Can high resolution UAV-derived weed maps replace camera-based targeted spray application?

Efficient weed mapping is critical for enabling site-specific weed management (SSWM) and ensuring effective targeted herbicide application in no-till dryland cropping systems. Producers are increasingly interested in using UAV (Unmanned Aerial Vehicle) based weed mapping to generate prescription maps as a lower-cost alternative to camera-based systems, yet the accuracy and operational feasibility of these UAV-derived methods remain uncertain. This study compared three UAV-based weed detection approaches, RGB Excess Green (ExG) index, NIR band thresholding, and a YOLOv11 deep learning model against a ground-based camera reference system across three mapping dates in a summer fallow field. UAV imagery was collected using RGB and multispectral sensors, and YOLO was trained on annotated RGB image tiles. Key performance metrics evaluated were prediction accuracy (MAE, RMSE, and R 2 ), omission rates (%) and Bland-Altman analysis (bias and Limits of Agreement-LoA). Results showed that all UAV methods systematically underpredicted weeds compared to camera-based, with bias and variability increasing with weed density and from early season (April 24) to late season (June 24). RGB consistently provided the lowest errors and most stable agreement, YOLO tracked spatial patterns but had the largest underestimation, and NIR performance varied according to soil and residue conditions. Weed omission rates ranged from 41–65% for RGB, 48–73.5% for NIR, and 78–85% for YOLO. These findings demonstrate that the effectiveness of UAV-based maps is constrained by image spatial resolution, which can limit the detection of small weeds that are more vulnerable to management practices. In contrast, real-time camera-based systems improve the detection of small weeds at application, increasing herbicide application efficiency in large-scale operations. Future research should focus on refining UAV image acquisition protocols and integrating advanced sensing technologies to improve weed detection reliability across the growing season and support more effective site-specific weed management.