Improved Spectral Segmentation in Wideband Data Using the Signed Cumulative Distribution Transform

Spectral segmentation, which seeks to detect narrowband modulation energy within wideband data, faces significant challenges in practical scenarios characterized by low signal-to-noise ratios (SNR), complex RF environments, as well as high temporal and frequency variability of embedded signals. Many statistical signal processing algorithms and recent machine learning methods rely on time-frequency analysis of the RF scene's complex baseband representation, for instance via direct spectrogram segmentation. In this paper, we make use of the mathematical properties of the signed cumulative distribution transform (SCDT) to recast the problem in a transform domain and perform spectral segmentation by combining SCDT with conventional classifiers. Extensive numerical experimentation demonstrates that the proposed approach outperforms both tra-ditional and state-of-the-art machine learning classifiers working directly in the time-frequency domain.