Robust principle component analysis (RPCA) for seismic

data denoising
Jinkun Cheng, Ke Chen and Mauricio D. Sacchi
Department of Physics, University of Alberta

Summary

Principal component analysis (PCA) is an effective tool for random noise attenuation. It has been
widely used in seismic data processing for the enhancement of the signal-to-noise ratio of seismic
data. However, PCA lacks robustness to outliers and, therefore, its applications to random noise
suppression has limitations. We present a robust PCA (RPCA) method that can be utilized in the
frequency-space domain to eliminate erratic noise. The method adopts a nuclear norm constraint
that exploits the low rank property of the desired data while using an `1 norm constraint to properly
estimate erratic (sparse) noise. The problem is then tackled via the first-order gradient iteration
with two steps of soft-thresholding. We use synthetic examples to illustrate the effectiveness of the
algorithm.

Introduction

Seismic data are always contaminated with noise. Therefore, signal-to-noise ratio enhancement
plays an important role in seismic data processing. A variety of methods that exploit the differences
between signal and noise have been developed. Noise removal techniques based on principal com-
ponent analysis (PCA) are of special interest to this article. Ulrych et al. (1999) illustrated how matrix
rank reduction methods can be utilized to eliminate incoherent noise from seismic records in time
domain. A related family of methods, the Karhunen-Loeve transform, has been introduced for the
enhancement of signal-to-noise ratio (Al-Yahya, 1991) as well. To handle dipping events, Trickett
(2003) proposed the f − x − y eigenimage filtering that applies matrix rank reduction to frequency
slices. Cadzow de-noising (Trickett and Burroughs, 2009), or singular spectrum analysis (SSA) (Sac-
chi, 2009; Oropeza and Sacchi, 2011) also operates in frequency space domain. However, Cadzow
denoising does not operate on the f − x data themselves but on Hankel matrices that are formed from
data in f − x domain.

Although the aforedescribed rank reduction methods are very effective for attenuating the random
Gaussian noise, their application to real data problems can be limited because they lack robustness
to erratic noise. In seismic data processing, erratic noise includes swell noise, power line noise and
artifacts caused by glitches in the recording instrument. Outliers tend to manifest as high-amplitude
isolated signals that do not obey the Gaussian distribution. Therefore, the conventional least squares
error criterion utilized by PCA will perform poorly (Golub and van Loan, 1996; Trickett et al., 2012;
Chen and Sacchi, 2014).

In this article, we present an algorithm named robust PCA (RPCA) to suppress erratic noise and
blending noise in simultaneous sources acquisition. We consider the minimization of a cost function
that combines a weighted nuclear norm and an `1 . The problem is then tackled via the gradient
iteration with two steps of soft-thresholding.

GeoConvention 2015: New Horizons 1

Theory

We consider to apply the proposed algorithm on a 3D seismic volume D(t, x, y). The Fourier transform
can be adopted to map data from t − x − y domain to ω − x − y domain. At a given frequency ω0 ,
the spatial data, D(ω0 − x − y), can be denoted via a matrix D(x, y). We can remove ω0 from the
analysis knowing that the algorithm operates for all frequency slices. As the desired signal is coherent
along two spatial directions, it can be estimated via a low rank matrix X. The observed data can be
expressed via
D = X + E, (1)
where E is a sparse matrix corresponding to the additive erratic noise. The RPCA method suggests
the following problem. Find X such is of low rank and ||E||0 is minimum. Where ||E||0 is the `0 norm
of E means the number of non-zero element in E. The latter is an NP-hard problem. In order to make
the problem tractable, we consider to the `1 norm, which is defined by the summation of absolute
values of the matrix elements, to replace the `0 norm and to use the nuclear norm which is defined
as the summation of singular values of the matrix, to replace the low rank constraint. The `1 norm is
the tightest convex relaxation of the `0 norm and the nuclear norm is the tightest convex relaxation
to the low rank constraint, respectively. We also introduce further relaxation with a Frobenius norm
constraint, ||D − X − E||2F , to tolerate the inclusion of Gaussian noise in the seismic data. The resulting
cost function is as follows
1
min J = ||D − X − E||2F + λ ||E||1 + ||X||∗ , (2)
2µ

where λ is a trade-off parameter that balances the sparsity and low rank constraints. The scalar µ is
a small constant that controls the inclusion of Gaussian noise.

