Professional Documents
Culture Documents
021436587:9<;>=;?3A@6B1:CDFE+;?CHG1:9<C;I9
J<KALNMOKAPQSRTPVUWYXZK\[]^KAL_RY`)KL_WZabKdceWYPSMO`)KLNMfRgQihfKjRSc
]^kYlgmonldapJ<RYq RYPrWTstRvunw]^RARYq<RdQyxzm{L_KAlrayhfRYq sTKL^apJf|~}t_YYrt}T}e
z
i^6oto/t{wA^Y66tt,Ao_6o+'t_<e_6fAt<eTA
eAf:wZ'Z/<'et{w?'<<'%6e%e<T2'A<_6<?pet{wAeY
AZ'At,r'AAt<'6_Ao6tAt6<t'e%6ef<_AAOT<o'Ao<'zTTo<Ato
A<ANVo_AAo^'<iA~<~'eTAoee)<ee$o<T6e^$A^Ao'
o'et{w$eTAZOt^_it~iOAevAoAo'<oATw4tew:to'
'66_6t'e%6A6TeY6'ZviA,T'<oTAofvAOt%A$:^_
6'Aoe)oA6eAtTV_pw6AAo^t
ptoTw%jA^Y6AtAAZ_wA8NZ6'A<'AT_<~At<eTAo'gAo
'6*A6Z/,'et{w_<#teA'g'AT^yw:'g)
*AZ
'e{6''f
)OAZ_wAN6!e_,Ze'^<Ao_/~_)e'e6'yt'eAe<Te<A^
<^'$6tA%A6TA6tON6'pAAZ'T^t<A'6T<Ae~'Ae
O_<tVt/TTot%e/<'6_oeetA%A6TA6tON6A$_<wZNAOAOt_oo'6_
oteT666T'66tONAo'A$voA^e6e:o%'o'VT66At'6Aeo_Ao'<yTw4te
AZ'teo2t'o^Y6,A<Z2'$TjO'o%j4~:V%N6tt''
o''%NA<tj:YfAoA<_et$/tv'\NZAZ'T^t
A<A''%AAovtA%A6TA6tON6<?OzAi6toO_oie'A64A<'
A<A'ozA<AT\'6tA%A6TA6tON6<t^,<N_N4t'ee
6<N:<i'
AtoeY6e~A<ATZ'T66At'6Aeo_Ao'<)e_~oAN'AAAeO_oo{w$_<~A^'
oAo'_f'et{w_<'Aoe%<ZA<AT64_6iAe%o<To<o6N
Ao'ozA<AvT66At'6Aeo_Ao'<'\i<Ai<_o\_AZ'te?6v6tONAo'v{t6t%AO_
N'e''6TeY6<tTt^A'$A<A^o^wZ_oO^>YrNYOV'<Ow
t'66AzT666T'66tONAo'\viOZ6AATONA^{t6t%AO_Ntvo
6tONAo'e_og!<
Ve_o^6Oi6/te6TtZtdNTZ_YAToNgrN,'toTw_<'{
Oe
:'<'AOA'<_v6tA'6_<te'!N_6tw*_A%%6T6o/tiNZAN
ANe^YAo'tjt'6Aeo_Ao''eTw$_<'AoT'OtZ_<t_<teAA^/o
N:ev_y'eTwAeY66Nt4AN'yt'/'e6<_ftA%A6TA6tON6Ao$A
6eAt<t':wZN6oeOA'AAo''<Z_<8'O_$6'T6Tzo~6_o{w
_p6eA6e<ToAtoo%AtAt6oAt<eTAo'<)<<tA^e6A<t'<{6<t
AtoeY6TSTA6tON6z'eTwz__o/AOI6TZ_o<?{Zp'6e_AeI6eA'oAo'N:eeNAt
N'ee<AOV6N'et%otA'6'<To'Z<e'e_Vt'ZNAo'_yT6<N
OvT_Z_o6A<_i'vT/'t%6<_'et{w'<_oOe
%NIto/t{w_<'Aoe/A^6
100 %
50 %
1
NORMALIZED AMPLITUDE
25 %
10 %
0.5
0.5
3 4 5 6 38P]ygq%zB<DA5XTA{3 rMM6X
!5X)-p
VELOCITY (km/s) (}m !&%RX M DT$:( |
% y:5* 1<}:5 i } RX M D$:( p
38P r M* } N }s
v%7 *Y% TX%: 3 T T %r 3 $
5 *3TX T
~D w$} } 7
M MT ) -1
TX%tr kklT%#pM(}T wq %TTX%uv
v%
$&5 5 hokl=pWqol=p8(} kl=1&OQwTTX, xXX 7X}X,
7%5 T5(# "X TX M D1 U %tw%6 ?y( 4 w %AwYeS ^(
N( 54 wZS
' ^
w'weNS?/^'rSe6y?y(e_w Aew^Sy'w
VA Oe zt%Zy {f z
E$e/TS {
{t
X4Y VZ /t SeN ?7 w %ZY 9:4 z ANZ
96 ^6% N
r %wZS ^u6w_ +
wNwA
ZN7{^(
N( 54wZS
^e (
Z9
4p%@ d Aw%_C $%0 ?y( e_n )eS e%YwzYe%TVSeZ $
VS(
Z6
rZtSw(
{t ^ + ' {tje 8 8$9%eSY^
Ytj d :%twZ^' BYT ^ +
Z\[g ^N_`VabEc5dfehgi_YjKk\l
mnal
Conventional
Selective 25%
Semblance
NORMALIZED AMPLITUDE
1
0.8
0.6
0.4
0.2
1 100 %
25 %
0.8 2
0.6
TRACE WEIGHTS
1.5
0.4
1
0.2
0 0.5
3 4 5 6
VELOCITY (km/s)
3 8PP}i}su
v%uMMX
! TX)-1
% ATX%:p55%55X%d kkl'%(}55 S 0
%55X%Tj,qolT%#1OQFTTX,*XXX}X,M*3%T 0 0.5 1 1.5 2 2.5
X T z
% *7M T
) A -1 OFFSET (km)
38P PA {$T 'jT5(# "X TF {*m $}TX%3T%< ^
TX%:X MD't56TX%
%DATXyFE]A* g{
/.1 2 je
K?7 ww%ZS.y? e Nw?Ze%j wZ?)je
{ 8 XA s
,5X* ') -p&}- 3rX= } } wo%* 1
wA'Z^Y!ww5gZ6sr8
{8
Z($39/eSZ,^,w( Bow
:% 6%4 Z^NwZNe%wZ{ 8%
w A
6 %$ w6V%t%%
9 Z61 x r ZeSw66wYwz
e z t Z + y ne Z%6
4 S6^Z +3U
X4 w%?y(e_www%zSe6Z0 w
wZ8
wM{ N9 6Sw6MN
yTSwZ
/ T{ ew%eNZw?Ze8 /.12
Tw%6y - w
$'e/wZ6OtB 6
(
N( 54
6_ e/
je $3{ R\? ASeZ6Twy? (e _wZ,%:/
X4 + t/
X4' V
6^
ZN ^%
6w^SSZtSw(
{t ^8$I'w
- t 8 w^S wA
^ (
N( 5 4 wZS
e?$ w %ZYwvZ^6wZ%6( '
iS%% ?\ew%%tS%Y
%6
??(^ Nwyw_% ZZr z ANZ
9 6^ A% { 6w%_Vtw:e%8
{ Z - 'w &(+ y%S
{
%6S {Vww%Z^
6_ e/
w66
X
4 ww%SeZZ6 wB/.12
tw6v tiw6
{N T 4
gt/
4' y Aw%_Vt%8
YVw)StZZ ?7 wB % Xoe
oeB 6^
ZN ^%
C
(
N( 54 e%
X'4 { + ??( ^Nw + g%ZY $?7 ww%6 %S e r8 ?y( ^Nw( $wZ@
yny ? (e _w'Z^ %ZVOt:er8 w5g Z6 ^f^
' ^ 'w ^Ys ?\ZZ /twZV ?7 ww/t f( w%6
w3 Nw6o6 (
ZNw<
eSwY - w
t$ e%v'wZS6 w)Z^
6_ e/
6weSSZeSw
T ^ +tw)S( N9
{eZ
e t ^v'wItw%6%Y$ twS%%\ ? 6{ - w '+U <SZZ6 iw ZN, ( ZN ! - w }* $C9')j
w%S(
Z6
rZtSw(
{t ^
w$ w%'wNZY wB
6w({
T^ (
X4 twB tSeZ 8W! - wZ e% }* 6 oe 3 96Sw63 SZ%tt 8 vw
w
^8r SAStZZzw/ewS%tS t/i Nw6 Y Tw T?yY z z + !E:ew:
Sw:
^NpASwe.
(
N( Z\e@
ab5a: c\a:Cl
l}abtk\,cCm\ab5c,k\mSk\bEdc5d Z[F
Conventional
Optimized Weights
NORMALIZED AMPLITUDE
0.8
0.6
0.4
0.2
0
3 4 5 6
VELOCITY (km/s)
38P^ P' } . } -p^
v%\ % TX%: PT 5%T 6
TX%j{ m7
C}
v%(}B$T'*jFAp|
v%3
XA s
,5X* ZM) A:5* 1*OQTTX, WXX X}X, { w%T
X T -1
38PK*?tFEmAS3 .C5gTX)I|TX%hA
SXG } } Jq*pS3 N:(} X V5(}} % Y(} c }#
:/ B*{k*1 o*1
100 %
One crosscorrelation wY6w6ZN^6w^SSZtSw(
{t ^" wZe%Y ! + + $
2 $C *20*,+U {K w SY
t6we)A%e e6 {wiZ
1
6wZ%6(
6( $w8 X6wZZzevw%r ' 9{v wY
NORMALIZED AMPLITUDE
1
100 %
0.8 25 %
0.6
0.4
0.2
0.2
0.4
3 3.5 4 4.5 5 5.5 6
VELOCITY (km/s)
38Ps^ e_ M* } . -p^
v%M:5h 3 mB^{k*1 o*p*r
X6% ATX%:\55%55X%vd kkl'%y(}5 S
%5 5X%T j,qolT%#1OQFTTX, *XX MX}}X C *3%T
X M D1
38P<rQB<DA5XFEAY%A@(XX)I|5X%
h < 0-p>{ p5(}} T55XT K} TT%TTX% J,t@(:(5
(
N( 54 +yx )6/T98
{ptSeZ w%AOe )5X% k'#% $$X A 1
zt%Z:ft
z4 %eSZ6^%AwZ%4s6w_B{w%:wZSZZ
4ZeNB /T{8t6wewA~te%8
z% S$fjt
ez
Z($ t/jSwer8
8{
6 P+ An 4%,' Ze8
{ A4tV6^NB /TwYV9: 4 ^8
) S +
NORMALIZED AMPLITUDE
1
6w<$ w%^ %
X wr % SzT ! z n * vZ 8 wY
:
100 %
tw%ft :%wIYi 8
% 'w%fw/66 ^e /
0.8 25 %
ww%:wNtt Yed :% twYO! *
/ '4w%:% S +
- ' w & ) S%%\? p) {
,d
ZN./.12^ tw%Ay?7 wB% Ze 0.6
ZN ^%?
Aw^SSZtSw({
t ^!$ 8
X48 8 e %6 VELOCITY (km/s)
wZS
^ SZZe/$89 twY_ \tw
NYtwZt9 pw 38P* yPf M* } B -pR
v%K m)A6XZ } } wp
w S^ (
N( 54!q\? 8{ B ?eI%ZZ6Swt
X4 w%\ZtSwZ6ye% * $ $5X%TX%:nT5%T5X%r kklT%
8^ 6j{ 8 w/Tve6( t4 w9tw 6w_{ i6 /T (}@ 5 SE%5 5X%m 5qolT%#1
98{
+ g
8 6w fw6jw($y? \%A^ %9 S6^ Yw/TIw\Y_
w%v( ^ Zoe9 tw Aw%_Z{ ({ Y\? %Zw%6 wpwZSZZp%ZtSw(
{twY S7 IZ/T9
yw)w%t
Z6_^ % SZ w v%tiwNz
! w(4:$\ * 2 e Z^
6_ e/
C
(
N( T 4t/
X4' +
8e %6 *,+ 1 ewZA
A%$ w%r{
NZt ^~pw0
(
N( T 4 Y
e ^ 6w%YvjSw^8
X4Y9:v 4 6w^SSZtSw({
t ^3oe/ w %StZZ
?7 w {
t^ 8X 6wZN{
w A
6 + ? ZY% Sw%zS(
Z6
%YO V:X ^(PO\% U V[ <% U ZO\(
ZtSw({
t ^ 6w%_ (
/ 64SeZ/ w:w/t4/%
iw({
(
X4s{
te X 6wZN{
C ^ Z' %$
7 (
/ 6\w^S4StZ U %ew6\S'wZi6Swew7 p
{N T46jzT ^ pjt
% ww/TzB{ %tww%Z^NS9' { e%B Z^
ZN ^%
8 jweS e% + g*w _
$w??Ze% {A
{
N T! 4 e/
X4 (I$ % eS ' 9 S6^ T ^!w%tw /.12 ^tw%As ?7 ww %Z6 ^2
ZNzZ^N {8 jt
w9% jwZSwS(
Z6
r ZtSw({
t ^d 8 6w_o ! ~ %S {t#e% * ?7 w#
{jwtS ^%
{t ejt/%T
ab5a: c\a:Cl
l}abtk\,cCm\ab5c,k\mSk\bEdc5d Z[F
'(
T ^ B ! - w & *,+UW d _Y
{twA
?8 R[ % U YOG% % U o(n / [ U&HU eYo V
8{8A^6e?pte
Y2w~wt/' S%:%ZZ
z8 jwtS {e%vw _Sw_ %ZB
{Tw6
y6eSw(
{t e w eI^
XV4 YeBN
Z3{ Nw6o6w7?7 w'z t Z($N(
eS(
X4
:
r8
{twYjt Z + yr _
\Otiw 4
N iz 8 TS
z t4 w,/Z6 e%iYtS A 6)Z^qS0
{4%e ZeN
/w %Z6wew 9(
A ?2 eZ%6 %y
( 4NAy ?7 w
N( 5$ 4 SZ6Sf$ %tS (8
T
X4"Ot0
%?.to{
%t + I
{N T 4 0 +
]\Te%f
( 4NAw 'ZS :+
$ "98 s twwY
Nw' @4 Nw6 %t$,%
Z($
t/w%iZ 9
6
Z tw n &A $8ew % Nw6 e
\ n
$ ? %A^ Ye Z%AtwY@'4 _w6 N/.12^ tw6n?7 wuz e:4
?
t/VY NS , ^6 ?p%
(
6M w% :%Z 4Y{ & + n +- 'w &% :'4 69
)w%Z6 ^
ZNw +
SA ?\w%iZe%6w6%Z( ^Z. Oe:w%ijT {6@ g jweSwYw e Z%AtwY2$e% S6 Z &% z T, 4 w
%Z6 ^C $ ^ % 8 Z^
6_ e/
6weSSZeSw(
{T {e! 8 %Z6 ^Vt/8{ Br8
8{
6s!E9 tw?7 w 2 SSe! jS
!S
{
% * $t/S(
ZA ^ rZeSw(
{T {e 8 !eSYp
% * 98 ^tSe f
Z * Oevw^d 8 Z% e/
3
(
N
oeK *2 f6w^SS6eSw(
{t e% + 54R _:
- w &%'
+ y%S
{ f
%u w{ Nw6e
- twZS%T $Ye z
A
Z '3 !E1 x* w(
{t
4
{
N T4q%6 ^% SZwA^ r^ (
N( { 60Oew
z
E$\z
Z %t 98
ww% Zz \jT {68 @ _'(
{ 8z w%z t4 (
ZNw($ t/w%%tS%Ye%r
wtS {e% +8U IS%% ?\s { - 'w &A $_Y _w w%Ze%6wZZ wK _wA^
t
{
N T4Bq%6 ^V% SYzwre Z%6Tw)w%KN
X
( '
Ztw7 98tSYVw% ?pT0
(
N( ZIw%t% $oew wY
YA
Z +\- ' ws S%%\ ? @'4 Nw%6 B/.12
tw6?7 w
^v %jT {6( $ttwI 9 ' ]\t ^6w%e)w%fZtSwZ6 S'wY^{
6e w!Y$t w i Nw6^ \
Y $3
6
{N T 4 + 4tw $y% ?y(
A% $?w/TwY _ vw%Ze%6 ? A^ (
6Yow(
Z%,4 & + $ e% z .
y \9
ZZv( ^Z:6 'wYV B9ewwiZ^
6_ e/
e% +n- ewZS O
/.1 2 t
p$ w%wtv
N9 6
S(
ZA ^ rZeSw(
{T {eV Aw%_VZ %( { $:9'yTIw8 { ?pw^8 6weSSZeSw
T ^%~6oe Y wZ
({
T{ e w
{N T 4"e
% +M- tw8 { _Sw(
\jT {8 @ SZAS$Ot ' YS{
(
N( 5 4 t/%ez
XT ({ ZZ
wtS {e%( $tS(
Z6
, d6eSw(
{t eV /t\ wA
ZK
(
N( 5 4 Ty ? y? %4 $ { &()+ y%\ZeSw6S^/ 6
(
N( 54 SZ
wZS
^A
6w/tI w6^
ZN ^%
%Aw^SSZtSw(
{t ^ SZ 8 wY" { 8 Ze
6_ e/
ZtSw(
{t ^z 8W! - w &%* {fSv' B /TwYN 9:4vw%w N
X
{
N T* 4 t/
'4 'NZ%es %AwYeS
(
N( 5 4 wZS
' ^
Z. ?\Zw%
/'Z\twijSwe ^Aw%ew%^Si t/
Sz/tn
S
N 8%ZZi^wvA%teAw6) w%6
' zT Z:w%t)zY z( 4p9Y {'ZN Y
Ot)S X
N( 5 4 SZASR! Z /tw - w6s ! 9 e%i9 *,+ tno{
zT {6 + O
Z ?\Z
/'Z4 y' zT Zie% r
X 9 Z6 ZwNz
y
ewZ8
V %6wYt we w
8
{6tw% :/
E$ %SZ%tT {eR (
'ZN^w%6Z6 Nw6e
E_$ %%y?
6%_$ ^4Yt, 66y_ St%Z% yY_ n! t
( ^ZNw + Xt6w * { w%
{
N T 4 SZ6S +
Z\[C ^N_`VabEc5dfehgi_YjKk\l
mnal
VELOCITY (km/s)
2 2.5 3 3.5 4 4.5
0
Primaries
Multiples
0.5
TIME (s)
1.5
2
{8PV^PAySE}5X%}WX MD@>
%F$}56(T/!5XX=XX-%M(}0 T$/v:mXX-%#1
P (YoS U| /T ! t6Vn
$y\? 6wu {w
N96)SeZ6V !w
yK%%ez
6Y6w^SSZtSw(
{t ^8K+ 'w6S{e
./ 12
Tw%6%$Ie% w%
r9AV)w8
Z0 w
K?7 /'A? * $?\%6wYt4''wZS ^! &A* SZ(\wz%ZZ
! &%* $%Yep 9n ?p SwZetZ%A 48 X6wZZ | t6S H N
Yt e% + y%:Z%A4
d8X6wZZSoe
^ H $%:% t ^o! * $A??(
6%$NZveN
{6/ wS(
XpwS(
Z
Kr!EDq >( G@;< * & ]
= T = H T I! *
(
X! 4 {
A{ 8
S iw%Aw^SSZtSw({
t ^pow w
8 +- eS /tw
4:
$ ' wZS ^! &A* ZYz%e?w%: w4t6
F F H r8
8
YT ^% + d \Ze9 i
Xz4 wp? SwZt
OM'wZS ^ ! * Z^
Z8 ZN e/ ZBe
X4
F (
(6_wZOt vw Ze%6w6%ZuYtwu9ZY%S
Kr!ED q >(TG@; < *3 = T (
= T I ! }*
ww({
t
(
X24 z
4
N96zve6t ^
Y F F
Zg] ^N_`VabEc5dfehgi_YjKk\l
mnal
38PG ?
C v %pQEA / 3 ]6( D0 )I|TX% Kg!&%zT wq %T TX%N 3 qol
T 5%T5X%*LA1|OQ
:mXXw$}A7HX MDY
*L$}56(Tp(}N7:eE}(XXw:y
w'* 5$1
ab5a: c\a:Cl
l}abtk\,cCm\ab5c,k\mSk\bEdc5d ZRw
ZN ^/
(
N( 5 4 e%
X'4 + y%V^
X4 ^Yu9' Z ZZ jt 8 jwtS {e%t/^8
e/' % {S $
Y% SY9
4 S{
66
(
X$ 4 '(
6 6w^SSZtSw({
t ^0 vw S(
ZA ^ rZeSw(
{T {eo
(
N( T 4,t/
4' /tzeZ e, 4Oe
: w( 6_Ot ewe/'0d e6ZZS" ( t4:$K98 w Zj zT {8 w%jt,
p
(
N( 5 4w/t6 Z %t 9
w
e%Zp ( eV4 w%: w( ZN. S{ ^ YeN + w%ti ?Ze
ZN ^/
6w^SS6eSw(
{t e +U Nw6wZj
i 9%S6 et es wZjw{ w/tfwASwew/tIjt Z
jweS ^t/e% % SB NSw_8 Zr 8 ,
Z
jt,
Y
(
N( 5 40{w
8 wS %ZZ%ZNI <w%6wZZN ^
SY {vw%S(
Z6
, d6eSw(
{t e0 6w%_ $(Ot 6wZ6_ ^Z
V
VV o/ w 9 '7 2 %w &%2s+ y:Z r
(
N( T 4SZ
{4 (
/8 8N zw%48 6weSSZeSw
T ^%S%SZs s
S OtB 9ew~S(
ZA ^ rZeSw(
{T {ei
{N T 4~e%
X4' e%
N( 5 4t/
4' ^8
X4)w%eS?ZtSw(
{t ^?7 wvw%fw(
{t
4 Ze
ZN ^/
(
N( 5 4e%
X4' Y v'6w6 B %Y
{t e(
X49:4
8ewZS
A ??6% $^%Ze 6wYeS
{N T 4wZS
8 6weSSZeSw
T ^%? SeZ\% w3 ?7 wVw
T
7
T ^:8 X6
^ oeitr ?7 wu 9ew S
{TwYe/ Nw6 o6 z(
ZNw( $ ZN {
CA
Z +
%
%8 t6^N B /TwY ?7 w" N
w /66 ^%( $ y 8
( ZNt ^ S(
ZA ^ rZeSw(
{T {e>
e%Z ~ S t% S $t/jT {) i jweS ^ + 'ZS6 9YV6wpZN
f^
X4V4
w _ Yt e Ze8
ZgMZ N^ _`VabEc5dfehgi_YjKk\l
mnal
6_ e/
{N T4e/
X4 e%w
%e (
ZM w %6wZeS
0
{N Tv4 wZS
etZ T ^/_
Z^jfw/T 60
%t98
vww/T: Ze^ ZN{ e/
{
N T 4 t/
'4 +
^ <W U Cx U V
y'w%ew6SZ({
X4#w%e823S z U tgNw(
6oOeu w
v%/ %eS: 3I
ewN /?(
( K ee8/Twj%8 Z:t
/?
