Multiple attenuation — g-Platform
Multiples attenuation

Multiple attenuation software

Multiples mask the subsurface image and degrade the stack, in land and marine 2D and 3D data alike. g-Platform carries proprietary techniques optimized for both CPU and GPU clusters — true-azimuth 3D SRME, horizon-driven prediction, high-resolution Radon and shallow-water de-reverberation — so the tool can be chosen to match the mechanism that generated the multiple.

Download free trial How multiples are predicted Watch Demos
Land & marine
2D and 3D, shallow to deep water
Predict or separate
Periodicity and moveout methods
CPU or GPU
Distributed across the cluster
Stack section before multiple attenuation in g-Platform, with multiple energy repeating below the strong reflector
The same stack section after multiple attenuation in g-Platform, with the repeated events removed and the primary reflectors continuous
In view
Before multiple attenuation
There is no single demultiple button, and that is the point. Water-bottom, peg-leg and interbed multiples each have their own character, and the algorithm has to be chosen for the one in front of you: onshore it is usually interbed energy bouncing inside the layers, in shallow water it is ringing, in deep water it is first- and second-order water-bottom multiples. g-Platform splits the problem two ways — predict the multiple and subtract it, using the periodicity of the surface-related bounce, or separate it by moveout, in the tau-p domain where primaries and multiples land in different places. Both routes end in the same place: an adaptive subtraction that matches the model to the data before removing it.
Prediction & subtraction

Model the multiple from the data, then take it away.

Surface-related multiples are predicted by auto-convolving the input data with itself — the recorded wavefield already contains everything the multiple bounce needs. The predicted energy is then adaptively subtracted, leaving the primaries.

Two ways to find a multiple
Periodicity-based methods predict the bounce from the data itself; moveout-based methods separate it by velocity. The generating mechanism decides which one you reach for.
3D SRME on the real geometry
Apertures are set independently along the source line and the receiver line, with a bounding-width ratio keeping the source-receiver combinations inside the area that matters — all combinations for completeness, or the lighter source-receiver-source-receiver scheme.
Near offsets filled in first
Missing near offsets are extrapolated before prediction, per streamer if needed, and the input can be interpolated onto a regular geometry with a first-derivative, Laplacian, F-K or Radon regularization scheme.
Mute above the first arrivals
Anything left above the water bottom is convolved into the model and comes back as a false multiple. The mute is driven by source water depth from the headers, or by a picked horizon, and applies to input and model alike.
Horizon-driven prediction
SRMP predicts water-bottom or peg-leg multiples from a picked horizon rather than from a full auto-convolution — imported, or picked on the reference stack inside horizontal velocity analysis.
Shallow-water de-reverberation
Single-channel shallow-water surveys arrive looking like a stack section and ring with water-bottom multiples. ZESMP models those orders from the picked water bottom on zero-offset data so they can be subtracted.
Test on one shot, not the survey
Pick a shot on the location map, calculate its multiples, and preview the adaptive subtraction in the same window — input, model and subtracted gather toggled one over the other before anything runs in production.
Ghosts as well as multiples
Source and receiver are towed metres below the surface, so every event carries a polarity-reversed ghost that notches the spectrum. Deghost models the ghost wavefield in F-K from the tow depths and adaptively removes it.
Moveout separation

In tau-p, primaries and multiples stop overlapping.

Correct a CMP gather with stacking velocity and the primaries flatten while the multiples stay under-corrected on an approximately parabolic moveout. Transform to intercept time and slowness and the two land in different places, so the multiple energy can be rejected before the inverse transform brings the gather back.

Tau-p transform
Radon - TauP

The workhorse: a de-aliased least-squares decomposition into user-defined parabolas, frequency band by frequency band. Parabolic, Foster-Mosher hyperbolic, linear and absolute-linear transforms are all available, with P and tau taper zones marking out what is saved and what is rejected.

High resolution
Radon - TauP - High resolution

Adds a sparse, re-weighted iterative focusing step after the transform. Energy is pulled towards where it would sit if sampling and aperture were unlimited, which sharpens the separation at near offsets and makes the result resistant to spatial aliasing — at the price of iterations, so the iteration count and tolerance are what you tune.

Sparse solver
HR Radon multiple attenuation

A high-resolution parabolic Radon on offset gathers solved by Gauss-Seidel iteration, with a sparsity weight and a damping factor that trade sharpness against stability. Traces must be in ascending absolute offset, with at least four per gather.

Velocity-guided
Radon multiple attenuation by velocity

Muting driven by a velocity model instead of a raw curvature range: a time-varying table of percentages of the primary velocity draws the boundary, and everything slower is treated as multiple. QC windows show the Radon spectrum and the velocity semblance before and after.

Apparent velocity
Recon

A sliding two-dimensional F-K window that suppresses energy standing out by apparent velocity — low-velocity water-bottom multiples and ground roll in one mode, steeply dipping interbed multiples in the other — and outputs the removed energy as its own multiple model.

Marine ghosts
Deghost

Builds the source and receiver ghost model in the F-K domain from the airgun and streamer tow depths and the water velocity, solving with an L1-norm iteration, then subtracts it. The output is low-frequency-rich, so a Q filter usually follows.

Workflow

Six steps from a ringing gather to a clean primary.

