FLORES, Guatemala — This March, rangers on patrol in the Maya Forest came across the feathers of hunted birds and paths that had been cleared through the trees. These led them to a 2-hectare (5-acre) opening in the forest where squatters likely planned to settle and then expand.
The people who’d cleared the forest were nowhere to be found. The deforestation had occurred around eight days before, the rangers guessed. Even with camera traps and other technology, there’d been almost no way to detect it in real time.
Rapid response has long been a challenge for conservationists in the Maya Biosphere Reserve, which spans 2.2 million hectares (5.3 million acres) across northern Guatemala. The reserve is a patchwork of national parks, logging concessions and biological corridors, some of them under pressure from cattle ranching and illegal logging.
“If we’re going out regularly to a site every two or three months, and something happens a day after the last visit, then two or three months will go by with no information,” said Rony García Anleu, director of biological research at the Guatemala office of the Wildlife Conservation Society (WCS).
A new project in the reserve aims to decrease ranger response times with bioacoustics devices that can “listen” for illegal activity, using AI models trained to identify sounds associated with logging, hunting and other crimes.
It’s part of the $100 million AI for Climate and Nature Grand Challenge, run by the Bezos Earth Fund for innovative uses of artificial intelligence for tackling biodiversity loss, climate change and food insecurity.
The project will also have a second component in the Pantanal wetland with WCS Brazil.
WCS, the Lab of Ornithology at Cornell University in the U.S., Chemnitz University of Technology in Germany, and Brazil’s Federal University of Mato Grosso do Sul submitted the proposal together. The group was one of 15 winners of the challenge, each given as much as $2 million.
“They’re basically giving us ears in the forest to detect and address threats in a much quicker and much more efficient manner,” said Jeremy Radachowsky, WCS director for Mesoamerica and the Caribbean.
WCS has already been using acoustic devices in the Maya Biosphere Reserve for around three years, but that technology is simpler and less effective. The devices collect only a few hours of sound per day, and require a ranger to physically travel to the location of each device — sometimes days away from the nearest station — and bring back the memory cards for review.
Researchers at Cornell are providing much more sophisticated devices, which are scheduled to be tested this year and installed early next year. They include a machine-learning model that can be trained to recognize the sounds of gunshots, chainsaws, engines and other human activity. Small data packages comprising short sound snippets and metadata — including location, date, time and other basic pieces of information — are sent via satellite to an online repository accessible by researchers and rangers.
The model can also learn species sounds. The team plans to include some biological sounds, such as the vocalizations of scarlet macaws (Ara macao), but will focus on environmental crime, several members said.
“You transmit the information so we can look at it and say, ‘Yeah, this was a gunshot. We should go out there and get ready and [look into] that.’ It’s a sensible approach,” said Holger Klinck, director of Cornell University’s K. Lisa Yang Center for Conservation Bioacoustics.
Klinck has built a career around bioacoustics technology, having spent years using underwater drones and other devices to monitor marine mammals in offshore areas where the U.S. Navy was conducting training activities.
Underwater sound travels efficiently and marine mammals tend to “chatter” to communicate, often with distinct and isolated acoustic signatures, Klinck said. In some cases, underwater projects require just one recording device. But terrestrial soundscapes can be much more challenging. In a high-biodiversity forest, there are multiple, overlapping animal sounds occurring at different frequencies, as well as the sounds of wind, chainsaws and engines. Many of those sounds can mask quieter species.
Others are extremely similar to each other, such as branches snapping and gunshots.
“We’re talking about areas that have some of the highest levels of biodiversity on the planet, so there’s a lot going on,” said Laura Figueroa, a professor of environmental conservation at the University of Massachusetts Amherst in the U.S., who isn’t associated with the project but uses bioacoustics for conservation ecology and entomology.
She added, “You have inadvertent sounds that are produced when organisms fly. You have bats trying to echolocate to find their prey. There’s just so much happening both at big scales and with small invertebrates that are also making sounds.”
To ensure accurate detections, the team will train the AI model before installing the devices. The process involves feeding it recordings of engines, chainsaws, firearms and other sounds so that it can “learn” how to differentiate their subtle acoustic signatures.
The algorithm should be able to handle between 50 and 100 sounds, Klinck said.
