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

net/publication/283196210

Ray-based seismic modeling of geologic models: Understanding and

analyzing seismic images efficiently

Article in Interpretation · November 2015

DOI: 10.1190/INT-2015-0061.1

CITATIONS READS

26 1,198

6 authors, including:

Isabelle Lecomte Paul Lubrano-Lavadera

University of Bergen SWECO
130 PUBLICATIONS 1,825 CITATIONS 10 PUBLICATIONS 51 CITATIONS

SEE PROFILE SEE PROFILE

Ingrid Anell Simon John Buckley

University of Oslo Norce Research
26 PUBLICATIONS 397 CITATIONS 98 PUBLICATIONS 1,786 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Tectonics View project

Trias North View project

All content following this page was uploaded by Isabelle Lecomte on 14 December 2015.

The user has requested enhancement of the downloaded file.

 t Special section: Geophysical modeling for interpreters

Ray-based seismic modeling of geologic models: Understanding

and analyzing seismic images efficiently
Isabelle Lecomte1, Paul Lubrano Lavadera2, Ingrid Anell3, Simon J. Buckley4,
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Daniel W. Schmid3, and Michael Heeremans3

Abstract
Often, interpreters only have access to seismic sections and, at times, well data, when making an interpre-
tation of structures and depositional features in the subsurface. The validity of the final interpretation is based
on how well the seismic data are able to reproduce the actual geology, and seismic modeling can help constrain
that. Ideally, modeling should create complete seismograms, which is often best achieved by finite-difference
modeling with postprocessing to produce synthetic seismic sections for comparison purposes. Such extensive
modeling is, however, not routinely affordable. A far more efficient option, using the simpler 1D convolution
model with reflectivity logs extracted along verticals in velocity models, generates poor modeling results when
lateral velocity variations are expected. A third and intermediate option is to use the various ray-based ap-
proaches available, which are efficient and flexible. However, standard ray methods, such as the normal-inci-
dence point for unmigrated poststack sections or image rays for simulating time-migrated poststack results,
cannot deal with complex and detailed targets, and will not reproduce the realistic (3D) resolution effects
of seismic imaging. Nevertheless, ray methods can also be used to estimate 3D spatial prestack convolution
operators, so-called point-spread functions. These are functions of the survey, velocity model, and wavelet,
among others, and therefore they include 3D angle-dependent illumination and resolution effects. Prestack
depth migration images are thus rapidly simulated by spatial convolution with detailed 3D reflectivity models,
which goes far beyond the limits of 1D convolution modeling. This 3D convolution modeling should allow geol-
ogists to better assess their interpretations and draw more definitive conclusions.

Introduction level. When faced with case-by-case interpretation is-

Seismic data are key to retrieving structural and pet- sues, an interpreter may want to be more confident
rological information about the earth’s subsurface about geologic models. Seismic modeling is therefore
within a targeted zone. This is most notably evident a very useful tool for conceptual testing, as long as
in oil and gas exploration, but it is also relevant to an appropriate method is used and all necessary input
CO2 sequestration, mining, geothermal, and various information is made available.
near-surface applications. Seismic interpreters are Ideally, seismic modeling of geologic models should
geoscientists with various backgrounds and need to be a prerequisite to validate seismic interpretation, as is
be acquainted with geology and geophysics. However, the case in any inversion approach in which data are
geophysicists involved in seismic imaging (Gray, 2001), used to infer a model. Such inversions compare the ob-
in charge of producing the seismic sections interpreters served data (here the seismic image) and the modeled
will work with, are seldom in direct touch with the lat- one. The discrepancy is then estimated and can be used
ter. Valuable information about what may lie behind the to update the model if the disagreement is too large, fur-
seismic images is not always provided, and although ther iterating the process until the best possible conver-
most interpreters are well aware of seismic imaging gence between observed and modeled data is found.
principles and potential pitfalls, e.g., illumination and Though (iterative) inversions may be carried out for
resolution issues, this often stays at the theoretical some stages of the complex seismic processing se-

1
NORSAR, Kjeller, Norway and University of Oslo, Department of Geosciences, Oslo, Norway. E-mail: [email protected];
[email protected].
2
NORSAR, Kjeller, Norway. E-mail: [email protected].
3
University of Oslo, Department of Geosciences, Oslo, Norway. E-mail: [email protected]; [email protected]; michael.heeremans@
geo.uio.no.
4
Uni Research, CIPR, Bergen, Norway. E-mail: [email protected].
Manuscript received by the Editor 22 March 2015; revised manuscript received 10 July 2015; published online 25 September 2015. This paper
appears in Interpretation, Vol. 3, No. 4 (November 2015); p. SAC71–SAC89, 20 FIGS., 2 TABLES.
http://dx.doi.org/10.1190/INT-2015-0061.1. © 2015 Society of Exploration Geophysicists and American Association of Petroleum Geologists. All rights reserved.

Interpretation / November 2015 SAC71

 quence, this is seldom performed at the interpretation Full-wavefield and ray modeling: Both
level. Key issues relate to the appropriate seismic mod- are needed
eling method to use and the cost of modeling, which is a Full-wavefield methods, such as FD, rely on direct
critical resource question during exploration and pro- numerical solutions to the wave equation, whereas
duction. RB approaches follow ray theory, i.e., high-frequency
Among the various existing modeling methods (e.g., approximation of this equation. The latter requires
Fagin, 1991; Behzad, 2012), we will only refer to two new equations to be solved: the eikonal equation for
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

main classes here: the so-called full-wavefield ones, rays and traveltimes and the transport equation for am-
among which finite-difference (FD) modeling is best plitudes along the calculated rays. Both modeling ap-
known and used — thus considered here as the refer- proaches are widely used, with each having inherent
ence for that class — and ray-based (RB) approaches. advantages and drawbacks. As such, they often comple-
These two classes cover methods of modeling general ment each other, neither being able to solve all model-
2D and 3D structures as is our interest here. Note that in ing needs alone. A detailed discussion on FD “versus”
1D models, i.e., horizontally stratified structures with- RB modeling is beyond the scope of this paper, but a
out lateral velocity variations, so-called reflectivity few key points should be mentioned.
methods (e.g., Sen, 2014) should be used instead be- Table 1 provides a brief comparison of the main fea-
cause these methods accurately calculate all wave tures and application domains of FD and RB modeling,
types with a relatively small computation effort. intended to show the complementary nature of the
Beyond conventional approaches, of interest for in- methods. FD modeling provides complete seismo-
terpretation work and which will thus be mentioned, grams, but with a high associated computing cost
RB modeling is also capable of estimating 3D spatial and in a “black-box” manner. A velocity model, survey,
prestack convolution operators called point-spread and wavelet information are input, and all possible
functions (PSFs; Lecomte and Gelius, 1998; Lecomte, waves are then automatically generated, with or with-
2008a). PSFs can then be used to rapidly simulate out box-boundary contributions (e.g., free-surface ef-
prestack depth migration (PSDM) images, without gen- fects). Analyzing the propagation may be difficult,
erating synthetic seismograms and processing them even if one can generate animations to visualize the
(Lecomte et al., 2003). That specific type of RB model- propagation, and only seismograms are produced. RB
ing is similar in application and efficiency to the well- modeling, in its traditional implementation, rapidly
known 1D convolution modeling method, extensively and accurately provides synthetic seismograms for
used in industry. However, such a 3D spatial prestack user-selected wave phases (P, S, reflection, transmis-
convolution approach goes beyond the major limits of sion, conversion, primaries, multiples, etc.). It may,
1D convolution and provides robust and flexible PSDM however, miss several wave types (e.g., headwaves,
modeling, including 3D angle-dependent illumination surface waves, and diffractions), although there are
and resolution effects. The latter are especially impor- various RB approximations to estimate them too
tant for a better constrained interpretation analysis, as (Gjøystdal et al., 2007). Despite this and in comparison
will be illustrated. The FD and RB modeling types are with FD, additional information, such as traveltime, am-
first briefly compared, for the sake of information, be- plitude, polarization, etc., is provided by RB modeling.
fore concentrating on the latter type, which is the main These propagation parameters are indeed useful be-
focus of the present paper. yond seismograms generation, such as for illumina-

Table 1. Comparison between RB and FD modeling.

