HONOLULU, Dec. 2, 2025 — Wildlife poaching remains a major conservation concern. Technological advancements have enabled webs of acoustic sensors to be deployed throughout rainforests, creating the possibility of real-time alerts to the sounds of gun-based poaching.
But the belly of the rainforest is loud, and sorting through a constant influx of sound data is computationally demanding. Detectors can distinguish a loud bang from the whistles, chirps, and rasps of birds and bugs. However, they often conflate the sounds of branches cracking, trees falling, or water dripping with gunshot noises, resulting in a high percentage of false positives for gunshot detectors.
Naveen Dhar, along with collaborators from Cornell University's K. Lisa Yang Center for Conservation Bioacoustics and Elephant Listening Project, aimed to develop a lightweight gunshot detection neural network that can accompany sensors and process signals in real-time to minimize false positives.
Dhar will present his model Tuesday, Dec. 2, at 3 p.m. HST as part of the Sixth Joint Meeting of the Acoustical Society of America and Acoustical Society of Japan, running Dec. 1-5 in Honolulu, Hawaii.
The model works with autonomous recording units (ARUs), which are power-efficient microphones that capture continuous, long-term soundscapes. The proposed system utilizes a web of ARUs deployed across the forest, each performing real-time detection, with a central hub that handles more complex processing.
An initial scan filters all audio for "gunshot likely" signals and sends them to the ARU's microprocessor, where the lightweight gunshot detection model lives. If confirmed as a gunshot by the microprocessor, the ARU passes the information to the central hub, initiating data collection from other devices in the web.
By determining if other sensors also hear a "gunshot likely" noise, the central hub then decides whether the event was a true gunshot or a potential false positive. If it determines a true positive, the central hub collates audio files from each sensor, allowing it to pinpoint the location of the gunshot and alert rangers with coordinates for immediate poaching intervention.
"Down the road, the device can be used as a tool for rangers and conservation managers, providing accurate and verifiable alerts for on-the-ground intervention along with low-latency data on the spatiotemporal trends of poachers," said Dhar.
He plans to expand the model to detect the type of gun that fires each gunshot and other anthropogenic activities, such as chainsaws or trucks, before field-testing the system, which is currently under development.
"I hope the device can coalesce with Internet of Things infrastructure innovations and cost reduction of materials to produce a low-cost, open-source framework for real-time detection usable in any part of the globe," said Dhar.
