Public Satellite Recordings Are a Free Test Bench for RF Models
Teams building machine learning for radio frequency work tend to stall at the same point: they have no data. Government sensor collections are restricted, vendor captures come with licenses, and synthetic signals make every model look better than it is. A new Satellite Hacking Handbook from RDN Security, dated 23 October 2026, suggests a way around that. It is mainly a learning guide to receiving and decoding spacecraft signals, but if you read it as a map of data sources, it shows how much the amateur community has already captured, labeled and published.
What is already out there
The handbook follows a chain that starts with a raw IQ recording and ends at the ground segment, and it names open tooling for each step. For anyone assembling a test set, three of those resources stand out.
Daniel Estévez keeps a repository of amateur satellite recordings that exists to test decoders. It holds roughly a hundred short 48 kHz mono WAV files, each containing one or a few packets that decode correctly, under an Unlicense. Every file ships with its answer, which is the packets inside it.
To get that answer out, there is gr-satellites, a GNU Radio module its documentation describes as a collection of telemetry decoders for amateur satellites. It accepts WAV, raw IQ and SigMF input, and the docs show the command for running it on those sample recordings. A detector or classifier can then be scored against what a trusted decoder recovers from the same file.
For your own captures, SigMF describes a recording with a JSON file kept next to the samples. It records sample rate, center frequency, capture time, hardware, license and annotations that mark where a signal sits in time and frequency, and its stated aim is to stop captures from losing the details that make them reusable.
The SatNOGS database adds slower, wider coverage: decoded telemetry frames from a volunteer network of ground stations, with frame counts per station and per satellite on its front page. We did not check its license terms, so confirm them before building anything commercial on it.
The Tianwen-1 case
The handbook’s most instructive example is China’s Mars probe Tianwen-1. Estévez’s July 2020 write-up says the probe launched on 23 July 2020 and that amateur Paul Marsh recorded its X-band downlink within two hours. No orbit data had been published, so the amateurs worked from the telemetry itself. A table of six-value vectors in the frames turned out to be position and velocity, spaced 32 seconds apart. Loaded into mission-analysis software, the vectors matched the antenna pointing that amateurs were using to track the spacecraft.
That is ground truth from an unexpected place. The recording is the input, and the pointing data from independent stations is the label. Nobody had to agree on anything beforehand.
The handbook adds a lesson about decoding. It says the telemetry used a convolutional code plus Reed-Solomon, and that this Reed-Solomon variant does not use the dual-basis representation the CCSDS standard assumes, so a standard CCSDS decoder fails until that option is switched off. The signal was fine. A reasonable default was wrong.
Estévez also reported an unresolved discrepancy. A later set of vectors differed from the amateur pointing data by about five degrees in right ascension, and he could not explain the gap. Public data comes with its own errors, and you have to budget time for finding them.
Why this helps a small team
What follows is our reading of those sources, not something they state.
A decoded packet is a label. A recording with a successful decode is a signal-present example with the answer attached, and labels like that are expensive to produce in this field.
The same files test your pipeline as well as your model. The Tianwen-1 failure was a wrong assumption in a decoding chain, and a classifier trained downstream of that chain would have inherited it without complaint. Running your detector and a reference decoder on one file shows where they disagree, and the disagreements are where the engineering work is.
Metadata is half the dataset. A recording with no sample rate, center frequency or capture time is close to unusable. If you adopt SigMF for your own captures, they will still make sense in two years and can go to a teammate or customer without a phone call.
Limits
Amateur recordings skew toward cooperative, well-documented signals from satellites that people enjoy receiving. That is a poor match for crowded, contested or deliberately hidden emitters. The sample files are short, so they exercise decoding and detection but say nothing about long-run behavior such as drift or intermittent interference. A model that does well here has shown it handles clean public signals, and it still needs data from the environment where it will run.
The handbook’s own ground rules apply too. It says receiving is one thing and using the content is another, that transmitting needs a license, and that attacks should be practiced in your own lab, in capture-the-flag events or on published recordings.
A starting point
Pick five recordings from the repository, run gr-satellites on each and save the decoded frames. Run your own detector on the same files and compare the results. Then wrap your first live capture in SigMF from day one. It takes an afternoon, and it gives you a regression test for every model you train afterward.