Seismic attributes extract information from seismic reflec-tion
data that can be used for quantitative and
qualitativeinterpretation. Attributes are used by geologists,
geophysi-cists,andpetrophysiciststomapfeaturesfrombasintoreservoir
scale. Some attributes, such as seismic
amplitude,envelope,rmsamplitude,spectralmagnitude,acousticimpedance,
elastic impedance, and AVO are directly sensi-tive to changes in
seismic impedance. Other attributes suchas peak-to-trough
thickness, peak frequency, and bandwidthare sensitive to layer
thicknesses. Both classes of attributescan be quantitatively
correlated to well control using mul-tivariate analysis,
geostatistics, or neural networks. Seismicattributes such as
coherence, Sobel filter-based edge detec-tors, amplitude gradients,
dip-azimuth, curvature, and
gray-levelco-occurrencematrixmeasuresaredirectlysensitivetoseismictexturesandmorphology.Geologicmodelsofdepositionandstructuraldeformationcoupledwithseis-micstratigraphyprinciplesandseismicgeomorphologyallow
us to qualitatively predict geologic facies.There has been a
virtualexplosion in seismic attributesin the last several years.
Oil and gas exploration companies,geoscience contractors, and
universities continue not onlyto develop new seismic attributes and
improve workflowsusing well-established attributes, but also to
minimize seis-mic artifacts and calibrate the attribute expression
of geo-logicfeaturesthatwereunrecognizedoroverlookedpreviously.
Three years ago, we were asked to provide a
historicalperspectiveonseismicattributestocelebratethe75thanniversaryoftheSocietyofExplorationGeophysicists(ChopraandMarfurt,2005).ThegoalofourcontributiontothisspecialsectionofTLE
istoupdatereadersontheprogress made since that time. As in 2005, we
will focus
moreonattributesusedinmappingstructureandstratigraphy,leaving
attributes used in lithologic estimation and the directdetection of
hydrocarbons to experts in those fields.Recent advances in seismic
processing. Seismic attributesenhance subtle amplitude and phase
variations in the seis-mic signal that are often quite difficult to
see on the
origi-naldata.Forthesamereason,seismicattributescanexacerbatesubtleamplitudeandphasevariationsinseis-mic
noise. With the exception of AVO and anisotropic veloc-ity
analysis, almost all attribute work is done on data thathave
already been migrated. From the seismic interpreterspoint of view,
there are two types of noise: those the inter-preter can address
through some relatively simple processapplied to the migrated data
volume, and those that requirereprocessing of prestack data. The
interpreter can addressnoise spikes, a limited degree of migration
operator alias-ing, small-velocity errors, and backscattered noise
that canresultinacquisitionfootprint,aswellasoverallrandomnoise
through band pass, kx-ky, and structure-oriented fil-tering. In
contrast, significant velocity errors will result inoverlapping
reflector signals, producing discontinuity andtuning artifacts that
may overwhelm corresponding eventsassociated with the subsurface
geology. Surface and interbedmultiples result in similar strong
artifacts. Our experiencehas been that if reflection events are
highly ambiguous (suchas what often occurs subsalt), attributes
have only limitedvalue. While the interpreter can play a crucial
role in
iden-tifyingprimariesandestimatingvelocitiesthroughinte-grating
well control and geologic models, fixing a prestackdata set
requires sending it back to a processing team.Suppression of
acquisition footprint. Acquisition footprintis defined as any
amplitude or phase anomaly closely
cor-relatedtothesurfaceacquisitiongeometryratherthantothe
subsurface geology. Spatially periodic changes in totalfold,
azimuths, and offsets give rise to spatial periodicity
inenhancement of seismic signal and rejection of seismic
noise.Attributes exacerbate these periodic changes, giving rise
toartifacts. Gulunay (2005) and others have shown that kx-kyfilters
can be very effective in reducing acquisition footprinton time
slices for regularly sampled surveys. Since footprintdue to fold,
offset, and azimuth tends to be organized ver-tically, while that
due to aliased migration artifacts is steeplydipping, kx-ky-w or 3D
running-window Radon filters mayprovide some additional
artifact-suppression leverage. Formore irregular acquisition
design, the noise estimated usingkx-kyor kx-ky-w filters can be
followed by an adaptive filter. For highly irregular footprint
(either due to the irregu-larity of the survey design or to the
irregularity of the near-surface conditions), filters need to be
more spatially adaptive.Cvetkovic et al. (2007) and Jervis (2006)
showed how
wavelettransformscansuppressacquisitionfootprintandthusimprove
attribute images. Wavelet transforms are similar torunning-window
kx-kyfilters but with one major
difference:theyarenonlinear.Theinterpreterdefinesnotonlythewavelength
of the noise to be suppressed, but also a thresh-old amplitude
value. If the amplitude of a given wavelength(the so-called scale)
component exceeds the threshold,
thatcomponentiseliminated;otherwiseitiseitherpassedorweighted with
an amplitude-dependent taper. In this man-ner, only the strongest
(noise, with some underlying signal)events are rejected while
moderate amplitude uncontami-nated signal is retained. Al-Bannagi
et al. (2005) used principal component analy-sis to quantify the
spatial pattern of acquisition footprint inthe shallow part of the
section. In this workflow, the inter-preter needs to define the
zone of interest where
footprintismostclearlydefinedandcomputeacovariancematrix.Nexttheinterpreterdetermineswhichprincipalcompo-nents
or eigenvectors (which spatial patterns) represent foot-print and
which might represent geologic signal. The noisepatterns are then
least-squares fit to and subtracted from eachtime slice. As with
other footprint-suppression techniques,the filter is applied to the
data and attributes are computedfrom the filtered
results.Structure-oriented filtering. Dip-steered mean filters
workwellonprestackdatainwhichdiscontinuitiesappearassmooth
diffractions, but smear faults and stratigraphic edgeson migrated
data. Dip-steered median and alpha-trimmedmean filters work
somewhat better but will still smear faults.
HoeckerandFehmers(2002)addressthisproblemthroughananisotropicdiffusionsmoothingalgorithm.Theanisotropicpartissonamedbecausethesmoothingtakesplaceparalleltothereflector,whilenosmoothingtakes
place perpendicular to the reflector. The diffusion
partofthenameimpliesthatthefilterisappliediteratively,much as an
interpreter would apply iterative smoothing toEmerging and future
trends in seismic attributesSATINDER CHOPRA, Arcis Corporation,
Calgary, CanadaKURT J. MARFURT, University of Oklahoma, Norman,
USA298 THE LEADING EDGE MARCH 2008a time-structure map. Most
important, no smoothing takesplace if a discontinuity is detected,
thereby preserving theappearance of major faults and stratigraphic
edges. Luo etal. (2002) proposed a competing method that uses a
multi-window (Kuwahara) filter to address the same problem.
Bothapproaches use a mean or median filter applied to data val-ues
that fall within a spatial analysis window with a thick-ness of one
sample.Marfurt(2006)describesamultiwindow(Kuwahara)principal
component filter that uses a small volume of datasamples to compute
the waveform that best represents theseismic data in the spatial
analysis window. Seismic proces-sors may be more familiar with the
principal component fil-ter as equivalent to the Kohonen-Loeve (or
simply KL) filter.Figure 1 shows the result of pc filtering on a
seismic dataset from Alberta. Notice not only the overall cleaner
look ofthe section after pc filtering, but also the sharpening of
thevertical faults. The filter was applied iteratively three
timessuch that the end result depends on 49 neighboring
traces.Figure2showsasimilarcomparisonoftimeslicesbeforeand after pc
filtering; improved event focusing and
reducedbackgroundnoiselevelsafterstructure-orientedfilteringare
clearly evident.Figures1and2use99overlappingwindowseachofwhich
consists of nine traces and 11 samples ( 10 ms) par-allel to the
dip/azimuth at the center of each window. Wethen apply our
principal component (pc) filter to the analy-sis point using the
window that contains the most coherentdata. Because it uses (for
our example 11 times) more data,the pc filter in general produces
significantly better resultsthan the corresponding mean and median
filters. We
advisethehopefulreaderthatthereisnosuchthingasasilverbullet in
seismic data processing. If the data are contami-nated by
high-amplitude spikes, then a median, alpha-trimmean, or other
nonlinear filter will provide superior results.Likewise, while the
pc filter will preserve amplitude varia-tions in coherent signal,
it may exacerbate acquisition foot-print amplitude artifacts,
whereas a mean filter will
smooththemout.Structure-orientedfilteringalsoexacerbatesthefault
shadow problem, which should properly be addressedthrough depth
migration.Depth migration and anisotropic migration. With the
excep-tion of complex overthrust terrains, most land seismic
datavolumes acquired in North America are processed
throughisotropic prestack time migration. The cost of prestack
depthmigration and of anisotropic time migration is primarily dueto
the iterations required to develop a more detailed veloc-ity model.
