An Edge-AI enabled UAV system for site-specific application targeting Palmer amaranth in corn and soybean fields

Current literature on the use of unmanned aerial vehicles (UAV) for site-specific herbicide spraying (SSHS) reveals several gaps. Many studies focus on minor customizations to detection and segmentation models, but few explore a complete framework for implementing these models on UAVs. Utilizing systems approach, this study identifies important components required for SSHS. We introduce 100k bounding box annotated Palmer amaranth dataset, which is the largest in current literature. The current study evaluates the performance of YOLO and SAM models for automating bounding box and segmentation annotations, which can reduce considerable manual resources and time required. Furthermore, the trained models are deployed on multiple (x4) edge devices for comparison of real-time performance for generation of weed maps. Our results show that the model-based auto-annotation approach enabled the annotation of 7.8 times more images, in approximately 0.03 % of the time, and with 87.5 % fewer personnel. The developed Palmer detection model achieved a high mAP@0.5 of 0.97 and inference speeds of 0.6–2.2 ms on a high-performance computer (HPC) and 3–5 ms on an edge device, when working with image tiles. Additionally, the segmentation model achieved a mAP@0.5 of 0.95 and inference speeds of 0.83 ms on HPC and 4 ms on an edge device. Running bounding box and instance segmentation inference on a large orthomosaic was achieved in 12.6 and 19.97 s respectively. To translate the findings of this study into a field deployable tool, we developed a CLI-based tool called the Palm Prescription Map, for streamlined stitching and weed map generation on an edge device in the field. These developments address efficient weed management using UAVs with the help of lightweight single shot models, automating weed mapping, and provides an open-source end-user tool for successful implementation in the field. • 100k Palmer amaranth dataset with 300k Palmer amaranth objects, largest in literature. • Palmer amaranth auto-annotate model with 99.97 % reduction in resource and time. • Comprehensive performance evaluation of YOLOv8–YOLOv12 for Palmer amaranth detection, mapping, and deployment. • Palm Prescription Map, an edge device enabled weed mapping tool.