Click
to article in PDF format.
PSAdvances in Seismic Fault Interpretation Automation*
By
Randolph Pepper 1 and Gaston Bejarano 1
Search and Discovery Article #40170 (2005)
Posted September 7, 2005
*Modified by the authors of their poster presentation at AAPG Annual Convention, June 19-22, 2005
1Schlumberger Stavanger Technology Center, Stavanger, Norway ([email protected]; [email protected])
Abstract
Since the first seismic trace was computer-rendered, automatic interpretation has been the promised panacea of the geo-science community. Twenty years later, we still struggle for a reasonable automatic interpretation methodology in structurally challenging areas.
While automated horizon tracking has become quite elegant, correlating across significant fault displacements remains an obstacle. Algorithms require human intervention to guide the tracking in newly encountered fault blocks. Constraining the horizon tracking to honor pre-existing faults helps, and knowing the fault displacement further enhances this process.
Advances in edge-detection algorithms have allowed direct illumination of faulting and seismically detectable fractures. These techniques improve manual interpretation, but only represent an entry point for automatic extraction of faults.
For some geologic plays, re-sampling of the enhanced edge attribute into a
geologic model property is a simple and effective method of un-biased automated
fault interpretation. Explicit methods to extract fault surfaces
can utilize an
automatically picked horizon indirectly through analysis of “non-picks” and
gradient trends, followed by spatial correlation for vertical connectivity.
Alternatively, using the familiar techniques of seeded auto-picking, on an edge
volume, shows great promise. Flexible editing is essential with these methods.
Finally, we
examine the recent work on fault system interpretation, which provides a
semiautomation of fault interpretation, elevating the interpreter’s task to the
analysis of fault systems. Incorporating new multi-horizon classification or
displacement attributes
allow inference about surface connectivity with fault
throw. The final assembly of these advanced methods as “bread and butter”
interpretation mechanics, while not completely in place, is visible on the
horizon
|
Historical Overview
The automatic tracking of seismic horizons has been widely available in commercial software since the early 1990s providing our first insight into the problem of interpretation automation for geologic faults. What is immediately obvious with a horizon auto-tracker is that the tracking frequently breaks down at fault boundaries. Depending on the tracker, and the parameter settings, we observe gaps in the resulting interpreted surface (non-picked areas) and possibly large time jumps where the auto-tracker picks an erroneous event. Consider the example where the horizon we are tracking encounters a fault that has a displacement equal (in time) to some multiple of our dominate seismic frequency (Figure 1). In this case, our algorithm cannot distinguish an unfortunate alignment of seismic character across the fault without additional information to “recognize” that we have encountered a faulted surface. Using a larger window, encompassing more of the wave train could potentially capture the offset on neighboring events. Or a more sophisticated approach could use simultaneous tracking of multiple horizons, reducing the likelihood for misalignment.
Most automatic horizon tracking applications include cross-correlation or waveform based tracking algorithms to capture the seismic character over a user controlled window length. These methods also compute a “quality factor” attribute associated with the horizon pick position, which give us a further indication on areas of faulting. The combination of interpretation gaps, large gradient trends, and connected regions of low quality factor can produce an excellent visual isolation of the fault geometry, relative to the background horizon structure.
While the fault expression was made visible from the horizon
auto-tracking method alone, as shown in Figure 2,
the means to extract this fault information directly and automatically
was not available. A clever approach to isolate the fault information
from an auto-picked horizon is to take the inverse of the surface, i.e.
show only areas where the interpretation does not exist.
Figure 3 shows an example of the inverse
operation on a surface. The fault boundaries for the structural extent
of the horizon are clearly visible. This technique must be applied to
each surface and then linked from between one surface to the next, if a
complete fault surface is required. Not really an automatic process, but
it does allow an un-bias extraction of faults from a statistically
consistent auto-tracker. Surface operations can be a powerful tool set
for deriving additional information from
An early effort for semi-automatic fault interpretation came from Landmark Graphics Corporation when they introduced FZAP! technology in 1997 (Hutchinson, Simpson et al., US Patent Number 5,537,320). This technique allowed users to begin their fault interpretation task by simply “seeding” one or more fault segments (sticks) on a vertical seismic section, and the automatic operation would perform a cross-correlation on a series of slanted traces derived parallel to the seeded fault segment. The method could be used for both tracking, where no previous fault interpretation existed, or snapping, where an existing fault interpretation would be corrected based on the slant trace cross-correlation algorithm. Each fault surface extracted would need an initial seed point.