The primary impact of anisotropic migration forTI media is to
increase the fold by including longer offsetsthat are
uncontaminated by ground roll and provide greaterleverage against
multiples. The primary impact of
prestackdepthmigrationistosharpenlateralterminations(bothstructuralandstratigraphic)directlyresultinginsharperattribute
images. A second effect of prestack depth
migra-tionistoremovemuchofthevelocitypull-upandpush-down effects of
the overburden which overprints
curvatureanomaliesofinterestinthedeepersection.Finally,sinceattributepractitionersareoftenstructuralgeologistswhowish
to map fractures, depth migration provides improvedimages of the
basement which often controls shallower fault-ing in the shallower
zone of interest (Aktepe and Marfurt,2007).Spectral decomposition.
Usually seismic interpreters workwith amplitude character that is
based on the dominant fre-quency in the source wavelet. At any
particular location
inthesubsurface,thedominant(orpeak)frequencymaybemodified by thin
layering, which further tunes the
seismicreflections.Notonlytheamplitudebutthephaseofthesourcewaveletisaffectedbythin-bedtuning.Spectraldecomposition
provides a means of examining subsurfacegeologic features at any
discrete frequency that falls
withintheacquisitionbandwidth.Mostcommonly,theseimagesareexaminedaseitherhorizonorstratalsliceswherelat-eral
changes in amplitude and phase can be directly corre-lated to
stratigraphic features of interest.
Byanimatingthroughasuiteoffrequenciesalonganinterpreted horizon, a
skilled interpreter can readily iden-tify where the strata are
thinning and thickening.
Spectraldecompositionisroutinelyusedforthicknessprediction(Partyka
et al., 1999), seismic geomorphology (Marfurt andKirlin, 2001) and
direct hydrocarbon detection (Castagna etal., 2003; Sinha et al.,
2005).Whiletheinitialalgorithmswerebasedonshort-win-dowdiscreteFourierandcontinuouswavelettransforms,Castagna
et al. (2003) and Castagna and Sun (2006)
devel-opedhigher-resolutionalgorithms,firstusingmatching pursuit
decomposition followed later by exponential decom-position
techniques. Others have used the Wigner Ville
dis-tribution(Rauch-DaviesandTalston,2005;Rauch-Daviesand Graham,
2006; Wankui et al., 2006) to predict fluid
type,fluiddistribution,reservoirquality,andreservoirdelin-eation.MARCH
2008 THE LEADING EDGE 299Figure 1. Segments of a seismic section(a)
before and (b) after pcfilteringfrom a 3D volume from Alberta,
Canada. Notice the cleanerbackground and focused amplitudes of the
seismic reflections after pcfiltering as well as the preserved
fault edges. (Data courtesy of ArcisCorporation.)Figure 2. Time
slices at 968 ms through the seismic volume generated(a) before and
(b) after pc filtering of the data in Figure 1. Notice thereduced
background noise and focused edges of the features after
pcfiltering. (Data courtesy of Arcis
Corporation.)Spectraldecompositionhasbeenappliednotonlytotime
domain, but to depth domain data as well. Montoyaet al. (2005)
demonstrate the application of spectral
decom-positioninthedepthdomaintoanareaintheGulfofMexico, which
helped understand the distribution and clas-sification of deepwater
geologic elements.Discontinuities in spectra to map unconformities.
In their
firstpaperoninstantaneousattributes,Taneretal.(1979)rec-ognized
that waveform interference gives rise to disconti-nuities in
instantaneous phase. These discontinuities in turngive rise to
singularities in the time derivative of
instanta-neousphase,orinstantaneousfrequency.Atfirstthey suppressed
these singularities by computing an envelope-weighted average
frequency, which emphasized changes
inphaseneartheenvelopepeakswherewaveforminterfer-ence is minimal.
Later, they enhanced these
discontinuitiesbysubtractingtheenvelope-weightedaveragefrequencyfrom
the instantaneous frequency, giving rise to what theycalled a
thin-bed indicator. Phase unwrapping and Wheeler sections. In the
absence ofwaveform interference associated with unconformities
andfaults, the phase of a seismic trace should always
monoto-nicallyincrease.Stark(2006)hasdevelopedameansofunwrapping
the phase in the presence of such
discontinu-ities,therebyallowinghimtogenerateaseismicWheelerdiagram
one that slices the data along a constant seismicphase, which is
closer to a constant moment in geologic time.Imaging stratigraphic
features along Wheeler surfaces
usingspectralcomponentsallowsonetoestimatetherelativethickness of
channel/levee complexes and also to delineatezones of nondeposition
and erosion (Figure 3).Lipschitz-Hlder exponent measures of
discontinuities. Ratherthan unwrap the phase about discontinuities,
Li and Liner(2008) fit spectral components with a Lipschitz-Hlder
expo-nent, resulting in images of unconformities and
condensedsections that are otherwise subtle but key to accurate
seis-mic stratigraphic analysis. In this workflow, the
interpreteruses commercial software to track unconformities that
fallbetweenseismicreflectionevents,ratherthanpeaksandtroughs of the
events themselves. Like the Wheeler sectionsdescribed above,
seismic data and attributes can then be flat-tened along these
unconformities.Usingspectraldecompositiontodesignband-passfilters.Spectraldecompositionallowsaninterpretertoquicklyidentifythatpartofthespectrumthatbestdelineatesthereservoir,
allowing the construction of simple band-pass fil-ters to optimize
the seismic image for stratigraphic and/orDHI interpretation (Fahmy
et al., 2005). The discovery welldrilledonaWest
Africafieldencountered45mofnetoilpay while a sidetracked well
encountered a downdip wetsand.
Anotherwellwaslaterproposedtoprovethepres-enceofathickerupdipreservoirtoevaluateitsquality.However,
since the seismic amplitude extraction did not
sup-portthepresenceofthickerorbetterqualitysandsupdipfrom the first
well, drilling was put on hold. Spectral
decom-positionusingthematchingpursuitalgorithmwascom-putedonseveralseismicprofilesthatintersectedtheproposed
well location. Analysis of the resultant frequencygathers showed
that the majority of the signals strength forthe reservoir
reflections was confined to a narrow band
offrequenciescenteredaround11Hzandextendingtojustbelow 20 Hz.
Consequently, a band-pass filter was designedand applied to the
seismic data to preserve the signal band-width of the reservoir and
attenuate all energy above 20 Hz.While the resultant data had lower
resolution, a pronouncedlow-impedance anomaly corresponding to the
reservoir
wasclearlyvisible,whichsignificantlyhelpedtheinterpreta-tion of the
stratigraphy. The well in question was later drilledand encountered
154 m of net oil sands in the target reser-voir, validating the
predrill prognosis of better sand faciesthan the first well. Q
estimation and Q compensation. Seismic resolution is thekey to
extraction of stratigraphic detail from seismic data.The broader
the frequency content of the seismic data, thegreater the level of
seismic detail that one can extract usingattributes. The effect of
finite Qis not only to attenuate seis-mic amplitudes but also to
rotate the phase. If the phase ofa given frequency is rotated by
more than 90 from a lower300 THE LEADING EDGE MARCH 2008Figure 3.