{t^' z AN
< {1p %6 +
'w:] @)
X%}p 1g1p(}=wTT-WpN{1pN *p^{TTX%JX MD
$&55c5X* t5}]}-iQ{ $*<D^5X-p{yrPp
q6Zq*1
=f\T -p 1pC(} "^} - pg 1g1p ZpgfgX M Du(} D*T
: N%0 X} T5 $DWiW{ $<D^5X p&r>Sp*q^ q6q^ wq*1
{X# }X* "D<p 1pW(}:* pAz1p8(}G B^T 6(A p&f'1p o*pSz%Z
5 .
$: Y(}G %T $%5 TX%A iR{ $D^5X -pgy|Sp
oY6k*1
/)Dp{BW1?/1p3( B^ *5%A pzBW1 1pp k*pzz) $$}TA(# 5
Xw q )I|5X%c )AN* 5X%(} X M Du
:TX%A iW{ $D*TX -pWy&yAp% koAj kY*1
wX XX p&L1|C1p )F)DpWC1|h1p @( )DpW1Whp/ ZpA{$T '*
Xz X M D=TT 65% -i|{ $*<D^5X -pr>Sp Y^ Sj Y*1
X* p*h1*BW1p(}= O( - p*EG1<Ow1p' e p&B< (} w(N(
}) D
%' 5X :(} M:5* iF{ $D^5X -p
e&Sp^6ok*1
B }^ < & - pWEG1p% Y*p&{$5 F m T5(# "J
v%Q 5
$/$$} 5X%> i&{ $D*TX pWPW p*^ Sk^ 1
B^* s MA pW1e 1p
(V ^ pNx1&w1p% q*p'%A5X< } {X M D
TT 65X%u(Z X J $}T TX iF{ $D^5X -p
eWAp|Z6s^1
O( pAEG1&OL1p
(N w< - p&}F1p Y*p&fCX M Dm$&55* i %
5 $ 5 %TX%K(n $$*XX,TX% C
X M Dw
!}}S
TX% -pe&rPpo6^ 1
O X} pC1Cg1p *pgfCX M D(: D*T { } J $X# "X* iF{Z
$D*TX -p&yrPp% ^ Sg *1
CWP-448
ABSTRACT
One-way wave operators are powerful tools for forward modeling and inversion.
Their implementation, however, involves introduction of the square-root of an
operator as a pseudo-dierential operator. Furthermore, a simple factoring of
the wave operator produces one-way wave equations that yield the same travel-
times of the full wave equation, but do not yield accurate amplitudes except for
homogeneous media and for almost all points in heterogeneous media. Here, we
present augmented one-way wave equations. We show that these equations yield
solutions for which the leading order asymptotic amplitude as well as the trav-
eltime agree with satisfy the same dierential equations as do the corresponding
functions for the full wave equation. Exact representations of the square-root
operator appearing in these dierential equations are elusive, except in cases in
which the heterogeneity of the medium is independent of the transverse-spatial
variables. Here, singling out depth as the preferred direction of propagation, we
introduce a representation of the square-root operator as an integral in which
a rational function of the transverse Laplacian appears in the integrand. This
allows us to carry out explicit asymptotic analysis of the resulting one-way wave
equations. To do this, we introduce an auxiliary function that satises a lower
dimensional wave equation in transverse-spatial variables only. We prove that
ray theory for these one-way wave equations leads to one-way eikonal equations
and the correct leading order transport equation for the full wave equation. We
then introduce appropriate boundary conditions at z = 0 to generate waves at
depth whose quotient leads to a reector map and estimate of the ray-theoretical
reection coecient on the reector. Thus, these true amplitude one-way wave
equations lead to a true amplitude wave equation migration (WEM) method.
In fact, we prove that applying the WEM imaging condition to these newly
dened waveelds in heterogeneous media leads to the Kirchho inversion for-
mula for common-shot data. This extension enhances the original WEM. The
objective of that technique was a reector map, only. The underlying theory
did not address amplitude issues. Computer output using numerically gener-
ated data conrms the accuracy of this inversion method. However, there are
practical limitations. The observed data must be a solution of the wave equa-
tion. Therefore, the data over the entire survey area must be collected from a
single common-shot experiment. Multi-experiment data, such as common-oset
data, cannot be used with this method as presently formulated. Research on
extending the method is ongoing at this time.
1 INTRODUCTION us to separate solutions of the wave equation into
downgoing and upgoing waves except in the limit of
One-way wave equations provide fast tools for mod-
near-horizontal propagation. The original one-way wave
eling and migration. These one-way equations allow
equations used for wave equation migration (WEM)
2 Y. Zhang, G. Zhang & N. Bleistein
[Claerbout, 1971, 1985] were designed to produce ac- the peak amplitude on the reector being in known pro-
curate traveltimes, but were never intended to produce portion to an angularly dependent reection coecient
accurate amplitudes, even at the level of leading order at a specular reection angle.
asymptotic WKBJ or ray-theoretic amplitudes. As such, This type of inversion requires common-shot data
that WEMprovides a reector map consistent with the with the receiver array covering the entire domain of
background propagation model, but with unreliable am- the survey. This is a serious obstacle for practical im-
plitude information. plementation; such data gathers are still relatively rare.
The objective of this paper is to describe a modi- To date, we do not have an extension of this true am-
cation of those one-way wave equations to produce equa- plitude WEM to other source/receiver congurations.
tions that provide accurate leading order WKBJ or ray- In the next section, we provide motivation for the
theoretic amplitude as well as accurate traveltime. The modication of the simple one-way wave equations with
necessary modication of the basic one-way wave equa- the objective of developing true amplitude one-way
tions can be motivated by considering depth-dependent equations for forward modeling. We start from homo-
(v(z)) medium. In this case, through the use of Fourier geneous media where the basic one-way equations do
transform in time and tranverse spatial coordinates provide true amplitude. We then proceed to analysis of
(x, y), we reduce the problem of modifying the one-way modeling for v(z) media. That leads to a modication
equations to the study of ordinary dierential equations. of the basic one-way equations in order to assure that
There, it is relatively simple to see how to modify the the appropriate transport equation is satised, as well.
equations used by Claerbout in order to obtain equa- Following that, we introduce the idea of using
tions that provide leading order WKBJ amplitudes, as the same equations for heterogeneous mediav =
well. This leading order amplitude is what we mean by v(x, y, z)with functions of transverse wave vectors
true amplitude for forward modeling. now reinterpreted as pseudo-dierential operators. It is
For heterogeneous media, v = v(x, y, z), the same for this new interpretation that we provide a conrma-
one-way wave equations still provide true amplitudes. tion that these one-way wave equations provide true am-
However, now the transverse wave vector (kx , ky ) must plitude forward modeling. The proof of the claim that
be interpreted as dierentiations in the corresponding this extension leads to the appropriate transport equa-
dual spatial variables. Further, our modied one-way tion is provided in Appendix A. That extension and the
wave equations involve square-roots and divisions by proof were originally developed by the second author in
functions of this transverse wave vector. We provide Zhang [1993]. Here, we present an update of that proof
an interpretation of these operators through some basic with attention to the application to WEM.
ideas from the theory of pseudo-dierential operators. Following the discussion of forward modeling, we
We provide a relatively simple representation of the develop true-amplitude WEM. This requires a modi-
one-way dierential operators. This, in turn allows us to cation of the basic one-way wave equations of Claer-
prove that the ray-theoretic solutions of these equations bouts WEM and also a modication of the boundary
satisfy the separate eikonal equations for downgoing (in- conditions of that WEM, which corrects the phase as
creasing z) and upgoing waves, but the leading order well as the amplitude of the downgoing wave used in
amplitudes also satisfy the same equationthe trans- WEM. We also show how to modify Claerbouts orig-
port equationas does the leading order amplitude for inal WEM equations in order to turn those into true
the full wave equation. It is in this sense that describe amplitude equations. There is a subtlety of scaling by
the solutions of these one-way wave equations as true a pseudo-dierential operator in the comparison. This
amplitude solutions. leads to a slightly dierent one-way wave equation for
Having these true amplitude one-way equations the downgoing waves in Claerbouts approach when
allows us to develop a true amplitude WEM for compared to the one-way equation that we use in the
heterogenous media. To date, we only have numerical new theory.
checks on this method for v(z) media, where the pseudo- We then provide a proof that this new approach
dierential operators revert to simple multiplications in leads to the same common-shot Kirchho inversion for-
the temporal/transverse-spatial Fourier domain. How- mula as is found in Bleistein [1987] and Bleistein et al.
ever, we are able to prove that the reection amplitude [2001], as expressed by Hanitzsch [1997].
agrees with the amplitudes generated by Kirchho in- Following that, we present our numerical check of
version (true amplitude Kirchho migration) as devel- true amplitude WEM. As noted above, the example we
oped by one of the authors [Bleistein, 1987, Bleistein et present is for the case of a v(z) medium where imple-
al. 2001] and colleagues. This proof is valid in hetero- mentation of the pseudo-dierential operators appear-
geneous media. Thus, at this time, the proof of validity ing in our method reduces to a multiplication in the
is ahead of the computer implementation in terms of temporal/transverse-spatial domain.
generality and it anticipates a reliable computer imple-
mentation in general heterogeneous media. It conrms
that the output of this method is a reector map with
True amplitude wave equation migration 3
2 MOTIVATION In particular, sign(/z) = 1 or sign(/z) = 1
indicates waves in the direction of increasing z; that
In this section we provide motivation for modifying the
is, downgoing. Of course, then, for upgoing waves
standard one-way wave equations that are used in WEM.
sign(/z) = 1 or sign(/z) = 1. We could
We do this by starting with the standard operator fac-
alternatively represent upgoing waves by
toring scheme for the wave equation in homogeneous
media, separating o the z-dependence so that we can W = A exp{i[t + (x, y, z, )]}
identify upgoing and downgoing waves. We identify the or
separate waves and conrm that they are solutions of W = A exp{i[t + (x, y, z)]}
rst order wave equations obtained by factoring the op- or
erator. We then show that the solutions of the derived W = A exp{i(x, y, z)},
one-way wave equations are no longer solutions of the
with sign(/z) = 1 or sign(/z) = 1. Both alter-
full (two-way) wave equation when the medium is al-
natives for representing upgoing waves are used in this
lowed to depend on z; that is, v = v(z). In a WKBJ
paper and in the literature.
solution in a v(z) medium, the amplitudes of the rst
For constant wavespeed, we can rewrite this equa-
order equations do not agree with the amplitudes of
tion in the frequency/wave-vector domain in the form
the two one-way solutions of the full wave equation. By
modifying the one-way wave equations, we obtain new 2W
LW = + kz2 W
equations whose eikonal and transport equations agree z 2
(4)
with the eikonal and transport equations for the full
wave equation; each one-way equation governing one- = ikz ikz W = 0.
way propagating waves of the two-way or full wave equa- z z
tion. Here,
We begin by introducing the wave equation,
2
2 2 v k
1 W
2 W = 0. (1) kz = sign() k2 = 1 , (5)
v 2 t2 v2 v 2
We will use the following denition of the Fourier trans- and k is the transverse wave vector,
form:
k = (kx , ky ), k2 = kx2 + ky2 . (6)
1
F (x, y, z, t) = dkx dky d F (kx , ky , z, )
(2)3 Further, the solutions of the full second order wave
(2)
equation are actually solutions of the two one-way wave
exp{i[t kx x ky y]}. equations
In this equation, the wave number integration ranges
over all space. The frequency domain integral has the
range < Re < , with Im large enough that ikz A exp{ikz z} = 0, (7)
z
the contour of integration passes below all singulari-
with the upper signs yield a downgoing solution and the
ties of the integrand in the complex -plane. Typically,
lower signs yield an upgoing solution.
this is just below the real axis in , with the integrand
We would like one-way wave equations for the het-
dened on the axis only through analytic continuation
erogeneous case, as well, similarly separating upward
from below. Consequently, when viewed as an integral
and downward propagating waves. For this generaliza-
on the real axis from to , we must interpret
tion, we will content ourselves with ray-theoretic solu-
multi-valued functions, such as square-roots, as if we
tions that yield the same leading order amplitude as
had passed under their singular pointsbranch points
does the two-way wave equation, using the leftmost ex-
and poles to go from one side of one of these points
pression in (4).
to the other.
First consider the case where v = v(z) only and
Consistent with the Fourier transform in (2), when
recast the leftmost dierential operator in (4) in a form
we write down WKBJ solutions, they will take the form,
that lends itself to ray theory analysis. To this end, we
W = A exp{i[t (x, y, z, )]} introduce the slowness vector p by setting
or k kz 1
p = , pz = = 1 (v(z)p)2 (8)
W = A exp{i[t (x, y, z)]} (3) v(z)
or and rewrite the wave equation in (4) as
W = A exp{i(x, y, z)},
2W
+ 2 p2z W = 0. (9)
depending on the context. In these forms, or z 2
points in the direction of propagation of the wavefronts. When we substitute the standard form
4 Y. Zhang, G. Zhang & N. Bleistein
W = A(z, p)ei(z,p) (10) 1 dv(z)
ipz 3 W = 0. (15)
z 2v (z)p2z dz
into (9), we nd
2
For these equations
2 d
p2z A
dz 1 dv(z)
ipz 3 A ei =
(11) z 2v (z)p2z dz
d dA d2
i 2 + A + O(1) ei = 0. d
dz dz dz 2 i pz A ei
dz
Here, O(1) is in order of powers of i.
This last equation leads to the familiar eikonal and dA 1 dv(z) i
+ 3 e = 0,
transport equations for the rays, except that we are in a dz 2v (z)p2z dz
Fourier transform domain in transverse slowness vector,
for which the transport equation in (14) is replaced by
(px , py ). Those equations are
2 dA
3
1 dv(z)
A = 0 (16)
d d
= p2z = = pz , dz 2v (z)p2z dz
dz dz
while the eikonal equation remains unchanged. We have
d dA d2 already seen that this is equivalent to the transport
and 2 + 2A=0
dz dz dz equation for the full wave equation. Hence, the one-way
wave operators in (15) will produce the upgoing and
dA 1 dv(z) downgoing traveltimes and leading order amplitudes of
or 2pz 3 A = 0. (12)
dz v (z)pz dz the full wave equation. In fact, these solutions are exact
for the case of a medium that depends only on z; that
dA 1 dv(z) is, when v = v(z).
or 3 A = 0.
dz 2v (z)p2z dz The question now arises as to how this insight
Here, we think of the upper sign solution as the one for can be extended to factoring the original wave oper-
which sign(pz ) = 1. In this case, the upper sign corre- ator in (1) of a fully heterogeneous medium, where
sponds to the downgoing wave and the lower sign cor- v = v(x, y, z). In this case, we could not have derived
responds to the upgoing wave. Note that the transport dierential equations in z alone through Fourier trans-
equation is the same for the two waves because only p2z form. If we disregard this obstacle, we could still exam-
appears in that equation. ine the one-way wave equations in (15) with the hope
Now consider the two one-way wave equations in of achieving a sensible interpretation. To that end, let
(7) and corresponding WKBJ or ray-theoretic solutions. us rst recast the one-way wave equations (15) in the
That is, consider original Fourier variables. Thus, we rst rewrite that
equation as
ipz A ei
z 2 v(x, y, z)
ikz 3 W = 0. (17)
d z 2v (x, y, z)kz2 z
= i pz A ei (13) For reasons that will become clear in the next section,
dz
we prefer writing the multiplier on this additional term
dA i as follows.
e +
= 0,
dz 2 v(x, y, z)
with consequent eikonal and transport equations, 2v 3 (x, y, z)kz2 z
d dA
= pz and = 0. (14) 1 v(x, y, z) (v(x, y, z)k)2
dz dz = 1+ 2
2v(x, y, z) z (v(x, y, z)k)2
While the eikonal equations in (12) and (14) agree,
the transport equations for amplitudes do not. Thus, we and then rewrite (17) as
must consider modifying the one-way wave equations in
1 v(x, y, z)
(13) if we are to make the latter transport equation ikz
agree with the former, while keeping the same eikonal z 2v(x, y, z) z
equation in both. We can achieve this goal for the one-
(18)
way wave equations if we modify them by adding a term (v(x, y, z)k)2
1+ 2 W = 0.
to the one-way operators appearing in the leftmost ex- (v(x, y, z)k)2
pression in (13). The key to doing this comes from ex- Let us now think of as a place-holder for the tem-
amining the last form of the trnasport equation in (12). poral derivative and k = (kx , ky ) as a place-holder
That is, we consider the new one-way wave equations for a transverse gradient operator; that is,
True amplitude wave equation migration 5
i /t and i(kx , ky ) (/x, /y). (vT x )2 = , z)T x )2
(v(
Then we could easily give meaning to the expression 2 2
(v(x, y, z)k)2 as follows. = v(
, z) + v(
, z)
x y
2
(v(x, y, z)k) v(/x, /y) (v(/x, /y)).
2 2
However, symbolically, kz involves taking the square- = v 2 (
, z) 2
+ (21)
root of a dierential operator, while the division in the x y 2
last term requires that we give meaning to the recip-
rocal of a dierential operator. Interpretation of such , z)
v( v(
, z)
+ v(
, z) +
expressions is what the theory of pseudo-dierential op- x x y y
erators is all about. Thus, in the next section we address
the interpretation of these terms in fully heterogeneous and
= (x, y).
media where v = v(x, y, z). With an appropriate inter-
However, in the current application, in the dienition
pretation, it turns out that these modied one-way wave
of , we need this operator to be valid to two orders
equations provide an asymptotic solution for fully het-
in , rather than the usual leading order. The reason is
erogenous mediav = v(x, y, z)for which the result-
that this operator is O(). Thus, as in the v(z) case, it
ing transport equations agree with the transport equa-
aects the eikonal equation to leading order, however,
tion for the full two-way wave equation.
its correction term, O(1) in is of the same order as
the other contributions to the transport equation. On
the other hand, is already O(1) in at leading order.
3 TRUE AMPLITUDE ONE-WAY WAVE Thus, the subtleties of lower order corrections to this
PROPAGATION term will only aect lower order amplitude corrections
Motivated by the discussion of the previous section, we that are not of interest to us here; we are only concerned
introduce the same one-way equations (18) for the het- with getting the leading order amplitude right.
Returning to the discussion of , we must distin-
erogeneous medium in which v = v(x, y, z). Here, we 2
will extend the denition of dierentiation through the guish between v k and (kv)2 for the operator (T x v)2
use of pseudo-dierential operators so that we can give dened by
meaning to the pseudo-dierential operator kz in (17)
(T x v)2 = (vT x )2 , (22)
and thereby give meaning to those one-wave equations
themselves. given in more detail by
Couched in the language of pseodo-dierential op-
(T x v)2 = , z))2
(T x v(
erator theory, those one-way wave equations are
2 2
L W = W W = 0 (19) = v(
, z) + v(
, z)
z x y
Here, and are pseudo-dierential operators with
symbols and , respectively: 2 2 2
= v (
, z)
x2
+
y 2
i (v k)2
= ikz = 1 ,
v 2 v(
, z) v(
, z)
(23)
2 +3v(
, z)
x x
+
y y
1 kz 1 v v k
= = 1+ 2
2kz z 2v z 2 v k 2 v(
, z) 2 v(
, z)
+v(
, z) 2
+
(20) x y 2
2
vz v k
2
2
= 1+ 2 , k (kx , ky ). v(
, z) v(
, z)
2v 2 v k + + .
x y
In this last expression, the subscripted variable vz con- Thus, we need to dene the symbols associated with
notes derivative respect to z, in contrast to its use in these operators more carefully by
kx , ky , kz where it distinguishes the three components 2
of the wave vector. v k v 2 k2 + v{kv} k,
The symbol notation is somewhat inadequate
here because it is really designed for leading order (kv)2 v 2 k2 + 3v{kv} k + v{k2 v} + {kv} {kv}.
accuracypropagation of discontinuitiesonly. The
2 In these equations, we use the curly brackets {} to indi-
notation v k is the symbol for the operator cate that the symbol only operates on the function con-
6 Y. Zhang, G. Zhang & N. Bleistein
tained within them. We think of the operator k applied 2
LT (s;
, z, t)q = s2 (vT x )2 q(s;
, z, t)
to W as contain terms that are O(), or O(1)k while k2 t2
(27)
applied to W as containing terms that are O( 2 ), O(),
2
O(1). We need to keep terms of the two leading orders = (vT x ) W (
, z; t) z > 0, t > 0.
in these operators when dealing with . This issue will
arise in the proof contained in Appendix A. Using this function plus the identity (24) for kz , W
The symbol has an exact representation [Zhang, and W can be expressed as
1
as .
x = (x, y, z), xs = (xs , ys , 0),
Also, we modify the imaging condition (33) to be
and the quotient of the wave elds pD and pU :
U = 0, pU (
x; )
z R(
x) = d. (36)
(32) pD (
x; )
xs ; ) = Q(x, y; ),
U ( See Zhang et al. [2001, 2002].
It is important to recognize that the boundary data
where D is the downgoing (source) waveeld and U is for pD in (34) involves a pseudo-dierential operator.
the upgoing (observed) waveeld. The image is then pro- Thus, it might be easier to solve a modied version of
duced as an impedance or reectivity function at every (31) for D, after adjusting that equation to be equivalent
image point dened by to the true amplitude equation (34) for pD .
U (
x; ) Let us, therefore, introduce a new D in (34),
R(x, y, z) = d. (33)
D(x; ) 1 1
D = 2pD , pD =
D, (37)
The key to this imaging method is that the construc- 2
tive/destructive interference between the phases of the for which the boundary data agrees with the data for
two waves produces a large amplitude where the re- D in (31). Now, let us substitute this choice into the
ectors reside and a small amplitude where they do dierential equation (34) for pD :
not. While this result produces a reector map, it does
D
not provide accurate amplitude information. To achieve + 1 D = 1 2 D
z z z
that, we use the solutions of our modied true ampli-
tude one-way wave equations (19). That is, we introduce +D 1 D = 0, (38)
pD and pU as solutions of the following problems.
D
+ D 1 + 1 D = 0.