The alternating first-order gradient method is utilized to estimate the low-rank data X as well as the
sparse matrix that represents the erratic noise E. We consider to minimize the cost function via an
iterative scheme. For this purpose we adopt a sub-gradient optimization technique (Zhou et al., 2010)

Xk+1 = min ||X||∗ + ||X − X̂k ||2F

(3)
Ek+1 = min λ ||E||1 + ||E − Êk ||2F ,

where X̂k and Êk are given by the gradient of the separable quadratic system

1
X̂k = Xk − (Xk + Ek − D)
2µ
(4)
1
Êk = Ek − (Xk + Ek − D) .
2µ

The two sub-problems defined by equation (3) are commonly seen in the field of compressive sensing
and matrix completion, respectively. The matrix Ek+1 is typically given by soft-thresholding the entries
of Êk as follows
λµ
Ek+1 (i, j) = max(|Êk (i, j)| − , 0) , (5)
2
where Êk (i, j) denotes the element in matrix Êk . Similarly, Xk+1 can be computed by applying soft-
thresholding to the singular values of X̂k (Fazel, 2002; Recht et al., 2010) as follows
µ
Σ̂(i, i) = max(|Σ(i, i)| − , 0) , (6)
2

where we assume X̂k = UΣV∗ and Xk+1 are then recovered by the new set of singular values Σ̂(i, j).
The resulting algorithm is summarized in Algorithm (1). In each iteration, we modify the current

GeoConvention 2015: New Horizons 2

estimate of the low-rank data and erratic noise in the opposite direction to the gradient of the quadratic
term. Then we apply two steps of soft-thresholding to the modified estimators . The convergence
of this algorithm is very similar to the convergence of the FISTA algorithm utilized for l1 − l2 inverse
problems (Zhou et al., 2010; Beck and Teboulle, 2009).

Algorithm 1 RPCA
Inputs:
Spectral matrix D, trade-off parameter λ and stopping criterion ε
Initialize:
X0 = 0; E0 = 0; k = 1
repeat
1
X̂k = Xk − 2µ (Xk + Ek − D)
1
Êk = Ek − 2µ (Xk + Ek − D)
[U, Σ, V] = svd[X̂k ]
Σ̂(i, i) = max(|Σ(i, i)| − µ2 , 0)
Xk+1 = UΣ̂V∗
Ek+1 (i, j) = max(|Êk (i, j)| − λ2µ , 0)
k = k+1
until kD − Xk − Ek k2F < ε

Examples

To test the effectiveness of the proposed algorithm, we first created a synthetic seismic section with
3D linear events (Figure 1a). We then added random, spike-like noise to mimic blending noise gener-
ated by simultaneous source shooting (Figure 1b). Figure (1c) shows the de-noising result and Figure
(1d) shows the error of the estimator. The proposed method effectively suppressed the incoherent
noise. We improve the the signal-to-noise ratio of data from −1.2 dB to a factor of 11.9 dB.

We also tested the algorithm on one common receiver gather of a synthetic 3D VSP data set. Figure
(2a) shows the centre shot line of one common receiver gather without noise. Figure (2b) shows the
same shot line contaminated with noisy observations. The traces are corrupted with erratic noise (the
amplitude of the erratic noise are about 3 times the maximum amplitude of the reflections). Figure
(2c) shows the resulting de-noised shot line with the proposed method. As a result, both Gaussian
and erratic noise were effectively removed. Through RPCA de-noising, we improve the quality of data
from −8.2 dB to a factor of 12 dB.