A “seedless” approach to fault segment extraction was presented by van Bemmel and Pepper (1999, US Patent Number 5,999,885), where the gaps and sharp gradients from a horizon interpretation are subjected to a connected body analysis followed by feature testing to deduce likely fault candidates. Through the analysis of multiple horizons, the entire fault framework could be extracted.
Seismic signal process advanced rapidly during the 1990s, allowing us to
approach the problem of fault interpretation automation in a similar
vein as we attack horizon interpretation. Bahorich and Farmer (1995)
present The Coherency Cube (US Patent Number 5,563,949), a seismic
attribute for imaging discontinuities. They note that fault
These new edge
The small additional step of executing seeded fault auto-picking on these edge volumes is just entering the mainstream in terms of a commercial software offering. The reason for this technology delay may be in our historical approach of using the seismic interpretation workstation to emulate our “paper” interpretation from yesteryear. We characteristically use the seismic workstations to pick “fault sticks” on vertical seismic sections and then link the intersection of these fault sticks with the interpreted seismic horizon to develop fault traces, in basically the same technique used historically for a paper-based interpretation. Fault contacts are transferred from their position in the vertical seismic section to their spatial position on a basemap for contouring of the seismic horizon. In this sense, the faults are disposable since we are really just interested in fault planes intrusion into the horizon map (surface inverse). Our seismic interpretation workstations simply emulate our manual interpretation process; see Figure 5. We manually draw our fault sticks on the seismic section, establish the fault contact points, and then see them posted on the basemap.
The current generation of geological modelling packages treat fault
Enabling Technologies
Many emerging technologies contribute to our understanding of subsurface faulting and fracturing. We recognize that much progress has been made in the use of the shear-wave component for fracture identification, but that’s a different story. For now, we shall focus on reviewing a collection of enabling technologies, which highlight the advances toward the interpretation automation of seismically resolvable faults and fractures. Our working definition of “seismically resolvable faults are fractures” means those features that express themselves through a spatially coherent measure derived from a typical 3D compressional-wave seismic survey. This measure could mean either a measure of discontinuity or another seismic attribute that allows cognitive identification and isolation of the fault feature.
We hope that by this point you can accept that discontinuity processing
of seismic data, via signal processing of the entire cube, or as a
by-product of horizon auto-tracking, enable us to visually isolate fault
features in the seismic data, particularly in a horizontal format
(either surface slices or time slices). This acceptance opens the door
that interpretation automation may be possible, but issues remain. Can
we improve the quality of our images sufficiently for algorithmic
extraction of the fault features? Our images contain a significant
amount of noise, or acquisition/processing artifacts that reduce their
quality for automated threshold type picking or extraction. A simple
example can demonstrate this point; consider the seismic horizon in
Figure 6a, for a horizon has been
auto-tracked. The tracking algorithm constantly encounters
discontinuities in the data that lead to cycle-skips, or holes in the
interpreted result unrelated to fault breaks in the data. For structural
tracking, we could consider a signal processing step of
Fast volumetric signal processing is becoming a basic element of the
geoscientist’s toolkit, as evident in the barrage of technical papers
and patents related to advanced signal processing on post-stack seismic
volumes. A good example of incorporating signal processing and seismic
interpretation are a pair of papers by Fehmers and Hocker (2002, 2003)
on fast structural interpretation with structure-oriented filtering.
They making a convincing argument that data conditioning before
automatic interpretation produces more complete areal coverage and
improved picking stability. Further, they describe their method to
reduce noise without degradation to the fault expression contained in
the original data. Randen et.al. (2000) demonstrated a collection of
seismic
Marfurt et al. (1999) further develop seismic discontinuity processing
in the presence of local structure using a smoothed local estimate. Chen
et al. (2003) offer an alternative method for imaging discontinuities
using dip-steering. Both are examples of processes, which benefit from a
priori knowledge of the local structure. Sudhakar et al. (2000)
familiarize us with the advantage of incorporating azimuthal variation
into our methodology for detecting faults and fractures. They
demonstrate the superior results obtainable by using restrictive
azimuthal volumes during processing and attribute generation. Most
commercial seismic attribute packages today offer some version of a
seismic dip and seismic azimuth
Many new signal-processing methods are being developed and entering commercial packages, exploiting properties of local curvature (Roberts, 2001), local frequency variability (Partyka et al., 1999), and seismic textures (Randen and Iske, 2005) for example. With this vast array of seismic attribute volumes, classification and neural network analysis are natural solutions for extraction or isolation of seismic objects.
