See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/284765618

Seismic data deconvolution using Kalman filter based on a new system model

Article in Geophysics · January 2016

DOI: 10.1190/geo2014-0611.1

CITATIONS READS

2 1,226

3 authors, including:

Zhengjun Zhang Dinghui Yang

University of Wisconsin–Madison Tsinghua University
126 PUBLICATIONS 2,201 CITATIONS 168 PUBLICATIONS 2,479 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Dinghui Yang on 05 October 2017.

The user has requested enhancement of the downloaded file.

GEOPHYSICS, VOL. 81, NO. 1 (JANUARY-FEBRUARY 2016); P. V31–V42, 10 FIGS., 1 TABLE.
10.1190/GEO2014-0611.1

Seismic data deconvolution using Kalman

filter based on a new system model

Xiaoying Deng1, Zhengjun Zhang2, and Dinghui Yang3

ABSTRACT select is fewer than the conventional. Furthermore, the number of

parameters can be reduced to only one by a theoretical demon-
Seismic resolution plays an important role in geologic interpre- stration for stationary noisy signals, which decreases the require-
tation and reservoir prediction. To improve the vertical resolution ment for multiple parameters selection in the conventional model.
of a seismic image, we have developed a new Kalman filter sys- The practical selection for this parameter should be a compromise
tem model for seismic deconvolution. Similar to the conventional between resolution improvement and noise amplification. Exper-
Kalman filter model for seismic deconvolution, our new Kalman imental results in the time and frequency domains on synthetic
model is also based on the common viewpoint that a reflected and field seismic records revealed that the Kalman filter based
seismic record can be regarded as a convolution of a seismic on the new model has the advantages of a higher resolution
wavelet with a reflection coefficient series. The new model uses and peak signal-to-noise ratio (PS/N) than the conventional Kal-
a reversed seismic wavelet to slide across a reflectivity function man filter for stationary and nonstationary signals, and it works
to achieve the convolution result, instead of using a reversed re- similarly to the Wiener filter for stationary signals, and it is supe-
flectivity function to slide across a seismic wavelet in the conven- rior to the Wiener filter in resolution and PS/N for nonstationary
tional Kalman filter model. A simpler state equation for the signals. The Kalman filter based on the new model can be applied
new model is achieved, and the number of parameters to to seismic resolution improvement.

INTRODUCTION imize the square sum of differences between the filter output and the
desired output to obtain the inverse filtering factor. We note that the
In seismic exploration, a reflected seismic record is convention- Wiener filter method conventionally assumes the desired output as a
ally considered as a convolution of a seismic wavelet and a reflec- unit impulse function and is only applied to stationary signals. Ul-
tivity function corresponding to the layered earth, which makes the rych (1971) proposes the homomorphic filtering method, which
spiky reflection coefficients blurred. Therefore, it is very important first transforms the seismic record into the logarithm spectrum do-
to remove the effects of the seismic wavelet’s complexity on the main, and then it separates the reflectivity function from the seismic
resulting seismic record and improve the seismic resolution in geo- record under the assumption of the separability of the seismic wave-
logic interpretation and reservoir prediction, especially for some let and the reflectivity function in the logarithm spectrum. This
fine seismic explorations such as the detecting and positioning method greatly depends on the separability degree of the wavelet
of thin beds. Many deconvolution techniques in the time or fre- and the reflectivity function, and it generally requires a seismic rec-
quency domain have been proposed to compress the seismic wave- ord with a high signal-to-noise ratio. The spectral whitening tech-
let to improve the resolution of the earth. Robinson (1957) uses the nique (Coppens and Mari, 1984) first computes the amplitude
Wiener filter theorem (Wiener, 1949) to extract the wavelet from the spectrum of each input record to determine a weighting function
seismic trace while leaving the reflection coefficients as a residual. to raise the energy of higher frequencies to the same level as that
Then, Rice (1962) and Robinson and Treitel (1967) propose to min- of lower frequencies without alteration of the phases. The function

Manuscript received by the Editor 25 December 2014; revised manuscript received 23 August 2015; published online 26 November 2015.
1
Beijing Institute of Technology, School of Information and Electronics, Beijing, China. E-mail: [email protected].
2
University of Wisconsin, Department of Statistics, Madison, Wisconsin, USA. E-mail: [email protected].
3
Tsinghua University, Department of Mathematical Sciences, Beijing, China. E-mail: [email protected].
© 2015 Society of Exploration Geophysicists. All rights reserved.

V31
V32 Deng et al.