Comparison of a conventional stratal slice through the seismic data
volume and a constant phase slice through a Wheeler volume.
Greenareas correspond to nondepositional hiatuses or erosional
unconformities. (After Stark, 2007).reference frequency, no amount
of amplitude compensationwill allow it to constructively contribute
to generating a broad-band reflection. In the processing shop,
poststack or
prestackspectralwhitening(orQ-compensationforseismicampli-tude) is
often applied to enhance the spectral bandwidth ofthe data prior to
interpretation. Recent advances in
comput-ingspectralcompon-entshaveleadtoimprovedestimationofseismicattenuation(1/Q)usingthe
well-established spectral ratiotechnique.WhileQ-compensa-tion is
routinely applied to seis-mic amplitudes,
Q-compensationappliedtoseismicphasesisamore recent development
(Wang,2006). Based on the same model-based physics as earlier
develop-ments by Taner and Treitel
(2003)andChopraetal.(2003)thatexploitedtheadditionalinfor-mationprovidedbyeitherwelllogsorVSPs,thismorerecentdevelopment
is based on the seis-mic data
alone.Spectraldecomposition-basedinversionforseismicreflectivity.Thin-bedspectralinversion(Chopra
et al., 2006) is a
processthatremovesthetime-variantwaveletfromtheseismicdataandextractsthereflectivitytoimage
thicknesses far below
seis-micresolution.Inadditiontoenhancedimagesofthinreser-voirs,
these frequency-enhancedinverse images have proven veryuseful in
mapping subtle
onlapsandofflaps,therebyfacilitatingthemappingofparasquenciesandthedirectionofsedimenttransport.Figure4acomparesaseg-ment
of a 580 Hz seismic sectionfromAlbertaanditsthin-bedreflectivity
inversion (Figure
4b).Noticetheincreaseddetailintermsofextracycles.Figure4cshowstheresultofconvolvingthe
thin-bed reflectivity with a 5-120 Hz band-pass wavelet
result-inginahigh-frequencysectionthat has more information for
theinterpreter. In Figure 5 we showtime slices through (a) the
seismicvolume,(b)thethin-bedreflec-tivity inversion and (c) the
reflec-tivityvolumeconvolvedwitha5-120 Hz band pass wavelet.
Theveryfinedetailonthethin-bedreflectivity time slice is too
highfor us to use conventional inter-pretation techniques. In
contrast,the5-120HzsectionshowninFigure5cissimilartoFigure5abutshowsfeatureswithmuchbetter
clarity. In Figure 6 we show stratal slices through attributes
com-puted from both the original and higher-frequency data.
Notethat the improved frequency resolution does not
significantlychange the curvature. In contrast, the impact on
coherence issignificant, where we note increased lateral resolution
of thechannel system.302 THE LEADING EDGE MARCH 2008Figure 4. (a) A
segment of a band-limitedseismic section extracted from a 3D
volume,(b) thin-bed reflectivity derived from thesection in (a),
and (c) a 5-120 Hz band passwavelet convolved with the reflectivity
in (b).Notice the higher resolution of the section in(c) as
compared with (a). (Data courtesy ofArcis Corporation.)Figure 5.
Time slices from (a) band-limited seismic volume, (b) thin-bed
reflectivity volume derived fromthe seismic volume in (a), and (c)
seismic volume obtained by convolving a 5120 Hz band pass
waveletwith the reflectivity in (b). The very high resolution in
(b) may not yield the expected information aboutgeology, but once a
high band-pass wavelet is put on the reflectivity volume, notice
the extra level ofdetail seen on the time slice in (c). (Data
courtesy of Arcis Corporation.)Volumetric computation of curvature.
Horizon-based cur-vature has been successfully used to predict
faults and
frac-tures,andshowntobecorrelatedwithopenfracturesmeasuredonoutcrops(Lisle,1994)ormeasuredbypro-duction
tests (Hart et al., 2002). Horizon-based curvature islimited not
only by the interpreters ability to pick, but
alsobytheexistenceofhorizonsofinterestattheappropriatelevel in 3D
seismic data volumes. Horizon picking can bechallenging in data
sets contaminated with noise and whererock interfaces do not
exhibit a consistent impedance con-trast amenable to human
interpretation.The trend now isto use volumetric computation of
curvature, which dispelsthe need for consistent horizons in the
zone of interest
(Al-DossaryandMarfurt,2006).Byfirstestimatingthevolu-metricreflectordipandazimuththatrepresentthebestsingle
dip for each sample in the volume, followed by
com-putationofcurvaturefromadjacentmeasuresofdipandazimuth, a full
3D volume of curvature values is produced.Many curvature measures
can be computed, but the most-positive and most-negative curvature
measures are perhapsthe most useful in that they tend to be most
easily relatedto geologic structures. Volumetric curvature
attributes arevaluableinmappingsubtleflexuresandfoldsassociatedwith
fractures in deformed strata. In addition to faults andfractures,
stratigraphic features such as levees and bars
aswellasdiageneticfeaturessuchaskarstcollapseandhydrothermally-altered
dolomites also appear
well-definedoncurvaturedisplays.Channelsappearwhendifferentialcompaction
has taken place.Multispectral curvature computation provides both
longwavelengthandshortwavelengthcurvatureestimates,whichenhancegeologicfeatureshavingdifferentscales.Curvatureimageshavingdifferentwavelengthsprovidedifferent
perspectives of the same geology (Bergbauer et al.,2003). Tight
(short-wavelength) curvature often
delineatesdetailswithinintense,highlylocalizedfracturesystems.Broad(longwavelength)curvatureoftenenhancessubtleflexuresonthescaleof100200tracesthataredifficulttosee
in conventional seismic, but are often correlated to frac-ture
zones that are below seismic resolution, as well as
tocollapsefeaturesanddiageneticalterationsthatresultinbroader
bowls.Figure 7 shows a time-structure map at 1600 ms
inter-pretedfroma3DseismicvolumeacquiredinAlberta,Canada. The
horizon surface was manually picked on a 10 10 grid, autopicked,
and smoothed using a 3 3
meanfilter,whichwasthenusedtogeneratethehorizon-basedcurvature
images shown in Figures 7b and 7c. Notice thatboth displays are
contaminated by acquisition footprint pat-terns (green ellipses).
These types of overprints are
artifactsanddonotmakeanygeologicsense.Horizonspickedonnoisy seismic
data contaminated with acquisition footprint,or when picked through
regions where no consistent imped-ance contrast exists (such as
channels, turbidites, mass
trans-portcomplexesandkarst)canleadtoinferiorcurvaturemeasures. In
spite of 3 3 spatial mean filtering, the
hori-zon-basedcurvatureestimatesstillsufferfromartifacts.Figures 7d
and 7e show volumetric most-positive and
most-negativecurvatureattributesextractedalongthepicked-horizon
surface in Figure 7a. These displays are free of theacquisition
footprint artifacts in 7b and 7c.In Figure 8, we depict strat-cube
displays through
vol-umetricestimatesofcoherence,most-positiveandmost-negative
estimates of curvature. Astrat-cube is a subvolumeof seismic data
or its attributes, either bounded by two hori-zons which may not
necessarily be parallel or covering seis-mic data above/or below a
given horizon. Notice the
claritywithwhichmostoftheNSfaultsstandoutonthecoher-ence display;
however, there is more fault fracture detail onthe most-positive
and most-negative curvature displays. Thelineaments in red seen on
the most-positive curvature displaywill correlate with the upthrown
signatures that one wouldMARCH 2008 THE LEADING EDGE 303Figure 6.
