You are on page 1of 5

ENHANCED SOURCE SEPARATION FOR PHASE-SEQUENCED

SIMULTANEOUS SOURCE MARINE VIBRATOR DATA


A. Nath1, A. JafarGandomi1, S. Grion1
1
Shearwater Geoservices

Summary
We present an approach to separate signal from blending noise, in particular the strong residual noise
from subsequent sources for the simultaneous source data acquired by marine vibrators. The proposed
approach assumes that the simultaneous source data are acquired with phase encoding. Phase encoding
provides the opportunity to perfectly separate sources and the respective residual shot noise up to a
certain interfering frequency. Beyond this frequency more elaborate methods for separation of signal
and blending noise is required. While the amplitude level of the blending noise is similar or weaker than
the signal, approaches such as matching pursuit can successfully remove the noise. However, when the
blending noise is significantly stronger than the underlying signal, strong artefacts are observed. We
propose to model a clean realisation of such strong noise and feed this into a separation flow. We
demonstrate significant improvements achieved by the proposed approach on synthetic data.

83rd EAGE Annual Conference & Exhibition


Enhanced source separation for phase-sequenced simultaneous source marine vibrator data

Introduction

In a conventional airgun survey, one of the factors that determine the shot separation is the time when
residual shot noise (RSN) amplitude decays sufficiently (Landrø 2008). For bigger sources, this delays
the survey since consecutive shots are restricted by RSN. Time that is lost in waiting for RSN to die
down can be gained with simultaneous source acquisition; this however comes with its own set of
separation problems. Many efforts have been made towards making simultaneous source acquisition
practical (Grion et. al., 2018; JafarGandomi et al., 2020).

Marine vibrators offer a higher level of control over the kind of signal that is emitted, one of them being
phase modulation. Laws et al. (2016) demonstrate via examples how phase modulation in successive
sweep generations helps in the RSN separation despite a short time duration between the consecutive
sweeps. After phase deconvolution the RSN neatly ends up in ‘the empty quarter’ in the frequency-
wavenumber (f-k) domain. In the illustrated examples RSN could easily be filtered out using f-k filter
up to the frequency where the signal cones in the f-k domain corresponding to each source are not
interfering. Beyond this frequency more elaborate de-blending methods need to be employed.

The RSN from the previous source actuation (N-1) is usually weaker than the signal and removing such
blending noise using phase modulation has been discussed in detail in previous work (Laws et al. 2019).
However, problems arise because of the RSN from the next source actuation (N+1). The high amplitude
RSN in a common receiver gather (CRG) is not so easy to remove via f-k filtering or even via more
sophisticated de-blending methods. Additionally, designing an elaborate phase sequencing is non-trivial
especially if you have multiple RSN sets in each gather.

In this paper we suggest a solution for the removal of the high amplitude RSN coming from the N+1
source actuation. First, we define a generalised phase encoding for marine vibrators which simplifies
the phase sequencing for elaborate scenarios. We then employ a multidimensional separation (MDS)
method that reconstructs the wavefield and separates the signal from RSN. This method is aided by a
pre-processing step that tactfully addresses the removal of high amplitude RSN from the subsequent
sweep through a data-driven modelling approach.

