Você está na página 1de 11

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.


IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Application of the Variational-Mode Decomposition


for Seismic Timefrequency Analysis
Ya-Juan Xue, Jun-Xing Cao, Da-Xing Wang, Hao-Kun Du, and Yao Yao

AbstractSeismic timefrequency analysis methods play an


important role in seismic interpretation for its superiority in
significantly revealing the frequency content of a seismic signal changes with time variation. Variational-mode decomposition
(VMD) is a newly developed methodology for decomposition on
adaptive and quasi-orthogonal signal and can decompose a seismic
signal into a number of band-limited quasi-orthogonal intrinsic
mode functions (IMFs). Each mode is an AMFM signal with the
narrow-band property and nonnegative smoothly varying instantaneous frequencies. Analysis on synthetic and real data shows
that this method is more robust to noise and has stronger local
decomposition ability than the empirical mode decomposition
(EMD)-based methods. Comparing with the short-time Fourier
transform (STFT) or wavelet transform (WT), instantaneous spectrum after VMD promises higher spectral and spatial resolution.
Application of the VMD on field data demonstrates that instantaneous spectrum after VMD targets the thickness variation in the
coal seam more sensitively than the conventional tools and highlights the fine details that might escape unnoticed. The technique
is more promising for seismic signal processing and interpretation.
Index TermsInstantaneous frequency, seismic interpretation,
seismic timefrequency analysis, variational-mode decomposition
(VMD).

I. I NTRODUCTION

URRENTLY, timefrequency analysis methods are


widely used in seismic data processing and interpretation for its superiority in significantly revealing how the
Manuscript received May 4, 2015; accepted February 9, 2016. This work was
supported in part by the National Natural Science Foundation of China under
Grant 41430323, Grant 41404102, Grant 41274128, and Grant 61401047, in
part by the Sichuan Youth Science and Technology Foundation under Grant
2016JQ0012, in part by the Key Project of Sichuan provincial Education
Department (16ZA0218), in part by the 2015 Annual Young Academic Leaders
Scientific Research Foundation of CUIT (J201507), and in part by the Project of
the Scientific Research Foundation of CUIT (KYTZ201503). (Corresponding
author: Ya-Juan Xue.)
Y.-J. Xue is with the School of Communication Engineering, Chengdu
University of Information Technology, Chengdu 610225, China, and also with
the State Key Laboratory of Oil and Gas Reservoir Geology and Exploitation,
the School of Geophysics in Chengdu University of Technology, Chengdu
610059, China (e-mail: xueyj0869@163.com).
J.-X. Cao is with the State Key Laboratory of Oil and Gas Reservoir Geology
and Exploitation, School of Geophysics, Chengdu University of Technology,
Chengdu 610059, China (e-mail: caojx@cdut.edu.cn).
D.-X. Wang is with the Exploration and Development Research Institute,
PetroChina Changqing Oilfield Company, Xian 710021, China (e-mail:
wdxl_cq@petrochina.com.cn).
H.-K. Du is with the Geophysical Institute of Zhongyuan Oilfield, Henan
457000, China (e-mail: dhk20052005@163.com).
Y. Yao is with the School of Communication Engineering, Chengdu
University of Information Technology, Chengdu 610225, China (e-mail:
yaoyao386@yahoo.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JSTARS.2016.2529702

frequency content of a seismic signal varies in time. The traditional seismic timefrequency analysis methods, which mainly
include short-time Fourier transform (STFT) (e.g., [1]), wavelet
transform (WT) (e.g., [2]), S-transform (ST) (e.g., [3]), and
WignerVille distribution (WVD) (e.g., [4]), are all successfully used in seismic timefrequency analysis. But they are all
especially limited by the timefrequency resolution due to their
decomposition basis [2], [5][8]. Most recently, empirical mode
decomposition (EMD)-based timefrequency analysis methods
which have significantly higher timefrequency resolution have
also been used in seismic interpretation and show their advantages in highlighting subtle geologic structure features and the
hydrocarbon information and removing the noise and so on
(e.g., [9][15]). Despite EMD which can recursively decompose a multicomponent seismic trace into a number of intrinsic
mode functions (IMFs) has progressed from EMD [16] to mode
mixing elimination methods such as ensemble EMD (EEMD)
[17] and complete EEMD (CEEMD) [18], there are still some
apparent limitations such as its sensitivity to noise and sampling
and lack of mathematical theory and so on.
The newly developed variational-mode decomposition
(VMD) method which is theoretically well founded by
Dragomiretskiy and Zosso is an entirely nonrecursive decomposition method for adaptive and quasi-orthogonal signal
decomposition and can decompose a multicomponent seismic
trace into a finite number of band-limited IMFs concurrently
[19]. Due to the Wiener filtering embedded to update the mode
directly in Fourier domain, the VMD algorithm is more robust
to noise than the EMD-based adaptive decomposition methods.
The objective of this paper is to demonstrate the ability of this
new VMD-based method for seismic timefrequency analysis.
This work is hoped to complement the current seismic processing methods with the addition of the VMD-based method.
First, we describe and illustrate the VMD procedures and its
equivalent filter property and comparatively study with EMD,
EEMD, and CEEMD using synthetic examples to show the better features of VMD. Next, we apply the VMD on field data
to highlight various geologic structure features and study the
instantaneous attributes comparatively with CEEMD and other
timefrequency tools.