Identification of faults by combining multi-attribute analysis with
neural network classification is another maturing area. Meldahl et.al.
(2001) remark that the trend is shifting from horizon-based towards
volume-based interpretation. We are replacing surface and fault drawing
with seismicobject detection methods, combining fit-for-purpose
attribute processing with pattern recognition technologies. Others
continue to exploit the horizon-based methods, but adopt a more global
approach by simultaneously operating on a collection of derived
A more sophisticated collection of
Borgos, et.al., (2003) take the analysis further by including a fault
displacement estimation by extrapolation of the classification results
onto existing fault
These examples suggest another important element in our quest. The integrated interpretation of faults and horizons, through iterative interpretation or simultaneous interpretation will help us converge on a more accurate structural framework. Tingdahl et al. (2002) offer one example of mapping faults and horizons concurrently, extending the work of Statoil’s seismic object detection technology (Meldahl et al., 2001).
S.I. Pedersen et.al. ( 2002, 2003) introduced a method known as ant-tracking, based on artificial swarm intelligence. This is an exciting method where many thousands of computational “agents” are deployed in a volume to extract a small patch of the discontinuity. The redundancy of agents over the same area reinforces and extends the extracted feature while increasing the confidence in estimate. Figure 9 shows the result of running ant-tracking on an edge volume to create both an enhanced edge volume and to automatically extract fault patches.
Another method offered by Goff et.al. (2003, US Patent Application 20030112704) extracts a fault network skeleton by utilizing a minimum path value and further subdividing a network into individual fault patches wherein the individual patches are the smallest, non-intersecting, nonbifurcating patches that lie on only one geologic fault. This introduction of a patch concept is exciting because it also introduces the idea of patch properties. We now have an additional means of segmenting our fault information. Interpretation Automation
Interpretation automation differs conceptually from automated interpretation. The goal of the first is to provide a tool to improve the quality and turn-around time for interpretation, whereas the latter implies a promise of providing an interpretation without human intervention. While a few corporate executives may like the idea of “click here to find oil”, the geoscientist needs a flexible software toolset which can automate where appropriate, supplemented with manual input when necessary, and most importantly offer a means of extracting the desired information easily.
This desired fault information can be classified in two different forms,
implicit or explicit. An explicit representation means
Looking at the explicit method in more detail, we can summarize an approach to leverage the enabling technologies previously discussed. We would like to move away from a basemap representation of our prospect to a true 3D model representation. One limitation in the past has been the difficulty to performing traditional interpretation, i.e. horizon and fault drawing, in a 3D canvas with the same ease they are currently performed in a 2D canvas. When emulating paper interpretation, a 2D view with polyline drawing functional is appropriate. If the interpretation paradigm changes from manual drawing to surface or volume extraction, the 3D canvas becomes the premier choice. An efficient presentation style for joint horizon/fault interpretation would be to show vertical plane through the seismic amplitude cube and a timeslice view of the discontinuity cube; see Figure 10.
For automatic extraction techniques, the seismic data must be
pre-conditioned either during the extraction process or as a preliminary
processing step. In addition, there may be multiple versions of the
seismic data or derived
For fault extraction, the construction of a discontinuity volume allows the direct detection of seismic faults. We again have the option to further condition the discontinuity data to emphasis large-scale features and/or the subtle detail. Digital processing libraries that offer directional filtering, connectivity filtering, volume segmentation, morphology operations, and multi-volume operations can all be utilized to further visually isolate our features of interest. Post-processing of the discontinuity volume can further isolate the interesting features. Processes such as skeletonizing, pruning, thinning, and erosion (Gonzalez and Woods, 1992) can be powerful filters. Other possibilities are iterative operations, such as running Ant-tracking on the results of Ant-tracking.
While the commercial market has a wonderful inventory of signal
processing methods for seismic volumes, the tools for surface extraction
from seismic volumes has been lacking. Seeded autotracking for faults is
not yet mainstream, but we can anticipate they will soon be widely
available. In addition, more sophisticated approaches for global
extraction of fault
These advances open the door for the geoscientist to work with the derived fault information in more meaningful ways. One of the greatest advantages of the migration from paper interpretation to the workstation was the opportunity to easily access the amplitude information from the seismic. This advantage can now be extended to faults. As previously mentioned, extracted fault patches can be filtered based on their properties (size, quality, orientation, average throw…) but this concept can also be extended to all fault objects regardless of the method used to extract them. Automatic and manual fault interpretation can be managed on a fault system level by filtering on one or more of the derived properties associated with the collection. New properties can be added to estimate fault connectivity, strike length, etc., which will be useful in support of well-based fracture network density analysis. Schlumberger Stavanger Research developed and presented interpretation workflows based on system level interpretation of faults by utilizing these collection of properties associated with extracted fault patches as visual filters, S.I. Pedersen et.al. ( 2002), Borgos, et.al. (2003), and Skov et.al. (2004). Simple histogram and orientation filtering allow the interpreter to reduce an automatically derived collection of fault patches into meaningful fault systems (Figure 12).