Strat-slices through (a) coherence, (b)
long-wavelengthmost-positive curvature, (c) long-wavelength
most-negative curvature, (d) short-wavelength most-positive
curvature, and (e)short-wavelength most-negative curvature volumes
(left) beforeand (right) after spectral inversion and convolution
with a 5120Hz wavelet.see on the seismic. Similarly, the lineaments
in blue seen
onthemost-negativecurvaturedisplaywillcorrelatewiththedownthrown
signatures.Not only faults and fractures, but stratigraphic
features areoften well-defined on volumetric curvature displays.
Figure9showsstrat-cubedisplaysthroughcoherence(Figure9a),most-positivecurvature(Figure9b)andmost-negativecur-vature
(Figure 9c). Notice the clarity with which the channelfeatures
stand out on the curvature displays. Figure 10
showsstrat-cube-seismic chair displays from (a) seismic
amplitude,(b) coherence, (c) most-positive curvature
(long-wavelength),(d) most-negative curvature (long wavelength),
(e) most-pos-itive curvature (short-wavelength), and (f)
most-negative
cur-vature(shortwavelength).Notice,thetopsurface,whichcorresponds
to a prominent horizon on the seismic, looks
fea-turelessonboththeseismicaswellascoherenceimages.However, the
most-positive and most-negative curvature dis-plays indicate fault
features that can be directly correlated tothe vertical seismic.
The short-wavelength curvature displaysshow the fault features as
crisp and
clear.Attributeanalysisofazimuth-limitedseismicvolumes.Aligned
subsurface vertical faults and fractures as well
asnonuniformambientstresscauseazimuthalvariationsinseismic
properties. 3D land surveys are usually acquired insuch a way that
(unknown) subsurface features are
illumi-natedatasmanydirectionsaspossible,givingrisetoahigh-fold
full-azimuth survey. Conventional processing ofseismic data
typically stacks all azimuths together, therebyobliterating the
azimuthal variation of amplitude and move-out. In order to better
illuminate subtle faults and fractures,Chopra et al. (2000) present
a method which first sorts thedata into azimuthal bins based on the
angle between
sourcesandreceiversontheEarthssurface.Ifthesubsurfaceisisotropic,faultsandfracturesarebestimagedbytheazimuths
perpendicular to them. Figure 11a, a time slice
froma3DOBCseismiccoherencevolumefromWestAfrica,showssomeNESWfaultsthataresomewhatsmeared.Missing
on this display is a NS fault which, based on pro-duction data, was
expected to
appear.Afterapplicationofzero-phasedeconvolutioninpro-cessing, the
data were sorted into azimuth bins of 22.567.5,67.5112.5,
112.5157.5, and 157.5202.5. Subsequently,each volume was processed
independently (through migra-tion,) followed by running the
coherence attribute. The
bot-tomofFigure11showstimeslicescorrespondingtotheall-azimuth time
slice shown in Figure 11a. Notice that dif-ferent azimuth coherence
slices show better alignment notonly NE-SW but in orthogonal
directions. A distinct
crossfault(indicatedbythegreenarrow)isseenontheEWazimuth volume
where it was expected. This
methodologyintuitivelyexploitstheazimuthalvariationoftheP-wave304
THE LEADING EDGE MARCH 2008Figure 7. (a) Time surface from a 3D
seismicdata volume from Alberta, and horizon-based(b) most-positive
and (c) most-negative curvature computed from the picked horizon
in(a). Volume-based (d) most-positive and (e)most-negative
curvature extracted along thepicked horizon shown in (a). Green
ellipsesindicate acquisition footprint not seen on thevolume-based
curvature displays. (Datacourtesy of Arcis Corporation.)306 THE
LEADING EDGE MARCH 2008Figure 8. Strat cubes through (a) coherence,
and long-wavelength (b) most-positive, and (c) most-negative
attribute volumes. (d) Color stack of allthree attributes. While
some NS faults are seen on the coherence display, the level of
detail is much higher on the curvature displays. (Datacourtesy of
Arcis Corporation.)seismic signal to image subtle faults/fractures
perpendicu-lar to their travel path. Anecessary requirement is that
theinput limited-azimuth volumes have sufficient (>12) fold
inordertohaveasufficientsignal-to-noiseratio.Whileimproved
delineation of faults and fractures were
obtainedinthisOBCexamplefromWestAfricaandothersfromAlberta,
azimuthally-limited volumes do not always resultin higher
resolution attribute images.Perez and Marfurt (2007) observe that
if there is signif-icantazimuthalanisotropy,scatteredenergysorted
bysurfacesource-receiverazi muthwoul dsti l l besmeared. They
introduce binsthattakeintoaccounttheazimuth of the travel paths
tothe image point from the sur-facesourceandreceiver.Migration with
this approachimagesdataintoazimuthsthat more accurately
representthedirectionofpropagation.Discriminationintoseparateorientationsintheazimuthdomainresultsinincreasedresolutionofperpendiculargeologic
features.Volumetric attributes suchascoherenceandcurvaturerunonazi
muth-l i mi
tedimagesbasedonthisap-proachresultinimproveddetectionandbetterresolu-tionoffaultsandassociated
fractures.Figure12afrom
theFortWorthBasin,Texas,showstimeslicesat1.36sthrough a
most-negative cur-vaturevolumecomputedfrom an azimuth-limited
par-tialstackusingconventionalMARCH 2008 THE LEADING EDGE 307Figure
9. Strat-slices through (a) coherence,and long wavelength (b)
most-positive and (c)most-negative curvature attribute
volumes.While some channels are seen on the coherencedisplay, their
definition and number is muchhigher on the curvature displays.
(Data courtesy of Arcis Corporation.)Figure 10. Zoom of chair
displays in which the vertical display is an inline from the 3D
seismic volume and the horizontal displays are strat-slicesthrough
(a) the seismic (b) coherence, long-wavelength (c) most-positive
and (d) most-negative curvature, and short wavelength, (e)
most-positiveand (f) most-negative curvature attribute volumes. The
fault lineaments correlate with the upthrown and downthrown
signatures on the seismic.(Data courtesy of Arcis
Corporation.)azimuthbinning.
Figure12bshowsthecorrespondingresultobtainedwiththemodifiedbin-ning.
Arrows indicate subtle faultsandfl
exuresthatarebetterresolvedusingthemodifiedbin-ni
ngmethodthattakesi ntoaccount the azimuth of the
travelpathtotheimagepointduringmigration.Attributevisualizationanddis-play.
Early attempts at 3D visual-i zati onbeganwi thi nl i
ne,crossline,time-andhorizon-sliceanimationsbyhighlytechnicalgeophysicistsusingexpensive,specializedcomputerhardware.Advancesinhardwareandsoft-warehavebroughtvolumeren-dering,geobodytracking,andvisualization,aswellasvirtualrealizationtoeveryinterpretersdesktop.Similaradvancesallowustorendermultipleattributeslicesorsubvolumes(includingtheoriginalseismic)eachwiththeirownuser-definedcolorbar(Figure
13). The interactive use ofcoloradjustment,directionof308 THE
LEADING EDGE MARCH 2008Figure 12. Time slicesat 1.36 s through
most-negative curvaturevolumes computedfrom (a) azimuth-limited
partial stacksusing conventionalazimuthal binning and(b) modified
azimuthalbinning. The bestimaged features strikeapproximately at
rightangles to the corre-sponding azimuthalorientation.