II. T HEORY
A. Variational-Mode Decomposition
VMD nonrecursively decomposes a multicomponent signal
into an ensemble of band-limited IMFs with specific sparsity
property. IMF u(t) here is defined as an AMFM signal with

1939-1404 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

the narrow-band property [19]


u(t) = A(t) cos ((t))

(1)

where the phase (t) is a nondecreasing function, and the envelope A(t) and the instantaneous frequency (t) =  (t) are
nonnegative and vary much slower than the (t).
VMD is a constrained variational problem represented by the
following equation [19]:


2 

 

j
j
t
k
t ((t) + ) uk (t) e

min


t
{uk },{k }
2
k

uk = f
(2)
subject to
k

where uk is the kth mode. k is a center frequency around which


uk is mostly compact. The bandwidth of each mode is determined by the squared H1 norm of its baseband-shifted Hilbert
complemented analytic signal with only positive frequencies.
A quadratic penalty and Lagrangian multipliers are introduced
to address the (2). The augmented Lagrangian is set using the
following equation:
2






j
jk t 

L (uk , k , ) =
t (t) + t uk (t) e

2
k
2







uk  + , f
(3)
+ f
uk
2

where is the balancing parameter of the data-fidelity constraint.


Alternate direction method of multipliers (ADMM) approach
is used to solve the variational problem in (3) and produce
different decomposed modes and the center frequency during
each shifting operation. Each mode obtained from solutions in
spectral domain can be represented as



() /2
f () i=k u
i () +
u
k () =
.
(4)
2
1 + 2( k )
VMD mainly includes the following steps.
() are updated as shown
1) Modes update. The modes u
n+1
k
in (5). Wiener filtering is embedded for update the mode
directly in Fourier domain with a filter tuned to the current
center frequency kn
u
n+1
() =
k




n () /2
n+1
() i>k u
ni () +
f () i<k u
i
1 + 2( kn )

(5)

2) Center frequencies update. The center frequencies kn+1


are updated as the center of gravity of the corresponding
modes power spectrum as shown in (6)
kn+1

2
  n+1
k () d
u
0
= 
.
2
u
n+1 () d
0

(6)

3) Dual ascent update. For all 0, the Lagrangian


n+1 is updated by (7) as dual ascent
multiplier
to enforce exact signal reconstruction until
2

2
 n+1 u
nk 2 / 
unk 2 <
k u
k



n+1
n+1
n
+ f

.
(7)
=
u

The detailed complete algorithm of VMD can be found in


[19]. The fact that wiener filtering is embedded for modes
update makes the VMD algorithm more robust to noise.
B. Instantaneous Attributes
For each IMF u(t), seismic instantaneous attributes can be
extracted using the IMF u(t) and its Hilbert transform y(t) by
computing its analytic signal
z(t) = u(t) + iy(t) = A(t) exp[i(t)]

)
where y(t) = 1 P u(
t d . P denotes the Cauchy principal
value.
Then, instantaneous amplitude A(t) and instantaneous phase
(t) and instantaneous frequency (t) can be computed by the
following equation:


2
2

A(t) = u (t) + y (t)


(8)
(t) = arctan(y(t)/u(t))

1 d(t)
(t) = 2 dt .
Generally, for avoiding ambiguities due to phase unwrapping
in (8), instantaneous frequency (t) can be computed instead
from the following equation:
(t) =

dy(t)
du(t)
1 u(t) dt dt y(t)
.
2
u2 (t) + y 2 (t)

(9)

Similar to the EMD, EEMD, and CEEMD methods, VMD


can produce one instantaneous frequency for each IMF at each
time sample and thus form a multitude of instantaneous frequencies at each time sample, by which we can give an in-depth
signal analysis. Since instantaneous frequency and instantaneous amplitude are all a function of time, we can define a
three-dimensional (3-D) space [t, (t), A(t)]. Let
 n



i i (t)dt
H(, t) = Re
(10)
Ai (t)e
i=1

where Re denotes taking the real part of the result. n is


the number of the modes. By turning two variables function H(, t) into three variables function [t, , H(, t)] among
which A(t) = H [(t), t], the 3-D space can be generalized.
Thus, the joint timefrequency distribution of the signal is
obtained and it is uniformly sampled in time but not in frequency. Based on the joint timefrequency distribution, we
can derive many useful instantaneous attributes of spectral
decomposition technique. Here, we also compute the peak frequency attribute. The peak frequency is the frequency where the
maximum energy in each time sample occurs, and it is aided for
interpreting the temporal thickness of the layer of interest.
For the theory about EMD, EEMD, and CEEMD, one can
found it in [16][18].

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
XUE et al.: APPLICATION OF THE VMD

Fig. 1. Synthetic signal composed of an initial 30-Hz cosine wave, superposed


100-Hz Ricker wavelets at 0.1 s, 70-Hz Ricker wavelets at 0.3 s, and two 40-Hz
Ricker wavelets at 0.49 and 0.51 s.

Fig. 3. EEMD result of the synthetic signal. The EEMD output is generated
with 10% added Gaussian white noise and 100 realizations. Some mode mixing
still can be found in IMF2 and IMF3. But the mode mixing is reduced to large
extent compared with the EMD results in Fig. 2.

Fig. 2. EMD result of the synthetic signal. Mode mixing phenomenon can be
found in IMF1 seriously. Thus, the following IMFs are all distorted.