RB modeling FD modeling

High-frequency approximation No high-frequency approximation

Flexible wavefield subselection Full-wavefield simulations
Missing some wave types, diffraction possible Contains headwaves, surface waves, diffractions, etc.
Smooth macromodels Gridded models
Fast and accurate Time and memory consuming
Target-oriented approach Whole model (black-box)
Synthetic seismograms for understanding real data Complete synthetic seismograms for processing tests
Analysis of wavefield parameters Propagation analyses by movies
Several wavelets Predefined frequency range
Synthetic seismograms and more Only synthetic seismograms
Illumination and miscellaneous studies Benchmark data set
Input parameters for Kirchhoff-based migration methods Used for full-waveform migration methods

SAC72 Interpretation / November 2015

 tion studies in survey planning and as input to migra- to obtain 3D spatial prestack ones, which allow for a
tion codes (e.g., Laurain et al., 2004; Gjøystdal et al., more in-depth understanding of seismic imaging. Some
2007). basic elements of seismic modeling by convolution ap-
There is no doubt that an ideal modeling approach proaches will be briefly reviewed. Because routine seis-
should consist of generating complete synthetic seismo- mic processing suppresses multiple arrivals, we limit
grams, before using them according to the processing ourselves to modeling of primary reflections in our ex-
sequence of field data. Complete synthetic seismo- amples, though some of the RB methods can deal with
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

grams in general 2D/3D models are best obtained by multiple reflections in addition. We also use only iso-
FD modeling. FD modeling is thus essential for bench- tropic cases for the sake of simplicity, although RB
marking studies (e.g., Arntsen et al., 2007; Johansen modeling handles anisotropy too.
et al., 2007) and development of advanced seismic
processing techniques. However, FD modeling is a de- 1D time convolution
manding task in terms of time, computer hardware, We need to come back to the basic principles of con-
software, and expertise, and efforts are often joined volving a 1D function with a wavelet to simulate a seis-
for 3D as is the case for the SEG Advanced Modeling mic trace. This is a correct mathematical principle,
Corporation (Pangman, 2007). Interpreters and other widely used in geophysics for decades (Robinson and
domain experts who use seismic data will seldom have Treitel, 1978). The method is mathematically simple
such possibilities, especially for routine work. We and thus easily understood and programmed, hence
therefore further discuss applications of RB modeling
its popularity. The key issues are, however, as follows:
for the specific needs of interpretation, and we include
(1) how to generate the 1D time function so that it prop-
cases that may contribute to improved interpreta-
erly represents some recorded/imaged elastic-wave re-
tion skills by applying modeling to diverse geologic
sponses, (2) the type of wavelet to use, and (3) the
models (e.g., outcrop analogs and numerical geologic
purpose of the generated trace. In standard reflection
models).
seismic modeling cases, the 1D time function consists
of “spikes” at specific times attached to reflected/dif-
Ray-based seismic modeling: Beyond fracted events, their strength being proportional to cal-
seismograms culated amplitude values; other propagation effects
We intentionally use the term “ray-based” instead of might be added too, such as attenuation. That 1D
“ray-tracing” modeling to broaden the concept. The time-convolution process is actually fundamental to
term “ray” is used here as a synonym to the underlying RB-modeled synthetic seismograms: the calculated
high-frequency approximation applied to the wave traveltimes of the considered waves, in prestack modes
equation, rather than specifically for the raypath, which (i.e., modeling “raw” recordings), define the spike loca-
is not always calculated in some of the RB methods.
tions of the time function and other parameters con-
This is, for example, the case of the eikonal solvers
strain the strength and shape of the signals for each
(e.g., Podvin and Lecomte, 1991), which aim to effi-
wave. This leads to synthetic data that can be used
ciently and robustly calculate the traveltime of the first
for processing tests, wave phase identifications, etc.
arrival for specific applications (refraction seismic,
In the following, however, we will first consider the spe-
earthquake locations, etc.). Rays are computed only
cific case of modeling poststack seismic, which is more
as a by-product if desired. These methods may be in-
suited for interpretation, showing how the convolution
deed sufficient for the convolution approaches we will
discuss here. Solving high-frequency approximations of principle is often used in that case.
wave equations in classic RB approaches, however,
leads to the calculation of other important parameters Modeling poststack seismic with 1D time
for each considered wave mode, i.e., not only traveltime convolution
but also amplitude, polarization, etc. These parameters In many cases, interpreters will work on poststack
are not specifically generated for synthetic seismo- time-migrated seismic sections. For flat structures,
grams; much can be learned from analyzing them as stacked-only sections can be used without any migra-
stand-alone products for a better understanding of tion, although that is not recommended. Performing a
wave propagation and waveforms, rays being also a stack means, in any case, generating seismic sections
useful visual support of these analyses. Interpreters will seemingly corresponding to zero-offset acquisition,
benefit from more systematically using RB modeling in i.e., with the source and receiver at the same location,
its standard forms because such modeling is fast and though data originally contain nonzero offset informa-
interactive. tion. To simulate such stacked seismic sections, mi-
Let us come back to the specific needs of inter- grated or not, three RB approaches may be used,
preters, which are to obtain rapid modeling results, two explicit (normal-incidence point [NIP] and image
while still accounting for real (complex) cases with ray [IR]) and one implicit, known as standard 1D con-
proper propagation and seismic imaging effects. This volution or 1D trace modeling, although the term of
kind of RB modeling relies on convolution approaches, modeling is not always well defined in that latter
starting from 1D poststack time cases, and progressing case.

Interpretation / November 2015 SAC73

 Standard 1D trace modeling cordingly. However, in 1D trace modeling, traveltimes
For decades, a dominating model has been the are often systematically extracted along “vertical”
horizontally stratified medium with elastic parame- lines, in the best case directly from a previously
ters only varying with depth, i.e., a 1D model. It time-migrated section, but often directly from a depth
has been used to design seismic acquisition, process- model or a well log. This may be misleading: zero-off-
ing, imaging, and inversion. With such a model, set rays are not vertical in non-1D models. Figure 1
it is relatively simple to obtain traveltimes of zero- illustrates 1D trace modeling in a simple synthetic
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

offset primary reflections as the sum of two-way ver- model in which reflectivity of a selected modeling
tical times in each layer, thus obtaining analytically target is extracted along vertical lines in depth; we
the time function needed for the convolution pro- will later see the consequences of that. Note first
cess. Zero-incidence reflection coefficients are also the structural differences between the true depth-
straightforward to calculate and, when combined version of the target and the simulated “time-
with the time function, a reflectivity log is formed, migrated” one.
ready to convolve with a wavelet. The obtained Let us now review the contribution of standard RB
zero-offset trace is then representative of a recorded methods for a more appropriate modeling of poststack
seismogram if the actual structure is 1D, and the seismic sections. A discussion on adapted models for
source and receiver are colocated. It can also be con- such modeling will be given in the next section.
sidered as an ideal time-migrated trace (no other am-
plitude effects than reflection strength), again if the Normal-incidence point method for poststack unmigrated
actual structure is 1D. However, even in this case, sections
although the vertical (time) resolution is properly es- The stacking process of reflection seismic intends to
timated (function of the frequency content of the con- simulate zero-offset acquisition from offset data, while
sidered wavelet), no lateral resolution effects are also reducing the data quantity and improving signal-to-
considered, effects which are inherent to the local fo- noise ratio. RB approaches are especially well adapted
cusing in seismic imaging. Horizontal resolution ef- to visualize “poststack” propagation. This is because in
fects are thus often ignored, even though they most cases, zero-offset propagation corresponds to the
influence amplitudes and structural information, in- same raypath, “down” and “up.” The reflected ray is
cluding in relatively flat cases. When the actual struc- perpendicular to the reflector (in isotropic cases), ac-
tures are not 1D at all, i.e., with lateral velocity cording to Snell’s law, and the reflection point is called
variations, the time function has to be adapted ac- the NIP. Figure 2 uses the model from Figure 1 and
shows such NIP rays for overburden
and target reflectors. The corresponding
traveltimes are represented as spiko-
grams (spike at the arrival time) for a
simple visualization of the correspond-
ing time horizons. Amplitudes and other
parameters are also available, however,
which can be analyzed and/or used to
generate complete stacked sections.
NIP modeling can also include diffrac-
tions generated at intersections of depth
horizons and which will appear as hy-
perbolic-like features on the spiko-
grams. The reflections on NIP sections
may look like geologic time horizons,
although there are well-known excep-
tions (e.g., the classic bow-tie example
of a syncline; see later). NIP modeling
is indeed only meant to be used for
simulating poststack unmigrated time
sections, as done in Anselmetti et al.
Figure 1. Synthetic model 1, trace modeling from depth model. (a) Overburden (1997).
with a top water layer and four constant-velocity layers with lateral structural
variations; (b) the target has a flat horizontal “erosion” level and three up- Image-ray method for poststack time-
ward-curved horizons truncated at the former, with velocity increasing at each migrated sections
reflector; (c) reflectivity logs in time are extracted along vertical (depth) lines for
the target area, converted to time, and convolved with a (time) wavelet; and
The IR method is used instead to sim-
(d) generating traces in panel (c) for several lateral positions and plotting them ulate poststack time-migrated sections,
side by side gives a time section that is supposed to approximate a poststack applying a ray concept introduced by
time-migrated one. Hubral (1977). Although the method re-

