Sugar beet detection using drone imagery

This study presents a scalable framework for sugar beet detection using full-resolution RGB drone imagery from the UAV-Sugarbeets public dataset, spanning four acquisition sessions that collectively cover six phenological growth stages. Unlike prior segmentation-based annotation efforts on this dataset, we introduce the first bounding-box object detection annotations covering all six growth stages, enabling instance-level plant detection. We train YOLOv8s at a custom resolution of 2048x1536 pixels, preserving spatial detail and achieving 96.5% mAP@0.5 and 70.0% mAP@0.5:0.95. An ablation study across YOLOv8 variants (nano, small, medium) confirms that YOLOv8s provides the optimal balance between accuracy and inference speed. The manually annotated subset of 725 full-resolution images was validated through inter-annotator agreement (Cohen's Kappa = 0.95), ensuring high annotation reliability. By prioritizing RGB-based detection, this work supports cost-aware crop monitoring using consumer-grade drone technology, while acknowledging that the current framework addresses single-class sugar beet detection and single-site evaluation only. The practical scope is plant counting, growth-stage monitoring, and stand establishment assessment, with multi-class extension including weed discrimination, cross-field validation, and multispectral fusion reserved for future work.