III. E XAMPLES
A. Equations Synthetic Data: VMD and EMD, EEMD,
CEEMD
In this section, we first use synthetic data to compare VMD
with the various EMD-based methods. Then, we investigated
the equivalent band-limited filter behavior of VMD. Finally,
the synthetic data with added noise are used to study the noise
robustness of VMD with the comparison of CEEMD.
The synthetic signal shown in Fig. 1 is composed of an initial 30-Hz cosine wave, superposed 100-Hz Ricker wavelets at
0.1 s, 70-Hz Ricker wavelets at 0.3 s, and two 40-Hz Ricker
wavelets at 0.49 and 0.51 s.
EMD decomposes the synthetic signal into six IMFs (Fig. 2).
But mode mixing phenomenon can be found in IMF1 seriously
and presents difficulties in recognizing the individual contributions of each component; thereby, the IMF lacks physical
meaning. And affected by the IMF1, the following IMFs are
all distorted.
The EEMD result is shown in Fig. 3. The EEMD output generates with 10% added Gaussian white noise and 100

Fig. 4. CEEMD result of the synthetic signal. The CEEMD output is generated
with 10% added Gaussian white noise and 100 realizations. The mode mixing
in CEEMD is slight and the IMFs are extracted more correctly than those in
EMD and EEMD (Figs. 2 and 3).

realizations. In IMF1, most of the 100-Hz Ricker wavelets at


0.1 s, 70-Hz Ricker wavelets at 0.3 s, and little of the two 40Hz Ricker wavelets at 0.49 and 0.51 s are retrieved. Although
some mode mixing still can be found in IMF2 and IMF3, the
mode mixing is reduced to large extent when compared with
the EMD results in Fig. 2.
Fig. 4 shows the CEEMD result of the synthetic signal. The
CEEMD output is generated with 10% added Gaussian white

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 5. VMD result of the synthetic signal. The decomposition level is set to 4.
The output is least affected by mode mixing compared with Figs. 24.

noise and 100 realizations. The resulting IMF1 in CEEMD is


similar to that in EEMD (Fig. 3), retrieving most of the 100-Hz
Ricker wavelets at 0.1 s, 70-Hz Ricker wavelets at 0.3 s, and
little of the two 40-Hz Ricker wavelets at 0.49 and 0.51 s. The
resulting IMF2 and IMF3 contain mostly the 100-Hz Ricker
wavelets at 0.1 s, 70-Hz Ricker wavelets at 0.3 s, and two 40-Hz
Ricker wavelets at 0.49 and 0.51 s. In IMF4, the background 30Hz cosine wave is mainly reflected. Although there is still mode
mixing in IMF4 due to the frequency of the background wave
close to the two 40-Hz Ricker wavelets at 0.49 and 0.51 s, the
mode mixing in CEEMD is slight and the IMFs are extracted
more correctly when compared with the results in EMD and
EEMD (Figs. 2 and 3).
The VDM result is shown in Fig. 5. The decomposition level
is set to 4. Thus, there are four IMFs generated. The background
wave is mainly reflected in IMF1. The 100-Hz Ricker wavelets
at 0.1 s, 70-Hz Ricker wavelets at 0.3 s and two 40-Hz Ricker
wavelets at 0.49 and 0.51 s are mainly reflected in IMF2 to
IMF4. Compared with the EMD, EEMD, and CEEMD outputs
(Figs. 24), the VDM output is least affected by mode mixing.
Fig. 6 shows the spectrums of the CEEMD output and the
VDM output. As shown in Fig. 6(a), for the IMFs in CEEMD
(only IMF1 to IMF4 are shown), there is an overlap of half
bandwidth between the two adjacent IMFs (suitable for the
IMF2, IMF3, and IMF4). For the high-frequency section (the
IMF1 and IMF2), their bandwidths are overlapped. But the
behavior of all the IMFs in VDM [Fig. 6(b)] shows the bandpass
filters characteristics evolving with the increasing predominant
center frequencies, even for the high-frequency section (the
IMF1 and IMF2).
Further for investing the noise robustness of VMD, we add
the Gaussian noise only distributed within the time 0.134
0.201 s [Fig. 7(a)] to the synthetic signal in Fig. 1. The resulting
noisy signal is shown in Fig. 7(b). The CEEMD output also
using 10% added Gaussian white noise and 100 realizations
is shown in Fig. 8. The background cosine wave is retrieved
in IMF4 with slight mode mixing around 0.5 s. The noise is
reflected in the IMF1 to IMF3 which mainly retrieve the 100Hz Ricker wavelets at 0.1 s, 70-Hz Ricker wavelets at 0.3 s and
two 40-Hz Ricker wavelets at 0.49 and 0.51 s. The noise is not
separated from the high-frequency section. For VMD method
shown in Fig. 9, we can find that the background cosine wave
is retrieved in IMF1.The noise is mainly reflected in IMF4. The

Fig. 6. Spectrums of the CEEMD output and the VMD output for the synthetic signal. (a) Spectrums of the CEEMD output. Note that only the spectra
of the first four IMFs are shown. CEEMD shows an overlap of half bandwidth between the two adjacent IMFs (suitable for the IMF2 to IMF4). For the
high-frequency section (the IMF1 and IMF2), their bandwidths are overlapped.
(b) Spectra of the VMD output. VMD shows the bandpass filters characteristics evolving with the increasing predominant center frequencies, even for the
high-frequency section (the IMF1 and IMF2).

Fig. 7. Noisy synthetic signal. (a) Gaussian noise only distributed within the
time 0.1340.201 s. (b) Resulting noisy signal.

Ricker wavelets are mainly retrieved in IMF2 and IMF3 with


a slight influence of noise. Compared with Fig. 8, VMD show
the noise robustness more than the CEEMD method. This also
can be seen in their spectrums. Fig. 10 shows the spectrums of
the CEEMD output and the VDM output for noisy signal. Note
that only the spectra of the first four IMFs are shown in CEEMD
[Fig. 10(a)]. As shown in Fig. 10(a), the bandwidths of IMF1
and IMF2 are overlapped. Due to the noise influence, the bandwidth of IMF1 becomes larger. For the spectrums of the VMD
output shown in Fig. 10(b), the bandwidth of IMF1 to IMF3 is
similar to that in Fig. 6(b). The noise is separated and mainly
reflected in IMF4. Compared Fig. 10(b) with CEEMD output
in Fig. 10(a), VMD shows better noise robustness.
B. Synthetic Data: Instantaneous Frequencies
In this section, a challenging synthetic data are used. We
first show that EMD-based methods will fail to distinguish the

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
XUE et al.: APPLICATION OF THE VMD

Fig. 10. Spectra of the CEEMD output and the VMD output for the noisy signal. (a) Spectra of the CEEMD output. Note that only the spectra of the first
four IMFs are shown. The bandwidths of IMF1 and IMF2 are overlapped. Due
to the noise influence, the bandwidth of IMF1 becomes larger. (b) Spectra of
the VMD output. The bandwidth of IMF1IMF3 is similar to that in Fig. 6(b).
The noise is separated and mainly reflected in IMF4. Compared with CEEMD
output in (a), VMD shows better noise robustness.

