Skip to content

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

3.2 KB
Raw Blame

Real-data recording clock and benchmark units

The Zenodo 23424278 record supplies 40 WAV recordings, annotated frequency/UTC pairs and flight tracks. Those UTC annotations are useful evidence, but they do not by themselves identify WAV sample zero. This audit retains the dataset's feature-benchmark scope; it does not promote it to an independent 100 m location benchmark.

The audit downloaded the all-case summary CSV, both analysis PDFs, the author's extraction HTML and the smallest four-case ZIP directly from the record. Every asset matches the upstream MD5, and local SHA256 values are recorded in the machine-readable receipt. No third-party audio or PDF is added to Git by this audit.

Case Annotated span WAV duration Difference
N456CS_R38 109.226 s 119.226667 s 10.000667 s
N60RS_R11 82.090 s 92.090667 s 10.000667 s
N748SP_R36 58.368 s 68.368021 s 10.000021 s
N814SR_R07 54.242 s 64.241833 s 9.999833 s

For all four cases, the summary's received_start_utc equals the first annotated frequency timestamp. The WAVs extend approximately ten seconds beyond the annotation span. That does not establish how the extra duration is split before and after the annotations. A five-second leading margin is a possible hypothesis, not verified timing metadata.

ffprobe -show_format -show_streams finds no format or stream tags in the sampled N814SR_R07 WAV. Its audio is mono 48 kHz PCM16. The tool supports manually entered recording timing and timing parsed from original filenames; the distributed WAV is named audio.wav. The ZIP's four KMLs contain flight tracks and no recorder Point, LookAt or Camera. The combined atlas presents the phone at the origin of a relative ground-track diagram. Text extraction of both PDFs and visual inspection of the atlas's N814SR_R07 page did not establish an absolute recorder coordinate or WAV start-time mapping. The remaining three raw ZIPs were not inspected here.

Independent location acceptance still needs a documented mapping from audio samples to UTC, an independently established recorder coordinate with uncertainty, and sufficient flight-data coverage. Annotation-based timing alignment or a reference reconstructed from flight geometry must retain its own evidence label.

Datetime-unit regression

At checkpoint 472a668, benchmark_real_features.py parsed UTC strings, converted their integer representation to numbers, and divided by 1e9. Pandas 3.1.0 parses these strings as microseconds. The 0.170 s interval between the first two N748SP_R36 annotations consequently became 0.00017000013 s, corrupting the annotation-fit diagnostics. This did not change the sound waveform or constitute a localization result.

Current main already fixes that conversion in b0d9cd1 by using the existing unix_time helper. This follow-up preserves the fix and adds six checks: identical SI-second intervals at second/millisecond/microsecond/nanosecond datetime resolution, the actual millisecond annotation example, and equivalent absolute times expressed with different timezone offsets. The checks allow floating-point Unix-second rounding below 0.5 microseconds; they do not claim nanosecond precision.