You are on page 1of 90

Quality factor inversion applied to 3D

seismic data in Colombia

Jorge Luis Ñustes Andrade


201424022
Department of Geosciences
Universidad de los Andes

This dissertation is submitted for the degree of


BSc in Geosciences

December 2017
For my loving parents.
Declaration

I declare that except where specific reference is made to the work of others, this thesis is an
original report of my research and has not been submitted for any previous degree in this, or
any other university.

Jorge Luis Ñustes Andrade


201424022
December 2017
Acknowledgements

I would like to acknowledge Dr Jean Baptiste Tary for helping me through all the process of
my research, and giving me important reviews of my thesis. I would also like to acknowledge
HOCOL S.A and especially to MSc Jaime Checa for believing on my research, and giving
me the seismic data used in this work. Special thanks to Halliburton for allowing me to
use their ProMax seismic proccessing software. I wish to thank MSc Camilo Gonzalez from
Halliburton for his invaluable help with the preprocessing of the data and for writing a
Matlab code to split the segy files. Finally, thanks to Dr Carl Reine for helping me with all the
questions I had through my investigation, he is the real inspiration behind this work.
Abstract

Seismic attenuation, usually quantified by the dimensionless parameter Q, contains valu-


able information of the petrophysical parameters of the subsurface, and presents a valuable
tool for reservoir characterization. Yet, due to the number of issues that must be addressed
to obtain reliable measurements, it is rarely calculated.

In this work I present a robust method to calculate attenuation on the τ − p domain from
prestack CMP gathers. The PSQI method (Reine, 2009), uses a variable-window time-
frequency transform, a simultaneous inversion scheme, and a τ − p domain transform
to get an accurate attenuation measurement in the form of 1/Q.

To test the robustness of the method, I apply the PSQI to a 3D dataset from Colombia.
Before doing the inversion, I describe a step by step preproccesing that must be done to
improve the attenuation measurements, and reduce the uncertainty of the inversion. The
final result proves the reliability of the PSQI method.
Abstract

La atenuación sísmica, generalmente cuantificada por medio del parámetro adimensional


Q, contiene información valiosa de los parámetros petrofísicos del subsuelo y representa
una herramienta valiosa para la caracterización de reservorios. Sin embargo, debido a la
cantidad de problemas que se deben considerar para obtener mediaciones confiables, rara
vez se calcula.

En este trabajo presento un método robusto para calcular la atenuación en el dominio


τ − p a partir de un arreglo CMP. El método PSQI (Reine, 2009), usa una transformada
tiempo-frequencia de ventana variable, un esquema de inversión simultánea, y una trans-
formación al dominio τ − p para obtener una mediación de atenuación precisa en la forma
de 1/Q.

Para probar la solidez del método, apliqué el procedimiento PSQI a un conjunto de datos
sísmicos 3D en Colombia. Antes de llevar a acabo la inversión, describo el paso a paso del
preprocesamiento que se debe realizar para mejorar las mediciones de atenuación y reducir
la incertidumbre de la inversión. El resultado final demuestra la fiabilidad del método PSQI.
Contents

List of Figures xvi

List of Tables xix

1 Introduction 1
1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Seismic attenuation 4
2.1 Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Attenuation and Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.2 Dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.3 Effective Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Types of Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 Intrinsic Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Apparent Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Methods to Calculate Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.1 Time domain methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.2 Frequency domain methods . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 The Spectral Ratio Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 Surface Seismic Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Pre-stack Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 Spectral Analysis 12
3.1 Non stationary signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Heisenberg-Gabor Uncertainty Principle . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Time-frequency Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3.1 Short-time Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.2 Gabor transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.3 S transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
xiv Contents

3.4 Synthetic examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17


3.4.1 Analysis of results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.4.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4 Inversion Scheme 22
4.1 Simultaneous inversion scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.1 Weighting functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.2 Weighted inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5 τ − p Domain 28
5.1 The τ − p transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.2 Artifacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1.3 Moveout and anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

6 Multimensional Interpolation 33
6.1 Seismic data interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.2 MWNI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.3 POSC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6.3.1 5D Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

7 The Prestack Q-Inversion Method 36


7.1 Preproccesing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
7.1.1 Static Corrections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.1.2 Velocity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.1.3 Band-Pass Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.1.4 Amplitude Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.1.5 5D Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.1.6 τ − p Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.2 Calculate Moveout Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.2.1 Interval Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.2.2 Equivalent Traveltime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.3 Spectral Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.3.1 Data Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.3.2 Spectral information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7.4 Inversion for 1/Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Contents xv

8 3D Data Example 44
8.1 Background Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.1.1 Geological Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.1.2 Survey Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.2 PSQI Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.2.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8.2.2 Boundary of Interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.2.3 Calculate Moveout Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
8.2.4 Spectral Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
8.2.5 Inversion for 1/Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
8.3 PSQI Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
8.3.1 Weighted . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
8.3.2 Unweighted . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

9 Conclusions 60
9.1 Recommended future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

Bibliography 63

Appendix A Other methods to measure Q 68


A.1 Risetime Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

Appendix B Multimensional algorithms 69


B.1 MWNI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
List of Figures

2.1 Classification scheme for common attenuation processes.From (Liner, 2012). 6


2.2 Schematic diagram showing the natural log spectral ratio data as a function of
frequency, from (Reine, 2009). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.1 Sampling of the time-frequency plane. Different forms of sampling: Shannon,


Fourier, Gabor, Wavelet , modified from (Flandrin, 1998). . . . . . . . . . . . . 15
3.2 The Short time Fourier transform, from (Gröchenig, 2013). . . . . . . . . . . . 16
3.3 Impulse signal (a) with the time-frequency representations of STFT (b) and
S-transform (c). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4 Sine signal (a) with the time-frequency representations of STFT (b) and S-
transform (c). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 Chirp signal (a) with the time-frequency representations of STFT (b) and S-
transform (c). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.1 Schematic drawing of the linear relationship of the natural log spectral ratio
data for all traces. The blue osculations represent some random noise. . . . . 23
4.2 Another visualization of equation 4.2 but in the ∆t ω space. Note that attenua-
tion causes the surface to descend as one coordinate increases relative to the
other. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.3 Solution to 4.6 with a variable intercept. Shows undulations in the data along
the ∆t coordinate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

5.1 Mapping of hyperbolic reflections in the t − x domain shown as ellipses in the


τ − p domain, and linear events like refractions in the t − x domain shown as
points in the τ − p domain. The vertical axis of the right figure is the intercept
time τ. From (Reine, 2009). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

7.1 Flow diagram of the PSQI process. Modified from (Reine, 2009). . . . . . . . . 37
List of Figures xvii

7.2 The horizontal slowness of the data must be truncated at the slowness of
refraction or maximum offset for the lowest interface. Here, the truncation
shown in gray occurs at the refraction of the red horizon. From (Reine, 2009). 42

8.1 Location map of the Sinú-San Jacinto Basin. From (Sánchez and Permanyer,
2006). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.2 Foldmap of the 3D survey. The black-dotted rectangle shows the subset where
the inversion was done. See that most of the fold values are consistent within
the subset, averaging a fold of 30. . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.3 Stratigraphic column of the Sinú-San Jacinto Basin. From (Barrero et al., 2007). 46
8.4 Raw CMP with static corrections applied. Look at the high-amplitude surface
waves indicated in red. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8.5 The average spectrum for the entire CMP (green), the surface waves (red), and
the reflections (blue). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
8.6 Band-pass filtered CMP gather. The surface waves, shown by the red triangle,
were effectively removed. The high frequency noise present on various traces
has also been removed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
8.7 The CMP gather from figure 8.4 after the multidimensional interpolation has
been applied. Noisy traces were removed, and anomalous amplitudes were
corrected. All this processes are necessary to reduce the artifacts during the
τ − p transform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8.8 (a) shows the offset distribution of a raw CMP gather and (b) shows the in-
terpolated offset distribution of that same CMP. Offset spacing in the original
data is regularized after the 5D interpolation. Moreover, offsets lower than 600
m (Including offset 0 m) that were not present in the raw data are created with
the interpolation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.9 τ − p transform of the raw data. The horizontal slowness values are in ms/m 51
8.10 τ − p transform of band-pass filtered data. The horizontal slowness values are
in ms/m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.11 τ − p transform of the interpolated data. The horizontal slowness values are in
ms/m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.12 Pre-stack time migration amplitude map. The circled zone indicates the area
of interest where high amplitude values are observed. The large amplitudes
are often associated with gas. The two horizons used for the inversion are
labeled as A and B. Courtesy of Hocol SA & PetroSeis LTDA . . . . . . . . . . . 53
8.13 Two-way travel times for (a) horizon A, and (b) horizon B. Time is in seconds. 53
xviii List of Figures

8.14 Stacking velocity profile (a), and its interval stacking velocity profile (b) calcu-
lated using Dix’s equation. Values are shown in m/s. . . . . . . . . . . . . . . . 54
8.15 CMP gather in the τ − p domain as (a) amplitude data, and (b) instantaneous
amplitude data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.16 Maximum horizontal slowness trace for horizon B . . . . . . . . . . . . . . . . 55
8.17 From left to right, the stacking velocity, interval stacking velocty, and critical
angle for each time sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
8.18 Natural log spectral ratio surface between horizons A and B. . . . . . . . . . . 56
8.19 The weighted PSQI map of 1/Q between horizons A and B is shown on (a). The
frequency bandwidth used is 15 Hz to 75 Hz, and the colorbar is clipped below
0. The uncertainty map of the weighted inversion between horizons A and B is
given in (b). (c) shows the unweighted PSQI map of 1/Q between horizons A
and B with the same frequency bandwidth used in the weighed inversion. The
uncertainty map of the unweighted solution is shown in figure (d). . . . . . . . 58
8.20 The filtered PSQI calculations of weighted 1/Q for the interval between hori-
zons A and B is shown in (a). likewise, the filtered PSQI calculations of the
unweighted 1/Q solution is shown in (b). The 1/Q color bar is clipped below
zero for both results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
List of Tables

2.1 Spectral amplitude ratios for determination of attenuation, modified from


(Bath, 1974) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Chapter 1

Introduction

Reliable measurements of seismic attenuation, usually presented as 1/Q, are desirable


for improving resolution of signal and reservoir characterization. As higher frequencies
attenuate faster than lower frequencies, dominant signal wavelength and period lowers
as the seismic wave propagates. Attenuation can be measured either in the time domain
(amplitude decrease) and the frequency domain (changes in frequency content). In both
cases, it is accompanied with a decrease in resolution (Zhang and Ulrych, 2002).

The attenuation coefficient is linked closely to petrophysical parameters that can be used in
reservoir characterization. For example, Winkler and Nur (1982) established that attenuation
may serve as an indicator of permeability, mobility of fluids, and fluids saturation. Depend-
ing on the properties of the fluid, such as viscosity, compressibility, and on the properties
of the pores, like porosity and permeability, amplitude losses due to attenuation may be
affected in different ways. For hydrocarbon exploration, it has been shown that attenuation
is highly dependent on the gas content in saturated rocks (White, 1975). Furthermore, at-
tenuation is sensitive to fractures and mobility of fluids, and can be used to monitor fluid
dynamics in a reservoir (Macride and Kanasewich, 1987).

Knowledge of quality factor is very desirable, yet due to the number of issues that must be
addressed to obtain reliable measurements, it is rarely calculated. Post-stacked is appealing
for Q factor calculation because of the optimum Signal to Noise ratio and the apparent zero-
offset. However, in post-stack data, each of the individual stacked traces have a different
ray-path for every case except for zero offset (Dasgupta and Clark, 1998). Additionally, for a
given reflection event, raypath geometries, spectral distortions, and angle dependent effects,
corrupt the measurement of attenuation and make estimation of Q from surface seismic
stacked data potentially erroneous.
2 Introduction

One method was proposed by Dasgupta and Clark (1998) to improve the measurements
of attenuation using surface pre-stack seismic data. The Q versus offset method (QVO),
exhibits real enhancements in comparison to other pre-stack and post-stacked common
methods used to calculate attenuation. Moreover, other studies have shown the applications
of QVO to determine the azimuthal discrimination of fractures and time-lapse changes in
anisotropic media (Clark et al., 2009). Despite the enhancements, attenuation measure-
ments obtained by the QVO method are still corrupted by the effects of spectral interference,
directivity and, anisotropy, making this technique less reliable for a robust calculation of the
attenuation factor.