Fig. 8. CEEMD result of the noisy synthetic signal. The CEEMD output is
generated with 10% added Gaussian white noise and 100 realizations. The noise
is reflected in the IMF1 to IMF3 which mainly retrieve the Ricker wavelets. The
background cosine wave is retrieved in IMF4 with slight mode mixing around
0.5 s.
Fig. 11. Synthetic data are composed of an initial 20-Hz cosine wave with maximum amplitude 1, superposed two 40-Hz Ricker wavelets at 0.7 and 0.8 s, and
two different frequency cosine waves between 1.2 and 1.8 s of, respectively,
2 Hz with maximum amplitude 1 and 80 Hz with maximum amplitude 0.25.

Fig. 9. VMD result of the noisy synthetic signal. The decomposition level is
set to 4. The noise is mainly reflected in IMF4. The Ricker wavelets are mainly
retrieved in IMF2 and IMF3 with a slight influence of noise. The background
cosine wave is retrieved in IMF1.

individual components of such signal for demonstrating the


strong local decomposition ability of VMD. Then we show
the instantaneous spectral analysis after VMD has higher
timefrequency resolution and greater precision than WTs and
CEEMD.
The synthetic data used here shown in Fig. 11 are composed
of an initial 20-Hz cosine wave with maximum amplitude 1,
superposed two 40-Hz Ricker wavelets at 0.7 and 0.8 s, and
two different frequency cosine waves between 1.2 and 1.8 s of,
respectively, 2 Hz with maximum amplitude 1 and 80 Hz with
maximum amplitude 0.25.
Figs. 1214, respectively, show the EMD, EEMD, and
CEEMD results. The EEMD output generates with 10% added

Fig. 12. EMD result of the synthetic data in Fig. 11. The EEMD output is
generated with 10% added Gaussian white noise and 100 realizations. Mode
mixing phenomenon can be found in IMF1 seriously. Thus, the following IMFs
are all distorted. EMD does not work for the synthetic data.

Gaussian white noise and 100 realizations. The CEEMD output is also generated with 10% added Gaussian white noise
and 100 realizations. As shown in Figs. 1214, IMF1 in EMD,
EEMD, and CEEMD does not solely extract the high-frequency
80-Hz cosine wave between 1.2 and 1.8 s but is polluted with
low-frequency components including the background 20-Hz
cosine wave and two 40-Hz Ricker wavelets at 0.7 and 0.8 s.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 15. VMD result of the synthetic data in Fig. 11. The decomposition level is
set to 4. The 2-Hz cosine wave between 1.2 and 1.8 s is first extracted in IMF1.
The background 20-Hz cosine wave is reflected in IMF2. Two 40-Hz Ricker
wavelets at 0.7 and 0.8 s are mainly reflected in IMF3. The 80-Hz cosine wave
between 1.2 and 1.8 s is mainly reflected in IMF4.

Fig. 13. EEMD result of the synthetic data in Fig. 11. Mode mixing phenomenon is seriously in IMF1. The following IMFs are all distorted. EEMD
does not work for the synthetic data.

Fig. 16. Time-frequency spectrum comparison for the synthetic data in Fig. 11.
(a) WT analysis. The morlet wavelet is used. It distinguishes 20-Hz background
cosine wave. But the time resolution of the 2-Hz cosine wave between 1.2 and
1.8 s is poor. The frequency resolution of the 80-Hz cosine wave and two 40Hz Ricker wavelets at 0.7 and 0.8 s is poor. (b) Instantaneous spectrum after
CEEMD. 20-Hz background cosine wave is distinguished. The time resolution
of the 2-Hz cosine wave between 1.2 and 1.8 s is higher than WT in Fig. 16.
But the frequency resolution of the two 40-Hz Ricker wavelets at 0.7 and 0.8 s
and the 80-Hz cosine wave between 1.2 and 1.8 s is still poor due to the mode
mixing. (c) Instantaneous spectrum after VMD has the highest timefrequency
resolution and precision and identifies all individual components. VMD shows
strong local decomposition ability.

Fig. 14. CEEMD result of the synthetic data in Fig. 11. The CEEMD output
is generated with 10% added Gaussian white noise and 100 realizations. Mode
mixing phenomenon is still seriously in IMF1 and the following IMFs are all
distorted. CEEMD does not work for the synthetic data.

