Presentation: 2025 ND EPSCoR Annual conference 

October 21, 2025, NDSU Memorial Union, Fargo, North Dakota

Reducing the Data Bottleneck: Real-Time Annotation for Precision Agriculture

Timely and accurate dataset annotation remains a critical bottleneck in deploying object detection models for precision agriculture, where rapid decision-making depends on large volumes of well-labeled field data. Conventional annotation workflows rely on offline, labor-intensive labeling of crop and weed images, slowing down experimentation and limiting responsiveness to dynamic field conditions such as weed emergence, pest outbreaks, or crop growth stages. To address this challenge, we introduce a real-time annotation framework that integrates YOLO-based detectors directly on agricultural edge devices, enabling immediate labeling during UAV or UGV scouting missions. The system is released as an open-source software package with both command-line and REST interfaces, supporting single- and multi-class mappings relevant to agricultural targets (e.g., weeds, pests, fruits, or diseases). It allows either automated or human-in-the-loop acceptance of detections and generates YOLO-format annotations, metadata logs, and training-ready datasets in real time, significantly reducing dataset preparation overhead in the field. To evaluate performance, we conducted a comparative study across three YOLO architectures (v5, v8, v12) under pretrained and scratch-based initializations, and in single-class versus multi-class annotation settings. Statistical analysis of learning dynamics and accuracy metrics shows that pretrained and single-class configurations achieve faster convergence, higher robustness, and improved annotation throughput. The results validate real-time annotation as a practical strategy for accelerating dataset development in precision agriculture, enabling more adaptive, scalable, and high-quality model deployment for tasks such as weed detection, disease monitoring.