A novel method to calculate attenuation from surface pre-stacked data was introduced
by Reine et al. (2012a) and proved to be better than other previous methods like QVO.
The pre-stack Q inversion (PSQI), uses a variable-window time frequency transform and
a scheme to invert for 1/Q from natural log spectral-ratio data, to overcome the effects
induced by spectral interference. Additionally, authors take the data to the τ − p domain
to reduce the angle-dependent effects like directivity, anisotropy, and raypath differences.
Reine et al. (2012a)

Reine et al. (2012a), shown that the combination of these techniques provides accurate
calculation of the quality factor by combining Q estimates using synthetic data and real 3D
data with VSP measurements. Results show the robustness of this new method to address
the main complications to calculate a precise value for the Q factor.

1.1 Objective
The main purpose of this research is to apply the PSQI method on 3D seismic data from
Colombia. To do that, a series of processing steps must be followed to get the precise value of
quality factor from pre-stack CMP surface seismic data. Considering that one of the reasons
of calculating attenuation, in the form of 1/Q, is to predict the properties of subsurface,
there are three major components that must be executed to achieve the expected results:

1. Transform common midpoint (CMP) data into the τ − p domain.


1.1 Objective 3

2. Calculate time-frequency transforms of the data using a variable-window time-frequency


transform.

3. Perform simultaneous inversion of the natural log spectral ratio data with respect to
frequency and time difference to get attenuation in the form of 1/Q.

In addition to the application of the PSQI method to seismic data, this research also includes
the application of pre-proccessing techniques required to obtain reliable measurements of
the attenuation factor. Furthermore, this thesis aims to provide significant results that can
be use to characterize the potential of a hydrocarbon reservoir in the area of study.
Chapter 2

Seismic attenuation

2.1 Q

2.1.1 Attenuation and Q


Q stands for quality factor. It is a dimensionless parameter that measures how much me-
chanical energy is converted to heat and fluid flow as a seismic wave propagates away
from its source (Costain and Çoruh, 2004; Müller et al., 2010). When the restoring force is
proportional to the amplitude of vibration and the dissipative force is proportional to the
velocity, Q can be defined as:
1 ∆E
=− (2.1)
Q(ω) 2πE
Where E is the elastic energy stored in the specimen when stress is maximum, and −∆E is
the energy dissipated in a specimen through a stress cycle (Aki and Richards, 2002).

In a medium with linear stress-strain relation, the measured amplitude of a wave A is pro-
1
portional to E 2 . By replacing the energy terms in equation (2.1) and assuming that Q ≫ 1,
the relationship can be expressed in a way that we can obtain the amplitude fluctuations
due to attenuation. Hence:
1 1 ∆A
=− (2.2)
Q(ω) π A
In order to observe the spatial decay of A we will assume that the direction of propagation
and maximum attenuation will occur along the x axis. Then, we can write (2.2) as:

d A ∆A
= (2.3)
dx λ
2.1 Q 5

Where λ is the wavelength given in terms of ω and phase velocity c by:

2πc
λ= (2.4)
ω

Replacing (2.4) on (2.3) we ended up with the exponential solution of A(x):

ωx
− 2cQ
A(x) = A 0 e (2.5)

The exponential term in (2.5) is known as the attenuation coefficient α, a quantity which
1
measures energy absorption, and is related to Q by:

ω
α(ω) = (2.6)
2c(ω)Q(ω)

Several attenuation measurements have been done in the laboratory over a wide range of
frequencies (Knopoff, 1964; Toksöz et al., 1979). The results showed that there exists a linear
relationship between attenuation coefficient and frequency. This leads to a quality factor Q
that is independent of frequency. Equation (2.6) can be simplified to:

ω
α= (2.7)
2cQ

Hence (2.5) is simply:


A(x) = A 0 e −αx (2.8)

2.1.2 Dispersion
Body wave dispersion can be defined as the dependence of body wave velocity on frequency
(Costain and Çoruh, 2004). This dependence causes each frequency component ω to travel
at a different phase velocity c. As the higher frequencies attenuate faster than the lower
frequencies, the wavelet changes its shape. This change in shape must be such that the
wavelet remains causal. It has been proved that for a linear attenuation theory the existence
of absorption implies some dispersion (Aki and Richards, 2002). Moreover, experiments done
by Futterman (1962) proved that phase shift of the wave is a direct cause of the dispersion
that guarantees a causal arrival of the signal. Depending on the attenuation model used, the
dispersion equation varies. Futterman (1962) introduced a dispersion relation given by:

ω ω
· µ ¶¸
1
k= 1− ln (2.9)
c0 πQ ω0

Where c 0 is the phase velocity at a reference frequency ω0 .


6 Seismic attenuation

2.1.3 Effective Q
In order to compute attenuation from real seismic data, effects like source spectrum, geomet-
rical spreading, anisotropy, scattering, and noise must be separated in advance (Tonn, 1991).
However some of them are difficult to handle, and often they cause a distorted measurement
of Q. This corrupted measure is known as the effective Q e , and can be defined as the sum of:

1 1 1
= + (2.10)
Qe Qi Q a

1 1
Where Qi is the intrinsic component and Qa is the apparent attenuation (Spencer et al.,
1982).

2.2 Types of Attenuation


Attenuation can be caused by a variety of physical phenomena that can be divided broadly
into elastic and inelastic processes. Figure 2.1 shows a scheme with the most common
causes of attenuation.

Figure 2.1 Classification scheme for common attenuation processes.From (Liner, 2012).

2.2.1 Intrinsic Attenuation


Intrinsic or inelastic attenuation (1/Q i ) is the energy loss in waves from the conversion of
the wave’s energy into heat and fluid flow (Raji and Rietbrock, 2013). It is mainly caused by
the presence of fluids in the pore space of rocks by a mechanism known as wave-induced
fluid flow (Müller et al., 2010). As a compressional wave propagates through a porous media,
a fluid pressure gradient occurs between the peaks and troughs of the wave, causing the
fluid to move from the peaks to the troughs until reaching an equilibrium state (Pride, 2005).
2.3 Methods to Calculate Q 7

This macroscopic flow is the main source of intrinsic attenuation described by the equations
of porous media acoustics introduced by Biot (1956a,b).

2.2.2 Apparent Attenuation


Contrary to intrinsic attenuation, apparent or elastic attenuation does not involve any lost
of energy to the wave. It is caused by the layered structure of the earth, and mainly involves
a redistribution of the energy by processes such as reflection, transmission, multiples,
scattering, and mode conversion (Liner, 2012).

2.3 Methods to Calculate Q


Tonn (1991) purposed 10 different methods for the computation of the quality factor Q.
These methods are separated into those in the time domain, and those in the frequency
domain. In this section I will briefly explain some of those methods, and then make a detail
explanation of the Spectral Ratio method which is the one used in the PSQI calculations.

2.3.1 Time domain methods


One of the simplest methods to compute Q in the time domain is the Amplitude decay
method. This method is based on the decay of amplitudes between two different distances
or times. For this method Q is defined as:

ω∆x A(x 1 ) −1
µ µ ¶¶
Q= ln (2.11)
2c A(x 2 )

Where f = 2πω is the dominant frequency and c is the phase velocity (Tonn, 1991). One of
the limitations of this method is that it requires true amplitude recordings, usually difficult
to obtain on exploration seismic data. Other methods like the rise-time, analytical signal,
and pulse-amplitude also demand real amplitude measurements (REFERENCE) (See further
information of these methods on Appendix A).

2.3.2 Frequency domain methods


The matching technique is a method to calculate Q based on the transfer function computed
on the frequency domain. It was introduced by Raikes and White (1984) to determine the
operator that transforms the down going pulse recorded at one level into that recorded at a
deeper level. The matching of a signal at depth 1 to a signal at depth 2 results on the transfer
8 Seismic attenuation

function H12 (ω) and the inverse transfer function H21 (ω). Tonn (1991) defines the ratio of
the transfer function as:
H21 (ω)
µ ¶
ln = k − mω (2.12)
H12 (ω)
Where k is a constant and m is the slope:

∆t
m= (2.13)
Q

Note that this method yields the same result as the spectral ratio method shown on equation
2.18, apart for a factor of 2. However, the PSQI method uses the spectral ratio method not just
because it is one of the best known methods for computation of Q, but due to its simplicity
and flexibility in how it is parametrized (Reine, 2009).

2.4 The Spectral Ratio Method


As described in section 2.3, spectral analysis is the most used technique to calculate seismic
wave attenuation. We can write amplitude of a recorded seismic wave as:

|A(ω, r )| = |S(ω)||D(θ)||G(r )||P (ω)||I (ω)| (2.14)

Where S(ω) corresponds to the source spectrum, D(θ) is the source space function (that
depends on the direction θ from the source), G(r ) stands for the propagation effects (ge-
ometrical spreading, attenuation,and dispersion), P (ω) is the effect of the interfaces on
the spectrum and I (ω) is the instrument response. The spectral ratio method estimates
attenuation by comparing the amplitude spectra of two seismic arrivals. Bath (1974) defined
3 spectral equalization techniques to calculate attenuation from spectral amplitude ratios.

Number of Number of Number of


Method
frequencies compared stations compared waves compared

Frequency ratio two or more one one


Station ratio one at a time two or more one
Wave ratio one at a time one two or more

Table 2.1 Spectral amplitude ratios for determination of attenuation, modified from (Bath,
1974)
2.4 The Spectral Ratio Method 9

For the PSQI method we will use the Frequency ratio method as the preferred spectral
equalization technique. Dasios et al. (2001) defines three conditions that must be set before
applying the Spectral Ratio Method:

1. Q must be independent of frequency.

2. The medium must be assumed to be non-dispersive.

3. Velocity is considered to be constant within the frequency range examined.

As stated in equation 2.8, the amplitude spectrum A(ω) of a spherical wave can be expressed
as an exponential decay process. However, 2.14 states that instrument response, geometrical
spreading, reflection-transmission interaction and directivity, must be considered for the
amplitude analysis. By now it will be only considered the amplitude loses due to geometric
spreading G and energy partitioning P . For an initial source A 0 (ω), the spectrum at a given
reflection will be:
− ωt
A(ω) = (PG)A 0 (ω)e 2Q (2.15)

Where t is the travel time for that event and Q is the average quality factor. To measure 1/Q
between two reflection events, the ratio of the two spectra A 1 (ω) and A 2 (ω) may be taken,
eliminating the need to know the source spectra (Tonn, 1991). Then applying the natural
logarithm to the spectral ratio we get the equation:

A 2 (ω) ω∆t
µ ¶
ln =− + ln (PG) (2.16)
A 1 (ω) 2Q

Function 2.16 is a linear equation with respect to frequency ω. Figure 2.2 shows that the
slope contains 1/Q. We could define it as:

y(ω) = E ω + F (2.17)
∆t
E =− (2.18)
2Q
F = ln (PG) (2.19)
10 Seismic attenuation

Figure 2.2 Schematic diagram showing the natural log spectral ratio data as a function of
frequency, from (Reine, 2009).

