Buzzdetect: A machine-learning tool for bioacoustic detection of flying insects

Pollinator research often requires in situ detection and quantification of pollinator activity. Many methods exist, none without its shortcomings. Visual observation and identification is direct, but labor intensive, requires expertise, and may not be possible when flowers are not in clear view. Traps such as bee bowls or sticky cards generate cumbersome samples with low temporal resolution and counts of trapped insects may be a loose proxy for true endpoints. To better detect pollinator activity we developed a machine learning tool called “buzzdetect” that detects the audible buzz produced by certain flying insects. While we are constantly refining buzzdetect’s performance and capabilities, our currently released model is able to detect buzzes from honey bees and similarly sized insects (bumble bees, wasps, flies, etc.) with roughly 87% precision and 78% recall. Our model is tuned for and tested on honey bees in agricultural environments, but buzzdetect also contains tools to create new models for different environments and focal organisms. buzzdetect facilitates simple, low-cost, non-destructive, and massively scalable acoustic surveys of insect flight activity.