is the inverse of the short-wavelength variation, which can be ob- X

M
tained by applying a low-pass filter to the amplitude spectrum. The yðnÞ ¼ bðmÞrðn − mÞ þ eðnÞ; n ¼ 1; 2; : : : ; N; (1)
limitation of the method is that it is only a crude attempt to balance m¼1
the amplitude spectrum of the seismic trace. Optimal seismic decon-
volution (Lainiotis et al., 1988) and its distributed algorithms where yðnÞ is the discrete seismic signal with a length of N, bðmÞ is
(Plataniotis et al., 1998) derive the new minimum variance decon- the seismic wavelet with a length of M, rðnÞ is the reflection co-
volution formulas that can use any filtering/smoothing scheme to efficient, and eðnÞ is additive noise. According to this convolution
obtain optimal estimates of the reflection coefficient sequence. This model of seismic signal, most of the seismic deconvolution methods
method is limited by the assumption that the reflection coefficient based on the Kalman filter used the following state equation and
sequence is random zero-mean and white. Inverse Q filtering measurement equation for extracting the reflection coefficients
(Wang, 2006) uses a reversal procedure of seismic wave propaga- (Crump, 1974):
tion to compensate the traveling attenuation quantified as the seis-
xðkÞ ¼ Φðk − 1Þxðk − 1Þ þ guðk − 1Þ (2)
mic quality Q factor. It is often difficult to derive an exact and
intrinsic stable Q model from the seismic data in practice. Gabor
and
deconvolution (Margrave et al., 2011) first estimates a time- and
frequency-dependent Gabor deconvolution operator formed from yðkÞ ¼ HðkÞxðkÞ þ vðkÞ; k ¼ 1;2; : : : ; N: (3)
the Gabor transform of a seismic trace by an iterative smoothing
process, then it estimates the Gabor transform of the reflectivity These two equations are called the Kalman filter system model. In
by spectral division in the time-frequency domain, and it finally state equation 2, the state vector xðkÞ ¼ ½ rðkÞ rðk − 1Þ · · ·
uses an inverse Gabor transform to recover the time-domain reflec- rðk − M þ 1ÞT represents parts of the desired reflection coeffi-
tivity. Subsequently, Wang et al. (2013) perform spectrum broad- cients sequence, g ¼ ½ 1 0 · · · 0T , the process noise
ening and amplitude compensation after molecular Gabor uðk − 1Þ is often supposed as the white noise sequence, and the state
transform. The inverse molecular Gabor transform is also needed transition matrix Φðk − 1Þ describes the relationship between two
to obtain a high-resolution, nonstationary seismic trace. In general, state vectors from time k − 1 to time k as follows:
the method requires nonstationary propagating wavelets. 2 3
The Kalman filter and its related models and computational for- φ1 ðk − 1Þ ··· φM−1 ðk − 1Þ φM ðk − 1Þ
mulas are widely applied in tracking, navigation, control, commu- 6 7
6 1 ··· 0 0 7
nications, and many other fields due to the seminal work of Kalman Φðk − 1Þ ¼ 6
6 .. .. .. .. 7;
7
(1960). In seismic exploration, the Kalman filter has been applied to 4 . . . . 5
the deconvolution of the seismic record for decades. For instance,
Crump (1974) first uses the discrete Kalman filter successfully for
0 ··· 1 0
the deconvolution of the seismic signals to generate an estimate of (4)
the reflectivity function. Mahalanabis et al. (1981, 1983) propose
P
fast and adaptive Kalman filter algorithms. Using the Kalman filter which means rðkÞ ¼ M i¼1 ½φi ðk − 1Þrðk − iÞ þ uðk − 1Þ; that is,
with different dimensions applied to seismic traces, Sayman (1992) the reflection coefficient at any sample time is assumed to be
uses optimal fixed-interval smoothing filters to attenuate the unex- the sum of a random quantity and a weighted sum of the reflec-
pected high-frequency energy caused by the original Kalman filter. tion coefficients at the M previous samples. Let ϕðk − 1Þ ¼
Song and Lv (2009) discuss in detail how to choose the initial state ½ φ1 ðk − 1Þ · · · φM−1 ðk − 1Þ φM ðk − 1Þ , and then ϕðk − 1Þ
vector and estimate the covariance of the Kalman filter. Kurniadi is a measure of the manner in which the reflection coefficients
and Nurhandoko (2012) give a different state equation by using are expected to change with the subsurface depth. Assigning
wavelet Laplace transforms. In the above approaches based on ϕðk − 1Þ ¼ ½ 0 0 · · · 0  is equivalent to assuming that the
the Kalman filter for seismic deconvolution, almost the same Kal- reflection coefficients are a white random sequence. In measure-
man model was built, and the parameters used were often difficult to ment equation 3, conditioning on the assumption that all traces
choose. are generated by the same wavelet, the observed matrix HðkÞ ¼
In this paper, a new Kalman filter system model for seismic de- ½ bð1Þ bð2Þ · · · bðMÞ  (i.e., the seismic wavelet) can be
convolution is proposed and compared with the conventional model known or estimated in advance; yðkÞ is the observation or measure-
and the Wiener filter. The parameter selection and performance ment; and vðkÞ is the measurement noise, which is also often as-
problems are investigated by simulations on the synthetic seismic sumed as the white noise sequence. Herein, yðkÞ and vðkÞ are
records. Finally, the practical processes on a single-channel field the single sample for a single-channel seismic signal, and this is
seismic record and a real seismic section are made and compared easily extended to the case for a multichannel seismic record.
with the conventional method and the Wiener filter.
New system model of Kalman filter
Back to the seismic signal convolution model (equation 1): Ac-
THEORY AND METHODOLOGY cording to the reciprocity property of convolution, it can also be
Conventional Kalman filter model written as follows:
for seismic deconvolution
X
N
In seismic exploration, a convolutional model (Robinson, 1957) yðnÞ ¼ rðmÞbðn − mÞ þ eðnÞ; n ¼ 1;2; : : : ; N; (5)
is often used to model a reflected seismic trace as follows: m¼1
 Deconvolution using a new Kalman model V33

where rðmÞ ðm ¼ 1;2; : : : ; NÞ is thought of as being invariant and Initialization

bðnÞ is reversed and shifted for different n, different from the case in
equation 1, where bðmÞ ðm ¼ 1;2; : : : ; MÞ is thought of as being Before iterative filtering, the initialization for the Kalman filter is
needed by computing the initial state estimate and the initial covari-
invariant and rðnÞ is reversed and shifted for different n. However,
ance matrix for the state estimate error as follows:
for a single-channel seismic record, the reflectivity function implies
the attributes of the vertical layered earth along the center point be-
x^ ð0Þ ¼ E½xð0Þ;
tween the shot and the geophone. Therefore, the reflectivity se-
quence is deterministic and not changeable in the whole Kalman Pð0Þ ¼ E½ðxð0Þ − Eðxð0ÞÞÞðxð0Þ − Eðxð0ÞÞÞT : (10)
recursive processes sample by sample. As a result, different from
the old Kalman filter system model, a new model can be built as In seismic deconvolution, generally by experience and need for sim-
follows: Let the state vector xðkÞ ¼ ½rð1Þrð2Þ · · · rðNÞT with plification, let
the same length as the seismic record, instead of the same length
as the seismic wavelet used in the old Kalman filter model. This x^ ð0Þ ¼ ½ 0 ··· 0 T ; Pð0Þ ¼ σ 20 IN ; (11)
state vector denotes the desired reflection coefficient sequence.
Based on the assumption that for a single-trace seismic record where IN is the N × N unit matrix and σ 20 is the variance of the re-
all seismic wavelets reflected from different layers are the same, flection coefficient error (here, we suppose that the initial variance
we can establish a new state equation as follows: of every reflection coefficient error is the same).

