Download as pdf or txt
Download as pdf or txt
You are on page 1of 305

Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.

org/
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

GeophysicalDataAnalysis:
UnderstandingInverseProblem
Theoryand Practice

Max A. Meju

Course Notes Series, No. 6


S. N. Domenico, Editor

ietyof ExplorationGeophysicists
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Thesecoursenotesare published
withoutthe normalSEGpeerreviews.
Theyhavenotbeenexaminedfor accuracy and clarity.Questionsor
commentsbythe readershouldbe referreddirectlyto the author.

Meju, Max A., 1959-


Geophysical
dataanalysis
/ Max A. Meju.
p. cm.- (Coursenotesseries;v. 6)
Includesbibliographical
references.
ISBN 1-56080-027-5 :$25.00
1. Geophysics•Observation. 2. Geophysics•Technique.
3. Numericalanalysis.
4. Inverseproblems(Differentialequations)•
Numerical solutions. I. Title. II. Series.
QC802.A1M45 1994
550•dc20

ISBN 0-931830-48-6 (Series)


ISBN 1-56080-027-5 (Volume)

Societyof Exploration
Geophysicists
RO. Box 702740
Tulsa,Oklahoma 74170-2740

¸ 1994 bythe Societyof Exploration


Geophysicists
All rightsreserved.
Thisbookor portionshereofmay
not be reproduced in anyformwithoutpermission
in writingfromthe publisher.

Published 1994

Reprinted1998, 2001
Printed in the United States of America
CONTENTS
1. Intrcxiuction 1
1.1 Basicconceptsanddefinitions 1
1.1.1Examples of problemsin whichinverse
theoryisused 2
1.1.2 Geophysicalprocessesandsystems 2
1.1.3 Geophysicalexploration
philosophyandinversetheo• 3
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1.1.4 Typesof geophysicaldata 3


1.1• Description
andcharacterisation
of geophysical
systems:
matl•matical
models 5
1.1.6 Discre 'Uzation and linearization 6

2. InverseProblems in Geophysics 10
2.1 Meanlagof iaverseproblems 10
2.2 Sampleinverseproblems 13

3. DescribingandFormulating
InverseProblems 14
3.1Typesof solution toinverseproblems:
whatcloweaskof a givendataset? 14
3.'2Classification
of inverseproblems 15
3.3 Discretization
andparameterization 15
3.4 Problem formulation 16
3.4.1 Density.distributionwithintheEarth 16
3.4.2 Boreholetemperatta'e measm'em• 18
3.4.3 Digital fil• designin seismiccleconvolution 19

4. Solving OverdeterminedLinear InverseProblems 2O


4.1 Simplelinear regression 2O
4.2Unconstrained linearleastsquaresinversion:
thegeneralized
mattSx
approac• 26
4.3 Obtainingmatrixinversesandleastsquares solutions 28
4.3.1 Matrix solutions
of a smallsetof equations 28
4.3.1.1 Cramer's rule 28
4.3.2 Matrix solutionof a largesetof equations 33
4.3.2.1Gausselimination method 33
4.3.2.2 Gauss-Jordan method 35
4.3.2.3LU (or Triangular)
decomposition
method 37
4.3.2.4 Generalized matrix inv6•ion 39
4.3.2.5The singularvaluedecomposition
(SVD) of a mirix 43
4.4 Sampleapplicationsof unco•ned inversion 45
4.5 .Problem session 53

5. Conswained
LinearLeastSquares
Inversion 54
5.1 Inversionwith priorixfformation 54
5.1.1 Implementation 54
5.1.2 Formulating
constrainingequations 54
5.1.3 Sampleapplications
of constrained
inversion 56
5.2 Inversion with smoothnessmeasures 61
5.2.1 Problem formulation 62
5.2_2 Problem solution 63
5.2.3 Geometrical
interpretation
of inversionwithsmoothness
measures 64
5.2.4 Sampleapplications
of smoothnessconstraints 66
5.2.4.1Seismicrefractiontime-termanalysis 66

6. Error Analysisin LinearInversion 70


6.1 Elaborate treatment of observational errors in inversion 70
6.2 Assessingthe qualityof a solution 71
6.2.1 Goodness-of-fit 71
6.2.2 Parameter resolution matrix 72
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

6.3 Errors/bounds
on theparameterestimates 76
6.3.1 Parameter covariance matrix 76
6.3.2 Extremeparametersets:extremalinversion 78
6.4Exampleof inversion
anddetailed
erroranalysis:
a recommended
practice 82

7. Solving Non-LinearInverseProblems 84
7.1 Charactet•ationof non-linearity 84
7.2 A generalstrategyfor handlingnon-linearproblems 87
7.2.1 The initial model and linearization 87
7.3 Uncons•d non-linear invexxion 89
7.3.1 Problem formulation 89
7.3.2 Problem solution: the Gauss-Newton method 89
7.3.3 Exampleof iterativeapplication
of Newton'smethod 9O
7.3.4 Problem session 92
7.3.5 Limitations of the Gauss-Newton method 92
7.3.6 The Steepest•escent (or gradient)method 92
7.3.7 Drawbackof the Steepest-descem method 93
7.3.8 Are thereanyremedies forinstabilityandnon-convergence
.9 93
7.4 Constrainedinversion:Ridgeregression or theMarquardt-Levenberg
method 93
7.4.1 Statemereof the minimizationproblem 94
7.4.2 Problem solution 94
7.4.3 Applicationof SVD in non-linearinversion 95
7.4.4 Determinationof dampingfactorsin ridgeregression 96
7.4.5 Ridgeregission in prac•ce 96

8. Non.Linear Biased Estimation 100


8.1Underlyingphilosophy 100
8.2Non-linearinversionwitha priori information 101
8.2.1 Retentionof k•own parametervaluesin thesolution 102
8.2.1.1 Problem formulation 102
8.2.1.2 Problem solution 103
8.2.2 Retention of known form of the solution 105
8.2.3 Consistencyanalysisof non-linearsolutions 106
8.3 Relationships
with standardmethods 107
8.4Implementation of constraints
in iterativeapplications 110
8.5 Problem session:constru•0n of effective biasedestimation scheme 111
8.6Biasedestimation
in practice:
illustrative
examples 112

9. SolutionAppraisalin Non-LinearInversion 122


9.1 Assessing
the qualityof thesolution 122
9.1.1 Goodness-of-fit 122
9.1.2 Parameter resolution matrix 122
9.1.3 Trend analysisof residuals 124
9.1.3.1 The Rnns test 127
9.2 Model bounds 127
9.2.1 Parameter covariance matrix 127
9.3 Extremetrarameter
sets:mostsquares
inversion 130
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

10. SampleSpecializedApplications
of InverseTheory 139
10.1Processingof time-or space-series
data 139
10.1.1 Adaptivesignalprocessing 139
10.1.2 Iterativealeconvolution
of seismograms
fromtipplefiredshots 146
10.1.3 Residual statics estimation 152
10.1.3.1Problemformulationfor linearsurveylines 154
10.1.3.2 Estimation of three-dimensional statics 156
10.1.4 Reductionof lineintersection
errorsin network-type
sun-veys 160
10.2Layered-Earthinversionof reflectionseismograms 165
10.2.1 Iterative inversion of normal incidence surface seismic traces 166
10.2.2 Inversionof verticalseismicpwfde CVSP)data 169
10.2.3 Limitations of normalincidence(zero-offset)inversion 174
10.2.4Plane-layer
inversionof pre-stack
seismo
• forvelocityand.densityprofiles 175
10.3Layered-Earth
inversion
of eleawmagnetic-and
electrical
resistivity
data 183
10.3.1 The inverseproblem 183
10.3.2 Linearizingparameterizafions 184
10.3.3 Forwardproblemsmxlcomputational considerations 184
10.3.3.1The TEM forwardproblem 184
10.3.3.2The MT forwardproblem 190
10.3.3.3The Schlumberger forwardproblem 194
10.3.3.4The Wennetproblem 197
10.3.3.5Approximatepartialderivatives 197
10.3.4 Resistivitymodelconstruction methods 198
10.3.5 Model appraisal 201
10.3.6 Sampleapplications in EM datainterpretation 201
10.3.6.1Jointinversionof TEM andMT soundings 205
10.3.6.2Dealingwith thestaticshiftproblemin MT interpretation 207
10.3.6.3Jointinversionof central-andcoincident-loop TEM soundings 210
10.3.7 Inversionof ticresistivitydata 210
10.4Estimationof effectivedepthof inferenceusingsmoothmodels 213
10.4.1 Aspectsof smoothmodelconstruction 213
10.4.2Geomeu•c interpretationof smooth inversion:
inUfitiveconceptsandimplications 222
10.4.3 Pooledmodelinterpretation: theeffectivectep•of investigation 226
10.5Simultaneous inversionof multi-stationdatawith spatialconstraints 229
10.6Inversionof potentialfield data 232
10.6.1 Interpretation
of aeromagnetic data 233
10.6.1.1 2-D interpretation
of intrabasinalanomalies 233
10.6.2 Interpretation
of gravitydata 236
10.6.2.1Intoration of intrabasin•anomalies usingPedersen's
2-D model 236
10.6.2.2Modellingof 2-D bodiesof arbitraryshapesusingT•waui's method 238
10.6.2.3Implementing a gra• inversionalgorithm 239
10.6.2.4 Aspectsof gravityinverseproblem 241
10.6.3 Jointinversionof magneticandgravitydata 243

III
References 248
AppendixA. SVDINV: A computerprogramforlinearinversionanddetailed
error analysis 253
AppendixB. WENINV: A simpledemonstration
programfor nonlinear
inversion
of Wennersoundings
by ridgeregression 27O
AppendixC. GRAVINV:A simpleprogram fortwo-dimensional
gravityinversion
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

by ridge regression 280

IV
Prefnce

Geophysical dataanalysisisa veryprac6• subject andthisbookisintended tobea


simpleguidetotechniques ofparameter estimationanderroranalysis. I haveplaced
emphasis ontilereconciliation
oftheory andpracticaldatatoenable themater
under•mnd howtotackletypical problems in dataanalysis.
It is hopedthata.mastery of
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

thesimple techniques
described inthisbookwillinspire confidence in the•.aderto
consultthemoreclassical treatmentof thesubje•It ispertinent 'iomention thatthis
bookdeveloped out'ofa highlysuccessfulcourseofthesame titlegiventosenior
undergraduate geophysics students
in Leicester
University since19.88anditspopularity
stemsfromthesuuctm•learning programs andthesimpleeffective practicalapproach
togeophysical
inver•on.
Thecom'se
is•veninthefirstseme.stm'
•oprovide
the
studentswiththebasictoolsforquantitativeanalysisof geoscientlfic
dataandconsists
of 10lectm-esandpractic•!
sessionsinvolvingthematerials givenin Chapm•1 m 9 in
thisbook.Chapter10derives fixxntopicalissuesof researchinterest
to m• andmy
inductioncoursesforpostgraduatestudentsin.geophysics and theindustry.Th•
contentsof thisbookareeasydigestibleandrequ• littlestatistical
ormathematic•
comrni•t.

Max Mcju
Leicester, 1994
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

This page has been intentionally left blank


lo INTl{ODUCTION

1.1 Basic Concurs and


In the g.eophysic•andrela•ed sciences,exp•ents •re usuallyperformedund•
controlledconditions({.e.,in a systematicm•n•er), •ud the outcome
may be :m'•m•'ical
v•lues'tha•
represent
ou•observatio•
at fixed(o•predetermined)
int•rvals,say.These
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

observations
of somepropertiesof the physic• woHda•e commonly•erred to asthe
EXPE•NTAL or OBSEI•VATIONAL DATA. Since such

systematicallycollected,
there must Se somewayof expl•i•i•g or coo•d•ati•g them.
To dr•w •uy i•erenc• from thesedata, we needto underst•udthe relations•p
betweenthe distributionof the propertiesof the physic• systemunde•study(e.g, the
E•-th)•udtheobservable
ß
geophysical
responses.
Thes•stem
ofequations
thatdescribe
this relationshipconstitutethe FOP•WAI• THEOP•Y.I•e•ence of the p•op•r•iesof
the physicalsystemfromoSservational datais the opposite. procedur•andwe needto
apply somequantitativetools collectivelyreferredto •s the •I•E THEOlaY.

Inversetheoryis •u organisedset of mathema•ica••ud statisticaltechniques(calculus,


matrixa•geb•,.st•tisticM
estin•tion•udinference,
e•c)forretrieving
usefi•
information about'a physic• system(or the physic• world) fromcon,rolled
observationson the system.It .is directly concernedwith the auMysisof expe•ental
dat•, the 'fittrig of mathem•ti• modelsto thesedata by estimatingthe unk•ow•
parametersof the models,•ud opt••l experimentaldesign.Inversetheoryis practiced
ß

by every sdentist that a•lyses data relating to the physicalworld which i• the caseof
geoscientlsts
istheEarth.As• m•tteroff•ct, anyone
thathasfitteda lineto a setof
m,mericaldata h•s practisedinversetheory.The level of'applicationof'inversetheory
m•y range•rom the simplestraight-linefitting of seismicref••ion d•ta, say to the
more sophisticated•coustictomographyor multi-d•ruension•l resistivitycurve
m•tchi•g.
Thiscourse
willemphasise
the•uMY•s
ofgeoscientific
data,butit should
be noted that the underlyingd•ta a•lysis procedures
.
a•e the s•me as thoseof mauy
other specialista•easeventhougha variety of termsis usedto describethe subjectin
the d•erent spheresof •pplic•tion (e.•., optimisationtheoryin the mathem•tic•l
sciences,tomography. in the medicalsciences, controltheoryin engin•g, •d
decision
theoryin man.agement
sdences).
Someofthe•reasin whichinverse
theory
finds •pplications are i.tem•sedbelow.
1.1.1 Fo?mpl• of probl•s in whichiz•verseth• is used.
1. Curve fitting.
2. Imageenhancement:
di•t•l filterdesign
anddeconvolution
of[seismograms.
3. Determinationof earthstructure•ud estimationof p•r•metersof ore depositsand
energyresource
acc•,mulations
fromgeophysic•
observations.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

4. Determinatfionof ear•hqu•e loc•tion usingwavearriv• t•mes.


5. Modellingof lithosphere's
response
to loading
orstrainratevariations
in
sedimentary basins.
6. Well (pump)testau•y• in hydrogeology.
7. F•ctor au•lysisin geology.
8. Geochronological
deter•i•atio• usinggeomagnetic
reversals
d•ts.
9. Satellite navigation.
10. Optimal controlof engin•g systems.
11. Medic• tomography.
12. Decisionm•g/Operational research
in management
audmineraleconomics.

However,ourdiscussions
willbel•ted to inverse
theoryandits applications
in
geophysic•
exploration
ofthesubst•ace
properties
oftheEarth.In allthefollowing
discussion•s,
it will be ass•ed that a forwardtheoryexistsfor any parti••r problem.
The typesof õeophysic•
processes
or Ea•h-syst• of pardicluff
interestto us
itemised below.

1.1.2 GeophysicalPx-ocesses
and Sy•'•
Seisrr•cor electromagneticwavepropagation throughtheeach andcurrentor fluid
flowin (porous) rocksareexamples of geophysical
processes.
TheEarthmaybe
describedin terms.of its physic•propertydistributions
whichd•e variousphysic•
systems that maybeinvestigated by'observing say,wavepropagation
in the subsurface.
For easeof discussion,
we will referto the followingfeatures•s constitutingou• geophy-
sicalsystem.•
(notethat thisuskge
may be different
fromtheconvention•approach)-
1. Density distributionwit]• the Earth.
2. Velocity distributionwit]• the Earth.
3. Temperaturedistribution
withinthe Earth.
4. l•esistivity
distribution
wit]• theEarth.
5. Distribution of radioactivemat•• within the Earth.
6. Magneticsusceptibility
vm-iations
wit]• theEarth.
1.1.3 Geophysicalexplorationphilosophyand inversethem7
Thegoalofgeophysical
exploration
istounderstand
orreconstruct
thestructure
of
the Earth from data recordedon, •bove or belowthe •roundsurface.To achievethis
goalwe oftenhaveto studyor makeobservations onthe variousgeophysical processes.
Often, the recordedd•t• m•y conta• additivenoiseor •re incomplete,and •-•,,•cient.
However,we needto õet someth• out.of the observations andsoproceedwith
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

processinE
•hed•t•. Inverse
theory
provides
• formSsin
bywhich
mauyofthe
questions
fund•,•entalto geøphysical
dataprocessing
maybeentertained-
e.g.,the
optim,•,• samp•g rate, how many more data are needed,and dedr• •ccu•acies.
Theoreticalmodelling
techniques
arethenusedastoolsto improveourunderstanS;•g
of the relationshipbetweenthe observeddata (Earth'sresponses
to someexcitations
whichmayBeartificial
ofnatural)
andthevarious
subsu_-4_ac•
physical
prope•;,•
changesor discontinuities
that may havegeneratedthem. TMs searchfor
subsurface
distribution
ofphysical
properties
is aga/naddressed
usinginversetheory.
Questionspertainingto the resolvingpower.ofthe dat•, the typesof modelsthat will
reproduce
theobservations,
and.theeffects
ofobservational
erw• cauallbeattacked
throu• inversetheory.Thus, it cannotbe empha•ed enou• that inversetheory'is of
paramount•,•por•ancein •eophysics,espec/•y wheredecisions
are takenbasedon
exper•,,•ental data.

1.1.4 Types of geophysic• dat•


Examples of data includemass•ud moment of inertia of the Earth, measur••ts of
travel t•,,•esof seismicwaves(½./•.,earthqu•e• andexplosions)•
•r•vity auom•es•
measurements0f apparentresistivitiesof the .$round,well dzaw-downdata etc.
Geophysic•dat• may be recor&of fieldexper•?,•ents
or laboratorymeasurements
(Fi•.
1.1). In/ield measurements,
the n,,•erical valuesobta/neddependon den•ty with/n
the earth,seismicvelodty,b,il• •ud shearmoduli,/roundresistivity,
rockpermeability,
ma/;netic susceptibilityetc-which are the physicalpropertiesthat charmme the
Earth model. In l•boratory exp.er•ments,
we may 8eneratethe responses
ß
of scaled-down
physicalmodelsof the earth whichare usefulezpedallyin situationswherethe
mathematical modelsare very complicatedand •l;•cult to work. There are other forms
of data. For example,previously
obta/n• valuesof somemodalparameters
may alsobe
ß
classified
asdataandsocanourquaurged
e•ectations
oftheformofthemodels
of the
systemunderconsideration.However,for easeof discussion,
weshalltalk mostlyabout
leophysicalfield measurements.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Gathering Data
Field Experiment:
iii

Controlled
Observed
Excitation Unknown '
(e.g. Inductive Earth System Response
(e.g. Resistivitydata)
onorgization)
iii i

Ex-tßreal input Geophysical Target Systematic Records

Laboratory Experiment:

Observed
Systematic Known Physical
Input Scale-model Response
ß (e.g. Seismicm•lel) ß(e.g. Seismic data)

Fig. 1.1 Commonmethodsof gatheringdata


When hza•g experimental data,it shouldalwaysbe bornein mind tha• e•h
observationd• c•u be regardedas a s,,mpledrawnfrom a setof equallylikely eventsor
values.If an experimentis repeatedseveraltimes,the observed valuesunder the
conditionsusuallyvary. Thesevariationscouldbe dueto instnunentalor h,r-,,- error.
The distributionof theses•mplesmay take the form of someprobabilitydistribution.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

For valuesrefexring to the s•mevariable,we couldconstruct a histogramand fit an


expression for theprobabilitydensityfunction.Notethat wehaveemployedthe method
of 'curvefitting' hereandhavethusappliedinversetheoryevenat the dat-•
ch•racterisation stage.Theta•õethereis to deter•-e themeanvalued• -- d•, its
standarddeviation,•r••ud a•y otherstatisticalparameters of interestthat cha•a•terise
the distribution.It is •oodpracticeto estimatethe d•ta errorsor uncertafmties,
ß
•r as
they havea bea_,4m• on the co,•denceto Beascribed to a•Y estimatedpa•,•eters. We
ß

will look •t the effects of observational errors on the i•verse solution late in the
course but it will s-f•ce to note that each • is a randam vaxiable whose mea•

1.1.5 Descriptionor Charaction of Geoph•• Processes:Math•,-,,_*icalModels


Most geophysical processescanbe describedm&them&tically. As mentiond eaxlier,the
set of equationsthat characterise
eachprocessor geophysical systemis kno• as the
forward •heory or simplythe MODEL. It is pertinent to mentionhere that the word
"model" has variousconorationsin the geoscientiicco,•,•,,,,ity. For insta.uce,it may
refer to CONCEPTLIALmodelsaswith manygeologists or to PHYSICAL (laboratory-
scale)•ud .MAT•MATICAL modelsasis cor-,•onin geophysics. In this coursewe
will beconcerned
mostlywithm•them•tical
models
butwillmakereferences
to
conceptualmodelswhere necessary.

A n-rner of geophysical c•u be described


processes mathematic•y by •u integral
equation of the form

wheredi is themeasurable or observable


response of thesystem to •u {ahexternal
input o• excitation(e.g.,explosions
or electricalcaxrentinjectioni•to the •round),p(z).
is a function
related
to some
aspect
oftheEarth's
structttre
orphysical
properties
(•õ.,
resistivity,densityor velocityas a functionof depthin a laterallyhomogeneous
Earth)
referred
to asthe.MODEL
PAR•METE
RS,andK• arecalled
theDATAKERNELS.
The data kernelsdescribe:he relationshipbetweenthe dat• aad'the Earth model
function
p(z).Themodel
parameters
(•.•., v.elodt7,resistivity,
deasitT)
mat be
continuousfancyionsof radiusor position.For =•mple, the travel-time • be•weea a
seismicsottrceaad receiveralon•a ray path L, for a continuous
velocityfield •(z,z), is
given by
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

•=f v(zl,
z)dl. (1.2)
A mathematical
description
ofa physical
system
suchasabove
is referredto asthe
POlq.WAP• THEORY. Forward•heorieshavebeendeveloped
ß
for martygeophysical
processes
a•tda•eroutinely
usedto predictthedataor responsesthat wewouldrecord
overa hypotheticMEaa'th-t• s•ructure.Theseclaraaretherefore
variouslycalled
THEO•TICAL, $YNTHETI0 or PI•DICTED DATA.

1.1.6 Discretizatkm aad r.;,,e•,,4,_•4_'an


In maaycasesofinterestto us,the each modelis a continuous
inactionof depthor
radius. Consider,for example,the massaad momentof 'inertiaof the earth. Both are
related to density within the earth by the formula

Mass
=4•r
Iar2p(r)dr
0

Moment
ofinertia
= ø Itr4p(r)dr

wherea is the eaxth'sradiusaadt•(r) corresponds


to p(z) in eq. (1.1), aad is the density
at radialdistance
r. Equations
(1.3a& 1'.3b)maybecombined
•o givethegeneral
expression

=Ia
o

whichis the sameaseq. (1.1).This i.utegralis easilyevaluatedusing-digitalcomputers.


This involvesthe useof discrete
mathematics.To appro•rnatethis integralon a
computer,
weusea sa•mmation
withp(•)drsettorna•dKi to•i ' say,•viag usthe
very use• computational
formula
ß
In this situation, we saythat the theoreticalproblemis DISCRETIZED. It is common
practicein •eophysicsto workwith discretem•T•bers. For tec-h•c• reasonsou• field or
exper/men•alobservations are recordedoverfinite intervals(e.õ.,discrete•requenciesor
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

fixedbandwidth)
instead
ofallobservations
intheran•e[0,co]required
to-•quely
ch•racterisethe e•rth-$ystem.Also,for computational simplicity,we often describe
an
otherwisecontinuousdistributionof the Earth'sphysic• properties p(z) by a •ite set
of parameters,e.g., a layeredstructttrewith ea• layer having• specificdensityand
thickess. This practiceis referredto as PARA_METERIZATION. For conveneace,
in thiscourse,
wewillbeConsidedx•g
onlydiscrete
models
aaddiscrete
pax•r•etersß

whichare e•sier to handlethan the continuousdistributions.We will thereforestudy


DISCRETE INVERSE THEORY instea• of CONTINUOUS INVERSE THEORY.

Let us define someother usef• term-•in inverseproblemtheory. In discreteform,


eq. (1.2) may be writtenas

= X:W
which is the formula usedin practic• applications.Notice that the travel-time is not
directly proportional to the model pax•meter • but to its inverse.The relationshipis
saidto be non-lineaxi• •. How.eve•, if we definethe modelpax-r•eter as c -- 1/•, where
c is the slownessof the seismicwave, then the problem caa be stated as

which is of the form & = •/m and the relatio•hip is now saidto be 1•,,_•,,.Sucha
trausformationoperationmay be referred to as I.;•,•.•.ri.i•g p•,•4_'_e•_
Now considerthe problemof catculatiaõthe apparentresistivitythat will be observed
overa two-layerEarth modelusin• the Schb,
mber•erelectrodeconfiguration
{Fi•. 1.2).
This is givenby (seePaxasnis,
1986)

whereL=AB/2 is the distax•ce to eachcurrentelectrodefrom the centralpoiat, J1 is


the Besselfunctionof order1, andK(A) is a function.ofthe paxameters
(layer
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 1.2
C1

2/-//////•///////////
ß
p• P•
••Z)•
surface

Schlumberger electrode configuration


Basement
Top layer
C2 Ground
resistivities
Pt,P=and toplayerthickness
t) of our systemandthe integrationvariable
•. K(•) is given by

=
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(P•--P•)
where
It easyto seethat we c,.nnotput eq. (1.8) in any simpleformzesembl•,•gE =Gm as
wed/dfor eq. (1.2). The zesis[•vity
depthsoundinõ problemiu saidto be highlynon-
linea•.Theusual'
method
ofde•g withsuch
probl_•n.•
involves
thederivatio•
of'Rue=
au•oguesusingTaylor'stheorem,a procedure
termed
2. INVEHSE PROBLEMS IN GEOPHYSICS
2.1 Me•i•g of Inverse Probl•
ß

• order •o •y •derst•d the me•g d the te• '•v•e proSly', it is •t••ive


to co•ider, first, theopposite
situation
dub• the'fo••d probl•'. Tr•tion•y,
.the•te•retation ofsome•physicd d•ta (e.•.,r•isti•ty d•th so••• dat•)
•volves comp•son •th th•reticd M•t• •• (or nomo••). ••e •• •e
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

computed
••t ce•• rel•tio•Mps
(math••ti• mode)
•,•t • p•i••
•tdbuti0n of the physi• prope•i• of the subside, i.a, the •• •e the
th•reti• re•o•es for ide•s• E•h-t•e st•••. Thepwced• is s•ple -

'Givensomei•orm•tion onthevaluesofthesetofparamete•(e.õ,nmnt•e•
of'layers,
thei• resisfi•fi• and thi•-l•esses)fo• a hypothetical
Ear•-modeh a them•cal
•o•/p .(mathematicalmodel)is usedto ded•e• val• o/some•le
CliP-fit/es(e.g., apparentresist/•es andphases)'.

This procedure constitutesthe FORWARD APPROACH a•d solveso• fo•vaxd


problem(seeFig 2.1). Note that what is generally
knownto ma•y geophysicists
as
•.Forward
modelling
b_.v
intera.cti.ve
computing'
isnøthi•g
butamore
vexsatile
extension
of the originalcurve-matchi-õtechnique.In interactiveforwardmode]];,•õ
the
theoreticalcurvesgeneratedfor aa input modelare displayedtogetherwith the field
curves, say, on a interactive term•_al. The model parametersare adjusted •ud the
operationsrepeated until a• acceptablevisualfit is oBt•ed betweenthe field and
theoretical curves.

In the inverseapproach, the Earth'sstructttre (or otherusdulin/ormation)is directly


retrievedfrom the.field dat• (Fig. 2.2). The inverseprocedureis described
as:
'Given someinf••• on the v•luesof somemea.m=• quautities(field or
experi,•ental data), we usea theoretical•tionship to d• the valuesof the set of
parametersthat explainsor reproducesour field observatioaa'.

We canillustratethe difference
betweenthe/orwsxdsad inverseapproach
usinga very
simpleex;r-ple. Considerthe temperatuxedistributionwithi,• the Earth and ass•,r,,e
that temperatuxeincreaseslinearlywith depthin the Earthsothat the relationship
canbe e.'vpressed
as.(e.g.,Menke,1984)

(2.1)

lO
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Forward Problem
i i ß

Given- Estimates or values of

• the
model
p•[rameters
.
Determine- Theoretical responses (data).
i i i

The Forward Process


i i

Numerical
Model Representation Computed
Parameters of System Dynamics Responses
(Fo•vaml Thee•y)
i

Input Operators Output

Figure 2.1
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Inverse Problem

Given: Field observations


(Earth system responses)

Determine: Psrameters of the e•rth-model

The Inversion Process


i ii !

Input Operatom Output

Figure 2.2
where• •ud b •re n•;r•ericalconstants.Giventhe v•luesof a aud b (-1.5 •ud 0.04
respectively,
say), we c•u c•lculatewhat the temperature T will be for •uy givendepth
valuez. In c•lculatingT at .variouslevels,we axesolvingthe forwardproblem.If, .on
the other h•ud, we measured• n,,•nberof temperaturesat variousdepthsdown a
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

borehole•ud wishto determ•i,•e a andb fromourobservations, thenwe axedea]iugwith


•u inverseproblem.The equation(2.1) relatingT amdz describes a straightline amd
we axethus facedwith the inverseproblemof fitting a straightline to the borehole
data.

2.2. SampleInverseProbl ß

Thereis a variety of inverseprobl_•-•-in•eophysics


•ud includethe deter•i,•afion of
earthstructure
fromobservational
data,deconvolution
ofseismograms,
earthquake
locationusing•riv• timesof waves,determi•.tionof trendsin timeseries
the deter•,•ation of optfin,,,,,s•,•pling rates•ud othersurveyparametersin data
ß

•cquisitionproblems,estimation of theresolvingP0.werof observational


data,.t•,•e
reversalsof the earth'sm•gnetic field(aninverse problem in geochronology,
s•r•c•o)
anddetermination
ofsubsudace
temperature
distribution
from'borehole
measurements.
3. DESCRIBING AND FOltMULA•G INVERSE PROBLEMS
Key Que•ons
A detaileddescription
or formulationof an inverse
problemwill addressthe following
nine goldenquestions'
1. What is .theapplicablepa•meterization-discrete
or continuous
?
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2. What is the natureof the geophysical


data;whatsrethe errorsin the observations
?
3. Ca.uwepose
theproblem
mathematic•y
andif sohowwellposed
•sit ?
4. Are there any physicalconstra•tson the problem?
5. What typesof solutionto the problemaredesirable;andto what accuracy ?
Are we lookingfor an approximate solution;boundsof the solutionor the exact
solution ?

6. Is 'the problemlinea•or non-linear?


7. Is it an overdetermined,underdetermined
or even-determined
problem?
8. What is thebestmethodof solvinõ theproblem7.
9. what •re the co,•dence]•ts of the solution
?
C•n it be appraised
by otherme•,•-•7
All the above issueswill be addressedin this course.

3.1 Typesof solution


t,oinverseprobl•,•-•' Whatdowe askof a given•lataset?
Thereis a plethoraof questions
that c•u beaddressed
usinginversetheory.Thismeans
that depending ontheproblemin hand:wemayhavea varietyof solution types.If we
•re •ndyzing geophysicaltime-series
for instance,
wemaybe interestedin deter•v,i•_g
the opti•,• s•r•plingratesor suppressing unwanted
signalor simplycalculating
removing
a trend
fromtime-sequential
events.
Tl•edesired
solution
inthiscase
isthe
best estimated d•ta from the processingsequence.
Wheninterpreting.dat•
forsubsurface
gtructure,
ihedesirable
solution
to ourproblem
will be the bestestimateof the subsur/ace
physicalpropertydistributionssuchas layer
resistivities
(orthereciprocal,
conductivities)
andthic•esses.
However,
owing
to the
fact that someof the solutionsare •herently non-,,,•que,it may be better to seeksome
combined propertyof thesubsurface
(e.g.,conductivity-thickness
product) which
are better resolved
by the'geophysical
datain useratherthan individualmodel
pa•r•eters.We mayalsoprefertheslowness(I/v) to thevelocity(v) of acoustic
waves
in the subsurfaceas ou• desiredsolutionbecauseof the •ttendant computational
advantages
offeredby suchp•r•r•eterizations.
In cert•,• situations,
it maybe of
interest to seekthe non-uniqueness
boundsdeemedby our data or • suite of .extreme
modelsthat definea pazticulazaspectof the modelor eventhe modelsp•cera•hera
ß

sLuõle
model
forthesubsurface.
Wecaz•alsoseek
'exact'
rathertha•'•ppro•rnate:
solutions
to problems.Thepazameterization involved(i.•., thechoice
of-tazge•
parmeters)will ofteninfluence
.theway •u inverseproblemis poseda•d the solution
proceduze
aswe sh•11seelater in •he course.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

.3.2 ClassiCcationof inverseprobl,•r-•


To a largee, ent,the•swersto someof the•bovequestions
or auyotherdemands
on a fi•ite set of observationaldat• will dependon the relative dim•io• of the
observationaldata and the modelpaz•neters. H the soughtmodelconsists
ß
of fewer
parameters.
th•u the n,•ber of field d•ta• thenthe inverseproblemis saidto be
OVERDETEKM•ED •ud c•u be formallysolvedusinõmethodsba•edon achieving
-.

• best fit to the dat•. If on the other hand the soughtmodelha• more pazamete:sth•u
thereaze'observ•ti0n•l d•ta, the inverseproblemis U1TDEKDETEI•W•NED.J.utkis
l•tt er case,theree.xists•u i•nity of modelsthat cansatisfythe sparsedata a•d sowe
ß

must find a meansOfsinglingout onepazti• modelout of •11possiblecandidates.


Thistypeofproblem
isbest•tt•cked
byconstructin
Smodels
whose
parameters
are
continuousfunctionof position.It will be shownlater that it is alsopossibleto usethe
methodsoriginallydevisedfor overdetermined proble_•-•to obtah•meaningful'smooth'
model• from underdeterm/nedcases. When there are very sm•11amountsof data a•d
we seeka comparablen•,•nberof modelpaz•eters from them, the problemis saidto
be EVENLY DET]••D •ud in this case,very s•rnplemodelsc,m be constracted
usingDIRECT IN'VERSIONschemes. The problemwith this type of problemis that
typicalfield dat• azerarelyindependentsuchthat the problemis in
underdeter•;,•ed.

3.3 Discretizatioa az•dpar•etetizatio•


For physical(or sometimes physico-chern•cal)res.sons,
we may expectthe dis:ribution
of somephysicalpropertyof the subsuzf•ceto be a continuous functionof depth.
Geophysical me•suremen:s azeusuallymaplein orderto determinethe subsurface
propertiesor structure.The distributiox• of the physica•properties
c,mbe •quely
determined if the measurements spanthe obse_•ational
ß
ba•d-•dth [0,oo]. However,
this is not possible owing:o :ecl•c•l ];•t•tions a•d we typic•y conductourfield
e.'rpe•en:s overa finite observationinterval: •ud the outcomesazediscrete::•erical
v•luesc•ed field da•'which •re incomplete
•ud ofteninco•isteut. For computational
simplicity we •o •end to s• the ••,,• set of p••et• that d••be o•
obse•atio• or the E•h:s st•••e. The •v•se problem • th•efore ••eds• •d
o• h•otheti• E•h-model is p••et•• •to • •te n•b• of p••et•.
ßG•physic• inverseth•• is th• conc•• •th the appro•tion of oth•e
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

•nt•uo• •ctio• •th a •te n•b• of p•-met•. The foregong•s••on •


be •ustrated belowus•g density•d t•perat•e mod• of the E•h.

3.4 Problem formation

An inverseproblem c•u be statedas: Given a set of exper4m•emtal


d•ta, determinethose
par•,•eter•
ofthee•r•h(ormodel
) thatwillexpls•u
theobservations.
Thetaskof
formulatingthis problemmathematic•y restson ou• ability to distinguishbetween
what representsthe experimentaldata •nd that whichconstitutesthe modelp•rameters
The processof selectingvariablesto representd•ta •ud modelpsrametersmaybe
ß

broadlyreferredto a• par•e•••m (m,•uyworkersusethis.termfor p•rameter


selectiononly). We needto posethe problemin the form J•gm which is
computationallym•uageable. For ou• discretesystem,we have that

The E•rth c•u be par•me•erized into a n-•ber of discrete layers ea• with its own
density(•i) or seismic
velocity(•i) or electric•resistivity(•i), say.In otherwords,
for
the densitydistributionproblem,•tead of want• densityas a faunion of radiuswe
may be interestedin deter•• the averagedensitiesof the core•ud mantle.Fo•
ß

illustration, we w•. considera few examples-ofproblem formuLstionnext.

3.4.1 D•ty distribution within the Earth


Considerthe problemof determi•i• the averagedensityof the core(Pt) andthat
the mantle (p•) (seeFiõ. 3.1)from measurements of the Eagh'smassandmoment
inertia.In this case.we havetwo data values(mass=g•and momentoœ inertia
andtwo modelparameters(•=rn• aad •2=rn•). The problemcaabe formttla•ed as

moment
ofinertia (3.2b)

16
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ß

•7
',a
••.

Figure3.1 A simple parametedzationof the Earth


a =6371km

km
ß . . "•=3485
which can be written in our favourite short-hand notation as

d•=•• xGi•rn• i=1,2. (3.3)

or in componentform •s
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ß el ß le ß ee• eeß Ii ßeet el ß ß ß

This is •u evendetermlnedsystemof equationsa•d must be solvedfor the soughtmodel

3.4.2 Borehole temperatu• me•urements


Now, considerthe straightline problemof fitting downhole
temperaturedata •iven by
eq. (2.1). Suppose
that wemaden temperaturemeasurements
wanttofit a straight
linetothedata.Here,• -- [Tt,Y:i...,
•'•]•, •udtheintercept
ß

•d slope5 •re the twomodelp•r•meteTM,i.e.,m ----[a,•]•.


Now, by the forwardtheorythe data T must satisfythe relation T i = a + bzi. So

T•=a+bz•
Ta=a+bza

T,•=a+bz,,

which in the matrix form d=Gm is simply

T! 1 z•
1 za
ß

(3.4)
1 %-1
I zn

whichis an overdeter•i,•edsystemof .equations.

18
3.4.3 Digital ffdt• design
in seismic
decnnvolution
Two sign•l.•a(t) andb(t)mayberelatedby convolution
with a filter f(t) in the form

a(t) = f(t), b(t) = (3.5)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

A question
thatoftencrops
upin signal
aualysis
is (seeMenke,1984)'lmowh•a(t) aud

To addressthis issue,let us discretize


the problem.H the time seriesis of length n and
thefilte•is of lengthp, thentheconvolution
integralof eq.(3.5)canbe conveniently
replaced
bythecomputational
formula
P
a,= At• f i hi_i +• (3.6)

whereb•= 0 if { < 1 or i > n, •ud At is the sampling int•al.


Noticethat equation (3.6)is linearin the,•-l•own filter coefficients
fi andcanbe
recastin the formd = Gin,wherem = f (the sou•l•tfilter), d-- a (the time series
data) aad

b• 0 0 0 0 ...... 0
b2 b• 0 0 0
b• b• b• 0 0
G= A• • • • b, 0

ß ß ß ß ß

b• b._x b._a b._3 ......... ba


o

wherek = n-p + 1. The abovesystemis overdeter•i,•ed,i.e., p < n a•d caube solved


for the soushtfilter coefficients.
We will nowex•r•i,•emethodsof solvingsuchlinea• syst_•r•.•
of equations
for tl•e
soughtp•rameter estimates.

19
4. SOLVIl•G OVERDETERMINED
LINEARINVERSEPltOBI,]!2•
4.1 SimpleLinea•
H an inverseproblemc• be r•r••t• •th the •Hdt •• ••tion •m
it is s•d to be L•A•. H a p•e• (or ••t) relatio••p •sts •tw• the
obs•atio• • •d tee mod• p••et•s m, th• we c• •e v• s•ple pro••• to
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

•v• o• men, emits for m. • m•y pra•i• situation, tM obs•• •ta may
not • he on a str•ht li•e (• Fig.4.1.1).H wed•de to fit a •e to thee dat• the
fitted •e may be someappredable •st•ce away•om somedata•u•. F• a
co•tion oI n datap•s {(z,,•), (•,y•), ... ,(z.,y.)}, the5tt•'!i,e (•o• • the
•G•SSION •e) • d••b• BytM ••tion
= +
•d e• data p• sati•• the ••on
•i = a + bxl+ ei (4.3)
w•e eiisthev•icM •t•ce •tw• the•a datapo•t •d •e re••on • (Fig.
4.1.1).Thequ•tity e•is•• tM •S•U•, •2F• or pr••on E•O•
•e •lu•• • •e ••t •e •• pmbl• • this • • n• • •

ov•det•i•. TMs t•e of probl• • $•••y solv• •• the LEAST


method.

lu the leastsquares
methodwetry to MIl•IMIZE the erro: 6 by cletermi,•iaõ
thoseparameters
a,b suchthatthes,,mofsquares
oftheerro•($) is mlnlm•l•i.•
Minimize
2

Miami•.ation is accomplished
by differentiating
$ with respectto the model
parameters
sadsettingthederivatives
eitualto zero.

The mi,•imization procedure


•all that I• = (a+bz) + e. Assumingthat the experimentalerrorsaxeonlyin the
fielddata!1•,thenequation 4.3holdsgood.Next, di•erentiate$ (in eq.4.3)with
respectto a sadb sad equatetheresultto zero,viz:

05 n
o,,= i----1 o
05'=2 = 0
Ob

2O
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

x (e.g. Geophonespacings)
Figure4.1 A-simplestraight-line problem
giving
Ea + Ebzi= Eyi (4.4•)

axed Eazi+ Ebz•= Ez•t• 12simultaneous


linea•
equations
(4.4b)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

whosesolutionfor a a•d bis verystra•õhtforward


a•d shownbelow.Re-statingthese
two equations,we have that

z• = ,• + (•)• (4.•)
•nd

.(•) = .(•)• + .(•)• (4.4bxn ) (4.5b)

Next, m•tiply eq. (4.5a)by (2z) sothat we-c•usubtr• 'it •rom eq. (4.5b) to fi-d b.
We will fi-d that:

(•.•) .r.• = .(•.•)• + .(r•)•


m•rms(4.5ax Xlz) r•• = .(•.•)• + (•)(•.•)•

E•V•S-,

T•ere•ore•

• slopeof the fitted line.


b= nI:• (y.,z) (4.6a)

Finny, find a by substitutingb into equation4.5a.

a=
'E•-bF•
n = •-b•
}interceptonthe y-a.xiz. (4.6b)

where• •ud • axethe meanvaluesof y and z respectively.


The formulafor the leastsquares
regression
(bestfitting) line is therefore

• = a + bz. (4.7)

The aboveconceptsareusedroutinelyin geophysical data ana•y• •ud especiallywhen


de•li•g with problemswith oneor two parameters(e.g., a .simplestraight-li- fitting)
and the tech•ique is termedlinear recession analysisor classicalLeast Souares•

22
The methodwasorigirrally
formulated
to providea solution
to the overdetermined
problem but thesameapproach c•ube adopted forunderdetermined probl•-•. The
'solution
wasoriginallygivenby Gaussin 1809.Whenwehavemorethautwomodel
parameters thenwerequire a s•mple
extension oftheabove method referredto as
Multiple-regression an•. Howevex, it is possible
to formulatea õe•eralized
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

relationship
that will beapplicable
to anydimension of d,xtaaudmodelparameters.
Thisapproach is commonlyadoptedin geophysics audtheprocedure usesmatrix
formulations
insteadandis aptlydubbedGen•l•.ed LeastSquares
or Matrix Inversion
(GMI). Asa prelude
to GMI,wewillnowdemonstrate
aaapplication
oftheclassical
leastsquares
methodof parameterestimation.

F.ar•ple 4.1: Application


oftheclassical
leastsquares
method
in s•_'•c
interpretation.

clistaUces
x•,andsuppose
thatweca• represent
theheadwaveequation

by a linear relationof the form

t--a-i-
b:r (a.•
Th,b--
The errorsaxeaSmlmed
to be in the pickedarrivalt•mes• only.The beztfit to the
datawith respectto minimizing
the s1• of the squares
of the residualz,
e•--ti-(a+bz•),
is givenby (seeeq. 4.6)

= D
aud

a •
n D

withthest•udarcl
.errors
X•andX• being
given
by

23
where
Notethat X: is estimated
astherootmeansquaxe deviationofthedat• f• fromthe
calculated
bestline(a+ b•) butwiththefactorn- :2in thedenom•-ator sincewehave
twomodel
parameters
(a andb)in thisproblem.
Thequ•utityIn- 2]isreferred
to as
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

the numberof degrees.


of freedom
in thisproblem-

Illustrative Exercise.
A tableof resultsfor seismic
traveltimes'isgivenbelow.Completethe table,deter•-
theslope
•ndintercept
onthet•me-axis
ofthebest
fi•ting
linetothed•ta,aud•hen
ßcompute
theerrors
associated
withtheestimated
parameters
oftheregression
line.

Trace zi(,,,) ti (,,,•) xiti •


1 2 5.1 10.2 4
2 4 9.2 36.8 16
3 6 11.9 71.4 36
4 8 14.9 119.2 64

Table4.1.1. Seismic
first axrivala(t•) at foux'geophone
positiom(z•).

Solutios
Thefirsttaskhereis thecomputation
ofthesums
Elz,•t,•xt,•z •. Having
caJculated
thesesums,we aimplyobtainb •ud a usingequation
.
4.6. Next,wecompute
the
residuals
andthes,,moftheirsquaxes
andfinally
calculate
•b andXaa
using
eq.4.8.
Forex.
axnple,
consid•g thesefourtraces,wehavethatn=4,•Iz=-20,•t=41.1,
•z•=237.6,Zz==120.Theleastsquares
solution
for a andbis thus:

b= 4(237.6)-20(41.1) so tha• tr=623.053


4(120)-20:•
= 1.605;
and

a=
Et-bElz =
n
41.1-1.605(20•
4
_ = 2.25.

2•
It is left to the reader to completethe exercise.

It is remarkedthattheabovecalculations
caubeeasilydoneoncomputers
using
standardsoftwarepackages.A simpledemonstration of how the aboveformulationscau
be implementedon a computeris givenin the Fortran prograznR.EGKESlisted below.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

program R•GP•S
c performssimplelineazregression.
estimatesparameters
c a and b (in the problemy=ax+b) •ud associated errors
c based.on data scatter,•err and bert. yhat=predictedy.
c Input file containsthe n datapairs:x(i),y(i)
c Author: M•x A. Meju
dimensionx(200),y(2O0),xx(200),x'y
(200),yhat(200)
character*20 i•]e,out•e
c initializations
s•rnx=0.0
s•rnxx=0.0

s•mxy=0.0
S•rny=0.0

open(unit=3,•le=i=•l e,status='old')
i=1

I read(3,,,err=99)x(i),y(i)
i=i+l

•oto 1
99 continue

n=i-1

close(,mit=3)
write(,,'(a)')' enteroutput51••e > '

open(,mit=3,Rle=outRle,status='new')
calculaterequiredquantitiesxx,x7 a•d
do 10 i=l,n
xx(i)=x(i),x(i)
xy(i)=x(i),y(i)

25
sumx=s-•x+x(.i)
s•,•x=s,,•xx+xx(i)
sumxy=•umxy+xy(i)
sumy=sumy+y(i)
10 continue
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

c••ate the commondeno•ator •d the des•edrecessionp•met•.


ßden=nßs•m•-( s•mx*s,,mx)
s=(•.•,,••-(s,,•.•,,•y))/
•= (stay. s,,• •- s,,•x. s,,••) / d•
cM•ate root me• squ•e •o•
sse=O.O

do 20 i=l,n
e=y(i)-a-(b*x(i))
sse=sse+(e*e)
20 couture
•q=•e/(=-2)

be==n.•q/d•
c computebestfitting •e •d •v•age •or
do 30 i=l,n
yhat(i)=a+b*x(i)
•=(•-==)+(•-•=).•(i)
•==(•+ •=) + (• +•=).•(i)
y•== (• (• •• )+ • (••))/=.
•t•(a,.)x(i),y•t(i),y••
30 cont•ue
close(•t=3)
stop
end

Fig.4.1.2A Fortra•program
fors•mple
linearrecession
analysis.

4.2 Unconstr•i,,ed
I.ine.
a• • SquaresInversion:
The Generalized
Mata• Approach
AsaJludedto in theprevioussection,
thelea.stsquares
methodofparameter estimation
maybeformulated
in matrixnotation
sothattheresulting
algorithm-•
canbeapplied
equallyin the inversion
of singleor severaldatasetsfor oneor severalmodel
parameters.
Thisapproach
willbeadopted
in therestof thecourse.

$•epsinvolved'.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

IProblem
deftnit.
ion ' matrixformulation ' Le•t squ.
aressolution
I

The linearproblemis posed in the õeneralized


matrix'formd--Gin.We nowwant to
solvefor m. For perfectdata(i.e.,noexperknental
errors),
m=G-•d.
However.,
G•uss(1809)suggested
that dueto experimental
errors,practicaldat• di
would •ot/it the model exactly, i.•.,
d = Gm + ei
iredthat the bestway to get• -,•que solutionfor the .modelparameters
ß
is to
the sumof squares of the residuals
el. Thisprocedure rnr-i•.es the •l•erences
betweenthe.fielddata •ud that predictedby the forwaxdtheory,the so-c•ed euclideau
distance. We thereforewish to rn•n•rn•zethe quautity that gaugesthe

q= Je----• (di- • Gijrnj)


s•l
2 j--l,..,p. (4.9)
As before,mi-imi•.ationis effectedby •li•exentiatinõq with respectto eachoI the
modelparameters andsettingtheresultsto zero(fora maxim,,mor mi-rm,m). The
resultingequationsc•n thenbe solvedfor rn as illustrated below.
First, we re-write eq. 4.9 as
q = (d-Grn)•(d-Gm)

T T T T T T

There/ore,3q= O[dd-d Grn-rn


G d+rnG GrnJ
cqrn•
=0

or

-drG - Grd + G?Grn + rnrGTG = 0


giving
2GTGra = 2GTd the so-callednormal e4vagoas

fromwhichwe obtainthe leastsqusxes


solutionfor the psx•meterestimatesdenotedby
,h aad given by

z7
m = [GrG]-•Gr•.

This is the unconstrainecl


leastsquaressolutionto the inverseproblemd = Gin. The
quantity[G'rG]-•Gr iscalledtheleastsquares
Generalized
Inverse
andoperates
onthe
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

data d to recover•. The leastsquaressolutiongivenaboveis referredto as an


unbiasede_•imatorof rn in statisticalParlaace.It is obvious,however,that we would
haveto solvea setoœ
equations
to obtain,•. Some
oœthecommonly
Usedmatrix
solutionmethodsare discussed
briefly in the/oHowingsection.

4.30bt•i,•i,•g Ma• Inv•


ß
and• Squares
Solutions
Noticethat eq. (4.I0) isoœ
theform z=A-•I• (thesolution
to theinverse
problem
y-=A•) whereA- • isequivalent
to [GrrG]-• andy to G•rd.Themaintaskhereisto
find A- • , the inverseof A (calledA-inverse).The methodof matrix inversionto be
adoptedmay dependonthe sizeanddegreeof symmetryof A aad the desired
nvmerical robustness.For illustration, we will explorethree commonlyused'matrix
inversion methods.

4.3.1 Matzix solution •[ a small set of equations


4.3.1.1 Cromer's Rule
To find A- • for a small set of.equations,the •ollowingstepsmay be •ollowe• (Stroud,
1986):
eva/uatethe determ•,•antof A 1,41
form the matrix of cofactors oœA denotedasC (the co{actorof aa dement is its
rni,•ortogetherwith its 'placesign')
findthetr•.•.spose
ofC (i.e.,theacljoi•tofA) cl•notecl
as C'r or add(A).
(iv)thencalculate
A- • = 1 .c•r
ß 1
The
desired
solution
isgiven
by= {JA]

Example ß Supposethat A a•d !/are given as

4 5 i 2

A' •-2_-,32
and
,=
-1

28
We •ve

4(4:-3)-5(- 2+9)+1(- 1+6)=-


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

so that

CllC12
'C13
• = |c•.c•: •
L C31C32C33
and

Cll C21 C31


C'r= cx• c• cs• { the matrixin applications;
programmed see
C13 C23 C33

wherecn--[-2-3 - l---7;

Thus,

C • .--11
13 13 -

The inverseof A is •iven by

A-•= 1 c'r
I.1 '

enablingus to deterre;hethe solutionvector

=:3 -1

It is easyto write a computer


programto inverta 3 x 3 matz/x.A simpleFortran
program
CKA.
MIN toeffect
the=boresolution
procedure
isgivenin Fig.4.3.1for

29
illustration•ud incorporates
• subroutine
for matrix inversionby Cramer'sP•ule.It is
reiterated
thatweh•vetoformthesquare
matrixA = •rG, andthevector• = •rj
in orderto solvethe leastsquares
problemusingmatrix inversionby Cramer's•ule.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

programC1LAMINV
c simple demonstrationprogramfor matrix inversion.
c solvesa line• systemof equationsy=A.x.
c input: (1) Matrix A, elements
re• rowby row.
c (2) Vector y, elements
readin oneline.
c output' solutionvector, x.
c method:A invertedusingCrmmer'sRule to give •uother matrix,Ainv.
c Ainv is then post-multipliedby y to givex.
c author: Max A. Meju
dimension'•(3,3),Ainv(3,3),x(3),y(3)
write(*,,)' Hello •here.Welcome•o • 3-by-3equationssolver!'
write(*,,)
write(,,,)' Enterelements of the A-matrix(orGTG), rowby row'
do 10 i=1,3
write(*,11) i
11 form•t(2x,' l•ead;ngl•w: ',i1,' <enterrowelementsbelow>')
read (.,.)(A(i,j)j=l,3)
10 continue

write(.,,)
c read data vector y
write(,,,)' Enterelements
of y-vector(GTd),•11i•l O]2.e
.
lille please'
read (*,.) (y(i),i=l,3)
c ca• matrix inversion routine

call cr•3x3(A,Ainv)
c calculatethe inne• productAinv.y aud return a• vector x.
callinprod(Ainv,y,x,3,3
)
c wri•:e ou•: solu•:ion

write(,,,)
write(.,.)' Here a•ethe orderedelements
of solutionvec•o•.x'
write(*,,)
stop

3O
end

subroutinecr•m3x3(A,Ainv)
dimeusionA(3,3),•mv(3,3),½t(3,3)
bl--(a(2,2)*a(3,3))-(a(3,2),•(2,3))
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

b2=(a(2,1)*a(3,3))-(a(3,1).a(2,3))
b3--(a(2,1)*a(3,2))-(a(3,1).a(2,2))
c calculate the determinant of A.

det=(a(1,1),bl)-(a(1,2),b2)+(a(1,a),b3)
write(.,.)' the determinaatof A-matrix is: '
write(.,.)det
c computethe elementsoœC-tr•spose (C=Matfix of Cofactors
of A)
C•(1,1)=(a(2,2),a(3,3))-(a(2,3),a(3,2))
C•(1,2)=-(a(1,2),a(3,3'))+(a(1,3),a(3,2))
C•(1,3)=(a(1,2),a(2,3))-(•(1,3).a(2,2))
C•(•,l)=-(a(2,1)*a(3,3))+(a(2,3),a(a,1)).
C•(2,2)=(a(1,1),a(a,a))-(a(1,3),a(3,1))
Ct(2,3)=-(a(.1,1).a(2,3))+(a(1,3).a(2,1))
Ct(3,1)--(a(•, 1).a(3,2))-{a(2,2).a(3,1))
ct(a,2)=-(•(1,1).•(a,2))+(a(1,2).a(3,1))
C•(3,3)=(a(1,1).a(2,2))-(a(1,2).a(2,1))
c computeA-inverse= Ct/det(A)
do i=1,3
do j=l,3
Ain(i,j)=C • (i,i)/det
end do
end do
return

end

SubroutineInprod(A,y,x,n,m)
c matrix by .vectormultiplicationor im•er product
dimensionA(n,m),x(m),y(n)
do 10 i=l,n
s•m---0.0

do 20 j=l,m
s• =s•rn+ A(i,j).y(j)
2O continue

X(I) =Sl,rrl
10 continue

return
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

end

Fig. 4.3.1 A program


forsolving
a small-size
(3 x 3) system
o•equations
using
matrix inversionby Crarner'srule.