Thus, the following IMFs are all distorted. This makes it difficult to distinguish the different contributions of each IMF,
thereby complicating signal analysis. For this synthetic data,
EMD-based methods all fail to work.
The VMD result of the synthetic data is shown in Fig. 15. The
decomposition level is set to 4. The resulting IMF1 is retrieving the 2-Hz cosine wave between 1.2 and 1.8 s completely.
The background 20-Hz cosine wave is reflected in IMF2. Two
40-Hz Ricker wavelets at 0.7 and 0.8 s are mainly reflected in
IMF3. The 80-Hz cosine wave between 1.2 and 1.8 s are mainly
reflected in IMF4. Comparing with the EMD-based methods

shown in Figs. 1214, VMD shows strong local decomposition


ability.
After the VMD decomposition, we can compute the instantaneous frequency and instantaneous amplitude of each IMF with
(8) and (9). Then the resulting timefrequency spectrum can be
generated. Next, we compare the resulting instantaneous spectrum with the WTs and CEEMD for the same synthetic data
shown in Fig. 11.
Fig. 16 shows the timefrequency distribution corresponding to the various tools. WT analysis is shown in Fig. 16(a).
The morlet wavelet is used. It distinguishes 20-Hz background
cosine wave. But the time resolution of the 2-Hz cosine wave
between 1.2 and 1.8 s is poor. For the two 40-Hz Ricker
wavelets at 0.7 and 0.8 s and the 80-Hz cosine wave between
1.2 and 1.8 s, the frequency resolution is poor.
Fig. 16(b) displays the instantaneous spectrum after
CEEMD. 20-Hz background cosine wave is distinguished. The

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
XUE et al.: APPLICATION OF THE VMD

Fig. 19. VMD result of trace of CDP 822 in Fig. 18. The decomposition level
is set to 3. Random noise is mainly reflected in IMF3.
Fig. 17. 2-D prestack seismic section with random noise interference. The data
are sampled at 2 ms. Strong reflection amplitudes exist in the area where coal
seam located between 1720 and 1746 ms.

Fig. 18. Trace of CDP 822 in Fig. 17.

time resolution of the 2-Hz cosine wave between 1.2 and 1.8 s
is higher than WT. But for the high-frequency components of
the two 40-Hz Ricker wavelets at 0.7 and 0.8 s and the 80-Hz
cosine wave between 1.2 and 1.8 s, their frequency resolution
is still poor due to the mode mixing.
Instantaneous spectrum after VMD is shown in Fig. 16(c).
VMD can identify all individual components. Thus, the highfrequency components of the two 40-Hz Ricker wavelets at 0.7
and 0.8 s and the 80-Hz cosine wave between 1.2 and 1.8 s are
all clearly seen in Fig. 18. VMD which has strong local decomposition ability shows the highest timefrequency resolution
and precision.
The two synthetic examples show the strong local decomposition ability and noise robustness and significantly higher
timefrequency resolution of VMD combined with instantaneous frequencies over that obtainable with WTs and CEEMD
combined with instantaneous frequencies.

IV. R EAL DATA


Next, we use the seismic data set from Ordos Basin in China
to demonstrate the features of VMD-based method. The twodimensional (2-D) prestack seismic section with random noise
interference (Fig. 17) consists of 401 seismic traces. The data
are sampled at 2 ms. There exist strong reflection amplitudes
where coal seam located between 1720 and 1746 ms. The
sandstone information is with weak amplitudes.
First, we take the trace for CDP 822 (Fig. 18) as an example to show the timefrequency distributions corresponding to
the different transforms. The VMD result is shown in Fig. 19.
The decomposition level is set to 3. Random noise is mainly
reflected in IMF3. For comparison, the CEEMD output with
10% added Gaussian white noise using 100 realizations is

Fig. 20. CEEMD output with 10% added Gaussian white noise using 100
realizations. Random noise is mainly reflected in IMF1.

shown in Fig. 20. Random noise with mainly high-frequency


components is mainly reflected in IMF1 for CEEMD. Fig. 21
shows the spectrum of the main IMFs for VMD results and
CEEMD results. Here, only the spectra of the first four IMFs
in CEEMD are shown. As shown in Fig. 21(a), the useful information mainly distributes below the 70 Hz. IMF1 which mainly
contains random noise in CEEMD distributes in the whole spectrum and some useful information located below the 70 Hz is
still retrieved in IMF1. But for VMD spectrum in Fig. 21(b),
random noise mainly reflects in IMF3 and is isolated completely in IMF3. IMF1 with the frequency below 30 Hz mainly
reflects the information of coal seam. Sandstone information is
mainly retrieved in IMF2 with the dominant frequency around
45 Hz. VMD output is more helpful for interpretation.
The results for the STFT with a 41-ms time window and the
WT and the instantaneous spectrum after VMD are shown in
Fig. 22. These tools show the similar feature that the strong
anomaly is visible between 1720 and 1746 ms due to the coal
seam. But instantaneous spectrum after VMD in Fig. 22(c)
provides a sparser image with much higher timefrequency

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 21. Spectrum of the main IMFs for VMD results and CEEMD results.
(a) Spectrum of the main IMFs for CEEMD results. Here, only the spectra of
the first four IMFs in CEEMD are shown. Random noise mainly reflected in
IMF1 and some useful information located below the 50 Hz is still retrieved in
IMF1. (b) Spectrum of the IMFs for VMD results. Random noise is isolated
completely in IMF3. IMF1 with the dominant frequency around 30 Hz mainly
reflects the information of coal seam. Sandstone information is mainly retrieved
in IMF2 with the dominant frequency around 45 Hz.

Fig. 23. IMFs from VMD for seismic section in Fig. 17. (a) IMF1. Strong
anomaly between 1720 and 1746 ms caused by coal seam is clearly visible. (b) IMF2. Sandstone information with less coal seam influence is mainly
reflected. (c) IMF3. Random noise is mainly reflected. (d) Reconstructed
section by IMF1 and IMF2. Compared with the original seismic section in
Fig. 17, random noise is suppressed and the fine details are more clearly visible.