+ pD (
x; ) = 0, z z
z
(34) We have omitted the constant factor of 2 throughout
1
xs ; ) = 1 (
pD ( x
xs ), these manipulation. For the last two terms, the lead-
2
ing order contributions contribute to the leading order
and amplitude and the corrections aect only lower order
pU (
x; ) = 0,
amplitudes. Therefore we can set
z (35) 1
pU (
xs ; ) = Q(x, y; ).
1 kz 1 kz
Here, we have not only modied the governing one-way 1 + = .
z kz z 2kz z
wave equations in accordance with the discussion of the
previous sections, but we have also modied the bound- Consequently, the true amplitude equation for D is
ary term for the downgoing wave from the source. The
+ + D = 0. (39)
reason for this is that the boundary value only accounts z
for the impulsive nature of the source in the transverse Within the factor of 2, this use of D agrees with the
direction. In the z-direction, we must account for an earlier papers cited in the references. Thus, we can
impulsive source by balancing the terms, avoid apply a pseudo-dierential operator to the two-
dimensional Dirac delta function in the boundary con-
dition for pD in (34) by solving for D, instead, using
8 Y. Zhang, G. Zhang & N. Bleistein
We start by considering the problem for pD as de-
120 140 160 180 200
ned by (34). Relying on the proof of Appendix A, we
0 know that this function is dynamically equivalent to the
downgoing (outward radiating) Greens function of the
full wave equation. Therefore,
pD (
x, xs ; ) = A( xs )ei(
x,
xs ) .
x, (40)
Here, is the solution of the full eikonal equation for
2 the full wave equation (1) with /z > 0 and A is
the solution of the transport equation for the full wave
equation. Equivalently, is a solution of the eikonal
equation
d 1
= p2x p2y ,
dz v 2 (
x)
4
Single shot record deduced from (18) with upper sign, and A is a solution
of the transport equation
2 A + A = 0,
Figure 1. Input model data.
deduced from (18) with upper sign. This is the essential
conclusion of the proof of Appendix A.
this equation and the boundary condition in (31). Of To derive a representation of the function pU , we
course, we could tie U and pU together in exactly the have to work a little harder. Again, however, we rely
same way. That is, on the equality between the leading order asymptotic
solutions of the one-way wave equation, (35), and the
U = 2pU , full wave equation. In Appendix C, we show that the
leading to the dierential equation Greens function representation for pU is given by
cos r
pU (
x; ) = 2i A( x)ei(
xr ,
x) dxr dyr ,
xr ,
+ U = 0. v(
xr )
z (41)
However, now the pseudo-dierential operator would xr = (xr , yr , 0).
appear in the boundary data for U ; that is,
In this equation, again, and A are the phase and am-
xs ; ) = 2Q(x, y; ).
U ( plitude of the free space Greens function for the full
wave equation. However, because pU is an upward (in-
Furthermore, the correct imaging condition now be-
coming) wave, we need the inward propagating Greens
comes
function. Hence the sign in the phase is opposite the sign
pU 1 U of the phase of pD dened by (40). Further, vr = v( xr )
R(
x) = d = d.
pD 1 D and r is the emergence angle of the ray with respect
to the normal.
We use this last result and (40) in equation (36) for
5 COMPARISON OF TRUE AMPLITUDE R and obtain
WEM OUTPUT AND KIRCHHOFF cos r A(xr , x)
INVERSION OUTPUT R(x) = 2 i
vr A( x, xs )
(42)
The previous section proposed the use of new one-way
propagators for the surface data, a modication of the ei{(
xr ,
x)+(
x,
xs )} dxr dyr d.
data for the downgoing wave and a new imaging condi- This is the result in Bleistein[1987] and Bleistein et al.
tion in equation (36) to achieve true amplitude WEM. [2001] as expressed by Hanitzsch [1997].
Here we show that the integration in this imaging condi- This representation of the imaging condition for
tion agrees with the integration for Kirchho inversion true amplitude WEM, being the same as the inver-
in Bleistein [1987] and Bleistein et al. [2001]. Actually, sion formula for Kirchho inversion, conrms the claim
we derive the representation of the Kirchho inversion of this paper, namely, that this formulation provides
formula as stated in Hanitzsch [1997]. In carrying out a true amplitude WEM. We can now be assured that
this comparison, we rely on the proof of Appendix A of the output of this new WEM will provide a reectiv-
the dynamic as well as the kinematic equivalence of the ity map with peak amplitude on the reector being a
solutions of the one-way wave equations and the solu- known multiple of the geometrical optics reection coef-
tions of the two-way wave equation. cient. The incidence angle in that reection coecient
True amplitude wave equation migration 9
is the angle dened by the specular pair of rays from a operator theory. We have provided proofs that these new
source/receiver pair. Note that this result is conrmed one-way wave equations provide solutions that agree dy-
for heterogeneous media v = v( x). Previous veri- namically, as well as kinematically, with the solutions
cations, as presented for example in Zhang et al. [2001, of the full wave equation. Further, we have proposed a
2002], were only for the case of depth dependent media new approach to WEM, transforming it into a true am-
v = v(z). plitude process, meaning that it produces an inversion
output that agrees asymptotically to Kirchho inver-
sion: it produces a reector map with peak amplitudes
6 NUMERICAL TEST on the reector in known proportion to the geometrical
optics reection coecient. We have provided a proof of
To show how true amplitude common-shot migration this claim. With the aid of a simple numerical example,
works, we apply it to a 2-D horizontal reector model we demonstrated that the migration method we pro-
in a medium with velocity v = 2000 + 0.3z. Recall from posed does calibrate common-shot migrations by cor-
the theory that in this case, the modeling and migration recting both their amplitude and phase behavior for an
can be carried out in the transverse spatial and temporal example in which the wave speed is depth-dependent
Fourier domains, with k being the simple transverse part v = v(z). The new method actually builds a bridge be-
of the wave vector. tween true amplitude common-shot Kirchho migration
The input data (Figure 1) is a single shot record and the migrations based on one-way waveeld extrap-
over four horizontal reectors from density contrast. olation.
Figure 2 left shows the migrated shot record us-
ing the conventional common-shot migration algorithm
(33). The peak amplitudes along the four migrated re-
ACKNOWLEDGMENTS
ectors are shown in Figure 2 right, normalized to the
geometrical optics reection coecient along the reec- G. Q. Zhangs research was partially supported by the
tor. This method has a phase error: note the multipli- Major State Basic Research Program of P. R. China
cation by i in on the right side in (34) as opposed (G1999032800). All three authors further acknowledge
to the lack of such a phase shifting factor on the right the support and approval for publication of Veritas
side of (31). The consequent phase error has been cor- DGC.
rected during the migration. However, the migrated am-
plitudes are poor, especially on the reector at depth
z = 1000m along which the reection angles vary over References
a wide range. (This method has incorrect angular de-
pendence when compared to true amplitude reectivity Bleistein, N., 1984, Mathematical Methods for Wave
or the geometrical optics reection coecient at each Phenomena: Academic Press, New York.
point.) The wide angle peak amplitudes decrease mono- Bleistein, N., 1987, On the imaging of reectors in the
tonically with increasing depth. The greatest error oc- earth. Geophysics 52, 931-942.
curs at wide angle, with the result along the shallowest Bleistein, N., J. K. Cohen and J. W. Stockwell, Jr.,
reector being the worst. However, the error is zero at 2001, Mathematics of Multidimensional Seismic Imag-
zero oset; in this limit, k = (0, 0) and cos r = 1. ing, Migration and Inversion: Springer-Verlag, New
Figure 3 left shows results of true amplitude York.
common-shot migration (36). The peak amplitudes Claerbout, J. F., 1971, Toward a unied theory of
along the reectors are shown in Figure 3 right. reector imaging: Geophysics, 36, 3, 467-481.
From this plot, we clearly see that the true ampli- Claerbout, J., 1985, Imaging the Earths Interior:
tude algorithm recovers the reectivity accurately, aside Blackwell Scientic Publications, Inc.
from the edge eects and small jitters caused by inter- Hanitzsch, C., 1997, Comparison of weights in
ference with wraparound artifacts. prestack amplitude-preserving Kirchho depth migra-
tion: Geophysics, 62, 1812-1816.
Zhang, G., 1993, System of coupled equations for up-
going and downgoing waves: Acta Math. Appl. Sinica,
7 CONCLUSIONS
16, 2, 251-263.
Common-shot migrations oer good potential of imag- Zhang, Y., Sun, J., Gray, S. H., Notfors, C., and
ing complex structures, but the conventional formula- Bleistein, N., 2001, Towards accurate amplitudes for
tions of such migrations produce incorrect migrated am- one-way waveeld extrapolation of 3D common-shot
plitudes. Here, we have described true-amplitude one- records: 71st Ann. Mtg., Soc. Expl. Geophys., Expanded
way wave equations that allow us to extend the stan- Abstracts.
dard method both for forward modeling and for wave Zhang, Y., Sun, J., Gray, S. H., Notfors, C., Bleistein,
equation migration. These modied one-way wave op- N., and Zhang, G., 2002, True amplitude migration us-
erators are developed with the aid of pseudo-dierential ing common-shot one-way waveeld extrapolation: 64th
10 Y. Zhang, G. Zhang & N. Bleistein
[p]
1.0
Amplitude
2
depth=1000m
depth=2000m
depth=3000m
0.5
depth=4000m
0.0
U/D imaging condition 51 61 71 81 91 101 111 121 131 141 151
Crossline
Figure 2. Left: nite dierence migration using (33) for imaging. Right: peak amplitudes along the four reectors. The wide
angle error decreases with depth of the reector.
[p]
1.0
Amplitude
depth=1000m
depth=2000m
depth=3000m
0.5
depth=4000m
0.0
Pu/Pd imaging condition 51 61 71 81 91 101 111 121 131 141 151
Crossline
Figure 3. Left :nite dierence migration using (36) for imaging. Right: peak amplitudes along the four reectors. The wide
angle error decreases with depth of the reector.
International Meeting of the EAGE, Expanded Ab- ned by (19) in two spatial dimensions (to make some
stracts. of the calculations simpler to follow). That equation is
repeated here as
Appendix A + D(x, y, z; t) + D = 0, (A-1)
z
In this appendix, we show that the one-way wave equa- Motivated by the source adjustment introduced in (34),
tions that we developed in this paper provided the same we introduce the downgoing wave of the inversion pro-
traveltime and amplitudes as does the full wave equa- cess through the scaling
tion (1). More specically, we will consider the one-way
pD = 1 D pD = D. (A-2)
wave equation for the downward propagating wave de-
True amplitude wave equation migration 11
Our objective is to prove that the eikonal equation F (x, y, z; t) ei(t(x,z)) Aj (x, z) j (A-7)
of (A-1) is the downgoing branch of the eikonal equation j=0
for the full wave equation (1) in heterogeneous media.
This equations will be specied below after we introduce A0 is simply written as A(F ) (F can be either D or qD ).
our notation for this asymptotic analysis. Clearly, we We simply denote pD by p, to be determined in the form
could dene a corresponding function pU and carry out
p(x, z; t) ei(t(x,z)) Aj+1 (j+1) . (A-8)
the same analysis for the upgoing wave. Below, we will
j=0
see that the eikonal equation for the downgoing wave
chooses the sign of the z derivative of the traveltime to By substituting these asymptotic expansions into
agree with the sign of , guaranteeing downward prop- (A-4), we nd that
agation. For pU , the sign of the z-derivative is opposite v 2 2x
to the sign of and that is the only dierent in the A(q(D))(s; )) = A(D) = e(s)A(D), (A-9)
1 s2 v 2 2x
analysis provided here.
We will then show that the leading order amplitude with
of pD satises the same transport equation as does the v 2 2x
amplitude of the full wave equation. Further, the sign e(s) = . (A-10)
1 s2 v 2 2x
of the z-derivative will be a common multiplier of all
terms of the derived transport equation, assuring that Now we substitute the asymptotic expansions into(A-5)
the same will be true for the solution pU . and (A-6) to obtain
We start from equation (29), specialized to down- A(D) = E(D)A(p), (A-11)
going waves in two spatial dimensions:
1
with
D 1 D 1 1
+ 1 s2 qD (s; x, z, t) ds E(D) = [1 v 2 2x ]1/2 . (A-12)
z v t v t 1 v
(A-3)
vz For e(s) in (A-10), we will also have need of
+ [D + qD (1; x, z, t)] = 0. e(s)/x. That result is
2v
Here, D = D(x, z, t) and qD (s; ) satises v 2 2x
e(s)x =
2 x 1 s2 v 2 2x
2
s2 v qD (s; x, y, z; t)
t2 x
1 1
(A-4) = 2 1
2 s x 1 s2 v 2 2x
= v D(x, y, z; t). (A-13)
x
1 s2 (v 2 2x )x
We use the denition of pD in (A-2) and the denition =
s2 (1 s2 v 2 2x )2
(28) for to write
1 1 1
1 (v 2 2x )x
pD 1 s2 qp (s; x, z, t)ds = .
v t v t (1 s2 v 2 2x )2
1 (A-5)
= D(x, z; t), A-2 Asymptotic solution of downgoing
where qp (s; ) satises one-way wave equation
2 By substituting the equation (A-8) into equations (A-3)
2
s2 v qp (s; x, z; t) and (A-4), we have
t2 x
(A-6)
1
2 i z A(D)
v
= v pD (x, y, z; t).
x
1
1
1 s2 A(qD (s; ))ds
A-1 High frequency asymptotic expansion v 1
(A-14)
We consider only the downgoing wave equation (A-1).
vz
For the upgoing wave equation, all results can be ob- + A(D)z + (A(D) + A(q(1; )))
tained by the same approach. 2v
We seek the solution of (A-1) in the form of follow-
1
ing asymptotic expansion + [ ] + = 0,
i
and
12 Y. Zhang, G. Zhang & N. Bleistein
2 (1 s2 v 2 2x )A(qD (s; )) v 2 2x A(D) and
1
1 s2 e(s)x ds
2
+ i s B(A(qD (s; ))) + B(A(D)) (A-15) I2 = 1 s2
1
1 s2 v 2 2x
+ [ ] + = 0, 1
(v 2 2x )x 1 v 2 2x ds
= 1 s2
where v 2 2x 1
(1 s2 v 2 2x )3
2 (A-21)
B(A) = 2v 2 x Ax + A v . (A-16) (v 2 2x )x
x = J3 (v 2 2x ) J2 (v 2 2x )
We will replace the integral operator on A(qD (s; )) v 2 2x
in (A-14) by the same type of operator on A(D), itself,
(v 2 2x )x
since the latter is independent of s. To do so, we inte- = 3 ,
grate (A-15) on the interval s (1, 1) with the weight 8(1 v 2 2x ) 2
1 s2 /(v(1 s2 v 2 2x )). Then, we add the result of with e(s) dened by (A-10).
that integration to (A-14) multiplied by i to obtain the Now we can proceed to simplify the various terms
equation for the asymptotic solution of the equation for in (A-15). According to the denition of B(A) in (A-16)
the downgoing one-way wave: and using (A-11), we have
2
2 1
z A(D) B(A(D)) = 2v 2 x A(D)x + A(D) v
v x
2
= 2v x (E(D)A(p))x
1
1
v 2 2x ds 2 (A-22)
1 s2 A(D) +E(D)A(p) v
v 1 s2 v 2 2x x
1
= E(D)B(A(p))
vz (A-17) +2v 2 x E(D)x A(p).
+i A(D)z + (A(D) + A(q(1; )))
2v
Similarly, using (A-2), we nd that
1
1 B(A(D)) + s2 B(A(qD (s; ))) B(A(qD (s; ))) = e(s)E(D)B(A(p))
+ 1 s2 ds
v 1
1 s2 v 2 2x (A-23)
+2v 2 x (e(s)E(D))x A(p).
+[ ] + = 0
So, for the order-i term in (A-15) we conclude that
The following integrals are needed
B(A(D)) + s2 B(A(qD ))
1
1 ds
Jn (b2 ) = 1 s2 . (A-18) = E(D)B(A(p))(1 + s2 e(s)) (A-24)
1
(1 b2 s2 )n
We have + 2v 2 x A(p)[(1 + s2 e(s))E(D)]x .
1
1 1 (1 b2 ) 2
Now we can carry out the integration in the order-i
J0 = , J1 (b2 ) = ,
2 b2 term in (A-17), using all of the results (A-19) - (A-22).
The result is
1 3(1 b2 ) + 1 1
1 s2 e(s)ds
I1 = 1 s2 1
1
1 s2 v 2 2x
(A-20) = E(D) B(A(p))J2 (v 2 2x ) + 2v 2 x A(p)I2
v
J2 (v 2 2x ) J1 (v 2 2x ),
= +2v 2 x E(D)x A(p)J2 (v 2 2x ) .
True amplitude wave equation migration 13
A-3 Eikonal equation and transport equation By setting the coecient of the order-i term in
for restored downgoing wave (A-17) we obtain the transport equation of the restored
downgoing wave A(p). We consider two parts in this
From the coecient of 2 in (A-17) we have
coecient. One part is an integral (A-25), restated and
1 expanded upon here:
z + 1
v 1
1
(A-26) I= 1 s2
1
v 1
1 v 2x
2
1 s2 ds A(D) = 0.
1
1 s2 v 2 2x B(A(D)) + s2 B(A(qD (s; )))
ds
We seek a nontrivial asymptotic solution, , so 1 s2 v 2 2x
A(D) = 0. Therefore we conclude that 2
=
E(D)
2v 2 x A(p)x + A(p) v
1 v x
z + 1
v
1
J2 (v 2 2x ) + 2v 2 x A(p)I2 (A-29)
1 v 2x
2
1 s2 ds
1
1 s2 v 2 2x
(A-27)
+2vx E(D)x A(p)J2 (v 2 2x )
1
= z + 1 v 2 2x J1 (v 2 2x )
v = 2vx E(D)A(p)x J2 (v 2 2x )
1 1/2 2
= z + 1 v 2 2x =0 E(D) 2
v + v J2 (v 2x ) 2
+ 2v x I2
v x
This is the eikonal equation of the downgoing one-way
wave equation (A-1). We can see this eikonal equation
+2vx E(D)x J2 (v 2 2x ) A(p).
(A-27) is just the one of two branches of eikonal equation
for the full wave equation (1) in 2D, namely, From (A-27), (A-12) and (A-21)
1
vx E(D)x vx
=
=
2
1 v x
2 2 1 v 2 2x (A-32)
! "
vx xx + vx /v 2
zx
.
1 v 2 2x
Here, we derive E(D)x from (A-6):
1 1
E(D)x = z + 1 v 2 2x
2 v x
! "
1 vx xx + vx /v 2
= zx
.
2 1 v 2 2x
By using these results, we can now simplify I in (A-29)
as follows.
14 Y. Zhang, G. Zhang & N. Bleistein
I = x A(p)x II = E(D)A(p)z
2 vz E(D)A(p)
v +E(D)z A(p) +
x x (v 2 2x )x 2v 1 v 2 2x
+ A(p) +
2v 2 4(1 v 2 2x )
= z A(p)z
! "
vx vx xx + vx /v 2
+
zx
+A(p) E(D)z +
vz E(D)
2 1 v 2 2x 1 v 2 2x 2v(1 v 2 2x )
1 = z A(p)z
= x A(p)x + A(p) (vvx x
2v 2 !
A(p) 1
vx x vx x v 2 2x 2 1 v 2 2x
+ +
v v(1 v 2 2x ) "
vz /v 2
vx zx vx x +
+
1 v 2 2x
1 v x
2 2 v(1 v 2 2x )
! " = z A(p)z
A(p) vx zx ! "
= x A(p)x + xx +
.
2 1 v 2 2x A(p) vx xz
+ zz
.
Another part in the coecient of term i in (A-17)
2 1 v 2 2x
is Consequently the coecient of term i in (A-17) is
vz
II = E(D)z + (A(D) + A(q(1; )). (A-34) I + II = x A(p)x
2v
From (A-5), (A-6), (A-27) and (A-2) we have ! "
A(p) vx xz
+ xx +
2 1 v 2 2x
+z A(p)z (A-36)
! "
A(p) vx xz
+ zz
2 1 v 2 2x
A(p)
= A(p) + .
2
This last expression, then, is the coecient of the order-
i term in (A-17) and, therefore, must be equal to zero.
This leads to the rst transport equation for the ampli-
tude of the Downgoing wave; that is,
2 A(pD ) + A(pD ) = 0. (A-37)
This is just the transport equation for the full wave
equation (1). Because is the downgoing traveltime,
True amplitude wave equation migration 15
1 + , after that rst circumnavigation is the same as
the integral before. Further, it is standard in complex
integration methodology to conrm that the integrals
on the circles of radius , shrink to zero as , 0. Thus,
calling this new contour of integration, C1 , we need only
introduce a factor of 1/2 to equate the integral on C1
to the original real integral on the interval (1, 1).
2
1
v k
I= 1 s2 2 ds. (B-2)
2 C1 2 s2 v k
We note that the integrand has two poles at s =
||/(v|k|) > 1. We propose to recast the integral as a
sum of residues at these poles plus an integral on a circle
of radius r > ||/(v|k|) Since we will now be concerned
Figure B-1. Contours of integration. C1 is the barbell con- with the region where |s| > 1, we prefer to rewrite the
tour. Deformation outward leads to the contour C2 .
integrand as
2 2
v k
v k
the solution will be the amplitude for the downgoing 1 s2 2 = i s2 1 2 (B-3)
wave. 2 s2 v k 2 s2 v k
A similar result can be derived for the upgoing one-
with
way wave equation, starting again from (29), but using
v k 1 I= D
pU (
x; )L G( x ; )
x,
s2 1 s and 2
2 s2 v k s2 (C-5)
to obtain G( x ; )L pU (
x, x; ) d3 x.
2
1
v k
s2 1 2 ds Here, D is a hemisphere of radius R centered at the
2i |s|=r 2 s2 v k origin, planar boundary at z = 0 and extending down
into the domain z > 0.
1 ds We use the wave equations in (C-1) and (C-3) to
2i |s|=r
s (B-6) replace the dierential operators in the integral in (C-5)
by the source terms and then conclude that
2
=
1
id = 1, x ; ).
I = pU ( (C-6)
2i 0
Now, we note that
s = rei , ds = irei d. pU ( x, x ; ) G(
x; )L G( x ; )L pU (
x, x; )
Combining this result with (B-5) in (B-4) we con-
clude that x; )2 G(
= pU ( x, x ; ) G( x ; )2 pU (
x, x; )
2 Next, we use Greens theorem [Bleistein, 1984] to recast
v k the integral in (C-5) as one over the boundary D of
I =1 1 . (B-7)
2 the domain D:
We only need to substitute this result into (24) to con- x, x ; )
G(
I= D
pU (
x; )
rm that identity, which is what we set out to do in this n
appendix. (C-7)
pU (
x; )
G( x ; )
x, dS.
n
Appendix C
Here, /n is the normal derivative in the outward di-
In this appendix, we derive the representation (41) for rection from D. This surface consists of two pieces.
pU . Because of the proof of Appendix A, we can use the First, there is the hemisphere of radius R on which
full wave equation to nd pU . That is, pU must satisfy /n = /r. The second part of D is the disk on the
the frequency domain equation plane at z = 0 and of radius R, centered at the origin.
