform speed of approximately 0.11 m/s in both indoor and field environments for real‐time target detection and detection‐based spraying. In indoor trials using simulated square targets, the system achieved a video detection accuracy of 98.9% mAP@50 and a spraying hit rate of 88.4%. Under natural field conditions, however, performance declined, yielding a video‐based mAP@50 of 48.9% and a spraying hit rate of 51.3%. These results highlight the challenges of AI model generalization in real‐time weed detection in uncontrolled field environments. Imperfections in nozzle configuration and a relatively large nozzle spacing (approximnately 51 mm) contributed primarily to off‐target spraying, and the timing errors for nozzle actuation also affected the spraying efficacy. Further improvements in both hardware and software, as well as extensive field validation, are needed before the prototype can evolve into a reliable, high‐resolution tool for vegetable weeding.