Fig. 24. Peak frequency attribute, respectively, based on STFT and VMD.
The black rectangle shows the area where coal seam locates between 1720
and 1746 ms. (a) Peak frequency attribute based on STFT with a 41-ms time
window. The output image is continuous and shows lower timefrequency resolution. (b) Peak frequency attribute based on VMD. The output image is much
sparser and has higher timefrequency resolution.
Fig. 22. Timefrequency spectrum comparison for trace of CDP 822. (a) STFT
with a 41-ms time window. The strong anomaly between 1720 and 1746 ms is
due to the coal seam. (b) Wavelet analysis. Strong anomaly is found between
1720 and 1746 ms. Vertical stripes at higher frequencies are due to an increased
time resolution but poorer frequency resolution. (c) Instantaneous spectrum
after VMD displaying the highest timefrequency resolution. The decomposition level of VMD is set to 3. Similar strong anomaly is found between 1720
and 1746 ms.

resolution than the STFT and WT. The timefrequency distribution provided by VMD can help these spectral anomalies
located more accurately and facilitated further interpretation.
IMFs from VMD for the seismic section in Fig. 17 are shown
in Fig. 23. Strong anomaly between 1720 and 1746 ms caused
by coal seam is clearly visible in IMF1 [Fig. 23(a)]. Sandstone
information with less coal seam influence is mainly reflected
in IMF2 [Fig. 23(b)]. IMF3 in Fig. 23(c) which contains the
high-frequency components mainly reflects random noise. The
reconstructed section by IMF1 and IMF2 is shown in Fig. 23(d).
Compared with the original seismic section in Fig. 17, random
noise is suppressed, and the fine details are more clearly visible.

Then, we extract the peak frequency attribute, respectively,


based on STFT with a 41-ms time window and VMD for comparison (Fig. 24). The black rectangle in Fig. 24 shows the area
where coal seam locates between 1720 and 1746 ms. Both the
STFT-based and VMD-based peak frequency attributes show
similar features: lower peak frequency at the two ends and
higher peak frequency at the middle in the black rectangle. They
reveal the thickness of coal seam at the two ends is thicker
than that in the middle part. As Fig. 24(a) shown, the output
image of the peak frequency attribute based on STFT is continuous and shows lower timefrequency resolution. The peak
frequency attribute based on VMD shown in Fig. 24(b) is much
sparser and has higher timefrequency resolution.
We finally extract the 18-, 20-, 25-, and 30-Hz common frequency sections after VMD (Fig. 25) and STFT (Fig. 26) to
illustrate the higher time-spatial resolution of the VMD-based
results and show the thickness variation in coal seam revealed
by the VMD-based results more clearly. A 55 Gaussian filter is applied to the output of the instantaneous spectrum after

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
XUE et al.: APPLICATION OF THE VMD

Fig. 25. Common frequency section after VMD for seismic section in Fig. 17.
A 55 Gaussian filter is applied to the output for display purpose. (a) 18Hz common frequency section after VMD. Some strong amplitude anomaly
between 1720 and 1746 ms is found due to the coal seam. (b) 20-Hz common frequency section after VMD. Strong amplitude anomaly between 1720
and 1746 ms is increased. (c) 25-Hz common frequency section after VMD.
Strong amplitude anomaly between 1720 and 1746 ms reaches the maximum.
(d) 30-Hz common frequency section after VMD. Most of the strong amplitude
anomaly between 1720 and 1746 ms is diminished. The thickness variation
in coal seam is clearly reflected and VMD shows the higher time and spatial
resolution.

Fig. 26. Common frequency section after STFT for seismic section in Fig. 17
with a 41-ms time window. (a) 18-Hz common frequency section after VMD.
Slightly stronger amplitude anomaly between 1720 and 1746 ms is found due
to the coal seam. (b) 20-Hz common frequency section after VMD. Amplitude
anomaly between 1720 and 1746 ms is stronger. (c) 25-Hz common frequency
section after VMD. Strong amplitude anomaly between 1720 and 1746 ms is
found. (d) 30-Hz common frequency section after VMD. Amplitude anomaly
between 1720 and 1746 ms is slightly diminished. STFT shows significantly
less amplitude variations due to the lower time and spatial resolution, and the
reflected thickness variation in coal seam is not fine compared with common
frequency section after VMD in Fig. 25.

VMD for generating a similar image as STFT. The instantaneous spectrum after VMD reveals the spectral characteristics
of the various reflections more clearly than the STFT results.
Some strong amplitude anomaly between 1720 and 1746 ms
is found caused by the coal seam in 18-Hz common frequency section [Fig. 25(a)]. Then, in 20-Hz common frequency
section [Fig. 25(b)], strong amplitude anomaly between 1720
and 1746 ms is increased. Strong amplitude anomaly between
1720 and 1746 ms reaches the maximum in 25-Hz common
frequency section [Fig. 25(c)]. In 30-Hz common frequency
section, most of them are diminished. Due to the low time and
spatial resolution, the instantaneous spectra after STFT with a
41-ms time window in Fig. 26 show significantly less amplitude
variations and it is unable to reveal the variation as the instantaneous spectrum after VMD does. VMD shows the higher time
and spatial resolution. The thickness variations of the coal seam
revealed by the VMD-based results are more clear.

The EMD-based methods use the local time-scale characteristics of the data to adaptively decompose the signal into IMFs
from high frequency to low frequency to tendency. Two primary
tasks fundamentally govern the EMD-based methods process.
First, extrema must be properly located to avoid masking the
detection of local extrema in any strong trend in various windows of the series. Second, the envelope to the upper and lower
extrema must be properly fitted to avoid either introducing artifacts or excluding intrinsic trends in the sifted result. Some
interpolation methods such as cubic and akima will dramatically affect the sifting results. The EMD-based methods are
sensitive to noise and sampling. The behavior of the EMDbased methods is proved to act essentially as an adaptive,
multiband overlapping filter bank [12]. But VMD is a generalization of the class wiener filter into multiple, adaptive bands
[19]. It decomposes the signal into a given number of quasiorthogonal IMFs either exactly or in a least-squares sense. IMFs
extracted from VMD are concurrent. VMD does not explicitly
deal with a global trend in the signal. The tight relations to the
wiener filter make VMD have some optimality in dealing with
noise. And VMD is made much more robust to sampling. Each
resulted IMF is compact around a center pulsation and contains
the information on the local characteristics of the data and has
certain physical meaning. The different treatments between the
EMD-based methods and VMD process explain why the 80-Hz
cosine wave with maximum amplitude 0.25 between 1.2 and
1.8 s is mixed with the 20-Hz background cosine wave and two
40-Hz Ricker wavelets at 0.7 and 0.8 s in IMF1 in the EMDbased methods while it is isolated in IMF4 in the VMD process
for our second synthetic signal.
VMD combined with instantaneous frequencies are easy to
implement. For display purpose, we can also apply the different

