buzzdetect: an open-source deep learning tool for automated bioacoustic pollinator monitoring

Abstract

Ecological studies of pollinators often require long-term and extensive monitoring, posing a significant cost and limitation to research. Traditional sampling methods of observation such as sweep netting and pan trapping provide valuable information on pollinator diversity, but scale poorly when conducting large sampling efforts across space and time. We introduce "buzzdetect", our tool to apply deep learning models to audio data for conducting passive acoustic monitoring on pollinators, and test our accompanying audio classification model. The model is capable of distinguishing the buzzing of insect flight from environmental noise with a sensitivity of 27% at a precision (true positive rate) of 95%. Application of buzzdetect to recordings made in pumpkin, watermelon, mustard, and soybean reveals differences in timing and intensity of foraging: foraging occurred the earliest in pumpkins and the latest for soybean while foraging activity was most intensive in mustard.