Deep learning driven edge inference for pest detection in potato crops using the AgriScout robot

• Field robot built a dataset of 832 geotagged images for beetle detection. • Six real-time object detectors were benchmarked with five random seeds. • Seed-to-seed variability was low, supporting robust small-dataset comparisons. • Top model balanced accuracy and stability while enabling real-time edge inference. • Edge system reached 46.5 frames per second for continuous row-to-row scouting. Early field-scale surveillance of Colorado potato beetle remains a persistent bottleneck for sustainable potato production because conventional scouting is labor-intensive and provides limited spatial resolution for timely intervention. Here, we present AgriScout, a battery-powered autonomous scouting robot equipped with RGB imaging, controlled lighting, and RTK-GPS geotagging for continuous row-to-row data collection. Using AgriScout, we curated a field dataset of 832 georeferenced images and manually annotated adult beetles with tight bounding boxes to support tiny-object detection under real canopy conditions. We benchmarked six YOLO object detectors (YOLOv5s, YOLOv8s, YOLOv9s, YOLOv10s, YOLOv11s, and YOLOv12s) using transfer learning, high-resolution inputs (1280 × 1280), and an augmentation strategy tailored to small targets (including mosaic, scaling, and translation). To address training variability on the modest dataset, models were evaluated across multiple random seeds (7, 42, 123, 999, and 2024) and compared using precision, recall, mAP, F1, confidence behavior, and statistical tests of between-model differences. Across runs, YOLOv11s provided the most reliable overall balance for deployment, exhibiting strong precision and robust localization performance. For edge deployment, inference throughput was measured on an NVIDIA Jetson Orin Nano across multiple export formats; TensorRT consistently delivered the highest FPS, reaching 46.5 FPS (YOLOv5s) and exceeding 40 FPS for several variants, confirming real-time feasibility under FP32 inference. Finally, YOLOv11s detections were fused with RTK-GPS coordinates to generate centimeter-level infestation maps that visualize spatial clustering of beetle activity and support hotspot-driven, targeted management. Collectively, this work demonstrates an end-to-end, robot-to-map pipeline for beetle monitoring and provides a reproducible benchmark of accuracy, stability, and edge deployability for YOLO-based pest detection in commercial potato systems.