2.5 Surface Seismic Measurements


Surface seismic Q measurements differ a lot from the direct measurements done through
core-laboratory and vertical seismic profiling (VSP) methods (Toksöz et al., 1979). While
these methods have the advantage of direct observations of the boundaries of interest, in
surface seismic measurements the layers cannot be isolated. This leads to a wide variety of
errors associated with the seismic wave propagating through a layered medium. However,
direct measurements present some problems that made them less suitable for reservoir
characterization. Laboratory Q measurements showed that attenuation in a single rock
type is highly dependent on the pressure, porosity and pore fluid (Winkler and Nur, 1982).
Furthermore, these measurements are usually done at ultrasonic frequencies (500-9000 Hz)
that are different for the seismic band (Raikes and White, 1984) and fail to simulate seismic
conditions accurately. On the other hand, VSP or check-shot surveys depend on the sparse
location of the well, and can only give a local measurement of the Quality factor.
2.6 Pre-stack Measurements 11

Despite the common sources of error like noise, geometrical spreading, and energy partition-
ing, a common midpoint (CMP) gather provides information in the time and offset domains,
allowing the extraction of information concerning structure, lithology, and material proper-
ties such as velocity and Q-factor of a dense spatial coverage (Zhang and Ulrych, 2002). This
is why robust methods (PSQI) for computing Q based on surface seismic measurements are
required to better understand attenuation at a reservoir scale.

2.6 Pre-stack Measurements


As described in equation 2.14 there are a number of factors that make the calculation of
attenuation elusive. These effects, combined with the obvious influences of noise and mul-
tiples, make estimation of Q even more difficult. A stacked seismic section might be used
for the computation of Q due to its optimum signal to noise ratio (S/N) and its similarity
with zero-offset VSP (Dasgupta and Clark, 1998). However, it has been proved that seismic
attenuation is highly corrupted by the normal processing steps done in stacking. Ebrom
(2004) described the stack related mechanisms that could affect the frequency content
of traces, and hence corrupt attenuation measurements. Some of the problems include:
Mis-stacking due to coarse velocity picking, mis-stacking due to nonhyperbolic moveout,
mis-stacking of converted shear waves and, mis-stacking of high amplitude multiples. All
of this effects lead to a spectral amplitude of the stacked trace with a changed frequency
content which is corrupting any attenuation signature.

Dasgupta and Clark (1998), introduced a novel method to improve surface seismic measure-
ments of attenuation based on prestack data. The Q versus offset method (QVO), uses the
variations of attenuation with offset to extract a zero offset attenuation value from prestack
gathers. Despite the effectiveness of this method, authors like Reine et al. (2012b) have
proved that QVO exhibits some issues that increase the uncertainty of the 1/Q measure-
ments and makes it less robust than the PSQI approach presented in this work. Some of
the disadvantages of the QVO method over the PSQI method include: high sensibility to
the bandwidth choice, corrupting influences related to the isochron between the horizons
of interest and, occurrence of spikes and ridges in the data due to the two step inversion
scheme.
Chapter 3

Spectral Analysis

The spectrum of a finite time series is a representation of how the total power is distributed
over frequency (Stoica and Moses, 1997). Attenuation measurements require a careful
analysis of the relevant spectra to be used. Hence, it is necessary to compare different types
of time-frequency (t-f) transforms and decide which are better for attenuation analyses.
On this chapter, I will discuss the general concepts of spectral analysis, I will expose some
relevant types of t-f transforms, and finally I will show synthetic examples comparing the
behavior of some t-f transforms to decide which one is suitable for the PSQI method.

3.1 Non stationary signals


The time representation of a signal is usually the first approach that we can have to any
signal obtained by receivers recording variations with time. However, the spectral analysis
involves the Fourier analysis and specially the use of the Fourier transform (FT), which
decomposes continuous signals, with infinite length and stationary frequency behavior into
their component sinusoidal basis functions (Bracewell, 1986). Furthermore, the Fourier
analysis plays an important role in signal processing because it is suited to common trans-
form methods , such as linear filtering (Flandrin, 1998).

The frequency representation obtained by the Fourier transform is:


Z ∞
X (f ) = x(t )e −i 2π f t d t (3.1)
−∞

Equation 3.1 shows that the Fourier analysis demands each frequency component to exist
with a constant amplitude over the entire time range, this is known as stationary, or time
invariant data (Bracewell, 1986). However, seismic wavelets are non-stationary, which is
3.2 Heisenberg-Gabor Uncertainty Principle 13

caused by the effects of wave front divergence, dispersion and attenuation (Zhou et al.,
2016).

3.2 Heisenberg-Gabor Uncertainty Principle


The deduction of this principle was well established by Flandrin (1998).
Defining the energy of a signal x(t ), as:
Z ∞
Ex = |x(t )|2 d t < +∞ (3.2)
−∞

Assuming that the signal x(t ) and its Fourier transform X ( f ) have a center of gravity that is:
Z ∞ Z ∞
2
t |x(t )| d t = 0 f |X ( f )|2 d f = 0 (3.3)
−∞ −∞

Introducing the respective moments of inertia:

1
Z ∞ 1
Z ∞
∆t 2 = t |x(t )|2 d t ∆f 2 = f |X ( f )|2 d f (3.4)
Ex −∞ Ex −∞

Defining the auxiliar quantuty I :


Z ∞ dx

I= t x(t ) (t )d t (3.5)
−∞ dt

We can use the Parsevals identity and find that:

∞ ∞
¯ ¯2
2 ¯dx
Z Z
2 2 2 2
(t )¯¯ d t = 4π2 E x2 ∆t 2 ∆ f 2
¯ ¯
[Re{I }] ≤ |I | ≤ t |x(t )| d t · t ¯ (3.6)
−∞ −∞ dt

Intergration by parts show that I is:

∞ ∗
dx
¯ Z ∗
∞ 2¯
I = t |x(t )| ¯ ∞ − E x − t x(t ) (t )d t = −E x − I (3.7)
− −∞ dt

Then:
Ex
Re{I } = − (3.8)
2
Replacing 3.8 into 3.6 and using the assumption that x(t ) decays so fast that t |x(t )|2 vanishes
at infinity, we get:
1
∆t ∆ f ≥ (3.9)

14 Spectral Analysis

Where ∆t is the duration time and ∆ f the frequency bandwidth (Flandrin, 1998), relate to
the resolution of adjacent signal components. If the standard deviations of ∆t and ∆ f are
measured, we can define the Heisenberg Gabor uncertainty principle as:

1
σt σ f ≥ (3.10)

Where σt and σ f are the standard deviations of time and frequency estimates respectively
(Hall, 2006). The final consequence of the Heisenberg Gabor uncertainty principle is that
for windows with improved time resolution there will be poorer frequency resolution, and
vice versa. Usually time dimension is typically fixed for spectral analysis, however it is not
a requirement. It is possible to use a time-window that changes with the frequency being
analyzed and still follow the inequality (Reine, 2009).

3.3 Time-frequency Representations


A spectrogram is defined as a real-valued, and non-negative distribution of the spectral
energy density (Bracewell, 1986). For a one dimensional (1D) time series, the time-frequency
transform is a 2D spectrogram with coordinates of time and frequency. Depending on the
transform used, the resolution in time or frequency domain varies following the Heisenberg
Gabor uncertainty principle.
3.3 Time-frequency Representations 15

Figure 3.1 Sampling of the time-frequency plane. Different forms of sampling: Shannon,
Fourier, Gabor, Wavelet , modified from (Flandrin, 1998).

3.3.1 Short-time Fourier transform


The short-time Fourier transform (STFT) is used to obtain local frequency spectrum by by
taking the Fourier transform in sliding windows over the signal. In it, the data trace s(t )
is gated by a sliding window function g (t ), and the Fourier transform is applied. A sharp
cut-off may introduce artificial discontinuities and create unwanted problems, so a smooth
16 Spectral Analysis

cut-off window function should be used (Gröchenig, 2013). The STFT is defined as:
Z ∞
S F (x, f ) = g (t − x)s(t )e −i 2πt f d t (3.11)
−∞

In equation 3.11 x represents the time lag to the center of the window function.

For signals sampled at a rate ∆t , this leads to a discrete STFT given by:

NX
−1 i 2πnm
S Fn,m = g n−x s n e − N (3.12)
n=0

Wheres x describes a discrete lag (Reine, 2009).

Figure 3.2 The Short time Fourier transform, from (Gröchenig, 2013).

3.3.2 Gabor transform


The STFT and the Gabor transform are known as fixed-window transforms. The STFT doesn’t
have an inherent window function. However, Gaussian functions are the only solutions that
3.4 Synthetic examples 17

minimize the duration-bandwidth product in the Heisenberg-Gabor sense (Flandrin, 1998).


Thus, the Gabor transform uses a window defined by a Gaussian function. This leads to a
transform defined by:
NX
−1 i 2πnm
SG n,m = g G n−x s n e − N (3.13)
n=0

Where the Gaussian window defined by (Zhou et al., 2016) is:

2
1 − t
g (t ) = p e 2α2 (3.14)
α 2π

Note that this window is parametirized by α, which acts as a reciprocal of the standard
deviation, a measure of the width of the Fourier transform (Harris, 1978).

3.3.3 S transform
When the size of the time-window is a function of the frequency, we refer to variable-window
transforms. By replacing equation 3.14 into 3.1 we get the The S-transform (Stockwell et al.,
1996): Z ∞ | f | −(τ−t )2 f 2
S(τ, f ) = s(t ) p e 2 e −i 2π f t d t (3.15)
−∞ 2π
Equation 3.15, where τ is the time lag. Note that unlike the normal Gaussian window, the S-
transform uses a Gaussian window that depends on frequency and time. Spectral-temporal
resolution trade off cannot be avoided since is a direct consequence of the Heisenberg
Gabor uncertainty principle. However, it can be optimized. It is desired to have good
frequency resolution at low frequencies (Resolve near signals at low frequencies) and, good
time resolution at high frequencies (Resolve near signals at high frequencies).

3.4 Synthetic examples


To better understand time-frequency transforms, 3 example signals (impulse, sine, chirp),
were generated and transformed using one fixed-window transform (STFT) and, one variable-
window transform (S-transform). All signals are 1000 ms long with a sample interval of 1 ms.
The fixed window transform uses a Hamming window of 128 points.
18 Spectral Analysis

(a) Input signal

(b) STFT

(c) S-transform

Figure 3.3 Impulse signal (a) with the time-frequency representations of STFT (b) and S-
transform (c).
3.4 Synthetic examples 19

(a) Input signal

(b) STFT

(c) S-transform

Figure 3.4 Sine signal (a) with the time-frequency representations of STFT (b) and S-
transform (c).
20 Spectral Analysis

(a) Input signal

(b) STFT

(c) S-transform

Figure 3.5 Chirp signal (a) with the time-frequency representations of STFT (b) and S-
transform (c).
3.4 Synthetic examples 21

3.4.1 Analysis of results


Impulse

The impulse function is a useful signal to analyze because it contains all frequency compo-
nents at a single time. This impulse has an amplitude of 1 at 0.5s. There is a remarkable
difference between the fixed-window transform and the variable-window transform. While
the STFT exposes a fixed time resolution for all frequencies, the S-transform has a better
time resolution for higher frequencies than for lower ones.

Sine Wave

Another function useful for time-frequency analysis is the sine wave, since it has a specific
frequency for all times. Figures 3.4b and 3.4c shows the time-frequency transform of an
input sine wave with a frequency of 125H z. In this case, both transforms have constant
frequency resolution for all times.

Chirp

A chirp signal presents a gradual change of frequency with time. Figure 3.5a shows a linear
chirp varying linearly from 25H z to 250H z. While the STFT shows a fixed frequency resolu-
tion, the S-transform presents a high frequency resolution at lower frequencies.

3.4.2 Conclusion
The results clearly state the advantage of a variable-window time frequency transform
(S-transform) over a fixed-window transform (STFT). Figure 3.5c shows that a frequency
dependent window is desirable for the detection of high frequency contents. Furthermore,
figure 3.3c reveals that the S-transform has a better time resolution at higher frequencies
compared to low frequencies. This is useful when trying to resolve two near signals at high
frequencies. Finally, the inverse frequency dependence of the localizing Gaussian window is
an improvement over the fixed width window used in the STFT (Stockwell et al., 1996).
Chapter 4

Inversion Scheme

In this work, inversion refers to the determination of 1/Q from the calculated natural log
spectral ratio data. The purpose of this chapter is to show the advantage of applying a
simultaneous inversion scheme with respect to frequency and time difference. Moreover,
this aims to demonstrate the effects of weighting the inverse problem in the presence of
high-frequency noise and random velocity fluctuations.

