Counting bees has long meant catching them. Researchers who wanted to know which species were visiting a field or garden often had to collect specimens, which killed them in the process. ([1])A team of ecologists at the University of Massachusetts Amherst is testing a different route. Their approach relies on sound, and it may let scientists track pollinators without collecting a single insect.
Listening instead of trapping
Traditional pollinator research depends on lethal traps to identify species and produce population counts. The Amherst researchers looked for a way to collect similar data from the bees’ buzzing. Their findings were published in Approaches in Entomology: Methods, Philosophies, and Ethics, and The Wildlife Society highlighted the work.
The core idea is simple. Each species produces a slightly different sound in flight, and field recordings can capture those differences. The challenge lies in sorting through hours of audio to pick out bees from everything else in the air, which is where artificial intelligence comes in.
How the models were built
The team developed two models using convolutional neural networks, a type of artificial intelligence that is most often applied to images or sound. The first model estimates overall bee activity. The second goes further and separates honeybees from bumble bees, two groups that many people would find difficult to tell apart by sound alone.
Training required a substantial body of real-world data. The models learned from 594 field recordings gathered over five years at 43 sites across the northeastern United States. That long timeline and wide geographic spread give the software a range of conditions to learn from, rather than a single meadow or season.
What the results show, and what they do not
The models were accurate in testing. They correctly identified between 93% and 97% of bees in a sample, and their output was highly correlated with what researchers recorded by sight in the field. Those figures suggest the software can track bee activity in a way that agrees with human observers.
Still, the models do not replace every tool in the field kit. Identifying many bee species still requires a microscope, so traditional lethal traps remain part of the process. The researchers presented the models as an addition to existing methods rather than a substitute.
| Area | What the study reports |
|---|---|
| Finding | Correct identification of 93% to 97% of bees in a sample |
| Limitation | Many species still need microscope examination |
| Next step | Using audio to monitor pollinators in places that are hard to sample |
Where the approach could lead
The most practical promise may lie in places that researchers struggle to reach. Remote or difficult terrain can make traditional sampling costly or impractical, and audio recorders can be left in position to gather information over time. The study suggests this could open new monitoring opportunities, though the published findings do not establish how far such methods can scale.
For now, the result is a more careful way to count bees. Scientists gain another route to pollinator data while avoiding the need to kill the insects they study, and the buzz that once served as background noise may now carry useful information.