Generalised phase sequencing

Phase sequencing, which is essential for the effective and accurate separation of simultaneous sources
acquired with the marine vibrator, consists of applying a series of phase rotations to the emitted sweeps.
The sources may operate simultaneously or consecutively one after another. Applying such phase
sequence results in phase encoding of sources within a common receiver gather. Once such a gather is
transformed to an alternative domain such as f-k, the signal from each encoded source occupies different
parts of the f-k spectrum, which consequently makes the source separation easier. It is possible to design
a sequence to optimally place signals from several sources. Laws et al. (2016) showed that a common-
receiver gather with phase gap of 𝜃 degrees between each consecutive trace lies in a f-k triangle centred
𝜃
at 360 × 𝑘𝑛𝑦𝑞𝑢𝑖𝑠𝑡 . Here we propose a generalized form of the phase sequence equation,
𝜃 360
∅(𝑖, 𝑛) = − [𝑛𝑥 + 𝑛(𝑛 + 1) ] + [𝑛(𝑖 − 1) ]; (1)
2 𝑚
where ∅(𝑖, 𝑛) is the output phase for a sweep with sweep-location index 𝑛; {𝑛 ∈ ℤ, 𝑛 ≥ 0} and source
number index 𝑖; {𝑖 ∈ ℕ} within a group of simultaneous sources. By sweep-location we are here
referring to the spatial location to which we deconvolve and source motion correct the recorded data. m
is the total number of simultaneous sources within that source group. θ is the desired phase gap between
traces of the previous RSN after phase deconvolution. For an optimum separation between the 𝑝 − 1
different RSN and the signal from each of the 𝑚 simultaneous sources we can write 𝜃 = 360/(𝑝 ∗ 𝑚).
𝑥 is the phase of the RSN from the previous sweep for ∅(1,0). If the output phase repeats itself every 𝑘
traces, where 𝑘 = 𝑞(360/𝑔𝑐𝑑(𝜃, 360))∀ q ∈ ℕ and 𝑔𝑐𝑑 refers to the greatest common divisor, then
a proposed value for 𝑥 is −(𝑘 + 1)𝜃⁄2. The choice of 𝑥 can determine the length of a phase sequence

83rd EAGE Annual Conference & Exhibition


(k) before it starts to repeat itself; this may have practical benefits in avoiding long phase sequences, as
other choices of 𝑥 may lead to longer sequences. The right-hand-side of equation (1) consists of two
terms. The first term in the brackets describes the phase encoding for a series of sources actuated one
after another with a certain time gap caused by source location spacing. The second term describes the
phase encoding for simultaneous sources without significant time delay. If the source activation time
between successive sweep-locations is larger than the maximum record length of interest, the first term
can be ignored. For example, if a dataset is acquired with one source (𝑚 = 1) and it is desirable to
remove two orders of RSN in a CRG (𝑝 = 3), then 𝜃 = 120o . We choose 𝑘 = 3, hence 𝑥 = −240O.
Using equation (1) we get the following output phase sequence:
∅ = [0°, 120°, 120°, 0°, 120°, 120°, 0°, 120°, 120°, … ].
This phase sequence is used to generate the synthetic example in the next section.

Source-separation and removal of residual shot noise

The first step in de-blending phase-encoded data is deconvolving the phase-rotation by applying a phase
−∅(𝑖, 𝑛) to the recorded data. We demonstrate the impact of phase encoding and subsequent separation
on a synthetic dataset created for an ocean bottom node (OBN) acquisition setup using the SEAM
(Society of Exploration Geophysicists (SEG) Advanced Modelling, www.seg.org/seam) phase 1 model.
A part of the SEAM model away from the main salt bodies was selected, and the receiver is just above
the seabed at a depth of 1900m. The finite-difference generated impulsive source data were generated
up to a maximum frequency of 64Hz. The source spacing is 10m and source depth is 10m. The 10m
spacing of impulsive source data provides us with unaliased input data. In the context of marine vibrator
data processing, this synthetic data is assumed to be post sweep deconvolution and source motion
correction.

Figure 1a shows blended synthetic OBN data gathered with phase encoding. Signals and blending noise
are reasonably separable in the f-k domain as can be seen in Figure 1b. Figure 1c and 1d show two
shallow and deep close-ups of Figure 1a. Figures 1e and 1f show the data in 1c and 1d after f-k filtering
(the filter is highlighted in Figure 1b). In the absence of spatial aliasing, an f-k fan-filter can be applied
to separate the signal perfectly up to an interfering frequency determined by the employed phase
sequence and the number of sources. Note that the effect of phase sequencing as observed in the f-k
spectrum is the same as spatial aliasing. In the presence of spatial aliasing, further interference of the f-
k spectrum of different sources is expected. Beyond the interference frequency, fan-filtering will no
longer be able to fully separate the sources and hence alternative methods must be used. Here we use a
MDS approach that employs a matching pursuit method (e.g., Mallet et. al., 1993) in the frequency-
slowness domain.

