Also known as: AMC, automatic modulation classification, automatic modulation recognition, AMR, modulation recognition
Automatic modulation classification (AMC) is the algorithmic problem of deciding which modulation scheme a received signal uses — BPSK, QPSK, 8PSK, 16-QAM, FSK, and so on — from the IQ samples alone, without prior knowledge of the transmitter.1 It is the tightest, most-studied slice of signal classification: the question is not what is on the air but specifically how are its symbols carried. Getting the modulation right is the prerequisite for everything downstream, because a demodulator built for the wrong scheme produces nothing but noise.
How it works
Every AMC method exploits the fact that different modulations leave different statistical fingerprints. A constellation of QPSK sits at four points; 8PSK spreads eight points around a circle; 16-QAM fills a 4×4 grid; an FSK signal has no stable constellation but a bimodal instantaneous frequency. The classifier’s job is to measure enough of that structure — through raw symbols, derived statistics, or a learned representation — to separate the candidates despite noise, timing offset, and unknown carrier phase. Because carrier and timing are usually unknown at classification time, robust AMC relies on quantities that survive those unknowns, such as amplitude/phase moments and rotation-invariant cumulants.
Variants
- Likelihood-based (LB). Compute the probability of the received samples under each candidate modulation’s model and pick the maximum. This is optimal in the Bayesian sense and sets the performance ceiling, but it needs accurate channel and synchronization parameters and is computationally heavy, so it is often a benchmark rather than a deployment.
- Feature-based (FB). Extract discriminating statistics — instantaneous amplitude, phase and frequency moments, spectral symmetry, and especially higher-order cumulants, which take characteristic values for each constellation and are insensitive to phase rotation — then feed them to a threshold rule or a small classifier. Nearly optimal in many cases, cheap, and interpretable.
- Deep-learning. Train a convolutional network on raw IQ (or a spectrogram) so it learns its own features, the approach popularized by the RadioML modulation-recognition work.2 It scales to many classes and low SNR but demands large labelled datasets and offline training — the RF machine learning route.
Performance for all three collapses as SNR drops and as the candidate set grows to include easily confused pairs (16-QAM vs 64-QAM, QPSK vs 8PSK), which is why practical systems narrow the list using known bandwidth and symbol rate before classifying.
Relevance to SDR
AMC is central to cognitive radio, adaptive links that switch modulation with channel quality, spectrum surveillance, and signal intelligence — anywhere a receiver must adapt to a signal it did not schedule. It is a recognition step, not a decode step: its output selects which demodulator to run next.
GopherTrunk does not perform automatic modulation classification. The systems it targets have fixed, standardized modulations — P25 Phase 1 is C4FM/π-4-DQPSK at 4800 symbols/s, DMR is 4-FSK, TETRA is π/4-DQPSK — so the modulation is a known constant of each protocol, not something to infer. GopherTrunk is told which system to decode and applies the matching, deterministic demodulator directly, with no classification stage, no cumulant features, and no neural network. AMC solves the upstream problem of identifying an unknown signal’s modulation, and could feed a decoder like GopherTrunk in a survey pipeline, but it is outside GopherTrunk’s own scope as a decoder of specified modes.
Sources
-
Automatic modulation recognition — Wikipedia, on likelihood-based and feature-based AMC and the role of higher-order cumulants. ↩
-
Convolutional Radio Modulation Recognition Networks — O’Shea, Corgan, Clancy (2016), the CNN-on-IQ approach and RadioML datasets behind deep-learning AMC. ↩