Aircraft noise can geolocate videos

TLDR

• YouTuber Colsto says faint aircraft noise in a recording, matched to flight data, located its camera within about 100 metres.¹²
• Colsto says the test took months, over 5,000 lines of code and 50 GB of data, and released no code.¹³⁴
• A Metabunk thread reports an AI-assisted Sitrec version, kept admin-only, built in about eight hours.⁵

YouTuber Colsto published a video on 2 November showing how propeller-aircraft noise too faint to hear in a recording can be matched to public flight tracks to find where the recording was made.¹³ YouTube's metadata showed about 1.9 million views when checked.¹

Why it matters: background audio becomes a geolocation clue for investigators and a privacy leak for people who post footage, although the demonstration depends on usable aircraft noise, flight data and a bounded search area.³⁴ Colsto says the odds of anyone doing this to a given person are slim and that there are probably a hundred easier ways to find them.²

How the test worked:

• The starting point was a blurry security-camera clip of trees and a riverbank. Colsto knew the date and that it was somewhere in the continental United States, but not the time.³⁴
• A spectrogram, which plots sound frequency over time, showed Doppler curves from two propeller aircraft about 920 seconds apart. Their shape gave ranges for each aircraft's speed and closest distance to the microphone.¹³
• ADS-B flight data for the day was filtered by aircraft type code to keep propeller types. For each point on a 200 metre grid, Colsto computed when each flyover would be heard, allowing for sound travel time and terrain elevation.¹³⁴
• The roughly 15 minutes of silence between the flyovers ruled out points where another aircraft should have been audible. Numerama reports this cut about 73,000 candidates to 15,000, and IBTimes puts the starting figure at about 70,000.³⁴
• Modelled Doppler curves for each candidate were scored against the real spectrogram. After the grid was refined to 100 metres, one location stood out.¹³ IBTimes says tree shapes and a utility pole in Google Street View matched the footage.⁴
• A second run used a single one-minute clip and a known recording time of 20:37 UTC. Numerama says it returned the same area across the continental United States.¹³

Defences: Colsto says common fixes such as high-pass filtering, drowning the audio in noise and phone-call compression did not remove the patterns in his demonstrations. IBTimes adds that he says cutting sections of a recording or extra noise cancellation could make the method harder to use.¹⁴

The big picture: no code or tools were released, but Metabunk's Mick West wrote that the idea would be easy for an AI agent to implement. He later posted "Success!" for a Sitrec implementation, now admin-only, and attached an AI-generated log of about eight hours of recorded work on 2 November.⁵ Other Metabunk posters raised misuse for stalking, and one asked how narrow the date and area must be. West replied that a wider data set takes more processing time and probably brings more false positives.⁵ Posters also pointed to Electric Network Frequency analysis, which matches mains hum in a recording against grid logs to estimate when it was made, as a related audio technique.⁵

Single-source note: the results rest on Colsto's own demonstration of one clip. TechRadar, Numerama and IBTimes relay his account, and none reports an independent test. The Sitrec result is the Metabunk administrator's own report, and the attached images were not reviewed here. Colsto's video was reviewed only through its YouTube description and the outlets' write-ups.