Written by Dr. Fabiana Fragoso
Imagine you’re standing in the middle of a soybean farm with thousands of acres surrounding you. What do you hear? Maybe a tractor plowing some fields, the leaves rustling in the wind? Or it may be a faint noise, barely noticeable among the background noise of the landscape. For most of us, outdoor sounds are usually just background noise. But what if you could tune your ears to discover hidden layers of data in your environment?
That’s exactly what researchers from the Ohio State University (OSU) Department of Entomology and Dartmouth College’s Department of Computer Science set out to do. In a study published in December, insect science journalthe team describes a new open-source tool called “buzzdetect” that uses machine learning and passive acoustic monitoring to detect pollinator activity.
Limitations of traditional sampling
As anyone who has done pollinator fieldwork knows, collecting data throughout the day or over an entire season is a significant logistical challenge. “Suppose you want to look at trends in daily pollination activity in 30-minute increments,” says Dr. Luke Hearon. student at OSU and lead author of the study. “That means we have to be on site from sunrise to sunset, taking samples with sweep nets, replacing hives, and visually counting them every 30 minutes. That’s not a bad way to spend the day, but at the end of the day, the effective sample size is n = 1.”
Multiply that effort over multiple locations, multiple days, or an entire growing season, and the limitations of traditional methods quickly become apparent. Passive acoustic monitoring provides a way around this bottleneck. “With passive acoustic monitoring, you can immediately deploy microphones at all sites and then continue recording continuously for several days,” says Hearon. “The temporal resolution of the data returned is virtually unlimited.” In other words, buzzdetect’s 24-hour “listen” capability allows researchers to continuously track pollinator activity across multiple locations and on a much finer time scale than traditional methods.
Turn buzz into data
Hearon and his colleagues created buzzdetect by applying deep learning models to field audio. Deep learning is a field of machine learning in which computer models inspired by the structure of the human brain learn to recognize patterns by being exposed to large amounts of data.
For BuzzDetect, the team started by setting up simple audio recorders in the field and listening to the recordings to note when they heard insect calls. Then, rather than starting from scratch, we used a shortcut known as transfer learning. They used YAMNet, a pre-trained Google audio model (already trained on many everyday noises), and tweaked it to specifically focus on the sounds of flying insects. The result is a model that can distinguish between insect wing sounds and environmental noise in seconds with 28% sensitivity and 95% accuracy.
To see how BuzzDetect works outside the lab, the researchers deployed microphones in farm fields and analyzed 24 hours of recordings from five different plants: pumpkin, watermelon, mustard, soybean, and chicory. The pattern they detected was broadly consistent with what previous studies have already shown about pollinator behavior. Chicory showed a clear early morning peak of activity, whereas soybean peaked later in the day, and overall activity was higher for mustard and soybean than for other crops. Buzzdetect also revealed how much variation exists even within a single crop. For example, in a watermelon field, some recorders captured more than 4,000 wingbeats per day, while others detected closer to 1,200, highlighting real differences in local pollinator activity rather than just noise in the model.
Despite its impressive performance, buzzdetect, like any automated system, can sometimes make mistakes. But, as Hearon points out, they are often reasonable. “Most of the false positives are ones where you can hear or see the spectrogram and think, ‘Oh, I can see why I heard bees in this.'”
However, some confusion remains memorable. “We’ve heard mysterious rattles and clicks, fierce arguments between squirrels, and a hundred different variations of insect calls,” he recalls. These examples highlight some of the challenges in labeling the training set. “We came up with a set of labels to use, such as ‘frog,’ ‘propeller plane,’ and ‘siren.’ And then at 2 a.m. you hear an unknown animal sniffing loudly at your microphone. Do you lump this in with ‘ambient noise’? Create a new label for ‘ambient sniffing’? Can you train a snort detection model? There are a lot of new questions in the field of bioacoustics.”
Tools open and accessible to many users
Hearon emphasized that the team is proud to have created a tool that is free, open source, and designed to run on relatively small amounts of hardware. “Just to be clear, ‘modest’ was a bit of a euphemism. The GPU we’re running our analysis on (GTX 1650) is one of the cheapest cards from four GPU generations ago. Our audio recorder is also a simple MP3 recorder, so we don’t have any expensive, complex scientific equipment in the pipeline,” he says.
This makes buzzdetect potentially useful far beyond academic research institutions. “Crop growers might measure pollinator activity in their fields before applying pesticides, public gardens could compare the attractiveness of different pollinator habitats, and citizen scientists could track the activity patterns of native bees nesting in backyards. There are many unanswered questions about how to apply and interpret bioacoustics, so the more people involved, the better,” Hearon says.
We also emphasize that tools like buzzdetect are becoming increasingly accessible to build and use. “Aside from deciphering some very confusing error messages from TensorFlow, the process is surprisingly easy,” says Hearon. “The barriers to entry for machine learning are getting lower every day, software packages are designed to be reasonably frictionless, and walkthroughs abound online. I didn’t know anything about Python until I started working on buzzdetect, and I encourage curious readers to jump in with both feet and see where it takes them.”
Buzzdetect, a new open-source AI tool, uses a simple microphone and machine learning to continuously detect pollinator activity. New research shows how this tool could provide researchers and growers with a low-cost way to “listen” to what their bees are doing in real time. Shown here is a series of buzz spectrograms recorded by buzzdetect. (Video provided by Dr. Luke Hearon)
complementary tools
uzzdetect offers a promising new method for automated large-scale pollinator monitoring, but Hearon and colleagues are careful to frame it as a complement to traditional methods, rather than a replacement.
“So far, the trends in buzzdetect are largely supported by previously published research, so there are no particular surprises,” he says. “All sampling methods, including bioacoustics, are subject to some bias. Therefore, future bioacoustic results may not match previously published trends, but it can be difficult to determine which trends are misleading. This is why we present buzzdetect as a complementary tool to existing methods. The strongest conclusions need to be supported by integrating multiple streams of evidence, including traditional sampling methods.”
Listen beyond the data
Beyond the practical application, Hearon says working on buzzdetect changed the way he thinks about sound in the natural environment. “One of the things I’ve realized over the course of this work is that there is so much you can learn by listening,” he says. “Training To build my dataset, I spent hours listening to audio in the field. Although the environment in the depths of a soybean field is quiet compared to the city, it’s still packed with noise. Our senses are filtered by attention and context, so these background sounds rarely register in our heads. But as you sit at your desk with your eyes closed and listen intently to the recording, you begin to notice the entire soundscape that has always been there: the chirping of birds, the sound of the wind blowing through the leaves, all of which sounds different during the quiet hours of the day than during the quiet hours of the night. The set is grouped under one label: Ambient Background. ”
Perhaps as important to Herron as the tools his research produced are the broader lessons he learned along the way. “I love training exciting new techniques and powerful models, but at the end of the day they are just models. The truth remains and always will be outside the world.”
Dr. Fabiana Fragosois an entomologist, biologist, translator and interpreter from Brazil, currently based in Italy. She recently served as a postdoctoral researcher at the U.S. Department of Agriculture, Agricultural Research Service, Madison, Wisconsin, USA. Email: fabianapfragoso@gmail.com.
related
Learn more about entomology today
Subscribe to receive the latest posts by email.
