Reading the map

Every aircraft that broadcasts ADS-B and comes within range of a sensor is recorded as a pass, whether or not anything was heard. The percentage is how many of those passes had a sound at the moment the sound would have arrived — distance divided by 343 m/s, not the moment the aircraft was overhead.

Counting the misses is the entire point. A network that only recorded its successes could not tell you its range, and the not-heard passes are what a classifier will eventually be trained against.

What it cannot tell you. Nothing on its own. A heard rate of 40% is excellent if background noise fills 10% of the time and meaningless if it fills 40%. Always read it against the bar below it.

The filled bar is the heard rate. The vertical line across it is the share of time a sound — any sound — was in progress. If sounds fill a fifth of the day then roughly a fifth of passes will overlap one by chance, with no detection involved at all.

The second bar is the trap, and it is a real afternoon from this network. A higher number can be worth less. Only the gap between the fill and the line means anything.

The 24-hour baseline above is an average, and averages hide weather. This figure is the share of the last hour with a sound in progress, for the busiest live sensor. Under 25% a match is worth something; over 50% most matches are coincidence whatever the sensors do.

What it cannot tell you. Whether the noise is wind, traffic, or rain. It measures how crowded the acoustic environment is, not what is crowding it.

Passes and detections grouped by airframe. The bar is the heard fraction; the figures on the right are heard over total.

What it cannot tell you. Almost anything, at present sample sizes. This view once showed the A220 at 1 of 17 and it was reported as the strongest finding in the project — engine loudness, apparently. It was an artefact of a 14-hour window when the agent’s clock had drifted and matching was inoperative. With those passes removed the same aircraft sits at 6 of 13. Treat any type with fewer than about fifty passes as noise.

Six bars, each an octave wide, from 50 Hz to 3.2 kHz, showing where the sound’s energy sat. The left two are drawn in a different colour because that is where wind lives.

What it cannot tell you. What made the sound. This was tested directly: normalise the octave shapes of wind, ADS-B-confirmed aircraft and unexplained sounds recorded here and the three curves lie on top of each other — everything falls away at 8–12 dB per octave. An earlier version of this page drew modelled reference curves for “jet” and “propeller” and confidently mislabelled an Airbus A321 as a propeller. The curves were removed. Read the bars as a fingerprint for comparing two sounds, not as evidence of a source.

Tilt is the 50–100 Hz band minus the 200–400 Hz band: how steeply the sound dies away with frequency. It is the one spectral quantity here that separates anything, and the bands show where three measured populations actually sit — each covering the middle eight tenths of that group.

Two things follow. The detector’s wind threshold of 20 dB was set from two gusts on the first day; it turns out to sit almost exactly between wind’s 10th percentile and confirmed aircraft’s 90th. And the unexplained sounds sit on the aircraft distribution, not the wind one — whatever they are, they are not the wind filter leaking.

What it cannot tell you. Separate one aircraft from another, or a drone from a lorry. It is a wind test.

A propeller or piston engine leaves a harmonic stack: tones at regular multiples of the blade-pass or firing rate. This measurement integrates the whole event, whitens the spectrum so only narrow peaks survive, and autocorrelates it to find the spacing. Spacing survives distance — air absorption removes the top of the stack but does not change the gap between what is left.

On synthetic signals at 0 dB SNR it separates rotor sources from wind, road noise and jet aircraft perfectly.

What it cannot tell you. Tell a drone from a scooter. Both are small engines turning something at around 100 Hz, and the same test scores that separation at chance. This is why the readout says “rotor signature” and never a percentage, and why nothing on this map turns red. A confidence figure here would be invented.

A ring around a sensor means it has recently heard something lasting more than twenty seconds with no transponder-equipped aircraft overhead. It turns slowly rather than flashing, because the sound has already finished by the time it reaches the map.

What it cannot tell you. Suggest a drone. Scooters, lorries, construction, rain on a roof and any aircraft without ADS-B all land here. “Unexplained” means unexplained by ADS-B, which is a much weaker statement than it sounds. Sounds that hit the four-minute event limit are excluded from the count entirely, because a continuous noise source would otherwise be reported as a new discovery every four minutes.

Rings collapse inward rather than sweeping outward, because that is what sound does — a microphone receives wavefronts, it does not emit them. They speed up as the sensor picks something up. Grey is offline, amber is listening, green is hearing something at this instant.