V. D ISCUSSION
One limitation of the EMD-based methods is that they
lack the mathematical foundation. IMF is explicitly defined
having the characteristics that the number of extrema and zerocrossings must either equal or differ at most by one in the
whole signal segment and the mean value of the envelope
defined by the local maxima and the envelope defined by the
local minima is zero at any point [16]. Unlike the EMD-based
methods, VMD algorithm is proposed with a mathematical
foundation. IMFs in VMD are only referred to AMFM signals with narrow-band property. The formula that relates the
parameters of the AMFM descriptors to the estimated signal
bandwidth is provided.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

sizes of the Gaussian filter to the instantaneous spectrum after


VMD for generating a similar distribution as the STFT, WT.
This also changes the timefrequency resolution and makes
the timefrequency resolution of the instantaneous spectrum
after VMD controllable. The real data example confirms that
instantaneous spectrum after VMD has higher timefrequency
resolution than traditional decompositions.
The peak frequency attribute is referred to the frequency volume in which the frequency value is at the maximum amplitude
for each time sample and this attribute helps for interpreting
the temporal thickness of the layer of interest. High-peak frequency indicates thin beds and low peak frequency indicates
thick beds. The real data example verifies the peak-frequency
attribute after VMD may vary more rapidly in time and space.
Thus, it will make the interpretation more challenging.
VI. C ONCLUSION
The ability of the VMD for quantifying meaningful geologic
information in the time and timefrequency domains is investigated in this work. VMD is more robust to noise and have the
strong local decomposition ability than EMD-based methods.
The band-limited IMFs with specific sparsity property obtained
from VMD are more convenient for seismic data processing.
Instantaneous frequency spectra after VMD represent geologic
targets with higher resolution in both time and frequency than
STFT and WTs on synthetic examples and field data. Therefore,
these characteristics make the technique more promising for
seismic signal processing and interpretation and allow for a
wide range of tools to be developed.
R EFERENCES
[1] G. Partyka, J. Gridley, and J. Lopez, Interpretational applications of
spectral decomposition in reservoir characterization, Lead. Edge, vol. 18,
no. 3, pp. 353360, Mar. 1999.
[2] A. Chakraborty and D. Okaya, Frequency-time decomposition of seismic data using wavelet-based methods, Geophysics, vol. 60, no. 6,
pp. 19061916, Nov. 1995.
[3] E. Odebeatu, J. Zhang, M. Chapman, E. Liu, and X. Y. Li, Application
of spectral decomposition to detection of dispersion anomalies associated
with gas saturation, Lead. Edge, vol. 25, no. 2, pp. 206210, Feb. 2006.
[4] Y. Li and X. Zheng, Spectral decomposition using wigner-ville distribution with applications to carbonate reservoir characterization, Lead.
Edge, vol. 27, no. 8, pp. 10501057, Aug. 2008.
[5] R. G. Stockwell, L. Mansinha, and R. P. Lowe, Localization of the complex spectrum: The S transform, IEEE Trans. Signal Process., vol. 44,
no. 4, pp. 9981001, Apr. 1996.
[6] S. Sinha, P. S. Routh, P. D. Anno, and J. P. Castagna, Spectral decomposition of seismic data with continuous-wavelet transform, Geophysics,
vol. 70, no. 6, pp. P19P25, Nov. 2005.
[7] S. Mallat, A Wavelet Tour of Signal Processing, 3rd ed. New York, NY,
USA: Academic press, 2009.
[8] X. Wu and T. Liu, Spectral decomposition of seismic data with reassigned smoothed pseudo wignerville distribution, J. Appl. Geophys.,
vol. 68, no. 3, pp. 386393, Jul. 2009.
[9] I. Magrin-Chagnolleau and R. G. Baraniuk, Empirical mode decomposition based time-frequency attributes, in 69th Ann. Internat. Mtg., Soc.
Expl. Geophys., Houston, Expanded Abstracts, 1999, pp. 19491952.
[10] B. M. Battista, C. Knapp, T. McGee, and V. Goebel, Application of the
empirical mode decomposition and hilbert-huang transform to seismic
reflection data, Geophysics, vol. 72, no. 2, pp. H29H37, Mar. 2007.
[11] M. Bekara and M. Van Der Baan, Random and coherent noise attenuation by empirical mode decomposition, Geophysics, vol. 74, no. 5,
pp. V89V98, Sep. 2009.
[12] T. Wang, M. Zhang, Q. Yu, and H. Zhang, Comparing the applications of
EMD and EEMD on timefrequency analysis of seismic signal, J. Appl.
Geophys., vol. 83, pp. 2934. Aug. 2012.

