Spectrogram Augmentation

A time-domain augmentation that randomly masks out contiguous horizontal blocks of time steps in a spectrogram. This forces the model to rely on partial temporal information rather than memorizing entire sequences.

• Mechanism: A random contiguous segment of t time frames is zeroed out or replaced with mean noise
• Parameter: Maximum mask width is typically set to a percentage of total time steps (e.g., 10-15%)
• RF Application: Simulates short-duration interference or transient dropouts common in bursty wireless channels
• Effect: Improves robustness to intermittent jamming and packet collisions

Originally introduced in SpecAugment for speech recognition, this technique directly transfers to RF spectrograms where temporal continuity of transmissions is a key feature.

A spectral augmentation that masks out contiguous horizontal blocks of frequency bins in a spectrogram. This teaches the model to classify signals even when portions of the spectrum are occluded.

• Mechanism: A random contiguous band of f frequency channels is zeroed out
• Parameter: Maximum mask width typically set to 10-20% of total frequency bins
• RF Application: Simulates narrowband interference, partial band jamming, or frequency-selective fading
• Effect: Encourages the network to learn frequency-invariant features and prevents overfitting to specific spectral peaks

Critical for cognitive radio systems operating in congested spectrum where co-channel interference is unpredictable.

A non-linear temporal deformation applied along the time axis of a spectrogram using a random warp function. This simulates variable transmission timing without altering spectral content.

• Mechanism: A random point along the time axis is chosen and the spectrogram is warped by a factor w around that point via interpolation
• Parameter: Warp factor typically bounded between 0.9 and 1.1 (logarithmic scale)
• RF Application: Simulates clock drift, Doppler compression/dilation, and variable symbol rates in real transmitters
• Effect: Improves resilience to hardware oscillator imperfections and motion-induced time dilation

Unlike simple time stretching, warping applies localized deformation that better approximates real-world timing jitter in low-cost RF front-ends.

The combined application of time masking, frequency masking, and time warping in a structured augmentation policy. This composite approach, originally from Google Brain's speech research, has become the de facto standard for spectrogram augmentation.

• Policy Structure: Sequential application of N time masks, M frequency masks, and optional time warping
• Typical Configuration: 2 time masks (max width 10%), 2 frequency masks (max width 10%), 1 time warp
• RF Adaptation: Mask parameters are tuned to match expected channel coherence time and coherence bandwidth
• Effect: Synergistic regularization that prevents co-adaptation of features across both dimensions

Modern RF implementations often add adaptive scheduling where augmentation intensity increases as training progresses, preventing early underfitting while maximizing final robustness.

A data interpolation technique that creates new training examples by taking convex combinations of pairs of spectrograms and their corresponding labels. This enforces linear behavior between classes in the feature space.

• Mechanism: x_new = λ * x_i + (1-λ) * x_j and y_new = λ * y_i + (1-λ) * y_j, where λ is sampled from a Beta distribution
• Parameter: Beta distribution alpha typically set between 0.2 and 0.4 for RF applications
• RF Application: Simulates signal superposition from overlapping transmissions in dense spectral environments
• Effect: Smooths decision boundaries and reduces adversarial vulnerability in modulation classification tasks

Particularly effective for automatic modulation classification where signals may partially overlap in time and frequency, creating ambiguous boundary cases.

A regional replacement strategy that cuts a rectangular patch from one spectrogram and pastes it onto another, with labels mixed proportionally to the patch area. Unlike Mixup, this preserves localized signal structure.

• Mechanism: A random bounding box is removed from spectrogram A and filled with the corresponding patch from spectrogram B
• Parameter: Patch area sampled uniformly, typically covering 25-50% of the spectrogram
• RF Application: Simulates burst interference where a foreign transmission suddenly appears within an ongoing signal
• Effect: Improves spatial attention and forces the model to identify signals from partial, corrupted views

More aggressive than time/frequency masking because it replaces content rather than zeroing it out, creating more challenging and realistic interference patterns.

A frequency-domain dropout technique that masks consecutive time steps in a spectrogram.

• Randomly zeros out a block of t consecutive time frames
• Forces the model to rely on non-contiguous temporal context
• Prevents overfitting to transient noise bursts
• Parameterized by mask count T and maximum mask width τ

Example: Masking 10-15 time frames in a 128-frame spectrogram simulates short-duration interference or fading dips.

A spectral occlusion technique that masks contiguous frequency bins in a spectrogram.

• Zeros out a block of f consecutive frequency channels
• Encourages the model to identify signals using partial spectral information
• Simulates narrowband interference or notched filtering
• Parameterized by mask count F and maximum mask height f

Example: Masking 20-30 mel-frequency bins forces a classifier to recognize modulation patterns from surviving sub-bands only.

A non-linear temporal deformation applied along the time axis of a spectrogram.

• Stretches or compresses random segments via a piecewise linear warp function
• Simulates variable symbol rates and Doppler-induced time compression/dilation
• Controlled by a maximum warp factor W along the time dimension
• Preserves frequency content while distorting temporal structure

Example: A warp factor of 0.1 applied to a 256-frame spectrogram shifts time boundaries by up to 25 frames, mimicking oscillator drift.

A convex interpolation technique that blends two spectrograms and their labels.

• Creates virtual training examples: x̃ = λxᵢ + (1-λ)xⱼ
• Labels are blended with the same λ sampled from a Beta distribution
• Encourages linear behavior between classes in the feature space
• Reduces adversarial vulnerability and improves calibration

Example: Blending a QPSK spectrogram (λ=0.7) with a 16-QAM spectrogram (λ=0.3) produces a hybrid signal that smooths decision boundaries.

The composite augmentation policy combining time masking, frequency masking, and time warping, originally developed for speech recognition and adapted for RF.

• Applied on-the-fly during training, not as a preprocessing step
• Policy: apply 2 frequency masks (max width 27 bins) and 2 time masks (max width 40 frames)
• Eliminates the need for explicit noise injection or channel simulation in some cases
• Demonstrated to reduce word error rate in speech and improve modulation recognition accuracy in RF by 5-15%

A regional replacement strategy that cuts a rectangular patch from one spectrogram and pastes it onto another.

• The label is mixed proportionally to the area ratio of the pasted patch
• Forces the model to attend to the full spatial extent of features
• More aggressive than Mixup; preserves localized signal structure
• Effective for recognizing signals in the presence of burst interference

Example: Pasting a BPSK burst into a noise-only spectrogram trains the model to detect short-duration transmissions.