Real-time sunflower detection using semi-supervised and self-supervised deep learning for precision agriculture

Accurate sunflower head detection is essential for precision agriculture, supporting timely monitoring and yield estimation. However, reliable detection under UAV settings remains challenging due to annotation scarcity, variable field conditions, and inconsistent localization across flowering stages. This study presents a unified framework that evaluates supervised, semi-supervised, and self-supervised learning strategies on UAV imagery collected under real field conditions. In the supervised setting, YOLOv12s achieved the strongest performance (mAP@50 ≈ 93 %), with stable convergence and focused visual attention, while RF-DETR showed lower recall and weaker localization. To reduce annotation requirements, a Pseudo-STAC teacher–student approach was evaluated across varying labeled-to-unlabeled ratios. Teacher models maintained high accuracy even with limited supervision (mAP@50 = 88.5–91.6 %), while student models approached teacher-level performance when 20–30 % of images were labeled. At extremely low label ratios (10 %), instability from pseudo-label noise was observed, though confidence-adaptive filtering alleviated some of these effects. Self-supervised learning (SSL) using DINOv2-style and BYOL pretraining further strengthened representation quality, consistently producing mAP@50 scores above 91 %. SSL-enhanced YOLOv12s generated compact and discriminative embeddings and exhibited smoother optimization, confirmed through loss curves, clustering analyses, and XAI visualizations. Finally, a real-time Streamlit application was developed, enabling image, video, and live-camera detection at up to 22 FPS, demonstrating the practical deployment potential of the proposed framework. This work demonstrates the potential of semi- and self-supervised learning to reduce annotation costs, enhance generalization, and deliver interpretable real-time solutions for precision agriculture.