Description

SWNA removes swell noise – the strong, low-frequency energy caused by sea-surface swell and current action on the receivers – from a gather. It scans the gather with a sliding time window and, at each frequency, statistically compares the spectral amplitude across traces to identify traces that stand out as abnormally noisy outliers rather than following the background amplitude distribution. Locations that are consistently flagged as noisy across the analysis windows are then attenuated, either by rescaling them toward the local background level or by replacing them with a spatial prediction from their neighbors, while genuine signal is left untouched. The module also provides "Before attenuation" and "After attenuation" sub-flow slots so that additional processing steps can be inserted immediately before or after the core noise attenuation, and it exposes QC displays (window spectra and signal masks) in its view panels to help verify the noise detection.

Input data

Seismic gather

The gather containing swell noise to be attenuated. Works with any gather type; results are generally best when traces within the gather are laterally ordered (for example by receiver or offset), since the algorithm compares neighboring traces to detect and repair noise.

Parameters

Time window

Length of the moving time window used to compute the spectrum and detect noise as the analysis slides down each trace. Shorter windows localize noise bursts that are limited in time more precisely but give a coarser frequency resolution and a noisier statistical estimate per window; longer windows give better frequency resolution and more stable statistics but average over more time, which can blur short noise bursts.

Default: 0.5 ms (minimum 0.001 ms).

Taper num traces

Number of extra, amplitude-tapered edge traces temporarily added to both sides of the gather before analysis. This padding reduces edge artifacts that would otherwise appear near the first and last traces of the gather. Increase it if you notice attenuation artifacts near gather edges; decrease it to reduce processing overhead.

Default: 20 traces (minimum 1).

Start time

Time, within the trace, at which the first analysis window begins. Set this above zero to skip noise detection and attenuation in the earliest part of the trace, for example to avoid the direct arrival or a shallow high-amplitude zone that should not be touched.

Default: 0 ms (minimum 0 ms).

Window step

Time step between the start of successive analysis windows as they slide down the trace. A step smaller than Time window makes windows overlap, giving smoother, more consistent detection results from window to window; a step equal to Time window uses non-overlapping windows and runs faster.

Default: 0.5 ms (minimum 0.001 ms).

Min frequency

Lower edge of the frequency band over which noise is detected and attenuated. Frequencies below this value are left unmodified. Swell noise is typically concentrated at low frequencies, so this can usually be left at or near zero.

Default: 0 Hz (minimum 0 Hz).

Max frequency

Upper edge of the frequency band over which noise is detected and attenuated. Frequencies above this value are left unmodified. Narrow the band toward the low end if swell noise occupies only a limited frequency range and you want to avoid touching higher-frequency signal.

Default: 1000 Hz (minimum 0 Hz).

Noise detection probability threshold

Statistical probability threshold used, at each frequency and time window, to decide whether a trace's spectral amplitude belongs to the noise population rather than the background/signal population. Expressed as a fraction between 0 and 1. Raising it toward 1 makes detection more conservative, only flagging amplitudes that are very confidently anomalous (less risk of damaging real signal, but some noise may be left behind); lowering it toward 0.5 flags more amplitudes as noise, giving more aggressive attenuation at greater risk of removing genuine signal.

Default: 0.5 (range 0.5 to 0.999).

Noise selection threshold

Percentile cutoff applied to the noise-detection results accumulated across all analysis windows, used to decide which trace/frequency locations are ultimately attenuated. Expressed as a fraction between 0 and 1. Raising it requires a location to have been flagged as noisy in a larger proportion of the analysis windows before it is attenuated (more conservative); lowering it attenuates locations flagged less consistently (more aggressive).

Default: 0.95 (range 0 to 1).

Noise attenuation method

Chooses how the trace/frequency values flagged as noise are actually corrected. "Scaling down" rescales the noisy amplitude at each flagged location to match the average amplitude of its non-noisy neighboring traces, suppressing the anomalous energy while keeping the trace's general character. "Prediction filter" instead predicts the value at each flagged location from its neighboring traces using a spatial prediction filter and substitutes the predicted value, similar to an interpolation-based repair. Prediction filter can give a cleaner result on strongly corrupted data but is more computationally demanding.

Default: Scaling down.

Scaling down window

Available when Noise attenuation method is "Scaling down". Number of neighboring traces on either side used to compute the average clean amplitude that a noisy trace is rescaled toward. A larger window averages over more traces for a more stable target level, but assumes the background amplitude is consistent over a wider area; a smaller window adapts to more local variations.

Default: 25 traces (minimum 1).

Prediction filter order

Available when Noise attenuation method is "Prediction filter". Number of neighboring traces on each side used by the spatial prediction filter that estimates noisy trace values from their neighbors. A longer filter can capture more complex spatial trends across traces but costs more compute and needs more clean neighboring traces to be reliable.

Default: 25 (minimum 1).

Number of prediction iterations

Available when Noise attenuation method is "Prediction filter". Number of times the spatial prediction filter is re-estimated and reapplied. More iterations let the filter progressively refine its prediction of the noisy values, at the cost of additional run time.

Default: 5 (minimum 1).

PDF estimation

PDF type

Statistical amplitude-distribution model assumed for the background and noise populations when estimating the noise-detection threshold at each frequency. "Exponential" models the amplitudes as a mixture of two exponential distributions; "Rayleigh" instead assumes two Rayleigh distributions, which is often a better statistical fit for the envelope of random background seismic noise. Try Rayleigh if Exponential produces an unstable or clearly incorrect-looking noise mask.

Default: Exponential.

Number of iterations (PDF estimation)

Maximum number of iterations allowed when statistically estimating the two-population amplitude distribution parameters at each frequency. Higher values give the estimation more chances to converge on difficult data at the cost of run time.

Default: 1000 (minimum 1).

Tolerance

Convergence tolerance that stops the distribution-parameter estimation early once the estimated parameters stop changing by more than this amount between iterations. Loosen it (raise the value) to stop iterating sooner if run time is a concern; tighten it for a more precisely converged estimate.

Default: 1e-6 (minimum 0).

Output data

Seismic gather

The gather with swell noise attenuated, having the same geometry as the input. Trace/frequency locations that were not flagged as noise are passed through unchanged; flagged locations are corrected using the selected Noise attenuation method.

References

Bekara, M. and van der Baan, M., 2010, High-amplitude noise detection by the expectation-maximization algorithm with application to swell-noise attenuation: Geophysics, 75, V39-V49.