Unraveling Honey Bee's Waggle Dances in Natural Conditions with Video-Based Deep Learning
Sylvain Grison,Rajath Siddaganga,Shrihari Hedge,James Burridge,Pieter M. Blok,Smitha Krishnan,Axel Brockmann,Wei Guo
DOI: https://doi.org/10.1101/2024.11.16.623982
2024-11-18
Abstract:1- Wild and managed honey bees, crucial pollinators for both agriculture and natural ecosystems, face challenges due to industrial agriculture and urbanization. Understanding how bee colonies utilize the landscape for foraging is essential for managing human-bee conflicts and protecting these pollinators to sustain their vital pollination services. To understand how the bees utilize their surroundings, researchers often decode bee waggle dances, which honey bee workers use to communicate navigational information of desirable food and nesting sites to their nest mates. This process is carried out manually, which is time-consuming, prone to human error and requires specialized skills.
2- We address this problem by introducing a novel deep learning-based pipeline that automatically detects and measures waggle runs, the core movement of the waggle dance, under natural recording conditions for the first time. We combined the capabilities of the action detector YOWOv2 and the DeepSORT tracking method, with the Principal Component Analysis to extract dancing bee bounding boxes and the angles and durations within waggle runs.
3- The presented pipeline works fully automatically with videos taken from wild Apis dorsata colonies in its natural environment, and can be used for any honey bee species. Comparison of our pipeline with analyses made by human experts revealed that our procedure was able to detect 93\% of waggle runs on the testing dataset, with a run duration Root Mean Squared Error (RMSE) of less than a second, and a run angle RMSE of 0.14 radians. We also assessed the generalizability of our pipeline to previously unseen recording conditions, successfully detecting 50% of waggle runs performed by Apis mellifera bees from a colony managed in Tokyo, Japan. In parallel, we discovered the most appropriate values of the model's hyperparameters for this task.
4- Our study demonstrates that a deep learning-based pipeline can successfully and automatically analyze the waggle runs of Apis dorsata in natural conditions and generalize to other bee species. This approach enables precise measurement of direction and duration, enabling the study of bee foraging behavior on an unprecedented scale compared to traditional manual methods contributing to preserving biodiversity and ecosystem services.
Ecology