On this disk, /n = /z. On the hemisphere, we
2
x; ) = 2 pU +
L pU ( pU = 0, write
v2 (C-1)
pU (x, y, 0; ) = Q(x, y; ).
Furthermore, pU is an upward propagating wave and
therefore it satises the inward propagating Sommerfeld
radiation conditions [Bleistein, 1984],
True amplitude wave equation migration 17
$ $
G( x ; )
x, pU (
x; ) G( x ; ) $$
x, $
= pU (
x; ) G( x ; )
x, 2i A( x )ei(
x,
x ) $
x, .(C-10)
n n z $ z z=0
z=0
G( x ; )
x, pU (
x; ) Now we use this result and (C-9) in (C-8) to conclude
= pU (
x; ) G( x ; )
x,
r r that
G( x ; )
x, i x ; )
pU (
= pU (
x; ) G( x ; )
x,
r v
$
$
= 2i Q(x, y; ) x, x )ei(
x,
x ) $
A( dxdy.
pU (
x; ) i z z=0
x, x ; )
G( pU . z=0
r v (C-11)
2
Note that we can write dS = R d in (C-7) for this To complete the story, we need to replace x by x and re-
hemisphere. Here, d is the dierential solid angle on x at z = 0 by xr . Furthermore, we use the fact that
place
the unit sphere. We need to evaluate the last expression /z is the z-component of the traveltime , which is
at r = R. We can pair up a multiplier of R from the a vector making the angle r with the vertical and hav-
dS multiplication with each of the factors in this last ing magnitude 1/v( xr ). The result of these substitutions
expression for . By using the Sommerfeld conditions is equation (41). This completes the verication of that
in (C-2) we conclude that equation.
$
G( x ; )
x, i $
RpU (
x; )R G( x ; ) $$
x, 0,
r v r=R
R
$
pU (
x; ) i $
RG( x ; )R
x, pU $$ 0,
r v r=R
R .
Since the domain of integration is bounded (by 2),
we conclude that the surface integral in (C-7) over the
hemisphere must approach zero as the radius of the
hemisphere approaches innity. Consequently, we are
left with the surface integral over the plane z = 0 in
that integral.
Here, we use the boundary data in (C-1) and (C-3)
plus the fact that /n = /z to conclude that
G( x ; )
x,
x ; )
pU ( = z=0
Q(x, y; ) dxdy. (C-8)
z
<>=@?BADCFEGCFEIHAKJMLKAKJ&ADNOQPSRUTVAXWYCZ[ADN!L9GN
\M]_^a`b]_cdfegch!ikjl]nmop]_^Sekq*]U^SilrsK]utvikcf`bq*]U^a`wexdzyw]{eft|op}k~xU~ur\Mek ekcig
eV(ope_ekMeud;^S]_~rywek g]U^pr\w
USkk
gvUrz|
B;B9
V_MF(Bzv+vg(F
p _MSSFb_S_2vSM|(UMvS;M
g(_FV
M_FSM(__FFSwlav_
Mvg_v
l
v_M
FgFFD
la_'zbU_Fbv
M
bFv
[F_pUv
SFFva_+S*4la( (
g2SM_u%!S(__Fv_(Fv2
g_(_
p_B*4_lSM_v_
-_S_(Fb'z_Y
__b
zUg_Sbl
vb__FFalU
DX pV_b
MMD'SFM_|S_F;F;_FM_FU
9z_'g_
v*nFlv*[(UFV_pgMF_F!_MaaSba_
%_FB
__b
wv
'_(_MBSFb_S_M!Mg_v*z_%FD;kSk
a_@
Mg(F9
_F+
Flz!ba
lS
g|_ba(wFMgp(_k
gMS_BpS_vSFF_S_(FbMSlS
_
l*MS%F(_;F
M_(Fp
z_gf(vlFMFp
*FVaSFvKS_MSFv_bgSlSg
nF
B_F'(__USv
z
BM49lF
(v4FvU(l'v_la(4z__FvU_S_p4k
_(_
p*nF
x(fMS__bgFbSpFDFbSMvv(MgMFS__Mg_bSF
_FUMaMS_BS*F'_vk(w_g__v;gM_v- _F
{la(
S(__MSMSFv[!gF(M
F_S_(Fbv_4(S_'_B(g
av;U_Mg_FlS_
Fgla_+
MS_FSbUM;S!_FS__SD_F
_vk(%SMFaMBFv*FvzS
'FFMFpa_M_p
M
(S_p'_FV(F
%Ma _lSM_v_bU_(kvvpbFF(
SM_vFvUl
(S_w
b4MaMF{D|(v
$!
pP6 ! ) v->($
l' MB &S!-(!5( Z> g? :@<AB
% '!#(!7
l' 3 ]!# 3 !"!"% * % _ v( f( *| surface of water
g'
! al( ,% ' !#( v ) !"% @ *|'
p !f*|3 ! water layer
@ 3
& n 3 !Eb) l' 6
depth
postrift
regional subsidence
! 9"# M!!%$ synrift
aY(4pl +) "!
! % = ) 2B F( Kp!# IS7]!# v'!(%2v*& 0214365)78:9<;F>=@?/A
BCB'DEB@FGAIHKJL'JB'MNB'L'ONP=@MQHKONR/A
B'?/S B@FA
HKJ=@?
J=UTKJf
V NO PAIONDKL%AIHKJWR
XKYZR
ONTKJ[DKP[J'\ V B'RIAI?
O F4A\K=@DKT]R/SZDZ?
O F4A^XKDKO AIR_
() 2&%l; p = ( / > a' )
n) !"%
8Z 3
ll ' 6>aY(
'&{ l; p 3 "3 ((
(z 3(; ( %@<; A 6}
) @ B Z lV !
6#HD)FVI\S s`
"= BD)S ( / <n( 7p- ! N l; p f7 K midpoint (km)
B \S %] " b Z<n( p l v f7 K
]
3.0 7.0 x10 4
depth (km)
III
ll
v ( /
flVf; (l
) !"!"%4 ) <n C"
2 D6 1.0
1000
aY(Y ' p p ! f# (l4 vf <= 3 @ l9 v3 U>
{a
( `! # vall(
$Xal> >
l' K3 3 v ; ) $&%+ 1500
500
depth (km)
500
depth (km)
1.0
1000
1.0
1000
1500
1500 a
2.0
2000
midpoint (km)
0214365)78;aB'bUbUB'Dp ONbc=@L'JLd=AIHKJ[?
R=F4A
J[?bUONL'?q=AIONB'D!rdJ[MNBCP[O AS x10 4
depth (km)
IV
1.0
1000 III
midpoint (km)
x10 4 1500
2.0
5.0
2.5 3.0 3.5
7.0
4.0 4.5 V
0.0 0
b
2.0
2000
depth (km)
1.0
1000
0214365)78 ;e =dmQON?
R/AikCb#B@FQAIHKJR/Aq=@P
k'JTETZJ V A
H!R
J[PAIONB'DE=FgAIJ[?
V I? J[RIA
=@PqkETKJ V AIHtbUONL'?
=A
ONB'D h2O AIH!AIHKJJR/A
ONbc=AIJ[T V =@?q=@bUJAIJ?IROND
YKMNBCPqkCRCezYmfRq=@bUJR
J[PA
ONB'DEh2O AIH"TZJMNONDZJ=AIJ[TtYZMNBlPqk]YB'XZDKTp
=@?IONJR_C2HZJo=@?
?IBh V B'ONDdAIRaA
BWA
HZJ{F =@XZM AaA
HK=A^R
J V =@?q=AIJ[R^YKMNBlP
klR>/I
2.0
2000
=@DKTx %_
midpoint (km)
0214365)78Q;aB'bUbUB'Dp ONbc=@L'JLd=AIHKJ[? R=F4A J[?bUONL'?q=AIONB'D!rdJ[MNBCP[O AS 0.0 2.0 4.0 6.0 8.0 x10 4
=@D=@M SlR
ONRxOND YKMNBCPqk XKR
ONDZL A
HKJ!Ah>B:JrdJ[DCAIREbc=@?
k'J[T YCS A
HKJ
1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5
0.0 0
2000
2.0
PSVM r x Xr
= JIS;<l
' T 3
^ +^ ^Jz@
'
PSRvM r r D;M r Xr^ ^^ ^-DV@
) UE M r +x Xr
^ J^ ^J v ^ ^ ^ ^I\
F (=
! a3 U >Bfl 4b A +(lf " v 9 0214365)7! 8 ;#"dAq=@Pqk'J[T RIJ[P[AIONB'D =F4AIJ? V ?IJ[RIA
=@Pqk TKJ V A
H bUONL'?q=AIONB'D
!(%f 3 S;64[D 7 +( 3 %- ' !=b +_ _9u! X h2O AIHAIHKJoJ[RIAIONbc=AIJT V =@?
=@bUJ[AIJ[?
R>OND=@MNM<KrdJoYKMNBCPqkCR_K^HKJo=@?
?
Bh2R
!#v ( <Y vKf (
(p pl(l (' 6 bc=@?Ik"AIHKJ]JrdJ[DdA
RUXKRIJ[T OND AIHKJrdJMNBCP[O AS =@D=@M SlR
ONR%FgB'?%YZMNBlPqk$%_
aY( ' !# "*|' !(% !# v@ $D [E
* !"lz > v -M<= 3 % ' !# ^HKJ>?IRIA^ezRIH=@MNMNBh{mbc=@?
k'J[T%?IJJ[P[AIB'?
ONRAIHKJ2YB@A/AIB'b B@F<A
HKJ2R
XZYp
m(M Xr r
^ \-D$^ ^J-S ( ) 4 ](9 ! a3 U v9 v (l' RIONTKJ[DKP[JXZDKO A\=@DZTtAIHKJxRIJ[PB'DZT ?
JJ[PA
B'?%ONRWAIHKJY<B@AIAIB'b B@F^A
HZJ
f 8z
-%&6 (" !Z - *| 'A 3
. l *|9 lf# V B'RIAI?
O F4A^XKDKO A&_ %2B@A
JAIH=AfAIHKJYKMNBCPqkY<B'XKDKT=@?IONJR2B@FQA
HZJordJMNBCP[O AS
! *|l @
p= ; Y # v& ]<n( la 3" J[MNTxTKBiDKB@A^FgB'MNMNBh AIHKJWL'J[B'MNB'L'ONP=@M bc=@?
k'J[?
R_
lZ
!fT
@ R US7 v' *|l& @ 2 4 " &Y EKfl
p ; *| l l ) BOC" + JIS;6.aY ; b (( 3
!a3 U>
s
5! ( )>!# 3 ! - )
Up) z C" 4 J6
aY(N _9u! U
l !(n 9X b 5 ! a; U ) N(*l'
$ 4g
]
J
midpoint (km)
borehole
3.0 x10 4
2.0 2.5 3.0 3.5
7.0
4.0 MVA
0.0 0
depth (km)
2.0
2000
4.0
4000
2.0
2000 4000
Flp& - ( ,% ,<ll2 (4 <4 Z!b '%a; T
( <
9
!a3 U ! pf_ # v& @<=; < ; b '
_9f! :@ M yAB M
'
( X Q^^^ S
lp@ )
2000
GJILKM 7x
V B JD'^ +
(7 *| ) ! *p' ! al( ,% @(@ v - B D I\G" -
3 S;6
4.0
4000
w zl ( 5 G IK '
* !"z (' P<V) ) =z(
3 p *|' ! al( ,%_
!(%f > f
A lf# !w *|l *
( ,<' *|l& z
Up) ' C" @ R7 :<l# p &
0214365)78);{J V A
H:RIJ[P[AIONB'D B@F2A
HZJxJ[RIAIONbc=AIJTtrdJ?/A
ONP=@M rdJ[MNBCP[O AS x v _V
U7 6MaY@ !#v ( lp- *| v) b !#! ) <n
J[MNT _^HKJ
TK=@R
HKJ[TWMNONDKJ[R)bc=@?
k^AIHKJaYKMNBCPqkoY<B'XKDKTK=@?
ONJ[R_^HKJGr@=@MNXKJ[R 3 v; l
Y 9v`P V M Xr x Xr
b = ! a; U ^ G ^ ^J@
)
ONDA
HKJMNJ[L'J[DKTE=@?IJoOND]byR_ PSR{M r r 7DM r r
^ ^ N^ ^D@
) EM Xr r +!
^ JIG N^ ^J-
^ "D N^ ^\-6 [
" !"!" 3 ' &%`(w 3 v3 U|) zfl pV C" Y G-'(Z
%
,<ll (5 (" 3+ v
!a3 USl p Y
midpoint (km)
!( V Z'
3z *p f" p( ) v& ' ' - ;* !O6
0.0
1.6 1.8
3.0
2.0 2.2
6.0
2.4 2.6
x10 4
z (" !' *|l& f { 9
! a; US
B
6
6# *|Z 4 v '
! C" @ R&S> @ !# p' !(%1 ). v _ !"! ) <
0.0 0
s
I
p'
0.2
depth (km)
2000
u9p& 3 v = 7 ! a3 U' (Z lf#
! *| + l' (fl p B
6 6"7
2.0
9
0.1
!a3 U NS! l' !(% (l ' vfl+ " &)T<=(3
V 0
' l$ &' f" v )
& 7!# 3 !>*|' ! al( ,% * b pH 64a (
(+ ($ p & N*|' ! al( :% p& (( ,% @
+
4.0
,9
4000
!a3 U
5 lfl[ !# 3 !"!(%$* % *|' ! al( :%&&<
=
!"! <) al' () ;6ZC( )9<
) @l 3 v3 U@( ;" 3 v
= Av vll
0214365)7
8 );{J V AIH R
J[PA
ONB'D,B@F2AIHKJcJ[RIAIONbc=A
J[T:=@DZJMNMNO V A
ONP[O AS V = f z) # ( (f A` *p ) !E*|' ! al( :%+ 45
?
=@bUJ[AIJ[?
_ v, ! a; U 9 G ILK M \\J^!" -
3 S;6( (l` D(
y9
B
lf# !7 *| K ! a3 U <4 vD f; *|) D# p &
Q=U7 a 3( lw p& (( ,%>v& < Y b
B :@ M yAB
M Q^^^w D\JG@+ [G ILK M
S;9<=
^!
\\J^
-g' 6 !( + 4
" <4 '%B " !#( lp& - 4 *|' ! a
( ,%>/( p ! lwv T vf (@ pA Y 3
E! [f * vll *v v 7f v a ;p !(%+(N( vf(@ p @lv- -(:%
fl{= GJILK 6= ]
!"!
( 5 3 ! 3 ; ' ! ( !# '%
B S;6 > vvf5!# '%a ln nll; B S;6
f l | | al ( '
a Y9 ; v ( l
@
) * ) !* '! (,% N
! 3U .< v p
3 ) 2 =
3 ; b
( lf n |l v ( l | ! _
Wx b ' p ll S v x Sy`
$ # ! *
! $,< 1
U )
P V M Xr Xr D G 7P R M r r
7DM r r
C " b & ,< R U U
! !"( V"3 v *|!a":% v !(%fvl;3 !"!(%V@*|l
EkM Xr T
x r _x
(
^ VJ N^ ^R ^ ^R N^ ^-D ^ "D'F N^ ^J
3 v;!l
v ) (f[( (lf! (
(.'
| p
(
^ ^ ^ ^J-6=C " C " _Q !b '%
lfl* p f ((
v * v&3 p**|T;,p
; !N^6"D6.aY z 3 " ! ( ,%' _ {m
(! *p f p b
'%
Z>
( lp& 3 !*' *|l& '6=a2p 3
MY< m f'
1l p(Y /v 5 v(l& T !(4 !# v <n( 7 ("
*p 1l(l) _9f! X g!
l3 ( &(
5!"U- ! " 4JIG> 4` p, .(T ! a; UM6(K 1l; vT bp
:(v
v Z)
!"p ( ( F; l& <Y '%(' !# ( !"!) < f' p- (
4 (" ( =
!(]!# v(4 ! f
l' f5 D- ' !E vf (@ p( + (5 ,< (- (T *S# (' B @ 3
& n b
p+ b =(N 3 v n "
! ,!
( 9 v 6 !"( !#U- ! (
BD'FF ISZal@ ]3 U| +x v 7 !a3 U ;6
z ,9\ m
!# ;3 ! *|' ! al( ,% * # v+ &
ll (& )Z( ! - !" 4N
9" p > v( f &%B $ !a3 U 9l) f
m <n >* %2
T *fa ( p E4 *| ) ! 6 l !#( 4[ 3 v )
/v,l( n(p -% ! [ *|
_9f!
!)
ulp& 3 v )| 4X !"v ( >Z (f# v) 2 (l ! +B !
8 +) Ff O,' _3 pA <; ' ! pf@
GJILK
f -%_/v *| ) !T
!# 3 !9 ; pl( ,%&H v S[ (" ( B' *|l& '6PaY *|' ! al( ,% >A 4
v
!#! v(f
>
(4
) ! a; UM6
3
9 4 (' !#9lv 3 v& <n( !' v3 (2 ! a3 U
_(@ vPSV YPSR 3 ( l& +
+< ;l
'
bp ($ &% ! + p2 5B *pl !'*|@ b
a DA ] <= ! <Y ' ! ' *|T 7 7l
^m
l7 A :< %
1l K l l& !(%f 3 ) ' P6
`
l' 3 & +) (f[ P([' /!) l& aY(;
p@ ; 7x C" D 7lp( 3 2
( b !*| !a" :% * b p9
(Y " w 3 " ! " :% (" 3 p> <n( (l
)z- (' !I 3
H 1
_9f! Blv@ ) <n( :p
v
; ( &l3 Sz ( l(2
d&x ( ,%2 ` 1l; v-' *|- '<=( ; lp '
> 9 ;
X
lfl Dp f 6TaY$ vl 3 v; % * @l !" lv $( ( ll <n f v ) `a B' (" 7 pa-
' ) 2&%K' ! pf[
3 P4 :(' !' *|[p2 7
'
p 7( *|' ! al( ,% ' !#- (_ ) -6 aY
)N/v
)>*| ) !&*|' ! al( ,%N<n( ' ( ! l
$fb ! *|l >b f *| 3 !' *|l& V A !" U| !"%
3 @BOC" A z D)^S;6 n ( !#2U|' ! P) < *| )( f) &% !("! l2
*|' ! al( :% * b p2
f) !
_9f! g= y=
)
(- + ; 3 f ! ' !(% <n( <Y )*| 3 %)
3 f
!"! Y
+ ([ ) ; A:% ) !P84X I )6
Z> "9 ; } 9 9 [9
X
lp ( @ *| ) ! *p' ! al( ,%2 lp& 3 ! `a Z!a3 US 3 3 . v p =! a3 U
$ 4g
]
J
midpoint (km)
x10 4
0.0
0.6 0.8
4.0
1.0 1.2
8.0
1.4 1.6
0.0 0
I
II
III
V
depth (km)
2.0
2000
IV VI
4.0
4000
midpoint (km) 0214365)78t9K` ; {J V AIH R
J[PAIONB'D B@FQA
HKJWJ[RIAIONbc=AIJTrdJ[?IAIONP=@M zrdJMNBCP[O AS
0.0 4.0 8.0 x10 4 J[MNT _K2HZJor'=@MNXKJ[R2ONDAIHKJWMNJ[L'JDZT =@?
JONDbyR_
0.6 0.8 1.0 1.2 1.4 1.6
0.0 0
I
IV
0.0 I
II III V
VI
depth (km)
0.15
4.0
4000 2.02000
b IV 0.10
0.05
VI
0
V I? J[RIA
=@PqkTZJ V A
HUbUONL'?
=A
ONB'DWh2O A
HA
HKJ>J[R/A
ONbc=AIJ[T V =@?q=@bUJAIJ[?
R lezYm
R
=@bUJ%R
J[PAIONB'Dth2O A
H!TKJ[MNONDKJ=AIJ[T,YKMNBCPqk!Y<B'XKDKT=@?IONJR_ ^HKJi=@?
?
Bh2R
V B'ONDdAAIB{AIHKJ^?
JJ[PA
B'?IRGXZR
J[T%ONDWA
HZJ _'aMNBCPqk)ONRh^=AIJ[?h2O AIH 0214365)78t9 ) ; {J V AIH RIJPAIONB'D]B@FQAIHKJJ[RIAIONbc=A
J[T V =@?q=@bUJA
J[?
_
s ILK u jv'w'wxbyR KAIHKJiJ[RIAIONbc=A
J[T V =@?
=@bUJA
J[?
RfFzB'?oYZMNBlPqk]If=@?
J
LI K
s e u kC V
b u ldwGbm u jv'w'wWbyR\d| u wZ} ]wZ} w WR G \
R
| u wZ} w ,wZ} wZj{R G \
u wZ} w twZ} w Z\K=@DK
T u wZ} w ,wZ} w
FgB'?EYKMNBCPqk /Iq\s e u Z} wQkl ILK b u 'v'wbm u j~ 'w byR\ z)
! f@ A " ) * 1l(lz w *|; l
V R 6}
!
| u wZ} wZ} lw cR
\| u wZ} wZ j wZ} wZjR
\
u wZ}Nj wZ} w Z\ <=(3! ]<'!"!a'-' (Z ' 2B Z lnZ !
6"
=@DK T u wZ} w wZ} w >FgB'?xYKMNBCPqk %\fs e u Z} vkC IK b u ! Tx a
D'FVI\&K 5!"U- ! " PTD)FF S;6 `*| lf
'%') f
$
V
j dw'wbm u 'w'wbyR\C| u wZ} !wZ} lw %R
\Z| R
u wZ} gw "wZ} w pB (N ( *# z lf' -<= 3 H(
P &) (ll
R
\
u wZ}Nj wZ} w Z\2=@DK T u wZ} gw wZ} w >FgB'?]YKMNBCPqk %\ (> pm L!
* !#(z _B U K ` !