The MDS approach has been used for apparition de-blending (Casasanta et. al., 2021), which employs
time or amplitude encoding. Here we extend this approach to include phase encoding with marine
vibrator data. Instead of filtering the undesirable noise from the data we try to reconstruct the signal
iteratively using multidimensional atoms 𝒃 in slowness-frequency (𝒑 − 𝑓) domain. The atom
𝒃𝑺𝒊 (𝒑, 𝑥𝑛 , 𝑦𝑚 , 𝑓) to model the source 𝑺𝒊 , for a given ray parameter 𝒑, sensor at 𝑥𝑛 , 𝑦𝑚 and frequency 𝑓
is,
𝒃𝑺𝒊 (𝒑, 𝑥𝑛 , 𝑦𝑚 , 𝑓) = 𝑒 −𝑗2𝜋𝑓𝒖(𝒑,𝑥𝑛 −𝑥0 ,𝑦𝑚−𝑦0 ) ∙ 𝑒 −𝑗∅𝑛 ; (2)
2
where j = -1, 𝒖(𝒑, 𝑥𝑛 − 𝑥0 , 𝑦𝑚 − 𝑦0 ) is the phase function for ray parameter 𝒑, and relative offset
between 𝑥𝑛 , 𝑦𝑚 and reference sensor at 𝑥0 , 𝑦0 . The phase function can be linear, parabolic or
hyperbolic, but for our application it is a linear function. The reconstruction problem is thereby solved
using an L1 approach:
min‖𝒎‖1 , subject to ‖𝒃𝒎 − 𝒅‖2 ≤ 𝜀, (3)
where 𝒎 is the modelled data for each source, 𝒅 is the recorded data and 𝜺 is the noise level. If 𝒃′𝑺𝒊 is
the phase deconvolved atom for source 𝑺𝒊 , separated data is given by forward modelling 𝒃′𝑺𝒊 𝒎𝑺𝒊 = 𝒅𝑺𝒊 .
Figure 2a shows the desired output (reference gather) in the lower half of the gather (see Figure 1d)
where the high amplitude RSN from the next sweep exists and Figure 2d is its corresponding f-k
transform. Figure 2b shows the MDS result on the same section of data. It can also be seen in its f-k
transform (figure 2e) that the matching pursuit method struggles to reconstruct the data in the vicinity

83rd EAGE Annual Conference & Exhibition


of the high-amplitude RSN. The RSN from the N+1 sweep has very high amplitude, especially around
the direct arrival and hence we see some leakage when using the matching pursuit algorithm in MDS.
The locally stronger events take dominance and produce artifacts. Using a moving block within the data
cube for MDS reduces the impact of locally stronger events and justifies the use of a linear radon kernel.
However, even with these conditions in place we still observe artifacts and residual noise around the
direct arrival event from the N+1 sweep (Figure 2b). This is in principle a signal-to-noise ratio problem:
removing weak noise in presence of strong signal is easier than removing strong noise in presence of
weak signal.

Figure 1 (a)Phase encoded impulsive source data after phase deconvolution and source motion
correction. (b)Frequency wavenumber (f-k) plot of data in a, the area within the dashed triangle
depicts the signal left after f-k fan-filtering. (c) and (d) shallow and deep close-up sections
highlighted in figure a. (e) and (f) are c and d after f-k fan-filtering.