Note that a slight modificationof the abovestrategyis requiredfor a 2 x 2 matrix.


It is obviousthat we cannotcalculateC •or sucha matrix usingthe methoddescribed
abovefor the 3 x 3 matrix. In general,for a squarem•trix A, if thexeis anothermatrix
F such
thatthematrix
product
• = I, thenF = 'A-t audisgiven
by

1 -B

provid•g that it exists,i.e, der(A)5• 0). The matrix B playsthe role of C in this case.
Let us see how easily B can be determined. Given'the matrix

A •

we have that der(A) = ad- bc.Now the matrix product

c) 0

c -c •

I• is obviousthat divi,t•g thisproductby de•(A) will yieldI. The desiredmat•, B


is thereforethe post-multiplyingmatrix on •he left-h•d sideof •he abovematrix
equations,i.e.,

d -b
B •

32
We couldelect to find the exactsolutionrn -- G-td or the leas•squaressolution
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

rn -- [Gr•] - •rd sincebothapproaches


willleadto thesameresul•for thisproblem.
Forthe{ormer
approach,
wehave
thatG-• = (1/de•(G)).
B where
def(G)= (-32 - 20)=-52 andB = -2 -4-10]
8

so that

156
].
Usingthe leastsquares
approach,
let us define

and

We have that

-72
def(A)
= (68
x116)
- (72
x72)=2704
and
B= -72
11668
yiel•l• the estimates

2704-72 68 -- [-5408J-- '

4.3.2 Matrix solutionoI'a large setof equaticma


4.3.2.1 Gauss •'.llmi,•,tion Method

Gausseliminationmethodproceeds
in threesteps:initi•li•.ation,forwardreductionby
elimintion aad backward substitution. We will illustrate theseoperational steps using
the problemIt = Az where
y= 7 a.udA= 4 5 -31 .
1 -2
1 3 -1 -2

Step 1: l•tia•zation.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

The solutionprocess
is initia•zed by formingthe partitionedor augmentedmatrix
4 5 1 ' 2

3 -1 -2 . 1

that is,

{=1,
2,
noting, however, that p---n.here.

Step 2: F,]•m•nation.
Successive
forwardreductionof the abovepartitionedmatrix is effecteda• this stage.
We want reduce the coefficientsof A• suchthat'for the original coefficientn•t• part,
the diagonalelementa(i,i) is ,•ty and the subdiagonM elementa(i,i-,,,) is zero.When
this operationis complete,the systemis saidto be triangula_•ed..To triangula.•e
dividetheiz•rowbytheelement
a(i,i) Wingtheequivalent
rowvectorr(•); thenfor
eachofthesuccessive
rows (i.e.,row•, •---i+• ton) multiply
theelements
ofr(•)by
the constantfactora(•, i) andsubtractthe resultfromthe • row of A.. Note that the
elementsof the i• rowof A. •re not moa•ed at this levelof operationsbut the rest
are.These
operations
•rerepeated
withthenextrowi +1furnish•,•g
•herequired
r(•'+1),
a•d so on. The procedurewill becomeclearwhenwe tacklethe abovesampleproblem.
Starting
fromrow1, rfl)--[1,1.25,0.25' 0.5]•udthemultiplicative
factors
a•egiven
by a(•, x)= 1 for row 2 •nd a(s,•)=3 for row 3 leadingto

4 5 [ 2
0 -3.25 -3.25 • 6.5 .
0 -4.75 -2.75 ] -0.5

Next,repeat
theoperations
using
row2,viz; divide
row2 by-3.25yielding
r12),
and
subtract
-4.75xr(•) fromrow3 to give
4 5 ! : 2
0 -3.25 -3.25 : 6.5 ..
o o 2 :

Thiscompletesthe triaugula•zation
process
forthechosen ex•rnle •ud the solution
for
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

z cannowbe foundby backward substitution(workin•up fromthe last row) asshown

Step3: Back-substitution.
Starting fromthelastrowofthereduced partitioned
matrix,wehavethat 2zs-- -10
or•s -- -5. Substitutin$thisbackintothesecond rowgives(3.25x5) - 3.25z• -- 6.5
or z2-- 3. Finally,forthefirstrow,wehavethat4z• + 5x2+ • -- 2 or zl -- -2.
The soultion is therefore


Noticethat the eliminationoperationcanbe described
by the formtda

aij.= aij -- akj kill:k/


{=k+l• k+2..
1,2,..
,-x
att .he. 0
and the back-substitution by

an• n -4-1
ann

lli,n+l -- • llijZj
lEk • _

Notethatthewholesolution
process
requires
aboutn:•/3multipUcati0ns,
us/3additions
a•d aboutn•/2 divisions.

4.3.2.2 Ga•-Jorda• Method


This is a• extelion of the GausseAimination
methodand consists
of.threeoperationz
(1)initi•ation, (2)norm•.ation
•ud(3)forward
andbackward
reduction
by
elimination.
The initi•zation processis asfor the Gaussel•-•,•tion method.In
normalization,
a permanent changeis madeto thepartitionedmatrixby scaling each
row by its pivot elementaii. The coefficientmatrix paxt of the partitioned matrix is
then reduced to an identity matrix by eliminatingthe offdiagonMelementsof each'row
ß

and column.The solutionis givenby the augment•g segmentof the final partitioned
matrix, i.e., the cob•rnucorresponding
to • in the in/tiM matrix.
Using •he previousex•ruple,we have that
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

4 5 1 ' 2
= = -s : .
3 -1 -2 :
ß
1

Next,normalize
(ord/vide e•chelementof)rowI by4 togiver/•);subtract
a• xr
fromrow2 anda• xr(•)'fromrow3 to give

I 1.25 .25 : .5
0 -3.25 -3.25 : 6.5 .
0 -4.75 -2.75 : -0.5

Similaxly,
divide
row2 by-3.25togiver(2)andsubtract
-4.75xr(2)fromrow3 (forwaxd
reduction).
Wealsosubtract
1.25xr•) •romrowI (backward
reduction)yiela•-g

I- I 0 -1 : 3
0 I I : -2 .
0 0 2 : -10

In thesamefashion, normalize
row3 by2. FinaJly,
subtract
1xr(•)fromrow2 a•d
-1 xr(•) •romrow1 (backward
reduction)
to givethedesired
partitioned
matrix

3. 0 0 : -2
0 • 0 : 3
0 0 1 : -.5

whose last column is the solution vector m. It is obvious that if the determ•,•aut of •4 is
zeroor if the pivot elementa, is zero,then the aboveoperations
are not possible.This
is also true for the Gaussel•m•,•ationmethod discussedin the previoussection.What is
donein many practicalsolutionschemesis to interchaugerowsor cob•m,,.•of A. if the
pivotelementsarefoundto be zeroor to usethe l•rgestcoefficient
in • rowor
aspivotelementsbut suchoperations requirea very $oodbook-keepingsystemfor the
interchanges
in orderto reconstructthe correctsolution.
4.3.2.3LU ( or Tx•angular)
Decom••• Method
A squ•rematrix ,4, canbewrittenasa produc•
oftwoothermatricesL andU, i.e.,

Lt• 0 0 uu • u•3
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

A=LU= L2i L22 0 0 •1,2:2


•1,23
L3! L32 Laa 0 0 u3a

Ll1•11 Ll1•12 LiltS13


L2•Ull L• u•2+ L•2u•2 L2•u•a+ L22u2a
Lsiui• Lsiui• + Ls•u• Lsiuis+ Ls•u• + L•uss

whereL is a lower-t•'••• m•t• (i.e., h• nonon-z•o•••t •ve the m•


•agon•) •d U is• upp•-t•••• mat• (•h nonon-zero ••• S•owthe
m• •a•on•). Noticethatboththeelects u, •d L, •e cont•• • them•
•agond. Sped•g eith• L, or•ii •ows • todete•inetheoth• elects ofU •d
L. For example,let uu= u2:=...=1 (or •t•atively, put Ln= •==1) • • c•••
practice, then

/111 al• /113


A •

/13! /132 /133

L2x L2!ux2
+ L22 L•!u13+
L31 L31u•2
+ LB2 L•I1/13
4'L32•34...•

enabliag
thevalues
ofLi•alldUijtObeeasilydetermi,•ed.
Theprocedure
issimple.
Starting
formlocation
aii(i=1),weworkfirstdown
thei•acob,
m,•aa•d
thenalong
the
itsrowfind•g thevalues
ofthesystemcomponents Li• andui• by comparison
withthe
givencoef•cients
ofA. Thisin essence
is Crout'salgorithmfordetermi,•i,•g
L a•d U.
Havingdecomposed A, howdowep•oceed to solve
thep•oblem • ?.Wenotethat
theequationA =/XI impliesthatAz = LUz andtherefore Az = L(Uz)-- !/ßWe

•7
therefore
put Uz - c andsolveLc = y to obtainc. Finally,wesolveU•: = c to obtaia
Let us demonstratehowthe methodworksby solvingthe problemAx=y where

3 -1 I
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Now, let ull-- u::-- ...=1 in the equationA=LU sothat

Lll Lllul2 Lllui3

A • L21 L2•ul2+ L22 L21ul•+ L22u•


3 -1 -2
La•ui• + L•2u2•+ L•

From the fixstcob•mnof A (startingat locationall), wenoticethat Lil:4 , L:I=I •ud


L3x=3;andsofor therestof thefirstrow(i.e.,positions ax=andaxe)we havethat
u•-=5/4=1.25, uls:1/4=0.25. Usingthesevaluesaad proceeding dowathe second
cob•mufrom locationa::, we fin that L= = -3.25 and La: =-4.75. World- alongthe
rest of the secondrow, we obtain u2• = 1. Finally, from the tkird coB•m•, we have that
La3=2.The decomposition
is there/ore,

4 0 0 1 1.25 2
1 - 3.25 0 0 1
3 - 4.75 2 0 0

Let us now proceedto determin the solution.We have that

4 0- 0
1 -3.25 0 cl
c: :. 27.
3 - 4.75 2 c3 I

Workin fromthe top(forwaxd


substitution),
wehavethat 4cl: 2 or c1: 0.5. From
the next row, we obtainc• = -2. The last -•lmown is then determinedaz Ca: -5.
Using
these
values
determined
forc,wesolve/or
z inthesystem
U•----c,
i.e.,

1.25
1
0

38
By back-substitution
(startingfrom the bo•tom),.wefind tha• z•= -5, z==3 andz•=-2
which•greeswith •heresultsobtainedusingthe othermethodsof solution.

It is possibleto developa computational formulafor the aboveLU decomposition


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

operations.1%ec•-•]that the first thingwe did duringthe decompositionof A wasto set


u•=l, for i=1 to n. We thenfoundthe valuesof œa(elements of thefirst
i=l,...,n) asLa= a•l andthe corresponding
rowelements
asuti=•ii/Li• . These,and
subsequent
col-mu androw m:•u•pulations,
are described
by the formula
j--1

Li•• ai•- Z Liklzt•J


fore•chi -- •,...,n •c•--x
i--1

j---i+l;...n.
•ZiJ
=• -- =
where the s•rar•ation terms are set equal to zeroin •;hef•rst cob•mnaud row
manipulations.No•ethat the schemebre•k.•downwhenthe determ•ua•to•[A (or œ•) is
zero. To preventany undesirableeffectsin LU decompositions,
it is thus •mportaut to
employ
a partia•
pivoting
technique
in which
rows
a•einterchanged.
This•saneffective
stabilization measure and is described in most standard texts.

4.3.2.4 Genm"•l;•cl matdx inversion

Althoughit is rewara•- •o understand the variouswaysof solvinga linearproblem,


quite o/ten we seeka generalstrategyfor solvingsuchproble•-•. The main task in the
development
ofa general
purpose
linearsolution
package
isfmai, a single
routine
for
invertingthe •ivenmatrixA or Grr•. Fortunately,
therearemanypublished
routines
in
the ligeratu•e•hat canaccomplish
this •a•k (seee.g., Presset al., 1986) making program
development
easy;wes•mplyprovidea control,•-•t or driver.pro,ram
ß
for the available
subroutines.For illustration,an inversionprogr• LINPACK incorporatinga
generalized
matrixinversion
routine
isgiven
inFig.4.3.2.

program LINPA CK
c A simple demonstration
of the amalgamationof existingsubroutines
c •nd a driverunitformiug
a general
purpose
linearinversio
n package.
c n= no. data(max=20);
m= no.par;,meters(max=10).

•9
c Data enteredon-lineor readfromdislcFi]e
containing
n linesof field data (d)
c followedby m rowsof the designmatrix components
(G).
c Author: Max A. Meju

a•e•io• a(20),c•(t0,e0),c(•0,t0),c•a(•0),==(t0)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a•e•io• ^(t 0,t 0),•v t(t 0),•p2 (10)


chazacter.1 ans,infile.20

write(,,,) ' Enterno. of rows& col-n inG' NROW,NCOL'


read(,,,)n,m
••(,,'(^,$)')' • •o• a• •o•d • • • ? •:•'
read(*,'(A)')•
ff(.•.eq.'Y'.or.•.eq.'y')thea
•te(,,*)' Enter •PUT filename'
read(,,' (A20)')i•e
open(nnit=3,file=•ule,stat•='old')
i••=3
ehe
i••=5

•te(*,,) ' Now ent• Field data d, oneat a t•e'

do 10 i=l,n
•a(••,,)
10 cont•ue

•(•.ne.'Y'.or.•.ne.'y')then
•te(*,*)' Now •t• •mat• row by row'

do 20 i=l•
r•a (i•m,*)(C(id) • = t •)
2O cont•ue

•(i••.EQ.3)th•
close( ,mit =3,state= 'k•p •)

c n• fo• G-tr•pose, get Gt,d, GtG=A, md A-•v•e


do 30 i=l•
do 40 j=l,m

•o
Gt(j,i)=G(i,j)
40 continue
30 continue

call inprod(10,20,m,n,G•,d•G•d)
call mxprod(10,20,m,n,Gt,G,A)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

call generalizedmatrix inversionroutine


CALL INVERT(10,m,A,wkspl,wksp2)
c fm•d]y, calculate solutionvector, xm
call inprod(10,20,m,n,A,G•d,xm)
c compute errors and print results
do 50 i=l,m
•t•(,,,)=•(i)
50 contimue

stop
end

subroutinehN•ROD(mm,•,•,m,n,A,x,y)
m••o• x(=),y(=),A(•,=)
c matrix A multiplied by vector x, returns vector y
do 30 i=l,m
s-m=0.0

do 40 j=l,n
s-,•,=x(j), A(i,j) +s,,•
40 continue

y(i)=sum
30 continue
return

end

subroutineMXPROD(• ?,•,m•,A,B,C)
c matrix multiplication: A,B=C
a••s•o• ,(mm,=),s (=,•),c (•,==•)
do 10 i=l,m
do 20 j=l,m
sum=0.0

do 30 k=l,n
sum=sum+A(i,k).B(k,j)
30 continue

C(i,j)=s-•
20 continue

10 continue
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

return

end

SUBROUTINE INVERT (NN,N,A,B,C)


C INVERSION OF NONSYMMETILIC MATR]• A

DIMENSION A(NN,NN),B (NN),C(NN)


C N=SIZE OF GIVEN MATthIX,N>I
C B AND C AR• WOR/{ING SPACE VECTORS
ß

c adapted •rom Akin, J.E., 1986: Finite Element Analysisfor


c undergraduates,AcademicPress,p.20.
Ni=N-1

x(•,•)=•./•(•,•)
DO 11 M=I,N1
K=M+I
I DO 3 I=I,M
SUM=0.0

DO 2 J=i,M
• SUm=SU•+X(I,J),•(•,K)
3 B(I)=SU•
D=0.0

DO 4 I=i,M
4 D=D+A(K,I),B(I)
D=-D + A(K,K)
A(K,K) = 1./D
DO 5 I=I,M
5 A(I,K)=-B (I),A(K,K)
DO 7 J=I,M
SUM=0.0

DO 6 I=i,M
• SUm=SU•+A(K,•),•(•,•)
• C(J)=SUM

1;2
DO 8 J=i,M
A(K,J)=-C(J).A(K,K)
DO 10 I=i,M
DO 9 J=i.M ,

9 A(I,J)=A(I,J)-B(I). A(K,J)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

10 CONTINUE
11 CONTINUE
RETURN
END

Fig. 4.3.2.A linearinversion


packageincorporating
pre-existing
routines.

It wasmootedpreviously that the GaussmethodsandLU decomposition techulque


requireadditionalstabilization measuresto be effectivein gen•M terms.What wasnot
mentionedthen'isthat theme algorithms dond.tBmzishuswith diagnostic i•ormation
about the'causeof any undesirable effects.A morei]l•,rni•&tillgalia robustmatrix
inversiontechniqueis •lmished by the singularva/uedecomposition method.For
example,singulazproblems-can sometimes be turnedinto non-singular onesby this
methodwithout any addedsophistication ,mllke the other solutionmethods(seePress
et M.,1986,p.20). As.we shallseelater on, it a/soallowsus to pin-poiutwhat the
problemis andis thusthefavoured approach in leastsquaresinversion of geophysical
data.

4.3.2.5The singula•valuedecompo•tion(,5'•D) of a ma•


An n xn or n xp matrixG, say,canbefactoredinto a productof threeother
matrices ß

where
forn dataandp parameters,
iY(n
xp) andV(pxp) arerespectively
the'data
ßspacemudparameterspaceeigenvectors, andA is a p xp diagonalmatrixcontaining
at mostr non-zeroeigenvalues
of G•, with r <_p. Thesediagonalentriesin A
...,Ap)
arecalled
thesin_uml•
values
ofG•.Thisfa•'to•ation
islmo• asthesingular
value(or •pectral)decomposition
SVD o{• (Lamczos,
1961).If theeigenvalues
of
matrix a•e small, the matrix is said to be ILL-CONDITIONED. The SVD method is
very popularwith geophysicaldata analystsbecause it is mathematicallyrobust and
numericallystable and alsoprovidesother vital informationon the state of the model
and data thus enablingmodelresolutionandcovariancestudies.

Applicationof SVD to Generalized Matrix Inver•on


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

The leastsquaressolutionto the inverseproblemof the form d=Gm is givenby

m = (G'rG)-•GTd

wherethe quantity[(G'rG)- •G• operates


ond to recover•. Thisoperator,theleast
squares
generalizedinverse,isal6u•o theG-• (G-inverse)forperfecgdata(described
above).The SVD technique is commonlyusedin geophysicsfor solvingthe NOKMAL
EQUATIONS for •he least squ•resestimates•.

Let us knowexpressthe aboveinversionformulain terms of the SVD of G. We have


that Gr'= VAU'r. Thus,

G•rG = VAU • ßUAW r = VAZVT

sinceUTU' I. The inverseofthismatrixis s•mply(G'rG)- • = VA- •V 'r. The


•eneralizedinverseis then givenby

(GTG)- XG
'r = VA- 2V'r-VAU'r = VA- 'U'r

since 1rrv = I.
Notethat the generalized
inversemay alsobe derivedin the followingsimpleœmhion:

(G'rG)-• = G- X(G
•) -• = VA- •UT. UA-•V'r = VA- :F'r
so that

(G'rG)- 'G'r = G- '(G'r)- :G•r • G-• = VA- tUT.

In •uy ca•e,the leastsquares


solutionis giv• by

m= (GTG)
- XGTd
= VA-XUTd]
orsimply
m= x(UT• (4.12)
ß

vectoT
r m•t•x
T vjtor
Equation(4.12) is progr•mrned
in routineapplications.
A comprehensive
linear
inversion
program
SVD• (seeMeju,1994a)usingtheSVI) methodis'givenin
AppendixA andw• beusedfor demonstrating
varioustecb,•ques
of parameter
estimation and uncertainty analysis.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

4.4 S•mpleApplications
of Unc•ed Invexskm
We will •pplyeq.(4.10)to threesimplecases
to Klustrate
its usmSKuess:-
(i) a simple
seismicrefractionstraightline problem,(•i) determ•-ation
of the averagedensityof
the Earth,and(iii) a simpledelay-time
seismic
experiment
(basictime-terms•lysis).

F•x,•ple4.4.1' Thestraight
llneproblem
' Intm'pr•fion
ofseh•mic
r•r•on
Giventhefollowing
seismic
refraction
travel-•ime
da•a( seeTable4.1.1),
ti
2 5.1

4 9.2

6 11.9
8 14.9

deter•-e the wavevelocityin the subsurface.


Compare
yourresultwith tha• obtained
g•phlca21y.

Solution:

Recall that the equationof • straightFreeis givenby y=a-l-bz. In the notation used

this section, this can be re-written as

di.= ra1 + •7•2•i

sothat the matrix equationd=Gm is of the form

'd• 1 xt
= 1.
We needthematrixproducts
GrrG aud•rd to determ•e•. Thesearegivenby

•5
and

d1
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Thus the least squaressolutionis givenby

m = [G•G]- 'Grrd= n Zzq


-1 Eldi'
How dowe obtain[G'rG]- • ?
Tnthiscase,weaxedea/ing
witha 2x 2 matx4_x
sinceEz,Exa,andn arejustsingle
n-tubers.We may therefore
usethe proced•eoutlinedin Section4.3.1to obtainthe
inverse of G'rG as

[GG] = iG•G[ ßB

where

and

z a - Zz
B •

Thtm,

•6
or s•mply(if wereplaced in GTdw/tht )

•.ud nExt-ExEt
Et•x•-ExExt.'
n•x•_(•x)•' --
Noticethat the abovesolution
is exactlythe s•meas •h•t o• eq.4.6 (Section4.1)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Thereforefrom Table4.1.1,n--4, E•---20,Et---41.1,Ext-•237.6a•d Ex•--120;andwe


findthatm•--2.25
•udm•--1.605.
Thus,
v - • -- 623.053
m/s.

E•xmple4.4.2 ' Determ•u•tion


of theaverage
d,•n.,dty
of the Earth.
Given that the earth'smassand momentof inertia are both relatedto the density
distribt[tion
withinit, canweestimate
theaverage
densityof theearthfrom
measurements
of its total masssad momentof inertia ? This inverseproblem(adapted
fromK.A.Whaler,
1986,
unpublished)
isoverdeterm•,•ed
sincetherearetwomeasured
data and we seek only one model par, meter.

P•l•m••ies:

Oux elementaxyforward theory tells us that

mass
ofearth,
M='•3TR3•
a•ld

moment

inertia,
I= •M/•2
=
whereM and I are measurable
quantities(whichserveas our data in this inverse
problem).Let usintroduce
a use211
practicalconcept
at thisjuncture.Noticefromthe
aboveforwardrelationsthat the two typesof data alf%r in ,•_•itude by a factor
about/•2. Sincewe aretryingto retrievea co,•monpara,•eter,• fromthe joint
measurements• M •ud I, it is desirable
to norm•li•.ethem to a co,•,•on-scale
by.
the measurementof the momentof inertiaby R2, say. Ottr worM,•gequationswould
then look like

d,=i•'/• a.ud
d•= _• =18•'•'/P•
.

Let usnowdefinethe interpretivedatafor this illustrativeexercise.


We will usethe
datagivenin L•-mbeck(1980,Table2.5) asourfieldmeasurements, namely:

•7
/{=6.371x 10•m,themeanradiusof theeaxtk;M=5.974x 1024kg;
andfor simplicity
we will ass,•rnethat I/MRS=0.33. Thus, our chosenset of data consistsof

•d
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

d:•
=mømentR?
inertia
= 1.97142x 1024k!7

This completesthe data preparationoperationsand we may now proceedto solvethe


inverseproblem.

Solution:

First, form d=Grn

1024
5.9740x
10
24
1.97142 x

T
d G rn

where
K= •-•rR
3= 2.166414
x1020
ma.
NextformGTGandfindthegeneralised
inv•e (•rG)- •

Noticethat G•rGis just a m•mberherea•d soits inverseis simply

29K2 ß

Now,obtai,,G•rdas
5.9740x 1024

G'rd
=/t'[5
2] 1.97142 x 10
24 =Kx33.813
x10
24
' just
are,tuber.
Flually,obtaintheleastsquares
solution
fortheaverage
density
•=• = (GTG•- zGrr.
d.
This is simply
-3

- / i }(Kx33.813x1024)
• = 29xK2 =5382
kgrn
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

F.,x•mple4.4.3 ' A sireplied seismicrefx•on t•ne-• arialye.


Consider
therefraction
seismology
del•y-time
experiment
inwhich
twoshots
St •nd
axerecordedat threereceiverpositionsR1, R2 •nd Rs asillustratedin Fig. 4.4.3.1.
The subsurfacestructureconsists
of two layersof velocityv0a•d vt (wherevt • v0).
The travel-timebetweena shotanda receiveron the sud•e •i is givenby

forward model
t•i = • 4-5•4-7i

wherezq is thehorizontal
dist•-cebetween
shot•ud receiverpositions
a•zd5i az•d7i
axethedelaythnesassociated
withtheizhshot•ud the 3•hreceiver
positions.

For a particularexperiment,
the followingresultswereobtaind (Hatton,Wor•hi-gton
and M•_'n 1986).

(i,J)
1,1 2.323 6.000
1,2 2.543 6.708
1,3 2.857 8.485
2,1 2.640 7.616
2,2 2.529 7.000
2,3 2.553 7.616

Table 4.4.3.1 Seismic travel-time data.

It is requiredto obtainthe velocityv• a•d the delaypar•nters 5 and7 from these


travel-t•me data. Formulateand solvethis problemusingthe generalised.matrix
inversion method.

•9
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Sl
S2

A simple

50
delay-time
experiment
R1

R•
2

(afterHatton
etal.,1986)
A_.su_•_ested
solution process:
Accordingto this model,for eachshot$i , the observation
at the threerecording
positions
R•must
satisfy
•i•= • + 5i+ ?•ßThus,
;gil
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

•i2

Consequently,
for the twoshots$x sad $2 themodelparameters
of the problemare
5x,
•2,7x,
72,
73,and
• (orvx)and
thematrix
formulation
oftheproblem
issimply

10100
i 0010
100. 01
0 i i 00
(4.13)
01010
01001
T
d

Notethatzi• appear
onlyin thedesign
matrix(kernel)
sincetheydescribe
thegeometry
o• the experiment•ud a•enot considered
• datain the equationd=GYr•
Now proceedto solvethe problem:

[•rst,calculate
the$VDof(•i]

andthenobt,,i,•m asVQI,U'rd.
Inext,A=VQI,find rfrd
find
--V
The resultsof the decomposition
oœU are s,,mmarised
in Fig. 4.4.3.2.below.

$bgulazVlues, A =
17.97092
1.732372
0.1539608

6.7762068E-07
,- •4 near-zero;
willcause
problems
wheninverted

51
1.420821
1.414207

U-tranpose
matrix(UT)
-0.3373460 -0.3764554 -0.4749875 -0.4267564 -0.3927588 -0.4271839
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

0.4197269 0.4190062 0.4152228 -0.3976767 -0.3971975 -0.3999271


-0.5317655 -0.00106414 0.4531686 0.4847122 -0.04687573 -0.5241370
0.2695436 -0.5769362 0.3073922 -0.2695442 0.5769363 -0.3073921
-0.3377683 -0.3011901 0.5493414 -0.3341068 -0.3062310 0.5366699
0.4904181 -0.5090613 0.01867663 0.4904131 -0.5090622 0.018678255

V-matrix dividedby A (V/A =VQI).


-0.0036809710 0.4178268 -3.360666 659978.2. -0.044392757
ß .
0.000016756792

-0.0038603232 -0.3981i71 :3.6407• -659976• -0.051352862 0.000016076374

-0.0023659754 0.00734722 -1.985073 .-659977.9 -0.3328176 0.4904165

-0.002381803 0.00726665 -2.022428 -65•74.2 -0.3008910 -0.5090644

-0.002793490 0.00509664 -2.993924 -659975.8 0.5379646 0.018677600

-.055216759 -0.00090729 0.8023227 -0.1106218' 0.0065730996 - 0.00000217039415

Fig. 4.4.3.2'P•esults
of SVD calculations.

The least squaresestimateso{ the souF•htpar•rnetersare:

[•=-$32.}•111"
•=-532.89,
,, 7•=õ33.63,
•:•=533.68,
, '73=533'.515,
I
Noticethat all but the slowness
parameterestimatesare-,,reaJistic.
The correct
estimatedvaluesof the modelparameter•are(c[. Hattonet al., 1986)ß

Sowhatwastheproblemin ourpresentcalculations
? A f,lanceat the valuesof the
singula•
values(,•)willreveal
thatthevaluecorresponding
to œourth
parazneter
ofthe
problemisnear-zero
(6.7762068E-07)
andthattheelements
oI thecorresponrli,•g
col-ranof the matrixvqI havealmostcollstalltvalues( ,•, 4-659978). The valueof 14
ismostprobably a.uartifactofthe computer
used(round-off error! ). Thishighlights

52
oneof the problemsfaceddurin• n•rnericalmatrix inversion- smalleiõenvaluescause
a greatdealof problemandthecorreponding
parameter
a•d cob•m•.q
or rowsof
associatedmatrices shouldbe deleted from the interpretation processor some form of
problem
regularisa•ion
measure
introduced.
Thissubject
willbeexplored
in detailin
Section 5.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

It is alsoobviousfrom theseresultsthat onlythe velocitycanbe accuratelydetermin-


ed.The delaytimesmaynotbe n•iquelydetermined
usingthe uncomtrmedsolution
process
required
byouroriginal
problem
formulation
(eq.4.13).Thisissobecause
an
alterationof the valueof/• followedby a compensating
adjustment
in the valuesof 3'
will leavethe systemofequations(4.13)unaffected
(seeHuttonet al., 1986).However,
it is comforting
to knowthat•heremedyto thisproblem is simplyto specify
a priori
the valueof oneof the delaytimes.Thishelpsconstrai•theleastsquazes solution
process
by actingmoreorlesslikea 'path-6•der'enabling
theresttobesafely
found
Thesubject
ofconstrained
inversionwillbeexamined inthefollowing
discussions
but
first,let usfind outhowmuchwehavelearntsofar in theformof anexercise.

4.5 PROB•M SESSION


4.5.1Giventhe following
boreholetemperature
data,findtheslopeandintercepton the
z-axisandhencepredictthe temperature
at a depthof 390m.

Depth,•(m) Temp,t(øc)
30 25.0
70 26.2
180 29.7
250 34.3
300 35.5

53
5. CONSTR.AINED LINEA• LEAST SQUARES INVE•ION
In many geophysical problems it is possible
to generatea setof completelydS•Ferent
solutionsthat adequatelyexpl• the exper•mentMdata, especiallywhere measureme-
nt errorsare present.Ultimately,onesolutionhasto be selectedasthe 'best' or most
feasibleanswerto the problem.To do this we have to add to the problem some
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

informationnot containedin the origh•alequationd=Gm. This extra in•om•tion is


referredto as a •r/or/•u•ormati• andservesto constrainour solutionssoas to satisfy
anyof ottrquantifiedexpectations of the modelparameters.A priorihfformationca•
ta•kesevera•forms. It may representpreviouslyobtained•eophysical,boreholeor
•eological
dataor maysimplybe dictatedby thephysics
of theprobere.Consequently,
constrainedinversion taJ•esmany forms.

5.1 Inversion wigh prior information


We canincorporate
previously
obtained•l•;onnation
aboutthesought
model
parameters
in ourproblem
formulation.
Thisexternalinformation
couldbe in theform
of resultsfrom previousexperiments
or quau'tifedexpectations
dictatedby the physics
of the problem.Generally,theseexternaldatahelpto singleoutß -,•lquesolution
kum
amongall equival•-tones(andit maybenotedthatourinitialdefinition
of whatmay
be termed 'dat•' hasnowbeenrevisedto includeprior parameterestimates).The
solutionprocess
is saidto beconstrained.
The procedure is s•_rnple.
The constraiuing
equations(data) are arranged
to forman expression
of the form
Dm=h

whereD is a matrix (withall theoff-diagonal


elements equalto zero)that operates
on
the modelparameters m to yieldor preserve
the the a priorivaluesof m t/hatare
containedin the vectorh. The equationDm -- h me,•u•that we axeemployinglinear
equalityc0nstr•,,-tsthat areto be satisfied
exaztly.The mathematical
development
is
straightforward.
We wishto bia• rni towardshi .
We simply rninimiz•

• = (• - Gm)•(d- Gra)+ •(Dm - h)•(Dm-

Setting
to zerothederivatives
of• withrespect
to themodelparzmeters
m wefind
that

2GTGm - 2GTd+ 2•2DTDm- 2•2DTh= 0

(G'rG +/?•D•D)rn= GTd+ f;2D•h ' Normalequations


or if D is the identity matrix

(GTG+ •I)rn = (G'rd+ •h) ' - NorraM equations


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

from which we obtain the constrained solution

= (rG + + (5.2)

This is the constrained linear inversion formula. The method is also referred to az the
biasedlinear estimationtec-h•ique.Its main advantageis that it he.•!•l
s singleout _a
unique
splution
ou•oft.hei•nitel¾m•uyplausible.,
solutions
•o a• overdetermined
pro.blem in. th..epresenceof ob.s.ervationM errorsor •mce•.•inti•. However, this
procedure
shouldonlybe usedwhenit i• reasonably justifiedsinceit may produce
undesirable
effectswhenh is -•rea•stic (seeTwomey,1977).

5.1.1 Impl•ment•tiom
The constraintsare implementedby arr•'ging the constr•ug equationsaz rowsin the
originaldat• equationd=Gm. The a.uxilim-y parameter• is chosenby trial and error. f/
is called an undeterminedor Lagrangemultiplier. This solutionprocessmay thus be
referred to as a Lagrangemultiplier method.

5.1.2 Formulati•õ ••,õ •tionz


The equation Dm•h is gener•y of the form

(5.3)

However,ghestructurecan be modi• asrequired.For ex•-mple,i/only one parameter


valueis known,we donot needthe full setlupabove.We simplyrequire

m1

- just a n,•mbet, not.a vector. (5.4)

55
H; on the other hand,we knowthe firs• andfourth parametervaluesfor a fottr-
parameterproblemsay,thefull constra/ninõ equationswill be
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1 h4

but in practiceonlythefirstandfourthrowsof the constr•uiugequations


arereally
necessary.The appropriate equations
arethenappended ontothei•tia/equation
Gm=d. Operationally, weonlyneedto addthe extrarowsof a pt/or/data ontoG and
d, e.g.,forthe caserepresented
by eq.5.4wesimplyaddtherow.[10 ... 0] ontothe
bottomof the G matrix andthe knownvalueof a parmeter [•,o•,] ontothe bottomof
the actual field data d. Where desired,both D and/• are premultipliedby •l (usua/ly
chose•to be lessthan or equa/to 1luSty).The constrained solutionis obtaiuedusing
the proceduresoutlinedin Section4.

5.1.3 S•rnpleApplications
of Constrained
Inversion
Examplesof applications of constrainedinversion
with priordataa•egivenbelowto
facilitatea practicalunderstanding
of all gheconcepts
highlighted
in the preceding
cliscussions.

F.ommple5.1.1: Constrainecl
fittingof a straightll-e t• s•ismicz•[rac'fion
dat•
Suppose that weareinterested in fittinga straightlined•---m•+rr•z•
( ozin collective
formd----Gm (wherem=[m•,rr%] • ) to seismicr•ractionfirstarriva/s
represented by the
dat• pairs({z•,•}, i=l,n). Nowlet usassume that wehaveprior•u•ormation o• reason
to believe thatthefit•edlinemustpassthrough a parti•ar pointwithcoordinates
(==,re), saytheorigin. Wecanconstra•thelea.st squares solution
processto satisfy
thisa pt/or/'infomarion
orass•rap•ion.
Theprocedure
isillustrated
below.

For the straightlineproblem,


wehave•womodelpara.metex•-
theintercept
mxaud
slope
rn•. Suppose thatwewantthelineto passthrough
(z•, t•). Thatis,wehaveone
constraint
(notethatyoumayimpose a n•mberofconstruth-rs
) ontheproblem.The
constrig equation
Drr•h will •herefore
be of the form
[1•,lI•]=[•]
T T ?

Thisequation
is thenappended
onto•heoriginaldataequations
d--Gin•ud thewhole
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

systemsolved
fortheconstrained
leastsquare•
solution(thedesirable
regression
line).
Note that this •ddition•l rowof extraneous
(constraining)
datawouldmeanthat our
original
expressions
forGTG•udGTJ in theunconstrained
inverse
problem
(Example
4.2)will needto bemodified
somewhat
to:
n •zi' 1

(CrG + •2;)= ß ß ee ß ee ß ee ß ee ß ß

augment•g equations
1 z•' 0
•ud

(a•d +•2h)=
augmentingequation

wherewe have assigned


a vmueoi ,,-ity to •.
The constrained
leastsquaxes
solution
forthestraightline.through
(me,
re)is therefore

]]l•tive Exercise.

ofthebest-fittLugli.ne
Usingthedatagivenin T•ble4.1.1,obtainthepaxameters
that passes
throughthepoint(•c----8,

Solutionprocedu• ß
The m•d-taskhereis to determinetheinverseof the 3 x 3 matrix(G•rG+•2/).This
canbedoneusingtheproceduxe outlined
in Section4.3.1.The elementsoI thematrix
of cofactors
of (G•rG+•:I) axe:

57
c ; C12--'-[•12:;
z•O ]-__-Z 2 C13=[ 1 z•
__--
0 [=z•

C22•[• c==-[n
1 Zzj=Zx_nx•
Xc ;
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

•,x 2
1
Xc I Zz
7l zl•
; C32------

Therefore,

2
x• (z•Ex- Ex2)

(x•Zx- Zz2) Ez- nz• nE=• - (Ex)•

Notice that

We then obtain the required determinant as

der (GerG+•I) = [ G'rG+•I [ = n(Y,•.O- •)- •z(O - x•) + 1.(z•x - •,x•)

= 2x•,x-Zz•-r•z•.

so tha• the constrainedlea.stsquaressolutionis given by

We cannowdeterminethe valuesof m that will satidy the constraint(z•--8, t•=14.9).


Usingthe data from Table 4.1.1, we havethat

41.1

58
20)-120--(4x/$4)=-56

L40 8oj
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

There/ore,

77lI

Sotheslopeofthebest-fitting
linepa•sing
through
the.point(•½,t•) is

-87.6: 1.564285.714
-L56 ; ( • = 1/m• = 639.27m/s )

the intercepton the Qme-axisgivenby'

-133.6
-56
= 2.385714286
.

usingthe relation t---m•+m•E. We havethat


Finally, test the solutionfor consistency

2.385714286
+ (1.564285714
x 8) = 14.9

which mea•2sthat our estimatesfor the modelpar;tmetersm safisfTthe co,•.•tra•ts.


These'resultscanbe ved•edusingthe inversion
progr•m.•Cramimv,Linpacko• Svtl•nv.
Notethat C•:,m•n.v
canbeusedto solvethisproblem;'
wehaveto input (•rdd-•//) as
theV-vector
andthematrix(GTi3d-•//)astheA-matrix,i.e.,

n Zzi 1 4 20 1

A-- ••'ß• •• = 120


8 8
0

Thiswillyield
thesolution:
I•h= 2.385714;
rr•= 1.564286;
• '..27'.143pi
!.

59
Example5.1.2 ' Cons•rx•ned
seismic
refract.
ion•q•e-t•m analysis.
(of Example4.4.3)
Considerthe refractionseisinologydelay-timeproblemof Example4.4.3. Let us
constrainthe problemby speci•ng oneof the delay-timeandseeif a ,,n•quesolution
can be obtainedfor all six par•Lmeters
of the inverseproblem.Ass,•rn•ng
that we know
the v•.lueof •, howwouldweformulate
•ud solvethe constra•edproblemforthe six
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

parameters?
Firstly,therequiredconstraining
equation
is sim•la•to eq.5.4 •ud givenby

['%1=
[•-
0000O]
0.433

The structureof the augmented


• a•d d is shownin Fig. 5.1.1.while
the resultsof the
application
of theSV'Dmethod
to thesystem
ofequations
aregivenin Fig.5.1.2.Note
here that/• = 1.

C-matrix plus bottomrow of constraintD plus one constr•,n•,•gdat•,• •r•


1,0,1,0,0,6. 2.323

1,0,0,1,0,6.708 2.543

2.857
1,0,0,0,1,8.485
2.640
0,1,1,0,0,7.616
0,1,0,1,0,7.0 2.529

0,1,0,0,1,7.616 2.553

1,0,0,0,0,0 •-•D

Fiõ.5.1.1Iupu•d•a structure.
No•e•hat/•=1 in th•_.•
example.

17.97105
1.898747
0.6343145

0.1198960817
,--•k4 nowbetterconditioned
1.421898
1.414207

6O
U-transpose
matrix (UT)
-0.3373508 -0.3764591 -0.4749891 -0.4267474 -0.3927506 -0.4271749 -.0036922784
0.3908712 0.3841310 0.3553412 -0.3553166 -0.3503931 -0.3689544 0.4338698
0.0977789 0.1944664 0.2359648 -0.06169433 -0.1618290 -0.2929147 -0.8838264
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

-0.5554522 -0.03372441 0.4192865 0.50200870-0.02007838-0.4823709 0.1715710


-0.3197534. -0.28229870 0.5672691 -0.35033420-0.3229672 0.5177553 -0.034040175
0.4903344 -0.5091351 0.0188053 0.4903353 -0.5091333 0.0188096 0.000004396

V-matx4xdividedby A (V/A =VQI)


-0.00369238 0.4338701 -0.8838334 0.1715780 -0.034040730 0.000004515
-0.00386020 -0.2980839 -1.2835310 -0.044986647 -0.076933742 0.00000761725
-0.00236593 0.0098617 0.0896802 -5.45729400 -0.33143010 0.4903358000

-0.00238176 0.0093580 0.08111537 -5.49373700 -0.2993690 -0.509135700

-0.00279343 -0.003776 -0.1415401 -6.44167400 0.5366624 0.018807005

-0.05521560-0.0088096 0.1686551 0.7883733 0.007619273 0.00000076958

Fig. 5.1.2SVD resultsfor.theconstrained


problem.

The calculated least squaresestimates axe:

6•=0.433,
•=0.346,
7•-----0.391,
7•-----0.433,
73-----0.304,
v-•-----0.250 !

Noticethat all the paxameters


havebeenaccuratelyestimatedthis tlvne.Notice
thatthe singular
valuecorresponding
to thefourthparameter
is nolongerclose•o zero
(0.098960817).
Onequestion
•ha•readilycomes to rni•clis: canflxesoln•ic•x
lm•ams bestabiTm••r
whatmaybe doneff no.priarestqm•,te
c• aneof[the d•b,y-tim,•is a•ailahle?
The aa'swerto this questioncanbe foundin the next section.

5.2 Inversion with Smoothness Meastues.


A veryeffective
wayofinverting
a finitecollection
o{inexactdatais to impose
the
constra•t that the desiredsolutionbe smooth.This measuremay be basedon the
physicsof theproblemoronpurely•eolo•ical
considerations.
Forinstance,consider
tryin•to inverttheseismic
dataforthedelay-times
asin Ex•,r•ple
5.1.2butwithout

61
anyknownestimates of a delay-time.
I• maybewondered whetherthe exactsolution
canbe foundby anymeaasothertha• that discussed in Section5.1 a•d we aJaos•w
how UXhst•ble
the solutionprocesscax•be for problem•wi•h very smallsingulaxvalues.
The simplesta.udcheapest remedyforsuchproblems or • prescription
for indetermln-
acy,or non-uniquenessin inversion
is givenbelow:
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

[ff in doubt SMOOTH[ remedy for non-uuSqueness

Note that the solutionobtainedwith smoothness


constraintsis a •ve es!i'm•te
of the actual modelandis the bestthat onecouldhopefor in the absenceoI rd•ble a
pt/or'/i•fformation.

5.2.1 Pxobl•n formul•fi•


Let us nowexaminehowa problemcaabe formulatedto yielda smoothsolution.
If it is desiredthat the modelpaxxrnetersvary slowlywith position,say,then we may
choose to •i•i•e the di•erencebetweenphysicallyadjacentpax•rnetexs (mx-rr•),

equations
in the formDm•/s, viz:

-1
ß

ße e

i-1 L•pj
T
D
T
rm

whereD is the difference


operator
]cno• hereasthe smooth,•_•__
matrixandDmb
the •m•th•• or F/•AT•E$$ of •he sola½ion
vectorm. No•eth• D ia of rli•nexmion
I xp since
thereare/=(p-t)firstai•erences
between
thep par•rneters
azxdhisof
dimemionI x 1. A detailedtreatmentof suchconstrxlnts
canbe foundin the excellent
textbyTWomey
(1977,
Chap.
6).
If themodelparc.meters
donotvarysmoothly
withposition,
thentheuseofconstrain-
ingequationsoftheform(cf. eq.5.3)

62
I

1.

T
D
T
m
T
k
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

is recommeded.In this case,D is the identity matrix of dimesion p x p aad/i is of


dimez•ionp x 1. It is interesting
to notethat it is thena formoi biasedes!itnmtioll
with
non-i•ormative • Fr/• in•ormafi• The operationeffectivelydampsthe lengthof the
solution(by forcingit intocoulo•ty with it) leadingto a stableinversion procezs.
To gaugethe smoothness of oursolution,
weusea quadratic measure ql(m!givenby

Let us defineaaothermatrix,H=DTD for the sakeof notationalsimplicity.

We state the constrained


problemas:From the incomplete,i•.,nrfflcient
and
inconsistent
fielddata,
fn• among•
allpebblesolutions
withresiduals
q•= [ d--Grn[2 , the smcx)thest
asjudgedby the measm•q2(m}.

Mathematically.
theabove statement
is equivalent
to: minimizeq2------mTHm
traderthe
condition
] d-Grn [• = q•ormoregenerally
[d-Gm[ 2_<qTwhereqwisthe
maxim-m tolerable residuals or mi•t.

Thecouztr•i•ed
problem
requires
thatwemi•imi•.eIId--Gin]12andq2(m)
together,
i.e.,minimize

• = (d-Gm)T(d-Gm)+ g•(m:rDTDm) (5.6c)

5.2.2 Problem solutiun


It is obviousthat we will usethe methodof La•rangemultipliersasbefore,where• is
largelyundet•i•ed (i.e.,0<• <_co). M;-ir-i•.ation
requires
thatforall j

O(dTd-mTGTd-dTGm+m
TGTGm+/•2m
THin)=0
Orn•
so that

(GTG+ ,82H)m= GTd The Normal Equations

63
f•om which we obtain the smoothest solution

m• = (G'r(•+ •H)-•GTd (5.7)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ff D is a first differenceoperatoror simply

,?,,= (c"r + (5.8)

if D = I. The abovesolutions
arebiasedlinearestimators,
sens•ziric•o (Me]u,1994d).
Equation
5.8ispopularly
known
astheDamped
I,emmt
Squares
solution
audismathem-
aticallyequivalentto add•g a positiveconstantbiasto.theeigenvaluesof the design
matrixto improveits condition beforeinversion-
the well-known Marqua•dt(19?0)
method.Let us explaiuthis latter approar.
h in moredeta• •or the sakeof clarity. •
that the unconstrainedleastsqauressolutionis givenin termso• the SVD oœG•as

(4.12)

Todsr•ptheabsolute
values
oftheparameter
estimates,
wesimply
adda smallbias,
to theeigenvalues,
i.e., thediagonal
mat•zxA- • isreplaced
operationally
by the
damped va•iant
A• _
- (A A (5'9)

yieldingthe constra•edinversion
computational
formula

It shouldbe notedthat we havenot appendedany zerosto the data vectord asin the
former
approach.
There
isthusa subtle
difference
between
thetwoschemes
andthey
donotyieldthes•meresults
andweshallseelater;theformermethod
is alsomore
flexible
aswedonotneedto.augment
allthediagonal
elements
ofGrr• to obtaina
reliable solution.

5.2.3GeometricalinterpretationeI inversionwith smooth,•_•measm-•


Thequadraticfactor• andthemes•ures q:andq• aretheessential
ingredients
in the
generation
of smooth modelsasillustrated
in thefollowing
interpretative
aualysis

6•
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

E
i
ß
ß

I
I
I

I
!

I
I
i

I
!

I
I

I
ß

I
I

,%

,%
%

%,%
I
%
I

I'
!

!
I
I
I
I I I
•,
i
,/
I
I

/
I
I
(acl•ptecl
from Twomey,1977,Chap.6) aimedat providinga clearerundersta.ucl•ug
of
the constrainedinversionprocess,and especiallythoseaspectsthat may h•ve practical
implications.Noticethat eq. (5.õc)may be writtenas

•b• (/1-1-'•2q2'
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Themeasure qeis zeroonlywhenm•=me=...=top=0, (seeeqs.5.6aand5.6b)orwhen


m•=rne=...=mpandD isa firstdifference
operator
suchthat Dm=//•0 (eq.5.6a).In
theseidealized conditions,q2is saidto haveaaaabsoluteminivroom a.udfor easeof
discussion,its locationwill be denotedby B in Fig. 5.2.1. Similarly,qxattxin.•a
minimumat someotherpointin functionspaced.enotedby A in thisfigure.If eq. (5.6½)
isminimized,
thesolution
travels
fromB toA as• isvaried
fromaninntely large
valueto zero, say.Sincethe quantitiesqxmu
d q• caaa
be computedfor the solution
obta•edfor muypaa•icular
valueof•3,thetravelpathofthesmooth
solutions
caube
monitored as fi is vaa-ied.Sucha solutiontrajectoryis illustratedin Fig. 5.2.1. The
contoursin this figure.•re the l-cllmensional.'hypersurfaces
of const•utq• andq2.If the
inverseproblemis posedasthe search for the solutionsatisfying
the constrains
andq2= '1,in thisfigure,say,it is clearthat thedesiredpointwherethe qx--qu:and
q:=7 contourstouch,C islocated somedistanceawayfromthe'vnlnima A andB. To
ß

getto thispointfromB requires a suitablechoiceof thequadraticfactor•; onecau


easilyovershootthepointC (thusmissing thedesired solution)
by usingcoarsestepsin
(orunrealistic
valuesof) •. Also,ff wespeci•eda largethresholdmisfit,sayqz-- qo,'we
mayobtainthe undesirable solutionC• whichalsosatisfies
the smoothness constraint.
Thus, a suitablechoiceof/• andan appropriate misfit criteriona•e a •in• quanon for
the constra•ed inversionprocess,especiallywhendexllngwith ill-posedlin• probl_•ms.
In routine inversion,an acceptablesolutionwill have the rnlnlrm•mcombinedmeazuxe
q=(q•+•q•)'However,
since/•isundetermi•.ed,
it is t•ood
practice
tofindall accept.able
solutionsand makea selection
fromamongthemon a basisotherthan a statisticalone
(seealsoMey•,1977, p.'400).

5.2.4 SampleApplicationsof smoothness


ctmstrain•
5.2.4.1 S,•i.--nicRefraction Time-Term analyais
The delay-timedatausedin theprevious
tel:rationseisinology
exercises
havebeen
invertedusingthetwotypesof smoothnessConstraintsrepresented
by equations
5.6a
and 5.6b for vaa'ious
valuesof •3.The typicaldata structurefor inversionwith •-
weightedfirst-diHexence opexztorsis illustratedin Figure 5.2.2. Notice that we have
incorporatedp-2 constraints insteadof p-1 as definedin eq. 5.6•. The ass•rnptionhere
is that the delay parametersaxeslowlyvaxyingspatially mudthe well-deterrnud
slowness parameter is left unconstrained.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

G-m•trix (•ugmentedwith d-vector

i0-2 rowsoœconstzai=t
•D) (•ugmented
with zeros)
1,0,1,0,0,6. 2.323

1,0,0,1,0,6.708 2.543

1,0,0,0,1,8.485 2.857

0,1,1,0,0,7.616 2.õ4

0,1,0,1,0,7.0 2.529

0,1,0,0,1,7.616 2.553

0.01,-0.01,0,0,0,0 • 0o0 •

0.0
0,0.01,-0.01,0,0,0,
T -
0,0,0.01,-0.01,0,0 •D
0,0,0,0,0.01,-0.01 •
o.o

F•g.5.2.2Inputdat• structure
with first-difference
operatorfor •--0.01.

The actu•-Iresultsobtainedwith •t specified


is re-statedbelowfor comparison

I &]--0.433,
&•=•1346,
'7•----0.391,
7==0.434,
'73=0.304,
•-••.250..J
and constitutesour target solution.

The resultsobta/med
onappendinõ
a/%weiõhted
px p identitymatrix (eq.5.6b)ontothe
bottom of •he C,-m•t• are s-•arised in Table 5.2.1 for • range of valuesof •. I• cau
beappreciated
'fromthistablethat theresults
for/•--0.0000055
•greewith thosederived
frominversion
usinõa knownvalueof • •-pa•r•eter (seeeq.5.11•).However,wMlethe
correct
solution
maybeob•a•ued
usinõ
thismethod
forSome
value
of/•,.selecting
•his
solutionon the basisof the computed
residuals
may not be a straightforward
•asksince
•hesolution
withtheleastresiduals
maynotbethecorrect
oneasshown
in Table5.2.1
•ud therem•y be more•hauonesolutionwith the s•e residuals.