The second form of extracting the fault information is an implicit representation, where the seismic is re-sampled into the geologic model as the container for the fault knowledge. A simple example here would be to take the fault expression from discontinuity processing (or further enhancement processing of faults), then re-sample this voxel information into the 3D property grid model (Figure 13). Incorporating implicit fault definitions with seismically constrained layer property population will yield high-resolution geologic models. Obviously, a voxel representation of a fault could be converted to an explicit surface representation through surface modeling options, i.e., gridding. Implicit methods can be made more sophisticated through advanced signal processing and custom workflows. It is not a great leap to appreciate that the seismic displacement field itself would be a valuable seismic attribute.
The 3D displacement field means that at any x,y,z location, we could determine the geologically equivalent position at all other locations in the prospect area. A novel means of constructing an implicit geologic model would be to stochastically populate a model at log resolution, but structurally guide the statistics along coherent orientation and across fault breaks from the displacement estimate. The displacement field would also be a welcome addition to volume restoration studies in support of structural geology interpretation. Dee et.al. (2005), acknowledge fault correlation from seismic as having immediate impact on structural geologic analysis best practices, but their perspective is from primarily manual interpretation methods, and does not include the orientation estimate available from seismic and the automation processes.
An automated means of producing this displacement field would require the combination of two separate elements. We could determine the displacement of a continuous seismic event by computing the local orientation of the horizon. With the dip and azimuth computation at a point, we could predict where the event will on the neighboring traces. But this only will work for continuous events. When we encounter a fault, the orientation estimate will not give us the fault throw, and in fact we will not get a reliable orientation estimate in the vicinity of a fault. Here we must introduce the second element of our automation approach, which is to compute the fault throw via some method of correlation of seismic events across the fault boundary. This step has made the bold assumption that we have a priori knowledge of where these faults are.
Much of this paper has been devoted to documenting the efforts to date in isolating the position of faults and a means of measuring the displacement across faults. See Figure 14. The various tools seem to be available to construct a workflow for creating the displacement field: Determine the location of faults Determine the areas of event continuity Compute the orientation in continuous areas Compute fault throw along fault planes Combine orientation displacement with fault throw displacement to get 3D displacement
Quality control to correct erroneous estimates will be necessary, but could potentially be reduced to manual intervention in a sub-set of the data set, focusing the interpreter’s time and energy on the difficult regions and let automation help us where appropriate.
Besides the attribute workflows, advances in 3D visualization and 3D
interaction capability are going to commoditize volume or geobody
extraction functionality which will include some combination of fault
extraction, horizon extraction, layer extraction, and confined volume
objects such as salt, carbonate build-ups, channels, fracture zones,
etc. These voxel bodies can be directly realized into our 3D geologic
models to freely share across the seismic to simulation activity. For
those that wish to continue with explicit representations, these can be
derived from the voxel presentation either as
ConclusionsWe hope that this paper has yielded some insight into the state of the art for geoscience interpretation automation in general, and also highlight the advances that are going to impact our ability to quickly and accurately interpret fault systems. Our limitation is not the computer hardware or visualization technology at the moment, but a lack of logical integration of the necessary interactive tools to intelligently extract the structural field from the seismic volume. While the technical pieces are all available, the commercial software offerings still lag behind. Many advances have been made and the research continues for both explicit and implicit methods of representing faulted structures. New algorithms for discontinuity estimation and subsequent feature identification are constantly arriving at the patent office and presented at international conferences. Let’s hope the wait is not long for these marvelous tools to reside on our workstation desktops.
