Description

DMSSA (Damped Multichannel Singular Spectrum Analysis) attenuates random and incoherent noise while preserving coherent reflection energy. It works in small overlapping space-time windows: within each window the data is transformed to the frequency domain and, at every frequency, a rank-reduction (singular value) decomposition is applied across the spatial traces to separate coherent, low-rank plane-wave energy from incoherent noise. A damping factor smooths the rank reduction so that useful amplitude variations are not truncated as abruptly as with a hard-rank filter. Use this module for general random-noise suppression on 2D lines or 3D volumes where you want to keep true reflection character while reducing background noise.

Input data

Seismic gather

The seismic gather to be denoised. The traces should carry valid inline/crossline (for 3D) or source/receiver line and station (for 2D) geometry, since this geometry is used to build the local analysis windows.

Input data type

Tells the module how the input traces are organized so it can build the correct local analysis windows. "2D common SRC" treats the data as a 2D line where traces within a common shot are ordered along the receiver line. "2D common RCV" treats the data as a 2D line ordered by source position within a common receiver gather. "3D" treats the data as a true 3D volume organized by inline and crossline position. The choice determines whether the Trace window parameter, or the Inline window / Crossline window pair, is used to size the analysis windows.

Default: 3D. Set this to match the actual survey geometry and the domain in which the data is currently sorted.

Parameters

Trace window

Number of adjacent traces along the line used in each local 2D analysis window (only used when Input data type is "2D common SRC" or "2D common RCV"). Windows overlap by 50% along the line. A larger window gives the rank-reduction step more spatial samples to separate coherent dipping events from noise, but increases run time and can blur laterally rapid changes; a smaller window reacts faster to local changes but gives a noisier estimate of the local signal rank.

Default: 40 traces (minimum 1).

Inline window

Number of traces along the inline direction included in each local 3D analysis window (only used when Input data type is "3D"). Works together with Crossline window to define the footprint of the moving analysis patch, which overlaps its neighbors by roughly 50%. Increase it to better resolve events with inline dip at the cost of longer run time; decrease it to preserve fine lateral detail.

Default: 40 traces (minimum 1).

Crossline window

Number of traces along the crossline direction included in each local 3D analysis window (only used when Input data type is "3D"). Same trade-off as Inline window: larger values better resolve crossline dip and improve the rank-reduction statistics, smaller values preserve local detail and reduce computation.

Default: 40 traces (minimum 1).

Time window

Length of the time window processed together, sliding down each trace with 50% overlap between consecutive windows. Shorter windows adapt better to signal content that changes quickly with time (non-stationary data) but are processed more times, adding run time; longer windows assume the frequency content is more stationary over the interval and reduce the number of windows, but may blend together zones with different signal character.

Default: 0.5 ms.

Min frequency

Lower edge of the frequency band over which the rank-reduction denoising is applied. Frequency components below this value are passed through unmodified. Raise this if the very low-frequency content (for example residual low-frequency noise or DC drift) should be left untouched by the filter.

Default: 0 Hz (minimum 0 Hz).

Max frequency

Upper edge of the frequency band over which the rank-reduction denoising is applied. Frequency components above this value are passed through unmodified. Set it to comfortably cover the useful signal bandwidth of the data; restricting the band to where genuine signal is expected avoids spending time filtering frequencies that contain mostly noise or no data, and keeps the filter from acting outside the signal band.

Default: 250 Hz (minimum 0 Hz).

Number of preserved eigen values

Number of dominant singular values (the assumed number of coherent plane-wave events, or "rank") kept in each local window during the decomposition. Lower values give stronger noise suppression but risk stripping out genuine weak or overlapping reflection events; higher values preserve more complex wavefield detail (crossing dips, closely spaced reflectors) but leave more residual noise in the output.

Default: 5 (minimum 1).

Damping factor

Controls how gently the singular values beyond the preserved rank are suppressed, instead of cutting them off abruptly. A larger damping factor behaves closer to a hard rank-truncation filter, producing a stronger but potentially harsher noise cut that can introduce filtering artifacts. A smaller value gives a smoother, more gradual roll-off between preserved and discarded singular values, which reduces artifacts at the cost of leaving a bit more residual noise. Adjust this together with Number of preserved eigen values to balance noise suppression against signal preservation.

Default: 1 (minimum 1e-9). Typical usable values are close to 1; reduce toward the minimum for the gentlest damping, or increase for a sharper rank cutoff.

Output data

Seismic gather

The denoised gather, with the same geometry as the input, produced by reconstructing the overlapping analysis windows (tapered and summed) after rank-reduction filtering in the frequency-space domain.

References

Oropeza, V. and Sacchi, M., 2011, Simultaneous seismic data denoising and reconstruction via multichannel singular spectrum analysis: Geophysics, 76, V25-V32.

Huang, W., Wang, R., Chen, Y., Li, H. and Gan, S., 2016, Damped multichannel singular spectrum analysis for 3D random noise attenuation: Geophysics, 81, V261-V270.