To effectively remove the RSN from the N+1 sweep we suggest a pre-processing method to specifically
target the high-amplitude RSN. Once we model and remove the high amplitude RSN from the gather,
MDS can be applied to the pre-conditioned data with a better chance of giving us signal with all the
RSN removed. For modelling a trace 𝒏 in a common receiver gather, the signal 𝑺(𝑛) holds the following
relationship with the RSN from the next sweep,
𝑵(𝑛)1 = 𝑺(𝑛 + 1) 𝑒 −𝑗2𝜋𝑓𝑡0 ; (4)
±𝒊
where 𝑡0 is the time interval between two consecutive sweeps and 𝑵(𝑛) is the RSN in trace n from
±𝑖 sweeps after or before the current sweep, which may be described as,
𝑵(𝑛)±𝒊 = 𝑺(𝑛 ± 𝑖) 𝑒 ±𝑗2𝜋𝑓(𝑖 𝑡0 ) . (5)
±𝒊
Therefore, we can model the noise 𝑵(𝑛) if we can isolate 𝑺(𝑛 ± 𝑖). The high amplitude events can
be isolated via windowing and filtering (e.g., median, f-k fan-filter or MDS) to remove the weaker
RSN within this window. The corresponding phase is then applied to this model to create modelled
RSN which can now be subtracted from the input data. This process removes the problematic high
amplitude RSN from the next sweeps, at least partially. In the resulting data, the amplitude of the
noise is much reduced and is thereby easier for the MDS to operate on, as illustrated by the result in
Figure 2c and its f-k plot in Figure 2f.

Conclusions

Phase encoding can enable a much-reduced sweep-location interval in marine vibrator surveys without
compromising on the residual shot noise in the signal. We show that phase encoding provides the
opportunity to perfectly separate sources up to a certain interfering frequency. Beyond this frequency
more elaborate methods for the separation of signal and blending noise is required. We demonstrate that

83rd EAGE Annual Conference & Exhibition


when the amplitude level of the blending noise is similar or weaker than the signal, approaches such as
matching pursuit can successfully remove the noise. However, when the blending noise is significantly
stronger than the underlying signal, strong artefacts are observed. Modelling a clean realisation of the
strong blending noise and feeding it to the separation algorithm through a pre-processing step
significantly enhances the quality of separated signal. This model-based process can be repeated for any
number of strong and problematic events.

Figure 2 (a) A section of the desired output data as reference. (b) Result after applying MDS
directly on data from figure 1a, and (c) result after applying MDS on a preprocessed data. Bottom
row shows corresponding f-k spectra. Note that the wavenumber axes are normalized by the trace
spacing (10m).

Acknowledgements

We thank Equinor and the Research Council of Norway for the financial support that makes this work
possible. We also thank Shearwater for permission to publish this work.

References
Casasanta, L. and Grion, S [2021]. October. Multidimensional Apparition De-Blending Through Sparse
Inversion 82nd EAGE Annual Conference & Exhibition 2021, Volume 2021, p.1 - 5.
Grion, S., Light, R. and Denny, S. [2018]. June. A seismic apparition experiment on towed streamer
seismic data. In 80th EAGE Conference and Exhibition 2018, pp. 1-5.
JafarGandomi, A., Holland, S. and Grion, S. [2021]. Adaptive de-blending of dithered simultaneous
sources. First Break, 39(5), pp.37-44.
Landrø, M. [2008]. The effect of noise generated by previous shots on seismic reflection data.
Geophysics, 73(3), Q9-Q17.
Laws, R. M., Halliday, D. F., Özbek, A., and Hopperstad, J. F. [2016]. Exploiting the control of phase
in marine vibrators. First Break, 34(11).
Laws, R.M., Halliday, D., Hopperstad, J-F., Gerez, D., Supawala, M., Özbek, A., Murray, T., and Kragh,
E. [2019]. Marine vibrators: the new phase of seismic exploration. Geophysical Prospecting.
67,1443-1471.
Mallat, S. G., and Zhang, Z. [1993]. Matching pursuits with time-frequency dictionaries. IEEE
Transactions on signal processing, 41(12), 3397-3415.
Moore, I., Dragoset, B., Ommundsen, T., Wilson, D., Ward, C. and Eke, D.[2008]. Simultaneous source
separation using dithered sources. In SEG Technical Program Expanded Abstracts 2008 (pp. 2806-
2810). Society of Exploration Geophysicists.

83rd EAGE Annual Conference & Exhibition

You might also like