AM FM Handcrafted Features vs. Learned Features in RF Modulation Classification bgilbert1984Download
We quantify the value of classical AM/FM and
spectral moments (e.g., amplitude-modulation index, frequency
deviation, spectral kurtosis/skewness) against learned representations in modern RF ensembles. Using a shared dataset interface,
we train a tree-based classical stack on hand-engineered features
and compare to a learned baseline of identical capacity on the
same samples. We provide (i) SHAP analyses over the classical
stack, (ii) per-family ablations, and (iii) SNR-stratified deltas.
Results show that a small set of physics-aware features recovers
most high-SNR accuracy while the learned model dominates in
low-SNR and burst-impaired regimes.