SAC74 Interpretation / November 2015

 lies on approximations, limiting its application to small 3D spatial prestack convolution for prestack
dips (Black and Brzostowski, 1994), it helps in under- depth migration modeling
standing time-migration effects. The NIP and IR meth- In complex geologic cases, interpreters should use
ods are very fast to compute and flexible, and are PSDM sections to avoid significant pitfalls produced
recommended, especially in 2D, to get a good insight by seismic imaging applied in the time domain. This
of wave propagation and the possibly complex relation- is assuming that the geophysicists are able to define
ships between time and depth domains, and between a proper interval-velocity model because PSDM is very
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

unmigrated and migrated time domains. Figure 3 shows sensitive to this input. As we are dealing with modeling,
the corresponding IR results for the
model used in Figures 1 and 2. The rays
may look unrealistic, which is a result of
them departing perpendicular to the ac-
quisition surface (for explanations, see
Hubral, 1977). By collecting two-way
traveltimes along these rays at each in-
tersection with a reflector and plotting
the latter in spikograms, a simulated
poststack time-migrated section is ob-
tained.
Figure 4 shows a comparison of the
IR section with that computed from
1D trace modeling (Figure 1). Although
the two sections are similar in shape,
there are significant misties of the
structures, laterally (up to 400 m) and
vertically (0.25 s). Although the three
diffraction points are at the correct lat-
eral positions in the 1D trace-modeling
section, resulting from the reflectivity
logs being extracted vertically from
the depth model, this is in no way proof
of good modeling. Poststack time migra-
tion will indeed mislocate these points
in a laterally varying velocity model,
as properly modeled on the IR section.
The only way that 1D trace modeling
could reproduce a poststack time-mi-
grated section in non-1D models would
be to extract the reflectivity log verti-
cally on the poststack time-migrated
section, after interpretation of key hori- Figure 2. Synthetic model 1, NIP ray tracing: (a) Velocity model, (b) NIP rays
zons in that domain. However, it is far for one overburden reflector, (c) NIP rays for one target reflector, (d) traveltimes
better to use an IR approach instead of all NIP rays (spikograms), (e) diffracted rays from one of the three diffraction
of 1D trace modeling, at least for target points, and (f) spikograms including all diffractions.
models that are not too detailed or com-
plex. To further illustrate NIP/IR sec-
tions, Figure 5 shows another model
with the same overburden as in Fig-
ures 1–4, but with a deep syncline as
the target. The corresponding NIP sec-
tion shows the well-known bow-tie ef-
fect, whereas the IR section shows a
closer resemblance to the geologic fea-
tures. Note, however, how the right side
of the syncline, which turns upward in
the depth model, appears to be turning
downward on the IR section, as may Figure 3. Synthetic model 1, IR tracing. (a) IR rays and (b) corresponding spiko-
happen with poststack time migration. grams.

Interpretation / November 2015 SAC75

 we assume in the following that the so-called back-
ground velocity model needed for PSDM is provided
a priori. We first review key elements of PSDM, before
introducing a 3D spatial prestack convolution approach
for direct modeling of PSDM sections.

Prestack depth migration: Basic principles

Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

We take a homogeneous isotropic velocity model

and consider a point scatterer P (Figure 6). P can also
be seen in the following as a reflection point. Because a
reflector can be approximated as a set of juxtaposed
elementary point scatterers (Huygens’ second princi-
ple), understanding the images of these is key to under-
standing seismic reflection images. For a given source/
receiver (S-R) pair, the point-scatterer response is re-
corded at a so-called scattering traveltime TS−P−R cor-
responding to the sum of the traveltimes from S to P
and from P to R; i.e., TS−P−R ¼ TS−P þ TP−R , whatever
paths the waves follow between the points (straight
lines in the homogeneous case of Figure 6). Any PSDM
method, either RB (Kirchhoff or diffraction-stack type;
e.g., Schneider, 1978) or full wavefield, consists of (1) a
compensation for the two paths SP and PR (the “back-
propagation” part of PSDM), before (2) focusing the re-
corded seismic energy at each potential scattering point
(the “imaging” part of PSDM). Note that for step (1), an
Figure 4. Synthetic model 1, simulation of poststack time-mi- accurate velocity model is necessary to perform mod-
grated sections. Compare (a) the trace modeling of Figure 1 eling of the wave propagation between S-P and P-R.
and (b) the IR section of Figure 3. Compare the positions of Figure 6a shows first synthetic input data for two
the three diffraction points between the two sections. simplified acquisition surveys, i.e., common shot and
zero offset, to later emphasize survey effects in seismic
imaging. Figure 6b shows the spatial loci of points cor-
responding to constant values of scatter-
ing traveltime TS−P−R for a given S-R
pair. These curves, called scattering iso-
chrones, are ellipses in a homogeneous
isotropic model. They are the result of
the sum of the circular isochrones
(wavefronts) TS−P and TP−R departing
from S and R, which are the focal points
of these ellipses. When S and R are at
the same location (zero offset and post-
stack equivalent), the ellipses become
circles. PSDM projects back the re-
corded seismic energy along these
scattering isochrones, but in a “blind”
manner in most methods, especially
commercial and industry standard. The
energy is spread all along the isochrones
because the specific direction the en-
ergy is coming from is unknown (Fig-
ure 6c). Due to the huge amount of
recorded data, one can expect in this
summation process a constructive inter-
ference at the actual point scatterer P
(Figure 6d), and destructive interfer-
Figure 5. Synthetic model 2. (a) Same overburden as in model 1 (Figures 1–3), ence elsewhere, although the data quan-
but with a syncline as the target; (b) NIP traveltimes with bow-tie effect; and tity in itself is no guarantee of this. The
(c) IR traveltimes — compare the syncline shape with panel (a). geometric distribution of the S-R pairs,

SAC76 Interpretation / November 2015

 in combination with the characteristics of wave propa- and are well known in optics, although in seismic imag-
gation through the velocity model, as well as the fre- ing, we generally deal with backscattering (two-way
quency bandwidth of the data, will also influence the energy path) and not direct transmission as in photog-
quality of the focusing (imaging) process at each con- raphy or astronomy (the latter case also applies to seis-
sidered P, as well as outside the scattering structures. mic imaging, i.e., to model tomography responses).
As an example, too coarse S-R sampling will make the Then, having RB-modeled PSFs of PSDM images and
scattering isochrones visible on the final seismic image, recalling Huygens’ second principle, which states that
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

due to a lack of destructive interference; this is the rea- reflectors are juxtaposed secondary sources (point
son for steep striping noise on many migrated seismic scatterers), direct modeling of seismic images is ob-
sections, which affects not only image quality, but also tained by a 3D spatial prestack convolution between
the amplitudes of picked reflections. the PSFs and reflectivity models (Lecomte et al.,
Figure 6c (left panel) shows the elementary PSDM 2003; Lecomte, 2006, 2008b, 2013a, 2013b). Other au-
image for one noncoincident S-R pair, with the ellipse thors have used full-wavefield modeled PSFs for similar
being smeared in a spatially varying pattern as a com- convolution approaches (e.g., Toxopeus et al., 2003). In
bination of the wavelet and a pulse-stretching effect. this work, we compute PSFs from RB algorithms be-
The latter is due to the opening angle between the in- cause it allows a faster and more flexible application
coming ray at P and the outgoing one: the larger that of the 3D spatial prestack convolution approach.
angle, the wider the zone around the scattering iso- It is beyond the scope of the present paper to detail
chrone is, hence the worse the resolution. An offset-de- the 3D spatial prestack convolution method of choice
pendent resolution issue will be illustrated later by (we refer to Lecomte, 2008a; Lecomte and Kaschwich,
modeling examples. In the zero-offset case (Figure 6c, 2008). Figure 7 summarizes the basic elements of such
right panel), the scattering isochrones remain circular modeling. A given velocity model and a survey are used
as S and R are coincident. Although the homogeneous to generate, via various RB modelings, so-called illumi-
point-scatterer model is the same in
both cases, the final images are different
(Figure 6d). The focusing patterns are
not only dissimilar around P, but also
away from it, with some large-scale
cross patterning due to the lack of de-
structive interferences of the scattering
isochrones attached to the extremities
of each survey (limited-aperture effect).
The simple homogeneous example of
Figure 6 illustrates several PSDM issues:
(1) images depend on survey geometry;
(2) resolution varies as a function of off-
set, as well as a function of opening an-
gle at each image point, and thus varies
along the scattering isochrones, even for
a constant offset; and (3) limited-aper-
ture effects may spread across a large
zone of the image. Therefore, having fast
and flexible modeling methods taking
into account all these issues, but with-
out performing the expensive “full-
wavefield modeling and processing,”
should help geologists and interpreters
to better understand the geology from
the seismic images they analyze.