xðkÞ ¼ xðk − 1Þ; (6) Prediction

For k ¼ 1 to N, in every iteration the state vector, observed vec-
and a measurement equation of
tor, and state error covariance matrix are predicted first. For the new
yðkÞ ¼ WðkÞxðkÞ þ vðkÞ; (7) system model, we have the following equations:

x^ ðk∕k − 1Þ ¼ x^ ðk − 1Þ; (12)

where the observed matrix WðkÞ ¼ ½ wð1Þ wð2Þ · · · wðNÞ 
and vðkÞ is the same as the one in equation 3, that is, the measure-
ment noise. We note that according to the convolution model (see
equation 5) of the reflected seismic signal, WðkÞ is truncated from y^ ðk∕k − 1Þ ¼ WðkÞ^xðk∕k − 1Þ; (13)
the extended and reversed seismic wavelet sequence B shown as
follows: and
 
0 ··· 0 bðMÞ ··· bð1Þ 0 ··· 0 Pðk∕k − 1Þ ¼ Pðk − 1Þ; (14)
B ¼ |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl} ;
N−bM2 c N−dM2 e
where x^ ðk∕k − 1Þ, y^ ðk∕k − 1Þ, and Pðk∕k − 1Þ denote the predic-
(8) tions for time k by means of the estimates at time k − 1.

where bð1Þ · · · bðMÞ can be known or estimated in advance, Gain and correction
and then during the iterations, let
The gain is computed as follows:
 
bðbM2 cÞ bðbM2 c − 1Þ ·· · bð1Þ 0 ·· · 0
Wð1Þ ¼ |fflfflfflfflffl{zfflfflfflfflffl} ; KðkÞ ¼ Pðk∕k − 1ÞWT ðkÞ½WðkÞPðk∕k − 1ÞWT ðkÞþ RðkÞ−1 ;
N−dM2 e
  (15)
Wð2Þ ¼ bðbM2 c þ 1Þ bðbM2 cÞ ·· · bð2Þ bð1Þ ·· · 0 ;
where RðkÞ is the variance of measurement noise vðkÞ, which can be
.. estimated in advance or in real time. Using the gain KðkÞ, the state
.
vector and its error covariance matrix can be corrected as follows:
 
0 · ·· 0 bðMÞ bðM − 1Þ · ·· bðdM2 e − 1Þ
WðNÞ ¼ |fflfflfflffl
ffl {zfflfflfflffl
ffl } . x^ ðkÞ ¼ x^ ðk∕k − 1Þ þ KðkÞ½yðkÞ − y^ ðk∕k − 1Þ (16)
N−bM2 c

(9) and

The reason for arranging WðkÞ in equation 9 is to remove the tran- PðkÞ ¼ Pðk∕k − 1Þ
sient state data at the beginning and end of the convolution. − KðkÞ½WðkÞPðk∕k − 1ÞWT ðkÞ þ RðkÞKT ðkÞ:
Iterative filtering process for Kalman filter (17)
According to the minimum-mean-square errors rule, the Kalman Up to now, an iteration is finished, then k is incremented by 1,
filtering equations can be derived. The Kalman filtering process in- and we go back to the “Prediction” subsection for the next iteration
cludes initialization, prediction, and correction. Prediction and cor- until k ¼ N. The final x^ ðNÞ is the desired output — the estimated
rection are repeated until all of the measurement data are used. reflection sequence.
V34 Deng et al.

Accordingly, for the conventional Kalman filter, the formulas for However, there are four parameters to be preset or estimated step
prediction and correction are as follows: by step with the iterations in the conventional model such as
RðkÞ; σ 20 ; σ 2u ðkÞ and ϕðk − 1Þ. How to optimize these parameters
x^ ðk∕k − 1Þ ¼ Φðk − 1Þ^xðk − 1Þ; (18) is very difficult, and filter trials and parameter testing are standard
practice in seismic processing. Theoretically, RðkÞ; σ 2u ðkÞ and ϕðk −
1Þ can be variant with iteration k to process the time-variant signal.
y^ ðk∕k − 1Þ ¼ HðkÞ^xðk∕k − 1Þ; (19) Their time-variant estimations are troublesome and are not the focus
of this paper. They are often considered invariant for k ¼
1; 2; : : : ; N. However, there are only two parameters in the new
Pðk∕k − 1Þ ¼ Φðk − 1ÞPðk − 1ÞΦT ðk − 1Þ þ Qðk − 1Þ; model such as RðkÞ and σ 20 . Furthermore, if the observed noise
is stationary, that is, RðkÞ is invariant with k (denoted by R), it
(20) can be found that for the new model, the final filtering result
x^ ðNÞ only depends on the parameter ratio λ ¼ R∕σ 20 instead of indi-
vidual parameters R and σ 20 , which can be proved according to equa-
tions 11–17. The detailed proof is given in Appendix A. Therefore,
KðkÞ ¼ Pðk∕k − 1ÞHT ðkÞ½HðkÞPðk∕k − 1ÞHT ðkÞ þ RðkÞ−1 ; for the new model, the number of parameters to select is fewer than
(21) the conventional.

SIMULATIONS ON SYNTHETIC RECORDS

x^ ðkÞ ¼ x^ ðk∕k − 1Þ þ KðkÞ½yðkÞ − y^ ðk∕k − 1Þ; (22)
In this section, seismic resolution is defined first. Then, simula-
and tions and comparisons with the conventional Kalman filter and Wie-
ner filter are presented to investigate the performance of the Kalman
PðkÞ ¼ Pðk∕k − 1Þ filter based on the proposed system model for the seismic decon-
volution. Subsequently, parameter selection for the proposed
− KðkÞ½HðkÞPðk∕k − 1ÞHT ðkÞ þ RðkÞKT ðkÞ; (23) method is discussed by experiments.

