Annotation-efficient weed detection using DINOv3-distilled YOLOv12.

Weed pressure causes global crop yield losses of 10-34%, while the deployment of deep learning-based weed detection systems at scale remains constrained by the high cost of bounding-box annotation across diverse field environments. This study addresses this annotation bottleneck in precision agriculture by proposing WEEDINO-YOLOv12, a label-efficient weed detection framework that transfers global-average-pooled feature distributions from a frozen DINOv3 ViT-B/16 teacher into a lightweight YOLOv12n backbone through feature-distribution distillation on unlabeled agricultural imagery, followed by supervised fine-tuning on a limited labeled subset. To rigorously evaluate the proposed framework, we present a controlled empirical benchmark comparing four training regimes: fully supervised YOLOv12n, semi-supervised Soft Teacher, self-supervised BYOL, and the proposed DINOv3 distillation approach. All methods are assessed using a common YOLOv12n backbone, consistent evaluation metrics, matched controls, and multi-seed reporting. External validation on the multi-class CottonWeedDet12 dataset further examines whether the observed label-efficient behaviour generalises beyond the single-class Roboflow Weeds benchmark. Across matched 20%-label settings, WEEDINO-YOLOv12 improved mAP@0.5:0.95 from 0.6402 ± 0.0271 to 0.6517 ± 0.0087 on the Roboflow fixed split and from 0.7987 ± 0.0154 to 0.8083 ± 0.0078 on CottonWeedDet12. Full-label supervision remained the strongest overall setting, indicating that the proposed method provides modest but consistent annotation-efficiency gains rather than replacing fully supervised training. High-resolution fine-tuning at 896 × 896 pixels is analysed separately because it can improve localisation independently of the distillation stage. A Streamlit-based deployment prototype further demonstrates the practical accessibility of the framework for agronomists and precision-agriculture users without requiring direct interaction with deep learning code.