KEY TAKEAWAYS
- Different bird species pose different food safety risks.
- Project harnesses artificial intelligence and machine learning to identify bird species.
- Based on the species present, a computer would decide whether to deploy various audible, optical and/or physical hazing methods.
- By mixing deterrents and durations, the hypothesis is it will slow bird habituation.
Referenced CPS research:
Not all bird species are created equal when it comes to the potential food safety risk they pose with produce. Chetan Badgujar, Ph.D., with the University of Tennessee, Knoxville, hopes to develop a digital toolbox that enlists digital sound surveillance, artificial intelligence and machine learning to identify bird species in a field. If previous research identified the species as high risk, a computer-based toolbox would automatically deploy a mix of hazing techniques.
By using deterrents only when high-risk birds are present, and deploying a wide variety of methods and at differing intervals, Badgujar said he hypothesizes it will slow birds from becoming habituated. At the same time, it should reduce the impacts on beneficial pest-consuming birds.
“What happens is the birds get habituated because the deterrents are triggered at set frequencies, maybe every 10 to 15 minutes,” he said. “They do not provide any feedback, and that’s why the birds become habituated. Our thought process is we can avoid the habituation problem by providing some type of feedback loop and only triggering the deterrents if we identify birds in the field.”
Not only does Badgujar hope the toolbox reduces losses caused when growers can’t harvest a portion of a field because of suspected bird contamination, but it also may help reduce direct losses caused by birds feeding on crops.
His proof-of-concept project builds on CPS-funded research led by Daniel Karp, Ph.D., with the University of California, Davis, that examined the prevalence of different bird species in agriculture and whether they carry and transmit food-borne pathogens. Karp’s work found that small insect-eating birds, such as swallows, posed a lower food safety risk than birds that flock near livestock, such as blackbirds and starlings that are more likely to transmit pathogens.
Rather than reinvent the proverbial wheel, Badgujar is using the BirdNET Sound ID, an artificial intelligence-powered bird recognition app developed by Cornell University. The app includes recording of about 950 different North American bird species.
Using a small microphone with a range of about 10 feet connected to a computer, an undergraduate student, Paul Chua,tested BirdNET’s accuracy in identifying species found in greenbelt areas around the UT campus. The collected data was used to develop a digital toolbox that can be placed near birds.
Currently, the researchers are using a directional microphone with a range of about 100 feet to gauge the toolbox’s accuracy in detecting the presence of birds and indentifying species. So far, the software has been 50% to 90% accurate in species identification, with better results for birds producing higher-pitched calls compared to those with lower-pitched calls.
One of the challenges has been background noise, which may interfere with capturing bird calls. Enter co-investigator Hao Gan, Ph.D., and an UT electrical engineer who specializes in developing micro-scale systems with a focus on AI-based modeling. His goal will be to filter out unnecessary noise.
Once the researchers achieve a confidence level of at least 70% in bird identification, Badgujar said they plan to add triggers to deploy various deterrents. He envisions three types: audio, such as firecrackers or gunfire sounds; visual, such as laser beams or flashing colored lights; and physical such as coils of reflective ribbons that unfurl and wave in the wind.
They plan to test the completed toolbox in farm-scale produce fields at the nearby UT Organic Crops Unit as well as in berry blocks at the Plateau AgResearch and Education Center about 90 minutes west of Knoxville.
The sedentary tool box is only effective for small fields of 2 to 5 acres.
For larger operations, Badgujar envisions the toolbox installed on an autonomous robot that roams fields or orchards to conduct bird monitoring and selective hazing methods in conjunction with pest scouting and other activities. The robotic set-up is something he’d like to pursue in the future.