The buzzing of the bees: Machine-learning bioacoustics as a targeted and non-lethal method for monitoring bees

Effectively monitoring populations of concern underlies successful conservation and restoration efforts, yet the limitations and possible negative side effects of current pollinator monitoring methods are increasingly being recognized. For example, many commonly used methods are lethal, result in bycatch, involve considerable manual effort to clean, sort, and pin the bees, and require highly skilled taxonomic expertise to identify the specimens. Use of machine learning to non-invasively monitor a broad diversity of animals, from elephants and whales to bats and katydids, is becoming increasingly common. Similarly, the characteristic buzzing sounds of bees can be used to train machine-learning bioacoustics models. Here we present a convolutional neural network model using audio samples of bees collected over a 5-year window across a broad diversity of environmental contexts. The model was trained on sounds produced primarily by two dominant bee taxa: the European honey bee (Apis mellifera) and bumble bees of northeastern North America (Bombus spp.). The model has high accuracy for those groups, especially when the bees fly within 50 cm of the audio recorder. The model will be available through the RavenPro platform, a user-friendly desktop program with step-by-step instructions available, to facilitate use. Given that all methods have strengths and limitations, the goal is for bioacoustics to be a part of a multifaceted monitoring toolkit available to researchers and practitioners interested in understanding pollinator foraging dynamics and phenology across broad spatial and temporal scales.