Prestack depth migration modeling by a 3D

spatial prestack convolution approach
In an attempt to provide interpreters
with a solution to the requirements pre-
sented above, RB modeling of PSDM Figure 6. PSDM principles. (a) Point scatterer in homogeneous background
velocity with Born-modeled synthetic (left) common-shot and (right) zero-offset
point-scatterer responses have been gathers; (b) scattering isochrones as sum of wavefronts calculated in back-
studied (Lecomte and Gelius, 1998; Le- ground velocity model for a shot (S) and receiver (R) couple, and at all im-
comte, 2008a). These point-scatterer re- age-point (IP); (c) PSDM response for one trace only and (d) for all traces,
sponses are the PSFs mentioned earlier for (left) common shot and (right) zero offset.

Interpretation / November 2015 SAC77

 nation vectors (ISR ) at a target point P (Figure 7a), i.e., point scatterers but also reflectors (Lecomte et al.,
using a method which can produce the required slow- 2003). The following examples will illustrate that.
ness vectors at that point (pS from S to P, and pR from P PSDM PSFs are spatially varying, even in a homo-
to R); ISR ¼ pR − pS . All possible ISR at P are collected. geneous background. However, for the sake of simpli-
Adding a wavelet (Figure 7b) and possibly other tech- fication, we will assume local targets in the present
nical parameters to account for various propagation/im- paper. In other words, we consider the target to be
aging effects, a so-called PSDM filter is generated in the small enough to correspond to a near-constant PSF
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

wavenumber domain (Figure 7c). An inverse Fast Fou- over the imaged zone. Simulated images will thus be
rier Transform (FFT) is applied to this filter to produce created by convolution between a reflectivity target
the (RB) PSF in the spatial domain (Figure 7d). The and one PSF calculated at the center of the image.
PSDM filter can be seen as an equivalent image-process- The analysis of imaging effects will thus be easier. In
ing operator describing the combined effect of propaga- addition, using a constant PSF over the (local) target
tion and imaging for a scattering structure at P. Its allows the spatial convolution to be performed as a
corresponding PSF is an estimation of a point-scatterer multiplication in the wavenumber domain between
response at P as imaged via PSDM. the PSDM filter and the FFT of the reflectivity. The in-
Figure 8 shows a comparison between the actual verse FFT of that product leads then to the desired spa-
point-scatterer response (i.e., PSDM applied to the syn- tial image. This is highly efficient and permits near-real-
thetic data, as in Figure 6) and the response computed time modeling, which is important for geologists and
from the RB-simulated approach, which assumes a interpreters who may want to test various models. How-
plane-wavefront approximation. As observed, the simu- ever, spatial convolution with space-varying (RB) PSFs
lated match with the exact responses at the point scat- can also be achieved for larger (full-field) targets but
terer and in its vicinity, but the larger cross-pattern with higher computing cost, although this remains far
effect is different, being more curved in the exact re- below the cost of a complete FD modeling followed
sponses. However, most critical is the response in by imaging. Such full-field convolution can be per-
and around the considered point. RB PSFs are thus al- formed at a later stage, if PSFs calculated over a grid
ready indicative of 3D resolution capabilities of the covering the imaging zone indicate clear spatial vari-
wave propagation and image construction during the ability. Prestudies via local-target convolution also help
PSDM, allowing quantitative assessment of the vertical to initiate the modeling parameters for a full-field one.
and lateral resolution of point scatterers. Moreover, Figure 9a shows a model with a homogeneous back-
when used in a spatial prestack convolution modeling, ground, and we consider three versions of a (local) tar-
RB PSFs are capable of reproducing resolution and il- get model at the same depth. The first two models
lumination effects on PSDM images, i.e., not only for consist of isolated point scatterers with different spac-

Figure 7. The 3D spatial prestack convolution and basic elements. (a) Velocity model and illumination vector (ISR ) defined as the
difference between two slowness vectors (pS and pR ) for a shot-receiver (SR) couple at a target point (green dot), (b) ISR for the
considered survey and combined with a selected wavelet to generate (c) the equivalent PSDM filter, and (d) the resulting PSF is
obtained by Fourier transform of panel (c).

SAC78 Interpretation / November 2015

 ing between them, and the third is a
truncated reflector. Let us consider a
survey with a vessel moving from left
to right and with two different offsets:
0 and 4 km. This case has no illumina-
tion issue; i.e., the truncated reflector
will be imaged by both offsets (later ex-
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

amples will illustrate the illumination

effects). Figure 9b and 9c shows the Figure 8. Point-scatterer case of Figure 6. Comparisons between PSDM results
corresponding simulated PSDM im- (red square) and RB-modeled PSFs (black square), for (left) zero-offset and
ages. The PSF footprint is visible on (right) common-shot configuration.
the point scatterer models, whereas it
is less obvious on the truncated reflector. However,
the latter image has to be analyzed as the result of con-
structive/destructive interference patterns of the PSFs
attached to (fictitious) dense point scatterers along the
reflector (the second Huygens’ principle), the two
point-scatterer models being used here to understand
such interferences. Following this, and because there
is no illumination issue, the migrated energy is focused
constructively along the reflector, the PSF footprint
explaining the reflector thickness (vertical resolution)
and some lateral smearing along it, as seen at the
extremities in the far-offset case (lateral resolution).
The latter effect explains why faults are not always
easily detected as simple discontinuities along reflec-
tions. Above and below the reflector itself, the off-
center parts of the PSF are expected to interfere de-
structively, which is observed in Figure 9 due to a
regular survey and dense enough S-R distribution.
Comparing Figure 9b and 9c also shows the difference
in PSFs between zero and far offset. It thus indicates
that resolution has to be seen as a survey-dependent
2D/3D pattern. In addition, the zero-offset case shows
that resolution is not only a vertical effect, as simplis-
tically modeled by 1D convolution, but also a lat-
eral one. Figure 9. PSF and seismic images. (a) Two point-scatterer
Let us now illustrate reflector-illumination issues, models (#1 and #2) and one truncated-reflector model (#3),
each model being located at the same depth and illuminated
i.e., which reflectors will be detected by seismic imag-
by a marine survey going from left to right; (b) simulated
ing. Figure 10a shows a synthetic case in which a flat PSDM images for each target model and with a zero-offset ac-
horizontal target reflector is below a horizontally lay- quisition; and (c) same as in panel (b) but with a 4-km offset.
ered overburden (above reflector point
A) with a salt structure (above reflector
point B). In Figure 10b and 10c, PSDM
filters and their corresponding PSFs
are given for both points. The salt struc-
ture strongly influences the illumination
of point B, reducing the ISR coverage at
that point, hence a narrow PSDM filter
compared with the one at A. As a conse-
quence, a more elongated PSF at B with
a different cross pattern is produced
(Figure 10b). PSDM filters also directly
indicate that the flat reflector at B will Figure 10. PSF and illumination. (a) Synthetic model: overburden containing a
not be illuminated, which is explained salt body above a flat target reflector (black arrow: normal to the reflector) with
two points (A and B) illuminated in a different manner; (b) PSDM filter at A and
by its normal (shown by the black verti- B, with the normal to the reflector being superimposed (black arrow) in the fig-
cal arrow in Figure 10b) not being ure; and (c) corresponding PSFs, the cross-pattern effects (blue and pink lines)
covered by the available ISR range (ar- being due to the truncation of the filter in panel (b) because of limited illumi-
row outside the orange area), whereas nation vector (ISR ) ranges (from Lecomte, 2008a).

