AutomaticFaultDetectionfor3DSeismicData

DavidGibson
1
,MichaelSpann
1
andJonathanTurner
2

1
SchoolofElectronicEngineering,UniversityofBirmingham,
Edgbaston,Birmingham,B152TT,U.K.
{gibsond,spannm}@eee.bham.ac.uk

2
SchoolofGeographyandEarthSciences,UniversityofBirmingham,
Edgbaston,Birmingham,B152TT,U.K.
J.P.Turner@bham.ac.uk
Abstract. Anovel approachtotheautomaticdetectionof faultsurfaceimages
in3Dseismicdatasetsispresented.Basedonthepremisethatseismicfaulting
introduces discontinuities into the rock layering (that is, the horizons), a
coherencymeasureisusedtodetectpointsofsignificanthorizondiscontinuity.
A highest confidence first (HCF) merging strategy is then combined with a
flexiblesurfacemodeltoestimatethe3Dfaultsurfacesiteratively.
1.Introduction
Largelyfuelledbythepetroleumindustry,majorresourcesaretargetedattheproblem
of imaging underground rock structures. Using acoustic reflection technology, large
3D datasets can be generated showing the changes in acoustic impedance as a
functionofdepth.Mostseismicreflectionsina3Dseismicdatasetaremanifestations
ofrocklayering,orhorizons.Theserepresentthesubhorizontalsedimentarylayering
within petroleum-bearing sedimentary basins. A typical vertical 2D slice through a
3D data set is shown in Fig. 1. (a). In this, the depth below ground is shown
vertically, with the horizontal axis representing a horizontal position on the ground.
The barcode-like pattern clearly shows the sedimentary layers of the rocks. To the
geologically-trained eye the patterns present in these 3D data sets provide useful
insightsintotherockformationandgeologicalstructure,aswellassuchthingsasthe
locationandshapeofgeologicalstructuresinwhichhydrocarbonsaretrapped.
Due to the complexity and size of these data sets it is perhaps not surprising that
image processing techniques have been brought to bear on a number of 3D seismic
datainterpretationproblems. Mostnotableofthesearehorizon-trackers[1].In this
context, horizons are surfaces representing individual sedimentary layers of rock.
Referring to Fig. 1(a), each black and white stripe represents an horizon (noting of
course that the horizon is 3D not 2D). A number of horizon trackers have been
successfully developed, and commercial products are available exploiting this
technology.Researchhasnowmovedtothemoredifficultproblemofseismicfault-
detection.Afaultiscausedbythenear-vertical,relativemovementofadjacentrocks,
resultingintheterminationofhorizons.ThiseffectisshowninFig.1(b).
821
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney

(a) (b)
Fig. 1. (a) 2D slice of 3D seismic data set, (b) Illustration of a seismic fault
(highlighted)

Seismic data sets typically contain a large number of faults at many different spatial
scales. Knowledge of the location of the faults is critical to understanding a
geological system. One effect that faults have, which is of real commercial
significance, is that they act as membranes to the movement of hydrocarbons.
Therefore having a good understanding of the fault positions is critical for the
effectiveplanningofdrilling sitesinorderto maximiseoutputefficiency.However,
despite the significant progress in the development of horizon autotrackers, current
approaches for picking faults are largely manual, and involve laborious hand-
picking of discontinuities on a slice-by-slice basis, one fault at a time. This is time
consuming resulting in hundreds of man-hours of work, performed by trained
geologists. It is estimated that for every six months saved in the work leading up to
the onset of production from a new oilfield, 5% will be saved from the total
production bill. Hence, there is a strong financial imperative for this work. Our
proposalisforanautomaticstrategyforfaultsurfacedetection[2,3].Inpracticethis
proves to be an extremely difficult problem to solve due to noise, imaging artefacts,
andlargenumbersofinteractingfaultsatdifferentspatialscales.
2.ProposedApproach
The proposed approach aims to exploit the idea that the presence of seismic faulting
will result in discontinuities in the horizons. Detection of the consistency of the
horizonsconstitutesthefirststepoftheprocess.Followingthis,smallplanarpatches
representing small parts of the fault surfaces are generated. Finally, these small
planarpatchesaremergedintolargersurfaces,usingahighestconfidencefirst(HCF)
mergingstrategy.Thesestepsaredescribedinmoredetailinthefollowingsections.