4.1 Simultaneous inversion scheme


Equation 2.16 establishes the relation between the natural log spectral ratio and the inverse
quality factor 1/Q. On each trace, the spectrum of the windowed event A 2 (ω) is divided by
that of the reference spectrum A 1 (ω) giving the following equation:

A 2 (∆t , ω) ω∆t
µ ¶
ln =− + ln (P (∆t , ω)G (∆t , ω)) (4.1)
A 1 (∆t , ω) 2Q

Note that 4.1 introduces the dependence of A 2 (∆t , ω), A 1 (∆t , ω), P (∆t , ω) and G (∆t , ω)
on ∆t (the difference in travel time between the two measurements) and ω (the angular
frequency). In section 2.4 is presented a linear solution to equation 2.16, here it extends to:

y (∆t ω) = E ∆t ω + F (∆t ω) (4.2)


4.1 Simultaneous inversion scheme 23

Where

A 2 (∆t , ω)
µ ¶
y (∆t ω) = ln
A 1 (∆t , ω)

1
E =−
2Q

F (∆t ω) = ln (P (∆t , ω)G (∆t , ω))

The inversion is considered as simultaneous because the product of ∆t and ω is treated as a


single variable (Reine, 2009).

As a first approximation, I will consider the term F (∆t ω) in equation 4.2 to be constant
with respect to the variable ∆t ω. However, this assumption introduces bias to the measure-
ment because both the reflectivity and the geometric spreading varies form trace to trace. At
a constant intercept F (∆t ω) = F .
Then, equation 4.2 can be written as:

y (∆t ω) = E ∆t ω + F (4.3)

Figure 4.1 Schematic drawing of the linear relationship of the natural log spectral ratio data
for all traces. The blue osculations represent some random noise.
24 Inversion Scheme

Figure 4.2 Another visualization of equation 4.2 but in the ∆t ω space. Note that attenuation
causes the surface to descend as one coordinate increases relative to the other.

To reduce the bias induced by the assumption of a constant ln (P (∆t , ω)G (∆t , ω)), it
is recommended to apply a geometric spreading correction to compensate for wavefront
divergence early in processing (Yilmaz, 2001). The energy partitioning term P is not easily
corrected, but some approximations may by applied to reduce its effects (Reine, 2009).

As a second and better approach, I will treat the inversion of equation 4.2 with variable trace
intercepts. In this case I will only consider a ∆t dependence of the ln(PG) term. Equation
4.2 takes the form:
y (∆t ω) = E ∆t ω + F (∆t ) (4.4)

Whit a matrix notation:


4.1 Simultaneous inversion scheme 25

   
y 11 ∆t 1 ω1 1 0 ... 0
 y 21   ∆t 2 ω1
   
   0 1 . . . 0 
 .   .. .. .. . . .. 
 ..  
 
   . . . . . E
 y N 1   ∆t N ω1 0 0 . . . 1  F 1 
    
    
 y   ∆t ω
 12  =  1 2 1 0 . . . 0   F2  (4.5)
 
 .   .. .. .. . . ..  
 . 
 .  
 .   . . . . .  .. 

   
 y N 2   ∆t N ω2
   0 . . . 1 BN
 .   .. .. .. . . .. 
 ..   . . . . .
   
yN M ∆t N ωM 0 0 ... 1

Equation 4.5 can be represented with the explicit linear equation:

d = Gm (4.6)

Equation 4.6 is considered as the foundation of the study of discrete inverse theory (Menke,
2012). In here,G stands for the data kernel, d is the data and, m refers to the model.

Since inverse theory is concerned with deducing knowledge from observational data that
has a discrete nature, d necessarily has to be discrete. Therefore, the discrete inverse theory
can be defined as:
M
X
d= G i j mi (4.7)
j =1

Equation 4.6 has no exact solution. However, a least square solution, is as a good approach
to estimate the solution based on the values of the model parameters that gave the best
approximate solution. Thus, Equation 4.6 has the least squares solution:

¤−1 T
m = GT G
£
G d (4.8)
26 Inversion Scheme

Figure 4.3 Solution to 4.6 with a variable intercept. Shows undulations in the data along the
∆t coordinate.

The solution of the inverse problem is a smoothly varying surface. The undulating values
of the surface shown in figure 4.3 correspond to the F ∆t term of equation 4.4.

4.2 Uncertainty
The two approaches for the inversion of equation 4.2 presented in this section (Variable
trace intercepts and constant trace intercepts), consider an equal degree of uncertainty over
all traces and frequencies. Nevertheless, this is not true for seismic data which is subjected
to frequency-dependent noise. As a result, each data point has its own uncertainty and can’t
be treated equally in the inversion. By applying a weighting function to the data, points
with low uncertainty will contribute more to the solution than points with high uncertainty
values (Taylor, 1997).

4.2.1 Weighting functions


If we assume that the measurements of the inversion solution obey a Gaussian distribution,
a good approach to provide an estimate of error is through the use of the standard deviation.
Taylor (1997) defines the weighting function w i as the reciprocal square of the corresponding
uncertainty σi .
1
wi = (4.9)
σi 2
4.2 Uncertainty 27

4.2.2 Weighted inversion


The inverse problem presented in equation 4.6 has a weighted least-squares solution defined
by Menke (2012) as:
¤−1 T
m = GT W G
£
G Wd (4.10)
Chapter 5

τ − p Domain

5.1 The τ − p transform

5.1.1 Definition
Huygens’ principle of superposition allows the synthesis of a downward plane wave from the
spherical waveforms of the multiple sources (Schultz and Claerbout, 1978). If an appropriate
time lag is chosen, we can create a downward plane wave at an arbitrary angle θ. Time lag
∆t may be defined as:
sin θ
∆t = ∆x (5.1)
v
Where θ is the incidence angle of the plane wave, v is its velocity, and ∆x is the separation
between sources. Snell’s law states that for a ray path in a medium where velocity depends
only on depth z, (sin θ/v) is constant, and the ray is confined to a vertical plane (Aki and
Richards, 2002). The ray parameter p is defined as:

sin θ(z)
p= = C onst ant (5.2)
v(z)

Where p is usually known as the ray parameter or horizontal slowness. Replacing equation
5.2 into 5.1:
∆t
p= (5.3)
∆x
Equation 5.3 reveals that the downward plane wave created by the superposition of spherical
waveforms is a line with constant slope in the time-distance (t − x) domain. Hence, a linear
moveout can be defined as:
t = px + τ (5.4)
5.1 The τ − p transform 29

Where τ is the zero-offset intercept time. (Turner, 1990) resumes the mathematical definition
of the τ − p transform. Z ∞
s x, τ + px d x
¡ ¢
ŝ(τ, p) = (5.5)
−∞

And for discrete data:


N ¡
s x i , τ + px i
X ¢
ŝ(τ, p) = (5.6)
i =1

In equation 5.6 N is the number of seismic traces used in the transform, and ŝ(τ, p) is the
amplitude at (τ, p) in the τ − p domain. For multiple layers, the τ − p mappings for the
reflections are the sums of ellipses, while the direct arrival and head waves are mapped
as points (Diebold and Stoffa, 1981). Figure 5.1 shows synthetic traveltime curves in t − x
domain, and their τ − p mapping for a model with plane homogeneous layers. See that
reflections on the τ − p domain are mapped as ellipses because their apparent velocity
changes with offset, while for the head and direct waves the apparent velocity is constant so
they are appear as points in the τ − p domain.

Figure 5.1 Mapping of hyperbolic reflections in the t − x domain shown as ellipses in the
τ − p domain, and linear events like refractions in the t − x domain shown as points in the
τ − p domain. The vertical axis of the right figure is the intercept time τ. From (Reine, 2009).
30 τ − p Domain

5.1.2 Artifacts
To create a perfect downgoing composite plane wave by the superposition of many spherical
wavefronts, there must be an infinite lateral extent of the shot array and an infinitesimal shot
separation. However, normal seismic acquisitions are of finite spatial and temporal extent,
and discretely sampled in space and time. Hence, an introduction of offset truncation
artifacts and aliasing are unavoidable (Schultz and Claerbout, 1978).

A relation between τ and p can be derived from equation 5.4.

τ = t − px (5.7)

Equation 5.7 shows that the slope of a linear artifact has a value equal to the negative of the
offset were it was produced.

In this work we will refer to aliasing as the appearance of energy at multiple values of
the transformed horizontal slowness. Yilmaz (2001) establishes that spatial aliasing occurs
when the wavefront separation in time ∆t equals half the dominant period T . To avoid
spatial aliasing, we require that:
v
∆x ≤ (5.8)
2 f max
Where ∆x is the spatial sample interval, v is the lowest horizontal phase velocity in the data,
and f max is the maximum frequency of interest (Stoffa et al., 1981). Using spatial sampling
principles, we can define a Nyquist ray parameter p N .

1
pN = (5.9)
2∆x f max

Horizontal slownesses below p N are mapped correctly, while higher slowness values are
aliased. To reduce the effects of truncation and aliasing, Schultz and Claerbout (1978)
suggest to use a limited portion of the horizontal slowness line so that the windowed data
will be the only included into the τ − p transform.

5.1.3 Moveout and anisotropy


Whereas traces are stacked over offset after a conventional t − x moveout correction, they
are stacked over slowness after a τ − p moveout correction (van der Baan, 2004). Figure 5.1
shows that for a reflection in a homogeneous and isotropic media, the moveout follows an
ellipse in the τ − p domain. For a layered media, the τ − p moveout is just the sum of ellipses
5.1 The τ − p transform 31

(Diebold and Stoffa, 1981).

For a plane wave traveling in a homogeneous and isotopic media the total slowness u
can be defined as:
1
q
u= = p2 + q2 (5.10)
v
Where v is the the wave velocity in the medium, p is the horizontal component of the
slowness and q the vertical one. For multiple layers, the intercept time τ can be found by
summing contributions from each layer.

N
τi = 2
X
qi zi (5.11)
i =1

With N the number of layers and z i the thickness of layers down to the reflector. Replacing
q from equation 5.10 into 5.11 we get:

τi = 2z i u i2 − p 2
¡ ¢
(5.12)

τi = τ0i 1 − v i2 p 2
¡ ¢
(5.13)

Where τ0i is the vertical incidence traveltime (Diebold and Stoffa, 1981).

Seismic anisotropy implies that the velocity of plane waves is direction-dependent (Hake,
1986). Moreover, it causes that the velocities of plane waves and the rays that carry the
energy to vary as a function of travel direction (Levin, 1990). Thus, ignoring the effects
of anisotropy may seriously affect the results of processing and interpretation steps, such
as normal moveout (NMO) correction (Thomsen, 1986). Tsvankin and Thomsen (1994)
describe the main distortions in reflection moveouts caused by the presence of anisotropy.

1. The short-spread moveout velocity is not equal to the root mean square (rms) velocity.
Hence, the difference between vertical rms and moveout velocities may lead to serious
errors in interval velocity, even for weak anisotropy.

2. Anisotropy leads to nonhyperbolic moveout, causing serious distortions in velocity


estimation.

The simplest anisotropic case of broad geophysical applicability is the transverse isotropy
(TI) with vertical symmetry axis. For this work, we will only study this particular case of
anisotropy, applied specially to weak anisotropy . Alkhalifah (1997) defines the η anisotropy
32 τ − p Domain

parameter to describe weak anisotropy.