Conclusions

We presented an algorithm for suppressing erratic noise via a rank-reduction method. The algorithm
relies on the low rank approximation of the spatial data at a given monochromatic frequency in the
f − x domain domain. A nuclear norm constraint for the data, as well as an l1 norm constraint for
the sparse erratic noise have been utilized to design the cost function of the problem. Through tests
with synthetic examples, we show that the proposed algorithm can be utilized for deblending and for
suppressing the erratic noise.

Acknowledgements

The authors are grateful to the sponsors of Signal Analysis and Imaging Group (SAIG) at the Univer-
sity of Alberta.

GeoConvention 2015: New Horizons 3

(a) (b) (c) (d)
hx (m) hx (m) hx (m) hx (m)
200 200 200 200
0 0 0 0

0.2 0.2 0.2 0.2

0.4 0.4 0.4 0.4

t(s)

t(s)

t(s)

t(s)
0.6 0.6 0.6 0.6

0.8 0.8 0.8 0.8

Figure 1 (a)Synthetic data with linear events in CMPx gather 15. (b) Synthetic data with 3 linear
events contaminated with randon, spike-like noise to mimic the pseudo-deblended CMP gather. (c)
CMPx gather 15 after RPCA de-noising, the SNR has been improved by 11.9 dBs. (d) The estimation
error between section (a) and (c).

(a) Source x position (km) (b) Source x position (km)

8 9 10 8 9 10
0 0
t(s)

t(s)

2 2

(c) Source x position (km) (d) Source x position (km)

8 9 10 8 9 10
0 0
t(s)

t(s)

2 2

Figure 2 (a) Centre shot line of the ideal clean common receiver gather of the 3D VSP data set. (b)
The centre shot line of the common receiver gather cotaminated with erratic noise. (c) The centre
shot line of the common receiver gather after RPCA de-noising. (d) The difference between section
(a) and (c).

GeoConvention 2015: New Horizons 4

References
Al-Yahya, K., 1991, Application of the partial Karhunen-Loéve transform to suppress random noise in seismic sections:
Geophysical Prospecting, 39, 77–93.
Beck, A. and M. Teboulle, 2009, A fast iterative shrinkage-thresholding algorithm for linear inverse problems: SIAM J.
Imaging Sciences, 2, 183–202.
Chen, K. and M. D. Sacchi, 2014, Robust reduced-rank filtering for erratic seismic noise attenuation: Geophysics, 80,
V1–V11.
Fazel, M., 2002, Matrix rank minimization with applications: Master’s thesis, Elec. Eng. Dept, Stanford University.
Golub, G. and C. van Loan, 1996, Matrix computations: London: The Johns Hopkins University Press, third edition edition.
Oropeza, V. and M. Sacchi, 2011, Simultaneous seismic data denoising and reconstruction via multichannel singular
spectrum analysis: Geophysics, 76, V25–V32.
Recht, B., M. Fazel, and P. Parrilo, 2010, Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm
minimization: SIAM Review, 52, 471–501.
Sacchi, M. D., 2009, F-x singular spectrum analysis: CSPG CSEG CWLS Convention, 392–395.
Trickett, S., L. Burroughs, and A. Milton, 2012, Robust rank-reduction filtering for erratic noise: 82nd Annual International
Meeting, SEG, Expanded Abstracts, 1–5.
Trickett, S. R., 2003, F-xy eigenimage noise supression: Geophysics, 68, 751–759.
Trickett, S. R. and L. Burroughs, 2009, Prestack rank-reducing noise suppression: Presented at the SEG Technical Program
Expanded Abstracts.
Ulrych, T. J., M. D. Sacchi, and S. L. M. Freire, 1999, Covariance analysis for seismic signal processing, chapter Eigenimage
processing of seismic sections, 241–274.
Zhou, Z., X. Li, J. Wright, E. Candes, and Y. Ma, 2010, Stable principal component pursuit: Information Theory Proceedings
(ISIT), IEEE International symposium, 1518–1522.

GeoConvention 2015: New Horizons 5