6"P\^^JS;6 5!(( $z
ILK
s e u ~Z} vQkC b u 'v'wQbm u jv'w'w]byR\ | u wZ} ' V
v wZ} w m " 9 .9
R
R
\| u wZ} w wZ} w 2R
\
u wZ}Nj!v wZ} w Z\=@DK" T u wZ} w !wZ} w
(" (* n ( 3 v& # !p !(% 7 ! a3 U 3 v
IK
=@DKT FgB'?2YKMNBlP
k {q\s e u Z} vQkC b u j 'v'wGbm u 'v'w'wibyR\
vl ! v ) ! 3 ( )
; ! l -l !"!T(
9
V| u wZ} '# R
v !wZ} w R
\l| u wZ}Nj #EwZ} w WR
$\
u wZ\l=@DK% T u wZ_ <Y )% +
! Kl' 3 Dl
p0 Z f# (l2
um ,9\
(( * _ ! a3 U ;6
v ; !a3 U 3-) < *| )- &( l 7
!# v +) >
B Ff @ ; p& 3 !K' *|l& ! ; p(
0'
p A p (l&N1V
U| f( (l& (" ( ' *pl& '6
! Tx
* 2
ll5(")<="! `5(l4 .Kfl v$'! p p C" <Y v f) &% !#U- ! (
BD'FF S v
} ( 9 6PSR M o r ! Tx
D)Q
+p(/2()6 Z!a3US ^ ^IQ@
;S v !#U- ! (
A !O69BD'FF S`" !"!" 3 2( @ *|l&
9 6PSR M r 7x &'
B
^ "D 4 G S (# (( f" p ) v& @!# ;3 ! *|' ! al( ,%
p /;+3 US
f -% & vl (- )6
QG IK
B
* # vH T<=( ; l !(B l )
f )<4 3( C
@ ! Y(
f alff + (lll' ()
(K# ! n Enfl vH 6S(l4 T ! ) <4 *|' ! al( ,% lp(K (>
v
A
!(>!# v( z# v& NQ-6 G+US v ('
! vf` >K
d&x Y !(4 !# v a(
'%v& Y lvD > J U BOC" DQS;-<= ; B = (fl& p+ D f * 2
J!
'
<) USf nbl
llp@ v(K
'&{ 5 !" 6 p
BOC" 2 D)Q S;6 5! f (z
d&x =
!( !# v(N =(
aY(%
p(( (4 @ 3
b m
_ ! +([f p$ ; ,<l# p & 4\A v +VAU v
:mTc
B U r
. ^ "D\ A^ ^R&S f
Xr !"!; .
N .b .( ! 6 ' Z>
lp& 3 !H 1l; K) *# v& 5J2US v $('
a (m
D(+ vff" ! + ' vfp
(
!) < * !#(lA ( \&6 Q U ) < @ *|) zlv& - (" :%Vp Y
f fl
! lw) fllp ! Y Y ! lfwlvf ! # )P v A( pP6
)
!# '% !# '! ]l vf" ! Y
' b Z( Fl; #*p4 v(f &% 8v@ fv Z( l 3 v3 U ' u" 3 )
p
( N<' !"!] ! " p' P6HaY <) U| A
f -%v fllp A C" D'VA v ` " 3 'D
pw A C" D'V
midpoint (km)
"p(l
$!b 3 !`*|'! al(,% * # p ! a;U W 9 Z(
- ((
!"'.<= ; ;'! ) !(%*S f"! # p &
x10 4
R US v
' \7USp ( ' # 3 )fl pP
0.0 4.0 8.0
^>d>
0.6 0.8 1.0 1.2 1.4 1.6
0.0 0
) / ;% 7 C" 2 D)V 6
I
II III
V
= ; (` p `3
&)
( F l(l :<l (
:<n
p N w Z v Nvf( vV -] ,<4 p(
'
depth (km)
IV
0
(l& (Y }
!# 3 ! -l& 7f pH 6 l) fY (
4.0
4000
VI 3 p$ (al( 4
ll & Yb 45 !b 3 !E* # v + (
*|; ' !P*|' ! al( ,%&&(lf` ,< 1l = ) 5(" ( 9v
(5 " 3 ';
p 6 aY = #
<4
4 !# p !"%+l *|)
+ C" D)V &%3 US V& vl (- Tnf" p ) v& 4!# &'
; { J V AIH RIJPAIONB'D B@FAIHKJWJR/A
ONbc=AIJ[T V =@?q=@bUJA
J[? C _ 3 !.*| !a" :%+* b p $ U9\
!a3 U ;6
! v'
0214365)78t9 K` <4
N 9) vf >( >! ( M zl
lz
(B
8 ]m :mTc Xr
* !"( B . ^ "D\S Bfp <=
Z) F; l& 5bp? pfA F!
' )7&% 5!"U- ! " BD)FF S` v Tx
midpoint (km)
! !#U- ! (
4 Tx :m c r <z
!
6YBD'FF S
B . ^ \S;6 @ ' ! ' *|2
0.0 4.0 8.0 x10 4
(9
7 ' vfp_ b N N f '
%5 lfl(l
0.0 0
0.6 0.8 1.0 1.2 1.4 1.6
&'
f" p( )
- $!# 3 !> pl(l( ,% !a3 US _ 9
9 v
;
_| Fa
I
depth (km)
0.10
& ! " :%' C" 7 D'V ( ) A >(@ ! ) p%
_9u!
2.0
2000
IV 0.05 AX . (- 7' !") 7+(f " # !*|' ! a
( ,%
* # v$ <n( ) ; f !" p7 v + v' 3
%Blp2
4.0
4000
VI 0 ) vfN b 5 ; ,p
l3 ( &(l' 6
9 x
} 7 Z> (? :@<A B
0214365)78t9 ; {
J V AIH RIJPAIONB'D B@FAIHKJWJR/A
ONbc=AIJ[T V =@?q=@bUJA
J[? _ Z%
vY<n"
l) MB &S
9 ;
`a [) # M( z vff"! $!(% vl(-`b
*|;' !P v !# 3 !H* # v (N
f >
*|' ! al(,%
"!"!"3 A/ (z @ *|l-
;(*|) -%$@X _9u! ' *#
'!#P6 D )<4 '
! ) A !b 3 ` K# 3 p *|' ! al( ,%-
}
@ l 3
; U2('#3 pP6 Zba'(f vlv
9 ;
!(%f * (- 7(lf" v) 7b `a K- ' ! [p@ pf'
- ( ,% f *p) >(9
'&{N
!(A!# v(:<l v > _9f!
) 7!# )%| n ! a3 U' 6aY 5 X !"p " (7 &
# v - 5\2
V2US? v f' *| 3 !H 1l f &( N *| l) ; U =
$a * v US_ 2= ]
B \^^RK SZ vY ,<
Z>
N' ! ) <%( ) .(w ( lv- 3 ! 1l .
b p & @ 3
& ( < b) l9 T( ( Zbp&lv *pl
R6 G_U v
( D6 RUS7 ] v
+
!(
!# v( p !H*|' ! al( ,%-: v !(%f4 d`
l3 ( -(l
# v - 4Q-6 G US v +( +J US7 6S l K" 3 v BD)Sf l*
pB .(N % "m
f zla - 2:%
7
) !#!(%
&( lp &% ; ! B
! p ' S@ *pl
vH -<* @ !l& )@ :< 3 v; fl !# v(l
midpoint (km) f)
p
pA I = ' * !" > 7&% ! v
x10 5
(p &% !
5(
fb ! *pl /( pH 6
z
0.2
3.0
0.4 0.6 0.8 1.0
8.0
1.2 1.4 1.6 1.8
0.0 (l$ (zfl !b
lz 3 5 v" ! ` l' fV w 3
0
# !
@ +)
! ( * # v <n( . Ff )
'%
(+ 0
lfl(l !# " :%' *p ! 8T!#
f *
('
Z'
- ) fl !# v(l> (lD vl ) &% U [ N !
6
depth (km)
2000
B \^^DE\^^I\Sn
'%+N (f) ) P6
2.0
depth (km)
2.0
2000 2.0
2000
4000
4.0 4000
4.0
a a
2000
2.0 2.0
2000
depth (km)
depth (km)
4000
4.0
4.0
4000
b b
0214365)78 9 ;"lJ[P[B'DKT RIJ[P[AIONB'D =FgAIJ[? e =dm ONR
B@AI?
B V ONP =@DKT+ezYm 0214365)78:9 ) ; "lJ[P[B'DKT"R
J[PA
ONB'D"=F4A
J[?%ez=dm{=@DKONRIB@A
?IB V ONP V H=@R
JfRIHKO F4A
=@DKONRIB@A
?IB V ONP V H=@RIJfR
HKO F4AUAIONbUJ]bUONL'?q=AIONB'D e4Fz?IB'b fMNklH=@MNO Fz=@H JA AIONbUJbUONL'?
=A
ONB'DteQONL'XK?IJxj_ Y<mo=@DKT:egY<m V ?IJR/Aq=@P
k TKJ V AIHtbUONL'?
=
=@M _N\>j dmq_aaB@AIH:R
J[PAIONB'DKR%h>J[?
JcM AIJ[?
J[T A
B"bc=AIPqH,AIHKJ =@b V MNO AIONB'De QONL'XK?
Jfj'jm_2HZJ^=@?
?
Bh2R V B'ONDdAAIBfAIHKJ>bc=@ONDWONb V ?
BrdJbUJ[DdA
R
AIXKTKJER V J[PAI?
XKb B@FoAIHKJ TKJ V AIpH bUONL'?
=A
J[T R
J[PA
ONB'D OND QONL'XK?IJtj'j =@PqHZONJ[rdJ[TYCSo= VKV M SlONDKL>A
HZJ :=@DKT V ?
J[RIA
=@PqkoTKJ V AIHbUONL'?q=AIONB'D _
=@DKT R/A
?IJ[AIPqHKJ[T A
B!TKJ V AIH XKR
ONDKLEAIHKJxrdJ[?/A
ONP=@M rdJ[MNBCPO AS,FzXKDZP[AIONB'D
Fg?
B'b QONL'XK?IJj Z_l^HKJo=@?I?
Bh2R V B'ONDdAaAIBAIHKJobc=@OND%ONb V ?
BrdJbUJ[DdA
R
=@PqHZONJ[rdJ[TEYdScAq=@kCONDKLU=@DKONRIB@A
?IB V SxONDdA
BU=@P[P[B'XKDdA_
C !#
v I ( /;$(3 "*|
v"3 v
| a":%@
!(%f pT Z(4
v; l) MB IS `a Z> ,? :@ A B 9 ;
7 7
* !
( l l)( ^6 \&6aY(`llv()
*|'! al(,% '!# ,' )" *S#(l v ( & + (
vb) !"
l' !
`!
!f9 ( ) lT(5lfl[ 3 v&# !M!# 3 !E; [ # !#!(%* %
f *|'! al(:% '!#('6 5!(
vll( :%% 2fv 5( !# '%a ' .<= ;<Y v2( vl(-' ( 7 *| ) ! *|' ! al( ,%) vf' !#(v lv 3 v ()
-% ( :(v
v _ "!
; ( &lz 5!"U- !( !O6 .=
&% ]O<Y *p` \'
1 p 3 2 !v ' !#: 3 9
@ !
Tx BD'FF S` v a !#(7; N !O6]BD'FFFIS;6 A ( ( ll <n f v f ff) _ ! ( ) N >(2
f @ v = lv
> ,9 ;
) Na 7- ( !)
@ v ! =
f -%@ v - A [' *)
*| ) !Y*|' ! al( ,% ' !#4 F 7 3 v; ' !T<Y )%
pl(l( ,%&v YX _9u! !"p
( - l) 4
l' 3 " !# v f B- (' ! b l 3
; U(' " 3 v
s% -i/214365+143
l
lz Z v f la l l& N
_ (pf <n( V
! 7
vH 6C )
lv
3
() ' *S !(% 2 @ (v
v &%' - *| " ' ) @-% )
f* -%&a + (N (flD . v (l'
>
*|l 7]! p ) 3 )*|' !( l7 p& 3 !' *|l& '6 *| !a" :%o&(@ H v vf) # (& vl' ( ; ,'
! K f 2 3 v3 U (' # 3 p_ <n( _( *|w 3 v ) Nbp
' _9u!
(Y X !"p ( ( ' ! pf !"%V
;(l
0'
lv ( )2*|' ! al( :% !b !()!
z ( w Mf# v( / (*' *p*lp@ )@/v? ' ( ! A 3 6
)
- @ *|ll& V %
.
p & !( :%lv@ ' ( _9u!
S # !# Z v& %9 ( 4X
l; ( &
lp
$!
fl p(z !"U ! (
> !
69BD)FF IS;6 x_ +x ' !# ) - @
(= ) )
! ) v; P l 3
; U
v N
!(zp (' " 3 )
p9 ) < ' " 3 pH <= 3 ,
U| !"p " ( 3
& b !"!"%
lv& - (" :%&$ v& ( $ !(@
7 ^>d>
/ ;%p ( s
X f" *|
( ; ,p
3
plN A >
<' !"!ba' (f) PP
p& 3 ! 1l l
u!
v; - (@ 5!"U- !( v
a * v U BD)FFIGS;
` <n( v(f ( !b '%a; N `pf( p() ! !#U- ! (
B +x
= !
6HBD'FF IS; v $L5l3 U z = !O6PB \^^I\S;6 aY
% 3 v3 U|) E6]aY$ vl' 3 (' %f) ! K 2*|' ! al( ,% :(v
v !#v ( ' ]() <' *|; ) - l; @
- ' ! 3 v& # !"!(%9; v v) ( E= (' !"D ( lfl9 Y!# 3 !*|' ! al( ,% * # vH P
*" !"
s
np[ (Dfl; v' <= N ! fA
(" (fllp ( !" 3 )$&%
*
@ ! l= *p 6
" 3 )
vN < A l *|) /;5( X _9f! v
! !( " 3 p@ llvlwfp <= Y
l 3 v3 U$' " 3 vH 6 lv ( (> lfl* ] v( f &% l) fz (
k S
^[
(l)f7 v f $3 '% 3 v ] -3 lp; #
) L 6 6 CT6"+C l(;P (z
6 E6"+
>
&' d`
LN 3 X7 ' UM
a`6"
f" v )
& p@ fp0 <n(0(lp@ D'FIQR t ! l-M"3 v+ v2*p'! al(,%
!(%f LDl
+
![" 3 p- 'PZ<=( ; (al;;; v p -%f l' * VDD - V\G&6
A
f ) # 6F& "!#!
M) (f9 9(A
!b p * (2
L5l3 [9
U 6"
a * v U P 36# D)FFV- C) vf" (# ! ( ,%
[ ](' (" 3 v; ' !#(
f &% -v &% ! *| - *| f pX 3
*| f !"%
_9f! ( _ ! < ;lv *p pl(l
<Y 3 + l X`
f V) (# MLDl &%f l' * FGQ - F F6 Tx
' _9f! *9 += +! * S=
X !)
- ' !
6]aYlf (( p ! X
%" 3 v% ' L5l3 U 6" l3 H 6" v a * v US H 36#\^^I\& O<4 )*|
U|D = !#v ( 5 lv A vlp@ 3
; US :*p' ! al( ,%+p 3 -%
Na ' (# LDl &%(' 6 9 ; `
= dx
3 ! Vf (>
l3 ( -(l' 6 p P6#
PDG&D D V-6
# N
H }
6#ALN ! " U|l' *M a`6" L5l3 U 6" W (K ff' H 9
6"
a * v USP 0
36"
\^^^- U!
%-
( v@ !
B : D
f"= ]O<Y *pT alff ` 9 [- ' !I/v Y L`!(
z
\} ^3 X! t S!
HX ,U| !#! W 6"I v LN v p
6" ' 6" [( f &%
4Y Z
s
3 !-5L5' Y !"! v p(a!#(a8T*-v X` \^^^ C3 v d`=
l' Slp- *|; )2<4 )*|l4
;) vfD ( l
a
l$ - N('!" ! vl p v_
U&
izg= Xz l) ( vz Y( fz
[[& x& p ! US 'p4
N 3 2 )* v" !# !N 28
!
6 `<!#+!"U|5
U ! * z ! t
K f -%_BOF NS;a 6 !
6MLDl &%' 6
l A( B f &%S:,a) v >(
z F= izg=
8l& "!
W " H 6# D)FFIQ& [ v !(% ) !> v3 " 3 v
*p 9(lvl B 8 ZS;T8 ! 3 (23 a ![ X l'
w *| d`
!a" :%7
!"%f ELD &%f l' ) .D)\JV D)\RF6
=f; !E f' ff p(= v z W ( >B 8 X S]b Y T d| *= z * ! 9a
U ) 6"8Y
3 p 6 6")
W v P 6"\^^-D
*S < ]V
v- f; # )6aY(V 5 5 < U$<Y v
= \
d`
*p' ! al( ,%$ v !(%(f MLDl -%f l' PD\VR D\FR 6
*S# (' +&%2(> 8p(f " ]d &{l; npl f u *p f
s izg= X| } w 9
U ) 6" Y (' !
<( 6-a`6" W ) 26" \^^I\& @' ! al( ,%
X (- ( b @8p@ ! - (
8
&%(
} _|
!(%f lfl' v@ ` ! ( * # p LDl
8T(l ) !VSl ll' =LDlvfl ll v pfll *" *(
a% i} | -%f l' 7x+x 7x PD R D Q\-6
t
f pP lz Y vf ( %l l(ll' 6 E 6 *' l& | } ! w ^>
U ) 6" Y U-! H A6# N3
(A6 W 6" \^^J-
t
( %&6
! K f - *| f p zfl f 3 _ ; lff' +
X` b 4aY(l Q}
% v &%f ) ![- (' ! % p )
T` Jx
v p( MLD &%f l' F^
Q^R 6
| X J!
%W [ U ) 6#
a *
US H ;6"\^^J- [ !(%f
p
! +x *9
5!"U- !( Ha`6"ID'FF !a":%>
!(%f .b 3 v(*|f'!(%
; Z> 9 ; `
@ dx
v; l) Na 7) # LDl &%f l' -\^D
X`^= ^|
f V) (# ] lf '8! 3 (B3a!9X l'6
\^I\G&6
T| v
S T p | al
!
5!"U- !( HPa`6"]D)FFIQ&Sl f
3
allff _ *|
f [ v Zl> F9 ;
U ) 6"I
( X` D f l
a * U H ;6"&\^^R X "3
* '! (,%
; `
a [) # HLDl &%f' \
) !#!(%' (p vl N xTx !(%
+
;
o(K l p v
)
( p
Na ) # L &% ' 7 -)6
x p
` D f l Jx
Xx
- UET856"`D'FF\- 1 3 &% 7
! v S ^9
QG-6
Tx z v v f L` Dl
"3 )$
ML &% ' V^ F\-6
5!"U- !( Ha`6" v +a * U H 36#ED'FFIG& @'! al(,%+
!" vfl
f $3
*|;f'!(% f( 4)b =LDl&%f l'
X` aY
f l M +xTx
H W 6"2D)FV - ) U '!# &% L
+xTx
%
! } ( v= 0!
-% ' PD)FGR D)F 6
D)GG^
D)G -6
Tx <9
6"9D)FF @' ! al( ,% v !(%f
p
3 v( *| f' !(% l v
f ` t p (
"!"! # P 6"4 8 ; 56"ND'FFF58Y -%
`
f ) # X (- ( ! %
%) vf % YaY L`
9 &% aY W ) 6EG-
t
p JIQ&D & JIQV-6
S
GVV -GFR 6
[Sl f f p l v v f
}
W )
t
p
f &% v
a * U
f z ( X` t f S [l l(l
H ;6"5\^^D " !(%
l p
| *) T`
Y !"!
L 6" D)FFIG-
= ^x
. 1 $ 3 $ ) # ! '*
f ` J l 3
; U" 3 vH | F
E v( &LDl ] !O68 6" 6
-%f l'
PD'RFIG
DG&D)J6
} Z l *E 6(A6"`8a; 3 vH `X6 6" 3 `L 6"5 v "| = t
<Y 3(' 2! |
6 6"HD'FV\-Mal p( l`
Bf) ( l& 3 v (
y!
S 4!#
- > 7 T`
( /[f &l(l M8Y [ p !# M- ( l 2! X` D ! [ _ z D
;lp& - 3 !
v @ pl ! %& !!\=
ZLlv( ) * 6(E 6 Y ; $ [ z
z US v |
7856 W 6 `3 U| ' 6 t aY( vf] ` l( =;'* !" A(A( &
8T ' #
6"
[[ ! ]X6#]\^^-DZaT< :( lf v ! p(N -l'7-%AX _9f! !"v( <=l *|(
*| !a" :%
v; - ' !
p@ /v>fl f = 1l; v !*|' ! al( :%! T
) 25 3 -+1
" #$ % /( p P' 6-8vf# ( (B
3 =-%N!a' &)
!( F l& # !af !#
l@ ) p ! ( Z> <`+! v > F9 ;
) Na ) " <n( 7> *p ) !H*|' ! al( ,%7(
p B %& ;
$) !] 3
f 'HLDl &%' 6 6 u- )6" J '
) v pG ILK 6B MgG IK O B
!& '
vf 9 1l; v
*x Z>
) B &S
.DR \ 6 *|l ><Y vA ll 3) bpl" & lv& 3 ! & vll
C ( ) W 66"ZD'FIG&DwSl f 2 *|' ! al( ,% v
( p `' B & * ! K ZG^^!7 6
' $ v +pl ! p A T`
MLD &%f l' .D'FI\ \^ 6 +x a llp( + *p ) !N*p' ! al( ,% * b pH T<
C ( ) W 6 6"&D)FIGJ *|' +! ! al( :%> ( pV ' !"( ! " (
7
)
z vl> <n( $lA lv& 3 !H v Z>
! <`
v * # v LDl -%f l' \Q-D \VV6 > l? 6B
) MB IS !# )%|
Up) - 0 C" AD) 6 f ( Y!
$ sKl
]
J
Midpoint (m)
Midpoint (m)
0 0
Depth (m)
Depth (m)
2000
2000
4000
4000
(a) (b)
-B-l
]Z+dqo>sda/Z$oZ qtKs-l6]Q1?faZaZ?-Zt
t
oqs t<QdZ<?qdqgoa?}qqdqK\yqZaaq
q
?dofqUZayqqooQg
asUUZqf/6qZZt/ZUq< ?
qZqZa?
tyxdqN]QosagaZa?sa?&6qZfadq
qa+ZZ
d$o??
s.odd
q
g&aZf yqa/oZl;?
a.
t6f
d
6a oaZlZ
a
oayqa a o a /]ol
q?
?qqoUo?f?y?Z
a
y
saUqodgyZZqSaZf-ZtdylUqZ
ZaZldqy]dod
-ZqoUoqo
o
d?
f_qZB od
qQZ NZZ
a?tsl/osa?_\aZ?f-Ztdo
-qd
aKyK
?aZfafaZNSt?ao> ;adqZqo@
oq
aqo.
a?ts?
ZqqoZZ Zd)
q?KdoZ qt
s]]aZBoa
qaU tf o a &qof yqa/oZBayqo
]Zs?
