Enhancing Pollinator Protection Using Real-Time Visual Recognition

Abstract

In the context of accelerating climate change and its detrimental impact on global food systems, the integration of advanced technologies for monitoring pollinator populations has become critical for ensuring environmental sustainability and food security. The preservation of pollinators is essential for human survival. This study investigates the application of Computer Vision and Object Detection techniques to automatically analyze bee activity through image data. A new dataset comprising 6,993 bee-containing images was curated from video footage and manually labeled using bounding boxes. The dataset was divided into training (5,203), validation (902), and testing (828) subsets. Evaluation of various fine-tuned YOLO models based on the COCO framework revealed that YOLO v5 m delivers the highest recognition accuracy. Nevertheless, YOLO v5 s emerged as the most efficient for real-time detection tasks, achieving an average inference time of 5.7 milliseconds per frame, albeit with a modest reduction in detection performance. The final model was embedded within an explainable AI system that generates time stamped visual summaries, allowing easy interpretation by non-technical users, including key stakeholders in the apiculture sector, to support sustainable practices in pollinator management.