where Qðk − 1Þ ¼ gσ 2u ðk − 1ÞgT and σ 2u ðk − 1Þ is the variance of Seismic resolution and reflection coefficients sequence
the random noise uðk − 1Þ in equation 2. After N iterations, every
component of the state vector is actually estimated M times (except Seismic resolution involves the vertical resolution and horizontal
for those near the end of the vector), and the reflection coefficients resolution. Only the vertical resolution is discussed in this paper.
can be obtained by extracting and combining the last estimated Vertical seismic resolution is the minimum vertical distance for
components. which we can discern two wavelets from the top and the bottom
of a thin bed. Once the thickness of a thin bed is less than the res-
Discussion on conventional and new Kalman filter olution limit, the two consecutive seismic wavelets will merge to-
models gether, and they will have the appearance of one seismic wavelet. In
general, the vertical resolution is given to be one quarter of the
In comparison, some differences and improvements can be found wavelength of the seismic wave. Ricker (1953) defines the seismic
in the new system model. First, the length of state vector xðkÞ in the wavelet resolution as follows:
new system model is changed from M (length of the seismic wave-
let) to N (length of the seismic trace). The change is reasonable
1
because the final desired output in the conventional model is also TR ¼ ; (24)
of N-length, and the new state vector adds clarity. In practice, the 3f m
piecewise processing can be used if the length of a seismic trace is
too long, so that computational time can be saved. Second, the state where f m is the dominant frequency of the seismic wavelet.
equation is simplified greatly from equation 2 to equation 6. The For testing and comparing the ability of improving the seismic
original state transition matrix Φðk − 1Þ and the random noise resolution by Kalman filters based on the conventional and pro-
uðk − 1Þ are removed, and the state equation changes into posed models and the Wiener filter, the reflection coefficient se-
xðkÞ ¼ xðk − 1Þ. The new state vector can be thought of as being quence shown in Figure 1a is chosen. Based on the seismic
unchangeable in the whole iterative computations. In fact, for a re- signal convolution model, Figure 1b shows the synthetic noise-free
flected seismic trace, the new state vector represents the reflection record by the Ricker wavelet with a dominant frequency of f m ¼
coefficients corresponding to an overview for a subsurface vertical 30 Hz and a sampling frequency of 1000 Hz. Requiring the length
line of the layered earth. As an objective fact, it is unchangeable in of the synthetic record to remain the same as the reflection coeffi-
the Kalman iterative filtering. In the conventional model, the prob- cients in Figure 1a, the M − 1 transient samples near two ends
lem of optimally estimating Φðk − 1Þ is very difficult to solve. of the convolved signal (M is the length of the seismic wavelet)
Nevertheless, this problem and the estimation for the variance of are removed in all simulations in this paper. The used Ricker wave-
the process noise uðk − 1Þ disappear in the new model, which let is
will make the estimation by the new model easier than that by
bðtÞ ¼ ð1 − 2π 2 f 2m t2 Þe−π
2 f 2 t2
the conventional.
m ; (25)
 Deconvolution using a new Kalman model V35

where f m is the dominant frequency of the Ricker wavelet. In this ½ 1∕M 1∕M · · · 1∕M ; where M equals the number of seismic
case, the seismic wavelet resolution of the synthetic record is wavelet samples (here, M ¼ 61 for the Ricker wavelet with a dom-
inant frequency of 30 Hz). For simplification, the above four param-
1 1 eters are fixed for k ¼ 1; 2; : : : ; N. For the proposed Kalman
TR ¼ ¼ ¼ 11 ms: (26)
3f m 3 × 30 model, let RðkÞ ¼ 0 (for all k); that is, the parameter ratio
λ ¼ 0. For the Wiener filter, the length of the inverse filter is set
There are 11 strong reflection coefficients in Figure 1a. The first as 1000 (the same as in the followed simulations) and the prewhit-
reflection coefficient corresponds to a single Ricker wavelet, and ening factor μ ¼ 0. After the inverse filter is obtained, we convolve
then five pairs of reflection coefficients simulate five thin beds with the inverse filter coefficients with the seismic record and remove the
different thickness. The time intervals between each pair of reflec- transient data at two ends of the convolution result, and then we
tion coefficients are 16, 14, 11, 10, and 8 ms, respectively. For ex- achieve the deconvolution result with length N.
ample, the time interval between the second coefficient and the third Figure 1c–1e shows the deconvolution results in the time domain
(marked by ① in Figure 1b) coefficient is 16 ms, and between the using three methods. We note that all plots in simulations are nor-
sixth and the seventh, it is 11 ms (marked by ③), which is just malized for easy comparison. Normalization means that all of the
the seismic wavelet resolution. In all of the following simulations, data are divided by the maximum absolute value in the data, and then
the reflection coefficient sequence shown in Figure 1a is used to the data are transformed into [−1, 1]. From Figure 1c–1e, the three
synthesize the seismic records. From Figure 1b, the wavelet pair methods can nearly compress every seismic wavelet into the ideal
labeled with ① is not separated very clearly, and neither is ②. In pulse shape. From the normalized spectra shown in Figure 1f–1j,
addition, the wavelet pair labeled as ③ looks like a wavelet, but ac- we can see that the spectrum of reflection coefficients (Figure 1f)
tually they are two wavelets that correspond to two different subsur- is wide enough to be full of the whole Nyquist frequency, and then
face reflected interfaces, and so does ④. The fifth pair of wavelets it becomes very narrow (Figure 1g) after convolution, only distrib-
implies two reflection coefficients with inverse polarities. In this uted in [0, 70] Hz. After deconvolution, the bandwidths are broad-
paper, the Kalman filter is used to try to compress the seismic wave- ened. We can note that the spectrum obtained by the proposed
lets and improve the seismic resolution, ideally obtaining the origi- method (Figure 1j) is closer to that of the reflection coefficient se-
nal reflection coefficient sequence. quence than the others, especially for the frequencies in [0,
30] Hz and [300, 320] Hz shown in the dotted boxes. Therefore,
Comparisons for seismic deconvolution for the noise-free signal, the proposed method works better.
For noise-free seismic record
For stationary noisy seismic record
For the synthetic noise-free seismic record shown in Figure 1b,
the Kalman filters based on the conventional and proposed models Stationary Gaussian white noise is added to the original noise-
and the Wiener filter are applied and compared. For the Kalman free record. We define the peak S/N (PS/N) of the noisy signal
filter, the observed matrix (H or W) uses the real seismic wavelet as follows:
data, which is used to synthesize the seismic record. Similarly, the
real wavelet is also used in the Wiener filter. In this simulation ex-  
A2m
ample, parameters for three methods are set as follows: For the con- PS∕N ¼ 10log10 dB; (27)
ventional Kalman model, let RðkÞ ¼ 0; σ 20 ¼ 1; σ 2u ðkÞ ¼ 10; ϕðkÞ ¼ σ 2n

