Also known as: signal classification, signal identification, RF classification
Signal classification is the task of deciding what an unknown radio signal is — its modulation, its protocol or waveform family, or simply whether a band is occupied — from the received samples alone, without a decoder that already assumes the answer.1 It sits one level above automatic modulation classification: AMC answers the narrow question “which modulation?”, while signal classification can also mean “is this LTE or Wi-Fi or a radar pulse?” or “is anything here at all?” The unifying idea is mapping raw IQ or a time-frequency image to one of a fixed set of labels.
How it works
There are two broad approaches, and modern systems often combine them.
Feature-based (expert) classification extracts hand-designed statistics that separate the classes and feeds them to a simple decision rule or shallow classifier. Useful features include occupied bandwidth, symbol rate, spectral shape and symmetry, the presence of a carrier or cyclostationary periodicities, envelope and phase moments, and higher-order cumulants. This route is interpretable and cheap, and works well when the candidate set is small and the features are chosen by someone who understands the signals.
Deep-learning classification skips the hand design: a convolutional network learns features directly from raw IQ or from a spectrogram image, trained on a labelled dataset — this is the RFML approach. It scales to many classes and can pick up subtle structure a human might miss, at the cost of needing lots of data, compute, and trust in a model whose reasoning is opaque.
Either way, the pipeline usually starts with detection — an energy detector or spectral search finds that something is present and isolates its band — before the classifier decides what it is. Both stages degrade as SNR falls, and both are only as good as the assumed class list: a signal outside the trained or modelled set will still be forced into one of the known labels, often with unwarranted confidence.
In practice
Signal classification is the front end of spectrum awareness: a monitoring receiver sweeps a band, detects each occupied slice on the waterfall, classifies it, and only then routes it to a matching decoder. It turns “there is energy at 462 MHz” into “that is an FM voice channel,” which is the decision that determines what to do next.
Relevance to SDR
Classification underpins cognitive radio, dynamic spectrum access, interference hunting, signal intelligence, and any survey tool meant to map what is on the air rather than decode one known channel. It is the recognition layer that a pile of deterministic decoders needs in front of it when the input is unknown.
GopherTrunk does not classify signals. It is a decoder for signals whose type is already known and configured — you point it at a P25, DMR, NXDN, or TETRA system and it demodulates and frames that specified mode. There is no learned model and no blind-recognition stage in its chain; the modulation and protocol are inputs, not outputs. In a broader monitoring system, a signal classifier would play the complementary role of deciding which signals are P25 or DMR in the first place and handing them to a decoder like GopherTrunk. That detection-and-classification front end is out of scope for GopherTrunk itself, which stays focused on decoding the modes it explicitly supports.
Sources
-
Statistical classification — Wikipedia, for the general task of assigning inputs to labels. See also Automatic modulation recognition for the closely related RF-specific problem. ↩