ϵ−δ
η= (5.14)
1 + 2δ

Where |ϵ| ≪ 1 and |δ| ≪ 1 are the weak anisotropic parameters defined by Thomsen (1986).
From equation 5.13, the anisotropic moveout in the τ − p domain within a given layer i is:

!1
p 2 v i2
Ã
2

τi = τ0 i 1 − (5.15)
1 − 2η i p 2 v i2

Where τ0 is the traveltime for a vertical incidence wave and v i is the interval stacking velocity
(van der Baan and Kendall, 2002). The parameter η for TI models with a vertical (VTI) axis of
symmetry can be obtained by inverting either NMO velocity from dip-moveout behavior of
P-wave surface seismic data, or nonhyperbolic moveout (Alkhalifah, 1997; Tsvankin et al.,
2010) For the isotropic case η = 0, equation 5.15 reduces to equation 5.13. To compute the
inversion of attenuation, the equivalent traveltime for a given reflection on the τ − p domain
is required.  

p2
 
t = τ
 X 
1 + ·µ ¶µ ¶¸ 1

 (5.16)
2
 1 1 p2 ¡ 2 2
¢ 
v i2
− p2 v i2
− 1−2η i p 2 v i2
1 − 2η i p 2 v i

(Reine, 2009). Equation 5.16 shows that for a given τ and p, the traveltime in the τ − p
domain can be calculated by knowing the interval values of v and η.
Chapter 6

Multimensional Interpolation

The most common preconditioning of seismic data is done to improve the signal-to-noise
(S/N) ratio by the removing of noise, the reduction of unwanted arrivals, and the regulariza-
tion of amplitude bursts. Although, this processing steps are sufficient to enhance seismic
images for interpretation (Brown, 2011), they usually ignore the effects of missing spatial
data. Missing offsets and azimuths, such as dead traces and lower-fold areas, may intro-
duce attribute artifacts and hence, have a negative impact in prestack inversion algorithms
(Chopra and Marfurt, 2013).

In this chapter I will discuss two interpolation techniques, the minimum weighted norm
interpolation (MWNI) method by Liu and Sacchi (2004), and the projection onto convex sets
(POSC) method by Abma and Kabir (2006). By the end of the chapter, I will choose the most
suitable algorithm to be used in a multidimensional reconstruction of seismic wavefields.

6.1 Seismic data interpolation


Seismic data reconstruction is crucial for any processing step requiring regular sampling.
Cary (1998) stated that the apparition of truncation artifacts and aliasing in the τ − p do-
main are mainly caused by gaps in the input data and missing data at near and far offsets.
Therefore, the proper infill of missing data prior to the transformation into τ − p domain,
can significantly reduce the presence of artifacts and aliasing.

Interpolation algorithms involving multiple spatial dimensions have many advantages


over one-dimensional methods (Trad, 2008). Particularly, interpolation in all five dimen-
sions (bin spacing along CMP X, bin spacing along CMP Y, bin spacing along azimuth, bin
34 Multimensional Interpolation

spacing along offset, time) is the most robust way to interpolate missing data that represents
the wavefield in azimuth and offset in a better way. For simplicity, I will expose the multidi-
mensional sampling functions for the 2D case, that can be generalized in a straightforward
manner to the N-D case (Naghizadeh and Sacchi, 2010). Consider the structure u with size
N x × N y given by

¡ ¢ n x ∈ 0 : N x − 1
u nx , n y = (6.1)
n ∈ 0 : N − 1
y y

The discrete Fourier transform (DFT) of u, is given by:

N y −1
à !
1 X 1 NX
x −1 ¡ ¢ −2i πk x n x −2i πk
Ny
y ny
U (k x , k y ) = u nx , n y e N x e (6.2)
N y n y =0 N x n x=0

(Naghizadeh and Sacchi, 2010). Assuming that M samples of the data are 0 (Death traces), a
signal with missing samples u z may be defined. Therefore, the DFT of the data with missing
samples U z , is obtained by
Uz = U ~ Q (6.3)

Where ~ is the 2D convolution operator, and Q is the 2D Fourier response given by

1 MX −1 −2i πk x h x (m) −2i πk y h y (m)


Ny
Q(k x , k y ) = e Nx e (6.4)
N x N y m=0

Where, h x and h y are the locations of the available samples on the X and Y axis respectively
(Naghizadeh and Sacchi, 2010).

6.2 MWNI
Minimum weighted norm interpolation is a band-limited frequency-domain interpolation
algorithm, which uses a conjugate-gradient iteration to invert for an interpolation solution
with minimum error (Liu and Sacchi, 2004).

The method starts with a multi-dimensional matrix of the specified output sampling that is
filled with ones or zeros depending on the availability of the traces. All traces are transformed
to frequency domain using the discrete Fourier transform, and each band-limited frequency
slice is processed through a conjugate gradient inversion.
6.3 POSC 35

The band-limiting for MWNI occurs in the wavenumber k space, and is defined through a
weighted norm. The final result is a denser data set, with fewer gaps, suitable for a variety of
prestack algorithms including the τ − p transform. A step by step explanation of the MWNI
algorithm is given in Appendix B.

6.3 POSC
The projection onto convex sets (POCS) is a frequency-domain interpolation algorithm,
which applies multidimensional Fourier transform to interpolate irregularly populated grids
of seismic data with a simple iterative method (Abma and Kabir, 2006).

The method defines a threshold value that varies linearly from a large value in the first
iteration to a low value on the last iteration. Each threshold is applied to the data to keep
only the highest amplitudes. Due to the thresholding, some amplitudes will fill areas that
were previously zeros while keeping the orginal trace implitudes. The iterative process
continues until a proper convergence is reached.

6.3.1 5D Interpolation
Interpolation in five dimensions usually refers to one of the following cases:

1. 5 dimenions: Bin spacing along CMP X, bin spacing along CMP Y, bin spacing along
azimuth, bin spacing along offset, time or depth.

2. 5 dimenions: Bin spacing along CMP X, bin spacing along CMP Y, bin spacing along
offset X, bin spacing along offset Y, time or depth.

Since MWNI algorithm uses a spatial band-limiting of the data and the POCS algorithm does
not, results obtained with the MWNI are more robust and have less noise. Additionally, the
objective of applying the interpolation in this work is to reduce the gaps between offsets .
Therefore, the azimuth dimension is required to compensate for any azimuthal errors in
the shooting geometry. As a conclusion, I will use the MWNI algorithm applied to the five
dimensions including the bin spacing along offset.
Chapter 7

The Prestack Q-Inversion Method

In the previous chapters I described the main components that make PSQI a robust method
to calculate attenuation from surface pre-stack seismic data. These components involved
using a variable-window time-frequency transform (S-transform), inverting the natural log
spectral ratio surface in frequency and time difference simultaneously, and operating in the
τ − p domain.

In this chapter I discuss the main seismic processing steps that must be done to accomplish
the best results with the PSQI method. Furthermore, I show how the three components
discussed in chapters 3, 4, and 5 may be integrated to form the PSQI method.

Reine (2009) designed the PSQI method to remain constant to the nature of the surface
seismic data, no matter if it is land or marine data, 2D or 3D surveys. Figure 7.1 shows an
overview of the PSQI process.

7.1 Preproccesing
The primary objective in seismic processing is to obtain an earth model in time with an ac-
companying earth image in time. To achieve this, some common steps in seismic processing
include deconvolution, CMP stacking, and migration (Yilmaz, 2001). Although these steps
produce an optimal image for interpretation, they corrupt the seismic spectrum which is
needed to measure attenuation. For this reason, the processing steps needed for attenuation
analysis differ from the common processing steps used for seismic interpretation.
7.1 Preproccesing 37

Figure 7.1 Flow diagram of the PSQI process. Modified from (Reine, 2009).
38 The Prestack Q-Inversion Method

7.1.1 Static Corrections


Proper alignment of reflections with the expected traveltime curves is highly sensitive to
static shifts due to variable topography, the presence of high-velocity rocks near surface,
and complex subsurface velocity (Bevc, 1997). Therefore, corrections for static shifts are
necessary for attenuation analysis that requires well defined traveltime curves for a correct
location of the target spectra.

Elevation and static corrections shift the datum of a collection of seismic traces from one
surface of arbitrary shape to another reference horizontal datum based on a near-surface
velocity model (Berryhill, 1979; Cox, 1999). Similarly, surface-consistent residual statics
compensates for the effects of time delays in the highly variable near-surface weathering
zone (Wiggins et al., 1976). Residual statics may be run in conjunction with velocity analysis
to improve both processes.

7.1.2 Velocity Analysis


To calculate the traveltime curves, the moveout velocity for each CMP must be known. Yil-
maz (2001) established that velocity analysis can be done in selected CMP gathers or group
of gathers. For the purpose of attenuation measurements, the main goal of the velocity
analysis is to provide the best match between the moveout curves and the data.

For anisotropic data, the velocity analysis may be performed directly on the τ − p domain
with the advantage of using phase velocity, which is the natural velocity used in VTI and
orthorhombic media analysis (Tsvankin, 1997; van der Baan and Kendall, 2002). If the
maximum offset is large enough, Alkhalifah (1997) defined a way to extract the effective
anisotropy parameter η, described in chapter 5, directly from the velocity analysis. Equation
5.15 shows that the availability of this parameter can significantly increase the accuracy of
the calculated moveout curves in the τ − p domain.

7.1.3 Band-Pass Filter


Surface waves can be a serious issue for attenuation analysis in land data. While shallow
events may not be affected by surface waves, usually deeper events will. Since these waves
have a lower frequency content than the reflections, the difference in the combined spectra
distorts the process of measuring attenuation (Reine, 2009). Furthermore, the high ampli-
tudes of the surface waves can introduce artifacts in the τ − p transformed data (Stoffa et al.,
1981).
7.1 Preproccesing 39

To correct these problems, a single band-pass filter may be applied to the data. This solution
works well to remove the surface waves, although some of them will still be present after the
filtering process. Surface waves are different from reflections in both temporal and spatial
frequency. For this reason, they are usually removed by applying a frequency-wave number
f − k filter. Nevertheless, when the amplitude of the ground roll is bigger than those of
the reflection signals, f − k filters cause a severe distortion in the signal, and may corrupt
attenuation measurements (Liu, 1999).

7.1.4 Amplitude Regularization


The main importance of having a trace to trace consistency of amplitudes stems for the
τ − p transform. Anomalously high amplitude traces introduce artifacts during the τ − p
transform process, while weak frequency traces do not properly contribute to the plane wave
synthesis. Noisy traces that can’t be removed after the band-pass filter must be removed
from the data set through the use of a mute.

After the elimination of noisy traces and ground roll, the effect of geometric spreading
must be removed too. Ursin (1989) derived an offset dependent geometrical spreading
correction as a function of two-way zero-offset traveltime, and the Root Mean Square (RMS)
velocity estimated in the velocity analysis. Finally, with the proper amplitude and geometri-
cal spreading corrections, the surface consistent amplitudes are calculated for all of the data
simultaneously.

7.1.5 5D Interpolation
The PSQI code (Reine, 2009; Reine et al., 2012a,b), uses supergathers to fill up the gaps
between offsets. Although, this solution efficiently reduce the gaps, a multidimensional
interpolation is much more efficient (Trad, 2008). The 5D interpolation approach, defined in
chapter 6, uses supergathers as input for the interpolation. The algorithm predicts missing
traces with data from the surrounding CMPs, rather than just creating a huge supergather
with N CMPs inside.

7.1.6 τ − p Transform
The final stage in the preprocessing is to transform the data into the τ− p domain. A number
of parameters must be specified, and some may vary depending on the transform used. It
40 The Prestack Q-Inversion Method

is important to make a proper choice of the parameters to avoid artifacts in the transform
process. For example, the maximum frequency used in the transform should be well above
the maximum frequency of the signal and the high-cut frequency of the band-pass filter
(Reine, 2009). A more detail explanation of the τ − p transform parameters used in this work
is given in section 8.