Figure 1. Deconvolution results for a noise-free seismic record: (a) reflection coefficient sequence, (b) synthetic noise-free seismic record, (c-
e) normalized results by the conventional model, the Wiener filter, and the proposed model, respectively, and (f-j) normalized spectra cor-
responding to (a-e), respectively.
V36 Deng et al.

where Am is the maximal amplitude of the seismic wavelet in the σ 20 ¼ 0.01 and R ¼ 2, σ 20 ¼ 100. We note that this parameter ratio
synthetic noise-free record, and σ 2n is the variance of the added is the same as that used in Figure 2e. We can see that Figure 2e, 2g,
Gaussian white noise. Figure 2b shows a noisy record with PS∕N ¼ and 2n is all the same, which verifies the above-mentioned conclu-
30 dB (Am ¼ 1 and σ 2n ¼ 0.001). By trial-and-error testing, we sion. In the same way, Figure 2f and 2m shows the results obtained
choose the following parameters for three methods to deconvolve by the conventional Kalman filter with the fixed parameter ratio
the noisy record: For the conventional method, let R ¼ 0.02, R∕σ 20 ¼ 0.02 and the other two parameters ðσ 2u ; ϕÞ unchanged,
σ 20 ¼ 1, σ 2u ¼ 1, and ϕ ¼ ½ 1∕61 1∕61 · · · 1∕61 ; for the Wie- but having different individual parameters R ¼ 0.0002, σ 20 ¼
ner filter, let μ ¼ 0.3; for the proposed method, we have λ ¼ 0.02. 0.01, and R ¼ 2; σ 20 ¼ 100, respectively. However, there are great
From Figure 2c–2e, the resolution in three deconvolution results is differences among the three results shown in Figure 2c, 2f, and 2m,
all improved. For example, the two pairs of wavelets labeled by ① even though they are obtained by the same parameter ratio R∕σ 20 ¼
and ② are separated more clearly. The third pair has been divided 0.02 with the other parameters remaining unchanged. That is, the
into two wavelets from “one wavelet,” and so is the fourth pair. The parameters R and σ 20 can individually influence the conventional
fifth pair of wavelets becomes narrower than the original. In the Kalman filter. These comparisons suggest that for the stationary
above-mentioned improvements, the proposed method shows better noisy signal, the proposed method works better and more easily
performance in terms of the obtained waveform, amplitude, and than the conventional Kalman filter and almost similarly to the Wie-
boosted noise than that of the conventional method, and it is almost ner filter.
the same as that of the Wiener filter. For instance, in the region
marked by the dotted box, the noise in Figure 2c is stronger than For nonstationary noisy seismic record
in the others. In addition, the bandwidths of the three deconvolution
results in the frequency domain are broadened from [10, 60] Hz in The Kalman filter can be applied to nonstationary environments
Figure 2i to [0, 100] Hz in Figure 2j–2l. The spectra by the proposed only by using the time-variant parameters without system modifi-
method and the Wiener filter are closer to the part of [0, 100] Hz in cation. However, the Wiener filter cannot be applied. The following
Figure 2h, especially the region in the dotted box. is an example for a nonstationary seismic record. Gaussian white
For verifying the conclusion that for a stationary signal, the pro- noise with different levels is added in the original noise-free record
posed Kalman filter only depends on parameter ratio R∕σ 20 instead to synthesize the nonstationary noisy record shown in Figure 3b.
of individual parameters R and σ 20 , Figure 2g and 2n shows the re- From Figure 3b, the noise becomes stronger and stronger over time.
sults obtained by the proposed method with the fixed parameter ra- For this seismic record, when using the conventional Kalman
tio R∕σ 20 ¼ 0.02 but different individual parameters R ¼ 0.0002, model, let σ 20 ¼ 1, σ 2u ¼ 1, ϕ ¼ ½ 1∕61 1∕61 · · · 1∕61 , and

Figure 2. Deconvolution example for stationary noisy seismic record with PS∕N ¼ 30 dB: (a) reflection coefficient sequence, (b) synthetic
stationary noisy seismic record, (c-e) normalized results by the conventional model, the Wiener filter, and the proposed model, respectively,
(h-l) normalized spectra corresponding to (a-e), respectively, (f) normalized result by the conventional model with R ¼ 0.0002;
σ 20 ¼ 0.01; σ 2u ¼ 1; ϕ ¼ ½ 1∕61 1∕61 · · · 1∕61 , (m) normalized result by the conventional model with R ¼ 2; σ 20 ¼ 100; σ 2u ¼ 1;
ϕ ¼ ½ 1∕61 1∕61 · · · 1∕61 , (g) normalized result by the proposed model with R ¼ 0.0002; σ 20 ¼ 0.01, and (n) normalized result by
the proposed model with R ¼ 2; σ 20 ¼ 100.
 Deconvolution using a new Kalman model V37