Forinstance, large-curvature zonesapproximately orientedSWNE
(indicated bythe yellow arrows inthe correspondingimage) are best
seen atSENW azimuths.Figure 11. Time slices from the
coherencevolumes run on (top) all-azimuth seismicvolume and
(bottom) restricted azimuthvolumes with the individual ranges
indicatedbelow each image. The NS fault indicatedby the green arrow
in the E-W azimuthimage shows up clearly, but is absent inother
displays. (After Chopra et al., 2000.)lighting, and opacity allows
the interpreter to highlight sub-tle stratigraphic detail that
otherwise could be
missed.Allcommercialworkstationshavetheabilitytoplotasingle seismic
or attribute volume against 256 discrete
col-orsdisplayedassinglegradational,dualgradational,orcyclical1Dcolorbars.Mostworkstationsallowtheinter-preter
to modulate this color barwitha1Dopacitycontrol,thereby enhancing
3D volumet-ric views of voxels whose
valuesfallwithauser-definedrange.Modern workstations also
allowthevisualizationofmultipleattribute vertical, time, horizon,or
stratal slices, as well as sub-volumes,inthesameimage,each with its
own unique
colorbar.Wedemonstratethiscapa-bilityinFigure13,wherewecorenderstratalslicesthroughcoherence,
most-positive curva-ture,andmost-negativecurva-turewi thseveral sei
smi csubvolumes. Such
co-renderingcapabilitiesarekeytovisuallycalibratinggeomorphologicalfeaturesandstructurallinea-ments
seen on stratal and
hori-zonsliceswiththeseismicamplitudesignatureseenonverti cal sl i
cestoascertai nwhether the features are
indeedgeologic,oraseismicacquisi-tionorprocessingartifact.Theinterpretercaninteractivelyrotate
the volume display in anydirectiontobetterunderstandthe data
disposition.Severalworkstationimple-mentations allow the
interpretertoplotdifferentattributesagainst red, green, and blue
pri-marycolors.Guoetal.(2008)find that this color model worksbest
when attributes are of
sim-ilartype,suchasnear-,mid-,andfar-offsetseismicampli-tudes.
Giroldi and Alegria (2005)use the RGB color model to
plotspectralcomponentsof20,25,and 30 Hz to generate the com-posite
image in Figure 14. Asys-tem of channels, clearly seen
onthemultiattributespectraldecomposition image, are diffi-cult to
see on the rms amplitudeimage even though both extrac-tions were
made using the sametimeintervalandcalculatedwith an equivalent
window size. Guo et al. show how the hue-lightness-saturation (HLS)
colormodel can be constructed to dis-play a second or third
attribute.The HLS color bar is displayedalternatively as a
three-axis
dualpyramid,sphere,orcylinderdevelopedbyMunsellearlyinthetwentiethcenturyandshowninFigure15.WhileRijks
and Jauffred (1991) used a 2D hue-lightness color
barover15yearsago,suchcapabilitieswerenotavailableincommercial
software until the introduction of attribute cal-310 THE LEADING
EDGE MARCH 2008Figure 13. Covisualization of seismic subvolumes and
strat cubes from (a) coherence (b) most-positiveand (c)
most-negative curvature volumes. Several channels are seen on the
coherence strat cube, but amore complete system is seen on the
most-negative curvature (which delineates the channel axes,
orthalwegs) and most-positive curvature (which delineates the
channel flanks). (Data courtesy of
ArcisCorporation.)culatorsdiscussedbelow.Figure16showspeakspectralamplitudeagainstlightnessandpeakspectralfrequencyagainst
hue using such an attribute calculator. Carlson andPeloso (2007)
use hue, lightness, saturation, and opacity todisplay four
attributes (Figure 17).Advances have been made in not only the
quantity andvariety of data that can be corendered but in the
flexibilityin which geoscientists can extract the most relevant
infor-mation using an optimized visual perspective (Roth et
al.,2005).Wallet and Marfurt (2008) use similar interpreter-dri-ven
projections in their analysis of spectral components dis-cussed in
a companion paper in this
issue.Textureattributes.Byconstruction,geometricattributesmeasure
lateral variation in reflector amplitude, phase,
andwaveformusingawell-definedandeasilyunderstoodmodel.Thusdipandazimuthmeasurethelateralchangeinphase(traveltime)whilecurvaturemeasureslateralchanges
in dip and azimuth (second derivatives of
travel-time).Spectralandotherdecompositionalgorithms(suchasthosebasedonChebyshevpolynomials)crosscorrelateeach
trace against a suite of precomputed standard
wave-form.Theinterpreterthenexaminesthecrosscorrelationcoefficients
(what we commonly call spectral components)in map view, and
interprets the observed patterns using con-cepts of seismic
geomorphology and thin-bed tuning. Waveform classification. An
alternative to using predefinedsine and cosine waveforms is to
compute waveforms thatbest represent the amplitude variation seen
about the
inter-pretedhorizon.Severalalternativeclusteringalgorithmsare
available, but the current method of choice is that of
self-organizing maps. Originally, the interpreter needs to
guessathowmanyclusterswereneededtorepresentthedata.Clever mapping
of the color spectrum against the orderedsuite of clusters
minimizes errors in making an inaccurateguess (for example, with
too many clusters resulting in
anextraindigoclusterfallingbetweentheblueandvio-letclusters). More
recently, Strecker et al. (2005) plotted
2Dself-organizedmapsagainsta2Dcolorbar,resultinginimages that look
significantly more geological than
earlier1Dclusteringplottedagainsta1Dcolorbar.Matosetal.(2007) show
that by using a large number of prototype clus-ters, the optimum
number of clusters can be estimated
bymeansofaU-matrixandotherstatisticalmeasures.Suchwaveformclassificationhasprovenaverypopularinter-pretation
tool, with the results interpreted in the context
ofseismicgeomorphology,andcorrelatingseismicfaciestoclustersthroughacombinationofdirectwellcontrol,orthrough
pseudo-well analysis using fluid substitution,
andlithologyperturbationfollowedbycarefulgenerationofsynthetics.Waveform
fragmentation. Carlson and Peloso (2007) findstatistical analysis
of seismic waveform lobes to be a
pow-erfulalternativetowaveformclassificationandspectraldecomposition.Waveformlobesaredefinedasfallingbetween
zero crossings. Each lobe is then represented by fourattributes or
measures: peak amplitude, shape
(symmetri-cal,top-loaded,orbottom-loaded)thickness,andstring-length
ratio (a measure of complexity). MARCH 2008 THE LEADING EDGE
311Figure 14. (a) Conventional poststack rms amplitude extraction
at agiven level in the analysis window. (b) A composite spectral
decomposi-tion image with spectral magnitude at 20, 25, and 30 Hz
plottedagainst red, green, and blue color components to form a
color stack.While the same overall picture of the channels emerges,
the spectraldecomposition image is much crisper and richer in
detail. (AfterGiroldi and Alegria, 2005).Figure 15. Munsells color
cylinder. Circumferential axis is hue, orcolor wavelength. Radial
axis is saturation or the degree color differsfrom gray. Vertical
axis is lightness (also called value), or brightness ofcolor.
(After Carlson and Peloso, 2007).Figure 17 shows an example over an
onshore US GulfCoast salt dome. The amplitude distribution
correspondingtotheproductionlevelshowshighnegativeamplitudesover a
large area, which makes it difficult to determine
thelocationofpocketsthataremoreprospectivethenothers.After
fragmentation analysis, the four attributes are
mappedintohue-lightness-saturation-opacityspaceinthevisual-izer,
with each attribute displayed according to the observedrange of
values in the data. Next, all voxels visually havingthe four
attributes within an interpreter-defined cutoff
rangeareextractedandconnectedintogeobodies.Figure17bshowssuchextractedgeobodies,whichmaylieaboveorbelow
the prospective reflector in Figure 17a. This methodyields
geobodies that show more details than those created312 THE LEADING
EDGE MARCH 2008Figure 16. An example of 2D color display using an
attribute calculatorin a commercial workstation. In this example we
combine two attributes into a single output attribute whose values
range between 0 and 255. Each value is associated with a color in a
2D color table.Attributes are peak spectral amplitude and frequency
at peak spectral amplitude. These colors allow the interpreter to
read off the tuning thickness of strong events.Figure 17. (a)
Amplitude map of a productive hydrocarbon zone over a salt dome.
