Buzzdetect: Open-Source AI Listens for Pollinators in Real Time

Imagine standing in a soybean field, acres stretching in every direction. Beyond the rustling leaves and distant machinery, a faint, almost imperceptible buzz hangs in the air. For most, it’s background noise. But researchers are now developing tools to “listen” to that buzz, unlocking a wealth of data about pollinator activity with a new open-source tool called “buzzdetect.”

The Challenge of Traditional Pollinator Monitoring

Traditional methods of tracking pollinators – sweep netting, pan trapping, and direct observation – are labor-intensive and limited in scale. “Let’s say you want to study the daily trend of pollination activity with a 30-minute resolution,” explains Luke Hearon, a Ph.D. Student at The Ohio State University and lead author of a study published in December in the Journal of Insect Science describing buzzdetect. “That means you need to be in the field from sunup to sundown, taking samples every 30 minutes. That’s not a bad way to spend a day, but your effective sample size after all that is n = 1.”

Scaling this effort across multiple sites and over extended periods quickly becomes impractical. Passive acoustic monitoring offers a solution. “With passive acoustic monitoring you can quickly deploy microphones at all of your sites and leave them to continuously record,” Hearon says. “The data you get back have practically unlimited temporal resolution.” Buzzdetect’s continuous listening capability allows researchers to track pollinator activity at a much finer scale than previously possible.

From Buzzes to Data: How Buzzdetect Works

The team developed buzzdetect by applying deep learning models to audio recordings. Deep learning, a subset of machine learning, uses computer models inspired by the human brain to recognize patterns from large datasets. Rather than building a model from scratch, the researchers leveraged a pre-trained Google audio model called YAMNet, originally designed to identify everyday sounds. They then “fine-tuned” YAMNet to specifically recognize the sound of insect flight.

The resulting model can distinguish insect buzzes from environmental noise with a sensitivity of 28% and a precision of 95%. This means that while it doesn’t catch every single instance of insect activity, it’s highly reliable when it *does* detect a buzz. The researchers tested buzzdetect by deploying microphones in agricultural fields, analyzing 24-hour recordings from pumpkin, watermelon, mustard, soybean, and chicory crops.

Validating the Results

The patterns detected by buzzdetect largely aligned with existing research. For example, chicory showed peak activity earlier in the morning than soybean. Overall activity was higher in mustard and soybean compared to the other crops. The tool also revealed variations within single crops – watermelon fields showed a range of buzz detections from 1,200 to over 4,000 in a single day, highlighting localized differences in pollinator activity.

Hearon acknowledges that, like any automated system, buzzdetect isn’t perfect. “Most false positives are things you can listen to or look at the spectrogram and think, ‘yeah, I see why it heard a bee in that.’” He also noted encountering unexpected sounds during data analysis, such as “mysterious rattles and clicks, heated arguments between squirrels,” highlighting the complexities of interpreting real-world audio data.

Open Source and Accessible

A key aspect of buzzdetect is its open-source nature and relatively low hardware requirements. “The GPU we run our analyses on was one of the cheapest cards four GPU generations ago,” Hearon explains. “Our audio recorders are also simple MP3 recorders, so there’s no expensive or complicated scientific gear in the pipeline.” This accessibility opens up possibilities for a wider range of users, including crop growers, public gardens, and citizen scientists.

Buzzdetect is available through its documentation website, offering a command-line interface, a Python API, and a graphical user interface. The team also encourages users to cite the published study when using the tool, and to cite the specific Zenodo DOI corresponding to the version of buzzdetect used for reproducibility.

A Complementary Approach

The researchers emphasize that buzzdetect is intended to complement, not replace, traditional pollinator monitoring methods. “Every sampling method is biased in some way, including bioacoustics,” Hearon says. “The strongest conclusions should be supported by synthesizing multiple streams of evidence, including traditional sampling methods.”

Beyond the practical applications, Hearon reflects on a shift in perspective that came with the project. “One thing that I realized during the course of this work is that there is a massive amount to be learned by listening,” he says. “You start to become aware of an entire soundscape that has always been there.”

Buzzdetect represents a step towards a more comprehensive understanding of pollinator activity, offering a new way to listen to the subtle sounds of the natural world and glean valuable insights from the data they contain.