[13] J. Han and M. Van Der Baan, Empirical mode decomposition for seismic
time-frequency analysis, Geophysics, vol. 78, no. 2, pp. O9O19, Mar.
2013.
[14] Y.-J. Xue, J.-X. Cao, and R.-F. Tian, EMD and teager-kaiser energy
applied to hydrocarbon detection in a carbonate reservoir, Geophys. J.
Int., vol. 197, no. 1, pp. 277291, Apr. 2014.
[15] Y.-J. Xue, J.-X. Cao, R.-F. Tian, H.-K. Du, and Y.-X. Shu, Application
of the empirical mode decomposition and wavelet transform to seismic
reflection frequency attenuation analysis, J. Petrol Sci. Eng., vol. 122,
pp. 360370, Oct. 2014.
[16] N. E. Huang et al., The empirical mode decomposition and the hilbert
spectrum for nonlinear and non-stationary time series analysis, Proc. R.
Soc. Lond. A: Math. Phys. Eng. Sci., vol. 454, no. 1971, pp. 903995,
Mar. 1998.
[17] Z. Wu and N. E. Huang, Ensemble empirical mode decomposition: A
noise-assisted data analysis method, Adv. Adapt. Data Anal., vol. 1,
no. 1, pp. 141, Jan. 2009.
[18] M. E. Torres, M. A. Colominas, G. Schlotthauer, and P. Flandrin, A complete ensemble empirical mode decomposition with adaptive noise, in
Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP11), 2011,
pp. 41444147.
[19] K. Dragomiretskiy and D. Zosso, Variational mode decomposition,
IEEE Trans. Signal Process., vol. 62, no. 3, pp. 531544, Feb. 2014.
Ya-Juan Xue was born in Inner Mongolia, China.
She received the B.S. degree in communications
engineering from the Southwest Jiaotong University,
Sichuan, China, in 2002, and the M.S. degree in communication and information system and the Ph.D.
degree in geophysical prospecting and information
technology from Chengdu University of Technology,
Chengdu, China, in 2007 and 2014, respectively.
She is currently an Associate Professor with the
School of Communication Engineering, Chengdu
University of Information Technology, Chengdu,
China, and a Postdoc with the School of Geophysics and the State Key
Laboratory of Oil and Gas Reservoir Geology and Exploitation, Chengdu
University of Technology. She has participated in several national research
programs under the National Natural Science Foundation of China (NSFC).
She is currently the Principal Investigator (PI) in several national and regional
research projects, especially the national young research projects funded by
NSFC, the Outstanding Young Academic Leaders Fund of Sichuan Province,
and the 2015 Annual Young Academic Leaders Scientific Research Fund of
CUIT. Her research interests include adaptive weak signal detection algorithms,
seismic signals timefrequency analysis algorithms, reservoir characteristics
and hydrocarbon detection methods, and other vibration signal processing
algorithms.
Dr. Xue is an active member of the Society of Exploration Geophysicists
(SEG) and regular member of American Geophysical Union (AGU). She is a
Review Editor of the journal Frontiers in Earth Science. She was the recipient
of the excellent doctoral dissertation from Chengdu University of Technology
in 2014 and the 2015 annual advanced technology personal award of CUIT in
2015.
Jun-Xing Cao was born in Shaanxi Province, China,
in 1961. He received the B.S. degree in rock mineralogy from the Northwest University, Shaanxi
Province, China, in 1983, and the M.S. degree in mineralogy and the Ph.D. degree in applied geophysics
from Chengdu University of Technology, Chengdu,
China, in 1989 and 1996, respectively.
He was worked as a Postdoctoral fellow of
Electronics and Communication with the University
of Electronic Science and Technology, Chengdu,
China, from 1996 to 1998. He has been a
Distinguished Visiting Professor of Geophysics at Massachusetts Institute of
Technology, Cambridge, MA, USA, in 2005. He was the Provost of the
Chengdu University of Technology. He is currently the Professor and Dean
of the School of Geophysics, Chengdu University of Technology, and a
Principal Investigator with the State Key Laboratory of Oil and Gas Reservoir
Geology and Exploitation in Chengdu University of Technology. He currently
hosts two national key research projects, funded by the National Natural
Science Foundation of China (NSFC). His research interests include petroleum
geology and exploration, reservoir characteristics, geophysical inversion and
tomography, deep learning of seismic data and earthquakes, and geodynamics.
Prof. Cao is a life member of Chinese Geophysical Society, regular member of the Society of Exploration Geophysicists (SEG), and regular member
of American Geophysical Union (AGU). He was the academic and technical
leader in Geophysics of Sichuan Province, China since 2004.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
XUE et al.: APPLICATION OF THE VMD

Da-Xing Wang was born in Gansu, China, in 1963.


He received the B.S. and M.S. degrees in geophysical exploration from the Southwest Petroleum
University, Sichuan, China, in 1983 and 1995, respectively, and the Ph.D. degree in solid geophysics from
the Institute of Geology and Geophysics, Chinese
Academy of Sciences, Beijing, China, in 2005.
He is currently a Senior Engineer with the
Exploration and Development Research Institute of
PetroChina Changqing Oilfield Company, Xian,
China. His research interests include reservoir prediction and hydrocarbon detection.
Hao-Kun Du was born in Sichuan, China, in
1987. He received the B.S., M.S., and Ph.D.
degrees in geophysics from Chengdu University of
Technology, Chengdu, China, in 2009, 2012, and
2015, respecitvely.
He is currently a Postdoc with the Geophysical
Institute of Zhongyuan Oilfield, Henan Province,
China. His research interests include seismic reservoir prediction and seismic interpretation.

11

Yao Yao was born in Sichuan, China. He received


the B.S. degree in microelectronics from the School
of Information Science and Technology, Southwest
Jiaotong University, Sichuan, China, in 2006, and
the Ph.D. degree in communication and information
system from the Southwest Jiaotong University, in
2013.
He is currently an Associate Professor with the
School of Communication Engineering, Chengdu
University of Information Technology, Chengdu,
China. His research interests include signal acquisition and processing especially sensor technology.

Você também pode gostar