67
• Solution SSE• SSE•+h

1.0 .088 -.0016 .065 .091-.070 .344 2.37232e-02 1.67160e-01


0.1 .382 .286 .250 ..291 .128 .279 1.27õ60e-03 5.96765e-03
0.01 .501 .414 .320 .363 .232 .250 2.61112e-07 7.75681e-05
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

0.001 .503 .416 .321 .364 .234 .250 5.91967e-09 7.84D60e-07


0.0001 .502 .416 .321 .364 .234 .250 5.87437e-09 1.36563e-08
0.00001 .481 .393 .343 .386 .256 .250 5.88756e-09 5.96562e-09
0.000006 .441 .355 .382 .425 .295 .250 5.89796e-09 5.92665e-09
0.0000056 .434 .347 .389 .432 .302 .250 5.89807e-09 5.92321e-09
0.0000055 .433 .346 .391 .434 .304 .250 5.87994e-09 5.90422e-09
0.0000054 .428 .341 .395 .438 .308 .250 5.87966e-09 5.90316e-09
0.000005 .419 .332 .405 .448 .318 .250 5.87778e-09 5.89812e-09:
0.000004 •362 .276 .461 .504 .374 .250 5.89932e-09 5.91334e-09
0.000001 -1.72 -1.81 2.55 2.59 2.46 .250 5.89733e-09 5.92285e-09

Table 5.2.1 Variation in smoothsolution•with • for D = I.. The p•rz•eter


estimatesaregivenin col-mu•2 - 7. SSE•is the s-m of squared
errors
corresponding
to theobservational
datawhileSSE•+h is thecombined
residuals
due to the actuM and a trriori data. The acceptableparameterestimatesare
shownin bold print while the solutionwith the rn•n•m•m_residualsis marked by

It would be desirable to be able to know when the correct estimates have been found. Is
fl• any solutionto th•.•choiceprobl_.•7 It is dear fromT•ble 5.2.1•hat the
slowness
(1/•) pa•meter
iswelldetermined
inmost
cases.
I• m•ythus
beus• •o
allowsucha parameterto "flo•t freely"whilethe variationsbetweenthe rest are
smoothedBy a •-2 first-•erence operator.The resultsof suchan approachare
s•marisedin Table5.2.2.It isobvious
fromthistablethatvery•oodestimates
ofthe
truepar2•eterscanBeobtained
forthisdatasetfor 0.0000110
< • < 0.0000114
usin•
thistechnique.
Notice
•/tom Table5.2.2th• the;raproved
behaviour
o[the
residualaof the combinedactualand a • data set may be usedto selecta• optimal
seto[ valuesfor the sou•:,ht
pa•m•: in •hlscase,the solutionwith the mi,•im,m
•lue of$b•a+aisin •c•d withthatderi•ed
usin•a known
•alueofa delay
pad-meter.
ASin theprevious
comparable
case,thestabilityof the-solution
process
over
a wideran•e o:[• is apparent.

68
Solution SSE• SSEd+;•

1.0 .560 .516 ß516 .528 .490 .211 5.78974e-03 9.36119e-03


0.1 .459 .373 .373 .415 .288 .249 4.91896e-06 2.57990e-04
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

0.01 .455 .368 .368 .411 .281 .249 6.43308e-09 2.63539e-06


0.001 .455 .368 .368 .411 .281 .250 5.89409e-09 3.21941e-08
0.0001 .455 .368 .369 .412 .282 .250 5.90245e-09 6.16546e-09
0.00005 .454 .367 .370 .413 .283 .250 5.89637e-09 5.96213e-09
0.000025 .451 .364 .373 .416 .286 .250 5.90745e-09 5.92394+09
0.0000114 .433 .346 .390 .433 .303 .250 5.89733e-09 5.90100e-09
0.0000111 ..433 .346 .391 .434 .304 .250 5.87488e-09 5.87837e-09 •--
0.00001105 .432 .345 .391 .434 .304 .250 5.8798,%-09 5.88330e-09
0.0000110 .432 .346 .391 .434 .304 .250 5.89847e-09 5.90019e-09
0.000010 .429 .342 .395 .438 .308 .250 5.90711e-09 5.91002e-09
0.00000• .346 .259 .478 .521 .391- .250 5.91007e-09 5.91193e-09
.

0.000.001 -2.29 -2.38 3.11 3.16 3.03 .250 5.94508e-09 5.97523e-09


0.0000001-180.3 -180.4 181.1 181.1 181.'0 .250 1.58823e-07 1.60129e-07

Table 5.2.2Vacation in.smoothsolutions wighdampingfactorusinga 1•2


first-difference
operator.The velocityparameter-was unconstrained in these
calculationsand the solutionwith the smallestresidualsis mazkedby az•azrow.

An ex•rnlnation of the solutionsderived.for va_•iousv•lues of/• with the conventional


Marquardt-type'dampingin which/3 is addeddirectly to the singularvaluesof the
problem(eq. 5.10) failed•o reproducethe correctvaluesof the parametersoverthe
range of vMuesof/3 usedin the tests.In fact the bestestimates•rom this tech,,•que
were

ß i

i i

It would thus appear that we may not recoverthe true solution via this avenue. All the
abovecalculations
wereperformedusinõthe computerpro•r•-• SVD• (listedin
AppendixA).

69
6. ERROR AN•YSIS IN LINEAR INVE1LSION
One questionthat is fundamentalto geophysical
data analysisis, how representativeof
the real geophysical
sy•em is ouzreconstructed
lemztsquaresmodelor how accurateis
ouz solutionto the given probl•- ? Recall that our initial a.ss•lrnption
was that the
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

experimental
data contah•errors(whichis whywecannotfit themexactly).Onemay
thereforebe interested{n how the exp•mental errorstranslateinto er•rs in the model
estimates.The answersobvioudycomefmom•t•tistics. Inversetheory not only provides
us with esthn•te of the relevantparametersbut Msofurnishesa plethoraof related
infom•tion th• enableus to g•ugethe "goodness" of the leastsquaressolutiongo•he
inverseproblem.Someof such"•mdliary parameters"are described following •
discussion
of how to incorporate•vail•ble observationalerrorsdirectlyin the inversion
process.

6.1 Elaborate Treatment of Observ•tion,al Enm• in Inversioa


It may be appreciatedfromourtreatmentof '•versionsofar that the process involves
Ku•g a solutionthat m•,•m•zesa suitablychosen quantity-the squareddistauce
betweenour solutionandthe givendata. Our desireshouldbe that the solutionbe
both nvmericallyand statisticallystable.Statisticalstability is important becauseof
the a•ffereninguncertaintiesassociatedwith our observations. If observationalerrors
are available,we canincorporatethem directlyin the problemformulationto obt•,• a
more acceptbleweightedsolution.Thus, ass,,m•uõ that the n st•nd•d data errors,
are Gaussianwith zero mean and are statistically independent,we can define a• n x
diagonalweightingmatrix, IV as

We can then re-statethe constrainedinversionproblemoœSection5.2, for example,as:


minimize

q= +

wheretheweighting
(orsc•l;ug)
ofeachdat•mbyitsassociated
observational
error
ensures
that tindueimportance
is not givento poorlyestimated(i.e., noisy)data.This
is a somewhatmore robust formulation and the solutionis identica•y,

7o
m, = ((WG)TW'G
+/•2H)-•(W'G)w]•rd

or simply
= +
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

if D = I. Notice that/? is just a singlenl•mberin the aboveformulation.If we axe


interested
in retai•g specified
valuesof thepax•meters,
i.e.,biasedestimation
proper
(Section
5.1)',a recommended
eleõa•tapproach
is to definea diagonal
matrixof
undeterminedmultipliers• and re-statethe problemas minimize

q = (Wd-WGm)T(Wd-WGm)+•(Drn-h)TfiT(Drn-h).

The solution to this problem is

•l,b----((WG)TW{•
"[-DT.BD)
- •'[(WG)TWd
.4-B.DTh] (6.2)

whereB = •Tfi. The diagonal


elementsof fi azeassigned
a constant
positivevaluefor
thoseparameters spedfled
a Prioriandtherest.aresetto nought.
Thesomewhat
identica/role oœthe matrices .W and fi is obviousaud we shall ex•e the statistical
implicationslater. Noticethat if'we defineanothermatrix E- w'rw, theneq. 6.2
may be expressedin anotherenlighte-•,•õform

ra•= (GTEG+ D•BD)-•[GTEd+ (6.3)

whichin every respectis equivalentto the so-calledBayesiauEstimatar(seee.g.,


Jacksonand Matsu'uxa,1985;Duijndam,1988).

6.2 Asse•si• the Quality oI a Solution


6.2.1

This is %hecommonlyusedmodel acceptancecrigerion.Ass•,m•g that our data d• are


normally
distributed
abouttheirexpected
values•ud withlmo• uncertainties
cr•( the
experimental
errors),we canazsess
thefit beween'
the observed
andpredi•eddataby
calculating
thestatistical
pax•eterq defined
by
n (d•._ Giirni)
•'
i=1
-

71
or simply
IIwd WGmII for a weighted solution.

For n independentobservationsamdp independent parameters, q is distributedasX2


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(Chi-square)wi•h (n-p)degreesof freedom.


In geophysical
inversion,we rejector accepta solution•o the problembeingconsidered
basedonthe valueof q. Theexpected valueof q is n (fronChi-square statistics)
but
inpractice
amodel
withn-p< q<•n+v•2n
I isacceptable.
However,
if q<<n,the
model is said to over-fit the observeddata and if q >> n, •he modelunder-fitsthe
data. Over-fittingmay leadto solutions
that containartifactsof the computer!
If experimental
errorsarenotavailable,
anunbiased
estimate
ofoa is givenby

cr
2_
-- (dTd-mrGrrGm)
n-p _
= s,,m
ofsquares
n-p ofresiduals
'

Anothermeasureof goochaess
of a solutionis the RootMean Square(rma) errorgiven
by
= In•
rms (d?*'-
Gffm./)
2
or simply
IIWd 11
= for a weightedsolution.
i---1

Obviouslyoneshouldnot'aimfor solutions with 1.0 << ms >> 1.0 Note that the
n,,,,•berof degreesof freedomfor a constrained
solutionwith I independent conatr•;,,t
equations
is {n - iv+ l} andweshouldsubstitute
thisqu•utityfor n- p in the above
formulations where appropriate.

6.2.2 Parameter Re•lution Matrix


For a linearsystem,we caaassess
the qualityof the modelderivedfrom a givendata
setbyex•m;,,i,,g
thepar:-•eter
resolution
matrix(jackson,19E).
Wewillshow
howit
is derivedfor the unbiasedleastsquares
solution.andthen applythe samestrategyto
ß

the caseswherewe haveconStrai,•ed the solutionprocess.

Caze 1: Unconstrained solution


Recallthatthetmconstrai•edleastsqu•res
solution
is m = {(G'rG-')
- •G•d whichin
termsof SVD is simplyrn= {VA-•[fr}d. Letusdenotethequaatityin braces
asH.

72
The matrix H is the generalized
inverseusedin the estimationand post-multiplyingit
by the applicable
design
matrix(or datakernel)givestheresolution
matrix R, i.e.,

R = HG = (GrG)- •G•. G = I (6.3a,)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

In terms of the SVD of G,

{VA-

R is of dimensionpxr (wherer is then-mberof non-zero


eigenvaluesof theproblem).
If R = I (identitymatrix),i.e.,r=p, theneachmodelParameter
is --iquelydeterm•-ed
The deviationof therowsof R fromthoseof the identitym•txix, I, is generally
ass-reedto be a measureof the l•ck of resolutionfor the correspona•-g.
model
parameters.

Caze 2 ßThe Marquardtdampedsolution


For the dampedleastsquares
solution
givenby eq. 5.10,wehavethat

a=(CrG
+•Z)-'C
r-G=Z+(••G) (6.3b)

which, in termsof the SVD of G, is simply

V•2V'r
a = VA• 21fr-VA:V• =

obtaiu• by simplya•J.d•.a constantbias


It is obviousthat the dampedsolution t:o

theeigenv•lues
ofa problem
does
nothavea perfe•resolution.

Case 3: In--on with a prior'/da•


by eq. 5.2, may be
The constraind inversionformularepresented written as

m, ' {(GrG+ •2D•D)-XG•}ßd + {(GrrG


+ •2DTD)-X•D?}'•h

andhencetheparameter
resolution
matrixmaybecalculated
as

a = (CrG+ •2DrD)-xcr-
G+ (•rG + •2DrD)-X•Dr'•D

73
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

T
2
T
co1.1 col.2 co1.3 col.4 col.5 col.6

1.000000 4.10187E-08 -2.0286F,-09 -7.94157E-09 -4.34869F•09 -6.02772E-09

4.10187E-08 1.000000 -1.6353E-08 - 5.59144E-09 -9.41843F•09 -9.81052E-08


-2.0286E-09 -1.6353,•,-08 1.000000 - 3.89008F•08 - 1.34107F•07 1.57051F•08
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

-7.9416E-09 -5.5914E-09 -3.8901E-08 1.000000 -9.6212F.•08 4.16797E-09

-4.3487E-09 - 9.4184E-09 - 1.3411E-07 -9.6212E-08 1.000000 1.50810F.-08

-6.0277E-09 -9.8105E-08 1.57051F_•08 4.16797E-09 1.50810E-08 .1.000000

Table 6.2.2.2. Resolutionmatrix, R for a constrainedsix-parameterproblem.


Notice in this casethat R is Msodelta-like suggestingi•er/ectresolution.

It is obviousthat eventhe unconstrainedproblem which we knownot to be determined


by the practicaldatahasa perfectresolution.
The calculated
R for an opt•mMdamped
solution
(i.e.,augmented
singular
values)
f6rthe'seismic
delay-time
problem
isgiven
i•
Table 6.2.2.3for the sakeof completeness.
Noticein this casethat R is not aaxidentity

co1.1 col.2 col.3 col.4 col.5 col.6


7.99918E-01 -2.00064E-01 1.99958E-01 1.99957•01.1.99939F_,-01 1.65436E-05
-2.00064E-01 7.99907E-01 1.99955F.,-01 1.99954F.,-01 1.99935F_,-01 1.78145E-05
1.99958E-01 1.99955E-01 7.99956E-01-2.00016F_,-01-2.00027F,-01 9.78324E-06
1.99957F•01 1.99954E-01 -2.00016E-01 7.99955E-01-2.00027•01 9.94825F_,-06
1.99939F•01 1.99935E-01 -2.0002TF.,-01 -2.00027•01 7.99929E-01 1.44205F_,-05
1.65436F•05 1.78145E-05 9.78324F.,-06 9.94825E-06 1.44205F,-05 9.99994•01

Table 6.2.2.3Resolutionmatrix for an optimald•mped solutionfor •= 0.00001.


Noticethat thediagonalelements corresponding to thefirstfiveparameters have
Valuesof about0.8whilethesixthpar•-eter hasa diagonal eleme•xt
of 1.0so
that R • I. Thevelocitypar•,•eteris thusthe onlyparameter that mayhave
beenwell determinedin thisinversion.

matrix which is in •ccord with ou• derivation in Section6.2.2. The information provided
by R hereis thattheslowness
(andby implication
thevelocity)
parameter
mayhave
beenwell deterr•ned. As we sawfor •he unconstr•-ed inve•on problem,a delta-llke

75
/• doesnot •u•rantee a mea•in• solution.A cautioususeof/• would thus seem
appropriatewhen deaJiu•with practica•da•a. For instance,it may be s•id tha•
provirileõthat the •eneralizedinverseusedin the c•lculationsexists,then the true
solution
maybe foundif/• - I;'a lackofperfectresolution
wouldtherefore
suõ•est
that
•he true solution may not be found.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

6.3 Erro•/Bo•• on the ParameterEstimates


An importantaspectof geophysical dataanMysis(or interpretation)is the
determ•uationof bounds(or confidence
limits) on the variousmodelparametersthat
are consistentwith the experimental data and their associatederrors.

6.3.1 Parameter Covariance mat•c

The simplestform of model error estimationis the •eger•ngion of the limits of •he
parametersfrom the Covari•ce matrix, Corem).The Covariaucen•trix dependson
the covarianceoœthe experimentMerrorsaud'the way in whichwe map the data errors
into parametererrors(Menke,1984).For illustration,let us expressthe estimatedleast
squ•res solution •

where œ is the •ener•d•.edinverseusedin-the.inversion.The aboveexpressionshows


that m•ø*is a lineartr•usformation
of d. The mathematic•expectation
(E) valueof
m is

E( m•') = S(I•) = œE(•)

If theexperimenta•
dataazeuncorrelated
az•dof equ•lvazi•ce o•, •hen(by L•w of
propagationof errors)

c,,,,.(,,,.,')= ß

C,,.,,(,,.,,'"')
= {(G'"G)-"G""}
[,:,"d = -" (6.4•)

since{(G'rG)-

76
For the Marquardt-t.vpedampedleas•squaressolution,

= (CrG + =

and the covazi•-•cematrix is givenby


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Co•(ra"•) = (G•G+ •I)- tG=[oal]G(G•G


+ •I) - •

= o•(G'rG+ •I)-•GrrG(•G +/•1')-• . (6.4b)

Usinga similarargument,
the covariance
matrixfor the solution
incorporating
information is identically

Cov(m
'0')= (G•G+ $•DVD)-,G,r[•al]C(G•rG+$=DTD)-

+ (•rG+$•D=D)- t$D=LS'I]
- •$D (G•G+$=DTD)-•

Coy(to'")= (GTG+$•DVD)
- •{•aGTG+ DTD}(GTG+$=D=D)
- '.

Note, however,that ff the-elaborate


treatmentof data errorsdescribed
in Section6.1
waseffected(seeeq. 6.2 & 6.3), then

Co,,(m'")= (GTEG+ DTBD)-'(WG)T{E[a2/]}WG(GTEG.+DTBD)- •


+ (GTEG+ DTBD)-•($D)T {B[B]-'} $D (GTEG+ DTBD)-•
or simply,

Co•(m
'ø')= (GTEG+ DTBD)-Z{GTEG
+ DTBD}(GTEG+ D•BD)-•

= (GTEG+ DTBD)-• (6.44)

wherethe symbols areaspreviouslydefined


aadthe covazi•cematricesof the actu•
anda pt/or/data in thisstandardized
frameworkareidentitymatrices.

Havingderivedworking
expressions
forthecovari•uce
matrices,
let ustry a•d see
whattheyrepresent.
Coy(m)
isa parameter-by-p•r,,r•eter
matrixwhose
•a diagonal
elementis the statistical
variance
of the•i par,,,•eterm•, andwhoseoff-diagonal

77
elements, the covariances,indicate the correlationsbetween the model parameters.
Large
off-diagonal
elements
Coviimenu
thatthei• andj• model
parameters
are
kigBlycorrelated.
Thesquare
rootsof thediagonal
elements
of Cot(m)aregenerally
referred to as the standarddeviationsof the least squaresparameterestimates and may
be used to estimate the boundsof the model parameters.Notice that eq. 6.4c reducesto
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

eq.6.4awhen/•=0 butyieldssmallervalueswhenfl • 0. NoteMsothat the variances


and coyaria,aces
of parameters constrainedto be equalto fixedvaluesin bi•ed
estimationare effectivelyzerowhile the correspondingquantitiesfor the free (i.e.,
unconstrained)parameters in the problemis reduced.

6.3.2 Extreme parametersets- F_a-tremal


inv•on.
We mayelectto determiue
a solution
withthema•ml•m tolerable
S11Trl
Ofsquared
resiclua/s.One method o{ extremal inversionis the Most Squat•
(1976)in whicha v•lueiscletermiuecl
foreachpaz•meter
whichismaadml,m
(or
miT•im•m)
under
theconstraint
thatthemisfit
oftheobserved
equal
tosomespecified
vMue.
Theessenti
alt•eiture
oIthemethod
isthatit produces
a
cl•s of models which d_•rl%es
a zone witbiu which the true solution,if it exists, may be
{ound.Themathematical
Iormttlatio
n ofthisextremal
inversion
problem
is
straightforward.
We statethe problemasIollows:

Givena• optirnxlleastsquares
solutionto a• inverseprobl,•n•m with residualsqr.s,
find.(onaccount
of theobservational
uncertainies)
othersolutions
whichfit the da•
to a spe•F;• threshold
residualqT;or equivMently,
extzmni• thelinearobjective
fimcfion mTb under the canstrainl

] d-Gm] 2=

wherethefunction • isa vector. ofzeros withthek• element(toberna•mi•.ed)


set
equalto unity;i.e.,/•T= [0,...,0,1
i, 0,...,0]andm isthevectorofmodel
parameters.
This searchfor extremesolutionscan be effectedsimply by mi,•irni•ingthe function

)- (6.6)

wherewehaveintroduced
the Lagrange
multiplier1/2/•. At the extremum,
we have

78
Or

GTGm = G•d - pb

•rom which we obtaiu the most squ•ressolution


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Thus the v•lue of • = 0 correspondsto the unconstrainedleastsquaressolution.


Howevex,since'eq. 6.8 must satis• eq.6.5, we have that

which simplifies to

so that

whereqLs-- (dTd- dTG[G•r•-•-


•G•rd)is thes,,rnofsquared
residuals
of the optimal
leastsquaressolution.Sh•cethere a•e two solutionsfor/• for each.modelpa•-eter,
therewill be 2p solutions
(forthe p par•r-eters).
whenevex q• > qLs.Note that setting
•ll theelements
ofthep•r•eterprojection
vector
bto.•,-•ty(i.e.,b= [1,1,..,1]T) will
yieldtwo (i.e., plus •nd minus)solutions
that may be interpretedaz the uppera•d
lower solution envelopesof our least squ•ressolution.If the errorson the data are
assumed,,•vaxiaut •ud uncorrelated,
it is expectedthat qr will ]•avea valuedoseto
the w,r•ber of the data, i.e., qr = n.
We c•u comparedirectlythe most squaresaud leastsqauressolutions.Equation õ.8
can be written as

79
=roDS qT--qLS
••11/•
:k•=(GrG)_(G'rG)-•b (6.10)

wherernLSis the leastsquaressolution.The mostsquazes


solutionenvelopemay thus
beinterpreted
astheco•dencel•rn•tsofthe leastsquares
solution.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

As we sawin Section4 (Ex. 4.4.3),the unconstr•ined


leastsquares
solutionprocess
is
unstablein certain situations.The sameproblemmay plaguethe mostsquaresmethod
(Eq.6.10)if thematrixGTGisill-conditioned.
Meju(199A<1)
su•eststhat a bound
couldbe'placedon thesize of thesolution•asin the conventional
inversionemploying
smoothnessconstraints.The constrainedproblemis de•ued simplyas:

Given amoptimal leastsquares


solutionto aminverseproblem,m with resid•
.

•;-d (onaccountof theobservational


un•i,•ties) the smoo• soluUotm (asgauged
by the measu• rnTDT•) whichit the data,to a •ed thrmhold_•_'_dual
qT'

Note in this casethat there a•e two types of data under consideration:
the actual
experimental
data andthe extraneous
con•%rainin•
data.The.threshol&
residualis
therefore'conveniently defined as

q• = [ d-Gini • +/•m•DrDm (6.11)

and the objective function is stated as:

(6.12)

so that

• (mTb+
I {mTGTGm-2mTGT•
or

[G•G + ,•2Dr.D]m= G'q- •& (6.13)

from whichwe obtainthe dam• mos• squax• solution

= [GG + (6.14)

8O
whereH = DrD. Now, equation6.14 mus%
satisfyeq. (6.11). Therefore,

q•r= did- dTGm- mZ•rd + ra•:OTGm


+ /•raTHra
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

= dVd_ dZG(G•G+•H)- '(G•d-pb)- (d•G_pbZ)(GrG+•H•-•G•d


+ (dZG-•bZ)(G•G+•I•-• {G•G+•H}(G•G+•2H)-, (GVd_•b)

which simpl•es to

did - dTG(GTG+•
2H)- 'GTd+ ;•bT(GTG+/•2H)
- •b

so that

qT-
qœ$)•/2 (•.•5)

whereqrs--dTd-drG(G'rG+t
•2H)-•G•rd''If D-- I, thenH = I. Equation(6.14)is a
stable inversionformula and the operationis equivaleatto add•-g • positive constaat
biasto the mafmdiagonal
of the matrixGTG.$ m•y be chosen
to be a smalln-mber
muchlessthmu•ty ( • 1) mudthe expected
va/ueof qv is n-l, wherethere are l
co•straiuts in the problem.

Note that we can alsoformulate the problemto preserveamya pr/ori paxaxneter


estimates as in Section5.1. Ia this casethe construed most squarem
solution is

,•.. = [G•G + •zr].-•[G•d + W•,- •] (6.10)

whereH = DrD (d. Eq. 5.2) a•d mustsatisfythe condition:

q'r= Id-GrnI =+ =

= dTd+ l•2h•h_
(dTG+$2hTD)m_
m•(GTd+•2DTh)
+ mT(GTG+fi2DTD)m.

Usingeq.(6.16)in theplaceoœ
m in eq.(6.17),wehavethat

81
qT= dTd+ • =hTh--( d•G + $'•hTD)(GrrG+$aH)
-'•(GCrd+$aDTh)
+/•bT(GTG+•aH)- •/•b

giving
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

qT
--qL$)•12 (6.18)

where
qLs:dTd+ • ahTh
- (dTG+i•hTD)(GrrG+i•=H)
- •(G'Td+$aDTA)
ß
The thresholdresidualmay be set to a valuecloseto n-l for a problemwith
constraints.

6.4 E•_•mpleof inversion


anddetail• erruranalTsis:
a recommended
p•ce
As an illustrationof a goodinversionpractice,let us a•xlysea simplelineaxproblem
(seealso,Meju 1994c).

Problem'
Given
thefonowing datapas (J•-•on,1976)
fora straight-liae
inverse problem:

-1.0o0000 -1.124600

-8.000000E-01 7.080000E-02

-6.000000E-01 -9.942000E-01

-4.000000F•01 -7.038000E-01

-2.000000E-01 9.637000E-01

O.000000E+00 5.810000E-02

2.000000E-01 -7.820000•02
4.000000E-01 -1.069000E-01

6.000000E-01 -9.231000F_•01

8.000000E-01 -7.819000F.•01
1.000000 -4.250000F.•02

deterrnlaethe leastsolutionfor the slopeaudintercept,computethe residuals,


the resolutionand•ovariaacematricesand the bounai•gmodels.

Solution:
Therearetwomodelpaxameters
in thisproblem:
m• (intercept)a•tdm•
(slope).
Using
theprocedure
outlined
forfittingstraight
linesto z,•/datapairs,we
will obtainthefollowing
unconstrained
lea•tsquares
estimates:
[..m•
=-3.32963õE-01,
rna= 1.074954E-01
!.

82
The computedd•t• misfit,q= I d-(•m I • -- 3.898074.

The resolution information is simply:


[1.000000-1.594859E-15],--row•
of
[-1.594859E-151.000000],--row•
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

To calculateC'ov(m),wewill ass•rnethat the experimentalerrorsin the data


axeG•ussia•, statisticallyindependent,of zeromean,and oœ•m•t variance.
The resultin• covaria•ce•formation is simply:
[9.090909F_,02
-3.624680E-16]+--rowz
of Cobra)
[-1.449872E-162.272727E-01]+--row=

The-squarerootsoœthe diagonalelements
of C•(m) are
the sta.ud•rd'deviationsof the estimates,viz:

!c',,h=
::t:
3.015113E-01,
-,,•= :!::
4.767313E-01•.
The most-squares extreme parameter sets are:
(a) plussolutions
(• positive)
[4.705472E-01
1.074954E-01],--m
t extre•.ed.
using
b----[1,0]
T
[-3.329636E-0[ 1.377958]•-m=
extremized
using
b= [0,].IT
so]io
[ -1.136474 1.'074954E-01]•--m•'
extremized
using
I, ----[1,0]
T
'[-3.329636F•0•-•.1629•7]•-• •x••,.•i u•• • = [0,1] T

The most-squaxes
solutionenvelopesaxe:
(a) plus solution
[9.653091E-02 1.181232J,--upper
envelope:
m extrem•.ed
using/•--
(b) rn•r,us solution
[-7.624581E-01-9.õ62411E-01],•lower
envelope

The threshold
residual,qTwassetto 11 (i.e.,•) in thesecalmtlations.
Noticethat the
rangeofparameters
provided
bythemost-squares
method
forthed•taare•reaterthau
that indicatedby the standarddeviations
ß
of thelea•t squares
estimates.
All the above
resultswereobtained usingtheinversion
pro,ramSVDIN• (whichincorporates the
subrøutine'MOSTSQ)
and•re in gooda•reement
withthosegiven
in Jackson(1976).
This Concludes
our discussion
of linear inversiontheory and practice.

83
7. SOLVING NON-LIIqEA.R INVERSE PII.OBLEMS
7.1 Charactuation of nonlin•ty
In mostof the interestinginverseproblemsin geophysics
the data •re relatedto the
modelparameters
ina nonl•neax
way(i.e.,notintheexplicit
formd=Gm).Recall
•hat
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

eventhe simplere•ractiondelay-time
problemex•4ued underChapters 4 and5 is
actuallynon-lineax
sincethetravelpathsof seismic
wavesthrougha layeredmedil•m
areproportionalto thepathlengthin thelayerbutinversely
proportional
to the
velocity, i.e.,

where• isthetotaltravelt•meforthe•'• ray,œ•isthei• raypathinthej•alayer


and• isthevelocityinthe• layer. Themodel parameters•reinversely,
rathe•t'hs•
]•nearly,
related
tothetravel
timedata;
andwehadtousetheslowness
(c-- 1•)
of v asthe parameter
of thelinearproblems
9f theprevious
sections.
It
remarkedhere that the way that a givenproblemis posedmay some•m_es determ;,,e
whether it is in effectline,• or not. For ;,•a•ce, i• we •re interestedin a gravity or
•etic modelwherethemodelparam• aretheanomalous den•tyof susceptibil-
ity in cellsof fixedposition,
thensucha problem
m•y be considered
asb•g ef[ec•vely
1in•. Another set of problemsthat may be easilymanipulatedaxethosein which the
forwardtheoryinvolves
simpleexponential
fxmctions.
For instance,in caxbondating
(using
radioactive
decay
dataforCaxbon-14),
theforward
theorystates
thatthe
fractionf of an originM•mountof Carbon-14
rem•i,•i•gaft• time•'ha• elapsed
is
given by

f = 1.Oe
mt (7.2)

wherem is the ,rolmown parameter. Now,fromexperimental andtheoretical studies,it


is kno• that a line• relationship
existsbetween• andlogjr. Takingthe logarithmof
both sidesof Eq. (7.2), we havethat

d = logf = rnt (7.3)

whichdescribes
theequation
of a straightline.The desired
parameter
maythenbe

8•
estimatedusing the 1/ne• methodsof the previoussections.On a d•erent level,
considerthe casewherewe areinterestedin investigating
the resistivitystructureof a
2-layerE•'th usingthe popularW•er electrodeConfiguration illustratedbelow.

current potential • Point electrodes


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Ground sudace
T

t Pl top layer (overburden)

P2 ' substrat-m(basement)

Wenuer electrodea•ray.

The forwardtheorystatesthat the apparentresistivitypoobservable


on the surface
of the abovehypotheticalEax•h-model
is siren by:

(v.4)
= + ] =

is the reflection coe/ficientat the interfacebetweenthe top layer and the basement.
An examinationof equations
(7.1)-(7.4)will revealthat we c,mnotpedorma simple
lineadzing
par;tmeterization
fortheresistivity
inverse
ß
problem.
In order
words,
it is
verydi/Scultto put thisexpression
in a formresembling
8----/7m
aswedidin the
seismicand radioactivedec•y examples.To illustratethis point,let usre-d_,•-eour
modelparameterandou• data kernelas

and

4 2

We canthen re-ex'press
the data-model
psx•tmeter
relationship
as

85
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
paxazneterestimationmethodsdeveloped for lineaxproblern.•
•ud to find appro•mate
me•u•nghfisolutions to the problems.The firstpa,-tof the abovestrategyis c•lled
lin•ation while the secondpart is termedmodelidentificationand appra/sah
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

7.2 A GeneralStrategyfor HahrilingNon-linearProbl•mn


Gausssuggested thatnon-linearproblems canbesolvedin successive
•ppro•mations
usingthe linearleastsquares
method.Thisinvolves
the conversion
the non-l•ear
probleminto •n approx/matelinearformby expanding
thefunctional
f(rn) in Taylor
seriesaboutan initial guessof the what thevaluesthe modelparameters
mightbe.
This is the standardstrategyadoptedin geophysical inversion.

7.2.1 The initial model and lin•ation


Most non-linearinversionschemesrequireas a staxt•g point, the provisionof some
approximate
(aSSlimed
or inferred)valuesof thedesired
modelparxr•eters,
whichwe
shaJl
simply
callmø.rnøisvar/ously
referred
toasthefxrst
guess,
l•tialmod
elor
start/rig
model).Th/sstart•gmodelmaybebased onmw/or/•or•tion (e.g.,
ßpreviousresults)ormaybesimplyanintelligen
t guess
andoftenspellsthe difference
betweensuccess
andfailurein findinga me•ugf• solution,or'betweenrapid a•d slow
convergence
to thesolution.Fromtheforwardtheorythe theoretical
responses
corresponding
to møfor the p modelparameters
are ß

or simply
do = (7.7)

Next,assuage
that fire) islinearaxound
møsuchthata sma•perturbation
ofthemodel
responses
about rnøcanbeexpressed usingTaylor'sthemeraas

ß
f •(mø)cgf•
8rnt
+ Orn• cgf•
6rn:
+ Ornat+ c9r•
Of•
5ms+ "'+ t9m•,
c3f•
•mv+higher
order
term•
i.e.,thefunction
d=fim)isexp•uded
about
thepoint(rn•ø,...,.rnø•)
in p•rameter
space.
Writingthe•boveTaylorseries
expamion
in short-hand
form,wehavethat

87
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
7.3 Unconstrai•ed non-linear inve•on
7.3.1 Problem formulation

We are now in a positionto formulateour inverseproblem.To reiterate,in thl. case,


we will be searching
for corrections
or perturbations
ß
to our initial model Sincewe
typically collectmoredata than therearemodelparameters, the problemis overdeter-
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

mir•ed.As before,the leastsquaresmethodwill be adoptedin suchmi•imi•.ation


problemsbecause of its mathematical
robustness whende;di•[ with baad-limlted•noisy
data. Let e be the residualerrorassociatedwith ouxpredictions(sinceour modelwill
not ii• the field dataexactly).We wishto rn{n{rn{ze
the objectiveirmaion

q= • = (,•- f•m•)• (,•- f•m•).


ß
(z.x:)

Usin• the resultsfromthe previous


section,
we canre-writethe problemas
minimize

q = eta = (!t-Az)T(!t-Az). (7.13)

7.3.2 Probl•n solution: The Gauss-Newton method.


As in the linearcasemi•irni•.ationis effected
by settingto zerothe derivatives
oœq
with respectto eachof the desiredparnmeters perturbations z•, i.e•,

=o
Oa:•

giving
-A•v - •='A+ AxAz + :,:='ATA
=0
Or

2.,4.
T• -- 2.,47!t= 0

from which we obta• ouxleastsquaressolutionfor the pax•tmeterpextttrbations

•:= (ATA)-'.4x• (7.•)

Thisperturbation
(z = 8rn)is•henapplied
toouxstarting
model
møtoyielda better
estimate of the solutionto our problem:

m•= mø + z (7.15)

89
However,thenewmodel m• maynot[it ourdataadequatelysuch
thatwemayneedto
repeattheprocedure
using
m• asthenewstartingmodel. Thesuccessiveapplication

oftkisprocedure
isdescribed
asunconstralned
iterative
leastsquares
fi•ting(orthe
Gauss-Newton
Method) . The iterativeformulawouldthen be
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

m•+z= m•+ (A:rA)-


•n:ry (7.16)
wherethe Jacobianmatrix A is evaluatedat m}. Let usnow consider
an idea•ed
problem to illustrate this method of iterative solution.

7.3.3 Example of iterative applicationof Newton'smethyl


Fora single
parameter
problem,
given
anequation
oftheformf(m)= m4.+m-1
= O,
find the solutionto this equationnea•the regionm = 1 and correctto two place•of
decimM.

Solution procedure:
If we werenot told to searchfor the solutionin the regionm = 1, it is instructiveto
tabulatea•d plotthefunction
graphically
soasto knowapproximately
(or guess)
where
the solutionmight be. A table of the.function.wouldlook.like

m -2 -1 0 1 2 •---x-axis

f(m) 17 1 -1 -1 13 : y-axis

Table 7.3.3.1Tabulatedvaluesof f(m)=0.

A.ux-y plot.wouldshowthat the curvecutsthe j•(m)=0 (i.e. y-) axisbetweeax


m=l aad
m=2 aad alsobetwee•0 md-1 on the m- (or x-) axis.
Now,ass,•ming
ß
that weknowthat thereis a solution
nearm=mø,expanding
•f(m)about
møa• in eq. (7.8) gives

.f(m)= .f(mø+6m)= .f(mø)+

O.fcm
ø)
• f(mø)
+ 'ore-

9o
Now, from the problem statement

.f(•o)+ '-b"•--6rn
= .f(m
ø)+ y(a)&n
=0
so that
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

=_ f(mø)
f'(mø)

whichis the solutionto be appliediterativelyto rnøuntil the a stablesolutionof the


desiredaccuracy'isobta•ued.Notice that we are actually solving• problem of the form
F=Az for z. Its solution'
for z is the s•measthat givenby eq. (7.14)audwill therefore
s-•ce to illustrate the iterative process.In the aboveformulation,d --0 by impl•catiom
ß

Tkisd•gressionazide,thenextstepis to updatethesolution thus:mx---mø-FSm.


Us•ugrn• asthe newguess solution, a new•m is found(a•er evaluat•ug
a newpartial
derivativewith respect
to rn•, i.e.,f(m x)) •ud •ddedontom• givingm:.
The process
is repeateduntil the desixed
solutionis found.Let uscompletethe exexcize.

Req•ed Steps'mø=l, f(mø)=4mZ-1

•ud so,

rn•=1+ (-if(1))='l+
f(1)• =1+0.3333=1.3333
•predicted response
rn:=
1.3333f(1.3333)
• 1.3333
_((1.333)•-(1.3333)-1•
+(-f'(1.3333)/= 4(1.3333)zL1 •pa•ial derivative

=1.3333 - 0.0973= 1.236,--- not goodenough(sincesecond decimalplace


is still non-zexo[u 8rn, soobtakunext iterate).

f(1.236)'•
1.236
+(-..f'(1.236)) = 1.236 - 0.0141

= 1.2214(second decimalplacestill cha•ging,soobta•unextiterate).


Usingma astheinitialmodel,wehavethat

= - 0.006(second
decim•placeis zeroin 6m,sostop).
6m= - f'(1.2214)•
Therefore, •he desiredsolutionis 1.22.

91
7.3.4 Problexn session

Find a solution for m correctto two placesof decimMof f(m) = ma-Sm+6 = 0


near m=2.

7.3.5 L•rrdtations of the Gauss-Newton method


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Considering
the assumptions
andapproximations
madein orderto derivethe above
algorithm,
onequestion
thatcomes
tomindisßW'• thi.•unconstr•ed
iterative
procedureaJwayssolveournon-linear problem7 The mafmdrawbacks of this technique
is that a goodappro•mationto the actu• model (•, a goodfn• guess)is requL-ed
for
theproceduxe
to converge
andthatthematrixATAmaybesingula•or near-sing-l•
producing
undoableeffects.
Thelatte•results
whenATAisill-conditioned
(i.e.,the
ei•envalues
are verysmallor neax-zero).
Aswesawin Section
4, the calculated
solution
becomesso large asnot to be physicallyre•d•.ablewhenthe eigenvalue•are verysma•.
The solution,in this case,is saidto 'ovex-shoot
the lin• range'.Thus in practiceit is
necessary
to correctrn•by onlya f•actionof= withtheattendant
decrease
in rateof
convergence.
Notealsothatevenif ATAisnon-singular, thesolution
maystilldiverge
or convergevery slowly.

7.3.6. The Steepest- Descent(or •ra•ent) method


The method of Steepestdescentis a simplegradientmethod.In this method, the
initial model is correctedin the direction of the netafire gradient of the objective
function q, i.e.,

aq

where/c is a suitably definedconstant.Rec,• that

q= (d-f(m))T(d-f(m))= i I

Thus,

='-2(d-f(m)). c9m-- (7.18)

sothat the parnrutercorrections


aregiven(usingeq.7.17) as

92
I• is obvious
that replacing
theconstant
factor[2k]in eq.(7.19)with [AU:A]
- • will lead
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

to the Gauss-Newton solutiongivemby eq. (7.14). The valueof k determ•u• the step-
sizeof the corrections.
Noticethat eq. (7.19) doesnot cont•-;.amyiuvex•ematrix. The
schemedoesnot divergeamdthisis amadvamtage over•heGauss-Newtonmethod.
providedk is smallenough.
It hasgoodiuiti•l convergeace
characteristics.

7.3.7Drawbackof the SteepestDescentme•od


A majorshortcoming oœ the methodof Steepest-descent is that the rate of convergence
decreases asthe leastsquares solutionis approached--;ke the Gauss-Newton method.
It is ofteninefficient,requiringa largem•ber oœ steps•ud is thereforenot recommen-
ded for practicalinversionapplications.

7.3.8 Are thereaay remedies


for •ability audnoa•vergence ?
Topreventunboundedsolution
growthwhenATAisill-conditioned,Levenberg
(1944)suggested
a methodof 'Dampedleastsquares
' to d•.r•ptheabsolute
valuesof
thepax•meter
perturbations
during.
successive
applications
of Taylor'sappro•rntions.
Levenberg
suõgested
that axbitraxy
positiveweights
be addedto •hem• diagonal oœ
ATA •ud •lso showedthat the directional derivative of the residual s-• of squaresq
hasa minim-• whenthe weightsaxeequal.This ideawasla•erusedby Maxquatdt
(1963,1970
) to develop
a veryuseI•non-linear
algorithrq.
Thistecl•queis caUed
RidõeRe,sion or •he Marqu•dt-Levenberg me•od andisthemostpopulax
inversion
procedure
withinthegeophysical
cor•-•y. I• is essentially
aninterpolation
between•he Gauss-Newton
•ud S•eepest-descent
methods•ud successfully
combines
•heir usefulpropertiesin •he hybridscheme.

7.4. Const•i-edInversion:
Ridge•oa or the Marqu•d•-L•enb•g Method
Theunderlying philosophy
hereis thattheparxmeterchaugesz• mayover-shoot
the
linearrauge(fora non-linear
problem)'if
theirabsolute
values
areleftunchecked.
A
boundis therefore
placedonthesizeof theperturbations
therebyconstraining
the
Step-length
ofthesolutions.
Operationally,'in
ridgeregression
werni•imi•.e
boththe

93
7.4.1 Statement of the m•m•,ation problem
Minimize the objective/unction

4 = • + •q• = ere + •(•r•-L•) (7.2o)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

wherewe axe•=•m•zing somecombination


•boI •heprediction
erm•azxdmlufion
len•h •d wehaveplaced
a bo•d L• ontheener• ofthep•amet• ••ges; •d • is
• •dete•ined (Lavage) m•tip•er that dete•es ther•ative impo••ce that
• be•ventoqx•d q2-Here,• isr•e•edtomthe •a•p• •. Notethereore
that •dge recession
is essenti•ya comtr•ed non-•e• lemtsqu• method.H a
vMuecloseto zerois •si•ed to t, the •olutionappro•at• theGa•s-Ne•on
solution.

7.4.2 Problem solution


Minimizationof eq.(7.20)is effected
in thesamemazmer
asfor theunconstrained
le•t squaxes
method.The procedure is simple'

Idifferentiate
•bwithrespect
toz,• nmult
tozero
aud
solve
for
a,/,
_ a{(y-.,•)=(y-•)+•(===-•))
Ox -
=o
Hint:

[expand
•het•nsinthen,,merat•r
o/eeI. 7.21
and
then
can
To•.ttheel'rffer•.
fiation
I.
We.will then find that

2ATAx - 2ATy -- 2•x---- 0


o•

(ATA+ l•l)x-- ATy , the n• equ•t;iox•

fromwhichweobtaintheridgeregression
or constrebled
or damped
leaatsquares
solutionfor the parxmeterperturbations

ß,.= [A•A+ l•l]-•ATy. (7.22)

9•
The solution• is thenusedin an iterativeprocess to/it our data.If our starting
model is m0, non-linearityis dealt with usin• the iterativeformula

m• +• = m• + [ATA+ •I]- •AT•/ (7.23)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

=d = +
for k•l iterations.
A comp•son of equatio• (7.16) •d (7.23)• rev• that the latt• is • ••ive
wayof de•g •th s•••ties • AZ• •dge re•sion is • •• a hybridtechn;que
in the sere that it combinesthe stypest d•c•t •d the G•s-Ne•on metho&. •e .

steep•t descentmethoddomin•t• wh• the st•t•g modelis • •om the solutionto.


theproblem
w•e theie• squ•es
method
becom•
do••t • thesolution
• appr•
ached.M•qu•dt(1970) showed
that ridgere•si
zed•version,•d that(ArA• •)-XA • approa••theg•••ed •v•e of
P•ose (1955)• • • •, •d intofeted themat•
set a •o• data.

7.4.3 Application of •/n non-linea• ß

The leas• squ•ressolutiongivenby eq. 7.14 ca• be written in

:.= [ArA]- 'A:r•= [vA- :v'r] v,w'r• = VA- 'trr• (7.2t)

wherewe haveass,,med
thatr=--iv
andthat theinverse
of ATA (i.e.,VA-:v'r). exists.
In •idgeregression,
wereplace
the1/I• in theA- • matrixbytheelement

sothat the ridgeregression


estimate,in termsof SVD is formally

whereA•-• is relatedto A- • by thetraudorm(7.25).The advautagesof using


thesetransfo•ed variables areobviousßwhenI i is muchl•rgertha• •, thequantity
(li+/•) diEers
verylittlefroml•; andwhenI i isdoseto zero,theaddition
ofa
positive
bias• willensure
that A'rAwillnotbesodoseto singulaxity.

95
7.4.4 Determination of d•rnpingfa•o• in •dge refreshen
For automatedinversion,the commonpracticeis to set/• first to a la•õe positive value
thusta•dngadvantageof the goodiniti• convergence propertie•of the steepestdescent
methoda•d thereafter/• is multipliedby a factorlessthat •u•ty a•er each iteration so
that the linear le•t squaresmethodpredominates nearthe solution.A va••t of this
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

procedure
(Joha.usen,1977)
azs•mes
a.s/9
thesraMlest
eigenvMue
ofATAmatrixa.ud
divergence
occurs:
it is replaced
by thenextlargesteigenvalue
untilthe solutionis
obtained.A moresophisticated
methodof da•npingh• beendeveloped
by Meju (1988,
1992)andis in effecta hybridofthe twomethods
higM•ghtecl
above.The dsmping
factorsa•e cleterm•ued
empiric•lly- f•om what might be likenedto appro•rnate
derivatives
of a Lagrangia•function(seeHerskovits,
1986)- andusedin
m•u•m•zation
sub-problem
ateach
iteration.
Theprocedure
forestimating
thedamping
factors
isi].lustrated
g•aphically
in Fig.7.1.Operationally,
thela•gesi
a•d smaller
eigenvalues
oftheproblem
a•emultiplied
by10.and0.1respectively
•iving
thata•eused
todetermine
ihecoef:fcients
ofa'Pazabola
fromwhich
tensxmples
ofthe
auxiliazy
factors
A• a•eobtained
using
thef6rmula
(Meju,1992)

A./,:
= ({100A•-
A•}+ {AI - ,k•},•)/99.: /c= 1,...,10. (7.:27)

Thefactors
,k/•aresquared
togivethe-d•nping
f•ctors/•required
in a !iueseazch
proceclttre
basedonequation
(7.23).Thistecl•u•que
involves
coarse
stepsin/• in the
re•ionof/• butthesteps
become
•nertowards/•(seeFig.7.1).
Thedaznping
factoris
setini½i•llyto thelargestvalue/•10sothatthe chazacteristics
ofthesCeepest descent
methoddominates a•d thensuccessively
smallervaluesoI •t a•eu•edthus enabling
the leastsquares methodto bedominantat the later stages of the line search'
procedure.Thesolutionbeforedivergence is return• asthenextiteratein .themain
callingpro•m. However, i• nodivergence
occurs clog thisintermediate search
œo•
•uopt•mM
solutio,•,
• value
ofze•oisassigned
to• yielding
theunconstr•ed
least
squaresestimate.

7.4.5 IlldEe •rea•ion in prac•ce


It will be a•s•rned here that ou• œelddata h•ve a•sociateclerrors.These error• •re then
usedto de•,•ea weightingmatrixW' asbefore.An effectiveridgere•ression
algorithm
thatusesa 1-Dlinesearch procedureisgivenbelowto illustrate
howthemethodmay
beimplemented in practical
applications.
TheMgorit•m issimply(Meju,1992)'

96
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2400
'3000

1200
1800

600
3600

10
8
•L=q2L

97
6

Figure 7.1
4

Index number (k)


I•s
2
Determination of damping factor

s
Step 1. Select • starting model, mo .
Step 2. Calculate
(wv) = wd- wi(mo)
and

q•- • (w•)•
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

StepS. If the computedresponses fit the experimentaldata d, or otherstoppiag


criteriamet (e.g.,q• >_qrr), STOP.
Step4. Obtaia the weightedpartial derivatives,(lirA) = A..
Step 5. Calculatethe SVD of A•.
Step6.' Obtain the mostfeasiblesolutionusinga liu•se•ch procedurei• a ridge
regressionsub-prograax.
(•) calculatedampingfactorsfor the 1/mesearchprocedure

x•= ({•ooa-•,}+ (;•,- x,}a)/•,., (v.•.v)


and hence

• = x•, •,..., •o.


whereA• and • •e the stalest •d'l•gest ••• vMu• of • m•tipH•
r•pectively by 0.1 •d 10.
(b) set•0 = 0.0
(c) peffom •e se•ch •th •dge re•sion
Loop (j=l to 11 md k=l 1-5)

Get dii•[•]• '1= di&•[•]i/(di&g[•]


2
• + •), /=1, •

Calculate

Co=•p•t•q•- i----1

If(
se•opt•,•] solution
•o mi_ t a•d qui•
else

se•optimalsolution
End Loop
$•ep7. Se••he op•im-mmodelfromstep6 asthenewi•er•e (i.e.,too).
Step 8. Go•o Step 2.

98
The aboveridgeregression procedureh•s beenfoundto be ve• •ective • res•ti•ty
dat• •version (e.g.,•e•u, 1992;•e•u •d Hu•ton,lDD2.).
No•icethat the comput•-
•on• req•rement• S•ep6 is••• once•hef•c•o•sa•ion
of• • b• done.•e
bterested'readermay no•e•ha• H••o•s (1986)pro•d• • v• •goro• ••is of •
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

tw•sta•e minimization
•go•t• • w•• a somewhat ••ogous ••se•• proced•e
•s•es the glob• converg•ceof the s&•e •d the fe•ib•ty of tM it•at•.
However,• sho• fo• •• problem,,the r•ts •om M•qu•dt •v•on may be a
smutbed versionof the truth. •e questionth• •s•' C• a n•i• solufi• •
det•• I• a n•llne•r p•bl• ? We •
.
now•mi•e somewa• of •d•ess•g tM
probl• of non-,miqueness
• nonllne• •ve•ion.

99
8. NON-L]/NEAR BIASRD ESTI1WA•ON

A Gener•l•.cl App•ar.]l
In interpretinga sca.utysetof inexactdata,conventional wisdomis to seekmodelsthat
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a•e in agreementwith a priori data derived•romsay,previousgeophysical studies,


boreholeor geologica• data. Theseextraneoms infomat/on helpsingleout • plamsible
solutionfrom amongstaJlpossibleonesadmittedby the inexactpracticaldata. In
Chapter 5, we sawhow the useof a priori data in 11ne• inversionhelpsresolvethe
problemof non-uniqueness. Un•orttmately,whilethe solutionto the linear inverse
problemincorporating a prioriinfomationiswellknown(seee.g.,Jackson, 1979),there
is no unequivocMtechnique• yet for resol.ringnon-,,n•queness in rionil-earinversion.A
n•ive approachto usinga priori informationin nonlinea•inversioninvolveshola•.g
const•t the v•lues of someof the soughtparameters withi• an iterativeinversion
scher•e.The formalapproach is to incorporate
the a priori datadirectlyi• the problem
fomulation. Much of the pubfishedfomal treatmentof a pt/or/informationin
nonlinear
inversion
adopta probabilistic
apl•roach
(e.g.,Gol'tsma•,1971,1975;
Ta•a•tola
& VMette,1982; Jackson& Matsu'ttra,1985;Pouset a1.,1987;Duijndam,1988)which,it
maybea•gued,
bestcha•acterises
thehug•va•iabifity
in practical
geophysical
mea.sttrements
(Meju, 1994d).We will •doptthe sameapproar.
h a• in ou• treatmentof
1in• inversionsinceit involves:'simple
matrix algebraand rn{r•:•rnaj..
statistical
commitment.FollowingMeju (1994d),•e problemof inver•on •if• a •
in•ol-mationwill be addressed
within a .•ifying fa•rneworkof biasedestimationwitix
emp• on simplicityandeffe•ve pxactical procedures.
In thisapproacha distinction
is madebetweenstar•ingmodelsand a i•/• informationfor clarity.We shallaim to
make the inversionschemeflexibleenoughto a/lowfor the construction of a variety of
leastsquaressolution•usingeitherreliableor diffusea vr/o• datathusmakingit a
useftd
practicaJ
toolforexploiting
pa•i•azgeophysical
situs%ions.
TOachieve
this ß

objective,we will needto considerseveraJformsof a pr•or/constraints


or solution
simplicitymeasures addtakethe observationalerrorsinto account.
Finally,we will
showthe relationships betweenthe ens,,ingalgorithmsand variousclassical
inver•on
algorithms.Thisstrategy, it is hopedwillenableusto analyseandunderstand themore
rigorous
algorithms
fornon-linear
inversion
(e.g.,Ta•autolag:Vahtte,1982).

8.1 Underlying Phil•hy


The biasedestimationphilosophy is simpleandthe undergitding
principles
azesimilar
to thosedescribed
in Chapter5. In somepracticalinverseproblems,we may havesome

100
ide• •bout the fore of the solution,the r•ngeof possible
valuesthat the sought
parameters
mightas•me,orhavereliable
priorestimates
ofsome
oftheparameters
of
the problem. Our goalin biasedestimationis to retain,where•vafd•ble,any reliable
prior estimatesor desirablestructuralformsin the finM solutionto an inverseproblem.
However, in caseswherethere are no realisticprior estimatesor structura• form of the
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

solutiona.udwe canonlymakean intelligentguessasto whatthey m•ght be, it is


desiredthat the solutionprocess be stableaadthat the finalsolutionbe independent
of
the initial guess.It is thusimperativethat wemust 'dmtiuguish
betweena prior•
•,•orma•ion,•i andinitial(orstarting)models,
møin •1•.,approach.
The components
of
the starting model m,,y includethe reliable a pr-k• estlm• towards which we wish to
biasthe final solution,but this is not obligatory.
.