(b) 3D display of extracted geobodies having the
calibratedsignature. Overlain color is the maximum amplitude, with
high-negative amplitudes in red. (After Carlson and Peloso,
2007).with single attributes.Statistical measures of texture. Not
all seismic features
canbeeasilydefinedbyasimpleverticalorlateralchangeinamplitudeandphase,givingrisetostatisticalratherthandeterministic
attributes. The simplest, oldest, and perhapsmost commonly used
statistical attribute is rms amplitudewhich simply computes the
root-mean-square amplitude ofthe trace within a user-defined
window. Generalizing this1D single-trace measure to a small 3D
window of traces
andsamplesgivesrisetowhatarecommonlycalledtextureattributes. The
human brain is very proficient at recogniz-ing textures. Skilled
interpreters correlate textures that
looklikesmilesandfrownstomigrationartifactsandtexturesthat look
like the wings of a seagull to channel-levee
com-plexes.Sometimesseismicfeaturesaredescribedasshin-gled,hummocky,wormy,orevenratty.Amalgamatedchannels,
karst, turbidites, mass transport complexes,
anddiagenetically-altereddolomitesallhavedistincttexturesandcannotbedefinedbysimplymappingpeaksandtroughs.Considerableprogresshasbeenmadein3Dsta-tistical
measures of seismic amplitude and phase, giving riseto what are
called texture
attributes.Mostpresent-daytextureattributesarebasedonthegraylevelco-occurrence(GLCM)matrix.Thecommonlyused
texture measures are energy (a measure of textural uni-formity),
entropy (a measure of disorder or complexity),
con-trast(ameasureoflocalvariation)andhomogeneity(ameasure of
overall smoothness).When applied to seismicdata, high-amplitude
continuous reflections generally asso-ciated with marine shale
deposits have relatively low energy,high contrast and low entropy
(Gao, 2003).
Low-amplitudediscontinuousreflectionsgenerallyassociatedwithmas-sive
sand or turbidite deposits have high energy, low
con-trast,andhighentropy.Low-frequencyhigh-amplitudeanomalies
generally indicative of hydrocarbon accumula-tions generally
exhibit high energy, low contrast, and
lowentropyrelativetononhydrocarbonsediments.Figure18compares three
textural attributes (homogeneity,
contrast,andrandomness)andthreemoreconventionalattributes(amplitude,
instantaneous frequency, and coherence).
Noticethetexturalattributesaresuperiortoconventionalattrib-utes in
terms of resolution, sensitivity, and significance tofacies
variations. West et al. (2002) computed such GLCM-matrix-based
textures to train a neural network to imitatean experienced seismic
interpreter. Seismicmorphologicalimaging. Morphologicattributeslike
size, orientation, geometric complexity (such as linearversus
meandering versus anastomosing), and planar cur-vature (describing
scroll bars in fluvial-deltaic
depositionalenvironments)canquantifydifferentmorphologicalfea-tures.
A
supervisedneuralnetworkcanbeusedtocorre-latethesemorphologicalattributestofaciesclasses.Thesupervisedlearningphasecomprisesconvertingsampledepositional
facies into facies categories labeled by arbitrarynumbers (1, 2,
3,...etc).After satisfactory training and val-idation, the full
seismic volume can be transformed into adepositional facies
volume.Another statistical pattern recognition technology tracesits
origins to biological sciences and was introduced to theindustry by
Durham (2001). This technology does not
requireanyneuralnetworktypeoftrainingphasebutoperatesdirectlyontheseismicdatatoyieldinformationonstra-tigraphyorisolatethoseregionsinthedatathatmaybe
geologicallypromising.Thistechnologycanquicklymineseismic data for
exploration leads.User-definedvolume-attributecalculators.
Severalcom-mercialworkstationsoftwarepackagesnowallowinter-preters
to define/generate their own attributes from seismicor attribute
data volumes and/or from picked surfaces
andhorizonattributeextractions.Whileslowertorunthanacompiled
program, this development allows any interpreterwhose programming
skills are not much beyond that of
fill-ingoutaspreadsheettodevelopinnovativeworkflows.Figure16showsanexampleofdisplayingpeakspectralamplitude
and frequency at the peak spectral amplitude ina single image via
the formula using a 2D hue-lightness colorbar described by Guo and
Marfurt
(2008):amp_freq_output=MIN(ROUND((peak_amp_input_Realized-peak_amp_input_Realized.min)/((peak_amp_input_Realized.max-peak_amp_input_Realized.min)/15)),15)*16+MAX(MIN(ROUND((peak_freq_input_Realized-peak_freq_input_Realized.min)/(peak_freq_input_Realized.max-peak_freq_input_Realized.min)*15),15),0)
where 314 THE LEADING EDGE MARCH 2008Figure 18. Comparison between
textural attributes and conventionalseismic attributes in map view
at a stratigraphic level in the upperMiocene interval. Both
conventional and textural attribute sets arederived from the same
3D seismic data set at the same stratigraphiclevel by using the
same processing parameter (Figure courtesy ofDengliang
Gao.)amp_freq_output isthenameoftheoutputdatavol-ume whose values
range between 0 and 255,peak_amp_input_Realized is the input peak
amplitudevolume stored in brick format (or realized in the
par-ticular software package used),peak_freq_input_Realized is the
input peak frequencyvolume stored in brick
format,thesuffixes.minand.maxdenotetheminimumand maximum values of
the specific volume,ROUND isafunctionthatroundsafloatingpointnumber
to the nearest integer (thereby placing it into aspecific hue or
lightness color bin),MIN and MAX clip the 2D color axes to fall
between0 and 15, and the 2D color bar consists of 16 hues and 16
light-nesses giving a total of 256
colors.Similarformulaecanbegeneratedtoplotthreeattributesagainst a
3D color bar or to apply logic where, for example,coherence is
plotted if the value of coherence is low, whileenvelope is plotted
otherwise, such as described by Chopra(2002) using processing
rather than interpretation
software.Forcompartmentalizedreservoirs,geometricattributes(indicating
faults/fractures) can be combined with oil andgas indicators
(envelope, impedance or AVO anomaly indi-cators).Image processing
for automatic fault detection. Faults
con-trolbothreservoircompartmentalizationandinternalplumbing. Fault
interpretation on 3D seismic data
volumesisverysubjectiveandlabor-intensive.Suchmanualinter-pretation
is most commonly performed on time slices andvertical slices
parallel to the fault strike. Coherence, curva-ture, and other
attributes sensitive to faults can greatly accel-erate this process
by providing images of the fault networkand, implicitly, the fault
hierarchy. However, converting theseimages into computer fault
plane objectsis still carried outby explicitly picking fault
segments on either time or verticalslices which are subsequently
joined together to form a faultsurface. During the past 35 years,
automated attribute-assistedfault object generation algorithms have
been deployed to com-mercial workstations. While 3D generalizations
of 2D Houghtransforms have been successfully demonstrated to
enhancefaults (AlBinHassan and Marfurt, 2003; Jacquemin and
Mallet,2005), such algorithms have not made it into the
interpreta-tion workstation marketplace. Rather, iterative
workflows
thataddressnonfaultfeaturessuchasstratigraphically-aligneddiscontinuitiesandacquisitionfootprintappearthemostpromising
practical
tools.Barnes(2005)startswithacoherenceimageandthenexploits several
well-established image-processing techniques.First, he suppresses
low-coherence geological (such as masstransport complexes and
condensed sections) or geophysical(such as zones where the
background noise overwhelms
theamplitudeofaweakreflector)featuresthatparallelstratig-raphy, and
enhances steeply dipping features (such as faults)through a suite
of simple filters. The result of such filters is
acollectionofmostlydisconnectedsteeplydippingpatches.Next, he
dilates these patches such thatas they grow longer,wider, and
fatter, they connect to each other. Dilation is fol-lowed by image
erosion to generate a surface that is onlyone pixel thick. This
skeletonized, fault-enhanced version ofcoherence can then be either
optically corendered with the orig-inal seismic, or digitized using
an autopicker. Dorn et al. (2005) present another workflow in which
thefirst step is a classical destriping operator applied to time
ordepthslicestoremoveanyremnantacquisitionfootprint.Once the linear
artifacts are suppressed, linear low-coherencefeatures (that we
hope are associated with faults) are exam-ined to see if they can
be linked to form a longer line segment.The result is a relative
probability volume, where each sam-ple represents the relative
probability that it belongs to a
hor-izontallinearfeature.Filterscanbeappliedtorestricttheazimuthrangeorexcludelinearfeatureslessthanauser-defined
length. The line-enhanced volume is next subjectedto a fault
enhancement process whereby linear features
hav-ingthesameazimutharenowlinkedintimeordepth.Vertically-limitedlinearfeaturescorrespondingtochannelboundaries,
pinchouts, and unconformities are filtered out atthis step.The
ant-tracking algorithm (Randen et al., 2002) is an iter-MARCH 2008
THE LEADING EDGE 315Figure 19. The result of ant-tracking applied
to a coherence volume computed over Teapot Dome, Wyoming. The
ant-tracker does a very good jobof delineating the major faults
(magenta arrows) but has no way to know that the low coherence
event indicated by the green arrow is an artifactdue to steep dip
(seen in (a) and (b).) (Figure courtesy of Rocky Mountain Oilfield
Technology
Center.)ativeschemethatprogressivelytriestoconnectadjacentzones of
low coherence that have first been filtered to elim-inate
horizontal features associated with stratigraphy.Theant tracking
algorithm draws an analogy from ants
findingtheshortestdistancebetweentheirnestandtheirfoodsource, and
communicating by making use of chemical
sub-stancescalledpheromones,whichattractotherants. Antsfollowing
the shortest path will reach their destination ear-lier and so
other ants are influenced by the pheromone onthe traversed path.