Interpretation / November 2015 SAC79

 in A, there is illumination (arrow inside the green along that reflector (Figure 11b) is first used as the in-
area). This example indicates that seismic illumination put of a 1D convolution approach, which does not con-
is not only a problem for steep dips, as is often assumed. sider any survey geometry, leading to an “ideal” image
Indeed, at location B, neither the flat reflector nor of the whole reflector (Figure 11c). In contrast, the si-
steeper reflectors dipping left would be detected, mulated images built for standard real survey geom-
whereas steeper reflectors dipping right would be im- etries and computed by 3D convolution of the
aged (as long as their normal is within the orange zone). corresponding PSFs with the same input reflectivity
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Illumination variability is here due to the survey geom- (Figure 11d–11g) show major differences in resolution,
etry in combination with wave propagation in the over- off-reflector imaging noise, and illumination. In particu-
burden. lar, imaging of the steep reflector parts clearly depends
Figure 11 illustrates the illumination effect caused on the survey geometry and its position.
only by the survey, by considering one selected reflec- Figure 12 illustrates the importance of accounting for
tor from a reservoir model below a homogeneous over- 3D effects in seismic imaging. This figure shows how
burden (Figure 11a). The corresponding reflectivity the same 3D reflectivity cube is seen by seismic imaging
if the acquisition consists of a monostreamer 2D line
(no crossline resolution at all, which affects the vertical
sections by off-side reflections being imaged too), a
multistreamer 2D line (some crossline features are
roughly perceived, whereas the vertical sections be-
come better resolved), or a 3D multistreamer survey
(the crossline structure is now properly imaged). These
are, in theory, well-known effects, though they cannot
be simulated using only a 1D convolution. Instead, a 3D
spatial prestack convolution approach takes into ac-
count wave propagation in the overburden in combina-
tion with the survey geometry and is much faster than
an FD-modeling and processing workflow. The local-
target examples shown here took just a few minutes
on a standard laptop, a computing time comparable
with performing 1D convolution.
So far, the given examples consist of various models,
for the target (the zone to image) and for the back-
ground/overburden (where the wave propagates). They
illustrate the versatility of the RB 3D spatial prestack
convolution approach, particularly in comparison with
1D convolution. Before further investigating the effects
of seismic-image construction through the use of 3D
convolution, we need to discuss the input model itself.

Which geologic models to use?

In standard seismic modeling (either RB or FD), the
user starts from a given global velocity model (one
model only), defines a seismic survey, chooses a pulse
to emit at the source points, and then produces syn-
thetic seismic traces. In such cases, the input velocity
model constrains the wave propagation and backscat-
tering of seismic energy. This is also the case for
the NIP and IR RB approaches reviewed in the first sec-
Figure 11. Survey effects. (a) Reservoir model with one of
tion for direct poststack modeling. FD methods can
the available parameters prior to seismic modeling and loca- deal with arbitrary property grids as models, the cell
tion of a 2D section to image (magenta line); (b) reflectivity being the basic property element. RB approaches
section for a selected reflector including faults; (c) 1D trace require instead so-called macromodels, due to the
modeling of the reflector in panel (b); both steep parts (faults, high-frequency assumption, with the elastic properties
pink and green) are imaged because the modeling cannot deal varying smoothly, possibly between smooth explicit in-
with illumination issues; (d-g) 3D convolution modeling of the terfaces (at which reflection, transmission, and conver-
reflector in panel (b) for various surveys: The resolution, off-
reflector patterns, and illumination issues vary according to sion can be considered), i.e., without geologic details
the survey, the steep part (green) being only illuminated by below a few wavelengths (Gjøystdal et al., 2007).
a vertical seismic profile with a well located on the right side For field-data PSDM, depth imagers first derive a
of the section. smooth background-velocity model from the seismic re-

SAC80 Interpretation / November 2015

 cords themselves (e.g., from standard velocity analyses, interfaces (e.g., salt bodies). However, the background-
reflection tomography, etc.). This model is then used to velocity model should remain smooth within the target.
simulate the wave propagation part during the PSDM. The target model used for 3D convolution, on the
After forward modeling, the imaging part of PSDM ex- contrary, can and should be as detailed as possible,
tracts the backscattered energy contained in the proc- being represented on grids as in FD methods. This is
essed data to “map” in depth the reflectors/diffractors a significant advantage compared with standard RB
onto the initial (smooth) velocity model. In the 3D spa- methods and resembles 1D trace modeling, where re-
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

tial prestack convolution approach using RB PSFs, we flectivity logs can be as detailed as a well log. We should
do not calculate synthetic data, which we process/im- indeed not a priori upscale the properties and let
age to get the desired PSDM section. As modelers, instead the PSF operator automatically perform a “blur-
we assume that the background velocity model (to cal- ring” of the structures according to the wave propaga-
culate the illumination vectors) and a model of the scat- tion and frequency content of the wavelet, resolution,
tering structures (the target to image) are given and illumination effects combined, as would elastic
information. Though these two parts should match waves in real cases. Arbitrarily suppressing geologic
where they overlap (in the target), and may indeed features in input because supposedly too small accord-
be part of the same global model, they are, however, ing to the “expected” resolution might be damaging.
used as two separate entities in the 3D convolution We can now represent detailed target properties on
modeling, thus allowing a flexible use of the target grids using various inputs. Besides synthetic models
model and the background model (e.g., fix one and vary used in earlier figures, we have already seen the case
the other). of reservoir models (Figures 11 and 12). For computa-
When dealing with real-data cases, with planned tional efficiency of reservoir modeling, these are usually
comparisons between actual and modeled seismic, very detailed vertically, but they have lower lateral res-
the background velocity model used to generate ISR olution and are thus not ideal geologic models. Reser-
should preferably be the smooth one used for PSDM voir models are, however, often assigned with sufficient
of the actual data. The aim is indeed to reproduce as parameters to derive the elastic properties needed to
well as possible the wave-propagation effects contained generate the input reflectivity grid, which is a major ad-
in the data. In pure synthetic cases, background-veloc- vantage. Another suitable target input is seismic inver-
ity models can be of a more general type, especially sion results because they are already on grids and
adapted for RB modeling, e.g., for overburden-related provide the necessary elastic parameters (e.g., Buland
effects with analyses of various wave phases at specific and Omre, 2003).

Figure 12. The 3D effects. (a) Reservoir model of Figure 11 with two local targets as examples, the 3D one being used in the
following, (b) geometry of 2D/3D surveys, (c) 3D reflectivity structure of the 3D target in panel (a), (d) modeled seismic for the 2D
line in panel (b) and one streamer (no crossline resolution), (e) modeled seismic for the same 2D line as panel (d) but with five
streamers (coarse crossline resolution), and (f) modeled seismic for the 3D multistreamer survey of panel (b) (proper crossline
resolution).

Interpretation / November 2015 SAC81

 Various geologic models can also be used as target temperature conditions of a buried reservoir. Going
models; e.g., derived from interpretation, outcrop ana- from outcrop to reservoir models prior to seismic mod-
logs, or any geologic modeling, assuming they can be eling is also a possibility (Rotevatn and Fossen, 2011;
gridded and populated with proper elastic parameters. Botter et al., 2015), provided that we do not lose too
Figure 13 is an example of a largely inaccessible cliff much of the lateral variability due to large pillars, as ear-
section displaying a series of growth faults in the sedi- lier mentioned. Figure 14 shows 3D numerical modeling
mentary succession (Anell et al., 2013; Osmundsen et al., of fault zones (Botter et al., 2014a, 2014b), using the 3D
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2014). Digital models, acquired using lidar or photo- spatial convolution to generate seismic cubes from such
grammetry analysis of such outcrops (e.g., Buckley et al., complex and detailed 3D scattering structures. The
2008), provide detailed spatial geometry of outcrop geo- elementary spheres of the geomechanical modeling (Fig-
logic features and structures, whose interpretation can ure 14a) are directly mapped on a cube to generate the
be used as direct input for modeling. The current analog input reflectivity model (Figure 14b), although one key
model of Figure 13 lacks, however, realistic elastic prop- issue again is assigning proper elastic properties. Finally,
erties. The latter should indeed reflect the expected Figure 15a and 15b shows actual geologic folds observed
variations contained in the observed outcrop struc- at all scales in nature, and Figure 15c shows numerical
tures, in combination with some fluids and pressure/ modeling results of such folds (Schmalholz and Schmid,
2012). Folds provide very interesting geo-
logic structures to model and image (Fig-
ure 15d and 15e; see the next section for
further analyses of that specific model),
even when assuming constant properties
within the layers.
The previous examples show that we
now have a large variety of possibilities
to define the (local) target model of the
3D spatial prestack convolution ap-
proach from detailed geologic struc-
tures. This allows seismic modeling of
various key near-vertical and/or com-
plex structures known to be problem-
atic in seismic imaging, such as fault
zones, gas chimneys, sand injectites, ba-
Figure 13. Outcrop modeling. (a) Outcrop analog (Kvalpynten); (b) geologic saltic intrusions, and carbonate res-
interpretation; (c-e) modeled PSDM images; (f) location map; (g) Barents sea ervoirs.
offshore seismic (similar structures). (c and d) Same illumination pattern, but
two different main frequencies; (e) same frequency as in panel (d), but with
a more limited illumination pattern (courtesy of TriasNorth project, University Seismic imaging effects
of Oslo). Seismic modeling of the numerical
fold model of Figure 15c was requested
by geologists because they suspected
that such complex geologic structures,
with high dip variations (from flat to ver-
tical), found at all scales and settings in
nature, would be represented poorly in
seismic images. The geologists wanted
to rapidly get synthetic seismic images,
not having the resources for exhaustive
modeling and processing. The 3D spatial
prestack convolution was thus applied,
using a complex background-velocity
model of a known Norwegian hydrocar-
bon field and thereby generating the
PSF in Figure 15d, which in turn yielded
the simulated PSDM image in Figure 15e.
This example will be further discussed
in later figures, but it already shows
Figure 14. The 3D numerical modeling of fault zones. (a) Geomechanical mod-
eling by the discrete element method for a 3D normal-fault zone with variable how seismic imaging can strongly dis-
displacement along strike, (b) generated reflectivity, (c) 3D convolution model- tort the actual geology, confirming what
ing of seismic, and (d) automatic attribute extraction (fault zone) from modeled the geologists suspected, but could not
seismic. Figure elements courtesy of C. Botter. easily demonstrate without modeling.