822
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney
2.1SemblanceMatching
Firstly we make the assumption that seismic horizons can be modelled as locally
planar,intheabsenceofseismicfaulting.Foragivenpoint(x,y,z)ontheimage,this
assumptionisthentested.Thisisdonebyconsideringaseriesofhslicesofseismic
data,centredaround(x,y,z),orientatedwithanormaln.ThisisillustratedinFig.2.

Fig.2.Illustrationofhslices ofdata,orientated with normaln,centredaroundpoint

(x,y,z)

The data contained in the w by w array of interpolated data for slice r (0rh-1) is
represented by the vector v
r
. A normalised measure of the data variability averaged
overthehslicesisthencomputedusing:
( )
( )

=
=
1
0
2
1
0
2
h
r
r
h
r
r
h
S
v
v
(1)
where the normal n is the estimated localised normal to the seismic horizons
estimatedfromthestructuraltensor[5]asfollows:

=
2
2
2
z z y z x
z y y y x
z x y x x
u u u u u
u u u u u
u u u u u
T


G T T * = (2)
where u
x
, u
y
, u
z
, are the respective seismic image gradients, G

is a Gaussian kernel,
*istheconvolutionoperation,andnistheeigenvectorcorrespondingtothelargest
eigenvalueofT

823
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney
Thenetresultisa semblance mapS which givesa measureofhorizoncontinuity,
with S(x,y,z)=0 representing the worst case fit, and S(x,y,z)=1 indicating the ideal
horizon continuity case. This semblance map, however, is only based on local
constraints,givenbythedimensionsoftheboxusedfortheestimation(h*w*w).As
fault surfaces, by definition, are not isolated phenomena, we can use this fact as an
additioncontextualconstraint.Inpracticethiscanbeachievedbyapplyingadirected
spatial filter to the semblance map S, as a post-processing step. The filter kernel is
directed to concentrate the filtering action along the direction of the fault surfaces,
which in turn is estimated using the same procedure as that used to estimate the
normaltothehorizons(usingequation(2),butthistimeusingthesemblancemapand
not the raw seismic data). The filtered result, S
f
is illustrated alongside the original
seismicsliceinFig.3,andshowsgoodcorrelationbetweentheareasoffaultinesson
theleftimage,withdarkenedareas(lowersemblancematch)ontherightimage.

(a) (b)
Fig. 3. (a) 2D slice of seismic data, (b) Equivalent semblance match (darker areas
indicateincreasinghorizondiscontinuity)
2.2PatchEstimation
Thesemblancedataisthenthresholdedagainstapre-definedthreshold.Thisisdone
on a point-by-point basis, with points having a semblance value below a predefined
threshold being labelled as fault points. This 3D binary map of fault points is then
spatiallysub-sampledtogiveasetofseedpointsasshowninFig.4.

824
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney

Fig.4.Seedpointsoverlaidontoaseismicslice

Given a set of seed points, these can be grouped into small planar patches,
representingsmallsectionsofafaultsurface.Moreformally,forasetofseedpoints
P,aneighbourhoodrelationshipisdefined,sothattwopointsp
1
andp
2
takenfromP
areneighboursiftheyfallwithinaspatialdistanced.
) (
2 1
p p and ) (
1 2
p p if
d p p <
2 1
(3)
whereN(p
x
)representsthesetofpointsneighbouringp
x
.
Aproposedplanarmodelcentredaroundeachpoint,p
x
Pisgeneratedbasedonp
x

andalloftheneighboursofp
x
.AcompatibilityfunctionC(p
x
)givesanindicationof
how well a group of points, centred on p
x
, would be represented by a planar model,
where,
) ( #
) , (
) (
) (
x
p N p
y x
x
p N
p p C
p C
x y