The shorter path will thus be more andmarked with pheromone.In the
software implementation, artificial electronic antsare distributed
in the seismic coherence attribute volume andallowed to follow
different paths.Ants following a fault willtrace the fault surface
for a specific distance before it fadesout.
Antsdeployedatdifferentpositionsinthecoherencevolumewilltraversethefaultsurfaceandmarkitwithpheromones.
In contrast surfaces that do not represent faultswill be weakly
marked and can be removed by setting a sub-sequent threshold
filter. As these ants traverse different
sur-facesinthecoherencevolume,theyalsoestimatetheorientation of
these surfaces, with both the attribute strengthand orientation
stored at each sample. These two measurespermit construction of
fault objects. Like the two previousmethodologies, filters can be
applied to reject small patches,and objects that have dips and
azimuths that can be attrib-uted to either acquisition footprint or
stratigraphic features. All three of these algorithms (and others
that are com-mercially available but not described in the
scientific
liter-ature)areonlyasgoodastheattributestheyactupon.Figure 19 shows
ant-tracking applied to a coherence attributecomputed from a survey
over Teapot Dome, Wyoming.
Thedelineationofthefaultsindicatedbythemagentaarrowsis excellent.
However, the ant tracking also picksa
coher-enceartifactassociatedwithsteepdip(indicatedbythegreen
arrow). Future advances. Seismic attribute technology will
advancethrough improvements in seismic data acquisition and
pro-cessing, through continued attribute calibration to well
dataandgeologicmodels,andthroughcontinuedalgorithmdevelopment and
interpretation workflows.Processing and acquisition. While
attributes are
routinelyappliedto3Dseismicvolumes,onlyrecentlyhavetheybeen applied
to common-offset and common-azimuth
vol-umes.WepredictacontinuedintegrationofvolumetricattributeandAVOtechnology,withlong-offsetvolumesdelineating
features associated with changes in shear-waveimpedance and/or the
presence of hydrocarbons.Converted-wave volumes have become more
economicfollowing the introduction of high-fidelity
multicomponentpi ezoel ectri cgeophones.
Four-componentshearwaveexplorationismoreeconomicfollowingtheintro-duction
of simultaneous vibrator sweeps on separate sourcecomponents. Thus,
volumetric attributes will be applied
tonotonlyP-wavevolumes,buttoconverted-wave,slow-shear, and
fast-shear volumes. The authors have seen pre-sentations where
horizons on P-wave and S-wave sectionsare correlated through
attribute-delineation of unique
hori-zon-specificstratigraphicfeatures.Whileconverted-andshear-wave
images are often of lower resolution than
cor-respondingcompressional-waveimages,theyaresignifi-cantly more
sensitive to changes in shear impedance
thanmoderate-offsetcompressionalwaves,therebyilluminat-ing
lithologies and hydrocarbons
differently.Asoilfieldsmature,time-lapse(4D)seismicisalsobecoming
more common. Attributes applied to time-lapseimages often show
changes of the production front, partic-ularly if the gas comes out
of solution with decreasing
pres-sure.Inadditiontoshowingproductionoffluids,suchimagescanilluminatepreviouslyunresolveablefaults,allowing
engineers to identify by-passed
pay.Sincetheyenhancesubtlechangesinamplitudeandphase,attributesareparticularlysensitivetoerrors(andenhancements)
in processing. Since their inception, attrib-utes such as coherence
have been used to evaluate
alterna-tiveprocessingflowsandparameterchoices.Attributetime-slice
images are particularly valuable in evaluating
theimpactofprocessingchoicesonsubtlestratigraphicanddiagenetic
features of interest that are difficult to see on ver-tical
seismic. Recent developments in fracture
illuminationbyFomeletal.(2007)bridgetheboundarybetweenpro-cessing
and attribute interpretation.Several processing parameters are one
and the same asseismic attributes. Thus, continuing advances in
measuringseismicanisotropyandseismicattenuationwillresultinattributesthat
we can incorporate directly into our inter-pretation
workflows.Calibration. We predict continued efforts in seismic
geo-morphology beyond the current emphasis on
fluvial-deltaic,deepwater turbidite, and carbonate systems to
include
analy-sisoffracturedbasement,volcanics,anddiageneticallyaltered
facies.Workflows for correlating seismic attributes to map
nat-uralorartificially-inducedfracturesarestillintheearlystages of
their application. While volumetric curvature
attrib-utesonpoststackdataareagreathelptotheinterpreter,development
of attributes sensitive to azimuthally-limitedfracture sets and
their applications will be seen in the nearfuture. Calibration of
such attributes with fracture-sensitivemeasurements including image
logs, tracer data, production(flow) rates, and microseismic
measurements will be seenmore and more. Application of attributes
to
distinguishingbetweenopenandsealedfracturesisinitsinfancy;weexpect
to see more attribute applications in this
area.Statisticalpatternrecognitiontechniqueswillbeusedmoreandmoretoautomatethesiftingoflargequantitiesof
data and observing target features, although the refine-ment of
such techniques could take some time. More neuralnet applications
using texture attributes could aid
charac-terizationoftargetreservoirzones.Wepredictcontinuedadvancement
in computer-assisted 3D seismic stratigraphy,with algorithms able
to delineate seismic stratigraphic
tex-turesincludingonlap,offlap,parallelism,sigmoidalfea-tures, and
hummocky clinoforms in 3D.Algorithm development and workflows.
Crossplots of attrib-utes have been used to identify and interpret
seismic pat-terns associated with target zones, using polygons to
capturethe interesting parts of data clusters. This methodology
isactivelyusedin AVOanalysisbuthasimmensepotentialfor application
with other attributes. After performing
prin-cipalcomponentanalysisonsuitableattributes,attemptshave been
made at crossplotting the first and second
prin-cipalcomponentstodistinguishfavorabletargetzoneswhich separate
as clusters on 2D crossplots. This is one areawe feel will be
explored more in terms of 3D
crossplottingusingthreedifferentattributescharacterizingpatternstoidentify
seismic objects in 3D domain.Just like the concept of textures was
adapted from imageprocessing, another promising concept is the
snakealgo-rithm. Snakes are active contour models that lock or
termi-nate at local edges in an image segment, thereby
localizingthem accurately. Originally introduced by Kass et al.