SAC82 Interpretation / November 2015

 The latter took just a few minutes on a standard laptop, keep a lateral thickness, being a pointlike structure
allowing many parameters to be varied, using the same of size defined only by the frequency band and the local
target model as a reference. In the following, we do not velocity. Thus, the corresponding image, though getting
change the target model itself because we want to dem- closer to folds, still contains imaging artifacts due to the
onstrate the importance of properly taking into account complex interference effects of the (here 2D) PSF along
various parameters influencing a seismic image, i.e., the reflectivity structure.
showing that the structure of the target itself is not
solely “responsible” for seismic distortions.
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Frequency and resolution/illumination

A common preconception is that seismic imaging
1D versus 3D problems can be improved with a high-frequency
Figure 16 is used to again demonstrate the short- and/or large-frequency band. This is only partly true be-
comings of 1D trace modeling for actual target struc- cause illumination plays an important role too: What is
tures embedded in any background-
velocity model. The velocity model
does not even need to be complex for
that 1D modeling to fail. Figure 16a is
the reflectivity grid used as input for Fig-
ure 15e and for all subsequent modeling
of the fold model (Figure 15c). An alter-
nating set of positive/negative reflec-
tions was assumed by assigning a
lower constant velocity in the black
layers compared with the white ones.
The background velocity model chosen
here was homogeneous, for emphasiz-
ing target-structure effects and allowing
a direct comparison between 1D trace
modeling (time domain) and 3D spatial Figure 15. Numeric geologic models. (a and b) Field fold examples, (c) numeri-
convolution (depth domain), i.e., for cal modeling of folds (see Schmalholz and Schmid, 2012), (d) PSF calculated in
avoiding the pulse-stretching effect of an actual complex velocity model of the Norwegian Continental Shelf, and (e) si-
PSDM due to velocity variations. One- mulated PSDM image when applying the PSF in panel (d) to a reflectivity model
extracted from panel (c).
dimensional trace modeling produces
the result displayed in Figure 16b. This
result corresponds in a sense to a “1D
PSF,” i.e., varying only in the vertical
direction. Except for the near-vertical
parts of the model, we do recognize
the folds, with resolution effects being
only of a vertical nature. Figure 16c,
however, shows a quite different image,
where folds are not so easily identified,
and with features looking like faults, flat
spots, etc. We were using the same fre-
quency content and background veloc-
ity of Figure 16b, but now modeling a
zero-offset 2D line acquisition above
the target, something 1D trace modeling
cannot represent. The image changes
are then due to the more realistic 2D
“footprint” of the combined wave-propa-
gation and image-formation effects (see
the corresponding PSF). The smearing
of the structures is also lateral and not
only vertical, the lateral resolution and
cross pattern being partly due to lim-
Figure 16. Fold model in homogeneous background. (a) Reflectivity (as used in
ited-aperture effect. Even in a perfect Figure 15), (b) 1D convolution (depth-converted for comparison), (c) 3D spatial
acquisition case (no limited aperture prestack convolution with actual (limited aperture) survey, and (d) 3D spatial
and all reflector dips illuminated), as prestack convolution with perfect illumination. The same wavelet is used for
in Figure 16d, the resulting PSF would all modeling experiments, and the PSFs are superimposed in panels (d and c).

Interpretation / November 2015 SAC83

 not illuminated cannot be seen, whatever the frequency changes between Figure 18a and 18b may be subtle,
that is used. Figure 17a is equivalent to Figure 16c, there are differences due to the lateral resolution effect.
whereas Figure 17b uses a much higher dominant fre- Such “geometric” effects will also occur between sea-
quency. In comparison with the input reflectivity of Fig- surface and sea-bottom acquisitions. One-dimensional
ure 16a, the higher frequency yields more structural trace modeling on the other hand would not reproduce
details, but zones with poor illumination display a very these differences.
weak signal. With actual data containing various
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

sources of noise, even after processing, the poorly illu- Survey and offsets
minated areas would disappear. Note that the cross pat- Figure 19 compares seismic images as seen by a zero
tern of the PSF explains the seismic-image “noise” seen offset (Figure 19a) and a larger offset (Figure 19b). The
between reflectors and which also exists on actual im- changes on the PSF are a decreasing lateral resolution
ages. A 1D trace modeling study would miss this inter- with increasing offset and also a change in orientation,
pretation as well as incorporation of illumination which indicates a difference in illumination capability.
effects. As a result, the seismic images are significantly differ-
ent. The observed differences in illumination as a func-
Target location tion of offset are of great importance when studying
Figure 18 uses the same homogeneous background amplitude-versus-offset effects.
velocity model, zero-offset 2D line acquisition, and fre-
quency band as Figure 17a, but the target itself is lo- Background velocity
cated at two different depths. For this geometric Finally, Figure 20 compares background-velocity
reason, the ISR range gets narrower with depth, which model effects between a homogeneous model and a
induces a decrease in lateral resolution. Although the complex field case (as in Figure 15d), with the zero-off-
set 2D-line acquisition and the frequency band being the
same as in Figures 17a, 18a, and 19a. The differences
between these two cases are significant, on PSFs and
on the quality of the seismic images, with the overall
decrease of resolution in Figure 20b due to the higher
velocity in the complex model, and the very different
illumination pattern due partly to near-turning waves
in the field model. This example helps support the argu-
ment that illumination/resolution effects due to the
velocity model need to be taken into consideration
when interpreting seismic images. This is once more
an effect that 1D trace modeling alone cannot take into
account.

Discussion
Though a 1D model representation of the earth’s
structures has been and is still a valuable model for
understanding wave propagation, and thus extracting
useful information out of seismic data, we should strive
to move away from this simplification, even if one is
interested in an efficient modeling approach. RB-gener-
ated PSFs and their subsequent use in 3D spatial pre-
stack convolution modeling of PSDM sections allow
near-interactive assessment of different interpretation
models, especially with a local-target version in a first
approach (one PSF for the considered imaging zone).
The full-field version with space-varying PSFs can then
be used to refine the final model. The 3D spatial convo-
lution is also valid in anisotropic cases (Lecomte and
Kaschwich, 2008) and accounts for diffraction effects
in PSDM (Kaschwich and Lecomte, 2010; Kaschwich
et al., 2011). Table 2 summarizes comparison points be-
tween 1D and 3D convolution modeling discussed in the
Figure 17. Fold model in homogeneous background and 3D present paper and thereby emphasizing the values of 3D
convolution with frequency effects. (a) 10-Hz modeling (same versus 1D convolution approaches.
as in Figure 16b) and (b) 30-Hz modeling. The target, depth, Although computing power continues to improve,
and offset are the same for both cases. the ratio of efficiency between the RB and FD ap-

SAC84 Interpretation / November 2015

 Figure 18. Fold model in a homogeneous
background: 3D convolution with depth ef-
fects. (a) Shallow depth and (b) deeper depth.
The same target, offset, and wavelet are used
for both cases.
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 19. Fold model in homogeneous

background: 3D convolution with offset ef-
fects. (a) Small offset and (b) large offset.
The same target, depth, and wavelet are used
for both cases.