We •lso striveto mainta• statisticalstabilityin ou•Solution


process.
Thus,owin• to
the nature of.practical data a•d if we ass,,methat the standarderrors •r associatedwith
otttn observationaldata,d arestatisticallyindependent,a• n xn diagonalweighting
matrix W=•r-•I may be definedandusedfor scalingourdataequationssothat undue
importanceis not givento poorlyestimate;:l data especiallywhenou• a Fr/.or/
ass•mptionsare not very accurateor reliable.This scalingoperationcan alsobe
interpretedasa process
of norm•li•.ation
of standardiz•on(asthe data equa•• a•e
thusrenderedrllm•od• aud uncorrelated).

8.2 Non-li,,ea• Inversion with a prio• information


In order to retain the desireda p,-io,'{estimatesh of the p soughtmodel parameters
in the œnal solution to aa inverseproblem involvingour baud-limited observational
ß

data, the necessaryconstrig equationsmay be developedin the form Dm =.h.


In thiscase,Dpxlbis anidentitymatrixthatoperates
on thep par•r•ete•,m to yield
or preservethe a priori data contain• in the vector/• i.e.,

I rn1

Dm= '. ee. -- e


. (8.1a)

L
Alternatively, if we ameinterestedin ret•i,•i,•ga particularform of the solutionthat is
in accordwith the physicsoœthe problem,say,we canalsodevelopthe appropriate
constraintsas in eq. (8.1a). For example,if we desirethat the modelparametersvary

lol
smoothlywkh position,thenweneedto minimizethe differences
between
physically
adjacent par•m•etersusingthe constraints

Dm=
1-1
1 -1
.m•
.. = '
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1 L,j
where I = p-l; and the I x p matrix D is a first differenceoperator.Such smoothness
(or derivativeregula•zation)measures
appearto havebeenintroduced
in da•ainve•ion
by Twomey(1963,1977)andTikhonov
(e.g.,Tikhonov,1963;
Tikhonov
& Glasko,1965,
1975;Tikhonov & Arsen•u,1977)and may be calledTwomey-T{khonovsmoothness
measures.Note that other regul•ation me•sttrescau be developedbut the two de•necl
above will Sl,•ce in this course.

8.2.1 R•tention oœknown parameter valuesifi the solution


8.2.1.1 Problem formulation

Our goal here is to bias m towardsh. W'e simply state the problemas:
"Given a t•ite collectionof ineaxac'tobservational
data, find the solutionamongstall
the 'equivalentones( on accountof data and modelm'zo•), that exp• the
obscrva•• and satizfie• the available reliable estimatesoI the model parameter".
Mathematicaj•y, the abovestatementis equivalentto mi•irai•uõ the predictionerror
ere and the departureoœthe solutionfrom the speci•edconstra•ts as discussedfor
llueaz problems. The mathematical developmentinvolving an elaboratetreatment of
the data errorsis straightforward.If reliableestimatesof m are givenin A, then we
simplymiulmi•.et-heLagramgeau function(Meiu, 1994cl)

z= +

where• is a diagonalmatrix of Laugraugemultipliersthat enablesus to pick out


thosepatmeters that shouldbe forcedinto conlo•ty with the prior estimates
leaving
therestunconstrained.
Thediagonal
elements
corresponcli•g
tø thepriordata
are assigneda constautvaluegreaterthan zerowhile thosecorresponding to the free
par•Lmetersare zero and may not be considered
in the interpretationprocess(Meju,
1994d).The mostapplicableconstraintsfor this kind of problemaregivenby eq. (8.1a).
H j•(m)is continuous
anddifferenti•ble,
expanding it aboutan imitiMmodelmø using

102
Taylor'stheorem•ives the1inearized
approximation
of eq. (8.2), i.e.,

(8.3)

, and we kave
where!•---• - J•(rnø),
.4isthenxpm•trix
withelements
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

usedthe •pproxim•tionß = m-mø.

8.2.1.2 Probl-m solution

M'mimizationis effectedby settingthe derivativesof œwith respectto z to zero as


usual.
[?•IT:It isalways
helpial
toexpand
equa•n
(8.3)befog,
ait••tiatioa.]
Thus,
dropping
the W termandclericalõ
thematrix• = •r• fornotational
simplicity,
we
have that

_ mOrDVa
+ zTDVDm
ø+ zTDTDz_
=TDVA_
AVDm
o_ ATDz
+ ATA}.
We can now carry out a simpledifferentiationwithout the risk o{ forgettingsome
terms. At the minire-m, we have that

•œ= -•TA- A=p+ A=Az+ zTA=A


+ møTD=BD
+ DTBDm
ø+ DTBDz

+ zTDTBD- D=BA- ATBD = O.

That .is,

--2AT!/+ 2ATAz + 2DTBDmø + 2DTBDz- 2DTBh= 0 (8.6)

yieldingthe norruMequations

ATA= + DTBDz = AT• + DTBA- DTBDmø (8.7)

whosebiasedsolution(with the W term reinstated)for the parametercorrectio• to be


appliedto mø is

(8.8)

103
sinceD=I (eq. 8.1a) and BI = B.
Nonlineazityis dealtwithusinganiterativeformulaof the form

• +• - • + [(w•)•w•, s]-• [(WA)•:W•


+ B(•,- (8.9)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

wherethe Jacobian
matrixA andthedatami•t vectory areevaluated
at m• andthe
iterationis begunat k=0.ThetermB{h-m•} ontheright-hand
sideofeq.(8.9)'helps
to forcethe solutioninto conforty with the specified
a priori parameterestimatesand
this inversionformulais dubbedthe ma•'•vm-bi• algorithmby Meju (1994d).

If the a priori.informationisdiffuse(or•mreliable), thenit maybedeskable to setthe


constraints equalt6 zero,i.e.,• - [0,0,...,0]•, andall theelements
of • aresetequalto
a constant value(0 < • < 1) sothatantheparameters haveequalimportauce. In this
case,• may beconveniently replaced bythesingle undeterminedmultiplierfl. We
then have that

(8.10a)

which canbe approximated


by the Ene•ed system

z = (wy- w•) ß (w•- w•)+ {/•(,•ø+-)•D•D(,-ø+-)}


ß (s.•0•)

The solutionfor the parametercorrections


is simply

and the iteration formula is identically

,•+• = .., + [(w,•)•w• + fi.•]-' [(w,,)•w•-

since
D=I. Here,•I serves
tocontrol
thestep-s•e
ofthesolution
while•ra •
helpsreduce
itsbiastowards
thenullvector
h andwemayrefertotheprocedure
asinversionwith smoothness.
constraint
or the .l•n;t-,,,l•-biasalgoritbl•.
8.2.2 Retention of Mao• form of the solution
Let us now turn to the situationwhere the bias in the solutionis dictated by physical
arg•rnents,say.In the casewhereh cont•,•sourspecifieddifferencesbetweenadjacent
par•rneters(e.g.,a fixed-step
model)•ud the applicable
constr•,,tsaregivenby
eq. (8.2).Here,all thepain-meters
haveequalimport•ce andweneeduseonlya single
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

undeterminedmultiplier•. That is, we needto minimize the function

• = (wd- wf•) ß (w•- wf•,•)+ f:(o,•- s)•(•- s) (•.13)

a•d the solution for the z that will lead to the minimization of the function œ is
identically

= = [(WA)TWA+ ,s:•.tr]-
•[(WA.)TW•
+ ,S:.O'r{a-.t:•o}].

where the m•trix H = DTD.

In manypractica•situations
wheretherea•egradational
cha•ges
in physical
prop•i•
• the subs•ace say,we may be •terested in obt•g the smother. solution•k
••zed •erences betw•nadjac•t p••eters.. For•t•ce, forph•i• re•o•
onec• ••e that there•e nosh•-•scont•ties in r•isti•ty • the subs•a• •d
elect •o •ve• the obse•ation•l data for smoothmod• (e.g., Co,table et M., 1987;
Meju•d Hutton,1992).• su• •es, wedonot rely have•y qu•t•able
est•ates of h then convention• •dom •ctates that we set the •••ts of h •
eq.(8.2)equM•o zero•d theappBc•leco, trots •e •v• by•. (8.lb). •
p•bl• • th• d•• • t• • f• •e •••• m•• (• ga•• by •

Mathematic•y, we minimize the ••ion

Z = (Wd - W/(•)) T (W•- Wf(w'•))


.-• (s.•)

whosesolutionfor the parametercorrections


is (Meju 1994d)

In thiscase,the term•H serves


to regularize
(Tikhonov,1963)thesolutionprocess
whilethe term-•]fm øhelpscontrolthe solution's
departure
fromthestipulatedform

lo5
a = [o, ....

8.2.3 Consistencyanalysisof non-linearsolutions


A good
practice
innon-linear
formulations
istoascertain
thatthesolution
process
will
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

yieldthelinearsolution
if the problemis s,,•cientlylinear,i.e., that the formulations
axeconsistentwith the well understoodlinearanalogues
(Meju, 1994d).R.ec• that in
linearinversionthe soughtmodelpaxameters weredirectlyretrievedfrom the data.In
non-linearestimation,we may Msosolveexphcitlyfor the soughtparametersinsteadof
the usualparametercorrections andthe solutionthusobtainedshouldreduceto the
linear solutionif the problemtackledis near-linear.
This approach
ß
is alsousefidwhen.
calculatingpaxa•etercovariances in non-linearestimationasshownlater. To achieve
thisgoal,
wesimply
goback
totheappropriate
nca-m•
equations,
rep• z withthe
quantity(m-mø) andthensolvedirectlyfor m.

Takingthedataerrors
intoaccount,
forthe-unconstrained
(Gauss-Newton)
solution
we
have that

((WA)TWA)(rn-rn
ø)= (WA)TW'•I

giving the unbiased'estimate

where{2} = {W• + WArn


ø}isa k•,•dofdataandthealgorithm-
canbeforced
to give
ottrunconstrained
linearleastsquaxes
solutionif the problemis linear.

Thenormalequations
formazdm,,rn-biaz
estimator
(eq.8.7)maybewrittenas

so that

,.= (AREA+ + nSh]

whereE andthenewdata• axeaspreviously


defined.
Theabove
estimator
caneasily

106
be shownto be consistent
with the linearestimator
•ven by eq.(6.2).

For the minimum-biassolutiongivenby eq. (8.11), we havethat

ATEAm+ f?2Im = ATEy-/•2Im0 + {ATEArn


ø + •2Irnø)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

givingthe explicit solutionfor the modelparameters

rn= (A•EA + •2i)- •(WA)•(•} (8.20)

whichis •lso consistent


with the linear •u•logue(eq. 6.lb).

For the solutionfor smoothmodelsgivenby eq. (8.16), we havethat

ATEArn+ •2Hm = ATEy--•2Hr• + {ATEAr•ø + •2Hmø}

giving the explicit relation

rn= (ATEA+

which is consistent with the.linear'solution for inversion with smoothness measuxes

(equations5.7 & 6.1a).

8.3 Re•tion.•hlps with atandard metlinda


A comparisonbetweenthe •lgorithmsderivedabovea=d a few popularLuvexaion
schemeswould perhapsthrow more light on the pr:mciplea•d tech•!que$of.blazed
estimation As in the lineaxmethods,if f• -, 0, then all the aboveiterative e•timatioxt
formulaereduceto the weightedclassicallea•t squa•ezupdateformul•

m•+ • = rn• + [(WA)TWA]- (8.22)

whereA andy areevaluated


at rn•, hig•$ht•g theimport•uce
off• h• biased
estimation. Thus the stability characteristicsaud the effectiveness
of the biased
estimationmethodsdependlargelyon f• and D.

107
Notice that eq. (8.11) diHers•rom the usuald•nped least squaxe$or ridge regression
updateformul• (seee.g.,Meju, 1992)

•-- [(WA)•WA + $I]-•(WA)•W• (8.23)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

by the -/•2rnøterm.Equation
(8.23)is theweightedvaxi•-uof theridgeregressio•
œormul• g•venby Maxquaxdt(1963,1970)•ud ttoerl& Ke•a•a (•970).It is obvious
that the •bove formul• couldonlybe derivedfxomeq. (8.10b)by repl•cinõthe l•st
term i• braceson theright-l•a•dsidewith •x•DTD•:; but thenwewouldnot be
dealin with the fullline•ed setofequations a•d thisha• implications
for •heintuit-
ive arguments employed in theusualridgeregression
formulations
wherethe singula•
goalistoplace• bound
ontheenergy
ofthepa•mete•increments.
It isperhaps
apropoz
to re-examine
thepopula•
ridgeregression
methodwit'hi•thef•amewo•k
of
biased estimation.
Le• us define•he constr;d•edinversionproblema.sminimize the L•raugeau f•mction

• = (wd - w f(..))ß (Wd- Wf(,..))+/•(.. - .•o)•(.,_ .•0) (8.24)

wherewe are m•y search•õ'forthat bounded


'perturbationto our i•tial modelthat
is op•al for fitt•g our'd•ta(c[.Marquardt,19õ3).Notethat weh•veneglected ottr
usua•operatorD sincei• is an.identitymatrixa•d all parameters
haveequMimportan-
ce.Linearizing eq. (8.24)a.susual.
gives.
thefollowingappro•ation to the'objective
function:

whosesolutionfor.x is •iven by eq. (8.23).To fully tmderst•udthissolution,let us


•u•yseisforconsistencywith.thelinear•uMogue
represented
bym = [GTG+fLr]-
•GTd
(i.e.,eq.5.10).Fromthenormalequations,wehave

(ATEA+ •l)[m-mø]= A?E•I


or

so that

108
wherethe quan•i•yd = (W!/+ W'Amø)•sa form of data.It is obvious that the above
relationwill not •eld the traditionaldampedleastsquares formulaIo• a linearsystem
(eq. 5.10). For • linearproblem,eq. (8.25)will reduceto some•h• in •he formoœthe
biasedestimation
formula
givenby eq.(5.2)if the term•rnøisequated
to •h, i.e.,
some'azs•rned' • p•c• model. Operationally,the main diHerencebetweenour biased
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

schemeand the Marquazdtalgorithmis that a new a pr•or•model(i.e., the previous


iterate) is usedat eachiterationin the latter a•d this strategyis not in accordwith the
presentphilosophyof biasedestimation. Paraphrased in a differentway, treatingthe
previousiterate as the a priori modelis nt consistentwith our ass•rntion that h is
independentof the observationaldata traderconsideration,sad may lead to undesirable
results. It is not •he favouredapproachto sxtdxessing the problemof non-•-quenessin
non-linearinversion,albeitthe mostpower• rnimizationscheme(seeMeju, 1992).

Usingthe quautityE in placeof W•W in eq. (8.9), wehavethat

m•+•= m• + {[A•SA
+ BI-•[A•E!t
+ B(h- m•)]}
whichis equivalentto the BayesianestimationschemeoœJackson& Matsu'ura(1985,
'eq.65-67) •ud comparable to Tarautola& Valette's (1982,eq. 49) no-i-ar algorithm
(seealsoPous et al., 1987,eq. 3) iIB is interpretedstatistica•yasthe inversea prior•
parameter corafierce matrix (Mejtt, 1994d).Thus, usinga simplealgebra,we have in
effect developeda schemethat is similarto the more mathematica•y rigorousschemes
ba•ed on a probabilistictreatment of nonlinearinversionwith prior data. However,note
that the inversionphilosophyand the usageof a priori informationin Taxautola&
Valette's lan••k method cliHexsomewhatfrom ouxs.We axeprima•y interestedin
forcingthe •nal solutioni•to closeconfo•ity with thoseparametere•timateslmowna
.

prior/and thus the last term on the right-handsideof eq. (8.26)is non-zerosinceo•y a
few parametersestmimatesmay be knowna priori in typical practicalsituations. In
Taratttola& Valette'salgorithm,h is simplythe actualinitial modelmø. In thiscase•as
inPøus
et al.,(1987,
eq.6),thelasttermineq.(8.26)
would
beequal
tozerointhefixst
iteration.We treatedh asbeingseparate
frommø,moreor lessasin Jack.son
&
Matsu'ura (1985). Note that in the algorithmof Jackson& Matsu'urathe quantityin
bracesin eq. (8.26) is multipliedby a suitablychosen
factorb (0 < b< 1). It may thusbe
contendedthat ouxsimpleapproachis somewhatgeneral.

109
8.4 Ir•plementat•on oœconstraints[u «t•rat«veapplications
The iterative constr•ned solutionprocessc• be effected• before•th the audited
equations

,
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

wh•e • = (h-Dm•),m• •e thep•meter est•ates'at thepre•ous(i.e.,


iteration,7 is the apphcable
,
Lavage m•tiph•s (fl or fl) •d D is the apphcable
problem
re••zation mat•. Notethatmaissimply
mø• the•t iteration
•d that
Aa maynotbe a n•-vectorev•thoughmømay•cludethem•bers of h.
optim,•mvMuefor • isfo•d by t•M •d e•or. However, a reco•end• proc••e
to selectthatv•ue that•ves • acceptable
residu•I d-f(m*) Ia •d •d• a re•t
t•at satisfiesthe •posed comtr•ints.

It is pertinentto mentionherethat in inversion


withsmoothness
memsttres
(eq.8.lb), if
we wish to retain knownestimatesof the soughtparameterswhilesmoothi,•gthe
variationsbetweenthe -,•knowns,then an alternativetechniquethat has beenfoundto
be effectivein practiceis to partitionthe augmenting
datafor A and• in the form

h.1

; h= (s.28)

andthe applicable.
constraining
equationis simplyDrn = h whereD is of dimension
ixp, •' isof dlm_•n.•ion
(p-•-l)xp andtherearet known
reliable
estimates
of •

A commonly used•echniquein geophysics


is theinversionof combineddatasets(Vozoff
andJupp,1975;Gol'tsmam,1976)andis anef•tive menus of deriving
reliableestimates
of thesoughtparameters.
Experience ha•shown thatthistechnique maynot Mways
yielduniqueparameterestimates
especiallyiI theconstituent
datacomponents
are
inherently
plaguedby non-,miqueness.
It maythusbemoreeffe•iveto incorporate
priorestimates
ofthesoughtpaxa,•eters
if available.
Thealgorith•n.•
described
hereare
flexibleand canbe usedin a joint inversion
scheme
with the dataandconstr•i•;•g
equations
axranged
in the form

11o
wherethe indicesI to • referto the weightedcontributions
fromthe • different
setsand the other symbolsare aspreviouslydefined.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

It is remarkedherethatwec•u alsodevelopa hybridpra•-tic•tech•;quethat combines


ridgeregression-type
solutionstabilization
measures
(e.•., Meju,1992)•ud biased
estimation
techniques
byadd•õ•n additi0n•lterm• to [A•A.] with7=1 in
eq.(8.27)where• is • Marqu•rdtd;•pin• factor.The •l•orith•-.• •ud practic•
techniquesdescribed above'
aresnmr-arizedin thefollowing•eneralalgorithm.

8.5 Problem Session: Constraction of E•ective


ß
Bzised F_,stimz•oa Sch•-me
Let us •ttempt to use•11that wehavelearnt sofar to developau effectivepractical
altorithm for nonlinearbiasedestimation.A .typicalinverseproblemis de•ed below.

Problem ß

Givendatad• (i=l,n), determi•e.opt•M


parameters,
m
informationh•.

Solu•on' Procedure:

Step1. Select•u initia•modelmø cont•g a pr-/or4 information.


Dete••e t•e of a p•o• co, trots •d Io• 7D; (wh•e 7=•, • s•e
•dete•ned m•tip• or a &agon•. mat• fi co•ist•g of z•os •c•t •t the
positio• co=espona•uto the r•able a •• •omation • mø wh•e it • se•
to a sm• positivevMue)

Step 2. Calculate the augmentingdata for y.


Loop(i = n+l,n+p; k=l,p)
=
End Loop

Step 3.' Clculte = [Wd - WL(mø)];

C•lcu•ate $SE =

111
( SSE is usedas a •toppingcxit•on)

Of •(mø)
Step 4. C•lculateA{i -

Form augmentedJacobiaum•trix
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Step 5. = = [a2 )

Step 6.

H(SSE' >_$SE) STOP.


else

Setmiterto mø
GotoStep2 (orStep3, if h=[0,0,...,0]• )

Notethat the solution


(of the-normal
equations)
for z at Step5 caube obtain• using
thesingulax
valuedecomposition
method(Eckaxt
andYoung,1939;
Lassos,1961).

8.6. Biased E•imafion in Practice:.m•,=•a•e


.
Examples
We • i•trate v•o• •pects of it•ative bi•• •t•ation ••g ma•etot••c
(MT) data.•e Mgofith•d••b• byeq.(8.11)isapph• to a s•thetic dataset
Figs.8.1a•d 8.1bto testits•e•iven•s. Notice• theefi•• tMt • the
p••et• oftheactuMmodal (A) weresat••to•y recov•• (•,•1 mod• lab••
F) =•g a•eer•t poorst••g moa (s) =d theprobl• re••ation (ord•p•g)
favor •=0.01. The•••-bi• Mgofithm
r••ed about7 it•atio• to ••e thee
r••ts. comP•abler•ts w•e &o obt•• •th •mø setto z• • eq. (8.11)to
i•tate the•u• approa• to non!ine•est•ation but req•ed a•ut the same
of iteration.T•s suggeststhatbothapproaches
co•d •sh r•able •o•ation
a data set of •gh qu•ty.

112.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

!
)---

$mm•

w- U
m
•mmm•
Em
m m

j j i &I I ]

Mld30

[ -Bep]
i i ßi

T
mml

E
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ii
ii

i J i i &i J

uJ
i-'lñct3O

ii

( ßI•ep)
ii
In Figs8.2a a•d 8.2bthe mJnjml•m-biastechnique is appliedto the well-known
practicaldata- COPROD(Jones & Hutton,1979)- usingdiHexent valuesof • but the
s_•me poorstaxtingmodel(labelledS). The resultfrom a conventionalridge-regression
inversionscheme (Meju,1992)is a•zoshown(asmodel1) for comparison i.uthese
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

figures.
All theoptimalmodels
haver.m.s.errorof1.0(X2 • 30)•udaxethus
equivalent.
It caz•
beseen
thatthemixdm•m-biaz results
varywith/••udaJso
d•er
fxomthe ridgeregression resultin thetop•ud bottomsectionz
of thegeoelectrJc
profile.
We axethusfacedwith the problemof non-,•iqueness sinceaJlthe modelsequaJly
satis• the givendata.Onewayof addxessing thisproblem
wouldbeto usea
data,if available,in the inversion process.
Forconve•ence a•zdto demonstrate
the
effectivenessof the ma•ml•m-bi• •lgorithmhzhau•g a pt/or/data, let us
thata previous
study
intheaxea
fu-m•shed
a resistivity
of500fl-mforthetop8krnof
the Eaxth'scrust.We wishto reta•uthisv•lue in the fin• solution(F) to the COPROD
problem.
Forplottingconvenience
theaZSllmed
parameter
estimate
isincluded
in the
staxting
model
(S)inthisfigure
•lthough
t•dsisnotmandatory
inthealgorithm.
The
resultof this illustrativeexercise
is shownin Fig. 8.3. Noticethat the algorithm
retaJ•ed
thedesired
priorestimate
forthetoplayer's
resistivity.
S•mila•
testsinvolving
the otherparameters
of themodelaJzopr6vedsuccessf•suggesting
that it is a potent
tech-quefor inversion
withpriordata.Notethat all the aboveexamples incorporated
the constr•nts givenby equation(8.1a).

To demonstrate
the versatilityof biazedestimation,the smoothness
constraitsgiven
byeq.(8.1b)havebeenemployed withtwoct•Herent valuesof f• in theestimation
proceduxeyieldingthesmooth resistivity
models shownin Fig.8.4. Here,theEaxthis
pax•rneterized
intoa succession
offortylayersofdifferentresistivitiesa•d predeterm-
ined•xed •hick•ess(hzlogspace).In bothcasesit is ass•med thatthexeaxenokno•
pr•orestimatesofthelayerresistivities
a•d a half-spacestaxting modeloœ•esistivity100
firnwasadopted •udtheinver•on process usedeq.(8.16).Notice-the goodfit ofthe
optima•models' responses
to theCOPI•ODdat•andthatall thema•usubsurine
featuresseenin the sparsely-paxamet•ed
modelsof Fig. 8.2 caualsobe ide.
uti•ed in
thesesmoothmodels.Howevex, thesesmoothmodelsaJzoshow• dependence on/•
(thesymbols
O audT inF•g.8.4represent
•esults
fordiHerent
f•values);
the
signficanceofthisbehaviour
willbeexplored
fuxtherin Chapter
10.In a•y•e, weaxe
againf•cedwith•hep•oblemofnon,•iqueness
in thislargely
underdeter•i-ed
(i.e.,densely-pax•rnterized)
problem.Wecanaddress suchproblems
using theblazed

115
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

.%

.
i

116
IN•!=Iblcldbl ('Be,p)
,!

•tSblHcl
A
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

i
.%
mm

•m

•m

iJ
i , , i l,J

uJ

117
Ii
ii

ml

[ -õep)
IIII
"'
J i
i
i
i
i
i

i
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

•mmm

e.
,m
mm--

i
i i i i i i!
, i

118
i

H.L•30
i
i

.LN3B•d•d
! i
i
i
i

[-Bep]
• i i i mi !

9'
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

m
,m
mm

,m.

mmmm

119
i
i
i
i
i
I
i j

I
z
1.1.,I
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

I
.%

i i i i

120
I

ß

( ßBep)
I
ß

•S.l::lHcl
'7
IE
estimation techniques.

For illustration,let usass-r•ethat the •opcrustis no• constrained


by the COPROD
databuttha•some
priorestimates
of•heresistivity
ofthissection
ofthe•eoelectric
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

profile
isavailable.
Wewish
todemonstrate
theeffectiveness
of•hepractical
•l•que.
described
by eq.(8.28).
Theresultobta•edusin••he• p•o• information
thatthetop
8krnof the E•rth's crusth•s a resistivityof 500F2m(Model1) is showntoõetherwith
thatobtainedusin•a h•f-space s•artinõmodelof 100•qmresistivity (Model2) in
Figare
8.5for•--0.6.No•ice
thatweareable
toforce
thefin• solution
toconform
with
thehypothetical
a/•or4 information
in thetop8/c•(upper
14l•yers)
ofthe•eoelectric
crustwhilethe restofthepro•e wassmoothed.
Model2 is a•aindi•erentfromthe•wo
modelsshown in Fi•. 8.4fordiCerent
values
of•. I• follows,
therefore,
that theuseof
hybridconstr•nts
(eq.'8.28)
maybethepractical
approach
toreducin•
theinherent
uncer•_•ies in •heinterpretation
of suchdensely-par•eter/zed
models.

121
9. SOLUTION APPRA•• IN NONL•A• INVERSION
An inversionprocess
is incompletewithoutan auMysisof the errorsin the constructed
solutions.When dealingwith inexactexperimentaldata, a fundamentMtask is the
determinationof howdoseto the truth (if thereis one!) or howrepresentative
of the
real world our inversesolutionis. To shedsomelight on this issue,we will examine
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

howthe experimentalerrorstranslateintoerrorsin themodelestimatesa.udusethe


well developed
toolsof linearappraisal
(Section6) to estimateapproximate
reliability
of the non-linear solutions.

9.1 Assessing
the Quality of the Solution
9.1.1
Asfor linearproblems,
andassl•rn•ng
that ourdatad• arenormallydistributedabout
theirexpectedvaluesandwithknownuncer•x•uties
(orexperimentalerrors),•r• the
goodness-of-fit
is defined
by thestatistical
parameter
q,calculated
as

i----1
(9.].)
Forn independent
observations
andp independent
parameters,
qisdistributed
a• X•
with(n--p)degrees
offreedom.
Theexpected
valueofq isn bu•in practice
a model
withIn-p< Q<•n+x/2n
! isacceptable.
TheRoot
Mean
Square
(rms)
error
given
by

{__! 0'2

or simply
= ,----1
II II= fo=wdgh.d
solution.(9.2)

9.1.2 Parameter Re•lution Matrix


Theparameter
resolution
matrixofa ll,•e•rsystem
canbeeasily
callated •udthat.
wasonemotivationforre-casting
ournon-linear
estimators
in equivalentli,•earform,
for the consistency
anaJyses
presented
in Section
8.2.3.We will applythe straight-
forward
rule( R =/irA, where
H isthegeneralized
inverse
used)to thr• rl•erent
c•es. We will assumethat the data equations
havebeenstandardized.

Case 1: Uncons•ra•ed solution


Using
therelation
m= (AREA)
- •(WA)T{•},
i.e.,eq.8.17,
wefindthat

122
• = (AVEA)- •(WA)•. WA = _r (9.3)

If the inverse.matrixexists,/• = •, meann that the modelparametersmay be fully or


perfectly resolvedby the data.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Case2: Ridge regressionsolution


Using
there-interpreted
Ma•quardt
solution
(m= [AREA+•I]
i.e, eq. (8.25),whichusesthe previous
iterateasthe a priori h•formation,
we havethat

= (A•A + •)-•(A•rA + •) # •. (•.4)

Thus the Marquardt solutiondoesnot have a perfectresolutionmatrix.

Case 3: Maxi•-•-bi• solution

•e-wr•ting equation 8.19 as

(9.5)

whereC = fiD, h. = f•h, a•d d is az previouslydeSned,we find that

= (A•EA + CTC)-•(A•EA • CT6• = I (9.6)

showingthat wehaveperfectresolution by incorporating


a pt/or/data in oursolution
process.
The rn•-•rn,,rn-bias
algorithmusesa setof•udepende•t a pr•o•/•ormation (i•
theformof a null vector)•ud eq.(9.6)is thusapplicable.
in thissituation.Notehere
thatthea pr•or•estimates
axeazs•med
tobe•ndependent
ofthedat•,,,•]{•einthe
Marquazdt casewherethe a pr•o,'/dataat successiveiterations
dependon the
experimentaldata.In general,/1is evalu,•ted
at an acceptable
m• R maybe
interpreted
asa gaugeoœ theb•la•cebetween the•nformation provided
by thedata aud
thoseass•rneda pr•or{.If • -- l, theneachmodelpax•rnetermaybe well.deter•{•ed.
As in the linear case,the deviationof the rowsof J• from thoseof the identity matrLx,
I, measures
the lackof resolution
for thecorresponding
modelparameters.

9.1.3 Trend Analy• of Resid•


The distributionof the misfitbetweenthe observed
data and thosecomputedfor the
optimalsolution
mayrevealthepresence
ofa trend(frompositive
tonegative
residuals,
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

say)orthelackofit. Recall
thatweintroducedthisconcept
dm4.1•g
theaa•-lysis
of
linearinversionwith smoothness
constraiuts.
The followingau•.iysislean•heavily on
thatprovided
by Constatinides
(1987,Chap.7).To examine
amydistribution
patterns,
it is instructiveto plottheresiduals
aga/mttheindependent
experimental
variable
(e.g.,geophone
orelectrode
spuclugs,
recording
frequencies)
andsometimes
alsothe
residualsversusobservational
data.The formerwill be mostlyusedhere.A correct
modelwill showsomerandomness in the distribution
of residuals
andthe lack of
randomnessmaysuggestthatsome
partofthemodelmayhavebeenbetterfittedthat
theothersandtherefore
theoptimalmodelisnott•hecorrect
one.Thefourexamples
shownin Figures
9.1a-9.1d
havebeenselected toillustrate
some
a•pects oftheproblem
andpossibleremedies
aresuggested.In Fig..õ.lait would
appear
at a glauce
thatthe
precticted
dataexplain
thefielddataadequately.However,amalysis
oftheresiduals
showtrendsof increasing
residuals
in proportion
to d. In thissituation,it may be wise
tonormalise
theresiduals
bydividing
them
bytheassociated
data(Constantiuides,
1987), i.e.,

andthe matrixA is scaledaccordingly


sothatwearein effectsolving
a weigthed
system
ofequations
fortheparameter
corrections.
A different
situation
isseenin Fig.
9.lb wherethe residuahshowa randomdistribution
aroundzero.The modelis deemed
satisfactory
in thiscase.
In Fig.9.1cis shown
another
interesting
situation
thatmight
appear
asa good
fit atfirstsight.
A caref•examination
w/11
reveal
thepresence
ofa•
oscilla•ory
patternin thedistribution
oftheseresiduals
around zero.Thismightsuggest
inadequate
prediction •udtheaddition
ofa termtha•willintroduce
anoscillatory
behaviour in theforwardmodelmayimprovethefit to the dat•. In somecases,wemay
findthat the residuat.q
showa trendfrompositive
to negative asin Fig. 9.1cl.The
situation
illustrated
in •hisfigureimplylowprediction
of d at smallvalues
of the
independent
experimental
variable•udhighprediction
ofd athighvalues ofthe
measurement
positions,
say.A correction
tothemodelto remedythissituation
seems
to bewarrantedbut wemustbearin miuclthatmodelresolution
decreases
with depth,

12q
u • -)!(. -)l(-
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(d
ø- , ........ •-•t
d
½
) ••, .•,. ......
• . •t

Figure 9.1 a
Analysisof residualsshowingtrend of increasing
residuals in proportion to d

(dO.
de
)t *
Figure 9, lb
Analysis of residualsshowing a random trend

125
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

d•

ii
•t

(dø dc
' )0 •t

Figure 9.1 c
Analysis of residuals showing osciilatorytrend

d•

I I I I I I i i

(dø-d½},•

ii
- i i i i

Figure 9.1 d
Analysisof residualsshowingtrendfrom
positiveto negative
(Fig.9.1 afterConstantinides,
1987)

126
•ussian, uncorrel•ted
andofequalvariance•. TheSasicruleis that if rn•ø'-- Ld,
thenbylawofpropagation
oferrors,
Corm"•) = •[Co•d)]LT ($e•Meye•,1976).

For the unconstra•ned


(unbiased)solutiongivenby equation8.17, we havethat
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

= = (AEA) (0.7)

in the st•udardizedframework.If the data equationswerenot normalizedby the


observationalerrorsat the outset,the the right-baudsideof eq. (9.7) mustbe preceded
by o• (the covariance
of d) •ud the E termwouldnotbepresent.It maybe notedthat
for the unbiasedstatisticallynormalizedsolution,thereis a simplerelationshipbetween
the covarianceand resolutionmatrices, i.e.,

1{ = Cov(rn•ø')A•'EA.

For the m•nimum-biassolutiongivenby eq. (8.20), we •_d that

Cormø')= {(A•EA+ •I)- •(WAT)}[E-Cov(•)]{(ATEA+ f;2/)-•(WA)•}•

= (A•'EA +/•2I)-•A'rEA (AREA+ •2I)-• (9.9)

wheretheusualpre-multiplier
o• (fromCov(d)=•I) hasdisappeared
as• consequence
of the wei•htin• procedureadoptedin the biasedestimationapproach.
For inversion with reliable a pr/ori estimates, the covaxi•mcefor the magma,m-bias
solution(eq. 8.19) is

Two interestingsituationsmay developthat would affect the 6•al form of the above
expression.
If there axeno variationsin the a prior{ pax•meters(whichare as.•med to
tobeindependent
ofthefitteddata),theneq.(9.10)
reduces
to

Cov(rn•l"
) = (A'rEA+ D•'BD)-ZA•'EA
(AREA+ DTBD)-• .

127
sayin mos•geophysica• exploration problems.
If oneis irrevocably
committedto
statistics,
thereis a simpletestfor randomness
of residualscMledthe RunsTest.

9.1.3.1 The R.•- test


This is usedto test the r•udomussof a distributionof residuaJz.
Let n• denotethe
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

total numberof positiveresiduals,n.•thetotaln•rnex of negativeresiduals,andr the


n•_•rnerof timesthe sequenceof residualscha.uges
sign.The qu•utity,r is lmo• •z
the num• of r•u•. The distributionof r is approximated
by the normaldistribution
a.udits mea.uand standarddeviationaxegivenby

The stn.udaxdizedform of the variable is

whichis distributedwith zeromea• a•d.•ut variance.To test the hypothesisthat Z is


random,weneedto compare Z withthe normalstandaxddistribution.If the valueof Z
I•l]• in theregionof •cceptance
of thistest,thenthehypothesis
that themodelis the
correc•one and that the residualsare raudo•y distributedc•u be accepted.If, on one
hand,Z is toolow,themodelisinadequate;•ud a two-sided
•estmust'beperformed.
On theotherbaud,if Z istoohigh,thenthedatacontainosc•ationswhichmustbe
zccotmtedfor by the model.

9.2. Model Bounds


The determination
of boundi-g
valuesontheoptimalmodelpaz•meters
for a givenset
experimental dataandassociated
errorsis•u import•utpartoftheinversion
process.
In fact,it should
belooked
uponasa logic• completion
of geophysic•
inversion.

9.2.1 Parameter Cova•iance m•tziz


Thecovariancematrices corresponding
to thethree.c•esmentioned
abovewillbe
derivedhereusingthe st•udard
procedure.
The ass-mptions
madeaxethat the dataaxe

128
However,in practicethe matrix $ is lazgelyundetermined (B = $v/•) a•d the optimal
set of va/uesrequired to forcethe solutionto s•tis•7 the a I•riori d•t• exactly c•u only
be dete_rm•nedby tria/-•ud-error.It follows•hat for e•ch setof vMuesuseda• the
components of/•, theresulting
pointestimates
maybed/fferen•. Put simply,•hedegree
to whichthe optima/solutionreta:ms
the specified/•va•eswi•h/•. It is possible
to
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

describethese vaziationsst•istica/ly a•d determ•uethe correspona•gcovazia•ce


m•trix, but we shallchooseto deduce•u equivalentmeasureusing• simpleintuitive
concept. RecaJlthat in theorigh•a/biasedest•m•tien
formulationsthe d•t• equations
wereweightedby theobserv•tiona/errors byw•y of thematrixW. In the same
formulations:the extr•ueouscl•t• equations
corresponding
to ou• a priori infom•tio•
werepre-multiplied bythem•trJx$. It iseasyto seethatthisoperation
is • weighting
procedure andthatboththeRr a•d/• matrices servethes•rne•unction
forthetwo
k•ucls
of cl•t• considered
in theproblem formulation.
Now,by cle•u•tion,
E-- w'rw,
audthereforeCo•(d) is simplygivenby E-•. The cova••ce of the variationsin A
isidentica•yB- • -• [$v$]-•. Theeffectof thesevariations
onC'o•m"*)in thesta•d•-
ardizeclf•amework may thus be •ppm•_'rn•tedby

Co•(m•")= (ATEA+ DTBD)


- •DTBD(ATEA+ DTBD)
-• ß (9.11b)

We m•y thendetermine
thefull cova.,'iance
relationgivenby eq.(9.10) fromthes,•rn
the two setsof covariances
definedineq. 9.11aand 9•11b.This.is simply

C•m(m'")
= (ATEA+ DTBD)- •[ATEA+ DTBD](ATEA+ DTBD)-•

= (ATEA + DTBD)- •. (9.12)

I• t[ollows
that for the ma••,,•-bi• solution,R andCov(m'ø•)arerelatedin the form

1t = C•.•(m"•)(A=EA+ DTBD).

Thesquare
rootsof thediagonal
elements
oI Cov(m
"•) maybetakenasthestandard
deviations
oI the leastsquares
paraxnetet
e•ima•esa•d maybe usedto estimate•he
botmdsof •hemodelparameters. In termsof confidence
llmi•, •he Covatiance
matxix
allowsusto define• con•dencehyperellipsoid
(mTCov(m•)m> 0 ][o•anym), in which
weare(100-a)%certain
thatthesolution
lies.Weusually
clefme
the95%co,•dence
interval by choosinga <_5.

129
9.3 Extreme Paxameter sets: Mos• sq• Inversion.
The iterative most squaxes
methodof extrema/inversion(Meju andHutton, 1992;
Meju, 1994d)will be adoptedasthe m•instayof ourmodelappra/sM.
We will determ•ue
thesolution
fora vaxiety
ofpractica/situations.
Theproblem
de•tions •re thesame
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a.sfor the lineaxc•sesaJxeady


considered
a•d the detailswill not be recountedherefor
brevity.Essentia•y,havingobta/ned•n optimalleastsquaxes
solutionto •u inverse
problem,
rnweextrern•ze
theobjective
fimction
mrbsubject
totheconstra/nt
thatthe
residualsq ( somecombination
of the dataprediction
errora•d solution
length)arenot
greaterth•u somethresholdvalueqT,wherebisthepaxmeterprojection vector.The
expected vMueof qTisn (then•mberof data)or n-I whenthesolution is effectively
constra/nedto satisfyl knownpaxameterestimates.We will chopthe weightingterm
herefor notationalsimplicitybut beaxin mindthat the data equations
have been
sta•daxdized.

Ca•e 1: Unconstr•ed problem


To ca/culatethe most-squ•esmodelsfor the Gauss-Newton
solution,we needto
extremize

,• +• [(•- f(,•))•(•
- f(.•))- q•] (•.•4)

This may be donein successive


stepsa.sfor nouliueax'least
squ•resmiuimi•.ation.
Linearization
oœeq. (9.14) aboutaniuitia/modelwill yieldthefunction

(.•0+ •)• +• [(•_x•)•(•_•) _


whichis •o be m/••zed or ma•m•.ed to yieldthe pax•metercorrections
•: •o be
applied
tomøuntilthecondition
q= qu:issatisfied
whereq= I d-f(m)I •. Themost-
squares
solution
forthepa.rxmeter
increments
issimply(Mejug•Hutton, 1992)

(9.16)

where 1

( qw-qœ$
• (9.17)
•,= ñ
a•d is determined
bytheconstra•tq= qw,a•d qcs
= yw•I-yWA(A•A)
-

•3o
Usinõthe optima/leastsquares
modelasourinitia/model,the relevantiterative
inversion formula is

m•:+• = m•:+ (A•A)-•(A•y.-•b) (9.z8)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

where
mkistherefined
model'it
iteration
k,andA andy axeevaluated
•t mk.The
expected
va/ueof qTisn, them•r•ber
ofthedat•.Notethat.ifwepreferred
to solve
explicitly
fortheparameter
estimates
rather
thantheperturbations,
thenwehave
that

,,-,= (A•A)- •(A•_ •) (9.•)

where• = (y + Arnø)andmaybeused in aniterative


fashion.
However, the•bove
inversion
formulae (9.18•ud9.19)are-•table •lgorithr•especially
if ATAis
ill-conditioned.
For this reason,Meju andHutton (1992)suggested
d•,•ping the
•bsolute
v•luesofthep•r•r,eterperturbations.
Thisand'
otherprescriptions
ford•g
with instabilityin iterativeapplications
are discussed
next.

Case 2: Damped solution


We alreadyknowthat theprocess
of damping
involvesaddingœctitious
(or extr•ueous)
data to the actualobserveddata and that d•rnpingcanbe achievedin two ways.Thus,
we can either define the statistica/messure q as

q= (,•_.f(,,,))%•_.f(,,))+
•(,,,_,,o)•
(..,_,,o) (•.20)

in line with the Marquaxdt-typeappros•Jx,or •s

q= (d-f(m))T(d-f(m))
+ ,82mTD'rDm (9.21)

in accordwith the biasedestimationphilosophy;


andin both caseswe have placedsome
boundon the sizeof the parameterperturbations.
Considering
theformerdefinition
of q (eq.9.20),thelinearized
qu,xutity
to be
minimizedis (cf. Meju & Hutton,1992)

(too+
•=)T•,
+ • {(•_ •)T(•_ A.•)+ •=•=- qT}ß (9.22)

E•ectingthe minimization
in the usua/m•er, wehavethat

131
or
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

[ATA+ ,fI]z- (ATy- (9.23)

•rom which we obtain the dampedmost-squaxes


parameterincrements

• = [ArA + •I]-•(A•- •) . (9.24)

Whenthequadratic
constraiut
( q= qr ) issatisfied,
wefindthat

•T = f• + =T[ATA+ /•=- 2=rAT•

Or

qT= yTy+ p2b'r(AT


A +/•I)-•b- y'rA(ATA
+/•I)-•AU:y (9.25)

h-ore wkic2 we obtain


1

P= -(qT--yTy+yTA(ATA+
T q-fiI]- •b•I)
bT[A -1ATy)

1

= q-[,b:r[ArA
+/•I]-• ß
Nonl•earityis deal•withusi.ug
theiterativeformula

,•} + • = m• + (,4T,4+ •)- •(A•y - (9.27)

where
A m2d
yareevaluated
atmk .
If weelectto solvedirectlyforparameter
estimates
ratherthanx, wehavein this case,

(9.28)

Adopting
thealternative
biased
estimation
strategy,
thelineaxized tobe
function
mjnm•.ed is simply
z = (,•o+=)=•+ •• {(•_ •t•)•(•_ •=) + f(•o+•)=m=u(mo+.)_ •=}. (9.29)

Differentiating œwith respectto z and settingequal•o zero yiedsthe normal


equations,
[•• + ••]• = (••_ ••.•0_ •)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

from whichwe can obtainthe solutionfoz the relevantpaxa•eter perturbations

(9.30)

whereH- DvD. If wechoose


to solveexplicitlyforthep•rameterestimates,
we fred
that

(9.31)

whered is aspreviouzlydefined.Sinceeq. (9.30)mustsatidy eq. (9.21), we have that


1

p= -4-bT[ATA
+ fhH].•

whereqJ•s
= drd + • 2reøTHinø
- (YTA - •2mrHr)(AZA+•2H)- •(AZY-f•2Hmø).

Nonlineazity is dealt w/th usingamiterative schemeas in' the previouscase.

Finally, if it is desiredto retain amyavailablereliableprior estimatesin the extremal


inversion, then we have to r-•rn•ze the/unction

(9.33)

The solution •s therefore

ß = [A•A + •• -• {A• + S/)•(h - •o) _


or equivalently
m= [ArA+ BH]- Z{Ar•+ BDrh- pb} (9.34.)

whereB = fir/i,

133
Most squares flow diagram

INPUT
ii

LEAST
SQUARES
MODEL
(too)
I
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ANDRESIDUAL
(%)

INITIAUZE
VARIABLES
• = qoxFACTOR,
j=0 iii

j=j+l,mi = mo

FORM PROJECTION

VECTOR,
bI

CALCULATE JACOBIAN AUGMENT


A
Awith 8D

CALCULATE .SVD OF
GET A

2
I---
X
I.IJ
OBTAIN
I.•,
Xd x

CONTROL
RATE
OF
I.i.I

CHANGEOF mi

i i i

X NEW ESTIMATE
NO
m ! = mj + Xd
q2qT
CALCULATE RESIDUAL
q =ZId - f (rn])12
YES
ß

NO YES NO CHANGE SIGN


SOLUTIONS
OBTAINED of I,L,
SETmi = mo
YES
YES
STOP
2

Figure 9.2
1

(9.35)

and

-(3/'rA- Bm•'H'r)(Z'rA
+ BH)-•(A•'y- BHmø).
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Themostsquares algorithm ca• be readilyimplementedaada flowdiagramof an


iterativeschemeusingthetwomethodsof dampingdiscussed aboveis shownin Fig.
9.2.Noticethat the fight-hand sidesof equations9.24and9.30containthe daxnped
leastsquaxessolution forall theparameters andan additional
termthat is specific
to
theparameter beingextremi•.ed. The methodusesa compensating relationshipbetween
theparameters andmaybeinterpreted asproviding theconfid•celimitsof theoptimM
leastsquaxes
model(Meju& Hutton,1992).Provided
thatq?> q;s, thereexists
two
solutions
for p for eachpaxametex;
a•d sincenoneof the modelpaxamege•s
is kept•txed
dtwingeachstepoftheiterative
process,
themethod
in effectgenerates
.
2p diCereat
models
forthe'pmodel
pax•,•eters
thataxeconsistent
withthegivendata.Thus,the
methodmay be usefulformappingout thepaxameter
space
for overdetermlned
n6nlkneax
problems.
Anapplication
ofthis'method
ofanMysis
totheCOPROD
{Jones
& Hutton,1979)magnetotelluric
datais shown
in Fig. 9.3 forilluztration.The
maxim-m
permissible
residual,
q• wassettOthen,•,•ber
ofdataa•dtheresults
show
that the modelsaxenot constrained
by the data in the top 10 krn of the Earth's crust
andat depthbelowabout300kmwhichagrees withPaxker's(1982)maxi•,,,• depthto
whichanymodelcaxtbeconstrained by the COPRODdata.It is obviotmthat themost-
squaxe•models t•ow some lightonthevariability
ofmodelspace andshowwhat
featuxes
of the modelsaxeimportantor consistent
with thefieldobservationa.
It maybe
notedalsothat in•teadof extremizing
eachsepaxate
parameter,
onecaa obtaina
solution
envelope
bysetting
theprojection
vector
azb= [1,1,...,1]
T (Jackaon,197õ).
Thiswouldyieldtheupperandlowerbounding modelsfor'theleastsquares
modelas
demonstrated
in Fig.9.4ausing magnetotellm'ic
data.Thisprocedure maybeuz•
whendeMirigwithsmooth, densely
paxameterisedmodels(eg.Constableet al., 1987;
Meju& Hutton,1992)asillustrated in Fig.9.4b.Notethatin gen•at•g thesmooth
models
shown in thisfigttre,
theEaxthwaspaxaraeterizedintoa succession
oflayer•of
equalthickness
(inlogspace)
andthatthesmoothnss
measures
wereonlyapplied
to
the layer resistivities.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

136
I
i

iN•Ba•B
II I I | i iI

['õep]
!•1
'

RSt:II-IJ
T
o

o
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/


Mmm
m

137
I--!..L
•':! a

(-6ep)
i I I i I iI

i
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

.%

t i i i i iI

u,I
,, i
i

i38
HJ.a•O
i
i
i

ii

,]
i
! i I 1I

['Bep]
i
' i
i
i

i
i
i
1
i

ii
i

•]g!:lNc:!
10. SAMPLE SPECIALIZED APPLICATIONS OF ••'E•E

10.1 Processing of Time- or Space-Series


Data
The Eazth has soundsfor thosewholistenbut the musicis alwaysdistortedby some
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

unwantedtunesandinverseproblemtheoryprovidesuswith the •ters for enhaaciag


and understandingthesesounds;
geephysicists
havem•steredsomeoœthesesoundsand
aze thereforein tune with the Earth. This is the •uthor's favourite descriptionof
geophysicalsignalprocessing
to the,•n•n•ia•ed•ud •hefull importoœ
•his simp•g
a•se•:tion
will becomeobvious
in •hefollowingdiscussions.

10.1.1 Adaptive Si[•n• Pr•.essimg.


Geophysicaltime- or sp•ce-series
clat• azetypicallybedeviledby noisefrom various
natural and artificial sources.These undesirablephenomenamanifest as interference
ß

noise•random noiseor nonlinear• in •he t•rne:seriesd•ta and .weWishto remove.


them or mi•ze their effectsusingthe toolsprovidedby inverseproblemtheory.
If the noisechaz•cteristic$(or somereferencedesirablesignals)a•e known,we could
elira•nate•he noisein the cl•t• by simple•geriugor by •d•ptive noisecancell•tio,•
operations
(e.g.,WidrowandHoff,1960;Wickow
et al.,1967;WangandTrei•el,1971).
Unfortunately,in typical geoexploration
situationswe do not know what the t.rue signal
is o• what constitutesnoiseand processingthe resultingtime-seriesdat• becomesa
3•cult task and the main goalis to suppressthe noise•s best as we can. We will
explainthe undergirdingprinciplesof •d•ptive signalp•ocessing followingsomeuseful
definitions.If the st•tistica•properties(i.e., mean,vazi•ce, correlationfunctions,etc)
chazacterising the d•ta do no• vary with •ecorrl•ugt•rue,then we m•y describesuch•
systema• beingst•tionazy.If however,.thesestatisticalcharacteristics vary with
measurementtime, the systemis s•id •o be non-st•tionaz7.

Ma•h,•m•tical Model fo• • S•ationaxyLine•


Consider• lineaz time-invazi•t but otherwise,,•kown systemwhosesampledvalues of
th•output
y-- (Y0,
Y•,Y•,---,Y•)•
aze•el•ted
tothecliscre•e
system
input
z•,...,z•)r = X • andtheimpulse
respo•efunction
• -- (w0,•,w•,...,•) • bythe
convolution

Y1-- • wiz.i-•-- wTx, i----0,1,2,


i:O
..... (10.1.1)
A seismictrace, for example,ca• be approximatedby the above'•ite discrete
convolutionand may be interpretedas representing the convolutionof a reflection
coefficientserieswith a reverberatingpulse•rain. In any case,it shouldbe noted that in
addition•o the usualinput signalz•; we needto supplyan ex-tran•• input signal{i.e.,
a.uestimateof the desiredresponse) d• containingthe adaptive61terclu_,iug
the
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

adaptationprocess,a kind of a priori information.

Inverseproblemformulatic•
The inverseproblemis simplystatedas:Giventhe signalsn:ia.uda• estimateof the
desiredresponse d•, find the impulseresponse w that is optim•-Iin the leastsqu•res
sense(i.e., the set of weiõhtstviyieldinõthe bestoutput).
'Let us de_•e the error betweenour desiredresponse d• (representedby the aAditional
inputsignal)andthe systemoutputy• (/or the weightestimates,
w•) as

ei = di - !ti

or simply

a= d- w'rX = d - XTw.

We wish to minimize the S111TI


of squaresof this error series,i.e.,

e•:e= (d- X'rw)X(d- xTw) = dTd- 2dX'rw+ w•rXX'rw.

The expectedvalue of e2, givenw is

e: = E(c•e) = E(dTd)- 2u•E(dTX) + w•E(XXT)w

whereE is theexpectationoperator.
Let the correlation
vectord'rX bedemot•by •a.
or simplyP •ud'the autocorrelation
matrixXX T bedenoted by •b• or R. Azs•mi•g
that •d. aad •.. approximate
theixexpectedvaluesfor this stationaryprocess,
we have
that

Miurni•.ationis effectedby settingto zerothe gradientof this quaxtratic


hmctionwith
respect to w.
Es•rn•t•õ gr•l•ents and
The error •dient is s•mply

v(• 2) = -2.ea=+ •=• = -2(•= - •=•o) (10.1.3)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

•nd equatingit to zerofor• mi•• le•d$.to thesolution


œer
•heoptim• weight
vector

(lO.1.4)

wkichis •he well-lmownWiener-HopIleastmenusquaxe(LMS) error•ter a•d ca• be


usedto find the min•rm•rnof the error surfacethroughsuccessive
ß
adjustmentof the
weightsw. Note that this•lgorithrncaz•be implemented in.differentwaysusingmatrix
inversionor gradient-type
schemes. An iterativeimplementation usingthe methodof
steepest
descent
givestheadaptive
process
•s (e.g.,Widrowet al., 1975,1976)

,,/+' = • + •(-v•) (•o.•.5)

wheretheestimated
chauge
in • isproportional
to thenegative
ofthe errorgradient
•ud • is • sc• co•t•t that controlsthe convergence •d stubbly of the
Substitutingthe e•ression for the ••ent vector(10.1.3) into t•s iterative Mgo•th•,
•el&

uP+• = u/+/•(2(d- x'rurf}X) = u• + 2f•e• (10.•.6)

whichis the popularWidzow-Hoff LMS algorithm' and doesnot requirethe calcalation


of •ba=or •==. The usehzloutputof thisiterativefilteringprocess
(i.e., the cleausignal)
is the error series•. It canbe shownthat the Wickow-Hoffalgorithrnconverges if
assigned
• value
between
0and•, where
;kisthelargest
eigenvalue
oftheinput
auto-
correlation
matrixXX 'r. Thealgorithm
alsoconverges
if
wheretrac•(A) is the s•rnof the elements
in themaindiagonalof A (Wang,1977).

