Investigating Traditional Machine Learning Models and the Utility of Audio Features for Lightweight Swarming Prediction in Beehives

Abstract

Remote monitoring of the status of beehives is essential for efficient beekeeping, leading to less workload on the beekeeper and, because of not opening the hives too frequently, to less stress for the colonies. Sound analysis, utilizing machine learning models of various paradigms, is a common feature of so-called smart hives. Most of these models are aimed at the task of swarming prediction. Swarming of a colony, a fundamental phenomenon in the reproductive process of bees, can cause substantial losses in the production of the apiary and, thus, its prediction is of utmost importance. However, especially in case of nomadic beekeeping where the apiary is moved to the country without access to electricity and good internet connection, the used prediction models should run “on-site” with as low energy consumption as possible and using internet connection only to send alerts to the beekeeper. For such, lightweight models are required which can be achieved by using simpler prediction models and/or only the most important audio features. In this paper, the importance of audio features for swarming prediction is investigated by using a genetic algorithm. Various Machine Learning models are trained, using the selected features, and used for predicting swarming on real-world data collected in one Hungarian apiary. This experimental evaluation is the main contribution of this paper. While genetic algorithms are commonly used for feature selection, however, to the best of the authors' knowledge, they have not yet been used in the beekeeping domain.