I/Q Spectrogram

Unlike a single power spectrum which averages frequency content over an entire recording, the spectrogram resolves when specific frequencies occur. This is critical for classifying modulation schemes with time-varying elements.

• Transient Capture: Visualizes short-duration events like frequency hops, preamble sequences, or symbol transitions that a global FFT would obscure.
• Resolution Trade-off: The STFT window size dictates the balance between time resolution and frequency resolution, a fundamental parameter that directly impacts downstream CNN classification accuracy.
• Overlap Processing: Successive FFT frames are typically overlapped (e.g., 50%) to ensure no signal transition is lost at the frame boundary, creating a smooth temporal representation.

Each modulation format produces a distinct visual fingerprint in the spectrogram domain, governed by its symbol rate, pulse shaping filter, and frequency deviation.

• Continuous Wave (CW): Appears as a single, constant horizontal line.
• Frequency Shift Keying (FSK): Manifests as discrete horizontal lines that switch abruptly between frequencies, creating a 'piano roll' pattern.
• Linear Frequency Modulation (LFM): Produces a characteristic diagonal 'chirp' line sweeping across the frequency band.
• Phase Shift Keying (PSK): Exhibits a constant frequency carrier with spectral broadening occurring at symbol transitions, visible as vertical 'bursts' of bandwidth.

The spectrogram converts a 1D time-series signal processing problem into a 2D image recognition task, unlocking the powerful spatial feature extraction capabilities of Convolutional Neural Networks.

• Spatial Hierarchies: CNNs can learn low-level features (edges, lines corresponding to tones) and compose them into high-level concepts (shapes corresponding to modulation patterns).
• Translation Invariance: Pooling layers provide robustness to slight shifts in time or frequency, accommodating minor synchronization errors.
• Transfer Learning: Pre-trained weights from massive image datasets (e.g., ImageNet) can be fine-tuned on spectrogram data, dramatically reducing the number of required labeled RF examples.

Real-world propagation effects leave distinct, visible artifacts in the spectrogram that a robust classifier must learn to see through or exploit.

• Multipath Fading: Appears as frequency-selective nulls—horizontal dark bands where destructive interference cancels the signal at specific frequencies.
• Co-Channel Interference: Manifests as a superimposed, structurally different spectral pattern overlapping the signal of interest.
• Impulsive Noise: Visible as short-duration, broadband vertical streaks that saturate the spectrum across all frequencies simultaneously.
• Carrier Frequency Offset (CFO): Shifts the entire spectral signature up or down by a constant frequency, requiring the CNN to learn frequency invariance.

The spectrogram's magnitude squared values represent the Power Spectral Density (PSD) within each time-frequency bin, providing a visual map of energy distribution.

• Dynamic Range: Spectrograms are typically displayed in decibels (dB) using a logarithmic scale to compress the large dynamic range of RF signals, making weak spectral components visible alongside strong carriers.
• Colormap Engineering: The choice of colormap (e.g., 'turbo', 'magma') for rendering is a critical preprocessing step, as it defines the input pixel intensity distribution the CNN learns from.
• Noise Floor Estimation: The granular texture of the background noise floor provides implicit SNR information that the network can use to gauge classification confidence.

A single spectrogram with a fixed STFT window may not optimally represent all signal features. Multi-resolution techniques combine information from multiple window sizes.

• Constant-Q Transform (CQT): Uses logarithmically spaced frequency bins, providing higher frequency resolution at low frequencies and higher time resolution at high frequencies, mimicking human auditory perception.
• Wavelet Scalograms: An alternative to the STFT that provides a multi-scale time-frequency tiling, often yielding sharper transient localization than fixed-window spectrograms.
• Multi-Channel Input: Stacking spectrograms computed with different window lengths as separate input channels to a CNN allows the network to jointly learn from coarse and fine time-frequency representations.