Additionally, the maximum horizontal slowness to be used in the transform should


exceed that of the reflection with the largest moveout. Equation 5.4 shows that this value
can be obtained by taking the derivative of time with respect to offset, this is equivalent to
differentiating the moveout equation with respect to offset. Alkhalifah and Tsvankin (1995)
derived a moveout equation for anisotropic VTI media.

à !1
x2 2ηx 4 2

t= τ20 + − 2 (7.1)
v r2ms v r ms t 02 v r2ms + 1 + 2η x 2
£ ¡ ¢ ¤

Where v r ms is the stacking velocity. Differentiating this function with respect to offset gives:
à !
1 + 2η x 2
¡¢
x 2ηx 3
p= + 2 £ 2 2 ¢ ¤ −2 (7.2)
t v r2ms
¡ ¢ ¤ £ 2 2 ¡
t v r ms t 0 v r ms + 1 + 2η x 2 t 0 v r ms + 1 + 2η x 2

(Reine, 2009). Note that if the medium is isotropic η = 0, equation 7.2 reduces to:

x
p= (7.3)
t v r2ms

Evaluating equation 7.2 or 7.3 at the maximum offset, deepest horizon, and maximum
velocity in the zone of interest, gives the value of the maximum horizontal slowness present
in the data.

7.2 Calculate Moveout Curves


The moveout curves are required to track reflections in the τ − p domain across the gather.
Hence, moveout curves should match the data as accurately as possible to correctly extract
the relevant spectra for each trace.
7.3 Spectral Decomposition 41

7.2.1 Interval Velocity


Velocity analysis carried out in the t − x domain yields a stacking velocity or approximate
RMS velocity (Yilmaz, 2001). In section 7.1.4, I examined the importance of doing the velocity
analysis in the time-offset domain to address the problem of geometrical spreading. How-
ever, it is necessary to convert this RMS velocities to interval stacking velocities. Equation
7.4 shows how to obtain interval velocities from RMS velocities.

v n2 t n − v n−1
¢ ¡ 2 ¢
t n−1
vi = (7.4)
t n − t n−1

Where v n and v n−1 are the RMS velocities at the layer boundaries n and n − 1 respectively,
and t n and t n−1 are the horizon times at these layer boundaries (Dix, 1955). Alternatively, if
the velocity analysis is done in the τ − p domain, interval stacking velocities can be obtained
directly in this domain (van der Baan, 2004).

7.2.2 Equivalent Traveltime


Equation 5.16 describes the equivalent traveltime for a given τ and p. This quantity is
necessary to calculate ∆t for the inversion of the natural log spectral ratio described in
chapter 4. The ∆t values are calculated by subtracting the equivalent traveltimes of the
relevant reflections.

7.3 Spectral Decomposition


In chapter 3, I mentioned the advantages of using a variable-window time-frequency
transform for the spectral analysis required in attenuation measurements. I picked the
S-transform because of its optimal time-frequency resolution, and the dependence of its
Gaussian window on time and frequency.

7.3.1 Data Range


The S-transform takes data with n time samples and m traces and produces an output of
n × n × m points. The result may be very huge and thus, computationally inefficient. For this
reason, the PSQI process requires a truncation of the data to avoid the transform of refracted
waves that are not properly recorded in the τ − p space. Therefore, the data is truncated
before the horizontal slowness of refraction at the lowest interface. To do so, a window
corresponding to approximately 1/4 of the dominant period is chosen. Then, the minimum
42 The Prestack Q-Inversion Method

critical angle, for each interface within the window, is calculated using the interval velocities
and the Snell’s law. Next, the extracted critical angle is converted into a value of horizontal
slowness.
vn
sin (θc ) = (7.5)
v n+1
sin (θc )
pc = (7.6)
v max
Where θc is the critical angle, p c is the critical slowness corresponding to the horizontal slow-
ness of refraction, and v max is the maximum interval velocity within the selected window
(Reine, 2009).

Additionally, the horizontal slowness corresponding to the maximum offset in the data
must also be considered. Equations 7.2 and 7.3 may be used to define the maximum hori-
zontal slowness at which the data has to be truncated. Figure 7.2 shows the truncation of
the data at the slowness of refraction.

Figure 7.2 The horizontal slowness of the data must be truncated at the slowness of refraction
or maximum offset for the lowest interface. Here, the truncation shown in gray occurs at the
refraction of the red horizon. From (Reine, 2009).
7.4 Inversion for 1/Q 43

7.3.2 Spectral information


Once the data has been truncated and properly conditioned, the S-transform is applied.
It is important to mention that only the positive frequencies of the transform need to be
calculated. The spectra relevant to each reflection are matched to the appropriate moveout
curve, and the natural log of the ratio of each reflection combination is calculated. The final
product is a natural log spectral ratio surface for each pair of reflections.

7.4 Inversion for 1/Q


The final result of the PSQI method is the effective attenuation in the form of 1/Q, discussed
in section 2.1.3. This result is obtained by a simultaneous inversion of the natural log spec-
tral ratio data with respect to frequency and time difference. In chapter 4, I mentioned
the advantage of using a variable trace intercept for the inversion scheme, rather than a
constant trace intercept. Furthermore, I discussed the use of a weighting function to reduce
the uncertainty in the inversion of 1/Q.

The inversion is performed in the natural log spectral ratio surface, restricted in the ∆t
coordinate by the truncation of horizontal slownesses in the data. Additionally, the inversion
is also restricted in the frequency coordinate by the choice of an appropriate bandwidth over
which the attenuation is to be measured. Once the 1/Q value for a single CMP is calculated,
the entire process is repeated for the remaining CMPs. By the end, a map of 1/Q is obtained
for all the data, as well as a map with the corresponding uncertainty measurements.
Chapter 8

PSQI applied to Seismic Data in Colombia

The main objective of this work is to apply the PSQI method on real 3D seismic data from
Colombia. The data used in this work was provided by the colombian oil and gas company
Hocol S.A.. The prospecting reservoir for attenuation analysis is located approximately at
650 ms. I first give a short description of the area’s geology and then, review the seismic
acquisition parameters of the survey. Next, I show how each stage of the preprocessing
affects the quality of the data and make a detailed explanation of each of the PSQI steps
applied to real data. At the end, I present the results in the form of 1/Q maps and their
uncertainties.

8.1 Background Information

8.1.1 Geological Background


The potential reservoir investigated in this work is located at the Sinú-San Jacinto Basin in
Northwest Colombia. Figure 8.1 shows a general location map of the Sinú-San Jacinto Basin.
Hydrocarbon generation in this basin, is mainly attributed to the Cretaceous Cansona For-
mation. This formation is favorable for the generation of liquid hydrocarbons and consists
mainly of organic rich kerogen types I-II with a total organic carbon (TOC) of 2-11% (Olaya,
1994). Furthermore, Marín et al. (2010) suggest the presence of heavy oil with American
Petroleum Institute (API) gravities greater than 30◦ .

The main zone of interest in this work is located somewhere between the Cienaga De
Oro Formation, a sandstone reservoir located at lower Miocene. Just above this formation,
lies the shale Carmen Formation that works as the regional seal rock of the hydrocarbon
8.1 Background Information 45

system in this zone Barrero et al. (2007). Figure 8.3 shows a stratigraphic column of the
Sinú-San Jacinto Basin.

Figure 8.1 Location map of the Sinú-San Jacinto Basin. From (Sánchez and Permanyer,
2006).

8.1.2 Survey Parameters


The 3D survey acquisition parameters were designed to image up to 8s, with a sample
interval of 2ms. The main zone of interest is located before 1s. A surface grid of explosive
sources and receivers was deployed with a natural bin size of 20 m × 40 m, with a fold of 30.
The entire survey is made of a 557 × 312 grid of CMPs, resulting in 92868 CMPs. However,
the CMP data that I use in this work comprises a subset of 104 × 32 CMPs. Figure 8.2 shows
the foldmap of the survey and the subset that is used in this work.

Figure 8.2 Foldmap of the 3D survey. The black-dotted rectangle shows the subset where
the inversion was done. See that most of the fold values are consistent within the subset,
averaging a fold of 30.
46 3D Data Example

Figure 8.3 Stratigraphic column of the Sinú-San Jacinto Basin. From (Barrero et al., 2007).

8.2 PSQI Process


In this section I will apply the PSQI process to the real 3D seismic survey described in section
8.1.2. The steps of the PSQI are resumed in figure 7.1 and explained in detail in chapter 7. All
the examples used in this section correspond to the CMP located at inline 281 and xline 176.
8.2 PSQI Process 47

8.2.1 Preprocessing
Static Corrections and Velocity Analysis

CMP sorting, surface consistent amplitude corrections, static corrections, random noise
attenuation, and velocity analysis were done by PetroSeis LTDA using ProMax. Elevation
and refraction static corrections were applied (Using the Gauss Seidel algorithm in ProMax,
with a reference datum of 0m and a correction velocity of 2200 m/s), and two passes of
surface-consistent residual static corrections were iteratively applied (Using the Max Power
Autostatic algorithm in ProMax) in conjunction with the velocity analysis. The raw CMP
gather with static corrections is shown in figure 8.4.

Figure 8.4 Raw CMP with static corrections applied. Look at the high-amplitude surface
waves indicated in red.

Band-Pass Filter

To apply a proper band-pass filter to the data, it is important to understand the different
components of the spectrum. The spectra of the surface waves, the reflections, and the
entire CMP gather are shown in figure 8.5. The peak of the surface wave spectrum is located
approximately at 5 Hz, with the majority of its energy ocurring between 4 Hz and 10 Hz. In
contrast, the spectrum of the reflections shows two peaks of frequency at 20 Hz and 30 Hz.
Likewise, the frequency peaks of the entire CMP occur at 20 Hz and 40 Hz approximately.

I design an Ormsby band-pass filter with corner frequencies 8 Hz, 12 Hz, 70 Hz, and 80
Hz. This values were picked to exclude the surface wave energy below its peak, and the
48 3D Data Example

low-amplitude high-frequency contents seen in the entire CMP spectrum. Figure 8.6 shows
the filtered CMP gather. It can be seen that much of the surface wave energy has been
eliminated.

Figure 8.5 The average spectrum for the entire CMP (green), the surface waves (red), and the
reflections (blue).

Amplitude Corrections

As I mentioned before, the amplitude corrections were done by PetroSeis LTDA in a surface
consistent manner (Taner and Koehler, 1981). Yet, noisy amplitudes capable of producing
artifacts during the τ − p transform were still present in some traces. To account for this
problem, I used the TFD Noise Rejection tool from ProMax. This tool uses the STFT to
transform each record into t − f space and then, filters out isolated noise by replacing it with
the amplitudes from adjacent traces. While this is not a spectrally preserved operation, I
applied it sparingly to avoid the creation of strong artifacts in the τ − p domain.
8.2 PSQI Process 49

Figure 8.6 Band-pass filtered CMP gather. The surface waves, shown by the red triangle,
were effectively removed. The high frequency noise present on various traces has also been
removed.

5D Interpolation

In chapter 6, I discussed the multidimensional sampling functions and described two of


the most common multidimensional interpolation methods. To interpolate missing data, I
used the ProMax 4D/5D Interpolation tool with MWNI algorithm. As an input to the tool, I
created a 3 × 3 supergather arrangement and defined a maximum expected fold of 34. The
dimensionality of computation was set to 5 (Bin spacing along CMP X, bin spacing along
CMP Y, bin spacing along azimuth, bin spacing along offset, and time). To improve the result
of the interpolation, I just pick the center CMP from the 3 × 3 supergather. Figure 8.7 shows
the interpolated gather.
50 3D Data Example

Figure 8.7 The CMP gather from figure 8.4 after the multidimensional interpolation has been
applied. Noisy traces were removed, and anomalous amplitudes were corrected. All this
processes are necessary to reduce the artifacts during the τ − p transform.

