A new cell-phone-sized device—which can be deployed in vast, remote areas—is using AI to identify and geolocate wildlife to help conservationists track endangered species, including wolves around Yellowstone National Park.
The battery-powered devices—dubbed GrizCams—are designed by a small Montana startup, Grizzly Systems. Together with biologists, they’re deploying a constellation of the devices across the Greater Yellowstone ecosystem to record audio and video of when and where wolves or wolf packs howl.
Once fully deployed, the data can help scientists and conservationists better understand wolf behavior and create new strategies for deterring wolves from attacking livestock.
Conservationists retrieve audio data from SD cards on remote recorders every few months. That data is fed into and analyzed by AI models trained using terabytes of data of howling wolves. The model—a convolutional neural network—converts the audio into a spectrogram, which analyzes the data, identifying different aspects of a wolf’s howl and geolocating where the sounds originated.
Grizzly Systems trained the model using NVIDIA A100 Tensor Core GPUs in the Azure cloud and PyTorch framework running NVIDIA CUDA-X libraries. For inferencing, they use NVIDIA Triton Inference Server and ONNX Runtime for model optimization, with an NVIDIA RTX 4090 for on-prem storage of sensitive data and local inference.
Video 1. A wolf pack recorded in 2023 in Yellowstone National Park vocalizes in chorus and asynchronously
Grizzly Systems CEO, Jeff Reed, PhD, highlighted how the system monitors large tracts of land 24 hours a day, every day of the year. The devices can help perennially under-resourced wildlife managers and state and federal agencies monitor lands that often lack personnel.
The AI model can identify varied pitches and intonations of wolf vocalizations, which can carry more than six kilometers from where they originate. Knowing where a pack moves by tracking their howls can help conservationists identify a wolf’s territorial boundaries.
While the model can’t yet identify individual wolves from their howls, Reed said future iterations of the technology will have that capability.
Figure 1. GrizCam collects sounds or video, runs through an on-device thin model layer, and analyzes data on a cloud-based LLM
Today, the GrizCams make up one part of a larger conservation effort aimed at balancing competing interests in the land.
These include the small but growing wolf population in Montana, which needs wild prey for food; the billion-dollar Yellowstone eco-tourism economy, which relies upon healthy wildlife populations; and the ranchers who need to protect their livestock, and whose land offers critical habitat for wildlife.
“Wolves, grizzlies, elk can be a hassle to a rancher because they might kill their livestock, or tear down their fences,” said Reed, who before starting up Grizzly Systems three years ago, spent his career working in the tech industry. “On the flip side, those ranches also provide critical habitat for wildlife on private lands around Yellowstone.
“If our devices can detect a lone wolf coming through a ranchland because we have AI on it, then we can playback the sound of guardian dogs barking, or a gunshot, or a large territorial wolf pack, which can “encourage” that wolf to move out of that area. But that requires vigilance throughout the day and the night—and nobody is sitting outside 24/7—which is where AI comes in.”
Another way AI is helping conservationists is by streamlining the data collection process.
The remote recorders—which can also be deployed with video capabilities—run a very thin-layer AI on-device, which weeds out most motion that would otherwise trigger false-positive recordings. The recorders can ignore wind rustling through grass or trees, or bright light reflecting off snow—two common stimuli that trigger false-positive recordings on remote devices.
As a result, the GrizCam’s batteries last longer and require less servicing by wildlife managers and landowners.
AI is also useful to conservationists as it quickly sifts through terabytes of recorded data to quickly identify and flag relevant audio or visual signatures.
While the on-device AI cuts down on unwanted recordings, it nevertheless records sounds and imagery of biological activity—including birds, elk, or bears moving across terrain and making noises.
“These acoustic recorders are gathering data with AI, they’re recording 24-7, every day for a year across 50 or so recorders,” said Reed. “With AI, we can crunch through the data, go through and identify wolves or other endangered species if they’re there, and then work with conservationists to say, ‘okay, we gotta go protect this area and do some additional conservation over there.’”
Grizzly Systems plans to continue its close collaboration with conservationists. It also foresees its rugged edge devices’ relevance for a variety of industrial use cases, including remote surveillance.
Reed points out that 97% of the Earth’s surface lacks access to an electrical outlet. A rugged recorder can monitor oil and gas rigs, as well as remote electrical transformers, which, in very rural areas, can attract vandals who take them offline.
“AI is a great example of how, if we can get it right, with battery life and ruggedness, we can monitor illegal activity that hurts us all,” Reed said. “Poaching, illegal wildlife trafficking, illegal logging or mining in the Amazon— this is activity that ends up hurting the vast majority of people and the planet–and which technology can help prevent.”
Read more about the partnership between Grizzly Systems and Yellowstone National Park.
Check out additional reporting on wolf conservation and decoding wolf verbalizations.
