Acoustic AI models monitor bee activity without lethal traps
Researchers at the University of Massachusetts Amherst developed two acoustic AI models that detect bee activity and distinguish honeybees from bumblebees in field recordings. Trained on 594 recordings from 43 northeastern U.S. sites, the models identified bees with 93%–97% precision and tracked closely with trained researchers’ visual observations.
The story
Laura Figueroa and colleagues at the University of Massachusetts Amherst developed two AI models that analyze bee sounds in field recordings. One model identifies overall bee activity at a recording site; the other separates honeybee sounds from bumblebee sounds. The team trained the models on 594 recordings collected over five years from 43 sites in the northeastern United States. In reported testing, the models correctly identified 93%–97% of bees in a sample and their results were highly correlated with visual observations made in the field by trained researchers. The models are available through Raven Pro, an acoustic-analysis tool used by scientists to examine species activity. The work is presented as a complement to, rather than a replacement for, established insect-monitoring methods. Traditional surveys often rely on lethal traps, such as colored cups containing soapy water, followed by laboratory processing. The researchers say acoustic monitoring can help estimate bee activity in settings or conditions where people may find sampling difficult, including tree canopies and heat waves. Their study, “What’s all the buzz about? Passive acoustic monitoring of bees,” was published in Approaches in Entomology: Methods, Philosophies, and Ethics.
Why it matters
Pollinator monitoring is necessary for assessing population conditions and whether conservation efforts are helping, but conventional trapping can also catch nontarget insects and requires substantial sorting after collection. An audio-based measure of activity offers a lower-interference source of field data that can be gathered repeatedly across locations. That may be especially useful when visual surveys are impractical, while allowing lethal collection to remain available for questions that require specimen-level examination.
Evidence and context
The researchers place the tool against documented pressures on pollinators, including pesticide use, habitat loss and climate change. The source says pollinators provide ecosystem services valued at more than $200 billion worldwide and are critical to the reproduction of most plant species. Yet monitoring presents a methodological trade-off: trained entomologists can collect specimens for close identification, but common lethal traps are indiscriminate and can kill large numbers of bees as well as other species. Passive acoustic monitoring shifts part of that work from collection and lab processing to recorded sound analysis. In this study, the models’ agreement with trained visual observations provides evidence that the recordings can reflect field bee activity.
Limits and unknowns
Lethal trapping remains important because many bee species can only be identified under a microscope, the researchers note. The reported models do not identify every bee species: one measures bee activity generally, while the other distinguishes only honeybees from bumblebees. Their reported precision and correlation with visual surveys come from recordings at 43 northeastern U.S. sites, so the article does not establish how the models perform in other regions, sound environments or taxonomic groups. The work therefore supports acoustic monitoring as an additional method, not a substitute for specimen-based identification or broader population assessment.
Follow the story
Share it, watch the explainer, or continue with the reporting trail.
Story timeline
Earlier insect-identification work
Laura Figueroa previously reported that some deep-learning neural-network models could identify insects, providing work the team used in developing the pollinator-focused models.
Acoustic AI models monitor bee activity without lethal traps
A short visual explainer for readers who want the story in a different rhythm.
Highlights
- Monitoring bees by sound
- Two models, two tasks
- Tested against field observations
- A way to check tree canopies
Transcript
Researchers have a new way to listen for bee activity without collecting the insects.
At UMass Amherst, Laura Figueroa's team built one model for activity, another distinguishing honeybees and bumblebees.
Trained on 594 northeastern recordings from 43 sites, the models reached 93 to 97 percent precision.
Field researchers could measure bee activity in hard-to-reach tree canopies without setting lethal traps.
Many bee species still need microscope identification; performance beyond northeastern sites hasn't been established.