Interpretation / November 2015 SAC85

 proaches will remain, RB modeling being a more effi- One key assumption made for the latter is a locally
cient and less demanding method. However, bench- plane-wavefront approximation, which explains the in-
marking with FD-modeled cases of various complex creasing mismatch of the cross patterns in Figure 8 fur-
structures requiring PSDM imaging is a necessity and ther away from the center (curved in actual PSDM and
could help to further improve RB-generated PSFs. straight in RB-modeled PSFs). However, RB methods
also give the necessary parameters for such curvature
effects (Gelius et al., 2002), and adding them would be
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

straightforward without a significant increase in com-

putation time. Implementing this would also give a bet-
ter match of the migration noise generated by the cross
patterns of the PSF (limited-aperture effect) and due to
imperfect interferences of the scattering isochrones
during imaging. Finally, though the RB-generated PSFs
are especially designed to simulate Kirchhoff-type mi-
gration results, we believe in adding specific prepro-
cessing of the PSDM filters to generate PSFs more
closely suiting full-wavefield migration images.
Key inputs in all modeling discussed earlier are
proper elastic parameters to obtain realistic reflectivity.
Although this major issue has not been specifically ad-
dressed here, it does remain a troublesome point if one
wishes to provide geologists and interpreters with easy-
to-use and efficient modeling tools, without getting in-
volved in complicated (e.g., rock physics) modeling. It
is indeed important to use adequate values of velocities
and density, and defining proper S-wave velocities too,
even if dealing with P-wave seismic (e.g., Øygarden
et al., 2015).
Industry-standard seismic inversion is still guided by
1D convolution modeling. Moving toward inversion us-
ing a 3D convolution operator for the modeling part is,
however, now possible with RB (or FD) PSFs. First-use
attempts of RB PSFs for better wavelet estimation have
been published elsewhere (Georgsen et al., 2010). In ad-
dition, turning the 1D inversion itself into a 3D RB-PSF
constrained inversion, following the Buland and Omre
(2003) inversion scheme, is also prototyped but not yet
Figure 20. Fold model in different background velocity published because thorough testing toward bench-
fields. (a) Homogeneous background and (b) actual velocity marked synthetics and actual PSDM data remain to
field of the Norwegian Continental Shelf. The same target, be completed. Similarly, using a full-wavefield ap-
depth, offset, and wavelet are used for both cases. proach, Letki et al. (2015) recently show that FD PSFs

Table 2. Comparison between 1D and 3D convolution modeling.

1D time poststack convolution 3D spatial poststack convolution

Features and Reflectivity and wavelet as input Reflectivity and wavelet as input
applications No propagation effects taken into account Propagation effects, with velocity model, survey, and other
parameters
Strictly limited to 1D models, i.e., vertical No model limitations, accepting lateral velocity variations
variations only
Fastest and robust Fast and robust
Simulates poststack time-migrated images Simulates PSDM images
Vertical time (1D) resolution 3D spatial resolution
No illumination effects 3D illumination effects
Limited sensitivity analyses Various sensitivity analyses

SAC86 Interpretation / November 2015

 can compensate for illumination effects in a depth-do- northern Barents Shelf during the Triassic: First Break,
main inversion approach, giving better amplitudes lead- 31, 67–76.
ing to improved interpretation. Anselmetti, F. S., G. P. Eberli, and D. Bernoulli, 1997,
Seismic modeling of a carbonate platform margin
Conclusions (Montagna della Maiella, Italy): Variations in seismic
Although 1D convolution modeling has been key to facies and implications for sequence stratigraphy,
seismic acquisition, processing, imaging, and inversion in F. J. Marfurt, and A. Palaz, eds, Carbonate seis-
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

developments, it is time to reconsider its use, espe- mology, Geophysical Development Series: SEG,
cially when dealing with complex earth structures re- 373–406.
quiring PSDM and more advanced interpretations. Arntsen, B., L. Wensaas, H. Løseth, and C. Hermanrud,
However, in such cases, it is not necessary to move 2007, Seismic modeling of gas chimneys: Geophysics,
directly from 1D convolution to FD modeling, the lat- 72, no. 5, SM251–SM259.
ter nevertheless being a requirement for producing Behzad, A., 2012, Seismic modeling of complex geological
complete synthetic data for advanced processing de- structures, in M. Kanao, ed., Seismic waves —
velopments in complex models. A middle ground of Research and analysis: InTech, 213–236.
RB modeling has proven to fill gaps in modeling needs Black, J. L., and M. A. Brzostowski, 1994, Systematic
by developing a 3D spatial prestack convolution ap- of time-migration errors: Geophysics, 59, 1419–1434,
proach. It allows more realistic modeling than 1D con- doi: 10.1190/1.1443699.
volution for geologists and interpreters, assuming that Botter, C., N. Cardozo, S. Hardy, I. Lecomte, and A. Esca-
the users take into account additional input elements, lona, 2014a, From mechanical modeling to seismic im-
such as velocity models, survey geometry, and other aging of faults: A synthetic workflow to study the
acquisition/processing parameters. Being able to pro- impact of faults on seismic: Marine and Petroleum
vide good estimates of PSDM PSFs, in a flexible man- Geology, 57, 187–207, doi: 10.1016/j.marpetgeo.2014
ner, and using them for 3D convolution with detailed .05.013.
target structures, makes RB modeling an improved Botter, C., N. Cardozo, S. Hardy, I. Lecomte, A. Escalona,
method to increase understanding and allow analysis
N. Cooke, and G. Paton, 2014b, From geomechanical
of various effects affecting seismic images. Although
modelling to seismic imaging of 3D faults: 76th Annual
FD modeling can also produce PSFs for such 3D con-
International Conference and Exhibition, EAGE, Ex-
volutions, RB modeling remains faster, more flexible
and interactive; hence, it is more accessible for routine tended Abstracts, A104 09.
work. However, work remains to further improve RB- Botter, C., N. Cardozo, I. Lecomte, A. Rotevatn, and G. Pa-
modeled PSFs, especially toward amplitude calibra- ton, 2015, The effect of fluid flow in relay ramps on seis-
tion and simplify modeling workflows for use in seis- mic images: 4th International Conference on Fault and
mic interpretation. Top Seals, EAGE, Extended Abstracts, Mo FTS 05.
Buckley, S. J., J. A. Howell, H. D. Enge, and T. H. Kurz,
2008, Terrestrial laser scanning in geology: Data acquis-
Acknowledgments ition, processing and accuracy considerations: Journal
The Research Council of Norway (RCN) is acknowl- of the Geological Society, 165, 625–638, doi: 10.1144/
edged for yearly basic funding supporting research ac- 0016-76492007-100.
tivities within seismic modeling and imaging at Buland, A., and H. Omre, 2003, Bayesian linearized AVO
NORSAR, and for project grant #234152 lead by the
inversion: Geophysics, 68, 185–198, doi: 10.1190/1
University of Oslo (Figure 13; “TriasNorth” project
.1543206.
cosponsored by DEA, Edison, Lundin, Statoil, and
Fagin, S. W., 1991, Seismic modeling of geologic structures:
Tullow). We are very grateful to three anonymous re-
Applications to exploration problems, Geophysical De-
viewers and associate editor Carlos Calderón for their
thorough reviews and useful corrections/comments. velopment Series: SEG.
The two first authors thank all NORSAR research Gelius, L.-J., I. Lecomte, and S.-E. Hamran, 2002, The con-
and development colleagues and programmers in cept of local parabolic wave imaging (LpI) in PSDM:
the seismic modeling department. Thanks also go to 72nd Annual International Meeting, SEG, Expanded Ab-
C. Botter for Figure 14, as part of her Ph.D. work stracts, 1184–1187.
on “Seismic imaging of fault zones” (RCN-Petromaks Georgsen, F., O. Kolbjørnsen, and I. Lecomte, 2010, A 3D
project, grant no. 210425/E30). Seismic Unix was used ray-based pulse estimation for seismic inversion of
for some of the cases and figures, especially to gener- PSDM data: 72nd Annual International Conference
ate reference data sets. and Exhibition, EAGE, Extended Abstracts, A031.
Gjøystdal, H., E. Iversen, I. Lecomte, T. Kaschwish, Å.
Drottning, and J. Mispel, 2007, Improved applicability
References of ray tracing in seismic acquisition, imaging, and inter-
Anell, I., A. Braathen, S. Olaussen, and P.T. Osmundsen, pretation: Geophysics, 72, no. 5, SM261–SM271, doi: 10
2013, Evidence of faulting contradicts a quiescent .1190/1.2736515.