?/odqZas Zy?q-dZoqQdq&qqZq
oyq+aaQ?
qtQ
a>Zqdaq?Ztt$q$ZZaa/asa?AdaZ?
qqt
q/yaqoZgZ
at
d\q.Notf
/aodqosaBfZq
qoaZZ]tQfos1qsaZg/?
f oqdqt?}aaqfo6aZgdqyf
qZ
asa>a?ZsK?Z Z]Zq
ofqdZ1oa
qada&qZ
dq%=dgoKdaZ&q
qa;oZdoAqZ
yq t6f o1 6a
Z-qZ-qf?a?+$ tf do+ a tgf?
odly?oq?tq& aZBqZ
d
yU
aoqfqKt
-aK6UQ?Uay}
qgay?aZ&sg@
ZZa$+;Udo?ZZZd ZKdqdZqdaq? K/a
y?Z/Uoya?faZBf-Ztdo
lsZS]?f]oddf
Sayo}q
?
/aU
qos?
aZZ
oZt;ot
dqaos.}aZaZ.ZBaNqqZqfq?
/]Zgasaq
qa?ZtdZgZayqoqg
/f]a<qZ.dqodqZ
dqa
o}aZaf
a
q
ao&aa
Z?osao+fZdq]Q.fofaZ
dd qt
fqsq$sdZlyyqdZZqtyoddya
a?sdZ<
tda?-ZZ\?qZya1/qqoq
asa?AdaZaqZa\ZsU
?
_a/Zq
tQ
qxaqZZZK6qZZo/
a?ts?
Zqq+]Zy}a?s?
Zos
dya
_fqsa <o\d?qqo)qgZas Zo\?]QaSd?o
d?
?
qadqZa?o-y?qayq+Q?
aqqZ?qZ?-ZtltZ
/$-s\?KdqoZdZtqdqa ?oZ qtSs?
ZQodd
q
as
q
!
% U% Q
!#" & :80>QP8u8HDWHPW|H0iQY0b@OtbQP0YP+bOb % Q0 V W YtPuQ8Y0WQ\8!HW|8bOWbW8w%WPrObFYbQ
F0||!WW)0>QP8sQrPHY8HOiQHWWPH0HPY.>bO8h -0#00Y.Pj ( ! W P80Wti%WYuY_WP8H -\P8!WH.QWHPP'0
W:||!F+0>QPM% "! UP8 O88FY%W# W !%$ O'& 0)8!HrQWHPPv0 W:||8WWP>QP)%
j
+* UP8 O8+8WWY%W,#
! >! !%$ ! *H !
( )
-0WP||:0bQQ|8W8sQ-.+* ! %W 0//1Hu% Q<! 2/1HHu%
# W !%$ b1 # <! !%$ ! 1 0 QP# >8 !%$ !
Q<! %
U
!#"!& :80>QP8_bQPbOb Q<! QP # <! 8H|P:8|WHd D%WYH[ UHPHYYbYP. ( W 8 W P80Wu%YH
0 % / &
U 0 % /0 %& U 0 % *0!-\P|:P8Y_0|iQ|Q W%sYYHPW,/ & P|PQiQ|!8W QW # W QP # Y8 QW
0:b+% 0
8HW88HQY0bb
%W U% `
!#"& :8.P>QP8WhQ0hO8 # >8 QP QW QPM [QP Q<! UHWP_|DP8Y YH0 /0 HbQirbQ WP_QWP8H0
QQ|!8QW
008M%\WP 8HWb!HQYP8 & 0 ( W! W P80WMi%WYYWP8H
0.132
0.13
0.128
0.126
0.124
(3)
0.122
0.12
0.118
0.116
0.114
0.112
10 20 30 40 50 60 70 80 90 100
S/N
!#" & .ADO8F8LQQYP8\Q # >8 UHHQW>QPY_YHPQ PHYi%WYP \- P_F0OWP!Y %sUHWP_|DP8Y YHP /e8W_8HDQ|
%W8)O
_Q0>QuPHYH
!#" & :8P>QP8WbQP_H_WPW88WYHUHjWP|QFH Q\L|DP88H|HW8QAu>bO8W_ W[ DQPY9):P8Wh!-\P
HFPHWPH0Y>bO8eWH YHPW-/uY bObW>QY+Ov +G-/.>bO8s W Q 0/0 %eQP # !%$ b1 & U i HWWPH0HPY
>8Ob M0 ( 8WbO_HWY8D%YHMQWP.F0||!WWP>Q0bs ! QPuWP.PQiQ|!8W QW /0 1% Q<! / * %
# W !%$ *H 0 # <! !%$ /0HQ0 # >8 !%$ b*0-\PQQ|!8sbFY%WYH'\Qe8FHW|b'PWYP) D%YHWH H 0YQrU QP
D%YHtF+/ HPYQeU
3 3
x 10 x 10
90 90 6
120 60 8 120 60
3 3
4
6
150 2 30 150 2 30 2
4
1 1
2 0
180 0 0 180 0
2
2
4
4
210 330 210 330
6 6
!#" & 8Y0WQY HQF0Y0_bO8 Y|8WYPQYUHWP|:0bQAeHWWPH0HPY\>bO8W ||YQYYHP8W:||!FsP>QP8
YHP-1Q -\P8YPPQY 8P8WbO WP|P ( 88P88.!A88WP|i8 WiQ88'bO8 Y|8 8H|P0W8UH_0WP|DP8QP'UH
WPebF
P|DP82H:iQYP8uUHdWPePHPO08HYYe|hO8H0YDO8WYHrQ8)H
2 :%WYH+FH UH # '& U :%WYH+F+/
#0:YPg ( W! W P80WMi%WYY\Q /0
: p>H
W : p 0
<! : p
W : p<Q
<!
: p 0
>8 : CWp P
>8 .
Z?d l Zo?d
! : p>H
W : p 0
<! +1H : p<H
W : p 0
<!
: CWp
>8 . : p :
>8
Uaa ?d Utdq?q?
/d
: p 0
W : p :
<! : p O
W : p D
<!
!y
: p<Q
>8 > 8 :.CWp P
_sU
dooa
?d sSdZ?od
d
: p>H
W : p 0
<! : p :
W : p<Q
<!
H (!
: p :
>8 : CWp
>8 . ^
tqqaqZooy?a?
??d $aftafZZtdgo
?q
??d
!#",+!& YQ2YOO8YHWbWP WHWP 8!WYHPWH 0+_ )DYPsPHY8H P#00Y.P ( ! 0b:Li%WY .- / $ %&
U P0/>QYW|H .- 2/ $ !%& P"1b8 + QPHQW .- X $ & QP@ WP YWYY0P>Q32P8H0UH| A .- $ -\P
QWh_Q r) WP.bFY%W8'PY8!YH'QP.8| 54FH#00 YQWP6 ' 8YYYPW & P:P.bD)hQ+QWhYWPQbY.0W[Q
WPs#00 Y Wu8W8!WP PHF
@
@
!#"Z$!& :8.P>QP8WbQPhO8 Q<! Q 0'% ! ! * UHWPW Pb!WYHPUHc WP=+ ):YP 0HY8H & 0 />QY|HW &
U WP 1b8 +QPHPQ & QPU WP Y WYY0>Q"2P8QUH| AON-0eQbY.0W[Y|hQWP88HPO88YD*)OYW WHdWP_WPWO!
#0:Y PYiuHYObQWA\QW2
!#"\$$& iQ!WP 8HPOWUH0 ( 8 WbOQ8WPWeQbY.0WPWHg8Y YH.YQH8Q0u8HeQQ :Ye%UW8 Q 0H / !! / 8 Y
.i)YY:b%W8w8YHuW+WP8PY ! ! *0 -\P)8WWY%8@W:||!F 0>QPuQ8Y.00UH ! ! t * QP ! H/
WH|8%%WiLWP PYW88WYHPQP ( W8PrQ!PW !Wh
W
f W<V
X h W 6
Fh W
?^qrQPJ
<!
f >!<V
X h <! 6
Fh >!
?^qr:OJ
>8
f W Xgf >! Xgh >8 ?i Xx f <! J
?i<VFh > 8 J%?i<Nf <! J
?^qrH0J
H
CWP-496P
ABSTRACT
Coda wave interferometry (CWI) utilizes multiply scattered waves to diagnose
small changes in a medium by using the scattering medium as an interferometer.
Since the medium is usually stationary over the duration of a seismic experi-
ment, different (non-overlapping) time windows in the coda allow for indepen-
dent estimates of the medium perturbation. If the seismograms are contami-
nated with noise, only those time windows can be used for which the amplitude
of the coda is above the ambient noise level. This limits the usable number
of independent time windows. Here, we show how bias due to noise in CWI
can be accounted for, by deriving a correction factor for the cross-correlation
coefficient. This correction factor allows more time windows further into the
decaying coda to be used, and hence allows for a reduction of the error bars on
the medium perturbation estimates. We demonstrate the validity of this correc-
tion factor by using data from a numerical experiment and field measurements.
These experiments involve the displacement of point scatterers and a change
in the source location, respectively. The application of our correction factor is
not limited to CWI, but can be used to correct for bias induced by noise in
any application that uses cross-correlation between different signals that are
contaminated with noise.
Key words: Scattering, seismic coda, monitoring, noise, earthquake location
correlation coefficient. We demonstrate its validity by (uu , up )(t,ts ,tw ) uu (t0 )up (t0 + ts )dt0 , (4)
ttw
using data from a numerical experiment involving the
displacement of point scatterers, and field data involv- where ts is the time-shift in the cross-correlation, and
ing the displacement of the source location. 2tw is the length of the time window. The maximum
The organization of this paper is as follows. We first of this correlation coefficient occurs at time-shift ts =
review the principles of CWI and then derive the noise h i, where h i is the mean traveltime change given by
correction factor. Subsequently we show the validity of (Snieder, 2002)
the correction factor using the numerical and field ex- P 2
T A T T
periments mentioned above. We conclude with a short h i P 2
. (5)
T AT
discussion of the results. In appendix A we derive a con-
dition for the reliability of the correction factor. This Using equations (1) and (2) in equation (3), Snieder et
condition can be used to determine the time windows al. (2002) and Snieder (2002) show that the maximum
(t,t )
in the coda where the correction factor is reliable. value of this cross-correlation function, rmaxw , is given
by
(t,tw ) 1 2 2
Coda wave interferometry rmax =1 , (6)
2
Coda wave interferometry is based on a path summa- where the frequency 2 is given by
tion approach, which is a generalization of the Neumann R t+tw 2 0 0
series solution of the Lippmann-Schwinger equation for tt
uu (t )dt
2
R t+tww . (7)
scattering of classical waves [e.g. Snieder (1999)]. This ttw
u2u (t0 )dt0
path summation can be represented as
X Here, u denotes the time derivative of u(t). The variance
uu (t) = AT (t) , (1) of the traveltime perturbation, 2 , is given by
T
A2T (T h i)2
P
2
where T denotes the different trajectories the waves T P 2
. (8)
T AT
have traveled, and the function AT (t) denotes the con-
Correcting for bias due to noise in CWI 81
In expression (8) and equation (5), the summations con- In equations (12)-(14), terms of the form (u, n)(t,ts ,tw )
cern only trajectories with arrival times in the window appear, where u = u(t) is a deterministic signal and
with central window time t and window duration 2tw . n = n(t) a realization of a stochastic process. Using
In deriving equation (6), it is assumed that the dou- that the noise signals have zero mean, we then find that
ble sums over trajectories T and T 0 with T 6= T 0 are
(u, n)(t,ts ,tw ) 0 , (15)
on average incoherent, and used a second-order Tay-
lor expansion of the cross-correlation function in the provided the window length tw is at least several dom-
quantity T h i (Snieder, 2002). The latter approx- inant periods of the noise-contaminated signal. For the
imation is valid when the traveltime perturbation due terms (nu , np )(t,ts ,tw ) and (np , nu )(t,ts ,tw ) in equations
to the medium perturbation is smaller than the domi- (12)-(14), we use the independence of the noise realiza-
nant period of the multiply scattered wavefields. Snieder tions to find
(2004b) showed that the double sums over trajectories
T
and T 0 with T 6= T 0 , are proportional to 1/ tw f , with (nu , np )(t,ts ,tw ) 0 , (16)
f the bandwidth of the signal, so that their impor- with the same assumption for the window length tw as
tance thus reduces for larger time windows and larger used in equation (15). Approximations (15) and (16)
bandwidth data. become more accurate with increasing window lengths.
Using equations (15) and (16) in equations (12)-
(14), it follows that
Correcting for the bias due to noise ` 0 0
uu , up (t,t ,t ) (uu , up )(t,ts ,tw ) , (17)
s w
So far, we assumed the waveforms uu (t) and up (t) to be ` 0 0
uu , uu (t,0,t ) (uu , uu )(t,0,tw )
free of noise. To study the influence of noise on CWI, we w
derive a correction factor for the correlation coefficient + (nu , nu )(t,0,tw ) , (18)
when the waveforms uu (t) and up (t) are contaminated `
u0p , u0p (t,0,t ) (up , up )(t,0,tw )
with noise. We define w
1 0 0 1 0 0 (20)
.
` 0 0
uu , up (t,t ,t ) (uu , uu )(t,0,tw )
`
0
r (t, ts , tw ) = q s w
. (10) up , up (t,0,t )
` w
(uu , uu )(t,0,tw ) u0p , u0p (t,0,t )
0 0
w This correction factor c(t, ts , tw ) depends on the un-
Our aim is to derive a correction factor c(t, ts , tw ) such known noise signals nu (t) and np (t). In practice, as-
that suming the noise is stationary, we can estimate these
noise signals using the recorded wavefields uu (t) and
r(t, ts , tw ) c(t, ts , tw )r0 (t, ts , tw ) . (11) up (t) before t = 0 of the experiment. This gives us es-
Throughout the remaining derivation, we assume that timated noise signals nu (t) and np (t). Replacing nu (t)
these noise signals are identically and independently and np (t) in equation (20) with their estimated coun-
distributed (i.i.d.), and wide-sense stationary with zero terparts, is a good approximation only if the terms of
mean [i.e., the mean is zero for all times, and the auto- the form (u, n)(t,ts ,tw ) and (nu , np )(t,ts ,tw ) in equations
correlation depends only on the time shift ts (Papoulis, (12)-(14) are small compared to the remaining terms.
1991, p. 298)]. Hence, for this replacement to be valid, we want the
Using equation (9), we find following inequalities to hold:
` 0 0
uu , up (t,t ,t ) = (uu , up )(t,ts ,tw ) + (uu , np )(t,ts ,tw ) |(uu , np )(t,ts ,tw ) + (up , nu )(t,ts ,tw ) + (nu , np )(t,ts ,tw ) |
s w
|(uu , up )(t,ts ,tw ) | , (21)
+ (up , nu )(t,ts ,tw )
2|(uu , nu )(t,0,tw ) |
+ (nu , np )(t,ts ,tw ) , (12)
` |(uu , uu )(t,0,tw ) + (nu , nu )(t,0,tw ) | , (22)
u0u , u0u (t,0,t ) = (uu , uu )(t,0,tw ) + 2 (uu , nu )(t,0,tw )
w 2|(up , np )(t,0,tw ) |
+ (nu , nu )(t,0,tw ) , (13) |(up , up )(t,0,tw ) + (np , np )(t,0,tw ) | . (23)
`
u0p , u0p (t,0,t ) = (up , up )(t,0,tw ) + 2 (up , np )(t,0,tw )
w In appendix A we rewrite these three inequalities into
+ (np , np )(t,0,tw ) . (14) one inequality that can be evaluated using only the noise
82 H. Douma & R. Snieder
Figure 1. Unperturbed (filled circles) and perturbed (open circles) locations of the point scatterers in the medium. For display
purposes the differences between the unperturbed and perturbed locations are magnified by a factor 10. The star denotes the
source and the triangles the receivers. The modeled seismograms are shown for three receivers. The drawn seismograms are
related to the unperturbed scatterer locations, and the dotted seismograms to the perturbedlocations. For times later than
t = 4.7 102 s (marked by the dotted line), the waves have scattered more than four times. The horizontal bar indicates the
window length for the time-windowed cross-correlations used to calculate the RMS displacement shown in figure 2.
contaminated signals u0u (t) and u0p (t), and an estimated equation (20) to calculate the correction factor. In this
noise signal n0 (t). The resulting inequality [equation case condition (A5) does not apply, and a new condi-
(A5)] then determines if a certain time window has a tion could be derived. To avoid belaboring the point, we
reliable correction factor associated with it, when the refrain from such a treatment.
unknown noise signals nu (t) and np (t) in equation (20)
are replaced with n0 (t). Replacing nu (t) and np (t) with
n0 (t) in equation (20), the correction factor is given by Displacement of the scatterers
c(t, ts , tw ) The problem of inferring the average displacement of
! 1 ! 1 scatterers in a strongly scattering medium from the mul-
(n0 , n0 )(t,0,tw ) 2
(n0 , n0 )(t,0,tw ) 2
1 0 0 1 ` 0 0 (24)
, tiply scattered wavefields, has been used to study flu-
(uu , uu )(t,0,tw ) up , up (t,0,t ) idized particle suspensions (Weitz & Pine, 1993; Heck-
w
meier & Maret, 1997; Page et al., 1999; Cowan et al.,
and is considered reliable if inequality (A5) is satisfied
2000; Page et al., 2000; Cowan et al., 2002). In geo-
with an appropriate value of [O(101 )].
physics, this problem may be relevant when a strongly
To derive condition (A5), we assume that the noise
scattering region in the earth is strained, causing the
is stationary and that nu (t) and np (t) have about the
scattering heterogeneities to move. In such a situation
same noise levels, (i.e., the same variance). If the noise
the displacement of the scatterers is not expected to
levels of nu (t) and np (t) are substantially different, sep-
be random, but will be correlated among scatterers.
arate noise estimates of nu (t) and np (t) can be used in
Correcting for bias due to noise in CWI 83
Here, we present a numerical experiment with point
scatterers in a homogeneous background model, where
we randomly perturb the scatterer locations and use
CWI to infer their root-mean-square (RMS) displace-
ment. Although this experiment is not directly related
to a changing strain in the earth, it serves the purpose
of testing the workings of our correction factor.
Snieder & Scales (1998) showed that for inde-
pendent perturbations of the scatterer positions and
isotropic scattering, the variance of the path length L
is given by
2
L = 2n 2 , (25)
where n is the number of scatterers in the path, and is
the RMS displacement of the scatterers in the direction
of either coordinate axis (horizontal and vertical for two
dimensions). Note that Snieder and Scales assume all Figure 2. Inferred RMS displacement as a function of the
directions of random displacement to be equally likely. center window time t (solid line) plus or minus one standard
deviation (dotted lines). The true value of is indicated by
As a result, the RMS displacement is the same in each
the horizontal solid line. The range of validity of CWI is indi-
direction (horizontal and vertical for2D), i.e., for 2D the
cated by the vertical lines. The half-window time tw used in
true RMS displacement would be 2. Using that the calculating the time windowed correlation coefficient equals
number of scatterers is on average given by n = vt/l , 0.01s.
with l the transport mean free path (Lagendijk & van
Tiggelen, 1996) and v the velocity, and using L = vt, it
follows that the variance of the traveltime perturbations frequency, 0 = 2f0 , and f0 = 600 Hz. The frequency
is given by band used was 400 800 Hz, with a resulting domi-
nant frequency of about 500 Hz due to tapering on ei-
2 2 t ther side of the spectrum. Since in our experiment we
2 = . (26)
vl have isotropic scattering, the transport mean free path
Inserting this in equation (6), it follows that the RMS equals the mean free path, i.e., l = l. The mean free
displacment can be found from path in our experiment was measured to be l = 17.6 m,
r vl which, using n = vt/l, can be used to infer that after
=
(t,t )
1 rmaxw . (27) t = 4.7 102 s the waves are on average scattered more
2 t than four times. This time is indicated in Figure 1 by the
Since the perturbations of the scatterer locations are dotted vertical line. The perturbed scatterer locations
assumed to be independent, and since the scattering is are indicated by the open circles in Figure 1. For display
assumed to be isotropic, the mean traveltime pertur- purposes the displacements are magnified by a factor 10.
bation h i = 0. This means that the maximum of the The actual RMS displacement true = 8102 m in both
(t,t )
cross-correlation function rmaxw occurs at zero lag, i.e., the horizontal and vertical direction. This displacement
ts = 0. equals just 1/38 of the dominant wavelength (the dom-
Figure 1 shows the setup of our numerical experi- inant wavelength = 3 m). The resulting waveforms
ment to test the inference of the scatterer displacements calculated using the displaced scatterers are shown for
from the seismic coda using equation (27). This exper- three receivers by the dotted lines.
iment was also outlined by Snieder et al. (2002). One Figure 2 shows the inferred value of [using equa-
hundred point scatterers (solid dots) are contained in tion (27)] as a function of the central window time t,
an area of 40 80 m2 , and the waveforms are calcu- where the estimated values for from all 21 receivers
lated using a numerical implementation (Groenenboom were averaged. [The receiver spacing was chosen such
& Snieder, 1995) of Foldys method (Foldy, 1945). The that the calculated multiply scattered waveforms were
resulting seismograms are shown by the solid lines for uncorrelated, meaning they can be treated as indepen-
three locations on the edge of the area, and the source dent.] The dotted lines show the average inferred value
location is indicated by the asterisk. In these calcula- of as a function of time plus or minus one standard
tions the scattering amplitude was set to 4i, in or- deviation. The half-window duration tw was 102 s, re-
der to get the maximum possible scattering strength sulting
p in a window length of 10 dominant periods, and
as constrained by the optical theorem [(Groenenboom = 3.66 103 rad/s. The vertical dotted lines indi-
2
& Snieder, 1995); in their notation we used = 4]. cated the range of validity of equation (27). For early
The background velocity equalled 1500 m/s and the times (i.e., t < 4.7 102 s) the relation n = vt/l
2 2
source spectrum S() = e /0 , with the angular for the number of scatterings used in the derivation
84 H. Douma & R. Snieder
of equation (27) is not valid, and for late times (i.e., mated average value of is too high, especially for later
t > 2.9101 s), the second-order Taylor approximation times, where the lower amplitude values of the coda
of the auto-correlation of the source signal is inaccurate result in lower signal-to-noise ratios. Figure 4c shows
by more than 15%. This latter time is indicated by the that the correction factor from equation (24) accurately
rightmost dotted vertical line in Figure 2. Within the accounts for the bias due to the noise. The noise es-
range of validity of equation (27), the true displacement timates were obtained using the signal before the main
is recovered within the range given by the average RMS first arrival in the seismograms. In the calculation of the
displacement plus or minus one standard deviation. For results shown in Figure 4c, we used condition (A5) for
late times, the correlation coefficient is close to zero and, each receiver, with = 0.125 to select the time windows
accordingr to equation (27), the inferred value of is then used for the inversion of [note that this is a different
than that used by Groenenboom & Snieder (1995)]. As
given by vl/ 2 t . This function is indicated by the a result, the number of receivers that had usable time
dashed line in Figure 2 and indeed agrees well with the windows for a given central window time t, varies for dif-
inferred value of for late times. For these times the in- ferent t. This causes the jagged appearance of the aver-
ferred value of is of course no longer a good estimate age inferred value of (and the standard deviation). We
of the true RMS displacement. used a time window only when at least seven receivers
As mentioned in the introduction, when the wave- (i.e., 30% of the receivers) satisfied condition (A5) at
fields are contaminated with noise, the number of in- the central window time t. After t 9 102 s fewer
dependent time windows that can be used is limited to than seven time windows satisfied condition (A5) with
traveltimes where the ambient noise level is small com- = 0.125, and hence the correction factor was judged
pared to the amplitudes of the multiply scattered waves. unreliable. As a result, the noise-corrected estimates of
In order to be able to use as many independent time the RMS displacement are not shown for times larger
windows as possible, and hence reduce the error bars on than t 9 102 s.
the inferred perturbation, it is important to correct the
cross-correlation function for the bias due to the noise.