A vazi•t oJ•the •bove adaptiveMgorithrnis'obta&ued by premultiplyingthe estimated


error gradientat e•ch iterationby an estimateof •he inverseof the •utocorrelation
matrix yielding(seea•so,Widrowet •1.,'1975,1976)
= + = .d +
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

which is an iterative inversionformula incorporatingsomekind of a priori data.

Tlme-vary/.ug LineaxSystexn• with Non-StationaryInputs


The task of suppressing or removingtmwa•tedsignals from geophysical time-seriesis a
difficult one sincethe data often displaynon-stationarycharacteristics
(suchthat their
expected
valuesaxenot constant).
Several
approacheshavebeendeveloped to address
suchproblems
(e.g.,Wang,1977)but wewill'discuss
a simpleandeffectiveadaptation
of the Widrow-HoffLMS algorithm-- the correlatedadaptivenoisecancellation
(CANC)algorithm of Hattingh(1988).
Here,therequireddatasetsaretheprimary
signalcontaining
thenoisethatis to bemizdmizeddi, anda reference
signalxi- Note
the notationalchazxge
fromthe previoustreatmentof adaptivenoisecaac•tion. l.a the
CANCalgorithm,
thereference
signal
inputio thedataadaptive
filter,X isactually
a tappeddelayt•meversionof the originalreference
signalwith the time delaybeiag
equal•o the filter lengthL, i.e.,

X = (xi, 2i_l, , 2i-L+l

whereL <_i <_n. As before,the referencesignalis merelyusedas a leaz•i- signalto


adaptthefiltercoefficients.
Thefilteroutput
at theithinstant
forthegiventoisthus

and •he error hmction to be mi•imi•.ed is simply

ei----d(i- 5/2)-- Yiß (10.1.9)

In applying
theWidrow
ad•ptive
a/gorithm
ur/ +t: ud+ 2fle•, wenoteherethat
1 <_j _<L. Therecovered
signal
fromthisoperation
comes
fromthefilteroutput'
Yi,aJzd
the extractednoiseis contained'
in the errorseriesei. A FortraI• translationof the
single-cha-,,el
CANCalgorithm
givenItattbgh(1988)
is givenin Fig.10.1.1a•d canbe
easilyextended
to bandiemulti-cha-•elLnput.
subroutine
CANC(d,xx,y,nd,ns,L,beta,itern,iscale)
c correlatedadaptivenoisecancellationroutine
c adaptedfroma• originalHP-Basicprogram
by Hattiu•h,
c Computers& Geosciences,
14, 467-480,1988.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

c d = primarysign•; xx = reference
signal;y = outputsignal
c X = delayed
referenesignal;w = filtercoefficients
(weights)
C e = error signal
c Ils = actualnumberof samples in thetime seriesd
c nd = physicaldimensionof thevectors d & xx in controlroutine
c L = filterlength(typicallyabout2 timessignalbandwidth)
c beta .= convergence{actor(typic•y 0<beta<<1)
c item = n-mber of iterations to be perfo•ed
c iscale= 0 (noscalinõ)
or I (eachsignal
normalised
by lazgest
coefficient)

reald(nd),xx(nd),y(nd),X(nd),w(nd),e(nd)
c initi•zations
iter=0

L2=L/2
do i=l,ns
x(i)=0.0

y(i)=0.0
,(i)=0.0
end do

calculateand removemeanfrom prim. •ud reI. sign•l.•

a(i)=a(i)-aean
xx(i) =xx(i)-xmea.u
end do

c note,at this stage,that for additionalstab'•lity,


we may employa
c scalingoperationof the signals usingthelargestcoefficientof each
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

c time-series.
fin•] outputmaybe de-scaled
asrequired(notimplemented).
if(iscale.eq.1)then
c find largestsamplevalue of eachsignal
bigd=d(1)
bigx=xx(1)
do i=2,ns
if(d(i).gt.bigd)then
bigd=d(i)
end if

if(xx(i). gt.bigx)then
bigx=xx(i)
end if
end do

c normalise the signals


do i=l,ns
d(i)=d(i)/bigd
xx(i)=xx(i)/bigx
end do
endif

c =. endof scalingOperations
**
c apply CA1NICMgorithra
10 continue
.

do i=l+L, ns
c form delayedreferencesignal
do k=l,L

end do

c filter delayedreferencesignal
do j=I,L
y(i)=y(i)+w(j).X(j)
end do

c delayprimaxysignalwith l•g L/2 andcalcttlateerror


e(i)=d(i-L2)-y(i)
cMcul•tenewfil•er coe•cientusingthe Widrow-HoffLMS algorittm•
do j=i,L
w(j)=w(j )+ 2., beta,e(i), X (j)
end do
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

end do

c update countfor next iteration


iter=iter+l

if(iter.LT.item)then
c zero filter output for next iteration
do i=l,ns
y(i)=0.0
end do

repeat operationsas specifiedby item


goto 10
else

compensatefor phasesh•t
do i--1,ms
if(i.LE.L2)then
y(i)=0.0

y(i)=0.0

else

y(i)=y(i+L2)
e(i)=e(i-•L2)
end if
end do
end if
return

end

Fig. 10.1.1 A Fortran CANC routine.

145
10.1.2 Iterative Deconvolution
of Seismogramsf•om PdppleFired Shots
The everincreazingenvironmenta/restrictions
onthe generation of highamplitudesur-
facevibrationsnowmake it d•cult for lazgeshotsto be fired in longrangecrustal
exploration
seisinology.
Theenergy
densityof a shotmaybereduced
by firinga number
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

of smallershotsspreadout in time that togetherwouldgenerateeffectsthat aze


comp•ableto a lazgeshot.Thisis knownastipplefiringofexplosives •nd is nowa
standazd practicein ma•y rockquazries.
Pdpplefiredshotsgenerate seismograms
conta/ning overlappingpulse-shapedevents(interfering
arrlva&s
of vazious
phases f•om
clifferent
shots)but the process
c•u bemodelled mathematicMly. Fromsuchcon•usecl
signMs, onehopes to deconvolve
lessnoisysignMs representing
theseismograrn that
wouldhave beenrecordedfrom a singlelazgeshot.This is a nonlinearinverseproblem
a.udit is possible
usingwave-shaping
aleconvolution
•ud inversion
techu•ques
to recover
a stacked,average sourcesigna/combining
eachincIividua/shot.
Thema/nusef•,l•ess
of
inversionhereis in opt_•rn•.ing
the input spiketime seriesfor the wave-shaping
deconvolution.The spiketime seriesis then deconvolved •rom the wholefield recordto
givea filteredsource
signa/witha muchreducedsignal-to-noise
ratio.Notethat a field
seismogr•rn c•uberegardedasa nonlineaz
ftmction
ß
ofthepositions a•d aznplitudes
of
the spikeinput.Ottrmodelis'theone-d;mensiona/convolution
of a sourcewaveletwith
a spiketime seriesof a T•te lengthplusadditivenoise,i.e.,

Then•mberofspikes in theinputs.ignMcorresponds•o then-rnberof shotsfired


the time separations
of the spikes
ß
corresponds
to thet•rnedelaysbetween •rst arriv•
f•om d•erent shots• • receiver.The time delaysconsistof the actual5•q.
o
ng delay
about25m• andthe delayowingto the differentpositions
of the shots.

Field recordsoœ
seismogr•-•.•
wereobtainedin an explosion
seisinology
experimen•using
quaz•blasts
at ChaZmwood
Lodge
in Leicestershire,
Englaud
(P.Magui•e,
1991unpub.).
I• wasintendedto fi• ghedatafor thesetipplefiredshots.The positions
•ud aznplitudes
of thespikesin theinputmodelwerechosen
asthemodelparameters.Themodel
responseis thesynthetic
seismogram
producedbyconvolution.
Thisinverseproblem
wassolved usingtheridgeregression
procedure
outlined
in section
7.4. Oneofthefield
records•ud thesta•ing modelfor the i•erativedeconvolution
a•e shownin Fig 10.1.2.
TheinitiMspiketimeseriesconsists of equispacedspikes
of equala.•plitudeaaxdis
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

z
rr

-1-

g'O
0

epnl!!d•v
g'o-

•E

s-o
(--

o
\
\

g'o-
aPnl!ldu•v
o
•E
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

LL!

L.U

I.
t

s'o
epn•,!ldLu¾
d

148
Z

LL!

LL!

oo •
I

OS

epn•,!lduJ¾
d
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

5'0
0
S'o-
•Pn•,!!dLu¾
z
I-
Z

>-
I. IJ
ß

•'o

o


•'o-
epnl!ldLu¾
o

o
ø
-•E
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

009
001•
epn•,!ldLu¾

150
LLI

•.OL.•
0 •-OL
epn•,!!dLu¾
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

..
-
I
I
' '1
i'
I
I
I
I
i
i

151
I
I

$JOJJ•3peJBnbs jo Luns
'1
I
I '
I
i

c)
o
showntogetherwith the spiketimeseriesalterthreei•era[ions
in Fig.10.1.3.The model
estimates •ter [hree i[erations mudafter 20 iterationsare show• in l•ig. 10.1.4. No•ice
the semblanceof theseestimatesto [he actualfieldseismogram
shownin Fig. 10.1.2.
The final spiketime seriesandthe deconvolvedstacked
signalareshownin Fig 10.1.5.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

The convergence characteristic


of thealgorithm
isshownin Figure10.1.6in orderthat
we cmuappreciate the effectiveness
oftheprocedure.
No•icethat[he sumof squares of
residuals
haddecreased
bymore
thantwo-fold
•oanacceptable
level
in12i•erations
a.udthat the improvementsalter iteration12 arenegligible.

10.1.3 1tesidual Statics Estimation


Whenprocessing
seismic
reflection
datait isimportant
to account
forthevelocity
va_•ia•ions
in the weathering
layeror wemay endupwith a distortedpictureof the
subsurface.Static •ime muom•lles
dueto the weathering
layer m•nifes• in the form of
signalmisaiignment
ontheseismic
traces
andleadto degraded
waveletcharacter
in
common-datum-point (CDP) stacks.For reliablestrnctu•alinterpretations,
we needto
estimatethe optimaldistribution
ß
of time shiftsrequiredto correctfor suchstatic
anom•es. Operationally,bulktime shiftsareappliedto individualseismic tracesto
compensate for thesenear-surface
velocityvariations.
However, a•er theb•lk time
shifts,residualtimedifferences
betweenthetraces may'persistsuchthat thedesired
reflectionhyperbola mayshowrandompatternsleading to scattering
of eventsalter
normalmove-out(NMO) corrections. H the timeshiftis smallin comparisonwith the
spreadlength,theappropriate eventsmaybelinedupin muoptimalwayby estimating
relativeshiftsalter NMO corrections
usiugTrace1, saymudthe windowsof strong
reflections.
This crudeapproach•o residuMstaticsmualysis
knownascoherency stac__k-iuõ
is notgoverned
by anyphysical
law.No•e,however,
thatreasonable
results
maybe
obtainedif the time shi•s are smallwith respectto the spread.length.
ln{ormagion
aboutthe near-surface weathering zonemay alsobe6tmishedby upholetestsin a
borehole
•hrouõh
theweathered
layer,
refraction
seismics
orrefractsion
firstbreaks
on
the reflection
records(thismaybea problem sincetherecording arraysin reflection
surveys aretypicallydesignedto eliminaterefractions).
It is possible
to setup a system
of equations forestimatingstatictimecorrections
for CDPseismic reflectiondata
(Ta.uner
et al.,1974).Wiggins
e• a1.(1976)
cast[heproblem
asa linearinverse
problem
andthe followingmu•-lysis
derives{romtheir lmurlmmrk
paper.

152
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

153
10.1.3.1 Problem formulation for linea• surveylines
Considerthe ray path from sourcepositionSi to a subsurface
reflector a•adthenceto a
surfacereceiverR• asillustratedin Fig. 10.1.7.Assuming surface
consistency (i.e.,
that the staticshiftassociated
with surfacepointS is independent
of the ray path and
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

of whetherS is • shotor receiver- provided thatdatumcorrectionsh•ve beenapplied)


andsubsurface
consistency
(i.e.,residual
NMOisassumed
to beindependent
ofoffset),
the total travel time from shotloc•fionSi to the buriedreflectorat the k• CDP to the
receiverlocationRi (i.e., a givenreflection
eventonthe ij trace)is givenby

whereTii is thetraveltimeforthechosen
horizon
ontraceij, Si is the traveltime
from•hesmtrce
tothedg•um
plane
fortheit• shot,
Ri isthe•ravel
timefromthe
receiverpositionR• to datumplane,N• is the normalincidence
two-way•raveltime
fromdatum•o t•hereflector
at thekthCDPposition,'M•
is•hetimeaveraged
residual
NMOandxidisthedistance
fromtheithshottothej•areceiver.
The aboverelationholdsfor everysource-receiver
combinationalonga surveyline. It
followsthereforethat for ns shotsandntre.ceivers,
therearen = nsxnt suchequations.
The desiredtime termsfor staticcorrectionsare$i,R•,N•: andM• andwe canposethe
problemin theformd =Gm whered istheyector
ofobserved
traveltimesTis,rnis
the vector of the soughttime termsand G is the couplingor coefficientmatrix for the n
equations.
Fig. 10.1.8For
for aillustration,
seismiclineconsider
shotoff-end.
the simple
The system
source-receiver
consistsof
configuration
four shot points
shownin

( i.e., n, = 4) andfourreceivers
(i.e.,the numberof tracesper shot,nt = 4) andthe

- = shot point
ß = receiver point

$4 83 S2 Sl
' R? 1:t.
6 l:[li R,4 1:1,
3
ß N•o ' Ns .... Ns ' N1

Fig. 10.1.8Illustrationof the CDP geometricalrelationships


(a•er Wigginset al., 1976)

154
shotsp•cinõis equalto the constantreceiversp•dng. We •hushavesixteenobserved
datavalues(•-- 16).It isobvious
fromFig. 10.1.8that therearesevenuniquereceiver
positions
(i.e.,nr-- n•+nt-1)
andtenunique
CDPpositions
(ng--nm--n•+2[ns-•])
sothai •hereare thirty-one(p •-ns+nr+n#+nm) modelparameters
for •hissimple
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

problem.
The aboveproblem
canbeposed
in matrixformwith

m -- [$•, ..., $4, R•, ...., R?,N•, ..., N•o, M•, ...,/Id•o]•

and the coefficientmatrix G as givenin Fig. 10.1.9.

ß . 1 ß ß ß 1 ........ 1 ......... 16 .....


ß ß 1 .... 1 ....... 1 ........ 9 ß ß .
ß .• ..... i ........ i . •...
..• ... • ........ i..i
... i ...i'. ........ • ......... •...
ß .. 1 .... i ......... 1 ......... 9..
ß .. 1 ..... 1 ........ 1 ........ 4
...• .- ..... i ...... i..i ......... i
T T T T T '1' ? T
S•...S• R• ... R• N• ... N,0M• ...

Fig. 10.1.9The G matrix for the probleminvolvingthe CDP geometry


shownin Fig. 10.1.8.The dots correspondto zerosaud the col-r-• of G
a•e identifiedwith the corresponding
parametersOfthe problemusingthe
usual symbols at the bottom(afterWiggins,et al. 1976).

The problemis undetermined sincen < p. For sucha short spread,increasingthe shot
pointswould increasethe indeterm•ncy. For instance,if ns is increasedto eight, we
have that p=55 and n=32. The problemis saidto be ill-posed•ud needsto be
regula•ized.
Note.however,
that if weemploya 24-tracereceiverg•oupsystem,the
problembecomesoverdeterminedforgreate••h•u threeshotsasshownin Table10.1.1.
It is stressedthat even in this o.verdeterm_
ined situation, there is still some indeterm-
inacyin the setup astherearenotenough
independent
equations
to constrain
the
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

individual parameters.

nt ns nr n9 p n

24 3 26 28 85 72
24 4 27 30 91 93
24 12 35 46 139 288
24 24 47 70 211 576

T•ble 10.1.1.Survey designparameters

Problem solutionaud usetiffcompuiationalstrategies


For suchunderconstrained problem.•,
we needto usea constrained
inversionmethod.
We may obtain the lea.stsquaressolution

m = [G•G + •]q- 'Grr•

usingeitherthebiased
estimation
•pproach
asin thecomparable
delay-time
problem
treatedin Chapter5 or by directaugmentation
of the singularvaluesof the problemfor
problemsofmoderatesizes.
Because of•hela•gesizeof equationsinvolved
for•ypical
fieldsurveys,
the Guass-Seidel
iterativemethodofsolution
is oftenemployedin the
solutionof the abovesystemof matrixequations.
While i• maybepracticable
to
augment
thematrixG wiihD = I, •heremaybeotherstralegies
ihatwillcutdown
•he
sizeof theproblem.
Forexample,
M• is,,,,likelyto varygreatlyfromoneCDPposition
to auo•herandit maythusbe usetiff
•o average
•he coefficients
oversome"distance
window"
along
asurvey
line(Hadron,
eial.,1986).
Thus,
coe•cients
Mi •oMi+•may
beaveraged
(i.e.,reduced
•o onecol,,tn-)to giveM• simplybyadding
iogether
col,•r•,•.• i to i+l oœthe matrix G.

10.1.3.2 Est•-ation of Three-,•,-•onal ResidualS•cs.


Whencomputing
residual
statics
fornon-linear
recording
linesof3-Dsurveys,
theeffect

156
Etleclive
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Survey
Line

di••
C C' Y

!
,,a
o• dip
direclion
R'H
=YijSinl,
s Atij2K'HDkYij
z Vk
(a)
(b)

Fig 10.1.10(a)A deviatedsurveylineS-G'anditsprojection


alongtheeffective
survey
line S-G. The distancex is the source-receiveroffsetand its midpointis C'. The midpoint
oftheCDPgatherisdenoted
byC. Theanglebetween
thesource-receiver
direction
and
the true dip.directionis denotedby ec.
(b)Across-section
taken
perpendicular
tothenonlinear
survey
lineattheKthCDP.
For
.thesinglecross-dipping
planarreflector,
Atiiisthetimedifference
inreflection
time
betweena tracewithitsmidpoint
on theeffectivesurveyline(normalto the planeof the
figure
anddenoted
byC)andonedisplaced
a distance
YiJ
from
it(denoted
byC').I;D,
is
the apparentreflectordip.
of dip on the residualNMO term M• andthe effectof deviationof source-receiver
midpoint
fromthek• CDPonthedipmoveout
termshould
betakenintoaccount.
As
in the 2-D case,we assume
that staticcorrectionsaxesurface•ud subsurface
consistent.
The observedtraveltimefor any.givenreflection
eventon the tracebeingconsidered
is
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

thengivenby thesumofsixterms(Zai-tia•,1988)

Ti)•:- Si+ R•+ Ni:+ Mi:z]+ M•:z•+ (10.1.12)

whereM• is thetime-averaged
residual
NMOcoefficient
alongthestrikedirection
at
thekzhCDPposition,M• isthetime-averaged
residual
NMO coefficient
alongthedip
direction
atCDPk, z,=z•ina,z•=z co.sa,
D•:=2•in.___•
istheapparent-dip
(orcross-
-dip)correction
ß
coefficient,
•/,--"y
sina,andM• andM• areconstant
fortheCDP
gather.
Here,z isthesource-receiver
offset,
• isthedistance
'between thesource-receiver
midpointandtheCDPpoint,v,,oistheR/VIS velocity
alongstrike,• isthetruedip
angle
ofthereflectoranda istheangle
between thetniedipdixection andthesotttce-
receive•direction(seeFig. 10.1.10).
Themeaning
of thecross-dip
termD•o canbeunderstood
fromFig.10.1.10a
and
10.1.10bwhereit canbe deduced that the time difference
dueto the deviationof the
source-receiver
from theeffective(or projected)surveylineis givenby

Ati-- 2R'H
or

Ati 2•Isin4.2•t
sins
coa(a
+•) 2

2!tsin4
.(- sina)
va = D}!l• (10.1.i3)

where•. = • cos(a+ '•.


To understandthemeaningofthetermsM• andM•, wesimplynotethatthetime-
dist•uce
relationfora 3-Dpointsource
canbewrittenas(Zai-tian,1988)

t +_
v... - ,i.=, -
wherea is the azimutho{anoffsetposition,
aois the azimuthof thetruedipdirection
andtheothersymbols
havebeenpreviously
defined.
Now,if a0= 0 in theabove
time-
distance
relation(10.1.14),wehavethat

158
(10.1.1,5)

Nowagain,since{1 -
it follows that.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

+ z 2co$2Q•
z=sin=e (10.1.16)

where v,,•, istheNMOvelocity


va=co•$ along
thetruedipdirection.
Thus,theresidual
NMO is givenby (Zai-tian,1988)

1 1
eat= ( 1)2rm•- + 1 ( 1.--V_•)2;2CO$20t
= Mt•z:•sin2ot
+ M•zaco.s:ta
= M•, a + M•z]

wherev•,•oandvgaretheuncorrected
velocities
applied
to NMOcorrection.
It should
benoted
thatv4=v•mo
when
•b---0
and ifin•ddition
a = 0,wehave
that
1 1 •z• .

sothat eq. (10.1.12)simplifies


to

Tii• = $• + R• + N•,+ M•z,a• (to.t.x9)

whichis equivalentto the expression


for thelinearsurveyproblem.
In thegeneral
3-Dsituation,
Si,R•, N•, M•, M• andD• axethesought
parameters
m, andeq.(10.1.12.)
is posedin theusualform

wherethe vector d containsthe observedreflectiontimes and G is slrnila• in foxIllto


that of the linearsurveyline(theelements
aremodifiedto accountfor thenew
structural effects).Theproblem canbesolvedasin the2-D caseusingtheLU or
singular-value
decomposition
techniques.
It should
benoted,however,
thattheproblem
is undersconstrained
andwill requiresomeproblemregulaxization
measures
to stabilize
thesolution
process.
It isremarked
thatthetypically
large
system
of'linear
equations
canbe economically
solved
in an approximate
m•nnerusingsimpleprojection
schemes.

159
10.1.4 .Reductionof Line IntersectionEtro• in Network-typeSurve•
ß

Land, marine find airborne•eophysicalsurveydatamay be in error dueto a numberof


causes.In typicalnetwork-typesurveys, navigational
inaccuracies
(including
altitude
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

differencesbetweenintersecting
lines),instrumental
biasanddrift, andtemporal
effectsarisingfromoceanographic,atmosphericand ionospheric
phenomena may
impartsystematic
and/ornon-systematic
biasonthefieldrecords. Consequently,
the
datarecorded
at coincident
locations
at clif•erent
timesin intersecting
survey
lines
may showcorresponding
discrepancies
dubbedcrossover
or intersection
errors(or simply
mis-ties). The magnitudeoœthesemis-tiesmay vary greatlydepending
on the causative o

factors(e.g.,navigational
errorsareoftendependent
on the qualityof thenavigation
systemusedin the survey)a•d the desiredlevelof acctttacyin a su•ey may dictate
what we considerto be a signi•cautsourceof error.For example,diurnM• may be
consideredthe main sourceof errorfor a + 1 gammaaeromagnetic
.
survey,with good
positional
controlusingDoppler
navigation
(orsection
roads).and
tracking
camera
for
accurate path recoverywhile in high sensitivity + 0.1 gamma •urvey, positionaland
ditwnal
• errors
become
equally
important
(Yarõer
etal.,1978).
Fora•eliable
interpretation, we need to adjust the raw data setsby some,rolmownamountwhic•
ß .

may be determined usingsomea pr/o• informationsucha.sbasestationmeasurements.

We are interested in crossoveradjustment procedureswithout complimentaxyhaze


station data. Traditionally, man• loop Closure methodsaxeappliedin seismic,
gravimetric, magnetic and lfeodeticsurveyin• but this practicehas now beenwidely
replaced
by automated
errorestimation
schemes.
Wewilllookat thebuilding
blocks
of
this mathematical-approach
andfor a better overallunderstanding
of the undergir•i•g
principles,we will try and blend the probl_•,-.•of network adjustmentin magnetic,
gravimetric and geodeticsurveyingusinga specificcaze.It is easyto extend'the
developments
oncetheProcedure
isunderstood
andit is therefore
instructive
to start
with a simple problem.

Considera small surveynetworkconsisting


of fourapproximately north-south (N)
traversesand tbx• approximately
east-west(E) tie-lines(Fig. 10.1.11).In eft•t, we
haveseven
lines(i0=7)andtwelveintersections
(i.e.,n=12).At the•h intersection,
there are two observations,'
rN• andr si but the actualvalueof the soughtphysical
quaatityat that point'may
bedifferent
frombothobservations.
Notethatthese

160
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Tie line

Intersection
2 37
r24 r3 r23 r4 r, r9 r21

T6
rl r2 r18 r5 r19 r8 r20 rll

1 T5

rl rl r15 r6 r14•r7 r13

T1, Start T2 T3 T4

Fig. 10.1.11 Sample survey plan. T1 to 1'7 are the tracklines and rl to r24
represent the readings at different times at.the twelve (1 to 12) points of
intersection

161
readings
couldBegravityormagnetic
fielddataorin theca•eof satellitea•timetry,the
distancefrom the satelliteto the earth'ssurfacedirectlyBelow.

If we are de•Hn• •rith altimeterdata, the problem•sstraiõhtforwaxd.


The measured
valuesof the satellite-to-earth
distance
r areknown; theheiõhth, of the satellite(above
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a referencedat•rn) is õenerally
constantalonõau orbitbut its exactvalue
The ground'selevationat anylocation•s givenby (h- r) and may a•sumediffferent
valuesfor cHfferent
tracks.The di$crep
.ancyin elevationestimatesat a l•ne intersection
•s given by (seeMe•e, 1984)

which, with the orderof the termscha•ged,leadsto

(10.1.20)

Now, sincethe quantity/•i is ouxdesiredparameterat a particular point, and the


crossover
observation
is •ven by d• = (r•v'- r•)•, wemay rewrite eq.(10.1.20)as

which is in the familiar form e = Grn - d.

For conventionalairbornepotentialfield measurements,the problem is somewhat


s•_rn•lar
but c•n be formulatedcl•fferentlydependinõ on the task at hand. In general,an
important objectiveof magnetic,gravlrnetrica•d indeedother geophysicalsurveys
is the determinationof accuratefield valuesat the observationalstations.The height of
the aircraft is usuallyfaJxlyconstantas in satellitealtimetry but usually known while
theexactva•ue
ofthedesired
physfc•fieldF •s-,•,•own.The'•--tie l•etween
track/c
whenit intersectstrackl, say,is simplyd = (r•v - r•) andthe soughtline adjustment
coef•cients areobtained
fromtheleastsquares problem Gm-- d wl•ere,asfor the
satellitealtimetryproblem,a•d for the sample7-linesurveynetwork(Fig. 10.1.11),the
coe•cient matx4.xand the discrepancy
vector a•e givenby

162
..

1 -1
1 -1
1 -1
-1
-1
-1
and d=
-1
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

-1
-1
-1
-1
-1

whereonly the non-zeroentriesof the 12x 7 coelicientmat•4_x


azeshown.
For small-sizesurveyssuchas the exampleadoptedhere,wemay solvethe problem
usinganyoœtheregulazized
matrixinversion
solution
tecl•,•ques
discussed
in the
previous
chapters.
Therequired'
quantities
forcalculating
thesolution
areG•r(•and
G•rd. Fo: the casetreatedhere,•-r(• is simply

3 0 0 0 -1 -1 -1
0 3 0 0 -1 -1 -1
0 0 3 0 -1 -1 -1
0 0 0 3 -1 -1 -1 (10.1.23)
-1 -1 -1 -1 4 0 0
-1 -1 -1 -1 0 4
-1 -1 -1 -1 0 0 4

A few structuralpointerscaube gleanedfromthe aboveformulations.Noticethat if the


matrix (7 is augmented
by a scaledvexsion
oI itself(i.e.,• :_•D) andthe discrepancy
vector
d iscorrespondin•y
aulPnentecl
withzeros,
weineffect
endupwitha smooth
invers/on formula-

(zo.z.2)

a•d, asin the refraction


seismolo•yexperimentin Chapter5, wecouldex•e the
trajectoryof the solutions
in •-space.Thedataweiihtin•termW is introduced
here
ßfor statisticalconsiderations
andis a flexiblestrategyfor controllingthe i•luence of a
•ivenobservation
onthefinalsolution.
Notethatthereareseveral
fo.rmsofwei•,hting.
In magnetic
surveys,
forexample,thedataequations maybeinversely
weiõhtecl
accordin•
to anestimate
of thelocalfieldgradient
usinõ,
say(seeRay,1985)

163
wherethe cor•espondinõ
r •ud • representthe field readingsand times•t • givenpoint
of intersection.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

It is obviousthat the structureof the m•trix G •lso •1ows the incorporationof • •v/•v/
constraints
in the solutionprocess.
For instaztce,
• highqu•li•y llne (o• ]•,aesegment)
couldbe held fixed simplyby •ddinõthe •ppropriateextrarow of constraints ontoG
and•ppending the wellestimated
a pt/or/dataonto•l asdescribed in Ch&pter 5.

Sofax,we h•ve concerned


ourselves
with the solution
o.fsmall-size
mat• equztions. ß

typicallaxge,
finelygridded
su•eys,thedimensions
of•hematrixG maybeof the
orderof 1000x 1000(or evengreater)sothat the application
oœsimple
decompositiontechniquesin routinep•ocessingbecomes auextremelyd•icul• task
usingpresent-d&y
computers.
Theusualapproax•h
uses
itexa•ive
gradieaxt-type
solutions
•nd the i•terestedreaderis •eferredto •he easily'digestible
schemes
described
in
Cloutlet(1•83) •xd Menke(1D84)arnOrig
others.Howevex,
i• maybepo•ted outthat
m•trixdecomposition
•ndinversion
oflaxge-size
surveys
isnotaax
impossible
task.An
ingeneous
ma_nipulation
ofthematrixG mayrenderthepzoblem
solvable
in auef•cient
manner.I• c•u be seen•romeq. (10.1.22)that a• moresurveylinesaxea•lded,the
m&t•x • increasesin size but maint• • consistentblock-structuxefor i•s non-ze•o
components
(B.ay, 1•85), viz-

[] (10.1.:26)

[]
Thestoa/1di•gons/blocks
•re dense
whilethela•geblockontherightha•d-side
is
typically
sparse.
Thisstructurecanbeexploited
in thesolution
scheme.
Forexample,
theproblem
maybesolved
bydirectapplicaQon
ofblockorthogonal
decomposiQon
usingridgeregression
(P,ay, 1985).
While the •bove simplistictreatment will serveto cementthe basictools of track
adjustment,
it is remarked
thatin routinemagnetic
processing
forexample,
theproblem
issetuptoremove
• bias/tom
a•lthelines
asabove
butwith
a•additional
goal
of
removing
tilt fxomoneof thet•ave•ses
andoneof the tie-lines.
In thissituation,there
would be an •ddition•l columnin G fo• eachof the lines usedfor the tilt correction.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

10.2 Layered-EazthInversionof Refl•on Seismo•,,,s


In this introduction to seismo•raminversion,we shall restrict our discussionto ca•esin
whichthe ea_•hcanbe considered
to consist
of horizontallayerscha•acterised
by
constant
materialproperties
(density
Pi, velocity
•i andthickness
hi) asdepicted
in
Figare10.2.1.The groundis excitedby someenergyinput derivedfrom an explosive

i . i i ml ii

ii i

,,, ½', •" -


•'n-,I-
2 O0

Fig. 10.2.1 An (n+ 2)-layered


Earthmodel.½i= •ivi is the acoustic
impedauce
ofthejtalayerofvelocity
v•, density
pi audthiclmess
hi.

source,say an.dpart of this energyis reflectedback towardsthe tarlace at ea• of'the


:

interfacesin the model.The signalscan be observedat the tarlace or in boreholesas


functionsof time a•d spatialposition.The reflectiontraveltime data recordedat a•
observationpoint constitutesa seismictrace. A typical seismog•:•mcollsistsof several
seismic
traces.
Wewilllookat twomeasurement
geometries:
(1) Hoffmont•lsource-receiver
sm'f•e arrayconfigurations
and
(2) Surface-•ud-borehole
sourceand receiverarrays.

165
10.2.1 Iterative Inv•on of Normal Incidence Surfa• Se;-•mlc Traces

Considerthe problemof recoveringthe acousticimpedance structureof the subsurface


fromnormalincidence(or zero-offset)reflection
seismograms recorded onthe surface
usinghorizontalsource-receiver
arrayconfigurations(e.g.,Cooke•ud $ctmeider,1983;
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Oldenburget al., 1983).The subsurfaceis p•ameterizedinto discreteintervalsof


constantimpedance.A commontype of pa•ameterization
usesequispaced
samplesof
the sxmedimension
asthe observed
data.If wecausimulateseismograms corresponding
to anygivendistribution
ofacoustic
impeda•ces,
thentheabove
problem
caubesolved
usinginversetheory. There are m•uy methodsof solvingsuchproblems•ud the
commontasksincludethe calculationof the forwardsolutionand the matrix of.partiM
derivatives(or gradientsof the objectivefunction)for linesrizedsolutionprocesses.

The forward problem


We sh•11adopt the familiar convolutionalmodelfor the forwardpr.ocess.
Essenti•y, a
layered-earthresponsefunctionis convolvedwith a seismicsource-waveletto simulate
thefiltera•d phase
shifteffects
oftheearthand'
recordinõ
system
yielding

$(t) =' (10.11.1)

where the P,(t) is the earth reflectivity function, ß is the convolutionaloperatorand


W(t) is the sourcewavelet. Usinõ the z-tran•orm method,a recursirerelationfor the
reflectivityfunctionis givenby (Cookeand Schneider, 1983)

(r.• + R• + •Z.)
j ----n,n-1,...,2,1 (10.2.2)
ii(Z)= (1+ r/ii+•Z)'
wherer i is the pressure
•mplitudereflection
coefficient
at theboundm•between
layers
• and• + 1 (seeFiõ. 10.2.1)andis definedas

-'
........ (10.12.3)
= +x+
andtheacoustic
imped•uce
½iistheproduct
ofthedensity
andvelocity
ofthej• layer.
If we defineR• + 1-- r• + 1 andthenapplytheaboverecurirerelations,
wewould
obta• the reflectivity functionP,(•) requiredin our convolutionalmodel.

Formulatingthe inverseproblem
Weareintereste
d infinding
theacoustic
impedance
structure
thatwillgenerate
the

166
syntheticseismogram
$(t) that fi•sourfieldseismogram
$(t) bestin the leastsqu•res
sense.That is, we waut to mi-imize the residual
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

subjectto physicalor geological constraints.


This is a nonlinearproblemin ( •ud we
thereforeopt for •u iterative1inea•zedsolutionprocess.

To start with •nd for notationalconsistency,


we notethat eq.(10.2.1)may be written as

so that the field recordsmay be expressedsimply as

s = .f(m) + e. (10.2.5)

Expa.u•g f(m) aboutaninitialmodels


møin Taylorseries
a•d neglecting
highorder
terms as usua/gives

O.f(mø). _o.
.f(m)
= f(mø)+ am
o •m-m•
or equivalently

sothat eq. (10.2.5)maybe writtenin our f•,•;l•ax matrix notationas

•=•+e

•nd is obtainedby differentiatin•equation(10.2.1)with respectto the choseumodel


par•r•eterization(i.e.,the •cousticimpedanceprofileor the reflectioncoefficient
series),
•nd Rø is the reflectivityfunctioncorresponding
to 'theinitial model.
parameters mo.
Notice that the sourcewaveletW is independentof the layer sequence
and is here
convolved
with •nothert•meseries
0Rø/5mø in the s•,•e way asforthe generation
of

167
a syntheticseismogram. We may therefore
developa recurrence
relationfor calculating
the partiM derivativesA.

Computationof paxti• derivatives


I/one is inverting• seismogramfor(, wenotefromeqs.(10.2.1to 10.2.3)that the
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

reflectivityfunctionR isa function


ofr whichis in turn • functionof ½.Thus,

aA _ aJza,-
a•' - a,-'a-•
(•o.2.6)
-2½•+ •
where
O//_(1+- r•//y+•Z)
Or- (1 (//•
+•Z)
:)• a•cl• =(½•+•+ ½•)=' ,

Solvingthe inverseproblem
It typicalpracticesituations,
wearerequired to solve
matrixequations
oflargesizes.
In thepresentformulation,it is theleast$qua:es
normalequations

ATA• = ATy

that haveto be solvedœor


the parametercorrections
z. Dependingon the auxi!lazy
constraints
imposed
onthesolution process,
wemayusetheconstrained matrix-
inversion
techniques
developed
in theprevious
chapters
or g•adient-type
algorithms.

Conjuõate-•radient
appzoar.
h
Gradient-typeschemes
a•eattractive
in thattheydonotrequire
matrixdecomposition
or inversion.
A staad•d c•njugate-iFadient
leastsq• 'mve•onalgorithm
is •iven
below for illustration:

Step1- Initializepara•neter
vector• anderrorvectore
set z ø = 0
•o = (Ar• ø -
•o =-eo
definea fixedpercentage
(e)0f eo[oruseasa stopping
criterion
set ,ol -- •l•øl
or Otherwise

SLep
2:Iterate/orbest/itmodel,
computing
newA a•d l/at eachiteration
LOOP (for i=l,n)

168
õe•:

At= d•_•b•
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

improveprevious
estimatein A-stepaxtdget newerror
get:
:• = Zi-1 + Aidi_!
ei = ei-1 + Aibi
convergence test
• levi< tol, STOP
Ehe

set•i = [ei_1[2
selectnew conjugatedirection
get di ='-el + •idi-•
END LOOP

Note that • new J•cobianmatrix A anddiscrepaucy


(or data m•sfit) vectory are
required
at eachiteration
in Step2 oftheabove •lgorithm.
In principle,
thealgorithm
neverdiverges
a•d thestopping criterion
is therefore
designedsuchthat theiteration
stopsoncetheresidualshavegonedownto a pretermiued tolerable
value,tol.

10.2.2 Inver•on of VerticalSeismicP• (VSP) Data


The seismicwavefieldgenerated by a sourceEredgt thesurface
(ora setofshotsset
off'gtvaryingoffsets
sadazimuth to a well-head)
cauberecordedat different
depths
in a borehole
(Fig.10.2.2).
It is alsocommon to'havesurIace
receivers
andborehole
sources.We will adoptthesurface-shot borehole-receivex
arraugementin thisdiscussion
but bearin mind that thevariousarraysaxegenerally
termedVerticalSeismicPro•les.
A majordi•erence
between
VSPdataa•d horizontal
surface
arraydataisthatthewave
fieldapproaches
theVSPreceivers
fromabove a•dbelow.It iscustomary
tosdpaxate
theupwaxda•d downward
propagating
w•vefields(e.g.,Seeman
a•d Horowicz,
1983)
a•d the.inverse
problem
caatherefore
beeasilytackledusinga•y of themethods
that
holdgood
forthecomparable
surface
arraydata.R,ec•11
thattheinversion
ofhorizontal
surface
arraydatasuffers
fromnon,
mlqueness
owingto a• arrayoffactorsof whichthe

169
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

D
U

,•m3=v3•t
z•m1=Vlsi
• rn2=V2aZ
•,

170
Rj
Rj+I

Rj+2

Fig. 10.2.2 A surface-sourceborehole-receiverarray system


uj

uj+l
bandlimitednatureof the da•ais mostproblematical.
This problemis somewhatlessin
the VSP p•oblemwherewe axeprovidedwith someusefulconstraints in the formof a
tooddescription
ofthedowngoing
wavelet
anditsamplitude
relative
to theupwaxd
propagating
wavefield
(Grivelet,1'985).
Letusnowexamine
some
common
applications
of inverse theory in VSP data processing.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Posing
theseparation
ofupgoing
anddowngoin•
wavefieldsasauin• problem
Consider
the hypothetical
situationillustratedin Fig:10.12.2
wheretherearefour
receivers
(goto g3)damped
against
thewallofa borehole
above
a horizontal
reflector
(characterised
byreflectivity
Rt) anda shotissetoffonthesurface.
Porsimplicity,
let
usassume
that aaycha•ges
in theshapeof thesource
waveletdiningpropagation
is
negligible
between thereceiver
positions
audthatthetimedelays
A• between
a•y two
contiguousreceiverpositions
a•ethes•me.Wecanthereforeexpress
therecorded
data
at positions
go•o g3as(see'
Hattonet al., 1986,p.1õ3)

so(t)= Uo(O+ do(t)

.st(t)= Uo(l-'•t) + do(t+•)

s•(t) = uo(t-2At) + do(t+2A/)

s3(t) = uo(/-3a/) + do(t+3At)

whereUoanddoare,respectively,
theupwardand'downwaxd propagating wavefieldsat
thechosen
reference
levelgo-It maybenotedthatin practicalsituations,
wewould
havean extra termfor additivenoisein the righthand
sideof eq.(10.2.7).

It is instructiveto obtainthe relationsfor the variousfrequencycomponents


of the
recorded
signals.
To dothis,wesimplyexpress
eq.(10.2.7)
in thefrequency
domain.
Thus, at any frequency,we havethat.

So(l) = U'o(,f)+ Do(l)

S2(f) = e:rp(-i•2•xt)Uo(f ) -t-e:cp(


i•o2
/xt)Do(f )

171
Sa(f) = ezp(-i•3At)Uo(f) + ezp(i•3At)Do(f) (10.2.8)

which,in our familiar matrix form,is simply


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

• -- i•oAt •i•At
• -- iu•2&t •iw2At
6 -- •3&t 6i•3At

In general,
eq.(10.2.9)isin theformd: Gmwhered cont•i,•.•
theFourier
of the data recordedat eachreceiverpositionandm is the solutionfor U0 a•d D 0 for all
frequency
components.
Thesolution
to theabove
problem
canbeobtained
using•uy of
the st•udaxdinversionalgorithm.•
employing
re.gula•zedC-mat• or solutionsi,•plicity
measures.
Forexample,
using
theridgeregressicOn
algorithm,
wehavethat

(10.2.10)

whereG. is the complexconjugate


of G. The desiredtime domainestimatesof the
upwarda•d downward
propagating
wavefields(% anddo)•re givenby the inve_•se
Fourier transform of m, i.e.,

(zo.2.zz)

where F- t denotes the inverse Fourier transform.


Havingseparated
theVSPwavefieldcomponents, andass•rni•g
thatthedatahave
beensubsequently
corrected
forspheric•spreading,
let usnowproceed
to showhow
usef• structuralinformationaboutthe subsurface
may be reconstructed
from the
upõoinõ waves.

•very of the aco•c irapeal,raceprofde


Our modelhere consists
of a stackof horizontsllayersof equaltwo-waytravel-time
excitedby compressional
pl•uewaves travelling
at normalincide.
uce,•ud the structure
canbech•racterised
asa time-series
ofreflectioncoe•cientsof•coustic
impedances •
forthecomparablesurfacereflection
problem. Wew•ut to•mdthemodelp•r-meters

172
(impedances
of reflection
coeffiecients)
that will 8erierate
a synthetic
seismogr•m
that
m•tches our field recordsbestin the least sqauressense.As before,we wouldrequire to
computesyntheticseismograms andtheirpartialderivatives (orgradients).
Fortheforwardproblem, wecompute •heupgoing anddowngoing wavesforan
elementarysurface(Fig. 10.2.2top right inset)as (Grivelet,1985)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

uj = -rjd• + (1 + r:•)u:•+•
aud

di+ x = (1 - r•)d• + rju• + t ß

Next,weobtaintheimpulse
response
sequence I asIj---uj (j=l,2,....,N). Theimpulse
•esponse
is thenconvolved
witha waveletW to yieldthesynthetic seismogram
ß
S(t).
Note that the actualobserved
downgoing
wavefield contains
the directpulsea•d the
associated
multiples
generated
above
therecording
levela•d ca•betransformed
intoa
zero-phase
waveletby application
of a wave-shaping
filter(Grivelet,1985).Thiswave-
shapeddowngoing
wavemaybeusedto 'apprbxima•e W"yielding

$(•) = W(t) .l(t).

Theinverse
problem
isthenstated
as:minimize
• I•'- $1• subject
toanygiven
constraints
where.• is theobservedseismogram for-thewave-shapedupgoing waves.
Theproblemis nonlineax andca• besolved usingaxxyofthestandaxdgeneralized
matxixinversion
methods or gradient-type
algorithms.
The initi• modelfor theitera-
ß

tire h•versionmay be obtainedusing,say,a recurslye


eventdetectionalgorithm
ß

(e.g.,Grivelet1985).Lineaxization
will yieldtheexpression

•=.Az+ ß
whel•e
ß t, 01ø (10.2.14)
A W()
ßaadis obtainedfromeq.(10.2.12)
for the chosen
modelpar•meterization
(i.e.,reflection
coefficients
r• or acoustic
impedances
½i),forex•,•ple,wemaywrite

a_r oI Or (10.2.15)
a½= • 'a½
-2(• + •
where
•r=-dj+u•+t
a•d•= (½•+1+ ½•)•'

173
10.2.3 I,;m•tationsof NormalIncidence(Zero-Offset)Inversion
Note from eq.(10.2.3) that if oneof the impedances
is known,(• say,then onemay
constructan impedancepro•e by rectu•veinversion usin• the relation(seee.õ.,
Lavergneand Willm, 1976)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

= (1+
-

where it is ass•amedthat the seismictrace is noisefree and has been rid of the source
wavelet,multiplesandeffectsof spherical
spreading,
transmission
losses
axtdscaling.
It is obviousthat all theseconditionsaxeseldommet in •outine data processing.
The
iterativeinversionapproachdescribed
aboveaxethusmorepopular,but they havetheix
!
ß

drawbackstoo. For example,the data are baad-!i•ited and the resultsaxenoxt•ique


without
theuseofa prioriinformation.
Thedatasuffer
fromthelackoflow-andhigh-
frequencyin.formationabou½
theimpedance structure.
Thiscanbeseensimplyby re-
statingeq.(10.2.1)as(01denbttrg
et M., 1983)

$(,) = F-

whereF- t is the inverseFouriertraaasform


operator'and R(I) and W(f) are the Fourier
transforms
of It(t) and W(t) respectively.
It is obviousthat the seismogram
contMsm
energy
onlyat thosefrequencies
whereR4I)andWif) arenon-zero.
Notethatrealdata
axeband-li,•itedin frequency.
onlybecause
of the measurement
system;the function
S(t)isbroadband
but W(t) isalways
band-limited
sothatS(0 doesnotcontain
much
in.fo•ation outsidethis bandwidth.It shouldbe pointedout herethat Oldenburget al.,
(1983)
developed
methods
forpredicting
themissing
low-andhigh-frequency
information from band-limited normal incidencereflection seismograms;in any case,
thismethoddoesnot holdgoodforverylowfrequencies
sincesuchfrequencies
{downto
zero)areonlypresent
in moveout
data(seeMcAulay,
1985).

It is a well knownfact tha• themostdesiredparameterin seismicdata processing


is
velocity.
In normalincidenceorpost-stack
inversion,
un--,•bigousdetermination
of
separate
layervelocities
a•d densities
isnotpossible
because velocityanddensity
arise
onlyasa productin thereflection
coefficient.
Moreover,it isdi•icultto model
the
exactmannerin which someof the observedundesirablefeaturessuchas interbed
multiples
arestacked.
Theinversion
ofpre-sta•datahelpsovercome
suchproblems
and will be introduced nex;.
10.2.4 Plane-Laye• Inver•on oI Pre-star.
k Seismo•• for Velocity and Density Profiles
In prestackor nonnormalincidenceconsiderations,
the traveltimesacross a •iven layer
dependuponthe anglesof thepla•e-wavecomponents whicha•e/unctionsof the
compressional andshearvelocities
in the layersbut not of density.This meansthat
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

velocitiesand densitiesmay be determinedseparatelyfor suchproblems.The la•dm•k


plane-layer
point-source
modelofMcAttlay(1985)will Beadoptedin ou•discussionof
seismo•r•minversion,for'velocity
and/ordensity
profiles.Essentially,
a point-sottrce
is
accountedfor usinga sphericalexpansion
into planewaves;a•d an extemsionof the
planewavereflectionresponseto thenon-normal
incidence
situation•nd the
incorporation
of surface
effects
completestheformalism.

A radiatingspherical
wavefroma pointsource
canbe expanded
into a set of pla•e
waveseachreflectedfrom'the succession
of horizontallayers (seeFig. 10.2.3),
,r / 2 - ioo

=i=of•=0
sin•)e.
xp[i(h,
+h,.)/%
½os•]
sin•
d• (10.2.18)
whereR(•, •) is the reflection
coefficient
Io • planewave at frequency • and striking
the stackof layersat aa aagle.
• fromthevertical,i -- V/ - 1, ko= •/vo is the wave-
w•mber for the medi,•mat the sudacehavinga velocityVo,h, and h• are respectively
the heightsof.thesource aadreceiverabovethe topmostinted•e, andJo is the zero
orde• Besselfunction. For a knownsourceof strenff•hP, the responseat a receiver
locatedat a speci•cpointin space
is determined
by S•rnmir• the e•ectsoverall all•les
in theabove
equation
andtr•-.•œormi•õ
fromfrequency
to timedomain.
Thelayered-
earthresponse
function//isdetermined sepaxately
for eachpla•ewaveandeach
frequencycomponentof thesottrce
field.

McA•y's alõørith'" for cal•. R


The non-normalincidence reflection
response//forthe compressional
wave caseis
considered
here.For•hejzalayerofdensity
•i andvelocity
vi, theupward
anddown-
wa•dcomponents
ofthewave
pressure
a•ewritten
interms
ofthata•thesame
posit{ion
in an adjacentlaver(j + 1). FromFig. 10.2.4,wehavethat

and

= c J::)+ trj" + . (10.2.20)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ii
i
i

i
ß

surface (after 54cArtlay,1985)

176
Ocean
Receiver

Fig. 10.2.3 Pointsourcemodelfor a horizontalstructurewith an ocean


Eartl•
Layered
Now, from eq.(10.2.19),

D• = D•+•- •17•+•) =

whichmaybe substituted
intoeq.(10.2.20)toyield
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

so
that
[•';1
•2=•I c•I c•I[7'J
L +
•'•+ (10.2.21)
where
• = -c•andt•t-•- c• = 1.Ther••ion c•ci•t c•fo•awave
atnon-
no•• incidence
ha•n• approached
the•t•ace •om theme•m ••• •p•••
• is•v• by
(•+l - (• (10.2.22)
c•= •+l + •
where
fortheplace-wave
whose
direction
is.atan a•gle0j 'withrespect
to thevertical,
the acousticimpedanceis definedas

p•v• p:•o (10.2.23)


• = cosO•P•
a•dthehorizontal
componen•
ofthewavefield
inlayer
j is
Forcomputational
e•e, wemaywritetheanõle
oftheplaaewavein eachlayerin
termsof the verticalcomponent
of the wavew•mber
at thesurface
medi,,m

whence
(10.2.24)

Fromtheset-up
in Fig.10.2.4,
thepropagation'from
onelayerontoanother
(across
the
f.• layer,
say)'ass-ming
timedependence
oftheformei't isdescribed
by

D• = D• e:zp(iv•5•p•)
and

177
i ii i

Z's U/OCEAN r

- •ø•urc Receiver
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1st LAYER
• Do•
• eo ' uø Oth
INTERFACE
1st INTERFACE

j-lth INTERFACE

VELOCrrr
DENSITY
Pi
jth LAYER THICKNESSvl•i
WAVE
NUMBER
I•
jth INTERFACE

j+ lth LAYER
Dj+ . ,, j+1 i + lth INTERFACE

ii i i i ii i

j• LA7ER INTERFACE

LOWER I+1
HALF
SPACE
Oj+ I ' D'j+ 1andUj+ 1' U•j+ 1 forEj+1=0

Fig. 10.2.4 Set up for non-normalincidencereflectioncoefficientcomputations


(after McAulay, 1985)

178
where5• is thenormal
incidence
traveltime
across
thej•hlayer.If wedenotethephase
delay acrossthis lkyer by the exponetialfactor
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

then we have that

thusallowingus to rewriteeq. (10.2.21)as

ID•I
=.•-•+•1 E•+• c•
• •/2
•E•+• 1 U)
+
•+
' (10.2.26)
l{ecursiveapplicationof equation(10.2.26)yields.therelationbetweenthe upward•ud
downwardwave pressures
just abovethe top interfaceU 0 sad D Oaad those one layer
tkicl•-essbelowthe bottominterface,U'j + x aadDj + •

= = qW +' (10.2.27)
= J+ J+

where

I ß, p•.=Ey+•
qi= .•,.,•
• w•/• cy a•dW=•w=•
w•lto••l'IoP
+• iE ) +• 1 w= '= • = PoP•...P•.
It is obviousthat P is a propagatormatrix that facilitatesthe computationof upgoing
ß

ßuddowngoing
wavecomponents
just•bovethej•hl•yer(i.e.,Uj andDi) fromthose
just abovethe (j + 1)-layer.

To obta/n the reflectionresponseR, we ass-methat no field entersinto the stackof


laye•sfromthe bottomhalf-space,
i.e., U• + •=0. Usingthisradiationcondition,//is
derivedfrom eq.(10.2.27)simplyas

= U0= w•2 (10.2.28)

179
Where a bodyof wateris present•bovethe stackof layers(Fig. 10.2.3),the f•eesurface
is ta.keninto •ccountsimplyby adjustingthe reflection
coe•cient/• to this surface,
i .
thus U! e•øhuø=e2'%•.
R (10.2.29)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

/tl=• = e-•ø•Do
wherepo=koCoS8o a.udh isthethickness of thewatercob•. For a shotfiredin water
•t a depthh, from thissuff•ce,thefieldreceived
at a receiver
position
in thesame
medi-m •t • depth h• from the free surfaceis givenby

R,=(1+ r0/•l)
'7 (10.2.30)
where

audwe h•ve neglected


the•ect w•ve •d i•ss•ace refl•tion; the me••e
--2
[•/(l•roR1)]e Poh•
istherespo•e
moa•ed-tor
thes•acereverb•ation
adj•t• to
thereceiver
•d for• s•acereflection
c•d•t of -to,.[1-r0e
2i•oh•]
isthere,ever
ghost
•d e•PO(h•-h')[1-roe
2•ø•')istheso'•ce
•ost'adj•t• totherec•v•depth.
It m•y benotedthat•co•oratings•ace •ects
to •cc•ately simwaresomeobse•edph•omena.Forexample, • practicMs•tuatio•
the reflectionener• ••ng •t the s••ce isr•e•ed •to the bodyof wat• b•ame
ofthe•gh refiecti•tyat thewater-•r•teff•e;
betw•n the s••ce •d the solid substratm on w•• the w•t• r•ts g•erat•g sev•e
m•tiples.The•go•thm described
above••tivley s••at• p•m••, •terb•
m•tiples,s•ace m•tipl•, ghosts,
l•t•M wav• (re•actio•)•d comb•tio• ofthe
waves.

