Skip to content

Python tool for aircraft Doppler audio geolocation using historical ADS-B flight data (Source-available)

7.8 KB
Raw Blame

Experimental coherent audio proposals

The ordinary per-frame peak linker can miss a continuous weak curve. Issue #341 evaluates an additional whole-file detector using digitalis-p/doppler-locate. The tested reference revision is ed84e0f, under its MIT license. This optional research script requires that package and its dependencies in the executing Python environment; the default DOPPLERGEIST installation is unchanged. Python 3.12 was used for the reference detector.

From this checkout, with an installed reference package:

PYTHONPATH=src OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 \
  python scripts/discover_coherent_audio.py original.mp3 \
  --output coherent-events.json --workers 2

Alternatively, add the reference checkout's src to PYTHONPATH. The script processes the entire original recording with the reference template bank. It captures actual measured ridge samples, skips unused straight-pass profile-range calculations, and exports the existing detected-event schema. It reuses the reference's 40-point and 1% relative-fit-residual requirements. It removes the later aircraft/close classification cutoffs at 30 m/s and 150 m. Duplicate and harmonic ridges are grouped; their count is not an independent aircraft count.

The reference template bank and refinement limits remain unchanged. In _refine_all, accepted fits require 6 < speed < 290 m/s and 35 < slant distance < 30000 m. The fitted center must stay within max(10 seconds, half-window / 2) of the template center, and the fit residual must not exceed 1.5 × harmonic order Hz. Tracing and fitting can also fail. Broad downstream matching priors cannot recover a ridge rejected by these upstream limits. This experimental script does not remove every speed or distance restriction requested in issue #341.

These are acoustic proposals. They can include ground traffic, and short curves can fit several unrelated archived flights. Neither a low residual nor ADS-B presence establishes audibility. Use the historical physical matcher and shared timing; never present these outputs alone as locations. This script does not change analyze, its discovery defaults, or existing job checkpoints.

The regression fixture contains only measured frequencies/times derived from XC1092456 by Valerie Heemstra, with the public recording link retained for attribution. It contains no audio, coordinates or flight IDs. The original recording is CC BY-NC-SA 4.0; see the source page for the recordist attribution.

Fixed real-clip development and held-out check

From the existing seeded AeroSonicDB sample, take the first two aircraft-labeled and first two background-labeled clips per publisher location for development, then the next two per class/location as a held-out set. No labels, flight IDs or locations enter the detector. The complete case outcomes are in the 24-row table. The table was committed at revision 69b11cf4, before the supported-overlap correction. That correction can change harmonic grouping, but not accepted traces or whether a clip has proposals. The recorded rows remain unchanged.

Partition and publisher label Clips Ordinary discovery Coherent adapter Union
Development aircraft 6 2 3 3
Held-out aircraft 6 3 3 3
Development background (duration-ineligible) 6 0 0 0
Held-out background (duration-ineligible) 6 0 0 0

Counts mean clips with proposals, not confirmed aircraft or independent events. The initial unrestricted reference-trace experiment proposed on 5/6 development aircraft clips and 1/6 backgrounds; the existing quality rules discard short or noisy traces. The final adapter adds one development clip and no held-out clip. All background clips are 10 s, while aircraft clips are 60 s. With a 1 s FFT and 0.25 s hop, 10 s supplies at most 37 frames, below the 40-point gate. The ordinary backend also requires 12 s. Thus the background zeros are structurally guaranteed and provide no false-alarm evidence. The duration-eligible scene and forest supplements below are separate checks. These small, source-defined aircraft strata do not establish improved general recall and do not justify changing normal upload defaults. See the AeroSonicDB source and sampling report.

All twelve development replays preserved the reference detector's measured traces exactly while omitting profile ranges. Summed wall time was 185.72 s with reference profiles and 172.24 s for the adapter, but the median paired ratio was 1.031. These single passes had different system load; there is no consistent measured speedup claim. Neither this clip test nor the Michigan regression establishes a location.

The standalone CLI also processed the entire 819.4 s XC891053 Kane County recording in 146.09 s, exporting seven proposal groups from nine refined traces. The exact location reference remained withheld. This checks actual whole-file script execution, not seven aircraft identities, a localization, or total search time.

Duration-eligible real scene controls

Four complete, approximately three-minute field recordings from TUT Sound Events 2016 development were fixed before running either detector: first two sorted files per home/residential scene. The original ZIP entries were downloaded and CRC-verified; recordings were not cut, padded, repeated or concatenated.

Source recording Duration Ordinary groups Coherent groups
home/a030 181.557 s 0 0
home/a031 180.663 s 0 0
residential_area/a001 183.270 s 0 1
residential_area/a002 184.105 s 0 0

The source annotates household sounds, traffic, birds and people. These are scene controls, not proof of exhaustive aircraft absence. The proposal in the traffic-annotated residential recording spans 170.25–182 s (48 measured points); its physical source identity remains unresolved. Thus fit quality alone is not an aircraft classifier. No threshold was retuned to remove this outcome. All four results are retained. This check addresses duration eligibility, and still does not justify normal upload integration or a population false-alarm estimate.

Duration-eligible Risoux controls

A separately predeclared supplement used twelve complete Risoux forest originals, three each labelled plane=0 with only silence, biophony, wind or rain. It excluded all 140 previously selected Risoux recordings. The application was pinned to cff7d164; the reference and scientific defaults remained unchanged.

All 24 ordinary/coherent arms completed without error or timeout. Every coherent input had 236 actual spectral frames, sufficient for the 40-point rule to be possible. Both methods reported zero groups on all twelve recordings. Three reference-refined traces reached the adapter with 38, 26 and 23 inliers; all failed the unchanged 40-point requirement, and two also exceeded 1% residual.

The complete fixed-cohort report and actual trace/filter outputs preserve all cases, the frozen protocol/driver, source identities and resource measurements. These are proposal counts on publisher-labelled forest controls, not proof of aircraft absence or a population false-positive rate. The result does not remove the reference's speed/distance limits, establish improved recall, or justify changing normal upload defaults. The earlier AeroSonic and TUT rows remain unchanged.