Abstract
Weeds remain a major biological threat to agriculture, reducing crop yields and causing substantial economic losses. Conventional blanket herbicide applications are costly, environmentally damaging, and accelerate the evolution of herbicide resistance. Machine vision–based precision weeding offers a sustainable alternative by enabling site-specific weed localization and removal. This study advances plant perception through YOLOv10 detection models and leverages a 1-cm nozzle array to achieve precise weed targeting that maximizes treatment efficacy while minimizing crop contact. A dedicated weed–crop dataset curated from vegetable fields further enhances model performance. Field testing was conducted in a lettuce field, and it achieved an average 80.0% mAP@50 in plant detection, and an 84.5% weed hit rate, with a moderate crop hit rate of 17.6% in real fields. The system also reached a high weed precision of 90.3% and low crop/soil precision of 7.9% and 15.7%, respectively. These results demonstrate the effectiveness of integrating an AI-driven perception system with a high-precision sprayer.
Keywords
Deep learning; Machine vision; Plant detection; Smart sprayer
DOI
10.5703/1288284318201
Field Test and Evaluation of A Smart Sprayer for Precision Weeding
Weeds remain a major biological threat to agriculture, reducing crop yields and causing substantial economic losses. Conventional blanket herbicide applications are costly, environmentally damaging, and accelerate the evolution of herbicide resistance. Machine vision–based precision weeding offers a sustainable alternative by enabling site-specific weed localization and removal. This study advances plant perception through YOLOv10 detection models and leverages a 1-cm nozzle array to achieve precise weed targeting that maximizes treatment efficacy while minimizing crop contact. A dedicated weed–crop dataset curated from vegetable fields further enhances model performance. Field testing was conducted in a lettuce field, and it achieved an average 80.0% mAP@50 in plant detection, and an 84.5% weed hit rate, with a moderate crop hit rate of 17.6% in real fields. The system also reached a high weed precision of 90.3% and low crop/soil precision of 7.9% and 15.7%, respectively. These results demonstrate the effectiveness of integrating an AI-driven perception system with a high-precision sprayer.