Estimating colony strength and pollination efficiency in honey bees using a novel dataset and deep learning-based models

• Non-invasive deep learning system monitors honey bee colonies in real time • Public dataset of 79,212 annotated bees across diverse real-world conditions • YOLOv10 and FrCNN achieved >93% accuracy for bee caste and pollen detection • Models robust to weather, lighting, and hive variability, unlike prior methods • Cost-effective system enables practical pollination and colony strength monitoring This research introduces a deep learning-based computer vision system for automated honey bee identification and categorization at hive entrances to estimate different castes. The presence of pollen foragers at hive entrances is used to estimate colony health, colony strength and pollination efficiency. However, no study has created a tool for monitoring completely unmodified hives in real-world conditions. The system was trained and evaluated on video frames captured over two years via GoPro cameras, which encompasses various lighting and weather conditions. Bees were annotated to three categories: 1. Worker bee – No pollen, 2. Worker bee – pollen and 3. Drone bee. Five object detection models were compared: YOLOv10, FrCNN, RetinaNet, FCOS, and SSD. While YOLOv10 (93.9% accuracy, 0.92 F1), FrCNN (93.4%, 0.91), FCOS (92.3%, 0.88), and RetinaNet (91.8%, 0.9) achieved excellent results, SSD (54.2%, 0.69) showed average performance. Best mAP scores were achieved with FrCNN and Friedman's test statistically confirmed a significant difference in model performances (Friedman’s statistics = 20.6667, p < 0.001). The proposed system provides a noninvasive, easy-to-use, and cost-effective method for monitoring honey bee behavior by providing valuable data for beekeepers and researchers. The study also contributes to a publicly available dataset of 4,590 frames with 79,212 bee annotations.