ReferencesAbbott, W., 1999, U.S. Patent Number 5,982,707 Method and Apparatus for Determining Geologic Relationships fFor Intersecting Faults. Admasu, F., and Toennies, K., 2004, Automatic method for correlating horizons across faults in 3D seismic data: IEEE Conference on Computer Vision and Pattern Recognition, Washington DC, June 2004. Alberts, P., Warner, M., and Lister, D., 2000, Artificial neural networks for simultaneous multi horizon tracking across discontinuities: 70th Annual Meeting SEG, Houston, 2000. Aurnhammer, M., and Tönnies, K., Image processing algorithm for matching horizons across faults in seismic data: Computer Vision Group, Otto-von-Guericke University (http://isgwww.cs.uni-magdeburg.de/bv/pub/pdf/IAMG_Melanie.pdf) Borgos, H., Skov, T., Randen, T., and Sønneland, L., 2003, Automated geometry extraction from 3D seismic data, in Expanded Abstracts, SEG Annual Meeting. Cheng, Y.C., Fairchild, L.H., Farre, J.A., and May, S.R., 2003, U.S. Patent Number 6,516,274, Method for Imaging Discontinuities in Seismic Data Using Dip-Steering. Dee,S., Freeman, B., Yielding, G., Roberts, A., and Bretan, P., 2005, Best practice in structural geological analysis: First Break, Vol. 23, April 2005. Bahorich, M., and Farmer, S., 1995, 3-D seismic discontinuity for faults and stratigraphic features: The coherency cube: The Leading Edge, Vol. 24.10, October 1995. Bahorich, M., and Farmer, S., U.S. Patent Number 5,563,949, Method of Seismic Signal Processing and Exploration, 1996.
Crawford, M., and Medwedeff, D., 1999, U.S. Patent Number
5,987,388, Automated Extraction Of Fault Goff, D.F., Vincent, L., Deal, K.L., Kowalik, W.S., Bombarde, S., Lee, S., Volz, W.R., and Jones, R.C., 2003, U.S. Patent Application Number 20030112704, Process for Interpreting Faults from a Fault-Enhanced 3-Dimensional Seismic Attribute Volume. Hocker, C., and Fehmers, G., 2002, Fast structural interpretation with structure-oriented Filtering: The Leading Edge, Vol. 21.3, March 2002. Gonzalez, R., and Woods, R., 1992, Digital Image Processing: Addison-Wesley Publishing Company. Hocker, C., and Fehmers, G., 2003, Fast structural interpretation with Structure-oriented Filtering: Geophysics, Vol. 68, No. 4, July-August 2003. Hutchinson, Suzi, 1997, FAZP! 1.0 offers automated fault picking (http://www.lgc.com/resources/MJ_97.pdf). Lees, J.A., “Constructing faults from seed picks by Voxel Tracking: The Leading Edge, Vol. 18.3, March 1999. Meldahl, P., Heggland, R., Bril, B., and de Groot, P., 2001, Identifying faults and gas chimneys using multiattributes and neural networks: The Leading Edge, Vol. 20.5, May 2001. Neff, D.B., Grismore, J.R, and Lucas, A.W., 2000, U.S. Patent Number 6,018,498, Automated Seismic Fault Detection and Picking. Partyka, G., Gridley, J., and Lopez, J., 1999, Interpretational applications of spectral decomposition in reservoir characterization: Leading Edge, Vol. 18.3, March 1999. Pedersen, S.I., Randen, T., Sonneland, L., and Steen, O., 2002, Automatic fault extraction using artificial ants: SEG International Conference. Pedersen, S.I., Skov, T., Hetlelid, A., Fayemendy, P., Randen, T., and Sønneland, L., 2003, New paradigm of fault interpretation: Expanded Abstracts, SEG Annual Meeting. Randen, T., and Iske, A., 2005, Mathematical Methods and Modelling in Hydrocarbon Exploration and Production: Springer Publishing.
Randen, T., Monsen, E., Signer, C., Abrahamsen, A.,
Hansen, J.O., Saeter, T., Schlaf, J., and Sonneland, L., 2000,
Three-dimensional texture
Roberts, A., 2001, Curvature
Simpson, A.L., Howard, R.E., 1996, U.S. Patent Number 5,537,320 , Method and Apparatus for Identifying Fault Curves in Seismic Data. Skov, T., Øygaren, M., Borgos, H., Nickel, M., and Sønneland, L., 2004, Analysis from 3D fault displacement extracted from seismic data, in Extended Abstracts, EAGE, Paris, June 2004. Sudhakar, V., Chopra, S., Larsen, G., Leong, H., 2000, New methodology for detection of faults and fractures: SEG International Meeting. Tingdahl, K.M., Bril, B., and de Groot, P., 2002, Simultaneous mapping of faults and horizons with the help of object probability cubes and dip-steering: SEG International Meeting. Van Bemmel, P., and Pepper, R., 1999, U.S. Patent Number 5,999,885, Method and Apparatus for Automatically Identifying Fault Cuts in Seismic Data Using a Horizon Time Structure. |