To test the correction factor in equation (24), we added Source separation
band-limited noise to the waveforms for all 21 receivers
from our numerical experiment. The bandwidth of the Snieder & Vrijlandt (2004) used CWI to estimate the
noise was the same as that of the noise-free signals (i.e., distance between seismic events having the same source
400 800 Hz). Figure 3 shows a waveform with and mechanism, that are recorded at a single station. They
without the added noise. Using the noise-contaminated derive the imprint of a change in source location on
waveforms, we again calculated the inferred values of the variance of the traveltime differences, and then use
both with and without the correction factor; see Figure equation (6) to infer this change from the maximum of
t,tw
4c and 4b respectively. For reference, Figure 4a shows the cross-correlation function, i.e., rmax . For two double-
the inferred value of when the noise-free signals were couple sources with a source separation in the fault
used. Figure 4b shows that the noise induces a bias in plane, they show that the relation between the source
the estimated value of ; the presence of noise reduces
the correlation between the unperturbed and perturbed
waveforms, and causes the inferred values of , calcu-
lated using equation (27), to be larger. For early times,
the true value of is embedded within the average value
of plus or minus one standard deviation, but the esti-
Correcting for bias due to noise in CWI 85
displacement s and the variance of the traveltimes 2
is given by
6 7 Pwave
+ 8 a)
8
2 = (s)2 , (28) 1
20 30 40
2 3
7 +
6 6 0.8
where and are the P and S wave velocities, respec- Cross Correlation
b)
tively. Different non-overlapping time-windows provide 0.6
20 30 40
independent estimates of the source separation. These 400
Distance
independent estimates in turn allow for the calculation 300
[m]
of error estimates of the source separation. 200
100
Figure 5a shows two seismograms (events 242003 c)
0
and 242020) from earthquakes on the Hayward fault, 20 30 40
California (Waldhauser & Ellsworth, 2000), recorded at time [s]
station CSP of the Northern California Seismic Net- Figure 5. Two seismograms from two earthquakes on
work. The recorded signal before the arrival of the P the Hayward fault, California, recorded at station CSP of
wave shows that the noise level is considerable. Figure the Northern California Seismic Network (a), their cross-
5b shows the maximum of the time-windowed cross- (t,t )
correlation maximum rmaxw (b), uncorrected (thin line) and
correlation function without the correction factor (thin corrected (thick line), and the inferred source displacement
line) and with the correction factor (thick line) applied, (c) using both the uncorrected (thin line) and corrected
(t,t )
and Figure 5c shows the inferred values of the source (thick line) values of rmaxw shown in (b). The horizontal
separation using equation (28). Here the estimate of line in (a) indicates the window length used in the cross-
the noise, i.e., n0 (t), was obtained from the waveforms correlation.
before the first arrivals. The half-window duration tw
used is 5 s (the full window length is indicated by the Conclusion
horizontal bar in Figure 5a), and the P and S wave ve-
locities used to calculate the source displacement are We have derived a factor that corrects for the influ-
= 5750 m/s and = 3320 m/s, respectively. Note ence of noise on the cross-correlation function, and have
that we used overlapping time windows, since we plot shown its accuracy using both numerical and field data.
(t,t )
rmaxw simply as a continuous function of the central The application of this correction factor is shown in the
window time t. Of course, non-overlapping windows context of CWI, for the inference of the RMS displace-
could be used to ensure independent estimates of the ment of scatterers and a displacement of the source,
source separation. from multiply scattered wavefields. For the displace-
(t,t )
Figure 5b shows that the corrected values of rmaxw ment of the scatterers, we showed that in the presence
maintain a fairly constant level for times t late in the of noise, a displacement of only 1/38 from the dominant
coda, whereas the uncorrected values decrease earlier in wavelength can be succesfully retrieved from the cross-
the coda because the noise decreases the similarity be- correlation between the unperturbed and perturbed sig-
tween both waveforms. As a result, the inferred values nals. This shows the power of CWI when compared to
of the source separation using equation (28) are more or methods that use singly-scattered waves only.
less constant for larger traveltimes when the corrected Since for both the scatterer displacement and
(t,t )
values of rmaxw are used (Figure 5c). This indicates source separation cases, the perturbation is independent
that the correction factor c(t, ts , tw ) given by equation of the traveltimes of the multiply scattered waves, using
(24), accurately corrects for the influence of the noise non-overlapping time windows to estimate the pertur-
on the cross-correlation function. For very large times bations provides a consistency check of the method, and
(say t > 40 s) the corrected values are more variable allows the calculation of error estimates. In this context,
because the correction factor becomes unreliable. We our correction factor is relevant, as it increases the num-
purposely showed the times where the corrected val- ber of usable time windows and hence allows for a re-
ues become variable, to indicate the level of variation duction of the error estimates. In addition the correction
caused by an unreliable correction factor. Of course, factor adjusts for bias in the cross-correlation induced
the time where the correction factor becomes unreliable by the noise. Since our factor depends on an estimate of
could have been estimated using condition (A5) with an the noise level in the data, we present a condition that
appropriate value of [O(101 )]. allows determination of the reliability of the correction.
This condition can be verified using only noise contam-
inated signals and an estimate of the noise level in the
data. Using this condition, the time windows used in the
86 H. Douma & R. Snieder
windowed cross-correlation can be judged to be reliable Maret, G., & Wolf, G.E. 1987. Effect of Brownian Motion of
or not. Scatterers. Z. Phys. B, 65, 409413.
The use of the proposed correction factor is of Page, J.H., Jones, I.P., Schriemer, H.P., Cowan, M.L., Sheng,
course not limited to CWI. Any application that uses P., & Weitz, D.A. 1999. Diffusive transport of acoustic
cross-correlations between different and noisy signals, waves in strongly scattering media. Physica B, 263-264,
37.
and needs to correct for bias induced by noise, can ben-
Page, J.H., Cowan, M.L., & Weitz, D.A. 2000. Diffus-
efit from the correction factor presented here. ing acoustic wave spectroscopy of fluidized suspensions.
Physica B, 279, 130133.
Papoulis, A. 1991. Probability, Random Variables, and
ACKNOWLEDGMENTS Stochastic Processes. McGraw-Hill Inc.
Pine, D.J., Weitz, D.A., Chaikin, P.M., & Herbolzheimer, E.
We thank Mark Vrijlandt for making Figure 5, and Ken 1988. Diffusing Wave Spectroscopy. Phys. Rev. Lett, 60,
Larner for his critical review. In addition, we thank 11341137.
Western Geophysical (now WesternGeco), in particular Snieder, R. 2004a. Coda wave interferometry. Pages 5456
Peter Nuttal, for funding a three-month visit of H.D. to of: 2004 McGraw-Hill yearbook of science & technology.
the Center for Wave Phenomena at the Colorado School New York: McGraw-Hill.
of Mines in the summer of 2000, that formed the basis Snieder, R. 2004b. Extracting the Greens function from the
for this work. This work was supported by the NSF correlation of coda waves: A derivation based on station-
ary phase. Phys. Rev. E, 69, 046610.
(grant EAR-0106668) and by the sponsors of the Con-
Snieder, R., & Vrijlandt, M. 2004. Constraining Relative
sortium Project on Seismic Inverse Methods for Com- Source Locations with Coda Wave Interferometry: The-
plex Structures at the Center for Wave Phenomena. ory and Application to Earthquake Doublets in the Hay-
ward Fault, California. submitted to J. Geophys. Res.
Snieder, R.K. 1999. Imaging and averaging in complex me-
REFERENCES dia. Pages 405454 of: Fouque, J.P. (ed), Diffuse waves
in complex media. Kluwer, Dordrecht.
Cowan, M.L., Page, J.H., & Weitz, D.A. 2000. Velocity fluc- Snieder, R.K. 2002. Coda wave interferometry and the equi-
tuations in fluidized suspensions probed by ultrasonic libration of energy in elastic media. Phys. Rev. E, 66,
correlation spectroscopy. Phys. Rev. Lett., 85, 453456. 0466151,8.
Cowan, M.L., Jones, I.P., Page, J.H., & Weitz, D.A. 2002. Snieder, R.K., & Scales, J.A. 1998. Time reversed imaging
Diffusing Acoustic Wave Spectroscopy. Phys. Rev. E, as a diagnostic of wave and particle chaos. Phys. Rev. E,
65, 0666051/11. 58, 56685675.
Derode, A., Roux, P., & Fink, M. 1995. Robust acoustic time Snieder, R.K., Gret, A., Douma, H., & Scales, J. 2002. Coda
reversal with high-order multiple scattering. Phys. Rev. Wave Interferometry for estimating nonlinear behavior in
Lett., 75, 42064209. seismic velocity. Science, 295(22 March), 22532255.
Derode, A., Tourin, A., & Fink, M. 1999. Ultrasonic pulse Waldhauser, F., & Ellsworth, W. 2000. A double-difference
compression with one-bit time reversal through multiple earthquake location algorithm: Method and application
scattering. J. Appl. Phys., 85, 63436352. to the northern Hayward fault, California. Bull. Seism.
Fink, M. 1997. Time-reversed Acoustics. Physics Today, 50, Soc. Am., 90, 13531368.
3440. Weitz, D.A., & Pine, D.J. 1993. Diffusing Wave Spec-
Foldy, L.O. 1945. The Multiple scattering of waves. Phys. troscopy. Page 652 of: Brown, W. (ed), Dynamic Light
Rev., 67, 107119. Scattering. Clarendon Press, Oxford.
Gret, A., Snieder, R., Aster, R.C., & Kyle, P.R. 2004a. Mon- Yodh, A.G., Kaplan, P.D., & Pine, D.J. 1990. Pulsed
itoring rapid temporal changes in a volcano with coda diffusing-wave spectroscopy: High resolution through
wave interferometry. Submitted to Geophys. Res. Lett. nonlinear optical gating. Phys. Rev. B, 42, 47444747.
Gret, A., Snieder, R., & Ozbay, U. 2004b. Monitoring stress
change in an underground mining environment with coda
wave interferometry. Submitted to the J. of Mining Ap-
plications. APPENDIX A: A CONDITION TO
Gret, A., Snieder, R., & Scales, J. 2004c. Time-lapse moni- ESTIMATE THE RELIABILITY OF THE
toring of rock properties with coda wave interferometry. CROSS-CORRELATION CORRECTION
Submitted to J. Geophys. Res. FACTOR
Groenenboom, J., & Snieder, R.K. 1995. Attenuation, disper-
sion, and anisotropy by multiple scattering of transmitted Equation (20) for the correction factor of the cross-
waves through distributions of scatterers. J. Acoust. Soc. correlation coefficient, depends on the unknown noise
Am., 98(6), 111. functions nu (t) and np (t). In practice, we dont know
Heckmeier, M., & Maret, G. 1997. Dark Speckle Imaging of these noise functions, and can at best estimate the aver-
Colloidal Suspensions in Multiple Light Scattering Me- age noise levels. Since in practice we often only have one
dia. Progr. Colloid. Polym. Sci., 104, 1216. estimate of the noise level, as opposed to some ensem-
Lagendijk, A., & van Tiggelen, B.A. 1996. Resonant Multiple ble average, we want the correction factor to be reliable
Scattering of Light. Phys. Rep., 270, 143215. when the unknown noise signals in equation (20) are
Correcting for bias due to noise in CWI 87
replaced by a single estimate of the noise. If inequalities (21)-(23) are satisfied, the correction factor will only weakly
depend on the estimate of the noise signals. Here we rewrite these inequalities into one inequality that can be verified
using the noise contaminated signals and an estimate (or single realization) of the noise n0 (t).
To write inequalities (21)-(23) as a single one, we first add the left- and right-hand sides of equations (21)-(23),
while multiplying equation (21) by two for convenience in the further derivation. This gives
`
2 |(uu , np )(t,ts ,tw ) + (up , nu )(t,ts ,tw ) + (nu , np )(t,ts ,tw ) | + |(uu , nu )(t,0,tw ) | + |(up , np )(t,0,tw ) |
2|(uu , up )(t,ts ,tw ) | + (uu , uu )(t,0,tw ) + (nu , nu )(t,0,tw ) + (up , up )(t,0,tw ) + (np , np )(t,0,tw ) , (A1)
where we have used that zero-lag auto-correlations are positive definite. If conditions (21)-(23) hold, equations (17)-
(19) from the main text are good approximations. We can use these approximations, together with the linearity of
( , )(t,ts ,tw ) and equation (9), to approximate inequality (A1) as
2|(u0u , u0p )(t,ts ,tw ) | + (u0u , u0u )(t,0,tw ) + (u0p , u0p )(t,0,tw ) . (A2)
Since in CWI we assume that the travel-time perturbations in the time window [t tw , t + tw ] are small, we
expect a positive cross-correlation between the signals u0u (t) and u0p (t) for times where CWI is valid. Using this, we
can write condition (A2) as
|(u0u + u0p , n0 )(t,ts ,tw ) | + |(u0u , n0 )(t,0,tw ) (n0 , n0 )(t,0,tw ) | + |(u0p , n0 )(t,0,tw ) (n0 , n0 )(t,0,tw ) |
2([u0u + u0p ]/2, [u0u + u0p ]/2)(t,0,tw ) (A3)
where
we substituted
for both noise signals nu (t) and np (t) the estimated noise signal n0 (t), and we assumed
(nu , np )(t,t ,t ) (u0u , np )(t,t ,t ) + (u0p , nu )(t,t ,t ) to eliminate the (nu , np )(t,t ,t ) term. The latter approxi-
s w s w s w s w
mation is more appropriate for larger signal-to-noise ratios, larger time windows, and uncorrelated noise realizations
nu (t) and np (t). Note that substituting a single noise signal for both unknown noise signals nu (t) and np (t) is ap-
propriate only, if both noise signals have similar noise levels. Dividing both sides of condition (A3) by (n 0 , n0 )(t,0,tw )
(which is positive definite), and defining the ratio
` 0
[uu + u0p ]/2, [u0u + u0p ]/2 (t,0,t )
(t, tw ) w
, (A4)
(n0 , n0 )(t,0,tw )
inequality (A3) leads to
1 (u0u + u0p , n0 )(t,ts ,tw ) (u0u , n0 )(t,0,tw )
0
(up , n0 )(t,0,tw )
+ 1 + 1 /(t, tw ) , (A5)
2 (n0 , n0 )(t,0,t ) w
(n0 , n0 )(t,0,t )
w
(n0 , n0 )(t,0,t )
w
p
where is O(101 ). Here (t, tw ) can be interpreted as the average signal-to-noise ratio.
Condition (A5) is satisfied only for time windows that have a large average signal-to-noise ratio. The l.h.s of
condition (A5) can be evaluated using the noise contaminated signals u0u (t) and up (t), and an estimated noise signal
n0 (t). This condition can thus be used as a selection criterion to determine which time windows have reliable correction
factors associated with them, when an estimated noise signal n0 (t) is used (i.e., when equation (24) is used to calculate
the correction factor). Note that can be interpreted as the inverse of the signal-to-noise ratio.
88 H. Douma & R. Snieder
CWP-498
ABSTRACT
We present a rational interpolation approach to nonhyperbolic moveout correc-
tion in the t x domain, for qP-waves in homogeneous transversely isotropic
media with a vertical symmetry axis. This method has no additional compu-
tational overhead compared to using expressions explicit in the relevant pa-
rameters, i.e., the anellipticity parameter and the (zero-dip) normal moveout
velocity VN M O . The lack of such additional overhead can be attributed to the
observation that, for a fixed value of and a fixed zero-offset two-way travel-
time t0 , the moveout curve for different values of VN M O can be calculated by
simple stretching and squeezing of the offset axis, where the amount of stretch
or squeeze depends on the change in VN M O . This observation is based on the
generally accepted assumptions that the traveltimes of qP-waves in transversely
isotropic media, depend mainly on and VN M O , and that the shear-wave ve-
locity along the symmetry axis has a negligible influence on these traveltimes.
The accuracy obtained with this method is as good as that of these approxima-
tions. The method can be tuned to be accurate to any offset range of interest,
by increasing the order of the interpolation, making it accurate for arbitrary
offsets.
We test the method using both synthetic and field data, and compare it with the
nonhyperbolic moveout equation of Alkhalifah and Tsvankin (1995). Both data
types confirm that for & 0.1 our method significantly outperforms the nonhy-
perbolic moveout equation in terms of combined unbiased parameter estimation
with accurate moveout correction. A comparison with the shifted hyperbola
equation of Fomel (2004) establishes almost identical accuracy of the rational
interpolation method and his equation. Under the above-mentioned approxima-
tions, we show that the correction factor for the Alkhalifah-Tsvankin equation
introduced by Grechka and Tsvankin (1998), is independent of VN M O , and we
present a method to estimate its optimal value in practice. This factor can be
used to maximize the performance of the Alkhalifah-Tsvankin approximation.
Key words: Moveout analysis, nonhyperbolic, anisotropy, anellipticity, VTI,
rational interpolation
t (% of t 0 )
Figure 1a show contours of the maximum absolute
traveltime difference (in percentage of t0 ) between ray-
traced traveltimes for a single horizontal VTI layer and
the nonhyperbolic moveout equation without the cor-
rection factor, for virtually all combinations of and
VN M O of practical interest. The maximum ODR is two,
and t0 = 1 s for all models considered. The ray-traced
a) b) c) traveltimes were determined with VS0 = 0 km/s and
= 0. The differences in the traveltimes are of the or-
Figure 1. Comparison of the accuracy of the nonhyperbolic der of a percent, which in this case amounts to 10 ms.
moveout equation without the correction factor, i.e., C = 1 For a dominant frequency of 50 Hz in surface seismic
(a), with the optimal correction factor C (b), and of the data, this amounts to a traveltime error of about half
[2/2] rational interpolation (c), for a range of models (com-
a dominant period. Such errors are not negligible, can
binations of and VN M O ) that spans most models of prac-
lead to substantial reduction in stacking power, and thus
tical interest. Contours are drawn for the maximum absolute
percentage error in traveltime (compared to ray-traced trav- cause errors in the estimation of and VN M O in velocity
eltimes where ray tracing was done with VS0 = 0 km/s and analysis. Note that here we use the true values of and
= 0) for a single horizontal VTI layer, over a range of ODR VN M O for the moveout analysis rather than the best-
values up to two. The contoured values are shown in percent- fit values, since we want to analyze the accuracy of the
age of t0 . Subfigures (d) through (f) show the same contours nonhyperbolic moveout equation for a known model. It
as in subfigures (a) through (c), except for a range of ODR is known that the traveltime differences shown in Figure
values up to four. The stars in subfigures (d) through (f) in- 1a can be reduced by using the best-fit values of and
dicate the model parameters used for the residual moveout VN M O , rather than the true ones.
plots in Figure 4.
Grechka and Tsvankin (1998) state that, for a single
horizontal VTI layer, introducing the coefficient C = 1.2
method and the shifted hyperbola method of Fomel minimizes deviations from the exact traveltimes for the
(2004) establishes the almost identical accuracy of the range of offset-to-depth ratio (ODR) 1.5 2.5. They
two methods. also mention that the correction factor can be used as
an optimization parameter by comparison of traveltimes
obtained using equation (2) and ray-tracing. Figure 1b
Accuracy of the nonhyperbolic moveout is as Figure 1a, except now the optimized correction
equation factor was determined for each combination of and
VN M O shown and then used in equation (2). Note that
The nonhyperbolic moveout equation of Tsvankin and indeed the introduction of the correction factor reduced
Thomsen (1994), rewritten in terms of the anellipticity the maximum errors in the traveltimes substantially, re-
parameter by Alkhalifah and Tsvankin (1995), is given sulting in errors of the order of a tenth of a percent. A
by straightforward application of the C = 1.2 for all mod-
els (not shown here), did reduce the errors somewhat,
x2 2x4
t2 (x) = t20 + ,(1) but the errors were still of the order of a percent. This
VN2 M O VN2 M O [t20 VN2 M O
+ (1 + 2) x2 ]
indicates that a straightforward application of C = 1.2
where t0 is the two-way traveltime at zero offset, x is is in general not recommended, except for models where
offset, and VN M O is the (zero-dip) NMO velocity. Note C = 1.2 may happen to be the optimal correction fac-
how the moveout reduces to hyperbolic moveout for el- tor. It is worth mentioning that in practice the opti-
liptical anisotropy ( = 0), as pointed out by Helbig mal value of C can be determined only approximately
(1983). This equation is exact at both zero offset and because the true values of and VN M O are unknown.
infinite offset. In an attempt to increase the accuracy Therefore, Figure 1b represents the best accuracy that
at intermediate offsets, Grechka and Tsvankin (1998) can be obtained using the nonhyperbolic moveout equa-
introduce a correction factor. Noting that for a single tion with the correction factor.
horizontal VTI layer the absolute error in is at least As pointed out by Alkhalifah (1997), a larger maxi-
twice as large as the relative error in the horizontal ve- mum ODR (say larger than 1.5) provides increased sta-
bility and resolution for the inversion for using non-
hyperbolic moveout. With current acquisition systems,
4 H. Douma, A. Calvert & E. Jenner
such values of ODR are feasible, especially for shallow Tsvankin (1995) obtained equation (1) from the nonhy-
targets. Moreover, the need for cost-effective acquisition perbolic moveout equation of Tsvankin and Thomsen
systems could, with time, increase the maximum ODR (1994), by also setting VS0 = 0. In addition, Alkhalifah
in acquisition systems. Also, in near-surface (or shallow) and Tsvankin (1995) showed that the time signatures
geophysical application, large ODR is common. Figures of qP-waves in homogeneous VTI media depend mainly
1d and 1e are as Figures 1a and 1b, respectively, except on the (zero-dip) normal-moveout velocity VN M O and
that here the maximum ODR is four. The introduction the anellipticity parameter , with an almost negligi-
of the optimal correction factor again reduces the errors ble influence of VP 0 . Since we are interested only in
in traveltime significantly, but the errors are of the order traveltime calculations, we can choose = 0, and thus
of a percent; that is, they are sizable. VP 0 = VN M O and = , in equation (5). That equation
then becomes
1
The dependence of nonhyperbolic moveout for V () = VN M O sin2 + 1 + ...