τ − p Transform

Figure 8.8 shows the offset distribution for the CMP data, as well as the distribution after
applying the 5D interpolation. It can be seen that with the multidimensional interpolation,
some offset gaps are filled, and the distribution is improved.

The τ − p transform that I use in this work is the ProMax forward linear Radon transform
described in chapter 5. I set a maximum transform slowness of such that it exceeds the
largest slowness of the shallow reflections. The number of horizontal slowness values is
the same as the maximum offset in the t − x space prior to the transform. The minimum
transform slowness is just the negative of the maximum slowness. Frequencies from 0 Hz to
200 Hz are considered, with a damping factor of 0.01, and a reference offset of 1800 m.

To demonstrate that various preprocessing steps are required, I compare the τ − p transform
of the data for each process. The τ − p transform of the raw CMP data is shown in Figure 8.9.
These data appears free from any near or far-offset truncation artifact. However, there are
some with very low amplitudes specifically below 300 ms.

The surface waves and high-frequency noise were removed by applying the band-pass
filter in the t − x domain. Figure 8.10 shows the result of the filtered τ − p transformed
data. The most significant change between figures 8.8 and 8.9 is the introduction of higher
8.2 PSQI Process 51

amplitudes at the early times.

To regularize the offset distribution, I applied a multidimensional interpolation to the


filtered data. Figure 8.11 shows the final τ − p transform for the interpolated data. The
amplitudes for the early times have significantly increased.

(a) (b)

Figure 8.8 (a) shows the offset distribution of a raw CMP gather and (b) shows the interpo-
lated offset distribution of that same CMP. Offset spacing in the original data is regularized
after the 5D interpolation. Moreover, offsets lower than 600 m (Including offset 0 m) that
were not present in the raw data are created with the interpolation.

Figure 8.9 τ − p transform of the raw data. The horizontal slowness values are in ms/m
52 3D Data Example

Figure 8.10 τ − p transform of band-pass filtered data. The horizontal slowness values are in
ms/m

Figure 8.11 τ − p transform of the interpolated data. The horizontal slowness values are in
ms/m

8.2.2 Boundary of Interest


The zone of interest is located approximately at 650 ms on a high amplitude area within
the survey. Reflectors dip approximately 30◦ . Two horizons were picked across the area of
interest (From now referred as horizons A and B). Figure 8.12 shows the seismic position of
8.2 PSQI Process 53

these horizons over a Pre-stack time migrated (PSTM) amplitude map. Likewise, figure 8.13
shows the two-way traveltimes for each event.

Figure 8.12 Pre-stack time migration amplitude map. The circled zone indicates the area of
interest where high amplitude values are observed. The large amplitudes are often associated
with gas. The two horizons used for the inversion are labeled as A and B. Courtesy of Hocol
SA & PetroSeis LTDA

(a) (b)

Figure 8.13 Two-way travel times for (a) horizon A, and (b) horizon B. Time is in seconds.

8.2.3 Calculate Moveout Curves


Because velocity field and zero-offset times are referenced to the original 20 m × 40 m bins
(the center CMP from the 3 × 3 supergather), the calculations are done directly on the τ − p
domain CMP gathers. Zero-offset traveltimes are calculated, and interval stacking velocities
are determined using Dix’s equation (7.4).
54 3D Data Example

With the interval velocity values and zero-offset travel times, I calculate the τ − p domain
moveout curves and the equivalent traveltimes (equations 5.15 and 5.16 respectively). Figure
8.14 shows RMS velocity profile and its converted interval stacking velocity section. Because
of the use of isotropic velocities, the calculated moveout curves might not align with the
actual data. To have a better alignment, moveout curves are snapped to the instantaneous
amplitude of the τ−p domain gathers. Figure 8.15 shows the CMP gather in the τ−p domain
as amplitude data, and as instantaneous amplitude data.

8.2.4 Spectral Decomposition


To reduce the total computational time, I limit the total number of traces to be transformed
into the time-frequency domain. Traces occurring at horizontal slownesses bigger than
that of the refraction of horizon B are first eliminated (See section 7.3.1). Additionally, I
determined the critical angle from the interval velocity data using search windows of length
10 ms. With the critical angle, the maximum horizontal slowness trace for horizon B is
calculated for each CMP using equation 7.6 (figure 8.16). Figure 8.17 shows the velocity data
and the critical angles calculated within a 10 ms search window.

Finally, I calculate the natural log spectral ratio for each of the reflection points between
horizons A and B. Each trace has its own traveltime difference calculated from the equivalent
traveltime curves. Figure 8.18 shows an example of the natural log spectral ratio surface
between horizons A and B. Each positive or negative notch in the surface has an associated
frequency position, which depends on the trace considered.

(a) (b)

Figure 8.14 Stacking velocity profile (a), and its interval stacking velocity profile (b) calculated
using Dix’s equation. Values are shown in m/s.
8.2 PSQI Process 55

(a) (b)

Figure 8.15 CMP gather in the τ − p domain as (a) amplitude data, and (b) instantaneous
amplitude data.

Figure 8.16 Maximum horizontal slowness trace for horizon B


56 3D Data Example

Figure 8.17 From left to right, the stacking velocity, interval stacking velocty, and critical
angle for each time sample.

Figure 8.18 Natural log spectral ratio surface between horizons A and B.

8.2.5 Inversion for 1/Q


The inversion for 1/Q is done with and without a weighting function. To compute the
weighted inversion, the natural log spectral ratios are calculated for each CMP individually,
and the mean and standard deviation are calculated for each CMP. The calculated mean is
8.3 PSQI Results 57

input into the inversion, which is weighted using equation 4.9.

The inversion of the natural log spectral ratio surface demands a bandwidth to be specified.
For this work I choose a bandwidth of 15 Hz to 75 Hz, which is coherent with the frequency
values of the reflections.

8.3 PSQI Results


Using the inversion solution, I calculate for each CMP the value of 1/Q and it´s uncertainty
(section 4). The final result is four maps of 1/Q representing the weighted and unweighted
inversion solutions, and its corresponding uncertainty maps.

8.3.1 Weighted
The maps of 1/Q and their uncertainty from a weighted inversion are shown in figure 8.19.
These maps show the attenuation measured between horizons A and B. To highlight the
results of the inversion, I apply a simple spatial filter to the mapped data. The filter smooths
the data using a mean filter over a rectangle of size (2Nr + 1) × (2Nc + 1), where Nr is the
number of points used to smooth rows, and Nc is the number of pints used to smooth
columns. The results of the spatial filtered map is shown in figure 8.20. The filtered map
shows various zones of high attenuation values, four high 1/Q areas are marked showing the
potential zones for hydrocarbon content within the reservoir.

8.3.2 Unweighted
Figure 8.19 shows the 1/Q map and its corresponding uncertainty using the unweighted
inversion. To understand better the results, I applied the same spatial filter used for the
weighted solution. The filtered map is displayed in figure 8.20. Although this result shows
many of the trends seen in figure filtered weighting solution, the attenuation values obtained
with the unweighed inversion are bigger. This may be due to interference effects that are not
corrected in the unweighted solution, allowing the presence of sparse high values.
58 3D Data Example

(a) (b)

(c) (d)

Figure 8.19 The weighted PSQI map of 1/Q between horizons A and B is shown on (a). The
frequency bandwidth used is 15 Hz to 75 Hz, and the colorbar is clipped below 0. The
uncertainty map of the weighted inversion between horizons A and B is given in (b). (c)
shows the unweighted PSQI map of 1/Q between horizons A and B with the same frequency
bandwidth used in the weighed inversion. The uncertainty map of the unweighted solution
is shown in figure (d).
8.3 PSQI Results 59

(a) (b)

Figure 8.20 The filtered PSQI calculations of weighted 1/Q for the interval between horizons
A and B is shown in (a). likewise, the filtered PSQI calculations of the unweighted 1/Q
solution is shown in (b). The 1/Q color bar is clipped below zero for both results.
Chapter 9

Conclusions

Attenuation contains important information of the petrophysical parameters of the rock


and serves as an indicator of the saturation and mobility of fluids. Unlike other seismic
attributes like AVO, the use of attenuation for reservoir characterization is rare and usually
limited to laboratory experiments that fail to recreate the conditions of a real reservoir.

In this work, I presented a robust method to calculate attenuation, in the form of 1/Q,
from prestack CMP gathers. The PSQI method, developed by Reine (2009) uses three main
components to obtain a robust measurement of effective attenuation.

1. Use of a variable-window time-frequency transform to obtain better spectral estimates.

2. A simultaneous inversion scheme with respect to frequency and time difference.

3. Operate into the τ − p domain to reduce the effects of different raypaths.

I described each of these components and demonstrate the reliability of the PSQI method
by applying it to a real 3D seismic dataset from Colombia. The results obtained, proved the
robustness of the method and define an important tool to be used in the future characteriza-
tion of reservoirs.

9.1 Recommended future work


Although I have proved the effectiveness of the PSQI method to deliver reliable measure-
ments of attenuation, further investigation should be done to make the approach much
stronger. For example, the inversion could be updated to accommodate a non-least-square
solution. Furthermore, the the incorporation of well-log data into the analysis can make the
attenuation measurements more robust. Moreover, applying the PSQI method over different
9.1 Recommended future work 61

sets of seismic data with other acquisition geometries, may give important results to test the
effectiveness of the method in other types of datasets.
Bibliography

Abma, R. and Kabir, N. (2006). 3d interpolation of irregular data with a pocs algorithm.
Geophysics, 71(6):E91–E97.

Aki, K. and Richards, P. G. (2002). Quantitative seismology.

Alkhalifah, T. (1997). Velocity analysis using nonhyperbolic moveout in transversely isotropic


media. Geophysics, 62(6):1839–1854.

Alkhalifah, T. and Tsvankin, I. (1995). Velocity analysis for transversely isotropic media.
Geophysics, 60(5):1550–1566.

Barrero, D., Pardo, A., Vargas, C., and Martínez, J. (2007). Colombian sedimentary basins:
Nomenclature, boundaries and petroleum geology, a new proposal. In Agencia Nacional
de Hidrocarburos, volume 1, page 92. ANH and B&M Exploration Ltd Bogotá.

Bath, M. (1974). Spectral analysis in geophysics: Developments in solid earth geophysics.

Berryhill, J. R. (1979). Wave-equation datuming. Geophysics, 44(8):1329–1344.

Bevc, D. (1997). Flooding the topography: Wave-equation datuming of land data with rugged
acquisition topography. Geophysics, 62(5):1558–1569.

Biot, M. A. (1956a). Theory of propagation of elastic waves in a fluid-saturated porous solid.


i. low-frequency range. The Journal of the acoustical Society of america, 28(2):168–178.

Biot, M. A. (1956b). Theory of propagation of elastic waves in a fluid-saturated porous solid.


ii. higher frequency range. the Journal of the Acoustical Society of America, 28(2):179–191.

Bracewell, R. N. (1986). The Fourier transform and its applications, volume 31999. McGraw-
Hill New York.

Brown, A. R. (2011). Interpretation of Three-Dimensional Seismic Data, 7th Edition. AAPG


memoir 42; Investigations in geophysics no. 9. Published jointly by the American Asso-
ciation of Petroleum Geologists and the Society of Exploration Geophysicists, seventh
edition edition.

Cary, P. W. (1998). The simplest discrete radon transform. In SEG Technical Program Ex-
panded Abstracts 1998, pages 1999–2002. Society of Exploration Geophysicists.

Chopra, S. and Marfurt, K. J. (2013). Preconditioning seismic data with 5d interpolation for
computing geometric attributes. The Leading Edge.
64 Bibliography

Clark, R. A., Benson, P. M., Carter, A. J., and Moreno, C. A. G. (2009). Anisotropic p-wave
attenuation measured from a multi-azimuth surface seismic reflection survey. Geophysical
Prospecting, 57(5):835–845.

Costain, J. K. and Çoruh, C. (2004). Basic Theory in Reflection Seismology: with MATHEMAT-
ICA Notebooks and Examples on CD-ROM, volume 1. Elsevier.