Interpretation / November 2015 SAC87

 Gray, S. H., 2001, Seismic imaging: Geophysics, 66, 15–17, imentation in the Triassic deposits of Eastern Svalbard:
doi: 10.1190/1.1444892. Norwegian Petroleum Directorate Bulletin, 11, 61–79.
Hubral, P., 1977, Time-migration: Some ray-theoretical as- Øygarden, B., H. Løseth, and S. Njerve, 2015, Rock proper-
pects: Geophysical Prospecting, 25, 738–745, doi: 10 ties of smectite- and ooze-rich claystones: Geophysics,
.1111/j.1365-2478.1977.tb01200.x. 80, D89–D98, doi: 10.1190/geo2013-0363.1.
Johansen, S., E. Granberg, D. Mellere, B. Arntsen, and T. Pangman, P., 2007, SEAM launched in March: The Leading
Olsen, 2007, Decoupling of seismic reflectors and strati- Edge, 26, 718–720, doi: 10.1190/tle26060718.1.
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

graphic timelines: A modeling study of Tertiary strata Podvin, P., and I. Lecomte, 1991, Finite difference compu-
from Svalbard: Geophysics, 72, no. 5, SM273–SM280, tation of traveltimes in very contrasted velocity models:
doi: 10.1190/1.2759479. A massively parallel approach and its associated tools:
Kaschwich, T., H. Gjøystdal, and I. Lecomte, 2011, Impact Geophysical Journal International, 105, 271–284, doi: 10
of diffraction on resolution of PSDM: 73rd Annual .1111/j.1365-246X.1991.tb03461.x.
International Conference and Exhibition, EAGE, Ex- Robinson, E. A., and S. Treitel, 1978, The fine structure of
tended Abstracts, P384. the normal incidence synthetic seismogram: Geophysi-
Kaschwich, T., and I. Lecomte, 2010, Improved ray-based cal Journal of the Royal Astronomical Society, 53, 289–
seismograms by combining modeling by demigration 309, doi: 10.1111/j.1365-246X.1978.tb03743.x.
with a prestack depth migration simulator: 72nd Annual Rotevatn, A., and H. Fossen, 2011, Simulating the effect of
International Conference and Exhibition, EAGE, Ex- subseismic fault tails and process zones in a siliciclastic
tended Abstracts, C041. reservoir analogue: Implications for aquifer support and
Laurain, R., L.-J. Gelius, V. Vinje, and I. Lecomte, 2004, A trap definition: Marine and Petroleum Geology, 28,
review of 3D illumination studies: Journal of Seismic 1648–1662, doi: 10.1016/j.marpetgeo.2011.07.005.
Exploration, 13, 17–37. Schmalholz, S. M., and D. W. Schmid, 2012, Folding in
Lecomte, I., 2006, Fremgangsmåte for simulering av lokale power-law viscous multi-layers: Philosophical Transac-
prestakk dypmigrerte seismiske bilder: Norway Patent tions of the Royal Society A: Mathematical, Physical,
322089. and Engineering Sciences, 370, 1798–1826, doi: 10
Lecomte, I., 2008a, Resolution and illumination analyses in .1098/rsta.2011.0421.
PSDM: A ray-based approach: The Leading Edge, 27, Schneider, W. A., 1978, Integral formulation for migration
650–663, doi: 10.1190/1.2919584. in two and three dimensions: Geophysics, 43, 49–76,
Lecomte, I., 2008b, Method simulating local prestack depth doi: 10.1190/1.1440828.
migrated seismic images: U.S. Patent 7,376,539. Sen, M., 2014, Seismic, reflectivity method, in H. Gupta,
Lecomte, I., 2013a, Method for simulating local pre- ed., Encyclopedia of solid earth geophysics: Springer,
stack depth migrated seismic images: Canada Patent 1269–1279.
2,521,919. Toxopeus, G., S. Petersen, and K. Wapenaar, 2003,
Lecomte, I., 2013b, Method for simulating local prestack Improving geological modeling and interpretation by si-
depth migrated seismic images: European Patent mulated migrated seismics: 65th Annual International
1611461, validated in France, Germany, The Nether- Conference and Exhibition, EAGE, Extended Abstracts,
lands, and United Kingdom. F34.
Lecomte, I., and L.-J. Gelius, 1998, Have a look at the res-
olution of prestack depth migration for any model, sur-
Isabelle Lecomte received M.S.
vey and wavefields: 68th Annual International Meeting,
(1987), Civ. Eng. (1988), and Ph.D.
SEG, Expanded Abstracts, 1112–1115. (1990; IFREMER grant) degrees, all
Lecomte, I., H. Gjøystdal, and Å. Drottning, 2003, Simu- in geophysics from the University of
lated prestack local imaging: A robust and efficient Strasbourg, France. In 1991–1992,
interpretation tool to control illumination, resolution, she was a postdoctoral fellow at NOR-
and time-lapse properties of reservoirs: 73rd Annual SAR, Norway (grants from EU, 1991;
International Meeting, SEG, Expanded Abstracts, Research Council of Norway, 1992).
1525–1528. She joined NORSAR permanently in
Lecomte, I., and T. Kaschwich, 2008, Closer to real earth in 2003 as senior research geophysicist in R&D seismic mod-
reservoir characterization: A 3D isotropic/anisotropic eling, and she is now principal research geophysicist. In
PSDM simulator: 78th Annual International Meeting, 2003–2012, she was a part-time researcher at the
International Centre for Geohazards (Oslo), coordinating
SEG, Expanded Abstracts, 1570–1574.
geophysics. Since 2012, she has been an adjunct associate
Letki, L. P., J. Tang, and X. Du, 2015, Depth domain inver-
professor at Oslo University, Norway, teaching near-sur-
sion case study in complex subsalt area: 77th Annual face geophysics. She received the 2001 EAGE Loránd Eöt-
International Conference and Exhibition, EAGE, Ex- vös award and the 2014 Norwegian Geophysical award.
tended Abstracts, N104 12. Her main research interests include seismic modeling
Osmundsen, P. T., A. Braathen, R. S. Rød, and I. B. Hynne, and imaging, with applications to seismic reflection, re-
2014, Styles of normal faulting and fault-controlled sed- fraction, tomography, resolution, and simulation of PSDM.

SAC88 Interpretation / November 2015

 Paul Lubrano Lavadera received Daniel W. Schmid received a Ph.D.
M.S. (2009), Civ. Eng. (2009), and (2002) in geology from the ETH
Ph.D. (2013) degrees, all in geo- Zurich, Switzerland. He is working
physics from the University of Stras- as a researcher at the Department
bourg, France. He is a postdoctoral of Geosciences, University of Oslo,
research fellow at NORSAR, Norway and as a consultant and software
(2014–present). His research interests developer for the oil industry. His
include geothermal, microseismics, main areas of expertise are structural
Downloaded 10/18/15 to 12.217.225.70. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

seismic modeling, and imaging. geology, geomechanics, and petro-

leum system modeling.

Ingrid Anell received a Ph.D. (2010)

in geology from the University of
Michael Heeremans received a
Copenhagen, Denmark. She is a post-
Ph.D. (1997) in structural geology
doctoral candidate with the Triassic
and tectonics from the Vrije Universi-
North project at the University of
teit, Amsterdam, Netherlands. He
Oslo. Her research interests include
started as a postdoctoral fellow
interpretation of seismic data, off-
(1997–2001) at the University of Oslo,
shore geometries of sedimentary
Norway, followed by a research posi-
features, and links to onshore deposi-
tion (2002–2008). Since 2008, he has
tional environments.
been working there as a senior engi-
neer, and his main focus has been on research-related
IT support. He has been teaching seismic interpretation
Simon Buckley received B.S. (1999) for more than 10 years.
and Ph.D. (2003) degrees in geomatics
from Newcastle University, UK, and
has had postdoctoral positions at the
University of Newcastle, Australia,
and the University of Bergen, Norway.
He is a senior researcher at Uni Re-
search CIPR, Bergen, Norway. His re-
search interests include developments
of geomatics methods, such as laser scanning and photo-
grammetry, in geoscience applications.

Interpretation / November 2015 SAC89

View publication stats