RðkÞ be the piecewise time variant; when using the proposed Kal- ary noise, the parameters can be reduced to only one; that is, the
man model, let σ 20 ¼ 1 and have the same RðkÞ as those in the con- parameter ratio λ ¼ R∕σ 20 , which has been proven in the previous
ventional method; let μ ¼ 0.2 when using the Wiener filter. From section. In this section, we discuss the influences of parameter λ and
Figure 3d, the result obtained by the Wiener filter greatly boosts the the used seismic wavelet bðmÞ on the performance of the proposed
noise, especially in [400, 500] ms. Table 1 lists the PS/N before and method by experiments.
after deconvolution by three methods during [0, 120] and [400,
500] ms. In [0, 120] ms, the result obtained by the old Kalman filter Parameter ratio λ
has a lower PS/N than that produced by the other two methods, and
in [400, 500] ms, the Wiener filter works very poorly because of the For a trace of seismic signal with PS∕N ¼ 25 dB (Am ¼ 1 and
very strong noise. In addition, the wavelet pair located at approx- σ 2n ¼ 0.0032), Figure 4 shows the results in the time and frequency
imately 300 ms in Figure 3c (the old Kalman filter) is not separated domains obtained by different λ, which are increased from 0.001 to
as clearly as that in Figure 3e (the new Kalman filter). However, the 1000 at the rate of 10 times. In the time domain, with the increase of
spectrum in Figure 3j is much closer to the original in Figure 3f than λ, the obtained wavelets become smoother, but the resolutions be-
in Figure 3h and 3i. Especially in the range [0, 70] Hz marked by the come lower, which suggests that a bigger λ plays the role of improv-
dotted rectangle, the spectral amplitude in Figure 3i is very weak. ing the S/N, and a smaller λ has the role of improving the resolution.
From the above, the Kalman filter can adapt to the nonstationary At the same time, we find that in the frequency domain, the smaller
noisy signal only by using the time-variant parameters, whereas parameter λ results in a wider bandwidth extension. As a particular
the Wiener filter cannot deal with the nonstationary case once example, the previous experiment for the case of a noise-free seis-
the inverse filter coefficients are obtained. In the future, we will fur- mic record, λ ¼ 0 (gets to the minimum) is chosen, and then the
ther study the time-variant wavelets in the measurement matrix of spectrum is nearly extended to the whole Nyquist frequency (Fig-
the Kalman filter to adapt to the nonstationary signal caused by the ure 1j). For the noisy record, the extension of the bandwidth also
changing propagating seismic wavelets with depth. Similarly, it can makes the noise boosted. For example, the strong noise in Figure 4c
be done without system modification. makes the result unavailable. From the above, a good parameter λ
From the above three simulations, for the stationary seismic rec- should compromise between resolution improvement and noise am-
ord, the new Kalman filter has the advantages of fewer parameters plification. In these experimental results, the result in Figure 4e
and better results than the conventional model, and it works sim- (λ ¼ 0.1) is optimal.
ilarly to the Wiener filter. However, for the nonstationary noisy rec- For investigating how parameter λ will change with different
ord, the proposed Kalman filter, without system modification, can noise levels, we try to find an optimal parameter value for every
obtain higher resolution and PS/N than the conventional Kalman level of noise by a grid search after balancing the resolution and
filter and the Wiener filter. S/N of the deconvolved result. Figure 5 shows the obtained opti-
mum parameters for cases with different stationary Gaussian white
Parameter selection for the proposed model noise levels (individual PS/Ns are also listed near every marker
point). From Figure 5, when the noise is weak (PS∕N > 25 dB),
Parameter selection for the Kalman filter has always been a tricky the optimal parameter changes slowly. However, when the noise
problem. For the proposed model, originally there are two param- becomes stronger, it changes a little faster. In all cases, the optimal
eters such as the observed noise variance RðkÞ and variance of the parameters for signals with the PS/N ½þ∞; 15 dB should almost be
reflection coefficient error σ 20. Furthermore, for the case of station- in [0, 1]. In practice, we can select a roughly corresponding param-

Figure 3. Deconvolution example for a nonstationary noisy seismic record: (a) reflection coefficients sequence, (b) synthetic nonstationary
noisy seismic record, (c-e) normalized results by the conventional model, the Wiener filter, and the proposed model, respectively, and
(f-j) normalized spectra corresponding to (a-e), respectively.
V38 Deng et al.

eter value according to Figure 5, and then we can finely search the rately. Therefore, it is very important to investigate the effects of
optimal parameter value centered on this initial value. different wavelets on the deconvolution results. Wavelet estimation
is another important topic, which is not the focus in this paper. We
Different seismic wavelets bðmÞ try to use different wavelets to do the deconvolution so that some
experience can be obtained. Figure 6 shows a set of experimental
In all of the above experiments, the real seismic wavelet that is results obtained by using different wavelets when a Ricker wavelet
used to synthesize the seismic record is used to do the deconvolu- with a dominant frequency of 30 Hz (Figure 6a) is used to synthe-
tion. In practice, the real seismic wavelet cannot be obtained accu- size a noisy seismic record with PS∕N ¼ 30 dB (Figure 6b). Ricker

Table 1. Comparison for PS/N before and after deconvolution by three methods

PS/N after deconvolution (dB)

Interval (ms) PS/N before deconvolution (dB) Old Kalman filter Wiener filter New Kalman filter

[0, 120] 31.0840 18.3371 20.0148 24.8427

[400, 500] 12.6810 13.4783 8.1777 13.1594

Figure 4. Results obtained by using the proposed method with different parameters λ: (a) reflection coefficients, (b) noisy record
(PS∕N ¼ 25 dB), (c) λ ¼ 0.001, (d) λ ¼ 0.01, (e) λ ¼ 0.1, (f) λ ¼ 1, (g) λ ¼ 10, (h) λ ¼ 100, (i) λ ¼ 1000, and (j-r) normalized spectra
corresponding to (a-i), respectively.
 Deconvolution using a new Kalman model V39

wavelets with different frequencies (Figure 6b–6d) and a normal- seismic record. Similarly, three damped sine waves with different
ized damped sine wave bðtÞ ¼ e−100 t sinð60πtÞ (Figure 6e) are frequencies, which are added by some zeros like the previous ones,
used for deconvolution. Some zeros are added in the damped sine are used for deconvolution. The main lobe of a Ricker wavelet is
wave to make the maximum of the damped sine wave locate in its used in Figure 7e. The deconvolution results shown in Figure 7g–7j
center. Then, the maximums of wavelets after deconvolution can seem better than those in Figure 6g–6j. The resolution is greatly
keep the same positions as those before deconvolution, which is improved. The reason is related to the attribute of the seismic wave-
a good choice when the real wavelet is unknown. From Figure 6, let. The damped sine wave is a minimum-phase wavelet, whereas
we can see that the time- and frequency-domain results (Figure 6h the Ricker wavelet is a zero-phase wavelet. From the above, the
and 6m) using the Ricker wavelet with the right frequency are the seismic wavelet similar to the real seismic wavelet is optimal. If
best; Ricker wavelets with lower or higher frequencies will limit the that is unable to be achieved, the seismic wavelet whose maximum
improvement of resolution. For example, the third and fourth pairs locates its center and whose predominant frequency is similar to that
of wavelets in Figure 6g and 6i are not separated. A similarly poor of the seismic record is preferred.
result is obtained when the damped sine wave is used. Conversely,
Figure 7 shows another set of experimental results obtained by dif- EXPERIMENTS ON FIELD SEISMIC DATA
ferent wavelets when a damped sine wave is used to synthesize a
For a single-channel field seismic record
Figure 8a shows a part of a field seismic record from a geophone,
and the sampling frequency is 1000 Hz. Figure 8b–8d shows the
deconvolution results by the conventional Kalman filter, the Wiener
filter, and the proposed method. First, the dominant frequency of the
signal is computed by fast Fourier transform, and then the Ricker
wavelet with the dominant frequency is used as the seismic wavelet
in the three methods. After repeated attempts, for the conventional
method, let σ 20 ¼ 1; σ 2u ¼ 1; ϕ ¼ ½ 1∕61 1∕61 · · · 1∕61  and
RðkÞ is piecewise time variant because the noise in Figure 8a looks
stronger and stronger over time; for the Wiener filter, the length of
the inverse filter is 1000 and μ ¼ 20; for the proposed method, let
σ 20 ¼ 1 and have the same RðkÞ as those in the conventional method.
From Figure 8, all wavelets after deconvolution almost become
narrower, and each of those wavelets that look like one wavelet indi-
vidually is divided into two or more wavelets. There are some
Figure 5. Optimal parameters λ for cases of different noise levels. differences among results from these three methods. For example,