Computationofpartialderivatives
,ofR
Thepartialderivatives
ofR withrespect to •hemodelparameters,
rn(i.e.,velocities,
anddensities,
p•),canbederived fromeqs.(10.2.28
- 10.2.30)
fortheappropriate
exp•ental set-up.Usingeq.(10.2.28),weh•ve(seeMcAulay, 1985)

(10.2.31)
c•m•-- w22
If thesurface
effects
aretaken
intoaccount.
thenweneed
t•he
paxtiM
derivatives

18o
OR, 7{(•+ roR•)(ORHOm
•)-roR•(OR•/Om•)}.
= (10.2.311)
(• + roR•,)
• (•+toRy)
•' Orn•
where

otll.= OR
Omiexp(2ipoh)
ami
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

aadPo:ko=• ß
Tocomplete
thecomputation
ofpartial
derivatives,
weneed
thefixaction
ORaad this

requix•
theevaluation
ofOre1
OWoxsimply
•OWa•dOW
Op•ß
Recallthat wedefined
thepropagator
matrixPi aza •a•ctionof Ei + x a•d c• sothat
P)_• isaf•mction
ofE• aadc)_•. Therefore,
dependiag
onthemodel
parameter
beiag
co•zidered,
Pi a•dP•_• axebothhmctio•z
ofm.Ia ge•exal,
wewrite

Now
(10.2.34)
b'• = [Oz•+•' Om• ac•
aad
OPi-•, OE)• OPi -.• Oc•
_•
0miJ
+ Ocs_•"
Oral (10.2.35)
where

OP•_• 0
0 I0 ' OP•_•
•+• OE• 1 0
0 a•d
Oci_
• E•
cj_t

Forthevelocity
paximeter,
wehavefromeq.(10.2.34)that

OE)+•=0 andOc) -2()+•(). a•2


since.
avj • = (½j
+t+ (•)•P}•;
andfor eq. (10.2.35),we have

OE
i -2i6E)p•
andOc•_

181
so tha,t

J
OW
rI Pk
Ov.i }I'=I=oPk
•+•0 ((i+z
+(i)•-'P k=•+l
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

k c./_t 0 0 (½•-•'+ G)'' • =•


Similarly,for the densityparameter,
we havethat

OPi _ i:?P
• . Oc
i
Opi- Ocj
where
Oci _ -2½•+•co

and

C•pj i•Cj_1 i•pj


where

•ivinõ

aw
=•-' { o • -2½•+,
½•1,(I •
+ •ri=oP}/• 0 ß(½•_•+•)•p •=•
rl P}.
It maybenotedthat thematrixofpartialderivatives
is of dimension
n xnp,i.e.,
m•mber
ofobservations
bynl•mber
0fmodel
parameters
where
in thefrequency-
wavem•mberdomainc•sethe n,,mberof observations
is equalto the n,,rnberof
frequencies
timesthen•,mber.
ofhorizontal
wavemlmbers
orangles.
ThepartiM
derivatives
arecomputed
during
theforward
evaluations
resulting
in considerable
savings
incomputing
costs.
Theinveme
problem
canbesolved
using
anyofthe
standardmethodsof iterativeinversion
but the conjugate-gradient
methodis
commollly
used(seeSection
10.2.1).
10.3Layered-EarthInversionof El•magne•c andElec•r/calRes/s•i•• Soun•l;•gs
Electromagnetic(EM) andelectricalresistivitymethodsof geophysical
explorationare
widelyusedin mining,
geothermal,
petrole-•,hydrogeological
aud•eotech•calinvesti-
lafionsa.swell asin geological
mapping.
The transientor time-domainelectromagnetic
(TEM or TDEM) andmagnetotelluric (MT) methods arethemostpopulax deep-
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

probint EM methods. Theobservational


dataaretypicallypresented
in •he formof
apparentresistivityand/orTEM voltagedecayorMT phasesoun•g cm'ves. The
$chlumberterandWennerdcresistivitytechniques arehugelypopularelectrica•
methodsa.udthe field record/nõs
are in the formof apparentresistivity(or simply
re$ista•ce
) sound/ng
curves.
Datainterpretation
wastrad/tionally
ca•ed outby curve
matching,a practicenowreplaced
by forwardmode!]•goncomputers
andautomated
inversion
techn/ques.
Letusnowex••e howsuch
del•th-sound/nõ
datamaybe
inverted
forsudsurface
resistivity
structure.

10.3.1 The Inverse Problem


The inverseproblemof.EM or electrical
depthsounding
involves
find•g a modelthat
fits the observedda•a bestor withi,• somepresetconditions.
The problemcan therefore
be formulated in a w,mberof ways dependingon the ultimate goalor type of
ß
ß

constraining
conditionsimposed
on theproblem.In general,
modelconstruction
involves
minimizingthe differences
betweenourn valuesof observational
data,d andthose
predicted
viathenonlinear
forward
functional,
f(m)whilst
satisfying
a•yco•straints
imposed
ontheproblem.
Theformofconstraints
depends
onourinforeedexpectations
of the subsurfaceresistivitystructureor the actualvaluesof the soughtmodel
parameters
derived
froTM
some
otherexperiments
at thes•r•elocation,
say.Thus,for
exnrnple,
if we expectthatthereaxenosharpdiscontinuities
in resistivity
in the
subsurface,
then it maybe wiseto seeksmoothmodels.If, onthe otherhand,we expect
somedistinctgeoelectrica•.•its
in accord
withgeologicalwisdom, thenit may be
desirable
to constructconventional
layered(i.e.,sparsely-par•rneterized)
models.Also,a
betterpictureof thesubsurface
may beobtained undercert•-• circ•rn-•tances
by
modellinga combinationof differentgeophysicaldatasets.Alongtheselines,wemay
alsoelectto ret2in somepriorestimatesof the soughtpar•vneters
in the fina• solutionto
ourinverseproblem.A .practicalconstruction a•gorithrn
shouldbeflexibleenoughto
accommodate
mostoftheinterpretational
features
highlighted
above.
It isalso
expected
that the inversesolutionprocess
be n•rnericallystable.
10.3.2 L•ne•vizi• P •r•metmSzatioms
The go• of layered-e•th inversionis to •d • e•th-t•e resisti•ty st•ct•e that •
reproduce o• field obse•ations.The e•h is p••eterized into a succ•sionof layers
of resisti•ties,g• •d t•c•esses, •(. The qu•tifies, g •d • •e the p••et•s
.
of
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

o• E•h-model •d may •s•e a •de r•ge of v•ues. It is w• •o• that the


app•ent resisti•ties•play a log-nomM•st•bution •d b•use a •de r•ge of
vMues•e consideredin t•icM •verse problem, theirlog•thm• •e used•t•d of
the o•nM data the•elv•. We •o •e the log•th• of p •d • • the sou•t
p•meters. For the app•ent resisti•ty data, we may •o re-de•e the obse•ationM
e=ors (e.g., Pealersen•d He••ce, 1986;Meju, 1992)•

(d, + (d,)

where •i0 is the i•hobserved


st•udaxd
deviations.
It isalsodesirable
to uselogaxit•c
diffezenceswhen calculatingpartiM derivatives.with a finite-differencescheme.

10.3.3 ForwardProblem•andComputati4nal
Co-•deratio•
To conducta seaxchfor modelsthat explain ouxfield data, we xxeedto be able to
simulate the observableresponses
of hypothetical eaxth-modelsto inductive
energizationso• current excitationsfor a givenmeazurementconfi•ation. This is
perhapsthe most important task in data invexsion•ud constitutesthe forward problem.
Thus, before implementing •u inversionscheme,it is instructive to ascextainthat there
exists•u efficient, stablem•mericalschemeIor solvingthe EM layered-earthforward
problem.For illustration,the inverseLaplace-trauorm methodoœI(ught and Raiche
(1982)and Raich4(1984)will be axiopted fo• the TEM methodand the algorithmgivexx
in Word,SmithandBostick(1970)wouldbeusedfo•theMT forwardproblem.Forthe
electricalresistivity simulations,we will usethe 1ineax
filter method(Koefoed,1970;
Ghosh,1971; Johau•en,1975).The •elev•ut theoriesa•e s-rnarized below.

10.3.3.1 The TEM Forward Probit,-

In the TEM-forward problem, we attempt to simulatethe grotmd'sresponses to


inductiveenergization.One'way of doingthis is to calct•atethe mutual impedance,Z
betweenthe traasmJtter(Tx) a•d P,eciever(Rx) loops,locatedon the ground'ssurface,
as a fuactionof measurement(or delay) timest. The • may be a smallmttlti-turn coil
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

0'2

6n
o'1

Gn+l
hi

hn
h2

185
a

Fig. 10.3.1 TEM loops on layered-ground


Z
placedat thecentreoftheTx loop(anarrangement referred
to asttxecentral-loop
or
in-loopconfiguration)
ortheTx canserveastheRx during thetransmitter-off
time(the
coincident-loop
measurementconfiguration).
In a layered-earth
situation,
fora Tx loop
of radiusa andRx loopofradius
b,(seeFig.10.3.1)wehavethat (KnightandtL•che,
982)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Z(t)= -•#ab L•'•{I(p)pAo(m,


o
p,A)}J•(Aa)J•(Ab)dA (10.3.1)

whereAo(m, p,A)is the layered-eaxth


impedance
fixnction,
rnrepresents
the paxameters
oftheEaxth-model
(i.e.,resistivities
Pi - ortheirreciprocals,
conductivities
aj - and
thicknesses,
hj ofthelayers),
p istheLaplace
transform
vaxiable
corresponding
to
where• is the ang•ax frequency,• is the integrationvm'iablefor the inverseHankel
transform, I(p) representsthe Laplacetransformof the norm•liged current waveform
andisequalto -p- 1forstepfunctionturn-off,J• is theBesselfunction oœorder1, and
L• 1 istheinverse
Laplacetran-•form
operator
withrespect. to p; p -- 4•r.10-7
henriesireis the magneticpermeabilityof free-space
.
and i is the imaginaxyunit.

Evaluationof eq.(10.3.1)canbefacilitatedusingthe Gaver-Steh/est


methodgivenin
R•iche and Knight and the linear filter m•thod. Note that the manner in which the
currentis switched
off influences
the TEM records
andmustbe accounted
for during
data.a•alysis.H we incorporatethe effectof Tx turn-off time, to in our calculations,
then eq. (10.3.1) simplifiesto (Ra•che,1984)

= •'•za
Z•"(•) 2to.A
sIo'=G(•"
'-(T m)J•t•)•d,• (10.3.2)
for the central-loopconfiguration,and

(].0.3.3)
toJo
for the coincident-loop
confi•ation. In equations
(10.3.2)•ud (10.3.3),we havethat

G(•2r,m)- F(•2r,) - F(•2r)

whereF(•2•') = -Lp [, t0isthe


ramp
(turn-off
time)
, .A
nisthe
effective
axea
of the R.xloop, .Ar is the axe•of the Tx loop, and the normalizedtimes r and r' are
given by

186
(t--to) for the SIROTEM-t•e systemsand

t (t+to)
for the GeonicsEM37-type systems.
oyta2 : o./za2

and cr=cqin the layered-earth


case.DigitalFalters
(e.g.,Anderson,1979)may be usedto
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

performthe inverseHankeltramforms.
Notethatequation(10.3.3)maybe re-written
with a different kernel as

= a'l.m
Z•att) to o{G((•r,m)j•(()}.i•(()d( (10.3.4)

to effectthe nttmerical
integratio
n withAnderson's
(1979)distaltiltera.Thekernel(irr
curleybrackets)will thusrequiretheevaluation
of the .I t Besselfunction.

Algorithmfor calo,l:ii-g .40


The layered-eaxth
impedance function,Aois foundby a back-substitution
process
(see
Knight
andR•che,1982;
P•che,1984).
Forastructure
withn layers
above
a -•orm
half-spaceof conductivity
tr•+• (Fig. 10.3.1),thereflectioncoefficient
at the interface
between
layer
j ofconductivity
tr•andlaye/j+loœ
conductivity
•j+•is

Rj si- si+i (10.3.5)


where
si = (A• - io•tiai)
ll2. Defining.
theexponential
factor
Ed= exp(-2sdhd)
and
workingup from the bottomof thesequence
of layers,we calfate
F.+l = 0 (10.3.6)
for the basal half-space•
F. = (10.3.7)
fortheoverlying
layern,and
(R• + F•+•)
(10.3.8)
Fj =.E•(1+ R•F•+t)
for any layer j (0 < j < n) on top of layer n.
Usingthis recttrsive
relation,theftmction,40issimply

(/to + F,)
(10.3.9)
Zo= (1+ aoF,)'

187
B_nrnp
functionapparentresisti•y approxlm•ttions
Formodelling
purposesit is customaxy
to define
anapparent
resistivity
(reciprocal
of
•pparentconductivity)
forwhichnosimple expression
exists
andmayc•õe with
measuremen•time. The •yrnptotic expressions
for the stepfunction•ppazent
resistivity
as• -+0 (theso-caJ.led
earlytime)anda• • -• oo(latetime)axe•espectively
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(SpiesandEggers:
1986)

and
wherev is thevoltageat the1• coil.I istheTx currentandthe othersymbols
have
their usuM meanings.
Following
Raiche
(1984),
theramp-adjusted
apparent
resistivity
caubeapproximated
fxomthe modifiedexpression
for Z which,forthecoincident-loop
geometry,
is

oo (_)•(2k+2)!X

Z•(•)= 2•--ga•r•
l•x•l•• k!(
k+1)!(k+2)!(21c+•
)T,(Ic)
where

T,(k)= to/t(k+•/•
1
(1- to/t)•+•1•
-'1 andX-
crAta•
- 4fo
Forthe central-loop
geometry,
wehavetkat (e.g.,Sandberg,
1988)

Z•"(t)
=(•)•/2
-2to-' •tn
.A• ••0 (-)• r,•¾•/•-
= 4•/c!(2k+3)(2k+Si r•+a/ - (z0.3.zz)
Iterative
procedures
(e.g.,SpiesandP•che,1980;
Sandberg
1988)axeusedto correct
theapproximations
uatilthefinalvalue
oftheapparent
resisti•tyisfouad.
Wewill
nowshowhowto develop
ß
analytical
expressions
forthepartialderivatives
required
ia
TI•,M data inversion.

Computation
ofpatti•lderivatives
of,4o
Thepa•iMderivatives
required
intheinversion
processes
canbecalctdated
from
equation
(10.3.1).
Ex•m•g thisequation,
it isobvious
thatthederivatives
with
respect
tothemodel parameters,
mwillonlyinvolve
thelayered-earth
function
A0.
Thuswerequirethederivatives,
c3Ao/Orni
which m•ybeobtainedbydifferentiat•g
equation
(10.3.9)..&s
inthepreceding
section,
arecu:sive
relationship
forthesepartial
derivatives
canbedeveloped.
Noting
thatrn•maybeeithercr• or hi, thenforthe:/

3.88
layers on top of Myer u restingon the basalhaJf-sp•ce
(Fig. 10.3.1),we may write

0,4o OAo OF• OF• OF• OF)_, OF•


o.• = o•" • "or;'o• ....-•-•f ' o%ß (m.3.•)
The fi•st term on the right-handsideof equation(10.3.12)is obtainedfrom the straiõht-
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

forward differentiationof eq. (10.3.9)as

a.,4o
OF• _
-- (1+RoF;)
(1+-ReFt)
Ro(P•
• +?•)= (1+
(1-P•)• '
RoFz) (10.3;.13)
To obtainexpressions
forcqF.i/am.i
andcqFi/aF•+•, wesimplydi•erentiate
equation
(10.3.8).
First,
letusfind
3F•/c9•r•.
Here
F1isofthe
functional
form
F=W•where
.W,U,D and alsoF arefunctionsof •r. We therefore
uselogarithmicdifferentiation.By
this technique,

so that
w +v

a?•
a• = •i•,•
?•{•• + [(s•
-+s•+,)-
2(R• (s•
++s•+•)
F•+d(s• • ß
[(s•-
so+•)-
(s•
+'s•+D]
2(1+ R•F•+,)(s•
+

,s•+•Fi++•s•+•):}'
= •- r• • -(n'•'+•'+•)(•+,•+,): (1+//•+•)(s• (10.3.14)
For oOFffcqh•,
we simplyobtafm

(zo.a.•)
cqh•- 2s•F•
and usingthe quotientrule i• differentiation,
wefindthat

(zo.a.lõ)

Alongsimilarlines,for layern •nd the basalhalf-space,


we havethat

0.4o OAo OF, OF: OF._• OF.


(•o.s.•v)
Om•,= OF•' OF:' OFa'" OF,, ' Om•,
wherem• couldbe h., •,., or •.+x. Notethat equation(10.3.17)usesthe interme-
diateresults
obta•ued
forOAo/Omj
•boveandwetherefore
requireonlyOF./Oo'.,

189
OF,.,/ah. and aF./Ocr.+• . Theseare respectively
givenby

O•r.= s. F. h.-2(s._s.+•)
+2(•.+•.+•)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(10.3.18)
-- $n+l)

OF.
= -
Oh.

and

OF.
1 = E.
•O'n+' {(s.+s.+•)-•s=
(•.+-(s.-•.+•)-•.+•)

2iw•E.s.
($2n 2 ):2
+ $-+1
(10.3.20)

The paxtiMderivatives of theTEM response


issimplydeterminedby replacing,4oin
equation(10.3.1)with 0Ao/0m.givenabove.Notethat sinceAnderson's•te• weights
(Anderson,1979,eq.[5])areindependent
ofthelayer-sequence,
theyca• beusedto
evaluate these partial derivatives.

10.3.3.2 The MT Forward Problem


In 'themagnetotell-uric
situation,
a well-lcnwnexpressionfor
computing
theEM
imped•uce measurableonthesurface ofann-layered
(homogeneous
•udisotropic)
earth
is (Word, Smith •ud.Bostick,1970)..

•. •(1 - E•) + Z•+,(1 + E•) (10.3.21)


z•==•.•(• +E•)+Z•+,(1-
1

where
theinduction
paxmeter
7i= (iwPø'•)
• , ?eexpon•ti-I
factor
Ei = exp(-27ihi)
andfortheterminating half-space
Z. = (i•p•.)•. Forease
ofdiscussionletussimplif•
thisflirther.For thejthlayerofresistivity
Pi andthickness
h•, definea• intrinsic
impedance
w•- •-•= (i••): and
theimpedance
atthetopofthis
layer
looking
down as

190
j---n-l,... ,1 (i0.3.22)
Z•= w•.1+ R•.E•'
wherethe reflectionCoe•icient
R• isgivenby

R./--w.•
- Z.•
w.•+ Z•+
+t• (10.3.23)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

andZ, = w,. P•ecursive


application
ofeq.(10.3.22)
stoa-ting
fromthebottom(j--n-l)
yields
theimpedanceat thesurface
Z• forthedesired
r•uõeofmeasurement
frequencies
ßud these axe complexvalues.
For modelling
purposes,
it is customaxy
to deanetheC•gni•rdapparent
resisti•ty

(10.3.24)

andphaseof the impedance

(10.3.25)
= I,a.(z'l)j'

Derivation of MT partial derivative•


Sincethe evaluationof theMT forward.problem
is •lrnosta trivial problemonmodern
computers,
forthe-•tiated thepartialderivatives
required
forinver•on
maybe
easily
computed
usin$
afinite
difference
scheme
(e.$.,Meju,1992).
HOWever,
wewill
showhowto derivethe necessary
an•y•ical e•ressionsfor the partialderivatives.
First,it maybenotedthatoneca=inverttheapparent
resistivity
orphase
•n•ormation
either
sepaxately
or'jointly.
Wetherefore
require
•' az•d
• where
the'
model
pax•.rneters
m•axethelayer
resist•v•ties
andthicknesses.
Now,recaJl
thatp, and•baxe
derivedfxomcomplex
values
Z1;wedeterm•v•e
theixpartialderivatives
a•

o':gm,
=OZ,
'am./=
• g[Z']
'/w'ei•m•.J'
+lm[Z•].
jr (10.3.26)
•d

(10.3.27)
or

191
•'•Z•, AsintheTEM
Toevaluate
•heabove
expression,
werequire
thequantity
•omi•
problem,•or j -- 1, ... ,n-l, wemay write
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Ogx
•OZa
Og•Ogi_
•}OZ• (10.3.:29)

We c•n compute
thequantities
in curleybrackets
fromeq.(10.3.22)
as

OZi+• = wi

4w}E• (10.3.30)
(w•+Z•+•)•(1
+R•E.i)
2
Now

a•d
OU OD
Oh
s= 2R/7/E/
and =
Therefore,

Oh:i
= (i+RiE:i)

or if wewantthe samedeno•ator asin eq.(10.3.30)for computational
ease

OZ• 4iw•Es[wi'
0%'= (1+ •sBs)'
- Z•+•]{w
i +Zs+•} 4/•øPE•(ur•-Z}+
l) • . (10.3.31)
[%+zs+•]{•s+zs+•}= (•+•sEs)•(ws+zs+•)
Ne•,•d OZt/Op
i using
eq.
(10.3.22).
H•eZiisofthe
•ctiond
fo• Z=W•where
W,U,D•d alsoZ •e •ctiom ofp.Wethgdore•e 10g•t•c •••tiation.
By tMs tec•que,
az=$(3_
aw (10.3.32)

so that
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

o/•= z 'op•+ (•-n•j' op• '-

(10.3.33)

which simplifiesfurtherto

OZ• Z• (•2iwI•R•E)Z•
•p•=y•- 2 2 ßh{i•
- I•E•)• + (• Z)+2•+•il
- Z•

(10.3.34)
(•+•f•)•(•.+Z•+0• - • - +z•+ .
For the terminatingsubstrat•m,
wehavethat

(1.0.3.35)

a.ud

(10.3.36)

which
a•erespectively
similax
toeqs.
(10.3.26)
a•d(10.3.27)
with
3Z._2OZ._•,• -(10.3.37)
az._•••-. } a•,.
where
thequantities
inbracket
areh=nished
byeq.(10.3.30)
a•d
.1

OZ..._.o.,
= iw• (iwP/P,)
2 7, (10.3.38)
o•,. 2•Vi= 2 '= r '

193
10.3.3.3 The Schl-mberõerForward Problem
For the Schlumberger
dcresistivity
method,thelayered-earth
forward
response
is given
by (Koefoed,'1970)

•a(L)
=œ•/ T(A)J•(•L)AdA (10.3.39)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

whereL is half the currentelectrodespacingand the resistivitytrausformT(A) is


calculatedrecursively.For an n-layeredearth(Fig. 10.3.2),we start

p• h•

- Ti(•)
p• h•
Ti+•(•)

P.-I
. - -T.(),)
Pn (DO

Fig. 10.3.2A layered-earth


modelfor the dc Schl-rnergerproblem.

from the bottom and obtain the.resistivitytransformat the top of the in'hire
substratumas T, = p,. For notationalsimplicity,let us definethe exponentialfactor

Ei =

aud the reflection coefficient

(pi - pi+•)
R•= (p•+•o•+•)
'
For layern-1 restingonthissubstrat,,rn
thetr•norm .is(Koef•, 1970)

1- R._•E._i
• (0.3.40)

For anyotheroverlying
layerj (i.e.,j = n-2,n-3,...,2,1),
W• + T i + •(•)
(•0.3.•)
:r•(•)- • + w••'•+
where
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

The aboverecursirerelationis usedto calculateT(A) = T•(A) requiredin eq. (10.3.39).


Eq.(10.3.39)
iseasily
evaluated
byconvolution
using
digital
filters(e.g.,Ghosh,
1971;
Johansen,
197.5)ona computer.
Theproblem
isrecast
in theform

Pa=,k = •krnin T(X)f} (10.3.42)

wherethe coefficients
of a movingaverage
filter arerepresented
by f•.

Partialderivatives
fortheSchlnmberger
probl',--
Usinglinearfiltersforevaluating
eq.(10.3.39),
thederivatives
required
fortheinverse
calculations
may be obtainedsimultaneously.
as

(10.3.43)
Now,for layersj = n - 2, n - 3, ...,1wehavethat

OT
: {OT•
3T: OT)_i}
Om• OT•OTa'''OT
i.'
•Ti Omi (10.3.44)
Fromeq.(10.3.41)
andusing
thequotient
tale,wehavethat
OT• ((• + W•T•+]lp}).(aU/ah)
- (Ws+ T•+])- (10.3.4:5)
•= (• + wff•+ x/ p22
•)

_2 OE
where
O[I_OW 3W
i .3Ei OW
i __
so that

3'r = (•+•)•'

195
Followingthe sameline of reasou•ng,
we have that

8V T•+• 8U

giving
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

0Tj=
0% {½(1-T•+x/p•)'[(•+%)
+ w•r•+•/4)5' 4ApjE•
•'
(10.3.46)

Proceeding
in a similarfashion,we havethat

OT• (1 - w•/•)
• • (10.3.47)
OT•+ • (1 + W•T•+ •,/ p•2
• ß
For OT/Op,we applythemethod
oflogarithmic
differentiation
andobtain

2(W•+ Ti+i)WjT•+t
•o•= {(•(•-•+,/4)[w•
(10.3.48)
+ wF•+,/4)•J'•-7
+ (• + w-•+,/•)•f .' . J

As before,we may write for the terminatinghalf-space

OT•(OT•
OT2
OT,.,
= 1 andOT._
noting
that
Thus,fortheinfinitesubstratum
andtheoverlying
layer,weobtain
from eq.(10.3.40)

0T._• _ T. _• 4R._,E._•p._tp. (10.3.49)


op.---E•,
- •'. -1 (x+ n._ 1E._1)2(p._
• + pJ

aT._• = 4AR}
_,E._,p.'• (10.3.50)
Oh._• (1 + R._ ,E._ •)•
and

OT._,= 4R.,,_•E._,p•_• (10.3.51)


o•,. (• + •._ •s._ •)•(p._• +

196
10.3.3.4 The W•,--er Problem

The We•er arr•õement ismodelled Sy • similarsetof equations


to thosegivenabove
for •he Schlumberõer
configuration.
The simpledemonstration programfor Weber
inversion•ENINV •iven in theappendixcontains the forwardroutineFW• which
uses•he filter coefficients
of Biwenand Barker(1994).
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

10.3.3.5ApproximateParti• Derivatives
In all the aboveforwardproblemdescriptions,
the expressions
for computingthe
requiredfirst partialderivatives
withrespect
to the modelparameters
wereobtained
ßanalytically.
In all cases,
t•te difference
approximations
mayalsobeused.Notethat
if afinitedifferenc
e approximatio
n isused
withthesaidlinea.,'izing
p•rameterizations
thenthe computational.formul•
for'thepa.,'tiaJ
derivatives
m•y beof theform

•t,s
= •og(r•(•s+,•))
&- •og(?•(rr,
s))
, (10.3.52)
for the forwarddifference
technque(e.g.,Meju, 1992)or

log(F•(mjq-&))
-log(F.•(mi-&)) (10.3.53)
for •he.central differencetechnique.
The formertech-que hasbeensuccessfixlly
applied
to EM andDC resistivityinversion
by ridgeregression
(seeMeju, !992)aadis
implemented
in the simpledemonstration programWENINV listedin Appendix B.
Theperturbation
factor• isdetermined
empirically.
In WENIN• (andtheMT program
MTINV givenin Meju,1992),forexample,
a valueof 0.03isusedandthepertuxbed
modelparameter is .simply'

(ms. ,•)= antiloa(lo•(rn


s)* 0.03). (10.3.54)

It maybepointedout'herethatthelineaxiziag
parameterizations
described
earlier
demandsthat we use as deri,•ativesfor the appaxentresistivities

A•= Oarn•
lnp,,,,
=•I arn•
OP,,
' (10.3.55)

197
10.3.4 ResistivityModelConstruction Methods
For the resistivityinverseproblem,we statethe modelconstruction problemas:
"Given a finite collectionof inexactobservation•data, fi,•d • statisticallyacceptable
model.that expl•in.•theobservational
dataandtheirassociated
uncertainties
and
satisfies
a•y constraintsimposed
ontheprobl,vrn
by physical
considerations".
Sincethe forwardfunctionalis nonlinearwith respectto the modelparametersm,
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

an iterativeprocedure
mustbeusedto solvetheproblem.Forinstance,
in inversion
withnullpriors
orsmoothness
mea•es,thef•nnaliterative
formula
isgivenas

Titk+l = + +
or

where
• = (W•/+WAm•).
ThepartiM
derivatives
A•i= Oi(m•)/c3mi
areevaluated
at
m• andthe iterationisbegunat/c--l).ThepartiMderivatives
maybe calculated
analyti-
callyusingtheexpressions
derived
aboveorbysimple n•,merical
finitedifferencing
a•d
thequahtitiesontheright-hand
sideofeq.(10.3.56)
areeasilydetermined using the
singular
va•uedecomposition
method.Thecomputationaladvantages acrtting
fromsuch
anapproach arewidely
known.Theaboveinversion
method isflexible.
Forexample, if
wewishto interpreta combination
ofEM andelectfica2
or someothercomplimentary
dataset(e.g.,Vozoff
andJupp,1975;
P•cheet a1.,1985),
thenwesimply
replacethe
matrix A and the vector• with theaugmented
analogues
,4, and•. whichare of the
form

x,= )i ;

wherethesubscripts
1 and2 indicate
thecontributions
fromtheseparate
but complim-
entarydatasetstobejointlyinverted.
Notethatmorethantwodatasetscanbe
combined and inverted simultaneously.

In generating
smooth
models,
if wewishtoretainknownestimates
ofthesought
parameters
whilesmoothing
thevariations
betweenthe,,nknowns,
thena• alternative
technique
thathasbeen
foundtobeeffective
inpractice
istopartition
theaugmenting
data for A and !/in the form

198
0 eeß
ß

ß
h.1
ß

...o 1 o...
; flh= ft
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a•d the applicable


constraining
equation
is simplyDm= h whereI is of dimension
ixp, •)is ofdimension
(p-/-1)x
p, b•= -{•)rn}jandk = (p-l-1) andtherearel
known reliable estimates of m.
A commonand very effectivederivativeregulaxization
measureis to add a stoa/1
biasing
constant
ß
to themaindiagonal
of(WA)rWA givingtheiterativeformula(e.g.,
Meju,1992)m• +• = m• + [(WA)rWA+ ft• - •(WA)•vY•/]which
isthe
standardizedversionof the verypopularridgeregression
formula.

UseSdadditional stabili•iagoperationsin the solutionprocess


Duringparameter
updating,
some
auxillaxy
physical
constraints
maycomeintoplayfor
addedstability.
It isrecommended toemploy thesmoothnss
criterion
ofjackson
(1973)whichensures thattherrnssizeofth.eparameter
perturbations
isnotgreater
than ,,-ty. The smoothness
criterion,scis definedas
p !

where
z• axethecalculated
parameter
corrections.
Forexample,
ateach
iteration.any
perturbation
• greater
thani isregarded
astmsucces•
orphysically
,mrealizable
andmaybemultiplied
bya factor(<1) thatdecreases
thelength
without
'chaughxg
its direction:Thus,theformalestimation
formulagivenby eq. (10.3.56)is replacedby
the practica/formula

rn•+•
= m• + sc{[(W'A)•WA
+i•=H]-•
[(WA)•Wy-
f12rn•]}(10.3.58)
where
0< sc< 1istheadhocmultiplification
factor.Notethatwhile
thisoperation
prevents
thesolution
fromwildly"over-shooting"
thelinearrange,
it doesslowthe
convergence.
Alsoforphysical
reasons,
a commonlyusedconstrai-t
isthatthemodel
parameters,
m axenon-negative
scalarfunctions.

It mayalsobedesirable
to scaleeachcob,ma
ofthematrixA by therootmeans,,mof

199
squaresvalue of the coeiScients
asit canspeedup •he convergence
oœ•he i•erative
process(Marquardg,1963).However,this shouldbe doneonlywhenthe elementsof one
row are markedly differentfrom those in anotherrow becauseof the round-off errors
incurredduringscaling (Meju•1992).A simpleroutineforeffecting this techu•que
is
lis•edin Fig. 10.3.3andcanbe called•t e•chi•era•ionin an inversionprogram.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

subroutine SCALEJ(ncol,nrow,A,s)
scalesthe elementsof eachcolumnof the matrix A by the
rootmeansumof squares
valueofthe coefficients,
s•.

effectthe scaJing
procedurefor every cob,ranof A
do j = 1,ncol
s-m=O.O

obtain•hescMing
ϥctors,
s•
do i=l,nzow
slim=slim•- A(ij)**2
end do

scale
each
element
ofthej,a colnrnn
bysi
do i= 1,n.row
A(i,j)=A(i,j)/s(j)
end do
end do
return

end

Fig. 10.3.3A simpleroutinefor scalingthe Jacobiaumatrix.

Notethat the originalmatrixA is effectively


transformed
to a scaledversionA'. This
scaled matrix is

where
ß=
and the diagonalscalingmatrix is of the form

2OO
C •
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Notethat if • scalingoperation
is effected,
thenthe resulting
leastsquaressolution
mustbe de-scaledto obtainthe actualparameterperturbationsor directestimates.
Let Z=W'A, whereW is theweighting matrixof the observational
errorsdefined
previously.
Thed•rnped
leastsquares
estimate
oœ thescaled
parameter
change
vector
ßz, (seeeq. 10.3.56)is

= [(½-Z)T(C -'z) + fi'ml-'(U-

= o[rz + fir]-

Therefore,
to obtaintheactualpar•rneter
perturbations
weh•veto re.scalethe resultant
solutionin the Iorm

10.3.5 Model Apprais•


Duringmodelseaxch,anaccept•uce
criterion
maybethatthes-m of squares
ofthe
errorweighted
residuals
is lessth•un + vf•, (i.e.,withi-one
standard
deviation
of
themeanof theassociated
random va.•iable)
(seeParker,1982).Thecovariauce
informationis calculatedusingthe relevantexpression
derivedpreviously.
However,it
maybenecesssary to control
therateofchange (orthesize)ofthemostsquares
solutions
(e.g.,Mejua•d Hutton,1992;Meju,1994<1).
Forex-rnple,in extremization
withnullpriorswe couldapplyJackson's
smoothnesscriterion
yieldingthepractical
mostsquaxes
formula

min. = •m. 4- + -

wherethe symbolsm:ea.spreviouslydefined.

10.3.6 S•mple Application•in EM Data Interpretation


In theabsence
of a Fr/o• information,
a useSxl
strategyin theinversion
of practical

201
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

--- .....-
]

m Nñ8•o

202
r- --- ----!

0
n•

0
0
b•
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

iI


........
I
.....

•1IAIiSIS3Em
ß

2O3
Ill I

'88W
I. I I i il III
........

(•W!:l/^n)•P/SP OøI
I
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/


I
I I I i IlJJ
i

IñIAIñSI$3B
I

•u HId30

ßddl:l

204
i
I
I
i IiiJ
I
I
I
I
m IliJ

(z'•!:l/An] :•P/BP •øI


I
data may involvethe construction
of optimalsmoothmodelsusing• featureless
initia•
model(i.e,a ha2f-space)
whicharethenusedto guidethesearch
foracceptable
simpler
models.The underlying
philosophy
in thismethodof interpretation
is that any
significantfeaturesretainedin the smoothmodelmay be presentin the subsurfaceand
thus warrautsfurther explorationusingsimplerp•rametricinversionprocedures.The
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

resultsof suchan application of inversetheoryareshownin Fig. 10.3.4for TEM data.


Noticethe concordaut
features
present
in bothkindsof subsurface
models.Applications
ofthemostsquares
appraisal
scheme
in thedetermination
ofthesolution
envelopes
aud the extreme classof modelsfor the optima2leastsquares5-layermodelfor the
TEMdataareshown
respectively
in Figs.10.3.5a
(broken
lines)a•d10.3.5b.
It is
obviousthat the most-squares
modelsadequatelydescribethe field observationsand
their associated
uncertainties
aswell ashighlightthe variabilityof modelspace.

10.3.6.1
Jointinversion
ofTEMandMT soun•lln•
To demonstrateanotheraspectof EM inversion,we undertakea joint inversionof TEM
andMT datarecorded
•t thesame
station.
Th• MT dataareverysimilar
in thetwo
orthogon•measurement
directions
•ud of goodquality(Meju et al., 1993)and oneset
of apparentresistivity and phasecurveshasbeen chosenfor this exercise.The MT dat•
are in accordwith the TEM data- plottedat their equivalent•requencies (Meju,
1994f)- a.sshownin the lefthandpanelsof Fig. 10.3.6. Note that for eachmethod,
there are two interpretabledata setsand any combinationof dat• betweenthe methods
would sufficein a joint interpretationscheme.However,the optimal model shownin
thisfigurewasobta•uedby simuRaneous inversion of all the availabledata (i.e.,
•pparentresistivity,TEM voltagedecayand MT phasecurves).I• is easyto show
invertingeachindividualEM data segwouldhogyield a moredetailedsubsurface
picture.In order wo.rds,we identifya better modelby combiningvariousgeophysical
observations (Gol'stman,1976).The abovecombinedd•ta interpretationstrategyholds
goodif'ghere•re no lateralchanges in gheearth'sresistivitystructurein the
neighbou•hood of the observationalstation.Ig is quReCommon to find the TEM •ud
MT apparentresis•ivitiesin disagreement and this oftensuglgests •hat •here are lateral
changesin the subsurface geology in the vicinigyof ghepar•i•ar stagionand presengs
differenginterpretationp•oblem. Sohowca• suchdiscrepancies arisingmainlyfrom
static
distortion
ofMTdat•(analogous
gostatic
gime
skiRs
inreflection
seismic.
s)be
resolved
using
inverse
problem
theory
?Wewilllook
atasimple
inversion
strategy
for
dealingwith problemsof ghiskind next.

205
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

_
I
ii
I
i
i
IIII
i
i

ur• ll'I AI IS I S 28
I
I
I
Ill

I II
,
I

' dd!::l

206
I

I
ii

I
I
I
I
I
ill
!,
I
I

(•mU/An)•P/SP
Ii
I
I
Ill
10.3.6.2 Dealingwith the staticshiftproblemin MT interpretation
Field and theoretical studieshave shownthat in the presenceof small-scalesurfaceor
ß

neat-surfaceinhomogeneities,
magnetotelluric MT apparentresistivitysounding
curvesare typicallyshiftedby a multiplicative
factorthat is constant
at all frequencies
withthe impedance phasedatabeingunaffected. MT staticshiftis essentially a
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

galvanic
response
associated
withthree-dimensional
(3-D)bodies
(e.g.,Berdichevsky
andDmitriev,1976)andthedirection
andpossibly
the amounto:[shi•'depends
onthe
resistivity
characteristics
ofthe3-Dbodies.
Two-dimensional.(2-D)
numericalmodelling
.studies
suggest
thatconductive
inhomogeneities
causea downward
shiftwhilethe
ßreverseis the casefor resistivebodieswhichalsohave a slightlylargermultiplicative
staticshift(JOnes,
1986).Wheninterpreting
MT soundings,
conventional
wisdom
is to
firstidentify
anypossible
shifts
andthenseek
some
efficient
corrective
measures
forthe
biased
apparent.
resistivity
curves.
ß
An approach'
to solving
thestaticshiftproblem
in
the absence
of anycomplimentary
information
is to shiftthe suspected
sounding
curve
to somereferencelevel determinedwith the aid of somestatisticalidentificationor
characterisation
criteria(e.g.,Jones,1988).A more,mlqueextension
.
of thispractice
usescomplimentaxT
surface
measurements
(e.g.,Steinberget al., 1988)or a pr/oridata
fromboreholes
{althoughaccess
to borehole
informationis oftenlimitedin maxty routine
surveys).

Thefavouredapproach involves
theusecomplimentary surface TEM databut unlike
theconv.entionmltwo-stagetechniques
involving
firsfly,the determinationof the correct
apparent resistivity
levelusing
TEM dataandshifting oftheentireMT apparent
resistivity
curveby a constant•mountbefore
interpretation (e.g.,Sternberg,1988;
Pellerinaxtcl
Hohm•, 1990),wewill adopta one-staõe tech,•iqueinvolvingonlythe
undistorted
andeqm.'valeut
MT phase information(Meju,1994e). Theunderlying
philosophy
hereisthatif therightlevelof'apparent resistivity
isfttmishedbysome
Othersurface
method,thenoneonlyneeds to jointlyinverttheMT phrase andthis
complimentarydatasettorecover thecorrect subsurfaceresistivity
distribution
thus
eliminating-the
needforapparent resistivity
datacorrection andleading %oa simpler
andmoreefficientinterpretation
scheme.
Therearesomeeffective
practical
considerations
to beborne
inmindwhenadoptinõ
thisapproach.
Forthejointinversion
problem,
ourpopular inversion
formula
forsingle
datasetsexpands
to (Meju,1994e)

207
Y' = = •,-f•,(m")] (dimension:
[nt+ nqb]
x 1)

and
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(dimensions:
[nt+

for •he nt ( > 1) TEM apparentresistivitydata,n• phasemeasurements,andnpmodel


parameters. Thesubscripts
t and• denotetherespective contributions
•o thesystem
of equations fromTEM andMT phasedata.I• shouldbe notedthat onemayalsoelec•
to inver• the phaseinformationfor the two orthogonal
measurement directions
wi•h •he
TEM data for an effectiveinterpretivemodelfor the stationof interestusingthe above
formulations.

It isgoodpracticein inversion
to scaleeachobservation
byits associated
uncerta/nty
so
thatpoorlyestimated datahaverelatively.
lessinfluence
onthesolution.
In this
situation,the weightedsolution
is simply(Meju,1994e)

T (A).
:,:= [(WA). W + 5:rl--1 -

wherethe diagonal
weighting
mat;fixW contains'
the-reciprocals
ofthe observational
errors.
The •boveparmetercorrections
ß areapplied
to møin successive
applications
to recover the desiredmodel parameters, m.
However,
i• maybenotedtha• thesetwosetsofdata•o beinvertedhavedifferent
magnitudes.
It istherefore instructive
•onorm•l•se
•hemto some common-scale
before
theyareassembled in It. &udA.. Now,asin conventional
simul•aueous
•versionofMT
apparent
resisfiviW andphase da•a,•helogari•hm•
oftheTEM da•aareconsidered
in
this schemewith the actualphasedata, •ha• is,

•h= {In d, -In ft(mø)}


and

A, = O{in i(raø)}/arnø.

Thecomponents ofrnarealsotakentobethelogarithms
oftheresistivities
and
thicknesses
ofthesought
subsurface
model.However,since
thephasedatacould vary
from0 to 90say,forsome
givenproblemwhilethetransformed
apparentresistivities

208
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/


I

I
I
I

L•
I
!

I
I II I

I
I!
u• H.Ld30

209
!

11IAIISIS3B 'ddld [zmld/An]:•P/BP


Søt
9*
wouldtypicallya.Ssl•rne&rangeof valuesthat is lessthat one-tenth
of that oi thephase
data, a usefulpractica•strategymightbeto normalizethe contributions from the phase
data to the systemof equations by someconstantfactor(~ 10).

An applicationof thisinversionstrategyto TEM andbiasedMT data from the same


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

stationin a sedimentarybasin(Meju, 1994b)is shownin Fig. 10.3.7.Noticethe goodfit


of the optimaljoint modelresponses to the fieldMT-phaseandTEM data. SinceMT
phasedata shouldbethesamein bothdirections in an ideallayeredearth situation,it
is a goodpracticeto seeka representative modelby invertingthe combinedphasedata
in areaswhere a horizontallystratifiedstructurecan be ass•medas a workingmodel.As
shown
in Meju(1994e), another
advantageofthisapproach
isthatTEM dataoflimited
bandwidth
(i.e.,withlittleornooverlap
withMT phase curves)
maysuffice
for
recovering
the trueresistivity
structurefromthe MT-phasedata.

10.3.6.3 Joint inversionof central- a•d coincident-loop


TEM sotmrlins
As a way of addressing
theproblemof lateralvariationswithinthe Tx loopareain
TEM inversions,it may be usehtlto definean effectiveresistivitydistribution
underneatha particularsiteby jointly in.vertingthe central-and coincident-loopdepth-
sounding
dataasillustrated
inFig.10.3.8a.
It isclear
fromthisfigure
thatthecentral
and coincidentloopdata arein excellentagreementsuggestingnegligiblelateral
variationswithin the areacoveredby the Tx loopfor the sampleddepth zones. A local
axealaveragemay Msobedefinedby thejoint inversion of multiple-station
(i.e.
combined) dataasillustrated
in Fig. 10.3.8b.Sucharealrepresentativemodels may
find usein regionalsurveys
wherethe gross features
of the subsurfaceresistivity
variations are desired.

10.3.7 Inversion of dc Resistivity Data


All the inversionapplications
demonstrated
aboveusingEM dataholdgoodfor dc
apparent resistivitydataandit wouldservenofurtherpurpose to recountthemhere.
However,it maybenotedthat it is common practiceto simultaneously
invertdcand
EM datafor bettermodelverification (e.g.,Vozof[& Jupp,1975).It mayalsobe
necessary to scaletheelements of the Jacobiau matrixif theydiffermarkedlyin sizefor
the differenttypesof datainvolvedin thejointinversion process.
The computer
program WENINV listedin theappendix performs onlysingle-stationdata'inversion
butcanbeeasily
modified
tohandle
multiple
datasets.

210
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Mmm
m

-
•_ -

• mm

Nmm m
ß
,mm

.m

,m

,m

,mm

I
i ßi ,
J i i iii ! i i i i iiiii i i i I illll i i i i ill[

m Hi•30

lB
II

9'
i ii

• i

• 11IAIISIS3B '88B [zmB/An] •P/BP m•oI

211
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

I
!

--
..
m

.,

m
•)mm•

m
mm

i,

mt• IIIAII•ISqE)

212
'ddW
I/

(zmU/An) •P/SP
10.4 Estimation of EffectiveDepth of InferenceUsin• SmoothModels
10.4.1 Aspectsof smoothmodelconstruction
P•ecallfrom Chapter 6 that smoothmodelsmay be constructed
by minimizingthe cost
function
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

= +/(mTlrm)

or the mathematically
equivalent
function(since/•is undetermi•ed)

(lo.4.2)

The constrainedlineaxizedsolutionis simply

(10.4:.3)
or

't]'l,
= [(W'A)T(W'n)-{-
•2.H]-1(W..t4•)T•i
.

where
d• = {(Wy)+ (WA)mo}
•d theothersymbols
havetheixnsua•
mea.uinõs.

Implementationof the smoothmodelalgorithm


The smoothmodelalgorithmis simpleandcanbe readilyimplementedas a biased
estimationprocedure.Basically,the first difference
operatorD is premultipliedby f•
(chosen
asa valuelessthanorequalto •mity)andappended
ontotheactualweighted
matrix of partial derivatives(WA). The weighteddiscrepancy
vectorW•t is also
augmented withtheextraneous
data(-/•Drnaor simplyzeros)formicõIt.- Themain
taskhereis thedeterminationoftheinversematxix[(wA)TWA+ •:H]- • •-
(A• A.) - •. UsingtheSVDtechnique,the(n+/)xp augmented matrix,A. is
decomposed intoa product
ofthreeothermatrices,
ß
i.e.,A, = U'AIfr , whereU'andV
axe'thedataspaceandparmeterspaceeiõenvectors respectivelyandA is a diagonal
matrixof the singularvaluesof A.. In termsof $VD, the constrained
leastsquares
solutiongivenby equation(10.4.3)is simply

Note that sincewe haveregularized


A by augment•git with'D, it is non-singular.
However,depending on thevalueof theregula•zation
parameter• used,someof the

213
singularv•ues of the regula.
rized derivativesmay still be quitesmall producing
undesirable
effects
oninvertingthematrixA•A•. A simple•ud effectiveroutinefor
implementingthis procedurePLOP, is givenin Fig. 10.4.1.It c•u be usedin •uy
conventionallayered-e•th inversionschemeto •enerate the initial model and the
augmentingequations for smoothmodels.It is calledoncebeforethe iterationprocess•
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

note that while the actualda•a segments


of A a,udy a•e modifiedwhen newpa•iM
derivatives•nd datamisfitvectora•eevaluated,theiraugmenting
data a•e no• changed
in the abovesolutionprocess.However,a routinefor appendingthe appropriate
extraneousdata ontoA •nd •, FSTDIF is includedin PP•P a•d may be calleda• each
iteration. The depthestimatesreturnedby PKEP shouldbe held const•ut in the
inversionprocesswhilethe modelresistivitiesa•e optimized.

subroutine
PKEP(n•ow,beta,A,Y,rm,
dm,nL,ncol,ndat)
constructs
a•uapproximate
initialmodel
forsmooih
MTinversion
c Input: jacobiaumatrixA, discrepancy
vectorY, actualnzowsin A
c and dampingfactorbeta.
c Output:rm,dm--initiMmodelparameters (nn--resistivi•y;am--depthsin metres);
c augmentedA,Y, ncoland nrow.ofnewA,y: uses•he routineFSTDIF.
c Returned modelhas nL layers •nd nD interlaces.
c Routinesetfor 50 datapointsand 20 layers.Thus A((50-1-50-1-20),(20-1-20))
c a,udY(50•50-1-20)to aJlowfor joint datasets.changea• requiredbut keep
c physica•dimensions of arrayssamea• in caJlingroutine.
c Nrow-•nfreqwhenfittingonly app.rhoor phase;n•ow--2.nfreqwhen
c fi•tingapprhoandphasesimultaueously.
Fielddatastoredin commonblock
c FDATA. AR.l=app.rho;Ar2=apprhoerrors;phi=phasein degrees.
c ph2=phasee=or•, •eq=frequenciesin Hz.
common/FDATA/Ireq,arl,a•2,phl,ph2,n•eq
realfreq(50),arl
(50),a•2(50),ph1(50),ph2(50),a(120,40)
re• y(120),m(20),dm(20),db(50)
character, 1 aus

ndat=n•ow
nL=20

nI.,2=nL-2
nD=nL-1
ncol=nL

nrow=nda. tq-nD

214
write(*,'(/A)')' pleasenote ßinitial modelis a half-space'
write(-,'(/A,$)') ' wantto specifyhalf-space
resistivity? y:n'

if(sms.eq.'¾'.
OR.•us.eq.'y')
then
")')
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

read(*,,)rhoH$
•1S•

ca.lcul•te•ver•ge resistivity•rom givend•t• a•l


rhoHS=0.0

do 1 i= 1,nfreq
rhoHS=rhoHS+arl(i)
continue

rhoHS=rhoHS/afxeq
endif

write(*,'(A,$)')' wantto specifyrain/max depths? y:n


read(,,'(A1)')ans
IF (ans.eq.'Y'. 0 tL•.•. eq.'y')then
write(,,,)' enter approx.top and bottomdepthsin metres'
11 read(, ,,) sma,big
if(big.LE.sma)then
write(*,*)' waxning:yourmax. depthis lessthan rain depth'
write(,,,)' enternew depthsto top •ud bottom of ha.If-space'
GOTO 11
endi/

top=alog10(sma)
bot=alog10(big)
ELSE

find minimum andmaxim,,m penetrationdepths


usingBostick's'
algorithm
pi=acos(-1.)
amu0=4.,pi,l.E-7
do 2 i= 1
f= 1./sqrt(2.,pi,•mu0,freq(i))
db(i)=sqrt(arl(i)),f
continue

big=rib(I)

215
sma=db(1)
do 10 i=l,n•eq
st=db(i)
if(st.GT.big)•hen
big=st
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

elseif(st. LT. sma)then


stoa,----st

endif
10 continue

sm•=sm•/nL
chanõe•bove line to $MA--10.**LD0-1-0.1 •f LD0 is lo• of
minimum plottin• depth•n your plottin• routine
top--aloõ10•sm•)
b'ig=big*l.8
bo•=alog10(big)
ENDIF

constructsmoothmodel.first put rm=rho-halfspace


do 20 I=l,nL
rm(i)=rhoHS
2o continue

next compute depth rangemudsubdivide•o ge•


.ND equi-spacedlayer boundaries
mg=bot-top
dd=rng/nL2
do •0 i=l,nL2
k=i+l

dt=dd, FLOAT{i)
dm(k)=10**(top+dt)
4o continue

am(1)=sm•
dm(nd)=biõ
c setup augmenting
equations
in jacobiaumatrixA and
c discrepancyvector y
callFSTDIF(nn,m2,beta,ncol,ndat,nrow,'A,Y)
return

end

216
subroutineFSTDIF(nn,m2,beta,ncol,ndat,nrow,A,Y)
c first differenceoperatorfor smoothmodels.•ugmentA and y.
real A(au,m2),Y(nn)
nfl=ndat+l
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

do 10 i=nfl ,nrow
Y(I)=0.0
I1Z=i-nda. t
IR2=IR+I.
do 20 j=l,r•col
20 A(i,j )=0.0
A(i,IR)=l.0,beta
A(i,ItL2)=-l.0.bet•
10 continue
return

end

Fig. 10.4.1 A simple.


datapreparation
routineforMT inversion
forsmooth
models.

Determintion of suitablestaxtingmodels
The main attractionsof inversionfor.smoothmodelsaxereflectedin the COrnrnnly held
viewsthat the optimalmodelsare independent
of the initial half-space
modeland •h•t
anysignificant
features
resulting
therefrom
maybejustifiedby the data.It is not
difficult to conceivethat inversionfor smooth modelsis most effective if the search is
initiated with • half-space
modelor a smoothmodelderivedfor instaucefrom a simple
resistivity-versus-depthdatatransformation
(Meju,1988);theminmizationpath
is well defined•ud the smoothnessconstraints,thoughcontradictoryin a senseto the
requirement
to satisfythed•ta, le•d to a stableinversion
ß
process.
Irnportautly,
since
the solutionis biasedtowardsthe initial a pt/or/information,the reconstructed
smooth
model will retain someof the desirablefeaturesof our initial smoothmodel; mudbecause
the data (providin theyareof reasonable quality)tendto predomintesomewhat'
ß
over
contrastinga prioriinformation wein effectretrievea smoothmodelthatis consistent
with our observations.It is easyto showthe relianceof suchmodelson the initial
conditionswhich is contrary.to the popular belief.
In Figs.10.4.2and10.4.3areshown theresults
ofinversion
of synthetic
magneto-
telluric(MT) datausingdifferent
initialmodels
withtheearth-model parameterised

Z17
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

I
I
I I Ill[
I

I
I
I
i I IIII

ur• illAIl$lS•W
I
i

m Nl•qO
11JI III

' ddW
['õ•P]
i I I I i111]

•SWN8
i i i IllllJ

z
u

n-
(:3
L•J

L•J
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

i
i
,

m• 11I AIISIS2W
' 88B

219
[.õep)
3St•Ncl
z
u
)...
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ß
J
LJ


i IllIll

I
I
I
L..J _

i Illill

I
l

111AII•I•3B
"i
m

I I I'lill
i

"' Hi,:!3Ci

'ddl:l

220
I illIll
I

I,,4
I i iiill
I
I I JJllll

(- 6sp] 3SJ:lHcl
!
.
o
o

o
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

i i i illJ
i
i

Ni,:l:3a

IN•! U I::1

221
i i i i

I
i
i

•1•!1::1 I:-•ep]
I
i i i illJ
into a succession
of twentylayersof constaut
thickness
(in logspace)audwi•h •
assigne
d a fixedv•lueoœ
0.1.In all thefigures,
theinitialmodelisrepresented
bylong.
dashedlinesandthe optimalmodels by solidlines.The truemodelusedto computethe
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

testd•ta is alsoshownin Fig.10.4.2a•ho•dashedlines).I• is obvious


that only
modelsreconstructedfromthesmoothiniti• models(Figs.10.4.2aand 10.4.2b)
retained the desiredsmoothfeaturesandresemblethe true modeladoptedfor the
exercise
(shownonri•hthaud plo•ofFigure10.4.2•).Asshown
in Fig. 10.4.3,thetest
employinga non-smooth
initiMmodelwasunsuccessful.

