Low-cost automated generation of application maps for control of Rumex Obtusifolius in grasslands

The majority of newly developed sprayers now feature advanced capabilities, allowing herbicide application with centimeter-level precision, potentially reducing herbicide use by up to 90%. However, accurately identifying the precise locations to spray, known as the application map, remains a significant research challenge. Recently, both commercial providers and research institutions have proposed various drone-based methods for generating application maps. Despite these advancements, practical adoption is limited, primarily due to regulatory constraints and the high costs associated with the technology. A promising approach to increasing the adoption of these technologies lies in the utilization of more cost-effective hardware solutions. In this paper, we introduce and evaluate a novel detection method specifically designed for identifying Rumex obtusifolius (sorrel) and for automatically generating application maps that are compatible with most GNSS-enabled sprayers. To this end, we present a new metric for treatment success, termed the treatment F1-score, and conduct a comparative analysis of the performance of the DJI Mini 2 and the DJI Matrice 350 RTK using our proposed system, achieving treatment F1-scores of 0.61% and 0.65% , respectively. The ability of this system to deliver good performance utilizing significantly less expensive hardware than typically employed in similar applications suggests a potential for broader adoption, particularly given the unexpectedly modest performance gap of only 4 percentage points in the treatment F1-score. Under controlled experimental conditions, we observed reductions in herbicide use of up to 97% without missing any targets. In practical applications within real-world meadows, a 40% reduction in herbicide consumption was achieved with a treatment accuracy of 85% . These findings underscore the substantial potential for future technological advancements. The standalone object detector achieves a mean Average Precision (mAP) of 67.4% and an F1-score of 62%, demonstrating robust performance even on out-of-distribution drone data collected by other researchers. Still, the performance of the object detection algorithm is identified as a critical bottleneck in the system. To facilitate further research and development in this domain, we have made our training dataset available for download.