Figure 6. Results obtained by using different wavelets for a seismic record synthesized with a Ricker wavelet: (a) Ricker wavelet with fm ¼
30 Hz and 61 points used for synthetic seismic record, (b) Ricker wavelet with f m ¼ 20 Hz and 61 points, (c) Ricker wavelet with f m ¼ 30 Hz
and 61 points, (d) Ricker wavelet with f m ¼ 40 Hz and 61 points, (e) damped sine wave (e−100t sinð60πtÞ) with 31 points, (f) synthetic noisy
seismic record (PS∕N ¼ 30 dB), (g-j) normalized results obtained by four wavelets corresponding to (b-e), respectively, and (k-o) normalized
spectra corresponding to (f-j), respectively.
V40 Deng et al.

the wavelets obtained by the conventional method in regions ① and there are five reflections in Figure 10b–10d. Furthermore, the
② have so small of an amplitude that they seem to miss some wave- continuity of the five reflections obtained by the proposed method
lets. The results by the Wiener filter in region ③ involve much more (Figure 10d) is better than those by others.
noise than the other two methods based on the Kalman filter, which In summary, the proposed method works better than the conven-
can also be seen from the spectra. There is much more high-fre- tional Kalman filter and the Wiener filter in real-world seismic data
quency noise in Figure 8g than in Figure 8f and 8h. So for this real processing.
seismic record, the proposed method can obtain higher resolution
and boost less noise than the Wiener filter and the conventional Kal-
man filter.

For a field seismic section

Figure 9a shows a part of a poststack field seismic section in Si-
chuan Province, China. The sampling interval is 2 ms, and the dis-
tance between the two geophones is 25 m. All images in Figures 9
and 10 are drawn in SeiSee software. For all three methods, the
Ricker wavelet with a dominant frequency of 17 Hz is chosen.
For the conventional method, let R ¼ 0.1; σ 20 ¼ 1; σ 2u ¼ 1;
ϕ ¼ ½ 1∕51 · · · 1∕51 ; for the Wiener filter, the length of the
inverse filter is 1000 and μ ¼ 0.5; for the proposed method,
λ ¼ 0.1. Here, the fixed R or λ is selected because the section in
Figure 9a looks stationary. The deconvolution is done trace by trace.
From Figure 9a, some reflected interfaces are very faint and mixed.
However, from Figure 9b–9d, the vertical resolution is improved
greatly after applying the three deconvolution methods. For exam-
ple, in the long strip region between [2300, 2400] ms, the original
interfaces in Figure 9a are too faint to see, whereas there are two
visible interfaces in each of Figure 9b–9d. In those marked regions
including the short strip at approximately 2700 ms, the oval approx-
imately 2500 ms, the square approximately 2600 ms, and near the Figure 8. Example of a single-channel field seismic record:
(a) original single-channel field seismic record, (b-d) normalized
arrow below, the original reflections are mixed and confused, results obtained by the conventional method, the Wiener filter,
whereas they are separated clearly after deconvolution. For exam- and the proposed method, respectively and (e-h) normalized spectra
ple, as shown in Figure 10 (the enlarged square region in Figure 9), corresponding to (a-d), respectively.

Figure 7. Results obtained by different wavelets for a seismic record synthesized with a damped sine wave: (a) damped sine wave with f ¼
30 Hz and 61 points used for a synthetic seismic record, (b) damped sine wave with f ¼ 20 Hz and 107 points, (c) damped sine wave with
f ¼ 30 Hz and 109 points, (d) damped sine wave with f ¼ 40 Hz and 111 points, (e) main lobe (no less than zero) of Ricker wavelet with
f ¼ 30 Hz and 15 points, (f) synthetic noisy seismic record (PS∕N ¼ 30 dB), (g-j) normalized results obtained by four wavelets correspond-
ing to (b-e), respectively, and (k-o) normalized spectra corresponding to (f-j), respectively.
 Deconvolution using a new Kalman model V41

Figure 9. Example on a field seismic section: (a) original field seismic section, (b-d) results obtained by the conventional method, the Wiener
filter, and the proposed method, respectively.

Figure 10. Enlarged square region in Figure 9: (a) original square region and (b-d) results obtained by the conventional method, the Wiener
filter, and the proposed method, respectively.

CONCLUSIONS based on the new model works better than the conventional Kalman
filter and the Wiener filter. Therefore, the Kalman filter based on the
To compress the seismic wavelet and improve the seismic reso- new model can be applied to seismic resolution improvement.
lution, this paper presents a new Kalman filter system model for the
seismic deconvolution. The new model considers that the reflectiv-
ity coefficients are not changeable in the whole iterative computa-
ACKNOWLEDGMENTS
tions of the Kalman filter for a single-channel seismic record. As a
result, the new model has the advantages of simpler equations and This work is supported by the National Natural Science Founda-
fewer free parameters. The theoretical derivation shows that the tion of China (41374114) and the China Scholarship Council. We
number of parameters can be reduced to only one for a stationary thank the associate editor and four anonymous reviewers for their
sequence. A good parameter should be a compromise between res- helpful comments and suggestions, J. Cao and H. Du for providing
olution improvement and noise amplification. Experimental results the field seismic data and kind help, and an anonymous person for
on synthetic and field seismic records show that the Kalman filter the language revision.
V42 Deng et al.