1
Name the multiple
Water bottom, peg leg or interbed; deep water, shallow ringing or an onshore bounce inside the beds. The mechanism decides whether you predict it or separate it by moveout.
2
Precondition the input
Sort to the domain the method wants, mute the direct and refracted arrivals, extrapolate missing near offsets and regularize the geometry where the survey is irregular.
3
Build the model
Auto-convolution over the source and receiver apertures for SRME, a picked water-bottom horizon for SRMP and ZESMP, or a tau-p decomposition where the separation is by moveout.
4
Test interactively
Select shots on the location map and calculate their multiples one at a time, checking the model and a preview subtraction before committing the survey to a production run.
5
Subtract adaptively
Match the model to the data in amplitude, phase and frequency inside sliding windows, then subtract. A second pass in another domain often cleans up the remnant.
6
Prove it on the difference
Read the difference gather and the velocity spectrum before and after: coherent primary energy in the difference means the parameters are too aggressive.
Subtraction & scale

The model is only half the job. Removing it is the other half.

Arithmetic subtraction ignores phase, amplitude and frequency, so a predicted model rarely cancels cleanly. Adaptive subtraction designs a matching filter that adjusts itself to the local character of the data, in time and in space, before the model is taken out.

Adaptive subtraction
How the matching filter is designed and applied
Solvers
SVD, Cholesky decomposition, least squares and FISTA — the last a sparsity-promoting iterative scheme built for large datasets.
Filter in time
Time-variant from a table of time-window pairs, constant across the record, or divided into frequency bands so each band gets its own operator.
Adaptation window
A horizontal window in traces and a vertical window in milliseconds, slid across the gather with an overlap — around half the window — so adjacent windows blend instead of leaving edges.
Shift search
Minimum, maximum and step for both the vertical shift in time and the spatial shift in traces, bounding how far the filter may move the model to fit.
Regularization
A lambda weight balances fitting the data against a smoother, less noise-sensitive estimate of the multiple.
Model variants
The raw predicted model, its Hilbert transform, a differential operator form, or Hilbert and differential combined for the harder cases.
Running it on a survey
Compute, limits and the QC that comes out
CPU or GPU
Prediction runs on either, selected per module, so the same flow moves between a workstation and an accelerated node.
Distributed
Calculation spread across the processing server with a per-machine bulk size in megabytes, thread limits on the nodes and an affinity tag to find your job in the server QC interface.
Calculation area
Restrict a run by sequence number or by inline and crossline range, so a parameter test costs a few lines rather than the whole survey.
Outputs
The multiple model or the subtracted gathers, plus the selected input, the extrapolated version and the pre-processed version saved alongside for comparison.
QC displays
Difference gathers, tau-p panels beside the gather they came from, and velocity semblance before and after the demultiple.
Bad-sample policy
Corrupted samples are fixed, reported and stopped on, or passed through — your choice per module, so a production job does not fail on one gather.
Modules in this group
The demultiple and de-ghost modules named on this page — not the full list
SRME 2D/3D SRMP ZESMP Radon - TauP Radon - TauP - High resolution HR Radon multiple attenuation Radon multiple attenuation by velocity Recon Deghost Adaptive subtraction
More g-Platform capabilities

Demultiple is one link in the processing chain.

The same project, the same batch processing and the same cluster carry a survey from field tape through to depth. These are the other topics in the g-Platform workflow.

g-Platform overview Signal processing & denoise Deconvolution Deghosting, debubble & designature Velocity analysis Stack imaging Time imaging Depth imaging FWI & RTM VSP processing Refraction statics Residual statics First-break picking Geometry assignment & QC Data & geometry formats
FAQ

Questions, answered.

Which demultiple method should I use?
It depends on how the multiple was generated. No single procedure attenuates every kind: water-bottom, peg-leg and interbed multiples all behave differently, so the method is chosen to match the mechanism. Surface-related multiples in marine data are predicted from the data itself with SRME, or from a water-bottom horizon with SRMP; shallow-water ringing on single-channel zero-offset data goes to ZESMP. Where primaries and multiples differ in moveout after NMO, the Radon family separates them in the tau-p domain. Energy that stands out by apparent velocity — low-velocity water-bottom multiples, or steeply dipping interbeds — can be suppressed with a sliding F-K reconstruction instead.
What does SRME need before it will predict a good multiple model?
Shot-sorted input, and a mute of everything above the first arrivals — direct and refracted energy left in the gather is convolved into the model and comes back as a false multiple. The mute itself is driven either by source water depth in the trace headers or by a picked water-bottom horizon. Missing near offsets should be extrapolated first, and the model is more accurate when the geometry is regularized, which the module can do internally with a first-derivative, Laplacian, F-K or Radon scheme.
Conventional Radon or high-resolution Radon?
Conventional parabolic Radon is the faster route and is enough when primaries and multiples are well separated in moveout. It loses quality where the moveout difference is small — at near offsets — and where the data are spatially aliased. The high-resolution transform adds a sparse, re-weighted iterative focusing step in the tau-p domain: fewer, sharper events, better discrimination between primary and multiple, and more tolerance of aliasing. The cost is computation, so the iteration count and the sparsity tolerance are the parameters to watch.
How is a predicted multiple model actually removed?
By adaptive subtraction rather than plain arithmetic. A matching filter is designed inside sliding time and trace windows so that the model is adjusted in amplitude, phase and frequency to the recorded data before it is subtracted. The filter can be solved by SVD, Cholesky decomposition, least squares or FISTA, and the operator can be held constant, made time-variant from a time table, or split across frequency bands. A difference gather shows what was taken out, which is the check that no primary energy went with it.
Get started

Try the demultiple on your own gathers.

Predict, separate and subtract — on land or marine data, in 2D or 3D. Talk to Geomage about a demo, or take g-Platform for a trial run.

Download free trial Request pricing