(1988),this algorithm is used for automated image segmentation.316
THE LEADING EDGE MARCH 2008When the edges of an image segment are
not continuous,low-level image processing may not help and so an
activecontour algorithm (or snake) algorithm is used. By
employ-ingpropertiessuchascontinuityandsmoothnesstothedesired
contour, the active contour also performs an
accu-ratejob.Thesnakealgorithmanditsdifferentimplemen-tationshavebecomeastandardtoolinmedicalimageanalysis.
We expect such algorithms can be generalized toauto-track channels
and other stratigraphic features in 3D.Suggested reading. Imaging
of basement control of shallowdeformation; application to Forth
Worth Basin by Aktepe andMarfurt(SEG2007
ExpandedAbstracts).Acquisitionfootprintsuppression via the
truncated SVD technique: Case studies fromSaudi Arabia by
Al-Bannagi et al. (TLE, 2005). Fault
detectionusingHoughtransformsbyAlBinHassanandMarfurt(SEG2003
ExpandedAbstracts).Multispectralestimatesofreflectorcurvature and
rotation by Al-Dossary and Marfurt (GEOPHYSICS,2006). Improving
curvature analyses of deformed horizons usingscale-dependent
filtering techniques by Bergbauer et al.
(AAPGBulletin,2003).Volume-basedcurvatureanalysisilluminatesfractureorientationsbyBlumentritt(AAPG2006AnnualMeeting).
Multi-attribute visual classification of continuous
andfragmentedseismicdatabyCarlsonandPeloso(SEG2007Expanded
Abstracts). Comparison of spectral decomposition
meth-odsbyCastagnaandSun(FirstBreak,2006).Instantaneousspectralanalysis:Detectionoflow-frequencyshadowsassoci-ated
with hydrocarbons by Castagna et al. (TLE, 2003). High-frequency
restoration of surface seismic data by Chopra et al.(TLE, 2003).
Curvature attribute applications to 3D seismic
databyChopraandMarfurt(TLE,2007).Seismicattributesforprospect
identification and reservoir characterization by Chopraand Marfurt
(SEG CE Course, 2007). Seismic attributesAhis-torical perspective
by Chopra and Marfurt (GEOPHYSICS, 2005).Seismic resolution and
thin-bed reflectivity inversion by Chopraet al. (CSEG Recorder,
2006). Practical aspects of curvature
com-putationsfromseismichorizonsbyChopraetal.
(SEG2006ExpandedAbstracts).Azimuth-basedcoherencefordetectingfaults
and fractures by Chopra et al. (World Oil, 2000). 2D
sta-tionarywavelet-basedacquisitionfootprintsuppressionbyCvetkovic
et al. (SEG 2007 Expanded Abstracts). Automatic faultextraction
(AFE) in 3D seismic data by Dorn et al. (CSEG 2005National
Convention). Technology unravels Genetic Code
of3Ddatatoimprovequality,speedofseismicexplorationbyDuncanandLatkiewicz(AmericanOilandGasReporter,2002).Successfulapplicationofspectraldecompositiontechnologytoward
drilling of a key offshore development well by Fahmyet al. (SEG
2005 Expanded Abstracts). Volume texture
extractionfor3DseismicvisualizationandinterpretationbyGao(GEOPHYSICS,2003).Mappingmultipleattributesto3-and4-component
color modelsAtutorial by Guo et al. (GEOPHYSICS,2008). Using
spectral decomposition to identify and character-ize glacial
valleys and fluvial channels within the carboniferoussection in
Bolivia by Giroldi and Alegria (TLE, 2005). Footprintsuppression
with wavenumber notch filtering for various
acqui-sitiongeometriesbyGulunayetal.(EAGE 2005
ExtendedAbstracts).3Dseismichorizon-basedapproachestofracture-swarm
sweet spot definition in tight-gas reservoirs by Hart etal. (TLE,
2002). Fast structural interpretation with structure-ori-ented
filtering by Hoecker and Fehmers (TLE, 2002). Automaticfault
extraction using the double Hough transform by Jacqueminand Mallet
(SEG 2005 Expanded Abstracts). Edge preserving fil-tering on 3D
seismic data using complex wavelet transforms byJervis(SEG2006
ExpandedAbstracts).Snakes:ActivecontourmodelsbyKassetal.(InternationalJournalofComputerVision,1988).
Wavelet-based detection of singularities in acoustic
imped-ancesfromsurfaceseismicreflectiondatabyLiandLiner(GEOPHYSICS,2008).DetectionofzonesofabnormalstrainsinstructuresusingGaussiancurvatureanalysisbyLisle(AAPGBulletin,
1994). Edge-preserving smoothing and applications
byLuoetal.(TLE,2002).Robustestimatesofreflectordipandazimuth by
Marfurt (GEOPHYSICS, 2006). Narrow-band spectralanalysis and
thin-bed tuning by Marfurt and Kirlin
(GEOPHYSICS,2001).Definitionofdepositionalgeologicalelementsindeep-water
minibasins of the Gulf of Mexico using spectral decompo-sition in
depth domain by Montoya et al. (SEG 2005 ExpandedAbstracts).
Interpretational applications of spectral decomposi-tion in
reservoir characterization by Partyka et al. (TLE, 1999).New
azimuthal binning for improved delineation of faults
andfracturesbyPerezandMarfurt(submittedtoGEOPHYSICS).Automatic
extraction of fault surfaces from
three-dimensionalseismicdatabyRandenetal.(SEG2001
ExpandedAbstracts).Using spectral decomposition and coherence for
reservoir
delin-eationandfluidpredictioninextensivelyexploredregionbyRauch-DaviesandGraham(SEG2006
ExpandedAbstracts).Spectral decompositiontransform methods and
fluid and reser-voir prediction case study by Rauch-Davies and
Ralston (EAGE2005 Extended Abstracts, 2007). Attribute extraction:
An impor-tant application in any 3D seismic interpretation by Rijks
andJauffred (TLE, 1991). Better understanding Wyoming
reservoirsthrough co-visualization and analysis of 3D seismic, VSP,
and engi-neering dataTeapot Dome, Powder River Basin by Roth et
al.(2005RMAG/DGS3DSeismicSymposium).Spectraldecom-position of
seismic data with continuous-wavelet transforms
bySinhaetal.(GEOPHYSICS,2005).Visualizationtechniquesforenhancingstratigraphicinferencesfrom3Dseismicdatavol-umes
by Stark (First Break, 2006). Teaching old attributes newtricks:
Implications of 3D instantaneous phase unwrapping byStark
(Geophysical Society of Houston, SEG Spring Symposium2007). Why
interpret with seismic attributes? Caveats, keynotes,and a case
study featuring multiple seismic attribute analysis
inhydrothermaldolomitebyStreckeretal.(2005SIPES3DSymposium).ComplexseismictraceanalysisbyTaneretal.(GEOPHYSICS,
1979). Arobust method for Q estimation by TanerandTreitel(SEG2003
ExpandedAbstracts).InverseQ-filterforseismicresolutionenhancementbyWang(GEOPHYSICS,2006).Successfulapplicationofspectraldecompositiontechniquetomap
deep gas reservoirs by Wankui et al. (SEG 2006 ExpandedAbstracts).
Interactive seismic facies classification using texturaland neural
networks by West et al. (TLE, 2002). TLEAcknowledgments: We thank
Arcis Corporation for permission to pub-lish the data in figures 1,
2, 4, 5, 7, 8, 9, 10, and 13. Figure 18 is cour-tesy of Dengliang
Gao of Marathon Oil and Figure 19 is courtesy of theRocky Mountain
Oilfield Technology Center.Corresponding author:
[email protected] THE LEADING EDGE MARCH 2008