Convergence
The biasedestimation approach described
aboveoffersa computational
advantage in
the sensethat a fixedv•lueof the quadratic
factor• is usedthroughoutau inversion
application.
Note,however,
thatin theaboveexamplesit wasdeemednecessary,
asa
practical
constrmt,toplace
a bound on.the
sizeoftheparameterpertl=bations
using
Jackson's
(i973)smoothness
criterion.
Thusthelengths
oftheperturbations'were
decrease•l-4o
about 30%of the actualvalueif they exceededsomefixed value. This
operation
decreased
thespeed
ofconvergence
butprovided additional
stability
to the
solution
process.
Thesmoo•hmodelshownin Fig.10.4.2aw• generated
afteronly
iterations.

Havingaccepted
thathalf-space
models
areconvenient
starting
points
forgenerating
smoothmodels,it is instructive
to seewhether
the algorithmwill converge
to thes•e
orsimilaroptim•lmodelfordifferenthalf-space
resistivities
fora givendataset. The
method wa•applied totheMT COPR. OD(JonesandHutton,1979)apparent resistivity
audphase
datausing
the100f•m
and1000f•m
initialmodels
andtheresults
areshown
inFig.10.4.4:.
Notice
thattheoptim•l
modeh
f•omthetwodifferent
start•g
models
are
exactly
thesame
(X•=29.3
forboth).A value
of0.õwasemployed
for• andtheearth
modelwa•par•meterized
intoa succession
offortylayersin theseexamples.

10.4.2Geom•c interpretation
ofsmooth
inversion:
intuitiveccn•ts audimpHcations
Thequadratic
factor
• andthemeasures
•/•and•/•a•etheessential
in••ents in the
generation
ofsmoothmodelsasillustrated
inthefollowing
interpretative
aualysisaimed
atprovid•r,g
a clearer
lmderst•.ud•,•g
ofthesmoothmodelconstntction
process,and
especially
thoseaspects
thatmayhave practical
implications.
Letusconsiderfirstthe
hypothetical
situation
wherethedesired
solution
canbefoundinonestepasthe
interpretative
analysis
wouldthenleauonthatprovided
forconstr•;l•ed
linearinversion.

222
Noticethat equations
(10.4•1)
and(10.4.2)maybewrittenin shorthand
œomas

=
and

(10.4.6)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Themeasureq•is zerowhenm•=rr•=...=mv, thatis,Drn:•O asforaz•initialha•.-


space
model(• situation
thatfulls oneof •heconditions
in ourproblem
definition).
For
easeof discussion,
it will be assumed
that thereis a pointin model-space whereq• is
effectively
minimized anddenoted by B'in Fig. 10.4.5.Similarly,q• artaims
a minimum
at someotherpointin function
space
denoted
by A in thisfigure.Thereaderis
reminded
thatin realitythecontours
(i.e,l-dimensional
hypersurfaces
ofconstant
andq•) in Fig.10.4.5canbehighlydistorted
in shape.
Geometrically,
thesituation may
be betterimaginedasa scenerywith chains
of mountains
andvalleysof variousform.•
•ud sizeswith the locations
of the minimaof ql (a•d sometimesq• ) completely
•,•l•,•own.Again,foreaseof discussion
ß
these•i•a arelikenedto twomaltavalleys
in the conceptualchain,oneof whichis a large-scale
ß
depr.ession
.
or regionalfeaturewith
thedeepest
pointcorresponding
to themi•m-min qt (rePresented
bypointA in Fig.
10.4.5).
Thedeepestpoint
intheothermajor
valley
wouldcorrespond
topointB.If
equation
(10.4.5)
or(10.4.
6)isru•tu•.ed
usin•
theinversion
formula
(10.4.4),
it is
conceivablethat for'smoothinitial models,the solutionwill in generaltravel from B to
ß
ß

A as• isvaried
from•u iu•u•tely
large
value
•ozero,
sayin • line-search
scheme.
Since
•hequ•utitiesq• •nd q• c•ubecomputed forthesolution
obtainedfor anyparticula•
v•lue of/•, the travelpathof the smoothsolutions
canbe monitoredas• is varied.
Sucha solutiontrajectoryis alsoillustratedin Fig. 10.4.5for the hypotheticalsituation.
A characteristicfeatureof nonlinearproblemsis •hat the travel path of •he solution
maybedottedwithminordepressions
and•idges(i.e.,a rugged
topography)
preventing
thejourney
sometimes
fromreaching
a successful
comPletiOn.
Another
undesirable
featureis that • f•lse trail maybe followedin the searchfor a m•uim•rnresultingin an
unwanted solution, i.e., thesearched
valleysaredifferent from.ourtargetvalleywhose
bottom-point
isA.However,
theuseof.a l•r/or/information
(J•ckson,
1979)
inthe
search processmayalle•atesomeof theseproblems assuchpriordataoftenhelpdefine
a fruitf• search
direction
aswella.sease[hetravelprocess.
In theabsenceof a priori
information,anyminim•-seekinõ
algorithmcan•t bes•hopeto •voidsomeof the
minorvalleys
in •hetravelpathof thesolution
butthereis Usually
noguarautee
thedeepest
valley
willbesearched
oritsdeepest
point
foun
d. Inthesmooth
model
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

I
I
I

I
!
I
I

I
I

!
I

I
I

224
problem:we haveincorporated priorconstraints on the desiredformof the solution.
Apropostherefore,is theselection
of suitableinitial modelsin inversion
with smooth-
nessconstraints.
Metaphorically speaking, thesmoothness measure f•(mrHrn) is a
good
pathfinder
butmostly
forarestricted
class
ofprospectors
- smooth
initial
models.
However,it may be pointedout that not all smoothmodelsconstitutegood.initial
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

models.Owingto the lineaxizing par•meterizations


employed in thesolutionprocess,
the suitabilityof the initialmodeldepends
onthe lengthof the potentialtravel path in
functionspace,•.e., between ourhypothetical
mi•ima A andB (a distancedubbedAB
here).Forsmoothinitialmodels,
thesizeof AB spellsthedifference between rapidand
slowconvergence
or between success
•ud failurein inversion
for there'is a threshold
dist•ucebeyondwhicha giveninitialmodelbecomes unsuitable
andthe algorithmmay.
fail to converge
to a desirable
solution.
If theinitialmodelis notsmooth,thenq2may
attain a minim,• at severallocationsin the p-dimensionalspaceand the solutionfaces
a moredifficult
journey
over•herugged
terrain.
Theendresultmaybeundesirable
and
the solutionprocedurewouldrequirespecialpath-findingmeasures
suchas'hybrid
inversion schemes.

Extendingthis over-simplified
analysis
to themoregeneralsituationwherethe solution
satisfying
the constraints
q•=q•axedq2= 7 sayis soughtin Fig. 10.4.5,it caxx
be
deduced
thatthedesired
point
inourq•valley
isposit
notatA butatsome
other
point
C, sayonthe valleysideneaxer B . To getto thispointfromB requires a suitable
choiceof the quadraticfactor•; onecane•ily overshoot the pointC (thusmissing the
desired solution)by usingcoaxsestepsin (or,mrealistic valuesof)•. Thedense•-
s;tmplinõ procedureadopted by Constable et al., (1987)maybe theIavouredstrategyin
this situation.However,sincethe solutionmustlie somewhere betweenA iredB in our
hypothetical analysis,
wemayrestrain therangeofvaluesof • required in a llne-search
process.
Now,depending
ontheinitialmodel(withql = qo,say),wemayobtainthe
undesirablesolution C'øwhich alsosatisfiesthe smoothnessconstraint. This may be the
case/ornon-smooth
initialmodels
(wherethepointB may beenclosed
by 'theql = q0
contour
inFig.10.415)
anda fortioriforusing
half-space
initialmodels.
In theforegoing analyses,
it hasbeenemphasised
that a suitable
choice
ß
of • andan
appropriate
initialmodelarea •/nequanonfortheinversionforsmoothmodels.
It isstressed,
however,
thatWhile
thiscontention
iscorrect
forthebiased
estimation
approach,
it doesnotimplythatnon-smooth
initialmodels
will always
leadto failureor
non-convergence
in smooth
model
construction
schemes
since
it ispossible
tø adapt
someothereffectivepractical
strategies
to suchproblems
(asdiscussed
later).
Subsurface characterisation at the datum-scale

On the interpretationalfront, the behaviourof the optimal solutionsreconstructedfrom


half-space
modelsforvariousvaluesof/• canalsobe visualisedwithin a geometrical
flamework.Here, theyare analysedfor consistency
and/or dependence on/• with a
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

viewto derivingsomemodelinterpretational
guides.
Notethatin everysense
generat-
ing the smoothestmodelentailsforcingthe solutiontowardsconformitywith h=0.
Thusinitiatingthe searchwith a half-space
model,asis commonpractice,is consistent
with the adopteda prior•constraintsin equations
(10.4.1)and(10.4.2). Sincethe form
of the solutionis welldefinedabinitio,thestabilized
successive
approximations
simply
serveto improvethefit to the givendata;the solutionis usuallysafelyguidedtowards
somesort of minim•m in ql. Obviously,the degreeto whichthis initial constraintis
satisfied
depends
onthevalueassigned
to/•. Interestingly,
asthe influenceexertedby/•
on the travel path of a particular solutionis counteractedby'the requirementto also
satisfy
thedatatObeinverted,
a useful
attendant
feature
isthatforsome
values
offl
the form of the resultantmodelwoulddiffersignificantlyat thosesubsurface positions
wherethe solutionpathsare not constrained by the observational data (i.e., the under-
determinedparts) but may be similarelsewhere. If this contentionis correct,then the
subsurfacemay be adequatelycharacterised at the dat,•m-scaleusingthe information
pooledfrom the smoothest modelsconstructed for a rangeof •. It may thus be possible
to estimate an approximatedepth of investigationfor a givendata set without recourse
to rigorousmathematicalanalysis(Meju, 1993)as demonstrated next.

10.4.3 Pooled model interpretation: the effe•ve depth of investigation


Let •s examinethe morphology of the solutions
in/•-space. The relationshipsbetween
the modelsderivedfor a rangeof/• valueswereshownin Fig. 10.4.(5for the COPROD
-.

data setsrespectively.bloticethat the modelsshowconcordantfeaturesin the depth


range ,,, 14kinto ~ 3001tinbut di•er signific•tly outsidethiszone.Meju(lf}93)
interpretedthe discordancein structureoutsidethis depth rangeas being due to the
fact that the modelsare not constrainedby the data at thosepositions.This interpre-
tation
isinaccord
with
thema••,• depth
determined
byParker
(1982)
to•vhich
any
modelcan be constrained
by the COPt•OD data and is alsoconsistentwith the most-
ß
squares
analysis
ofthedata(MejuandHutton,1992).l•otethatsuchconsistent
patternshave beenobtainedfor a varietyof resistivitydepthsoundingdata. In the MT
situation,the patternmay be diffuseat depthfor thosedata setsir• which the models

226
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

D_

i
I I I i I III
i
i

227
ß
I I I I III

[ -6ep)
[ I
I
i
I Ill
I
i
I
o

o
c)

r-I
0
0

0
o

o
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

mm

I
I I I IillJ
i


i I i Ii!11

Hilda

::::'28
-
,,

i11

,/

(•WU/Art):)P/EIP
i
9'
E
o)
o
showa mono•onic•llydecreasing resistivitywith depth;bu• evenin suchcases,some
discernible
trend is presentespecially
in the.uppersectionof the geoelectricprofiles.
Sucha methodmay thushaveapplications in simpledeterminations of the approximate
depth
ofinference
fora given
setOfpractical
data.Anobvious
method
ofefficiently
derivingthis usefulinformationwill be to generateoptimal smoothmodelsfor two
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

v•luesof fi that are aboutan orderof magnitudeapart as illustratedfor TEM data in


Fig. 10.4.7. The agreemen•
between
thisTEM interpretation
andthe resistivity-versus-
depth information derivedusinga recentlydevelopeddirect-depthTEM data
transformation
technique
(seeMeju, 19994a,19941')
is obviousin thisfigure.

10.5 SimultaneousInversionof Multi-Station data with Spatial Constr:•ints


Consider.
the processof interpretinggeophysical
soundingsfrom a profileacrossa
sedimentary
basin.In the absence
of anydeformational
activities,•ud assuming
that
the subsurface consists of a successionof horizontal or near-horizontal beds of sedime-
ntary rocks- an ideal situation- we expectth• physicalpropertyof a givenrock bed to
befairly uniformlaterally,onlydifferin•significantly
from that of the traderlyingor
overlyingbed. The interpretationof field measurements •donga profile in this caseis
o

a straightforwardtask. Unfortunately,in most basin exploration situationswe find that


there are lateral changesin the E•'th's structure•ud this introducessomebias in the
field data. For example,in resistivitysurveying,the recordedapparentresistivity
soundingcurvesat somelocations •1on•the s•id profilemay be shiftedalongthe
verticalaxesor showa l•r•e scatterin theirdistribution.
In the magnetotelluric
(MT)
situation
thismaymanifest
asa static(frequency
independent)
shiftoftheapparent
resistivitycurvesbut not the EM impedancephasedata but the directionand
magnitudeof the shift axenot knownexceptwherethere are someothercomplementary
informationavailable at the givenobservationalsite. The inversionof sucha biased
•nd/or noisedegraded datasetwill notyielda reliablepict.ureof the subsurface
at the
artectedsite. In the MT case,it is knownthat a plethoraof modelscaureproduce•
givenphasecurveandsomeinformationnotcontained
in thedatafromtheaffe•ed
observational
sitemustbeusedto determine
thecorrectlevelsof.theapparent
resistivity
curves
ortheappropriate
modelparameters.

The recoveryof the correctparametersfor a biaseddat• set selectedfrom a pool of


otherrelated
da•asets,some
ofwhich
axeunbiased,
canbeposed
asaninverse
problem
(Meju,1988).We simplyusethereliabledataasa priori information
andimpose
some

229
penaltieson the solution
for the biasedsite.An effectivepenaltyIor sucha site is that
the solutionbe ascloseaspossible to thoseof the neiõhbouring
observational locations.
We thereforestatethe optimizationproblemas:
Givensetsof conflicting observationsfxomdifferentpositionsalonga surveyline,
someof whichsatisfysomenormaJity conditions,
fLud.thesmoothest lateraldistribution
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

of physica2propertiesthat explainsthe observations.

Here,we axeinterested
in obtaininga l•terallysmoothinterpretivecross-section.
We
startoffwith • smooth
profile(i.e.,thes•memodelis usedforall sites)andinvertall
the variousdat• setssimultaneously
with the constraintthat differences
in the layer
parameters betweenphysicallyadjacentstations
beminimalandthat thesolutionsbe
statistica•y
stable.This,in effect,maybetermed a l•-dimensional
problem andthe
inversiontechnique
willbedemonstrated usinga two-station
example(extension
to
higherdimensionsisstraightforward)whichin theMT casemaybeeffected by the
simultaneous
inversion
of the phasedatafroma problematical
siteandthe unbiased
•pparent
resistivity
andphase
datafromtheneighb0uring
location.
Thedataand
constraining
equations
arepartitioned
in.the
formA.z.= y. andgivenby(Meju,
1988)

(10.5.1)

where
A. isan(,+n'+p)x(p+p)
global
matrix
containing
theparti•derivatives
A•,A•
plusthepx2psmoothness
matrixforthetwostations,z. contains
thesolution
vectors
zl andz2eachofdimensionp
x1, andthevector
It. contains
thediscrepancy
vectors
Itl(ux1}
andY2(n'
xl)plus
theaugmenting
datahi,...,h•,
which
arethedesired
differences
in thevalues
ofthecorresponding
layerparameters.
In thisexample
h isthenullvector.

Thecoupled
optimization
problem
isstatedasminimize
theLagraugean
function

23o
- - (,,. n.o.•.) (•0.•.•)

or its lineaxizedequiva/en•

0 0
.tlm.+x.) = (•. -&z.)ß(•.-&z.)+•2{(•ø.+z.)•n• (•0.•.3)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

wheref/is an undeterminedmultiplier, m. is the soughtjoint setof modelpaxameters


•ucl rn.o is our commonstazting
model.We havethat

oz _ -2AI•. + •AIA.x. + ••r,z ø.+ 2••z. = 0

or

A.••x. + f Hx. = A.U.


• - $•Hmø
.

from which we obtainthe solution

(•0.5.4)

We cana/soviewthe problem
froma naive(lessformal)but moreilluminatingangle
and state the lineaxizedinverseproblemaz the searchfor the smoothest
solution
amongstall possible
solutions
with [ y.- A.z.[2 <_qT(themaximrim tolerable
misfit),
azjudgedby thethequaAratic
measure [ x•-z•[•. We define
theLagr•ugeaa
hm'ctionin
component
form az (Meju, 1988)

wherep is an undetermlnedmultiplierandthe subscripts


1 aad2 indicatethe observa-
tionalstationcontributing
the data;thesedatahavebeenstandardiz•but the W term
isneglected
forclarity.Setting
thederivatives
ofœwithrespect
tothemodel
paxarneters
to zero,weobtainforthefirstlinearapproximation
(Meiu,1988),

0œ(•:•,•)
ß = • - x• + • -'(A;A, - A;•,) = 0
.•ving
and

O.l.(x• , xa)
,

- =o
giving
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

so that

and

:r.a
= (A•A:+ I,tl)-•(Aa•yz
+'l,t•).

Theaboveequations (10.5.6and10.5.7)mustbesolved simultaneously forthesolutions


andmaybeapphed successively torecover
thedesiredmodel. It isdearthattheabove
solutions
axecoupled via/l. If theconstr•i.nng
equationsareweighted moreheavily
thanthedataequations thenthedifferencesbetweenthevalues oftheparameters axe
minimizedat theexpense ofincreasingtheprediction
erroroftheotherequations. A
verysmallvalueof•twould producetheusualroughmodels. Theregulaxizationfactorg
isdeterminedby trialanderroranda valueisaccepted if thepermissiblemisfitq•ris
satisfied.
Thistechnique wassuggested forthemagnetotelluricstaticshit problem
(Meju,1988}
andmaybeinterpreted
asmoving
a resolving
kernel
across
thesurface
instead
of downthegeoelectric
section
asiscustomary
in geophysical
inversion
ofone-
dimensional
depth-soundings.
Withappropriate
constraints
forexample,
onemight
deducea smootherandlessdistorted
resistivitycross-section
for a setof depth
soundings
thanwouldbeobtained
using
conventional
interpretation
strategies.
Wewill
nowproceed
to proper
multi-dimensional
inversion
ofgeophysical
data.

10.6 Inversion of Potenti• Field Data


Wehaveshown
howpotential
fielddatamaybecorrected
fortemporal
andspatial
variations
duringroutine
processing
byapplication
ofinverse
Problem
theory.
We
shall
lookat another
aspect
ofdataauMysis
inwhichgeological
targets
ofinterest
a•e
delineated
using
fielddata.Thisprocess
willbereferred
to asdatainterpretation.

2•2
10.6.1 Interpretation of AeromagneticData
'Ar•p«dandcost-effective
w•y of exploring
vastlandaxeaz
is by airborne
m•gnetic
surveys. Aeromagneticsurveysfind •pplication•n mineral explorationin basement
areasandin petroleumexploration of sedimentarybasinsamongotheruses.Consider
the problemof interpreting•erom•gneticd•t• from • sediment• basinwith the
objectiveof determiningthe depth to the bazementand any structuralfeatures•h•t
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

m•y be important •n petroleumresourceevaluation.The indrabasement


anom•es often
targetedfor analysism•y be clueto severalindependent
•wo-dimensional
(2-D) or three-
dimensional/3-D) bodiesofarbitraryshapes someof whichcanbe•pproximated by
prismshaving arbitrary orientationsandm•gnetiz•tions.

10.6.1.12-D Interpretation of Intrab• Anomalies


Theobservable
response
dueto a simple2-Dbody(e.g.,a dyke-like
feature)maybe
simulated
asthatofa vertical
semi-in•inlte
2-Dprism
(see
Fig.10.õ•1).
Thus,
shallow
sourcesdue to faults and vertica•intrusionsin a sedimentarybasincould be modelled as
thin prismswhile the topography
of the basement
may be appro•rnatedby deep
sources
in theformofthickpris
ms. Considering
thevertical
prismmodelshown
in
Fig.10.6.1,
thereaxesixparameters
ofinterest
•ous,namely
thedepth
tothetopof
theprismzs, thehorizontal
positions
of[hetwoedges
oftheprism• andx•+x,the
intensityof magnetisation
Jj, theinclination
ofmagnetisation
I• andthe general
datum
•erm ATo. In typical explorationsitua[ions,there could be several2-D bodies in the
surveyedarea suchthat our field observations may have contributionsfrom more than
one prismsin our model for the basin. For this idealizedscenery,we note that the
observable
magnetic
fieldintensity
at a pointxi dueto thejta prismmaybe obtained
as (Leite & Leao, 1985)

andthetotalcontribution
duetok independent
bodies
fortheitsdat•m(i = 1,2, ... ,n)
issimpl
y

= zX.To

whereAo•= aoa) - 7o7•,Bo)= ao7•+ a57o,Ci• = :r• - zi, Di5= :rj+ • -


E•)= Di• + z•2 andFi• = C• + z2.
Thequantities a• and7• axethedirection
cosines
o{themagnetization
vectorofthe

233
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Fig.10.6.1Vertical
I
xI

J] -
t

I
!

I
Xi+l

prismmodelforthemagnetic
extendsovera longdistancein thestrikedirection,Y.
•>X

computations.
Theprism
body while ao and 70 are thoseof the geomagnetic
field vectorof magnitudeT a.udare
givenby
a• = cosI• cosD• (10.6.3)
and

sinI• (10.6.4)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

for j = 0,..,k, whereD is the declination.

C•lcudation of Partial Derivatives

The partial derivatives


requiredfor the inverseproblemof determining
the parameters
of the 2-D bodiesfrompracticaldataaxeobtainedby differentiating
eq. (10.6.2)with
respectto eachof the sixmodelparameters. Thus,asin LeiteandLeao,we havethat

OAT(•)
az.•=2J•{Ao.•(D,ff
S,.•-
Ci.•/Fi.i)+
Bo.•(z.i/.E,.i-
zffl;'i.i)
} (10.6.,.5)

SAT(::,)=
2J•{Ao•(z
ffF,j)- Bo•(C,•/F,•)} (10.6.6)

OAT(,)
= Bo(Du/Eu)
} (10.6.7)

c3AT(zi)
t•Js
= AT•(zi)
J•
(10.6.8)

azso=
and

OAT(z•).
5I• = 2J•{A•
•{tan-l(C•/z•)
- tan-l(D•g/z•)}
+0.SB•
In(E•/F•)}
(10.6.10)
where
A•S= (-aosinI scosDj) - 70cosI s andB•S= (aocosI• - 7osinI scosDS).

A typicalinversion•lgorithmwouldrequireasinput'the geomagnetic
field anda guess
modelassuminõ
th• •hefielddatahavealready
beensubjected
to thest•udard
processing
techniques
(e.g.,smoothing
andregional-residual
separation).
The solutionto
theinverse
problem
minimizeq(m)=I d-f(m) I • maybeobtained
usinganystandard
optimizationmethodbut the mostpopularschemes are the regularizedmatrix inversion
methodsand the conjugate-gradienttechniques.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

10.6.2 lnt•retation of Gravity Data


The interpretationof gravityanomaliesdueto 2-D and 3-D structuresof arbitrary
shapesis an inverseproblem that can be handledusingthe generalizedmatrix inversion
techniquesdiscussedpreviously.The key task here is the solutionof the forward
problem.If the forwardsolutionis knownfor bodiesof uniformdensitiesandirregulax
shapes,then we caneazilyretrievethe subsurface
densitydistribution(or the parame-
tersof the2-D/3-Dbodies)
thatwillexplain
a givensetofobservations.
In the
applicationof inversetheory, we need to computethe paxtiM derivativesof the models
and have to incorporate someform of constraintson the solution processto avoid
instabilities.The gravity problemis generaally
non-•_•nique,
a.udit is thus desirableto
incorporateany available(and reliable)a priori informationin the problem
formulation. Sincethe forwardproblemholdsthe key to a successful interpretation,it is
appropriatethat we start by considering2-D forwardmodelsof the subsurface.The
well-knownPedersen(1977) model and someapplicationsof the hugely popular
Talwani method (Talw•= et al., 1959) will be•tiscussed.

10.6.2.1 Interpretation of lntrabazinM Anomalies• Pedcrsen's2-D Model


Consider the problem of sedimentary basin evaluation for hydrocarbonresources.
An important interpretationproblemhereis the determinationof a single2-D
interface between the sediments and the basement as this info•ation can be used

to delineate axeasworthy of further detailedexploration.For sucha task, we ca•


adopt Pederson'smathematicalmodel.

Let p• and p• be the densitiesof the sedimentsandbasementrespectivelyand consider


the observedgravity field alonga profileto be the sumof the signaturescontributed
by a stack of prismsor overlappingdyke-likebodieswhich extendinfi•itely in the
strike directionand whoseuppersurfacedefinethe sediment-basement bounda• (see
Fig. 10.6.2). The densitycontrastbetweenthe basementand the sedimentsis
At>= p•-p•. In many practicalsituations,the gravity data are processed to yield the

236
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Fig.
10.6.2
overlying
sediments
!

!
,

:
x]

Contiguous
ß '•

ofdensity
•11'
prisms

237
X]+l

.
+1

representing
i
i
Xp

basement
i lie
x

ofdensity
•2 with
residualanoma/iesimplyingthat the regionalgravityeffectassodated
with some
referencedepthzo hasbeenremoved.In thissituation,for prismaticbodieswith the
jta prismhavingedges
zi andzj +1 (Fig.10.6.2),
theobservable
gravityresponse
someprofile positionzi is •ven by
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Ag(zi)
: -GAff__
•C•i
InC•i
ji +
C2 + z}
Z•_Di
iInD•i
•+
D2 + z}
z• + 2zj(ian-•
zj - tan
zi 2

- :0- - +
+26•zo
tan-x
Dii
zo- 2•jzj tan-• •} (10
ß6.11)

whereC'i•= (zi+• - z•)andDi• = (zd- z•),G istheuniversal


gravitational
constant,
and•i isequalto-1 if j=l andequalto1if j=m (notethatinthelastsnmmation
j is
either 1 or m).
If Pt,P2andz0 arespecified
andit isintended
to findthe 2-Dtopography,
i.e.,theset
ofvalues
zi thatwould
explain
thefieldobservations,
thenthepartialderivatives
ofthe
modeldatawithrespect
to theparameters
zi aregivenby (Pedersen,
1977)

, j = 1,2,...,m-1

'
, j = 1,m (10.õ.12)

wherei=l,2,...,n aretheindices
of theobservational
stations.Fora moregeneral
2-D
interpretation,weshallexaminetheTalwa,•imodelnext.

10.6.2.2Mod•lll,,gof2-Dbodie•of arbitrary
shapes
usingTalwani's
method
A 2-D subsurfacefeatureofirregular
shape maybeconceptually
viewedasa polygon
havinga similar
shape
andphysicalproperty
- a kindofscaled
down
modelofthered
body.Thegravitational
attraction
ofsuch
ann-sidedpolygon
canbecalculated
and
constitutes
theforward
problem.Consider
the2-Dbodyshown
in Fig.10.6.3.
If we

238
, o

•Z1)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Fig. 10.6.3Cross-section
ofa ploygon
representing
a 2-Dsubsurface
body.

assume
thatit extendsto greatdistances
in thestrikedirection
(whichis mutually
perpendicular
to thex- andz-directions),
thentheattraction at anobservational
position,
O alongthex-axis
isg/yenby(seealsoGraut& West,1967)
nsides(

where• = (z, + 1-z,)/.(z•+•-z,), b•= (z,z,+i-z, +iz•)/(z•+•-z,), C isthe


•versM •a•iationM const•t, •d • is thede•ity controtof the•ody•th the
su•o••ng me••.
Thep•me•ers of themodel•e thedensitycontrot•th thes•o•&ng me&,,m•p
•d the(ho•zont••d vertic•)coordinteso•then-sided2-Dbody.Thecomputation
o• thep•i• de•vatives
o•themodelisstr•ght•o••d •d • 'thec•e o•thedensity
p••eter we simplyh•ve that

10.6.2.3Implementing
a GravityInver•onAlgorithm
The'ideahereis to pointoutthatapartfromtheforwardproblem,
themai• difference
betweenone-dimensional
inversiona•d multi-dimensional
inversionis in the dimensions
of the vectorsand matricesinvolved.An optimizationor minmizationroutineusedfor
resistivity
inversion
say,caneasilybeadapted
to handlethegravityormagnetic
problem.
Forillustration,
theridgeregression
routine
usedin the1-DWenner
resistivity

239
inversion(Sec. 10.3.3.4)is adaptedto a 2-D gravityinversionproblemusing a variant of
the popularTaiwan5forwardalgorithm.The program(GRAVINV) is listed in
AppendixC andis self-explanatorybutw• willexplainoneimportant practicalaspect
of gravitydata analysis.
A common, but by nomeansmandatory, pre-interpretation
practiceis to separate
outregionaleffectsfromthe observed
dat•. The resultingdata
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

set is referredto as the residualgravity anomaly.It is thereforea goodpracticeto build


this aspectof data processing
into an inversionprogram.Thereare severalwaysof
estimatingregionalgradients but we will restrictourselves
to simpletrend estimation
usingthe methodofleastsquares. A simpleimplementation of theclassicalleast
squares schemeis illustrated
in the sub-program SIMPREGlistedbelow.

subroutineSIMPREG(/t,x,y,y2,ey2)
c a simpleroutineforestimating
theregional
trendforpogential
field data
c usinga linearregression
technique.
estimates
regional
gradient
c and the associatederrorsbasedon data sca•er, aerr and bert.
c input: x-position;y=potentialfieldda•a;n--numberof data.
c output: y2-- regionaltrend; ey2--estimation
errorin y2.

dimension
x(100),y(100)
,xx(100),x'y
(100),ey2
(100),y2(100)
c initi•llzations
sumx=O.O
sumxx--O.O

sumxy=O.O
sumy=O.O
calculaterequiredquantitiesxx,xy and sums
do 10 i=l, n
xx(i)=x(i),x(i)
xy(i)=x(i),y(i)
sumx=sumx+x(i)
sumxx=s,,,•xx+xx(i)
sumxy=su••+•(i)
sumy=sumy+y(i)
10 continue
calculatethe commondenominatorand the desiredregression
parametera.
den=n,s-rnxx-(sumx*s,,mx)
b=(n,sumxy-(
sumx, sumy))
/ den

240
a----(
snmy* su•,cx-$uxr•x,
s,,mxy)/ den
cMcul•te root mean squareerrorska •,b.
sse=O.O

do 20 i=l,n
e=y(i)-a-(b*x(i))
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

sse=sse+(e*e)
20 continue

rmsq=sse/(n-2)
aerr=rmsq*snmxx/den
berr=n*rmsq/den
do 30 i=l,n
y2(i)=a+b*x(i)
hnin=(a-aerr)+ (b-berr),x(i)
fmax=(a+ aerr)+ (b+berr),x(i)
ey2(i)=( abs(f'min)
+abs({max))/2.
30 continue
retur•

end

Thissubroutine in usedin thegravityinversion


programlistedin •he appendix
where
the residualanomalyis simplytakenasthe difference
between the observed
gravity
a.uomaly
andthecalculatedregional
trend.Thecomputer program
allows
forthe
interpretation
ofeithertheoriginal
fielddataortheresidual
anomaly.

10.6.2.4Asp•s of GravityInverseProblem
Thegeneral
inverse
problemisnonlinea•
bu•wecanreduce
thenon-linear/•y
by
choosing
anappropriate
parameter/zation.
Consider
•heinverse
problem
inwhichthe
dimensions
of a bodyarespecified
andwearerequired
•o determine
i•sdensity
from
somegivengravity
observations.
Asdemonstrated
bythe•est$•mamaxi$•
ill Figure
10.6.4a,
as•hebulkdensity
ofa bodyoffLxed
geomegrical
rt•mensions
isvaried, the
gravity
response
atanobserva[ional
point,
0 situated
ata distance
dfromthebody
showsa 1/near
relationship.
Theslope
ofthislinearchange
in gravityresponse
depends
ontheseparation,
d. Ontheother
hand,if thedensity
ofa bodyisspecified
(e.g.,from
intersecting
borehole
measurements)
andwearerequired
•odetermine
i•sshape
from
theohserred
•avity data,thenwearedealing
wi•ha different
kindofinverse
problem.
(a)
regal
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

.. Kg/m3
Variable density
fixed dimension

(b)
mGai

metres
Fixed density
variable dimension

Fig.10.6.4Gravity
response at a point0 as(a)•Oisvadedwithblocksizeheld
constantand(b)• isfixedasblock sizeisvaded

2•2
As illustratedfor a simplebodyof fixed densityin Fig. 10.6.4b,the responseat point O
as the physicaldimensions of the test bodyis variedby extendingor shorteningi%sbase
showsa non-linearrelationship.The •avity effectof varyingboth the shapeand
densi%y to fit a givensetof e.-vperimental
datawill obviouslyshowa non-linear
behaviourand is left to the imaginationof the reader.In any case,it canbe gleaned
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

from thesesimpleteststhat we couldin principlereducethe inherentnon-l•uearityin


the interpretationof practicaldata by usinga priori infomarionderivedfrom say,
geological
dataor othergeophysical
results
fromthe area.Let usnowdemonstrate
the
aboveconceptsusingtypicalcrustalex-ploration
data.

In Fig. 10.6.5is shown


a 2-D modelformagnetotelluric
(MT) andgravitydataacross
the CaJedonide$
of Scotlaud
(Meju, 1988).The geometrica•
shapes
in the model
were determinedby separate2-D modellingstud/esof MT data across.
the re/ion. The
resulting
modelwasusedinitiallyto frxthegeometry of thegravityinterpretation
problemwhichreduced it to the searchforthe optimMdensitydistribution
that will
explainthe gravityanomalies. At somelocati6.ns
not constrained
by the MT data,the
shapesof the bodieswereinferredfromgeological considerations.Thisis the simplest
approachto gravityinversion. Noticethe goodfit of the integratedmodelresponse to
the data. The same•avity data wereinvertedfor the subsurface densitydistribution,
this time allowingboththe shapesand densitycontrastsof the bodiesto be varied.The
resultsare shownin Fig. 10.6.6.This operationrequireda powerfultwo-stage
rni••zation by ridgeregression asdescribedin Chapter7 (seealsoMeju, 1992).Note
the goodfit betweenthemodelresponse andthefielddataa•d thestructural
concordance with the modelshownin Fig. 10.6:5.The agreementbetweenthe model
shownin Fig. 10.6.5andhighresolution seismic reflectionimagesof the deep
crustanduppermantlein the reg/onwasdemonstrated in Meju (1988)andunderlines
the needfor an integratedapproachto geophysical data analysis.

10.6.3 Joint Inversionof Magneticand Gravity Data


Magnetic
andgravityanomaJies
mayhavethesamesource
(or at leastthesignatures
maybeclueto structures
whichhavesomecommongeometric
elements) anduseful
information can be obtxi•_edwhen both data sets•re simultaneouslyinverted for the
shapeandpositionof thecausative bodiescharacterised by constant
uuknown
density
andsusceptibility
contrasts.
Thisis a nonlinear inverseproblemandmaybe formulated
asin ourprevioustreatmentof joint inversion
of resistivitydata.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

• 0

• -12-
m=-18-
*--' -24-

• -30-
r,D o
6 +.025
".02
1
+.025 +.025
18
8k >80k
E 24
• 30

2- 30k

Density contrast: g/cc


Resistivity:ohm-m

Fig.10.6.5 An integrated
geophysicalmodelfortheCaledonidesof northern
Scotland(Meju,1988 ). (a) Bouguergravityanomaly;
(b) MT/gravity
model.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

• o .-"talc'
ß

-30
' ! I I I I I I
4'(j 60 80 100 120 1'40 160 180 200 ;•Z;?..O
Km
MTZ GGF HBF
2.71

E 1 2.82
2.78 2.8 2.8

2.85 2.85 2.85


3.3

Rg. 10.6.6 A 213regionalgravitymodelfortheCaledonides


of Northern
Scotland

2•5
l.f the observedmagneticandgravity data ared,. anddo respectively,
they axerelated
to •he np modelpaxmeters m character/sing the anomalous subsurface bodiesby

d• = f.•(m) + e•
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

and

da - fa(m) + %

wherethe nonl•neax
forwardfunctiona•f• andf• havebeendescribed
in the preceding
sections
ande• andeaaxethevectors
of additive
noisein themeasurements.
Thejoint
da•a set is simply

andis of dimension(nrn +ng) x 1 wherethereaxenm magneticobservations


andng
gravitydata.Similarly,thejoint prediction
modelmaybe expressed
as

f(m)
leading to the relation

d = f(m) + e.

Lineaxizing
aboutaninitialmodelmøleadsto thefamiliarproblem

y=Ax+e

wherex is the soughtpaxmeter correctionvector,

y= and A= .

Here,It,.- a,,-f,,(mø),v, = d•-f,(mø),A•= 0f,•(mø)/Om


øandA,= Of•(mø)/Om
ø.
Notethat thematrixof partialderivatives,
A is of dimension
(nm+ ng)xnp and•he
colum•vectorIt is of dimension
(nm+ ng)x1. The systemofequations
maybesolved
usinganyof•hes•andardprocedures.
I• shou/d
bepointed out•hatsince•hegravity
andmagneticda•aarein different
-ni•s,a useful
s•abilizing
s•rategy
in jointinversion
is

2q6
to employsomeformof scaling
for y,, •t,, A• amdA• beforetheyareassembled
in •he
jointmatrixA andjointvector•t.Aftersolving theleastsquares
normalequations,
the
resultingsolutionmust be appropr/atelyre-scaledto yield the correctionvector x. A
constrainedinversion formulais recommended
forsolving
theinverseproblema.udany
availablea priori information
maybescaledandappended ontothesystemof equations
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

for better model identification.

Acknowledgements:Someof •he d•ta usedin the exercisesin this coursenote were


takenfrom publisheda•d unpublished
accounts
andthe authoris eternMlygratefulto
the vaxious sources.

247
tLEFER.EN CES

Anderson. W.L., 1979. Numericalintegration of related Hankel traax•oms of orders 0


and 1 by adaptivedigi%al
filtering.Geophysics,
44, 1287-1305.
Backus,G.E. & Gilbert,J.F., 1967.Nnmer/calapplicationof a formalismfor
geophysicalinverseproblems. Geophys. J.R. Astr. Soc.,13,247-276.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Backus,G.E. & Gilbert,J.F., 1968.The resolvingpowerof grossearthdata.


Geophys.J. R. Astr. Soc.,16, 169-205.
Backus,G.E. & Gilbert,J.F., 1970.Uniqueness
in the inversionof inaccurategross
earth da•a. Phil. Trans. 1{. Soc. A, 266, 123-192.
Biwen,X.. and Barker1{.,1993.Filter coeffiden•s for offsetWennerandWeuuer
resistivkysounding interpretation.Computers &:Geosciences (in press).
Cloutier,J.I:L, 1983.A technique for reducinglow-frequency, time-dependent errors
presentin network-•ype surveys.J. Geophys. Res.,88, 659-663.
Constable.S.C., Parker, R..L.,i: Constable,C.G.,1987. Occam'sinversion:A practical
algorithmfor generating smooth models fromelectromagnetic soundingdata,
Geophysics,52, 289-300. -
Constantin/des,
A.. 1987.AppliedNumericalMethodswith personalcomputers.
McGraw-Hill Book Co.

Cooke,D.A., and Schneider,


W.A., 1983.Generalized
linearinversion
of reflection
seismicdata. Geophysics,48, 665-676.
Duijndam,A.J.W., 1988.Bayesian
estimation
in seismic
inversion.
Part I ßPrinciples,
GeophysicalProspecting,36, 878-898.
Duijnda.m,
A.J.W., 1988.Bayesian
estimation
in seismic
inversion.
Part II: Uncertainty
Analysis,GeophysicalProspecting,
36, 899-918.
Eckart,C. & Young,G.,1939.
A pr:mdpal
axistransformation
fornon-Hermitian
matrices, Amer. Math. Soc.Bull., 45, 118-121.
Gaver,D.P., 1966.Observing
stochastic
processes
andapproximate
trausform
inversion.
Oper. Res., 14, •.•!-459.
Ghosh,D.P, 1971.The application
oflinearfiltertheoryto the.direct
interpretation
of
geoelectrical.
resistivity
sounaiug
mea.suremen•s.
Geophys.
Prosp.,
19,192-217.
Gol'stm•u,F.M., 1976.Thecombination of observations
in the identification
of
geophysical objects,
Izv.,EarthPhysics,no.7, 40-54.
Gol'tsman, F.M., 1975.S•atistical
theoryfor•heinterpretation of geophysical
field,
Izv., E•rth Physics,
no.1,29-53.,translated
by M.N.Pillai.
Gol'tsm•u.
F.M., 1971.Statistical
models
of.interpretation.
Science
(Nauka)Press,
mosco•,v.
Grant,F.S., andWest,G.F., 1967.Interprecation
theoryin appliedgeophysics.
McGraw Hill. New York.
Grivelet,P.A., 1985.Inversion
of verticalseismic
profilesby iterativemodeling.
Geophysics,
50: 924-930.
Hattin•h,M., 1988.A newdataadaptive
filterin•progrzmto remove
noisefrom
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

geophysical time-orspace-series
data.Computers L:Geosciences,
14,467-480.
Hatton,L., Worthington,
M.H. L:Makin,J., 1986.Seismic
DataProcessing:
Theory
and Practice. Blackwell Scientific Publications,pp.139-163.
Hoerl,,A.E.L: Kennard,
R.W.,1970.Ridgeregression: biased
estimation
for
nonorthogonal problems,
Technometrics,12,55-67.
Jackson,D.D., 1979.Theuseof a prioridatato resolve
non-uniqueness
in l•near
inversion,Geophys.J. R. astr.Soc.,57,137-157.
Jackson,
D. D, 1973.Marginal
solutions
to quasi-linear
inverse
problem.•
in geophysics'
theedgehog
method.
Geophys.
J.Roy.
Astr.
Soc.,
35,121-1'36.
Jackson,
D.D.,1972.
Interpretation
ofinaccurate,
ins•cient, andinconsistent
data,
Geophys.J. R. astr.Soc.,28, 97-109.
Jackson,
D.D. & Matsu'ura,
M., 1985.A Bayesian
approach
gononlinear
inversion,
J. Geophys.Res., 90, 581-591.
Jobarisen,
H.K.,1975.Aninteractice
computer/graphic-disPlay-ter
minalsystem
for
interpretation
of resistitivity
soundings.
Geephys.
Prosp.,23,449-458.
Jones,A.G. & Hutton,R.,1979.A multi-station
magnetotelluric
studyin southern
ScotlandII. Monte-Carloinversionof the data and its geophysical
and tectonic
ßimplication,
Geophys.
J. R..astr.Soc.,56,351-358.
Knight,J. H. andtLa•che,
A. P., 1982.
Tran.•ient
electromagnetic
calculations
using
the Gaver-SgeMest
inverseLaplacetr•n•orm method.Geophysics,
47, 47-50.
Koefoed,
0., 1970.A fastmethod
fordetermining
thelayerdistribution
fromtheraised
kernelfunction. Geophys.Prosp., 18, 564-570.
Lainbeck,
K.,1980.TheEarth'svaxiable
rotation:
geophysica•
causes
andconsequencies.
CambridgeUniversityPress,449p.
Lanczos,
C., 1961.Linear
Differential
Operators,
Chap.3, VanNosfraud,
London.
Lavergne,
M., andWi•m,C.,1976.Inversion
ofseismogram.•
andpseudo
velocity
logs.
Geophys.Prosp.,25,231-250.
Leige,
L.W.B.,andLeao,J.W.D.,1985.Pddge
regression
applied
go•heinversion
of
two-dimensional
aeromagnetic
anoma•es.
Geophysics,
50, 1294-I306.
Levenberg,
K.,1944.
A method
forthesolution
ofcertain
nonlinear
problem•
in lea.st
squares:
Quart.Appl.Math., 2, 164-168.

2•9
Marquardt, D.W..1970. Generalizedinverses,ridge regression,biasedlinear estimation
and nonlSne• estimation. Technometr/cs, 12, 591-612.
Matsu'ura, M. & Hirata, N., 1982. GenerMizedleast-squares
solutionsto quasi-linear
inverse problem with a priori information, J. Phys. Earth, 30, 451-468.
McAulay, A.D.,1985. Prestackinversionwith plane-layerpoint-source
modelling.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Geophysics,50, 77-89.
Meju, M.A., 1994f. Simpleeffectiveresistivity-depthtransformationsfor in-field or
real-time data processing.
First Break (in review).
Meju, M.A.,1994e. Jointinversionof TEM and MT-phasedata: A simpleeffective
remedy for staticshiftin MT sound•gs.Geophysics (in review).
Meju, M.A.,1994d. Biasedestimation:a simpleframeworkfor inversion•ud uncertainty
analysiswith priorinformation.Geophys. J. Int., (in press).
Meju, M.A., 1994c.A generalprogramfor linear parameterestimationand uncertainty
ß

analysis.
Computers
•zGeOsciences,
20,197-220.
Meju, M.A., 1994b.Joint TEM and AMT surveys'cost-effective
exploration in frontier
basinregions.
Proceedings
of 10thPetr•le•mCongress
andExhibitionof Turkey,
Ankara; Geophysicsvob•me, 157-176.
Meju, M.A.,1994a. Assessingthe role of infield resistivityimageprocessingin shallow
subsurfaceinvestigations.Proceedingsof the 'Symposil•mon the Application of
Geophysicsto Engineer/ngand EnvironmentalProblems,Boston;19-40.
Meju, M.A.,1992. An effectiveridge regression procedurefor resistivitydata inversion,
Computers & Geoscience, 18, 99-118.
Meju, M.A.,1988. The deepelectricalstructureof the Great GlenFault, Scotland.
PhD thesis,Un/versityof Edinburgh(Jan. 1988)
Meju, M.A., Fontes,S.L. andOliveira,M.F.B., 1993.JointTEM/AMT feasibility
studiesin P aznaibaBasin,Brazil: geoelectrostratigraphy
and groundwaterresource
evaluation in Piau/State. 3rd Iutemational Congressof the Brazilian Geophysical
Society, Rio de Janeiro.ExpandedAbstracts,1373-1378.
Meju, M.A. and Hutton,V.1•.S.,1992.Iterative most-squares
inversionßapplicationto
magnetotelluricdata,Geophys.J. Int. 108,758-766.
MenIre,1984.Introductionto Geophysical Data Analysis'Discreteinversetheory,
Academic PressInc., Orlando,Florida.
Meyer,S.L.,1976.Data Analysis for Scientistsand Engineers.
JohnWHey,New York.
Oldenburg,D.W., Scheuer, T., andLevy,S., 1983.R.ecoveryof the acousticimpedance
from reflectionseismograms. Geophysics, 48, 1318-1337.
Parker, R..L.,1982.The existenceof a regioninaccessible
to magnetotellur/csounding,

250
Geophys. J. It. Astr. Soc., 68,165-170.
Parasnis,D.S.,1986. Principlesof appliedgeophysics.ChapmaumudHall.
Peclersen,L.B., 1977. Luterpret•gion
of pogentia.l
field clara,• generalizedinverse
•pproach. Geophys.Prosp.,25, 199-230.
Peclersen,J. a•d Herm•uce: J. F., 1986. Least-squaresinversionof one-dimensional
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

m•gnetotelluricd•t•: •n assessmentof procedures


employedby BrownUniversity.
Surveysin Geophysics, 8(2), 187-231.
Pous,J., Marcuello, A. •z Queralg,P.,1987.Resistivity inversionwith a priori
information, Geophys.Prosp., 35, 590-603.
Press,W.H., Fl•.nnery,B.P., Teukolsky,S.A., and Vetterling,W.T., 1986.Numerical
recipes:the art of scientificcomputing.C•rnbridgeUniversityPress,Cambridge.
818p.
P•che, A. P., 1984.The effectof rampfamctionturn-offon the TEM response
of
layered earth. Expl. Geophys.,15, 37-41.
R•che, A. P., Jupp, D.L.B, Rutter H. mudVozoff, K., 1985.The joint useof coinci-
dent looptrmusientelectromagnetic
and SchI,,mberger
soun3;ug
to resolvelayered
structures. Geophysics,50, 1618-1627.
P•ay,R.D., 1985. Correctionof systematicerror in magneticsurveys-An application
of.ridgeregression andsparsematrixtheory.Geophysics, 50, 1721-1731.
Seeman,B., and Horowicz,L., 1983.Vertical.seismicprofil•ug:separationof upgoing
and downgoing acoustic
wavesin a stratifiedme(]i,,m.Geophysics,48, 555-568.
Spies,B. R. and.Eggers,D.E., 1986.The useand m•suse-of. apparent'
resistivityin
electromagneticmethods.Geophysics, 51, 1462-1471.
S•elffest,H., 1970b.Remarkon Algorithm368, N,,•erical inversion of Laplace
tramsforms.Comm. Assn.Comp. Mach., 13, 624.
Stehfest,H., 1970a.Algorithm368,N,,mericalinversionof Laplacetrmusform.•.
Comm. Assm Comp. Mach., 13, 47-49.
S%roud,K.A., 1986.FurtherEngineering Mathematics:Progrxmmes mudProblems.
Macrn•llanEducationL•d. (reprinted1987).
Talw•l•, M., Worzel,J.L., andLandisman, M., 1959.l•pid gravitycompuSalions
for
two-d•men$ionalbodieswith application•o •he Mendocinosubmarinefracturezone.
J. Geophys.Res., 64, 49-59.
Taxme.
r, M.T., Koehler,F., andA]h;lali,K.A., 1974.Estimalionmudcorrection
of near-
surfacetime anomalies,Geophysics,39, 441-463.
Tarmutola,A. •z Valette, B., 1982.Generated nonlbae•rinverseproblems
solvedusing
the leastsquarescriterion,Reviewof Geophysics and SpacePhysics,20, 219-232.

251
Tikhonov: A.N., 1963. Regularizationof ill-posedproblems,Doklady Akad. Nauk
SSSt{, 153, 1-6.
Tikhonov.A.N. & Arsenin,V.¾., 1977.Solutionof ill-posedproblems,JohnWiley and
Sons. Inc.

Tikhonov, A.N. & Glazko, V.B., 1965. Applicationof • regulariz•tion method to


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

nonlinear problems,J. Comp. Math. and Math. Physics,5, no. 3, 1965.


Tikhonov, A.N. & Glasko,V.B., 1975. Applicationof the regula•zation method to
geophysical
interpretationproblems,Izv., Earth Physics,no.1,38-48,tra•uslated
by
M.N. Pillar.

Twomey, 1977.An Introductionto the Mathematics


of Inversionin l•emoteSensing
and In•rect Measurements,ElsevierScientificPubl. Co., Amsterdam.
Vozoff, K., i: Jupp,D.L.B.,1975.Jointinversionof geopliysic•d
data, Geophys.J. R.
astr. Soc.,42, 977-991.
Wa.ug,R.J., 1977.Adaptivepre4ictivefieconvolution
of seismic
data.Geophys.
Prosp.,
25, 342-381.
Wa.ug,R.J., andTrietel,$., 1971.Adapti.ve
signalprocessing
throughstoch•tic
approx•ruation.Geophys.Prosp., 19, 718-728
WicLrow,B., a•d l:ioff,M., J•., 1960.Ad•ptive switchingcircuits.IR• WESCON Cony.
Rec., pt. 4, 96-104.
Wickow, B., Glover,J.R..J•., McCool,J.M., Ka,,n•z, J., Williams, C.S., He•, R..•.,
Zeicller,J.l•., Dong,E. Jr., and Goodlin1•., 1975.Adaptivenoisecancelling:principles
and applications.Proc. IEEE, 63, .1692-1716.
Wid.row,B., Mantey, P., Gri•tks, L., and Goode,B., 1967.Adative a.ute•a sysiem.•.
Proc. IEEE, 55, 2143-2159.
Widrow, B., McCool,J.M., Larimore,M.G., and Johnson, C.R..,Jr., 1976.Stationary
and nonstationarycharacteristic•of the LMS adaptive•ter. Proc. IEEE, 64, 1151-
1162.

Wiggins,R..,Larner,K., a.udWisecup,
R.D., 1976.Residual
statics•uMysis• a general
linear inversep•oblem. Geophysics,41, 922-938.
Word, D.H., Smith,H.W., a•d Bostick,F. X., Jr., 1970.An investigation
of themag•e-
totellurictenso•imped•ucemethod.Tech. Rep.No. 82, Electr. Geophys.R•. L•b.,
University of Texas, Austin. 264p.
Y•ger, H.L.,Robertson,
R.I•., •ud Wenritual,
R.L.,1978.Diurnal• removal
f•om
aeromagnetic datausinglea.stsquares.
Geophysics,
46, 1148-1156.
Z•i-tia•, Ma. 1988,Problemsof •hree-•mensionM
seismicdataprocessing.
Chinese
J. Geophys.,31,147-156.