Cox, M. (1999). Static corrections for seismic reflection surveys. Society of Exploration
Geophysicists.

Dasgupta, R. and Clark, R. A. (1998). Estimation of q from surface seismic reflection data.
Geophysics, 63(6):2120–2128.

Dasios, A., Astin, T., and McCann, C. (2001). Compressional-wave q estimation from full-
waveform sonic data. Geophysical Prospecting, 49(3):353–373.

Diebold, J. B. and Stoffa, P. L. (1981). The traveltime equation, tau-p mapping, and inversion
of common midpoint data. Geophysics, 46(3):238–254.

Dix, C. H. (1955). Seismic velocities from surface measurements. Geophysics, 20(1):68–86.

Ebrom, D. (2004). The low-frequency gas shadow on seismic sections. The Leading Edge,
23(8):772–772.

Flandrin, P. (1998). Time-frequency/time-scale analysis, volume 10. Academic press.

Futterman, W. I. (1962). Dispersive body waves. Journal of Geophysical research, 67(13):5279–


5291.

Gladwin, M. T. and Stacey, F. (1974). Anelastic degradation of acoustic pulses in rock. Physics
of the Earth and Planetary Interiors, 8(4):332–336.

Gröchenig, K. (2013). Foundations of time-frequency analysis. Springer Science & Business


Media.

Hake, H. (1986). Slant stacking and its significance for anisotropy. Geophysical Prospecting,
34(4):595–608.

Hall, M. (2006). Resolution and uncertainty in spectral decomposition. First Break, 24(12):43–
47.

Harris, F. J. (1978). On the use of windows for harmonic analysis with the discrete fourier
transform. Proceedings of the IEEE, 66(1):51–83.

Jannsen, D., Voss, J., and Theilen, F. (1985). Comparison of methods to determine q in shallow
sediments from vertical reflection seismograms. Geophysical Prospecting, 33(4):479–497.

Knopoff, L. (1964). Department of physics and institute of geophysics and planetary physics
university of california, los angeles. Rev Geophys, 2(4):625–660.

Levin, F. K. (1990). Reflection from a dipping plane—transversely isotropic solid. Geophysics,


55(7):851–855.
Bibliography 65

Liner, C. L. (2012). Elements of seismic dispersion: A somewhat practical guide to frequency-


dependent phenomena. Society of Exploration Geophysicists.

Liu, B. and Sacchi, M. D. (2004). Minimum weighted norm interpolation of seismic records.
Geophysics, 69(6):1560–1568.

Liu, X. (1999). Ground roll supression using the karhunen-loeve transform. Geophysics,
64(2):564–566.

Macride, C. G. and Kanasewich, E. F. (1987). Seismic attenuation and poisson’s ratios in oil
sands from cfiosshole measurements’.

Marín, J. P., Darío Bermúdez, H., Aguilera, R., Jaramillo, J. M., Rodríguez, J. V., Ruiz, E. C., and
Cerón, M. R. (2010). Evaluación geológica y prospectividad sector sinú-urabá. Boletín de
Geología, 32(1).

Menke, W. (2012). Geophysical data analysis: discrete inverse theory: MATLAB edition,
volume 45. Academic press.

Müller, T. M., Gurevich, B., and Lebedev, M. (2010). Seismic wave attenuation and dispersion
resulting from wave-induced flow in porous rocks—a review. Geophysics, 75(5):75A147–
75A164.

Naghizadeh, M. and Sacchi, M. D. (2010). On sampling functions and fourier reconstruction


methods. Geophysics, 75(6):WB137–WB151.

Olaya, I. (1994). Proyecto cuencas fronteras cuenca sinú - san jacinto. Ecopetrol-ICP.

Pride, S. R. (2005). Relationships between seismic and hydrological properties. Hydrogeo-


physics, pages 253–290.

Raikes, S. and White, R. (1984). Measurements of earth attenuation from downhole and
surface seismic recordings. Geophysical Prospecting, 32(5):892–919.

Raji, W. and Rietbrock, A. (2013). Attenuation (1/q) estimation in reflection seismic records.
Journal of Geophysics and Engineering, 10(4):045012.

Reine, C. (2009). A Robust Prestack Q-Inversion In The tp Domain Using Variable-Window


Spectral Estimates. PhD thesis.

Reine, C., Clark, R., and van der Baan, M. (2012a). Robust prestack q-determination using
surface seismic data: Part 1—method and synthetic examples. Geophysics.

Reine, C., Clark, R., and van der Baan, M. (2012b). Robust prestack q-determination using
surface seismic data: Part 2—3d case study. Geophysics.

Sánchez, C. and Permanyer, A. (2006). Origin and alteration of oils and oil seeps from the
sinú-san jacinto basin, colombia. Organic geochemistry, 37(12):1831–1845.

Schultz, P. S. and Claerbout, J. F. (1978). Velocity estimation and downward continuation by


wavefront synthesis. Geophysics, 43(4):691–714.
66 Bibliography

Spencer, T., Sonnad, J., and Butler, T. (1982). Seismic q—stratigraphy or dissipation. Geo-
physics, 47(1):16–24.

Stockwell, R. G., Mansinha, L., and Lowe, R. (1996). Localization of the complex spectrum:
the s transform. IEEE transactions on signal processing, 44(4):998–1001.

Stoffa, P. L., Buhl, P., Diebold, J. B., and Wenzel, F. (1981). Direct mapping of seismic data to the
domain of intercept time and ray parameter—a plane-wave decomposition. Geophysics,
46(3):255–267.

Stoica, P. and Moses, R. L. (1997). Introduction to spectral analysis, volume 1. Prentice hall
Upper Saddle River, NJ.

Taner, M. T. and Koehler, F. (1981). Surface consistent corrections. Geophysics, 46(1):17–22.

Taylor, J. (1997). Introduction to error analysis, the study of uncertainties in physical mea-
surements.

Thomsen, L. (1986). Weak elastic anisotropy. Geophysics, 51(10):1954–1966.

Toksöz, M., Johnston, D. H., and Timur, A. (1979). Attenuation of seismic waves in dry and
saturated rocks: I. laboratory measurements. Geophysics, 44(4):681–690.

Tonn, R. (1991). The determination of the seismic quality factor q from vsp data: A compari-
son of different computational methods. Geophysical Prospecting, 39(1):1–27.

Trad, D. (2008). Five dimensional seismic data interpolation. In SEG Technical Program
Expanded Abstracts 2008, pages 978–982. Society of Exploration Geophysicists.

Tsvankin, I. (1997). Anisotropic parameters and p-wave velocity for orthorhombic media.
Geophysics, 62(4):1292–1309.

Tsvankin, I., Gaiser, J., Grechka, V., Van Der Baan, M., and Thomsen, L. (2010). Seis-
mic anisotropy in exploration and reservoir characterization: An overview. Geophysics,
75(5):75A15–75A29.

Tsvankin, I. and Thomsen, L. (1994). Nonhyperbolic reflection moveout in anisotropic


media. Geophysics, 59(8):1290–1304.

Turner, G. (1990). Aliasing in the tau-p transform and the removal of spatially aliased
coherent noise. Geophysics, 55(11):1496–1503.

Ursin, B. (1989). Offset-dependent geometrical spreading in a layered medium. In SEG


Technical Program Expanded Abstracts 1989, pages 1156–1159. Society of Exploration
Geophysicists.

van der Baan, M. (2004). Processing of anisotropic data in the τ-p domain: I—geometric
spreading and moveout corrections. Geophysics, 69(3):719–730.

van der Baan, M. and Kendall, J. M. (2002). Estimating anisotropy parameters and traveltimes
in the τ-p domain. Geophysics, 67(4):1076–1086.
Bibliography 67

White, J. (1975). Computed seismic speeds and attenuation in rocks with partial gas satura-
tion. Geophysics, 40(2):224–232.

Wiggins, R. A., Larner, K. L., and Wisecup, R. D. (1976). Residual statics analysis as a general
linear inverse problem. Geophysics, 41(5):922–938.

Winkler, K. W. and Nur, A. (1982). Seismic attenuation: Effects of pore fluids and frictional-
sliding. Geophysics, 47(1):1–15.

Yilmaz, Ö. (2001). Seismic data analysis: Processing, inversion, and interpretation of seismic
data. Society of exploration geophysicists.

Zhang, C. and Ulrych, T. J. (2002). Estimation of quality factors from cmp records. Geophysics,
67(5):1542–1547.

Zhou, H.-l., Wang, C.-c., Marfurt, K. J., Jiang, Y.-w., and Bi, J.-x. (2016). Enhancing the
resolution of non-stationary seismic data using improved time–frequency spectral mod-
elling. Geophysical Supplements to the Monthly Notices of the Royal Astronomical Society,
205(1):203–219.
Appendix A

Other methods to measure Q

A.1 Risetime Method


This is an empirical method based on the measurement of acoustic pulses propagating in
massive rocks. (Gladwin and Stacey, 1974) defines τ as the pulse rise time, and t as the time
1
of propagation of a pulse. This times are related to Q by:

Z t 1
τ = τ0 +C dt (A.1)
0 Q

Where C is a constant whose value is estimated experimentally. Equation A.1 can also be
expressed as the difference between two times, τN for a layer N and τ0 . This is given by:

N T
1
τ N − τ0 = C
X
(A.2)
i =1 Q i

Where Ti is the traveltime within layer i , and Q i is the quality factor or layer i (Jannsen
et al., 1985). From A.2 the effective quality factor Q e described on equation 2.10 for the layer
between N − 1 and N is:
T N − T N −1
µ ¶
Q eN =C (A.3)
τN − τN −1
Appendix B

Multimensional algorithms

B.1 MWNI
This method begins with a vector x of length N sampled on a regular grid.

x = x 1 , x 2 , x 3 , ...., x N

With a vector y of observed data given by:

¡ ¢T
y = x n(1) , x n(2) , x n(3) , ...., x n(N )

And a multi-dimensional matrix T of the specified output sampling defined by:

Ti , j = δn(i ), j

Where δ is the Kronecker operator. Liu and Sacchi (2004) defined a linear relationship
between the data and the observations given by the linear system:

y =Tx (B.1)

Since this is a frequency-domain interpolation algorithm, the discrete Fourier transform


(DFT) is introduced.
1 X N −ı2π(m−1)(k−1)
Xk = p xn e N (B.2)
N n=1
With inverse discrete Fourier transform:

1 X N −ı2π(m−1)(k−1)
xn = p Kk e N (B.3)
N k=1
70 Multimensional algorithms

(Bracewell, 1986). In chapter 4, I discussed the non-uniqueness nature of inverse linear


theory. To approximate the best solution, I defined a PSQI weighed inversion. Likewise,
2
MWNI uses a solution that minimizes a weighted model norm: ||x||W . The wavenumber-
domain norm is defined as:
2
X X k∗ X k
||x||W = (B.4)
k∈κ P k2
Where P k represents the spectral power at wavenumber k, and κ indicates the region of
spectral support of the signal. A diagonal matrix ∧ is introduced as:

P 2
k
k ∈κ
∧k = (B.5)
0 k✚
∈κ

Hence, the wavenumber-domain norm can be expressed as:

2
||x||W = X H ∧s X (B.6)

Where X H is the the complex conjugate transform of X , and ∧s is the pseudoinverse of the
2
diagonal matrix ∧. Then, ||x||W may be expressed as:

2
||x||W = xHQsx (B.7)

Where Q s = F H ∧s F . And F is the discrete Fourier transform. The minimum norm solution
is found by minimizing the folloing cost function:

J = b T T x − y + ||x||W
2
¡ ¢
(B.8)

Where b is the vector of Lagrange multipliers (Liu and Sacchi, 2004). Finally, minimizing J
with respect to x leads to the MWNI solution:

¢−1
x = QT T T QT T
¡
y (B.9)

You might also like