=
(4)
This results in the mean compatibility value between p
x
and all of its neighbours,
where C(p
x,
p
y
) defines the compatibility of a pair of points p
x
and p
y
and can be
calculatedas:
( )( )
y x y x y x
p p C n n r n r n . . . 1 ) , ( =
(5)
where n
x
and n
y
are the estimated normal unit-vectors to the points p
x
and p
y

respectively(ascalculatedearlier),andristheunitpositionvectorbetweenp
x
andp
y
.
C(p
x,
p
y
)tendstowardsunitywhentheestimatedsurfacenormalstothepointsp
x,
and
p
y
lineup,andareperpendiculartothepositionvectorbetweenp
x,
andp
y
.
Starting with the planar model with the corresponding highest confidence value,
theplanarmodelisconstructed,andthepointsremovedfromP.Thisisrepeateduntil
theconfidencevalueofallremainingclustersisbelowapre-specifiedthreshold.
825
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney
2.3SurfaceMerging
Although a planar model can well describe a small patch of a much larger fault
surface, it is inadequate for describing larger segments, or complete faults. To
describe larger fault surfaces we make use of a combined parametric and residual
field [4], the parametric model (in our case planar) models the basic structure of the
faultsurface,withtheresidualfieldmodellingthesurfaceirregularitiesandcurvature.
More formally, a planar model is described by a set of vectors {
x
,
y
,
n
,
o
},
where
x
,
y
and
n
, are mutually perpendicular vectors, with
n
representing the
normal to the plane. A known point on the plane is denoted
o
, and represents the
approximate centre of the subsequent fault surface model. This is illustrated in Fig.
5(a).
In combination with this planar model is a residual field, constructed using a 2D
meshofvectors,normaltotheplaneandofvaryinglengths(seeFig.5(b)).

i
n j i y x o
f j i + + +
,
] [

] [
y x o
j i + +
j

(a) (b)
Fig.5.(a)Illustrationoftheplanarmodel,(b)Planarmodelplusresidualmeshfield

Theresultant surfaceis then formedbyinterpolatingthis mesh,anillustratedin Fig.

6.

Fig. 6. Illustration of the surface formation as an interpolation of the residual field

overlaidontotheplanarmodel

826
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney
Given a set a points, P, a surface model can be generated by firstly computing the
planar model based on a simple least squares fit of the points in P. The distance of
each point p
x
P from the planar surface is then computed, and is denoted
x
. The
residualfieldcanbedescribedasameshofvectorsM,asshownbythearrowsinFig.
5(b).Theendpositionofthe(i,j)
th
vectorinMwillthenbeoftheform,
n j i y x o j i
f j i + + + =
, ,
] [ m
(6)
where f
i,j
represents the length of the vector m
i,j
, [
o
+i
x
+ j
y
] is the start
position, on the plane, of the vector corresponding to the index (i, j), and is a
constant which specifies the residual field mesh spacing. The values of f
i,j
for all
valuesofiandjcanthenbecalculatedasaweightedaverage,

=
P p
x
P p
x x
j i
x
x
p w
p w
f
) (
) (
,

(7)
wherew(p
x
)isaweightingfunctionwhichdiminisheswithdistance,
) / exp( ) (
2 2
=
x
p w (8)
forwhichisthedistancebetweenthepointp
x
andthepoint[
o
+i
x
+j
y
].The
netresultisacompletelydefinedfaultsurfacemadeupofacombinationoftheplanar
andresidualmodels.
Givenasurfacemodeldefinition,amergingstrategybasedontheHCFprincipleis
adopted, in much the same manner as in the last section. Firstly, all of the planar
patches from the previous section are promoted into surfaces. All possible pairs of
surfacesareconsideredformerging,withacompatibilityfunctionfortwosurfacess
1

ands
2
definedas,
( ) ( )
( ) ) & (
) ( ) (
) , (
2 1
2 1
2 1
s s mean
s mean s mean K
s s C