APPENDIX A The values Kð1Þ and P ð1Þ only depend on parameter ratio R∕σ 20
from equations A-3 and A-4. Let λ ¼ R∕σ 20 . Then together with
PARAMETER RATIO FOR A STATIONARY NOISY equations A-1 and A-2, we can conclude that x^ ðNÞ is only related
RECORD to λ, not the individual parameters R and σ 20 .
For stationary observed noise, RðkÞ is invariant with k (denoted
by R). For the kth iteration, let P ðkÞ ¼ PðkÞ
R and then equation 15 REFERENCES
becomes
Coppens, F., and J. L. Mari, 1984, Spectral equalization: A means of im-
proving the quality of seismic data: Geophysical Prospecting, 32, 258–
KðkÞ ¼ RP ðk − 1ÞWT ðkÞ½WðkÞRP ðk − 1ÞWT ðkÞ þ R−1 281, doi: 10.1111/j.1365-2478.1984.tb00731.x.
Crump, N. D., 1974, A Kalman filter approach to the deconvolution of seis-
¼ P ðk − 1ÞWT ðkÞ½WðkÞP ðk − 1ÞWT ðkÞ þ 1−1 mic signals: Geophysics, 39, 1–13, doi: 10.1190/1.1440408.
Kalman, R. E., 1960, A new approach to linear filtering and prediction prob-
(A-1) lems: Journal of Basic Engineering, 82, 35–45, doi: 10.1115/1.3662552.
Kurniadi, R., and B. E. B. Nurhandoko, 2012, The discrete Kalman filtering
approach for seismic signals deconvolution: American Institute of Physics
and equation 17 becomes International Conference Proceedings on Physics and Its Applications
1454, AIP, 91–94.
Lainiotis, D. G., S. K. Katsikas, and S. D. Likothanassis, 1988, Optimal
RP ðkÞ ¼ RP ðk − 1Þ seismic deconvolution: Signal Processing, 15, 375–404, doi: 10.1109/
36.673671.
− KðkÞ½WðkÞRP ðk − 1ÞWT ðkÞ þ RKT ðkÞ; Mahalanabis, A. K., S. Prasad, and K. P. Mohandas, 1981, A fast optimal
deconvolution algorithm for real seismic data using Kalman predictor
model: IEEE Transactions on Geoscience and Remote Sensing, GE-
i:e: ; P ðkÞ ¼ P ðk − 1Þ 19, 216–221, doi: 10.1109/TGRS.1981.350375.
Mahalanabis, A. K., S. Prasad, and K. P. Mohandas, 1983, On the applica-
− KðkÞ½WðkÞP ðk − 1ÞWT ðkÞ þ 1KT ðkÞ: tion of the fast Kalman algorithm to adaptive deconvolution of seismic
data: IEEE Transactions on Geoscience and Remote Sensing, GE-21,
(A-2) 426–433, doi: 10.1109/TGRS.1983.350503.
Margrave, G. F., M. P. Lamoureux, and D. C. Henley, 2011, Gabor decon-
volution: Estimating reflectivity by nonstationary deconvolution of seis-
From equations A-1 and A-2, KðkÞ and P ðkÞ do not involve an mic data: Geophysics, 76, no. 3, W15–W30, doi: 10.1190/1.3560167.
Plataniotis, K. N., S. K. Katsikas, D. G. Lainiotis, and A. N. Venetsanopou-
explicit parameter R; from equation 16, x^ ðkÞ is related to KðkÞ, los, 1998, Optimal seismic deconvolution: Distributed algorithms: IEEE
which depends on P ðk − 1Þ. From iteration N back to iteration Transactions on Geoscience and Remote Sensing, 36, 779–792, doi: 10
1, we have the following equations: .1109/36.673671.
Rice, R. B., 1962, Inverse convolution filters: Geophysics, 27, 4–18, doi: 10
.1190/1.1438979.
Ricker, N., 1953, Wavelet contraction, wavelet expansion, and the control of
Kð1Þ ¼ P ð0ÞWT ð1Þ½Wð1ÞP ð0ÞWT ð1Þ þ 1−1 seismic resolution: Geophysics, 18, 769–792, doi: 10.1190/1.1437927.
 −1 Robinson, E. A., 1957, Predictive decomposition of seismic traces: Geo-
σ2 σ2
¼ 0 I N WT ð1Þ Wð1Þ 0 I N WT ð1Þ þ 1 physics, 22, 767–778, doi: 10.1190/1.1438415.
R R Robinson, E. A., and S. Treitel, 1967, Principles of digital Wiener filtering:
  Geophysical Prospecting, 15, 311–332, doi: 10.1111/j.1365-2478.1967
T T R −1 .tb01793.x.
¼ IN W ð1Þ Wð1ÞIN W ð1Þ þ 2 ; (A-3) Sayman, A., 1992, Application of Kalman filter to synthetic seismic traces:
σ0 Jeofizik, 6, 67–75.
Song, W. Q., and S. C. Lv, 2009, Seismic deconvolution method based on
Kalman optimal estimation theory: Geophysical Prospecting for Petro-
and leum, 48, 342–346, doi: 1000-1441(2009)04-0342-05.
Ulrych, T. J., 1971, Application of homomorphic deconvolution to seismol-
ogy: Geophysics, 36, 650–660, doi: 10.1190/1.1440202.
P ð1Þ ¼ P ð0Þ − Kð1Þ½Wð1ÞP ð0ÞWT ð1Þ þ 1KT ð1Þ Wang, L. L., J. H. Gao, W. Zhao, and X. D. Jiang, 2013, Enhancing reso-
  lution of nonstationary seismic data by molecular-Gabor transform: Geo-
σ2 σ2 physics, 78, no. 1, V31–V41, doi: 10.1190/geo2011-0450.1.
¼ 0 IN − Kð1Þ Wð1Þ 0 IN WT ð1Þ þ 1 KT ð1Þ: Wang, Y. H., 2006, Inverse Q-filter for seismic resolution enhancement:
R R Geophysics, 71, no. 3, V51–V60, doi: 10.1190/1.2192912.
Wiener, N., 1949, Extrapolation, interpolation, and smoothing of stationary
(A-4) time series: John Wiley and Sons.

View publication stats