Automated Detection and Classification of Pollen Grains Using YOLO-Based Deep Learning Models

Authentication of high-value honey, particularly from the biodiverse Anzer Valley, relies on melissopalynology to verify botanical origin. However, traditional manual analysis is time-consuming and subject to observer bias. This study developed an automated system for the simultaneous detection and classification of 53 pollen grains belonging to ecologically and apiculturally important taxa distributed in Anzer Valley using advanced deep learning architectures. A novel dataset was constructed from approximately 2095 light microscopy images of the pollen grains, enhanced to 19,860 images via geometric and photometric data augmentation to improve model robustness. Two state-of-the-art object detection models, YOLOv11 and YOLOv12, were trained and compared. The YOLOv11 architecture outperformed YOLOv12, demonstrating superior stability and precision with a mean Average Precision (mAP@0.5) of 93.2% and an F1-score of 0.896. The model showed exceptional generalization on unseen data (recall 90.6%), successfully handling the visual complexity of raw microscopic samples. Misclassifications were minimal, primarily occurring among pollen grains belonging to morphologically similar taxa of Geraniaceae and Fabaceae. These findings demonstrate that YOLOv11 provides a reliable, high-throughput framework for automated pollen analysis. By significantly reducing processing time while maintaining high accuracy, this system offers a scalable tool for the identification of pollen grains, overcoming the bottlenecks of manual microscopy. Beyond its technical contribution, this approach may support more objective, standardized, and scalable regulatory practices for honey authentication and geographical origin verification.