252
APPENDIX A

SVI)•: A computerprogramfor linear inversionand detailederror analysis


programSMDINV
c a detailed illustration of generalisedmatrix inversion
c usingthe singularvaluedecomposition (svd) method.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

c solvesthe linear inverseproblemoœthe form d--gin


c and performsuncertaintya.ualyses (modelresolution,
c parameter covariancesand lea•t squaresstaad•d
c deviations,most-squaressolutionenvelopesand extreme
c paxmeter sets,i.e., modelbounds).
c adaptedfrom M.A.Meju, 1994a, Comp. & Geosc.,20,197-220

real g(100,40),u(100,40),ut(40,100),v(40,40),q(40),d(100)
+ ,vqi(40,40),utd(40),gm(100),vt(40,40),r(40,40),cov(40,40)
+ ,stclev(40),qi(40),x(100),m(40)
character, 16 in•,outfil, au•, 1
c describeprogram capabilities
write(*,,)
write(*,,)' This programca.ut•ckle the followinglinear problem.•'
write(,,,)' (1) simplestraight-linefitting for intercept a.ud'
write(*,,)' slope(i.e.,two-parameter probl_-_•ms);
Ix,y]
write(,,*)' pairsrequiredas input in this situation'
write(,,,)' (2) multiple-parameter rme• inverseproblems;
write(*,,)' vector d and designm•trix G requiredas input'
c determine •orm of d•t• and select •ppropriate input style
[¾:N]'

if(axm.eq.'Y'.or.ans.eq.'y')
the=
c read Ix,y] data pairsaadform G-matrix
call getdat(ndat,x,d)
ncol=2

do i=l,ndat
•(i,•)=•.
g(i,2)=x(i)
end do

else
c read data and componentsof G-matrix from a diskfile
write(*,'(a,$)') ' enterndat,nparm:> '
read(* ,. ) ndat,ncol
write(* ,'(a,$)')' enterinput filenamefor d & G '> '
•e•a(•,'(•)')i•!
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

open(tmi•-- 1,file= infi•,status='old')


write(*,'(/a)') ' readingdata vectord '
do i=l,ndat
•a(•,.) a(i)
end do

write(*,'(/a)')' reactingG matrix, row by row'


do i= 1,ndat
read(1,*)( g(i,j),j= 1,ncol)
end do

close(1)
endif

c open result file


write(.,'(//a,$)')' enter output fAenamefor'results-> '
read(*,'(a)')outftl
open(unit=3,file=outfil,status='new')
c select mode of operation
write(*,'(/a,$)')' usedampingtechnique? [Y:•I '
read(.,'(al)')•ns
if(azm.eq.'Y'.or.•.•.'y') th•
••(.,'(/•/)')' •l• d••m• •:• •o• •o•l•m ••l••io•'
••(*,*)' to• •moo•h• •-•=•(• •ol•io•), •• It]'
write(.,.)' for biasestimationproper(D=I), enter[21'
write(*,*)' for Mazquaxdt-Levenberg damping, enter[3]'
write(.,'(/a,$)')' enter1, 2 or 3 '> '
reaa(*,'(I•)')iopt
write(*,'(•,$)')' enterdampingfactor(beta) '>'
read(*,*)beta
constrainingequationsfor the designmatrix G. Note: if last parameteris
c well determined, it may be allowed to float with the othersconstrained.
if(iopt.eq.l.or.iopt.eq.2)then
,•4t•(.,'(•,$)')' co•t•i•: ALL [11,aii but last [21? [1:•.1>'
read(* ,:(I1)')minus
if(iop•.eq.2)then
minus=minus-1

endif

nrow=ndat +ncol-minus
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

k=O
do i=nda•+ 1.nrow
k=k+l

d(i)=O.O
do j-l,ncol
g(i,j)=O
end do

if(iop•.eq:2)then
g(i,k)=beta, l.O
elseif(iopt.eq.1)•hen
g(i,k)=beta. 1.0
g(i,k+l)=-beta, l.O
endif
end do

elseif(iopt.eq.3)then
•row=ndat
endif

else

c solution damping not required


beta=l.

nrow=ndat

endif

call svd(nrow,ncol,g,u,¾,
q)
do i=l,ncol
calculate u-transpose
do j=l,nrow
ut(i,j)=u(j,i)
end do

calculate v divided by lambda


do k=l.ncol

if(iopt.eq.3)then

255
c effect MarquardbLevenberg-typedampingof •he singularvalues
qi(i) =q(i) / (q(i)+beta)•* 2.
vqi(k,i)= v(k,i).qi (i)
else

vqi(k,i) =v(k,i) / q(i)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

endif

end do
end do

calculate the inner product ut*d


call inprod(40,100,ncol,nrow,ut,d,u•d)
calculate the regressionestimatesm
cM1inprod(40,40,ncol,ncol,vqi,utd,m)
write(3,'(/a)') ' herearethe leastsqu•resestimates'
write(3,.) (m(i),i=l,ncol)
write(3,'(/a)')
call inprod(lOO,40•nrow,ncol,g,m,gm)'
calculated•ta misfit,sumd=!d-Gm[**2& total misfit,s-mt=sl,md+[Dm-h[**2
s•,rnt=O.O

do i= 1 ,arow
sumt= (d(i)-gm(i)),,2+s,,mt
if(i.le.ndat)then
m, md=sumt
endif
end do

write(a,,)' datami•t d+constr.misfit dampingf•ctor'


write(a,,)Sl,.md,s•mt,beta
calculate model resolution matrix, r
c fnmt,form v-transpose;
then get matrix-productv.vt
do i=l,ncol
do j = 1,ncol
if(iopt.eq.a)then
v(j ,i)=v(j ,i), q(i), qi(i)
endif

vt(i,j)=v(j,i)
end do
end do

256
ca•lrexprod (40,40,ncol,ncol,v,•,r)
write(3,'(/a)'): resolution
information'
do i=l,ncol
write(3,. )(r(i,j):j--1,ncol)
end do
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

calculatepaxasneter
covaxiance
matrix, coy
••(.,'(/•,$)')' aoso. •o• • ••.• o•a• ½=o=?[¾:N]'
re&d(.,'(al)')ans
if(ans.eq.'Y'.or.a=s.eq.'y')then
•t•(..'(•.$)')' •t• t• ••• ->'
read(.,.)sigma2
else

if(ndat.gt.ncol)theu
sigm•=s•md/(ndat-ncol)
else

sigm•=s•d
endff
endJ•

do i=l,ncol
do j=l,ncol
if(iopt.eq.3)•hen
cov(j,i)= sigma2.r(j,i)/ (q(i)+beta)/ (q(i)+beta)
else

cov(j,i)=sigma2.r(j,i)/q(i)./ q(i)
endif
end do

calculate
s•andard
deviations
(sqrtof diagonal
elements
of cov-ma•x4x)
stdev(i)=sqrt(cov(i,i))
end do

••(a,'(/•) ')' co••• •o••io•'


do i=1,ncol
write(3,.)(cov(i,j)•j--•,ncol)
end do

write(3,'(/a)')'
st•uaa•d
cle•i•tion•
ofestim•te•'
wdte(3,.,)(stdev(i),i-l,ncol)
cMculateextremep•ra•netersetsor solutionenvelopes

257
callmostsq(iopt,nrow,ncol,g,d,m,sl,mt,v,vt,q,
qi)
close(3)
stop
end
ß

C--'
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

subroutine
GETDAT(ndat,x,y)
ß

dimensionx( 100),y(100)
character.20 inf•e,outfile
nerr•0

write(.,'(//a)')
write(.,.)' [x,y]datap•s forline-fitting
to bereadnext!'
write(*,'(//a,$)')'a•ethedatastored
asa aisk•le? [Y:N]
•a(.,'(•)')a•
if(ans.eq.'y'.or.ans.eq.'Y' )then
goto 96
97 write(*,95)in•le
nerr----nerr+l

95 format(/' !!!errorin opening


di$1cfile:
',a/0,' try again
if(nerr.gt.5)return
96 write(.,'(/a,$)')'enternameofdatafile(max.20ch•s)
read(ß,' (a2O)')i-•ile
open(unit
=3,err=97,file=inTile,status='
old')

read(3,,,err=98)x(n),y(n)
n=n+ 1

goto 1
98 naa•--n-.i
write(,, 1O)infile,ndat
10 format(/lx,a,20,"
n,,mberof dataread= ',i3)
close(unit=3,status='keep')
else

write(,,,)
write(,,,)' please
enteryourdatalinebyline:x,y <retura>
write(,,,)' terrn•,•ate
dataentrywith anycharacter,
e.g.,
n=l

read(.,*,err=99)x(n),y(n)

258
n=n+l

goto 2
99 nda•=n-1

write(x:l 1)ndat
11 œorrna•(/lx,' pleasenote:nuntuber
oœdata read - ',i3)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

write(•,'(/a,$)')' savethesedatain a clislc•le


? [Y:N] >'
read(.,'(al)')ans
if(ans.eq.'y'. or.ans.eq.'Y')
then
write(*,'(/a,$)')' enternamefor disld•le(max.20chars.) >'
read(*;:(a20)')outme
open(unit=3,•e=outœ1e,status='new')
do i=l.ndat
ß

write(a:.) x(i),y(i)
end do

close(,,•it=3 ,status='keep')
endif

return

end

subroutineINPROD (mm,nn,m,n,xm
x,vl,v2)
c multiplies the matrix amx by vector vl a•d retrainsvector v2
dimensionvl (au),v2(•),•(••)
do i=l,m
s•m----O.O

do j=l,n
sum=vl (j).amx(i,j)+s,,m
end do

v2(i)=s,,,•
end do
return

end

subroutineMXPROD (rnm,na,m..n,a,b,c)
c performsmatrLx multiplication
dimension
do i=l.m ,

do j=l,m
sum=O.O

do k=l.n

sum=sum+a(i,k),b(k,j)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

end do

c(i,j)=sm
end do
end do
return

end

subroutineM 0 STSQ(iopt,ndat,ncol,
G,d,m,qlz,v,vt,q,qi)
c linear extremal inversionusingthe most-squaresmethod.
real g(lOO,40),d(lOO),v(40,40),•(40,40),q(40),qi(40)
+,vqi2(40,40),b(40),vtb(40),vqb(40),y:m(lOO),yp(lOO),xm(40),x-p(40)
reM mu,mu1,m(40),msp
(40,40),msm(40,40)
character.1 opt
logical lenv
write(*,'(/a)')' do you want solutionenvelopes(E)or extreme
+parameter sets(S) ? '
wr/te(*,'(a,$)')' pleame enter E or S '> '
read(, ,' (al)')opt
if(opt .eq.'E'. or.opt.eq.'e')then
lenv=.true.
nkol=l
else

lenv=.false.
nkol=ncol
endif

c initi•lise solution arrays


do i=l,akol
do j = 1,ncol
m•p(i,j)=m(j)
m.m(i,j)=m(j)
end do

26O
end do

iflag=0
calculate the limits of model parameters consistengwith •he data
c by a method ak/n to Jackson(1976)-JGP•,v.81,no.5,p.1027-1030.
c..... main loop for nkol e.,ctremizations
.........
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

do ns:l,nkol
c fi•st, form projectionvector,b
do i: 1,ncol
if(lenv)then

endif
end do

if(ifl•g.eq.O)then
c get the max/m•rn permissible mis•t; (i.e., thresholdresidual)
write(.,35)qls
35 format(/lx,' notetha• leastsquaresresiduals, qls =',f14.10)
write(.,'(/a,$)')' enterdesiredthresholdresidualQt (>Qls)'>.'
read(.,.)qt
c if Qt lessthat optimal leastsquaresmisfit (Ql.q),skip analysis.
if(qt.lt.qls)then
write(.,.)' Qt < Qls, searchterminated'
return

endif
endif

iflag=l
calculate relevant qu•utities

do j =l,ncol
do i=l,ncol
if(iopt.eq.3)•hen
ßvqi2(i,j)=v (i •j). qi(j)** 2.
else

vqi2(i,j) =v (i,j) / q(i)/ q(i)


endif

261
e•d do
end do

call i=prod(40.40,ncol,ncol,vt,b,vtb)
call inprod(40,40,ncol,.ncol,vqi2,vtb,vqb)
mul =0.0
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

do i=l,nco1
if(iopt.eq.3)then
mul=(v• b(i)ßqi(i))** 2+m•l
else

mul=(v•b (i)/ q(i))* • 2+real


endif

end do

mu----s
qrt( (qt-qls) / mul )
do i=l,ncol
xms=mu.vqb(i)
calculate and savemost-squaresestimat4s
msm(ns,i)=m(i)-xms
m(i)=msm(nq,i)
msp(ns,i)=m(i)+xms
xp(i)=msp(ns,i)
end do

calculate residuals for estimates

call inprod(lOO,40,ndat,ncol,g,xm,ym)
call inprod(100,40,ndat,ncol,g,xp,yp)
sump=O.O
summ--O.O

do i= 1,ndat ß

sump=sl•mp+(d(i)-yp(i))**2.

end do

write(,,*)' paraxn.no.misfitiplus misfitimi•us Qt'


write(,,,) rm.$11mp,$•mrn,qt
end do

c ...... end of main loop............


if(lenv)then
write(S,'(/a)')' most-squares
upperandlowersolutionenvelopes'

262
else

write(3,'(/a) •)' most-squaresextreme parameter sets '


endif

write(3,•)' plussolutions'
do ns-l.nkol ß
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

w-it( )
end do

write(3,.)' minussolutions'
do ns=l,nkol
write('3,.)(msm(ns,i),i=l,ncol)
end do
return

end

subroutineSVD(n,m,a,u,v,q)
c Singula•Value Decomposition routine. Basedbn R..L.Parker'sFortran
c tr•.•lation of an originalAlgol programfrom: Willd•son & l•in.•ch,1971,
c Handbook for Automatic Computation Vol 2- Linear Algebra, pp140-144.
c Method: The matrix A(n,m) is decomposed.Singularvaluesin Q,
c Pre-matrix in U. Post-matrix in V. The array ]• is usedas working space.
REAL a(100,40),u(100,40),v(40,40),q(40),e(400)
tol = 1.0e-77

eps = 1.0e-10
do i=l,n
do j=l,m
u(i,i) = a(i,j)
end do
end do

c Householderreductionto hi-diagonal{orm
g=O.O
x=O.O

do i=l,m
e(i) = g
s = 0.0

1 = i+l

do j=i.n

263
s - u(j,i)•2 + s
end do

if ($ .ge. tol) then


f - u(i:i)
g =-sign(sqrt(s),f)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

h - f,g-s
u(i,i) -
if (1.1e.m)then
do j=l,m
s - 0.0

do k=i,n
s - u(k,i),u(k,j) +s
end do

f - s/h
do k--i,n
u(k,j) - u(k,j) + f.u(k,i)
end do
end do

end if
else

g=O.O
end if

q(i) - g
s = 0.0

if (1.1e.m)then
do j=!,m
s = u(i,j)**2 + s
end do

end if

if (s.ge.tol)then
f= u(i,i+l)
g - -sign(sqrt(s),f)
h = f.g-s
u(i,i+l) - f-g
if (1.1e.m)then
do j-l.m

26•
½(j) =
end do

end if

if (1.g•.n)go •o 40
do j--1,n
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

s=O.O

if (1.1e.m)then
do k=l,m
s= u(j,k)*u(i,k) + s
end do

do k=l,m
u(j,k) = u(j,k) + s.e(k)
end do
end if
end do

else

end if

4O y = abs(q(i)) + abs(e(i))
if (y .•. x) x=y
end do

Accumulationof right-handtransform•(v)
do iback= 1,m
i = m+l-iback
if (g .he. 0.0) then
h=
if (1.1e.m)then
do j--1,m
v(j,i) =
end do

do j=l,m
s- 0.0

do k=l,m
s = u(i,k).v(k,j) + s
end do
do k=l.m

265
v(k,j) = v(k,j) + s.v(k,i)
end do

end do

end if
end if
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

if (1.1e.m)•hen
do j=l,m
v(j,i) = 0.0
v(i..j) = 0.0
end do
end if

v(i,i) = 1.0
g = e(i)
1 =i

end do
Acc?•mulation of lePc-hand t;ransform.q

do iback= 1,m
i = m+l-iback
1 =i+l

g = q(i)
if (1.1e.m)•hen
do j=l,m
u(i:j) = 0.0
end do

end if

if (g .he. 0.0) then


h - u(i,i),g
if (1.1e.m)then
do j=l,m
s=O.O

do k=l,n
s = u(k,i),u(kd) +s
end do

f = s/h
do k=i,n
u(k•j) = u(k,j) + f,u(k,i)

266
end do

end do

end if

do j=i,n
u(j,i) = u(j,i)/g
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

end do

else

do j=i,n
•O,i) = 0.0
end do

end if

u(i,i) = u(i,i) + 1.0


end do

Di•gona.llzationof bi-diagonalform
eps: eps*x

do kb•ck= 1,m
k = m+l-kback
Tes• F-splitting
10 do lb•ck=l,k
1 = k+.l-lback

if (abs(e(1)).le.eps)goto30
if (abs(q(1-1)).le.eps)goto20
end do

Ca.ucenationof e(1),if (1.gt;.1)


2o c --0.0

s --1

11 = 1-1

do i=l,k
f = s*e(i)
e(i) = c*e(i)
i• (•b•(O .g•. ep•)
g = q(i)
q(i) = sqrt(f,f + g,g)
h -- q(i)
c =
s -- -f/h
do j=l•n
y = u(j,11)
z = u(j,i)
u(j,11) = y*c + z,s
u(j,i). =-y,s + z,c
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

end do

end if

end do

Test F-convergence
30 z = q(k)
if (1 .he. k) then
Shift from bottom 2 x 2 minor

x=
y=
g: e(k-1)
n= •(k)
f= ((y-z),(y+z) + (g-h),(g+h))/(2.0,h,y)
g= sqrt(f,f + 1.0)
f= ((x-z),(x+•) + h,(y/(f + sign(g,f))-h))/x
Ne• q-r transfo•ation
c = 1.0

s = 1.0

lplus = 1 + 1
do i=lplus,k
g - e(i)
y - q(i)
h = s,g
g = c,g

z = sqrt(f,f + h,h)
e(i-1) = z
c = f/z
s = h/z
f = x,c+g,s
g =-x,s+g*c
h - y,s
y = y*c

268
do j-l,m
x - v(j,i-1)
z -- v(j,i)
v(j,i-1) = x•c+z*s
v(j,i) =-x,s+z,c
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

end do

z = sqrt(f.f + h-h)
q(i-1) = z
c =.f/z
s = h/z
f = c.g+s*y
x =-s*g+c,y
do j=l,m
z=•fi,i-•)
z=u(j,i)
u(j,i-1)=y*c + z*s
u(j,i)=-y,s + z*c
end 'do

end do

e(1) = 0.0
e(k) = f
q(k) = x
goto 10
end if

Co,,vergemce
if (z .It. O.O).theu
q is made non-negative
q(k) = -z
do j=l,m
v(j,k) = -v(j,k)
end do
end if
end do
return

end

269
APPENDIX B

WEI••: A simpledemonstration
programfor nonlinearinversionof
Wenet soundings
by ridgeregression.
programVV-ENINV
C A simpledemonstration of non]Jnea•
dc resistivity
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

C inversionby ridgeregression
usingthe SYD method.
C The Wennet/Offset
Wennetconfixationin assnmed
and
C inverts data ar spacings
of 0.5,1,2,4,8,1õ,
......
C Handles a maxirn-m of 5 layersin the presentform.
C Alteration •o handlemore layersis trivial.
C Author: Meju, M.A.
COMMON/SOLUTION/MOD,N2
P•AL RM(5),DM(4),Y(22),MOD(9),DOBS(22),AB(22),DCALC(22)
P•AL A(22,9),U(22,9),V(9,9),Q(9),DMB(4),RMB(5)
INTEGER EXTIT
CttA•CTER*20 INFILE
C setup flagsandperforminitializations
SSQ=l.E+10
TOL=0.0003
EXTIT = 0
INTIT=0
N2=l

c readfielddatastoredas[ab,d]pairs'ab=exper/mental
geometry,
d=data.
WRXTE(,,'(/A,$)')' ENTEt• DATA FILENAME'> '
PmAD(,,' (A0)')IFmV,
OPEN(UNIT=3,FILE=LNFIZE,
STATUS=' OLD')
N=I

t0 m•A•(a,,,ENV=0•) ^n(N),nOnS(N)
N=N+I
GOTO 10
99 CONT•XIUE
NDAT=N-1

CLOSE(UNIT=a,STATUS='KEEP')
WPaTE(,20)nmv,,NnAT
20 FORMAT(/2X,A20,
' NUMBER OFDATAR•AD = ',i3)
C readguess
model;
inputparameters
•reNLAYlayerresistivities,
tLM

27O
U and NLAY-1 depths to layer boundariesfrom the surface, DM.
WRITE(*,'(/A,$)') 'ENTER NUMBER OF LAYEI?•SIN GUESSMODEL :> '
P•AD(*,-) NLAY
NCOL=NLAY,2-1

WRITE(*,'(/A)') 'NOW ENTER LAYER I•SISTMTIES (ohm-m):'


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

READ(• ,- )( P•M(I),I= 1,NLAY)


WRITE(*,'(/A)') ' NowenterDEPTHS(metres)to layerBOUNDAi•?F,S-'
READ(.,,)(DM(I),I=I,NLAY-1)
C initiate iterative refinementof Kuessmodel
1 CONTINUE

c first solvethe forwardproblem•ving DCALC andform cliscrepa•cy


c vector Y = Dobs-DcaJc.'Obtain S,,rnof SquaresError, SSE=Y**2.
CALL MISFIT(NLAY,RM,DM,NDAT,DOBS,DCALC,Y,SSE)
WI•ITE (. ,30) EXTIT,SSE
30 FOPsMAT(/' Iteration n•mbcr = ',I4," Sumof squareserror= ',F10.4)
C checkfor convergence. are stoppingcriteria-satisfied
?
•F(SSQ.LT.SSE)ThEN
WI•ITE(*,,)' DIVERGENCE' ITERATION STOPPED'
GOTO 999
END IF.

i• (•SS(SSQ-SSE).LE.TOL) Tn•,N
W•UTE(*,,)' SLOW CONVE•ENCE: ITEP•AT•ON ABO•TEn'
GOTO 999
ENDIF

SSQ=SSE
C save best solution in arrays KMB and D MB
• o 40 •= :,NnAY-:
m• (i)=m•(I)
DM•(I)=DM(I)
4O CONTh'qUE

Pa•S(NL•Y) =m•(NLAY)
C write out current model

WPaTE(.,'(/^)')' cum•NT •OnEt IS ASFOLLOWS'


W•TE (•,•.)(m•B (I),•= •,NnAY)
WPaTE(-,,)(DM•(I),I=•.•AY-•)
C calculatepartial derivativesfor the inverseproblem

271
CALL PAltTIAL(ND AT,DCALC,NLAY,1tM,DM,A)
C calculate SVD of A

CALL SVD(NDAT,NCOL,A,U,V,Q)
c calculatemodel correctionby ridgeregressionand return updatedmodel
c MOD in commonblock/SOLUTION/
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

CALL 1tlDGE(Y,NCOL,NDAT,U,V,Q,SSQ,INTIT,NLAY,R_M,D
M,DOBS)
C checkthat controlflag is in order and update parameters
IF(N2.EQ.0)THEN
GOTO 999

ELSEIF(N2.EQ.•)THEN
DO 50 I=I,NLAY-1
J=I+NLAY

aM(•)=MO•(•)
DM(I)=MOD(J)
5O CONTINUE

tLM(NLA¾)=M 0 D(NLAY)
ENDIF
EXTIT=EXTIT+I
C repeat iterative operation
GOTO 1
999 CONTINUE

WRITE(.,'(/A/)')' *****INVEILSION COMPLETED******'


WltlTE(.,*)' Here are the optimumparameters:R.HOSAND DEPTHS'
WtLIT E ( * ,. )(RMB (I),I= 1,NLAY)
WRITE( .,. )(D MB(I),I= 1,NLAY-1)
STOP
END

subroutineBIDGE(Y,NCOL,NROW,U,V,Q,SSQ0,INTIT,Nlt,R.M,DM,DOBS)
C RIDGE REGRESSION 1tOUTIhrE. M.A.MEJU 1986

C 1tef: Meju,M.A.,1992. Computers& Geoscience,vol. 18,99-118.


COMMON/SOLUTION/P,N2
I•AL Q(9),v(9,9),U(22,9),X(9),QK(O:10)
+, Y(22),YT(22),UT(9,22),VQ(9,9),UTY(9)
+, DOBS(22),DCALC(22)
+, RM(5),D M(4),l•T(5),DT(4),It(4),P•D (9),P(9)

272
C INITIAL CONSTANTS AND FLAGS
ND=NR-1

FC=•./•0.
N2=0

SSQT0=SSQ0
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ND=NR-1
NPAP•I=NR+ND
C SET MARQUARDT DAMPING FACTORS FOR R2DGE l•GP•SSION
C FIND SMALLEST/LARGEST SING•AR VALUES QS AND QL
QL=O.00000001
QS=10000000.
DO 10 I=i,NCOL
IF(Q(I).GT. QL)TttEN
QL=Q(I)
ELSEIF(Q(I) .LT. QS)TttEN
QS=Q(I)
END IF
10 CONTINUE

QL=10,QL
QS=QS/lO.
CALCULATE TEN SAMPLES OF TI• FUNCTION QK = A + BK**2
DO 20 K=l,10
QK(K)=( 100,QS-QL+ (QL-QS),FLOAT(K)** 2)/ 99.
2O CONTINUE

QK(0)=0.0
DO 30 IK--1,11.
INTIT=INTIT+I
CALCULATE DAMPING FACTOK
BETA=QK(ll-IK)**2
C dampQ to avoidsingultrix, get Q-inverse,andU-tr•n•pose
DO 40 I=I,NCOL

DO 50 J=I,NROW
UT(I,J)=U(J,I)
50 CONTINUE

CALCULATE V/LAMBDA

273
DO 60 K=I,NCOL
VQ(K,I)=Q1.V(K,I)
6O CONTINUE
4O CONTINUE

CALCULATE INNER PRODUCT UT.Y


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

CALL INPROD (9,22,NCOL,NROW,UT,Y,UTY)


CALCULATE I•GRESSION ESTIMATE X.

CALL INPROD (9,9,NCOL,NCOL,VQ,UTY,X)


NN=0

c place boundson sizeof perturbationsusingthe Smoothness


Criterion of
c JACKSON(1973) GJP•S 35,121-136.
c If SQRT(SUMSQ(X)/NCOL).GT.1, solutionnot physicallyreaJizable.
c therefore, DECI•ASE SIZE OF X WITttOUT CttANG•G ITS DIRECTION
72 SMC=0.0

DO 80 M=i,NCOL
SMC=X(M)**2+SMC
C IF X TOO LARGE,SET FLAG NN=l
IF( AB S(X (M)). GT. 3.0)NN= 1
8O CONTINUE

SMC=S QR.T(SMC/FLO AT(N COL))


IF(SMC.GT.I'.0)TItEN
C, DECt•ASE MAGNITUDE OF PAtLiMETER INCI•MENTS
DO 73 K=I,NCOL
73 X(K)=X(K),0.9
GOTO 72
END IF

C**END OF AMPLITUDE CHECK; PROCEED WITH LINE SEAR.CIt IF NN=0


IF(NN.EQ.0)TYiEN
DD=0.0

H(1)=DM(1)
DO 75 I=2,ND
H(I)=DM(I)-DM(I-1)
75 CONTINUE
C UPDATE PAILSMETER. VALUE AFTER RF,-SCALING X

CONTROL RATE OF CHANGE OF PARAMETERSBY FACTOR.,FC (30%)


DO 9O I=I,NR

27•
PdvI
D (•) = •0.**( AL0 C10(RM(I)) +X (I)* FC)
IF(I.LE.ND)THEN
J=I+NR

tt (I) = 10.**( ALOG10(I-I(I)) +X(J), F C)


DD=DD+H(I)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

tLMD(J)=DD
ENDIF
90 CONTINUE
CHECK FOR. NEGATIVE MODEL PARAMETERS
DO 92 I=I,NCOL
IF (tLMD(I).L Z. 0.)TttEN
WRITE(*, (A)) NEGATIVE PARAMETEI• FOUND"
P•TUR. N
END IF
92 coNTINUE
C passthe elementsof tLMD backinto resistivi'ty
and depthparameters
DO 95 I=i,ND
J=I+NR

RT(I)=R/VlD(I)
DT(I)=R.MD($)
95 CONTINUE

I{T(NR)=RMD(NR)
CALL MISFIT (NR,RT,DT,NROW,DOBS,DCALC,YT,S
SQT)
IF(INTIT.LE.11)TWEN
IF(IK.EQ.1)TtIEN
WRITE(,,'(A)')' **STAGE2 (INTER•AL) ITERATIONS:RIDGE**'
WRITE(,,'(" ESTIMATEDMISFIT AND DAMPING FACTOI{")')
END IF

WR/TE(,,,)SSQT,BETA
ENDIF
C STOPPING CRITERION

IF( SSQT.GT.SSQTD)THEN
RETUR/•
ELSE

SSQT0=SSQT
CHECK THAT INTEltFACE DEPTHS ARE IN ORDER

275
DO 100 J=2,.•'D

I• (•T( •- • ). G•,.• T (•))T••


WR2TE(,,'(A)') ' ABNORMAL INTEI•ACE POSITION FOUND '
tLETURN
END IF
100 CONTINUE
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

C SAVE BEST-FIT SOLUTION

DO =,NPAaM

110 CONTINUE
N2=l
BETA0=BETA
END IF
END IF
3O CONTINUE
I•TUR.N
END

subroutineINP R0 D (• • ,n•,m,n,matrix,vec,x)
c multiplies a matrix by vector vec •nd returns vector x
real vec(n.u),x(rnm),matrix(mm,an)
do 10 i=l,m
s,,m=0.0

do 20 j=l,n
s,,m=vec(j),mat•x(i,j) +s-m
20 continue

10 continue
retro

end

subroutineMISFIT (NLAY,tLM,DM,•AT,D OBS,DCALC,Y,SSF.)


C call forward routineF•, obtainsmodelresponses FCALC
C and discrepancyvectorY as well as residualerror SSE.
REAL tLM(5),DM(4),Y(22),DOBS
(22),DCALC(22)
CALL FWP• (NLAY,tkM,DM,D CALC)

276
SSE=0.0

DO 10 I=i:NDAT
Y (I)= ALOG10(DOBS(I)/D CALC(I))
SSE=SSE+Y(•),,•
lO CONTINUE
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

CONTINUE
END

subroutineFWl{D(nlay,rm,am,Azho)
c Computeslayeredearthresponse for the DC Weuner/O•et Wennetcase.
c P•esponses
= apparentresistivities
(Arho)at spacings
a =0.5,1,2,4,...1024.
c modelp•rameters arerm andhrn. rm = layerresis.; h = layerthic•esses;
c t = resistivity
traudorms; c -- filtercoef•cients
ofBiwenandBarker(1994).
realt(21), c(11),rm(5), h•(4), Arho(22),din(4)
data c/-0.000409,-0.009047,0.065211,-0.1741571,
0.38731,
&:-0.0431,
2.093407,-1.622729.,
0.3383152,-0.039963,
0.004206/
get layer t;hicknesses
•,,(•) = am(•)
if(nlay.gt.2)then
do 100 i=2, nlay-1
•,•(i)=a•(i)-a•(i-•)
100 continue

ca/••te resistivity •ra.usform,


==0:•/64.
do 20O nk=l,2]•
x;r----xr* 2.

xx=a/og(xr)+0.246844
x=,•(x•)
•=•(m•y)
•(m•y.•t.•)t•,•
do 300 kk= 1,nl•y- 1
r•=rm(nlay-kk)
zw=•(m•.v-•)/x

•= (• + m,•)/( 1.+ m, • / •)

277
300 continue
endif

t(nk)=bb
200 continue

c **evaluate convolution*•

do 400 nk=l,ll
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

do 500 nkk=l,ll
•=r•+ •( • +m•-• ),c( • 2-•k)
500 continue

j•=nk
Art•o(j•)=rrm
400 continue
re%urn

end

subroutine
PAP•TIAL
(NDAT,DCALC,NLA
Y,RM,D
M,A)
Compute partiM derivativeby nnmericMforwardcH•erencimg
c perturb each model,find forwardresponse,
subtractcalc (urnperturbed
c responsefrom subroutineMISFIT) from perturbedresponse mudfill up
c correspondingcol•wnnso• A-matrix. NOTE: for rapid convergetrace,
forward
c differencesnot divided by perturbation factor D1ELbut solution scmled
c accordinglyin R1'DGElater.
REAL A(22,9),DCALC(22),PDATA(22),RM(5),DM(4)
DEL=0.03

D O 10 J= 1,NLAY
SAFE=R.M(J)
C perturb by del-m in log space
R.M(J) = 10.**(ALOG10(RM(J))+DEL)
CALL FWRD (NLAY,RM,DM,PDATA)
C FILL•G COLUMNS OF MATRIX A

D O 20 I= 1,ND AT
A(I,J) = ALOG10(PDATA(I)/DCALC(I))
2O CONT•UE

R.M(J) = SAFE
10 CONTINUE

278
C repeat for depth parameters
DO 30 J = 1,N•LAY-1
SAFE -- DM(J)
DM(J)=10.,-(ALOG10(DM(J))+DEL)
CALL FWtLD(NLAY,tL¾I,DM,PD ATA)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

C FILLIN G C 0 LUMN S 0 F MATRIX A

DO 40 I=I,NDAT
A(I,J+NLAY) = ALOG10(PDATA(I)/DCALC(I))
4O CONT•N•

DM(J) = SAFE
3O CONTINUE
KETUKN
END

subroutineSVD(n,m,a,u,v,q)
C insert originalroutine5om SVDINV here .tousethis program.
return

end

279
APPENDIX C

G••I1VV: ,4,simple
program
for•wo-cHm_•m.•iona•
gravi•y•nversion
by ridgeregression

prograzn G1R.
AVINV
c forward mudinversemodellingprograznfor 2-dimension•linterpretation
c of gravity dat•. The forwazdproblemis solvedusing%heTa2wami
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

c methodmudthe inverseproblemusesridgeregression techniques.


c The matrix i=versionemploys%hesingulazvMue decomposition method.
c Ch•.nnelAssignments: chain.1 (input)-fielddata amdmodelpazazneters
c chain.3 (output)-modelling results
c lamguage:FORT•N77
c Author ßM.A.Meju,1993
c Modification History
c Version1.0' ori•na• development,
1993July; 50 prismseachhaving
c a maxim•,m of 50 vertices.

reMfix(
100
),fz(100),x(
50,50),z
(50,50)
,rho
(50),gcal
(100)
+ ,rhoc( 50),rhob(50),amora
(100),y(100),ey(100),reg(100),
+
integer extit,nsides(50)
common /sol/rhoc,nm2
common/model/x,z;npol,n•idem,rho
common/fdata/amom,•,fz,ndat
chazacter*1 head(20),az•s,opt,
d•et,a•l •dtype
logical liuv
c

c set up plotting parameters


c call pltparm(usersupplied)
isc=O

liuv=.false.

c readfield data (stationlocationsand gravity•uom•l•es)


19 continue

call getdata(fx,fz,•nom,ey,ndat,head,nc)
c computeregionaltrend. residualanomalyand displayaJldata sets
write(•,'(//a)') "NB: 2 formsof the Field Data displayednext:'
write(•,,)' the actualBougueranomalyand the residuals'

28O
17 continue

call plottin• routine.NB: g•avplotshouldcall SIMPI•G hereto estimate


c the regionalfield KEG (seeSect.10.6.2.3).
call g•avplot(npol,nsides,x,z,ndat,fx,anom,•cM,head,nc,0,
q- 0,reg,ey)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

write(,,'(a,$)')' chan•eplotscale? [Y:N]>'


•a(.,'(•)')a•
if(ans.eq.'Y'. or.a•s.eq.'y')then
c call chpltsc(usersupplied)
isc=l

endif

write(,,'(/a,$)')' anotherfielddatadisplay? [Y:N] >'


•d(,,'(•)')a•
i-f(ams.eq.'Y'. or.ams.
eq.'y')then
write(.,'(/•,$)')' S•e [S]or A.uother
[A]dat•-set? [S:A]•'
re•d(.,'(•)')aset
if(dset.eq.'S'.or.dset.eq.'s')tken
•oto 17
elseif(dset.eq.'A'.o•.dset.eq.'a')then
goto 19
endif

endif

c determine what form of data is to be interpreted


write(.,'(/a)')' note:eitherof the 2 formsof the data may be'
write(*,,)' interpreted:i.e., Bouquetor I•idual anom•y. '
w•te(,,'(/•,$)'): model•ou•uer(B)orre•iaua•(l•)aat• ? [B:i•]• '
read(,,'(al)')dtype
if(dtype.eq.'19.'.or.dtype.eq.'r')then
do i:l,nd•t
aaom(i):anom(i)-reg(i)
end do
endif

c setreg to zero as this constantbiasis addedontoforwardresponse


do i=l,ndat
reg(i)=O.O
end do

281
c determine relevantoperationßinversionof forwardmodelling
write(.,'(/a)')' pleaseselectdesiredoperation'
write(*,'(a,$)')'EnterI (Inversion)or F (Forwardmodelling):>'
r•a(.,'(,•)')
if( aris1.eq.:I'. or.aris1.eq.'i ')then
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

linv=.true.
else
linv=.œa2se.
endiœ

write(,,'(/A)') ' note:constructed


moae•readnextfroma
53 cM1getmod(npol,nsides,x,z,rho)
nparm--npol
18 continue

if(.not.linv)then
25 continue

write(,,'(•) ')' Yourcurrentmodeli• asfollows:'


call print(npol,rho)
call miser(ey:y,•uom,reg,ndat,ssq,npo!,nsides,x,z,•,•z,rho,gcM,
+ a)
call•ravplot(npol,nsides,x,z,ndat,fx,anom,gcal,hesd,ncjl,0,
9- reg,ey)
if(isc.eq.0)then
•te(.,'(•,$)')' c•a•½ plot•• ? [¾:•] • '
•a(.,'(•)')a•
if(ans.eq.'¾'.or.ams.eq.'y')then
c call chpl.•sc(user
supplied)
isc=l

endif
isc--1

endi•

c model ch•n•g
write(,,'(a,$)')' stopinteractive
modellin•
now? y:n•'
•a(.,'(•)•) a•
i/(am.eq.'Y'.OR..ans.eq.'y')then
write(,;'(A:$)')'
PROCEED
TO AUTOMATICINV•I:LSION
•a(.,'(•l)') opt

282
if(opt.eq.'Y'. 0 R.opt.eq."y')then
goto 35
else

wr•te(-,'(/a/)')' Endof modell•r•õ


sess•or•:
Have a niceday !'
do •-- 1,npol
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

rhob(i)=rho(i)
end do

goto 5
endif
endif

call aajust(rho)
goto 25
endif
35' continue
c**END OF INTERACTWE FORWARD ,MODELL•G**
nrow=ndat

ncol=npol
c initi•llze variables
tol=O.001
stol=10000000.
NN2=i

SSQ0=10000000.
intit=0
extit=O

c STAllT THE ITEtLATIVE PROCESS


i continue

w•te( * ,102) extit


102 format(/' EXT. ITEKATION NUMBEB.-',I2)
c print current model
iI(extit.LE.15)callprint(npol,rho)
extit=extit+l
c calculate relative difference between field data mad initial model
c response madform discrepancy vector,Y. Vectorscaledto reflect
c differingobservationalerrors(wheregiven).Calculatesumo•
c squaresof residuals,$SQ
callmisfit(ey,y,anom,reg,ndat,ssq,npol,nsides,x,z,ix,iz,rho,gcal

283
+ ,a)
write(•,103)SSQ
103 format(/' SSQ= ',Fll.5)
C STOPPING CtLITE•ON 1: if $SQ largerthan previousiterate's, STOP
if( SSQ0.LT. SSQ)•hen
goto 2000
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

emdif

c save best solutionin array R•OB


SSQ0=SSQ
do 40 i=l,npol

40 continue

ckisq=SSQ0
flag=stol-ckisq
c STOPPING CRITEtLION •:

if(fiag.GE.tol)then
goto 80
else

goto 2000

80 continue

c calculate the SVI) of A

call svd(nrow,ncol,A,u,v,q)
c apply ridge regression for optimal solution
c...this is a two stagesolutionprocess...
call ridge(y,ey,ncol,arow,a,u,v,q,ssq0,•tit,reg)
c STOPPING CRITEtLION 3: if dampingfails first time, skip•qdge
c regression
if(NN2.EQ.0)then
goto 2000
elself(NN2.EQ.1)the=
do k-- 1,mpol
rho(k)=rhoc(k)
enddo
endif

stol=chisq
GOTO 1
2000 CONTINUE
c...INVEP•ION COMPLETE, PRINT RESULTS
write(*,'(/A)')' •-END OF.MODEL SEARCH** '
write(,,'(/)')
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

write(*,;("EXT.ITERATIONSANT.IT SUM-SQUAtLE
MISFIT")')
write(ß,• )extit- 1,intit, chisq
6 contLuue
call •vplot (npol,•sides,x,z,ndat,fx,a•om,gc•l,head,nc,
+ 1,1,reg,ey)
write(,,'(A,$)')' CHANGEPLOT SCALE? [Y:N] >'
read(,,'(A1)')•.•
if(ans.EQ.'Y'. OR.ans.EQ.'y')then
c call chpltsc(u•ersupplied)
isc=l

goto 6
endif
C...WRITE OUT I•SULTS TO OUTPUT FILE.

5 write(,,'(A,$)')' WRITE RESULTSTO A FILE ? [Y:N] >'


read(,,'(A1)')ams
if( ams
.EQ.'Y'. OR.•,• .•.EQ.'y')•hexx
c•11output(head,chisq,npol,rhob)
endif

write(,,'(A,$)')' SAVE tLESULTSAS A PLOTFILE.? [Y:N] >'


re•d(,,'(A1)')a•
if( a•.E Q.'Y'. OR.•.•.EQ.'y') •h•
call •vp lo• (npol,asides,x,z•uda•,•x,a•om,gc•l,he•d,,•c,
+ 1,1,•eg,ey)
endi•

write(,,'(A,$)')' TERMINATE MODELLING SESSION? [Y:N] • '

if(ams.
EQ.'N'. 0 R.a•. EQ.'n')then
write(,,'(A,$)')' models•e(S)or areother(A)
data-set? [S:A]>'
read(*,'(A1)') •et
i•(dset.EQ.'A '.0 P•.dset.EQ.'•')them
go•o 19

285
elseif(riser.EQ. 'S'.0 R. dest.E Q.'s')then
write(*,'(A)')' returnto Forwardmodeling(F),
Inversion(I)?'
re•d(*,'(X•)')a•
if( ams1.EQ.'F'. 0 R.ans1.EQ.'f') then
linv= .false.
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

goto 18
elself(a•us
1.EQ.'I'. 0 R.ans1.EQ.'i') then
write(*,'(A)')' YOUR CUI:[•NT MODEL IS AS FOLLOWS-'
call print(npol;rho)
call adjust(rho)
goto 35
endif
endif
endif
52 write(,,'(/a/)')' Endofmodelling
session:
Havea niceday!'
stop
end

**SUBROUTINES**

subroutinegetdata(fx,fz,anom,error,ndat,head,nc)
c reads•ravity measurement locatiom•d fi•d obse•atiom
c •d •sociated •c••ties

c •put Stmct•e: twodataforests•e accepted


•d de•• •a controlCONS.
c (1) postion(m)•d •omMy(mg•) forw•ch CONS= 0 OR
c (2) postion(m),
•omMy(mg•) •d top,height(m)for
wM• CONS=1
reM •( 1O0),k( 1O0),•om (1O0) ,e:or(100)
••acter.20 ingle

••acter, 1 head(20)
c read project header
•ie(.,'(//a/)') ' EnterProme/proj•t

' >'
read(,,'(•O)') •le

286
open(unit=l.file=i•ile,status='old')
c read input data-typecontrol
read(1,•)cons
do n=l,1000
if(cons.eq.O)then
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

readil,,,end=999) fx(n),anom(n)

else

endif
end do

999 ndat=n-1

write(,,10) head,ndat
10 format(/' ',20al,'number
ofgravitystations
read=.',i3/)
close(1,status='keep')
do i= 1,20
k=21-I

20
end do
20 nc=k

ca2culateweightingf•ctorsto be usedfor data standardization


c set errorsto ,m•ty **•ctual errorsnot read in this version**
do i= 1,nd•t
error(i)=l.0
end do
return

end

subroutine
getmod(npol,nsides,x,z,rho)
c readsinputmodelgeometry.
c INPUT FILE structure:

c Dat• 1' n-mber of polygons,NPOL.


c (polygons
assigned
n,•mber
1 to NPOLforidentification
purposes,POLNUM)
c Data 2' for eachpolgonthere axetwo l•-es of input.
c Linel'bodyid(POLNUM),no.ofsides(NSIDES),
den.contrast(g/cc)RttO
c Line2:nsidespairsof x- andz-coordinates
of bodyreadclock•se

287
and closingat staxtLugpoint.

• x(•o,5o),z(•O,•o),•o(•O)
integer nsides(50)
character, 20 infile
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

c read input filename


write(,,:(//A:$)') ' enterinput-modelfrlenme (a•0)- >'
read(,,'(a20)') ingle
open(•mit= 1,file=iufile,status=' old')
c read numberof prismsand topo info
read(1,,) npol
c read model parameters
do k=1,npol
read(1,,)polm•m,nsides(k),rho(k)
read(1,, )(x(i,k),z(i,k),i= 1,nsides(k))
end do

dose(1,status= 'keep')
retura

end

subroutineadjust(rho)
real rho(50)
chaxacter. 1 •n•

wri•e(,,'(a,$)')' changethe demity o• a polygon7 Y:N ) '


•a(,,'(•)') a•
if( ans.eq.'Y'. or.a•.eq. 'y')then
wri•e(•,'(A,$)') ' •ype in polygonn•rnber • '
read(,,,)•lay
write(.,•(•,$)')' •nt• n•w a•nsi•y contrast:> '
read(,,,) rho(nlay)
write(.,'(A,$)') 'changeanotherpolygon7 Y:N > '
goto 28
e•d if
return

end

288
subroutine
t a12d•d(reg,ndat,npol,x,z,fx,fz,rho,n.•ides,gcal,a)
a modified versionof the originalTalw•-n•algorithm.adaptedfrom
R.Hipkin: Universit7of Edinburgh
integer nsides(50)
r• .•(50,50),z(50,50),•(50),z•(50),•(50),•ho(50),a•(•00)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

-+- ,fx(100),•(100),spdelz(50),sselz(100),thet•(100),gcal(100)
+ ,•( • 00,• 0),•( • 00)
pi2•8,•tan(1.0d0)
pi--4*•tan(1.0d0)
pih--2,ataui
1.0d0
)
do k= 1,ndat
•t•(u)--•(u)
sselz(k)=0.0
end do

main loop
do 11 nl--1,,•pol
nvert••ides(•)ol
do 420 k= 1,nd•t
sdelz---0

•xx(X)--x(x,m)-m(u)
•(•)_-•(•,m)4z(u)
•(x)-_•(1)**2+z•(1)**2
if(exx(1).lt.0.0)then
if(zee(X).•t.0.0)then
t•t•(1)--•t•(z•(X)/•x•(1))-pi
else

theta(1)--atau(zee(
endif

elseif(exx(1).eq.O.O)then
if(zee(1).lt.O.O)
then
theta(1):-l.570796327
elseif(zee(1).eq.O.O)then
theta(1):O.O
elseif(zee(1).gt.0.0)then
theta(i)= 1.570796327
endi•

289
elseif(exx(1).gr.O.O)then
theta(1)= atan(zee(1)/ exx(1))
endif
do 410 i=l.nvert,

exx(i+ 1)= x(i+ 1,nl)-fx(k)


Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

zee(i+ • )=z(i + • ,m)-f•(k)


r(i+ 1)=exx(I + 1). - 2.+ zee(I+1)**2.
if(exx(i+ 1).it .0.0)then
if(zee(I+ 1).lt.O.0)then
•hct• (•+ • )= •t •(z• (i+ • )/ •=(•+ • ))-•i

thet•(•+ • )=•t'•(z•(i+ •)/e=(i + •))+pi


endif

eheif(•,=(i+•).eq.O.O)the=
if(zee(i+ 1).lt.O.O)then
theta(i+l)=-pih
elseif(zee(i+
theta(i+l)=O.O
d,d•(ze•(i+
theta(i+l)=pih
endif

elseif(exx(i+l).gt.O.O)then
thera(i+ • ): ataa(zee(i+ • )/ exx(i+ • ))
endif

check=exx(i),zee(i+
if(check.eq.O.O)then
adz(i)=o.0
else

omega:theta(
i)-thera(i+
if((abs(omega)-pi).le.O.O
dtheta=omega
elseif((abs
(omega)-pi).
gt.O.O)thea
if(omegaAt.
O.O)then
dtheta=omega+pi2
else

dtheta:omeg•-pi2

290
endif

endif

delz(i)= (check/((exx(i+l)-exx(i))•2+(zee(i+1)-zee(i))•2)). (((ex


+ x(i+l)-exx(i))•d•he•s)+O.5.(zee(i+l)-zee(i)).•log(r(i+l)/r(i)))
endif

sdelz=sdelz+delz(i)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

exx(i)=e.xx(i+l)
zee(i)=zee(i+l)
r(i)=r(i+l)
t beta(i )= t beta(i+ 1)
410 continue

420 continue

c main loop ends


continue
return

end

subroutinemisfit(ey,y,anom,reg,ndat
,ssq,npol,nsides,x•z,/x,fz,
+ rho,gc•l,a)
integer.mides(50)
real anom(1O0),y(100),x(50,50),z(50,50) ,rho(50),ix(1O0)
+ ,iz(100),gcM.(100),a(100,50),ey(100),reg(100)
compute model response
ca21ta12d/wd(reg,nd•,npol.x.,/x,fz,rhog•sides,gcal•a)
s-•=O.O

computediscrepancy.
vectory, •ud s,,• of squarederror ssq
c *•activate as necessary:
data norma•ed by associatedstandarderrors,ey
do i=l,ndat
c.- y(i)=(anom(i)-gcal(i))/ey(i)
y(i)=•uom(i)-gcal(i)
s•=sum+y(i)•2.
end do

291
do k=l,npol
C•* do i=l,ndat
a(i,k) = a(i,k)/ ey(i)
end do

C•* e,•d do
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ssq=sum
return

subroutine
ridge(y,ey,ncol,nrow,a,u,v,q,ssq0,inti•,reg)
c ridgeregression
routine.MaxMeju,1986.Seealso
c M.A.Meju, 1992,Computers & Geoscience,
vol.18,no.2/3,pp99-118.
common/ sol/p,n2
common/model-/xx,zz,npol,nsides,rho
common/fdata/anom,f:x,fz,ndat
realfx.(100)
,fz(100),xx(50,50),zz(50,5O),rho(50),p(50),thom(50)
+ ,amora(100),y(100) ,ey(100),a(
100,50),u(100,50),v(50,50),q(50)
+ ,x(50),qk(O:1O),yt(1O0),ut(50,1O0),vq(50,50),uty(5O),ax(100)
+
integer nsides(50)
c set initial constantsmadflags
fc----1./10.
n2=O

ssqtO=ssqO
c set Marquardtd•mpingfactorsfor ridgeregression
c find smallest& largestsingular
valuesQs&:Q1.
ql=O.00000001
qs:10000000.
do 10 I=l,ncol
if(q(i).gt.ql)then
ql-,q(i)
elseif(q(i).lt.qs)•hen
qs:q(i)
end if
10 continue

ql=10*ql

292
qs=qs/10.
calculate10 saxnples
of thedampinõfunctionQk = a + bk*-2
do 20 k=1,10
qk(k)=(100•qs-ql+(ql-qs).float(k).-
2)/99.
20 continue
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

qk(O):O.O
c---main optimizationloop--
do 30 ik--- 1.11
intit =intit + 1

calculatedampingfactorbeta
beta=qk(11-ik)**2
c dampq to avoidsingul•ities,
õetinverse
q a•du-traazpose
do 40 i= 1,ncol
qi=q(i)/(q(i)*,2+be•a)
do 50 j = 1,nxow
ut(i,j)=u(j,i)
50 continue

compute v/q
do õ0 k---1,ncol
vq(k,i)=qi*v(k,i)
60 continue
40 continue

compute i,•,aer productut.y


call inprod(50,100,ncol,nxow,ut,y,uty)
calculate regressionestimatex
call inprod(50,50,ncol,ncol,vq,uty,x)
call inprod(100,50,nrow,ncol,a,x,ax)

c placeboundsonthesizeof perturbations
usingthe smoothness
c criterionof Jackson(1973)GJRAS 35,121-136.
c if sqrt(s,,msq(x)/ncol).gt.1.),
solution
notphysically
realizable
ß

c therefore.decrease
sixeof x without changingits direction
72 smc--0.0
do 80 m= 1.ncol

smc=smc+x(m)**2
c if too laxge.setflagnn to 1

293
•(,,.1:)
• (x (•)). gr.3.o)•= ].
8O continue

s•: • •t (•=•/•o •t (•o•))


•f($mc.gt.l.O)then
do 73 k=l,ncol
73 x(k)=x(k)*0.g
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

goto 72
end if

c endof amplitudecheck.proceedwith 1-d line searchif nn=0.


c updateparametervalueafterre-scaling x.
c controlrate of changeof parameters
by •actorf½(30%).
do 90 i=l,ncol
thom(I)= rho(i)+x(i) ,fc
90 continue

callmisfit(ey,yt,anom,reg,ndat',ssqt,npol,nsides,xx,
zz,fx,fz,rho,
+ gcal,a)
if(intit .le.11)then
if(ik.eq.1)then
write(*,'(/a)')' *Sgage2 Mi_•mi•.ation
by Ridgeregression**
'
write(*,'(" Estimated.MisfitandDampingfactor")')
end if

write(.,.)ssqt,beta
endif

c stoppingcritehon
if(ssqt.gt.ssqtO)then
re•mm

ehe

ssqtO----ssqt
c save best solution in array p
do 110 i----1,ncol
p(i)--rhom(i)
110 contkuue
m2=l

endif
30 continue
returm

294
end

subroutineinprod(mm,nn,m,n:vx,y,a)
c matrix vx multiplied by vectory gives vector a
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

do 10 i=l.m ,

s,,m=O.O

do 20 j=l,n
sum=s,,m+y(j),vx(i.j)
20 continue

•(i)=•um
10 continue
return

end

subroutineoutput(head,ssq,npol,rho)
rlo(50)
character*20 outfile
character*1 head(20)
c openoutput file
write(*,'(a,$)')' typein output.filename
(•20)'> '

open(,•t=3,•le=outFsle,st•tus='new')
write(3,10) head
10 format([1X,'2-D
GravityModelforSiteß',20M)
write(3,'(/A)') '.BestFit Model'
wri•e(3,15)ssq
15 format(iX,'chi-squared
misfit= ',f9.3)
write(3,20)
20 form•t(/1X,'polygon
no. density
cons•ras•
')
do i-- 1,npol
write(3,,)i, rho(i)
end do

close{3)
return

end

295
subroutineprint(npol,rho)
c prints curren• parameter estimates
real rho(50)

10 format(/1X::poly•on no. densi%y


con%ras%
•)
Downloaded 06/21/14 to 134.153.184.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

do 20 I=l,npol
ro(i)
20 con%inue
ret•

end

c subroutinesvd(n,m,•,u,v,q)
c **insert •he routine listed in SVDINV here***
c return
c end

c ** plot routines suppliedby user **

c subroutinegrayplot(npol,nsides,x,z,ndat
;f:x,anom,gcal,head,nc,O,
c q- O,reg,ey)
c ***User to supplyplotting routine**
c return

c end

296

You might also like