+ +
=
(9)
wheremean(|(s
x
)|)isthemeanresidualfieldmagnitudevalue,(s
1
&s
2
)representsthe
combination of surfaces, and K is a constant which encourages small surfaces to
merge.
The merging of surfaces is continued iteratively until the confidence value falls
belowapredefinedvalue.
3.Results
Natural fault systems are present at a wide range of spatial scales, from faults that
covertheentiredatasetunderconsideration,tofaultsthataresmallerthanthespatial
resolutionofthedata.Ouraimistoestimatethesetoflargefaultsthatmakeupthe
generalstructureofthefaultfield.Inordertoevaluatetheperformanceofthemethod
a manually labelled set of fault surfaces have been identified by a geologist. An
example set or results showing vertical 2D slices of the seismic data are presented
827
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney
providing a comparison between the automatically generated results produced using
theproposedmethod,withmanuallylabelledimages.TheseresultsareshowninFig.
7, withtheautomatically generatedand manually generatedresultsoverlaidin white
and black respectively. It should be noted, of course, that the lines shown on the
figures are the result of slicing the surfaces. In effect, the surfaces can be imagined
comingoutof,andgoingintothepage.

(a) (b)
Fig.7.(a)Estimatedfaults,(b)Equivalent,manuallylabelledfaults

As can be seen from the comparative results there is a broad agreement between
the placement of many of the faults from the automatic detection and manually
labelled methods. A 3D view of the detected surfaces along with the equivalent
manuallylabelledfaultsisshowninFig.8.
Theproposedmethodperformswellatdetectingthelargerfaults,althoughitmisses
some of the smaller, less well-defined faults. Problems can also occur if two faults
are spatially very close together. This can result in over-merging of faults, the net
resultbeingonesurfacepoorlyrepresentingtwofaults.
4.ConclusionsandFutureWork
A method has been presented to tackle the difficult and resource consuming task of
fault detection in 3D seismic datasets. Based on a multi-stage approach, it first
detects points of horizon discontinuity, and progressively groups these points into
larger surfaces. The final surface representation is a combined parametric and
residual field model, which allows for a highly flexible surface representation.
Comparativeresultswithmanuallylabelledfaultsshowpromisingresults.
One of the key questions not addressed in this work is that of combining the
automaticfaultdetectionapproachwithsomeelementofhumaninput,togiveasemi-
automaticfaultdetectiontool.Webelievethatsuchanabilitytoedit,refine,direct,or
imposepriorconstraintsontheproblemcouldprovideconsiderableproductivitygains
over and above a manual approach, whilst allowing the flexibility required for
complexdatasetswithlowsignaltonoiseratioandmultipleinteractingfaulting.
828
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney

(a)

(b)
Fig.8.(a)Estimatedfaults,(b)Equivalent,manuallylabelledfaults

829
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney
References
1. Aurnhammer, M., T nnies, K.D., Mayoral, R.: A Genetic Algorithm for Constrained
Seismic Horizon Correlation. Proceedings of the International Conference on Computer
Vision Pattern Recognition and Image Processing (CVPRIP 2002), pp. 859-862, Durham,
NorthCarolinaUSA,8-14March,2002.
2. Randen,T.,Pedersen,S.I.,Signer,C.,S nneland,L.:ImageProcessingToolsforGeologic
Unconformity Extraction. IEEE SYMPOSIUM, Sem Gjestegrd, Asker, Norge, 9-11,
September,1999.
3. Tingdahl,K.T.,Steen, .,Meldahl,P.,Herald,J.:Semi-automaticdetectionoffaultsin3-D
seismicsignals.SEG/SanAntonio2001.
4. Black, M.J., Jepson, A.: Estimating Optical Flow in Segmented Images using Variable-
OrderParametricModelswithLocalDeformations.IEEETransactionsonPatternAnalysis
andMachineIntelligence,Vol.18,No.10,Oct.1996,pp.972-986.
5. Fehmers,G.,Hocker,F.W.:FastStructuralInterpretationwithStructure-orientatedFiltering.
Geophysics,Vol.68,No.4,August,2003,pp.1286-1293.

830
Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney