An Embedded Deep Learning System for Real-Time Detection of Pollen-Bearing Bees in Colombian Agroecological Contexts

Bee pollination is crucial for biodiversity and agriculture. However, there is a gap in the efficient monitoring of pollen collection activity, which indicates colony health and productivity. This study presents the development and evaluation of an automated embedded system for real-time detection of pollen-carrying bees in Colombian agroecological contexts. A deep learning approach was used, implementing a Convolutional Neural Network based on the DetectNet-V2 architecture, trained with a dataset of 1,070 manually labeled images and deployed on an NVIDIA Jetson Nano platform. The system was validated in field conditions by analyzing 301 annotated instances, demonstrating an overall accuracy of 83.6%, a recall of 88.8%, and an F1 score of 86.1%. These results confirm the feasibility and effectiveness of the system to reliably discriminate between pollen-carrying and non-pollen-carrying bees reliably, providing a valuable tool for intelligent beekeeping, optimization of hive management, and generation of empirical data on foraging activity.