2
a single VTI layer on and VN M O q` ff1/2
2
From simple geometric considerations, it follows that for 1 + 2 sin2 2 sin2 2 , (8)
a homogeneous horizontal VTI layer, the traveltime is
given by while equations (3) and (4) for the traveltime t and the
associated offset x, become
VP 0 t 0
t= , (3) VN M O t 0
v cos t= , (9)
v cos
where VP 0 is the P-wave traveltime along the vertical
symmetry axis, is the group angle, and v is the group and
velocity for propagation in direction . The offset x as- x = VN M O t0 tan , (10)
sociated with this traveltime is given by
respectively. Note that the phase velocity V () now
x = VP 0 t0 tan . (4) depends linearly on VN M O . This linearity causes the
1 dV
In a TI medium, the phase velocity is given by term in equations (6) and (7) to be indepen-
V () d
(Tsvankin, 2001, p.22) dent of VN M O . Since the dependence of the group angle
f
on the anisotropic parameters is governed by the term
V () = VP 0 1 + sin2 1 . . . 1 dV
2 [cf. equation (6)], the group angle is inde-
191/2 V () d
pendent of VN M O and depends only on . In addition,
s 2
2 sin2 2 ( ) sin2 2 A
=
1+ (5), it follows from equation (7) that the group velocity v
f f ; depends linearly on VN M O since the phase velocity de-
pends linearly on VN M O . From equations (9) and (10) it
where is the phase angle, and are the Thomsen
2 then follows that the traveltime t becomes independent of
parameters, and f 1 VS0 /VP20 , with VS0 the S-
VN M O and that the associated offset x depends linearly
wave phase velocity along the symmetry axis. For such
on VN M O . Also, t0 is a simple scaling factor for both
a medium, the group angle is related to the phase angle
the traveltime t and the associated offset x. In a single
through (Tsvankin, 2001, p.29)
horizontal VTI layer, this means that for fixed and t0 ,
1 dV the moveout curve for different values of VN M O can be
tan +
V () d calculated by simple horizontal stretching and squeezing
tan = , (6)
tan dV along the offset axis (see Figure 2). This important ob-
1
V () d servation is a straightforward consequence of the neg-
ligible influence of VS0 on qP-wave traveltimes in TI
while the group velocity v is given by
s media and the fact that the kinematics of qP-waves in
2 homogeneous VTI media depend mainly on VN M O and
1 dV
v = V () 1 + . (7) . To make explicit the independence of the traveltimes
V () d
of VN M O , we rewrite equation (9) as
It is known that the qP-wave phase-velocity in TI t0
media depends only weakly on VS0 (Tsvankin & Thom- t= , (11)
v|VN M O =1 cos
sen, 1994; Tsvankin, 1996; Alkhalifah, 1998). For all
kinematic problems regarding qP waves, VS0 is there- where v|VN M O =1 denotes the group velocity calculated
fore usually ignored. We likewise set VS0 = 0 (or f = 1) for VN M O = 1 km/s.
in equation (5). This is the acoustic approximation Figure 1a (as well as b-e) shows that the con-
from Alkhalifah (1998; 2000). Note that Alkhalifah and toured maximum traveltime differences between ray-
traced traveltimes and traveltimes calculated using the
Nonhyperbolic moveout analysis using rational interpolation 5
x
t
0
V nmo
t
g) h) i)
d) e) f)
Figure 5. Accuracy of the [2/2] rational interpolation as a
function of the separation in ODR of the interpolation trav-
eltimes for = 0.3 and = 1.0. In both cases VN M O = 2.4
km/s. Even if the separation in ODR is two, i.e., the max-
imum ODR is 8, the percentage errors in traveltime are of
the order 101 .
C=1.0 optimal C RI
1 1.00 2097 2101 0.00 1.00 2097 2101 0.00 1.00 2099 2101 0.00
2 0.98 2058 2179 0.06 0.98 2056 2188 0.07 0.98 2058 2188 0.07
3 1.00 2271 2839 0.28 0.98 2253 2996 0.38 1.00 2252 3022 0.40
4 0.99 2313 3012 0.35 0.93 2290 3255 0.51 1.00 2282 3285 0.54
ODR = 4
C=1.0 optimal C RI
1 1.00 2094 2099 0.00 1.00 2094 2099 0.00 1.00 2096 2099 0.00
2 0.95 2029 2207 0.09 0.95 2025 2218 0.10 0.96 2035 2218 0.09
3 0.57 2238 2894 0.34 0.54 2299 2892 0.29 0.84 2189 3158 0.54
4 0.57 2320 3038 0.36 0.57 2300 3152 0.44 0.87 2241 3391 0.64
Table 1. Comparison of obtained semblance values (s), VN M O , Vhor , and ,for the three methods, for maximum ODR of
two (top) and four (bottom), for a layered medium. The model consists of four horizontal homogeneous VTI layers with the
following values of VN M O and . The parameters for each layer are given in the main text.
but the same values of VN M O and (solid). The an- Figures 6 and 7, showed basically no difference between
sitropy parameters are VP 0 = 3094 m/s, VS0 = 1510 both methods.
m/s, = 0.05, and = 0.256 (i.e., VN M O = 2935 m/s
and = 0.34), and t0 = 646.5 ms. Except for the solid
line, Figure 8 reproduces Figure 7 of Fomels (2004) pa- Application to horizontally layered VTI media
per. Note that the accuracy of the shifted hyperbola
approximation is marginally better than the ray-traced Up to this point, we have treated only a single horizontal
traveltimes with VS0 = 0 km/s and = 0 but the same VTI layer. In this section we test the applicability of the
values of VN M O and . The maximum difference be- rational interpolation method to a horizontally layered
tween these two traveltime approximations is about 1 VTI medium, and compare its accuracy to that of the
ms (i.e., 0.15% of t0 ) for this particular model. From nonhyperbolic moveout equation with and without the
a practical point of view, the two approximations are correction factor. Based on the work of Tsvankin and
therefore identical. Even though we have not calculated Thomsen (1994), Grechka and Tsvankin (1998) showed
the solid line with a rational interpolation, we can ap- that the nonhyperbolic moveout equation (1) remains
proximate the solid line to almost arbitrary precision valid in horizontally layered media provided the param-
with rational interpolation (i.e., with [M/N ] rational eters and VN M O are replaced by effective values that
interpolation where M, N > 2). For ODR up to four, are some average over the vertically heterogeneous over-
we showed this in Figure 1f using a [2/2] rational inter- burden. Also, they rewrite the Dix-type inversion pro-
polation. Since the rational interpolation method that cedure, originally introduced by Tsvankin and Thomsen
we propose uses traveltimes calculated with VS0 = 0 (1994), in terms of and VN M O . In that method, the
km/s and = 0, we conclude that the accuracy of our effective values of and VN M O , obtained from apply-
method is basically identical to that of the shifted hy- ing the single-layer equation (1) to data from a layered
perbola approximation of Fomel (2004). Comparison of medium, are used to estimate the interval values of
semblance scans and moveout corrections calculated us- and VN M O .
ing equation (21) and the rational interpolation method We compare the effective values of and VN M O ob-
(not shown here), for the models and offsets studied in tained using the nonhyperbolic moveout equation, with
12 H. Douma, A. Calvert & E. Jenner
[t]
a) b) c)
Figure 9. Field data used to test the rational interpolation method. Inline (a), crossline (b), and plan view of the residual
topography of the event used for testing (indicated by the arrows in a and b) after removal of the regional dip (c). The dotted
lines in c denote the locations of the inline and crossline sections shown in a and b. The dashed line in c denotes the location
of a regional fault.
and without the correction factor, and the rational in- tor gives comparable values of and VN M O to those
terpolation method, through application of all methods obtained with the rational interpolation method, for
to a horizontally layered medium with the following pa- all layers, just as for the single-layer case (see Figure
rameters: 6). Also, the semblance values from the nonhyperbolic
moveout equation with the correction factor slightly de-
layer 1 : h = 1 km (h is the depth of the bottom of
crease with increasing depth (or level of anellipticity),
the layer), VN M O = 2098 m/s, Vhor = 2098 m/s, = 0,
just as for the single-layer case. For maximum ODR=4,
(VP 0 = 2000 m/s, = 0.05, = 0.05),
the nonhyperbolic moveout equation with and without
layer 2 : h = 2 km, VN M O = 2000 m/s, Vhor = 2300
the correction factor result in decreasing semblance with
m/s, = 0.16, (VP 0 = 2000 m/s, = 0.16, = 0),
depth (i.e., increasing level of anellipticity), indicating
layer 3 : h = 3 km, VN M O = 2892 m/s, Vhor = 3745
a lack of ability to accurately moveout-correct the data.
m/s, = 0.34, (VP 0 = 3048 m/s, = 0.255, = 0.05),
Again this is analogous to the results from the single-
layer 4 : h = 4 km, VN M O = 2460 m/s, Vhor = 3880
layer case (Figure 7). That the rational interpolation
m/s, = 0.74, (VP 0 = 3292 m/s, = 0.195, = 0.22).
method has substantially larger semblance values than
For all models we used VS0 = 0 km/s. Note that here do the other two methods, for all layers, indicates that
the first layer is elliptically anisotropic, and the second this method is able to flatten the gathers best, with the
through fourth layers have the same model parameters least residual moveout. From a practical point of view,
as those in the models studied in the first through third the method that flattens the gathers best, is expected
row of Figures 6 and 7, respectively. to give the most confidence in the estimated values of
Table 1 shows the results from all three methods and VN M O . From this point of view, the rational in-
for a maximum ODR of two (top) and four (bottom). terpolation therefore provides more robust estimates of
These results closely resemble the results obtained from and VN M O than do the other two methods, for maxi-
the single layer numerical tests shown in Figures 6 and mum ODR=4.
7. Since all methods are for a single-layer VTI medium It thus seems that the rational interpolation
only, and because we apply these methods to layered method, which is based on a single horizontal VTI
VTI media, the , VN M O , and Vhor values in Table 1 are layer, is suitable for application to a horizontally lay-
all effective values. For maximum ODR=2, the nonhy- ered medium, at least up to maximum ODR=4 and ar-
perbolic moveout equation without the correction fac- bitrary levels of anellipticity. For maximum ODR=2,
tor gives consistently lower estimates of than do the the nonhyperbolic moveout equation with the correc-
other methods, and the differences increase with depth, tion factor gives comparable estimates of the effective
i.e., as we increase the level of anellipticity. For that and VN M O and comparable semblance values to those of
method, the semblance values (indicated with s in Ta- the rational interpolation approach. This indicates that
ble 1) are high for all layers, indicating high quality for maximum ODR=2, even for extreme levels of anel-
moveout correction. Use of the optimal correction fac- lipticity the effective values of and VN M O estimated
Nonhyperbolic moveout analysis using rational interpolation 13
Figure 11. Map view of derived from the event of interest in the field data with maximum offset of 2250 m, i.e., ODR 2
(a-c), using semblance scans with the nonhyperbolic moveout equation without (a) and with the optimal correction factor (b),
and the [2/2] rational interpolation method (c). Subfigures d-f are as a-c, except that the maximum offset is 4500 m, i.e.,
ODR 4. Estimates of at locations with poor offset distributions were set to white.
tained using nonhyperbolic moveout equation without in large estimates of , compared to those of the other
the correction factor (a), are generally smaller than the methods. From the synthetic tests, we would expect this
estimated values from both other methods, just as method to give values comparable to the estimated
in our synthetic data examples (cf. Figure 6). The es- values obtained using rational interpolation. None of
timates of obtained using the rational interpolation our synthetic tests, however, included amplitude-versus-
method (c) and the optimal correction factor approach offset (AVO) variations, whereas the field data example
(b) are quite comparable. This striking similarity is con- clearly does (cf. Figure 10). Some additional synthetic
sistent with the results from the synthetic data also. The tests (not shown), indicate that the rational interpo-
same comparison for VN M O (not shown) showed that all lation method is less sensitive to such AVO variations
three methods gave similar estimates of VN M O , just as than is the nonhyperbolic moveout equation method,
in the synthetic examples. with or without the correction factor; this can be ex-
Figure 11d, e, and f, are as a, b and c, but here the plained by the high accuracy of the rational interpo-
maximum ODR is four. Notice, that the estimated val- lation method . Therefore, the lack of resemblance be-
ues are spatially less variable, for all methods. This can tween the estimated values for the correction factor
be understood in light of the improved resolution in for approach and the rational interpolation may be due to
larger ODR, together with the lack of evidence of sub-
stantial lateral heterogeneity from the seismic data (cf.
We did not include phase changes with offset in the tests
Figure 9). Here, the correction-factor approach results
concerning sensitivity to AVO variation.
Nonhyperbolic moveout analysis using rational interpolation 15
the AVO variations in the field data. Note that the pres-
ence of AVO variations causes semblance based moveout
correction to be biased towards the offsets with higher
amplitudes. Such bias is especially noticeable when an
inaccurate moveout approximation is used to moveout
correct the data.
The values resulting from the rational interpola-
tion method are on average slightly higher than those
from the nonhyperbolic moveout equation method with-
out the correction factor; av = 0.28 for the rational
interpolation method, and av = 0.25 for the nonhy-
perbolic moveout equation method. The lower values
of for the nonhyperbolic moveout equation method
without the correction factor are consistent with the re-
sults from the synthetic-data tests (cf. Figure 7). For
all methods, the values are high compared to values
typically reported from nonhyperbolic moveout analysis
[e.g. Toldi et al. (1999)]. Of course the relatively high
values are likely a result of the particular lithology
of this area. However, most studies reporting estimates
of have been done on marine data, where a substan-
tial waterlayer reduces the effective values of , hence
introducing a bias in what are traditionally considered
acceptable values of . Figure 12. Normalized semblance difference between the
The [2/2] rational interpolation resulted in more nonhyperbolic moveout equation with, and without the cor-
spatially smooth and continuous values of both (and rection factor (a), and between the [2/2] rational inter-
VN M O ) that in some parts correlate somewhat with the polation method and the nonhyperbolic moveout equation
geologic trend (cf. Figure 9c). It is important to note method without the correction factor (b), for a maximum
offset of 2250 m, i.e., ODR 2. Subfigures c and d are as a
that this spatial continuity was not imposed, but fol-
and b, respectively, except for a maximum offset of 4500 m,
lowed from a straightforward application of the ratio-
i.e., ODR 4.
nal interpolation method presented here. Since the seis-
mic data show no indication of substantial lateral het-
erogeneity (cf. Figure 9), the relatively smooth spatial tion has on average 10% higher semblance values. This
variation of is consistent with the seismic data. This supports (for the whole dataset) our findings from the
increases our confidence in the values obtained using synthetic results that the correction factor approach in-
the rational interpolation method. creases distortion in the moveout correction and that
Figure 12a shows the normalized semblance differ- the rational interpolation method results in more accu-
ence between nonhyperbolic moveout equation method rate moveout correction.
with and without the optimal correction factor, for max- In conclusion, for a maximum ODR larger than two,
imum ODR=2. The normalized semblance difference is the [2/2] rational interpolation method provided gener-
the difference between the semblances from both meth- ally improved moveout correction, improved semblance
ods divided by the semblance for the nonhyperbolic values, and more spatially continuous estimates of for
moveout equation method without the correction factor. this field data example. This supports the applicability
Figure 12b is as Figure 12a, but it shows the normalized of the rational interpolation method to a horizontally
semblance difference between the [2/2] rational inter- layered VTI medium, observed on synthetic data in the
polation method and the nonhyperbolic moveout equa- previous section. As the numerical tests showed similar
tion without the correction factor. For these offsets, all results, we expect these findings to generalize to other
methods result in similar semblances, indicating similar datasets. For maximum ODR up to two, all methods
ability to flatten the gathers. This supports our findings had similar ability to flatten the gathers, but the non-
from the numerical tests that all methods obtain com- hyperbolic moveout equation approach gave comparable
parable results in terms of moveout correction, for an estimates of to those of the the rational interpolation
ODR range up to two (cf. Figure 6). Figures 12c and d method only when the optimal correction factor was
are as 12a and b, except that here the maximum ODR used.
is four. For this range of ODR, the correction-factor ap-
proach has on average 7% lower semblance values than
does the nonhyperbolic moveout equation without the
correction factor, whereas the [2/2] rational interpola-
16 H. Douma, A. Calvert & E. Jenner
Discussion equivalents of steeper dips in time-migration. There-
fore, the increased accuracy for large maximum ODR
We have shown that for ODR values up to 8, the
provided by the rational interpolation for traveltimes
[2/2] rational interpolation results in an accuracy of
of qP-waves in such media suggests improved accuracy
O(101 )% of t0 for most models of practical interest. If
when the rational interpolation scheme is used in imag-
higher accuracy is desired, or accuracy up to larger ODR
ing of steep reflectors in the context of pre and post-
is needed (e.g., in near-surface seismic experiments),
stack time-migration in VTI media.
higher-order polynomials can be used in the rational ap-
proximation. The added computation time for inclusion
of several extra terms is negligible; hence the efficiency
Conclusions
of the proposed method remains essentially the same.
We have not done any numerical testing for polynomi- We have presented a rational interpolation approach to
als of order higher than two. nonhyperbolic moveout correction of qP-waves in VTI
In the field data example, we treat the overburden media. The accuracy of the method was tested using
of the reflection event of interest as a single horizon- both synthetic and field data and compared with that of
tal homogeneous VTI layer. As a result, the estimated the nonhyperbolic moveout equation, which is the cur-
values of and VN M O are effective parameters. The geo- rent standard in seismic data processing. Both synthetic
logical significance of such effective quantities is difficult and field data results confirm that our method signifi-
to establish, and an approach assuming a layered over- cantly outperforms the nonhyperbolic moveout equation
burden and resulting interval estimates of and VN M O in both unbiased parameter estimation and the quality
would remove this difficulty. Although we do not demon- of moveout correction when a maximum ODR larger
strate such an approach in this paper, we believe that than two is used. For a single horizontal VTI layer, and
such an approach is feasible by applying rational in- for a maximum ODR up to four, the errors from the
terpolation in a layer-stripping fashion. We leave the [2/2] rational interpolation are O(102 )% of t0 or less;
verification of this idea to a future study. Meanwhile, this is one to two orders of magnitude more accurate
the single-layer approach outlined here can be used to than the nonhyperbolic moveout equation of Tsvankin
obtain more accurate estimates of the average (or ef- and Thomsen (1994). Even for a maximum ODR up to
fective) values of and VN M O in layered media. Using 8, the traveltime errors resulting from the [2/2] ratio-
these values in current Dix-type averaging procedures nal interpolation are O(101 )% of t0 , for virtually all
could lead to better interval estimates of these parame- models of practical interest.
ters. Under the customary assumptions that traveltimes
We showed that our rational interpolation ap- of qP-waves in VTI media depend mainly on and
proach achieves accuracy almost identical to that of VN M O , and that the influence of VS0 on traveltimes of
the shifted hyperbola approximation of Fomel (2004). qP-waves in TI media is negligible, we found that the
Since we show the applicability of the rational interpo- traveltimes in a single horizontal VTI layer, for fixed
lation method to horizontally layered VTI media, the group-angle, , and t0 , are independent of VN M O , while
almost identical accuracy of both methods implies the the associated offsets are linear in VN M O . As a conse-
applicability of the shifted hyperbola approximation of quence, therefore, the nonhyperbolic moveout curve for
Fomel to such media also. For horizontally layered me- different values of VN M O , but fixed and t0 , can be
dia, a plane-wave decomposition, obtained through a calculated by simple horizontal stretching or squeezing,
p transform, is the natural decomposition of the i.e., along the offset axis, where the amount of stretch or
data. This fact was successfully used by Van der Baan squeeze is determined by the change in VN M O . This ob-
and Kendall (2002) and Van der Baan (2004), to obtain servation allows us to calculate the traveltimes needed
interval estimates of and VN M O from moveout in the for the interpolation from a small number of traveltimes
p domain. Even though they successfully estimate for a reference value of t0 (conveniently tref
0 = 1 s) and
interval values of and VN M O , the acquisition geome- values ranging from -0.2 to 1.0 in steps of, say, 0.01.
try does not always allow a straightforward p trans- This range of covers most models of practical interest.
form. Therefore, there is room to try to extend our t x The few hundred traveltimes for a reference value of t0
based rational interpolation method to render interval and values ranging from -0.2 to 1.0, can be precom-
estimates of and VN M O . puted in about a minute on a modern PC and stored
The geometry of straight rays involved in velocity in a table. Therefore, the rational interpolation method
analysis for a horizontal homogeneous layer and the ge- has no additional computational overhead compared to
ometry of the rays associated with point scattering in that of the nonhyperbolic moveout equation method.
a medium with constant velocity are identical. There- We show that the observation of the stretch-squeeze
fore, the rational interpolation method proposed here influence of VN M O on the nonhyperbolic moveout causes
is immediately applicable to the problem of post-stack the correction factor C in the Alkhalifah-Tsvankin non-
and pre-stack time-migration in VTI media. Larger off- hyperbolic moveout equation (2) to be independent
sets in the context of moveout velocity analysis are the of VN M O , under the above-mentioned approximations.
Nonhyperbolic moveout analysis using rational interpolation 17
This correction factor therefore depends on and maxi- Vasconcelos. The first author thanks Matt Haney for
mum ODR only. We calculated this correction factor for pointing out that the nonhyperbolic moveout approx-
0 1 and maximum ODR between one and four, imation of Tsvankin and Thomsen (1994) resembles a
and presented a contour plot of C(, ODR). If one wants Pade approximation, thus leading him to study rational
to get the best use out of the Alkhalifah-Tsvankin non- approximants.
hyperbolic moveout equation (2), this plot can be used
in practice to determine the optimal correction factor in
the following way; first estimate and VN M O with the REFERENCES
nonhyperbolic moveout equation without the correction
factor, then use the estimated VN M O to estimate the Alkhalifah, T. 1997. Velocity analysis using nonhyperbolic
moveout in transversely isotropic media. Geophysics, 62,
maximum ODR using equation (14), and subsequently
18391854.
determine the optimal value of C from Figure 3 using Alkhalifah, T. 1998. Acoustic approximations for processing
the estimated k and . For maximum ODR up to four in transversely isotropic media. Geophysics, 63, 623631.
and large levels of anellipticity, tests with synthetic data Alkhalifah, T. 2000. An acoustic wave equation for
showed that this correction factor works well to reduce anisotropic media. Geophysics, 65, 12391250.
the bias in that would be obtained if the nonhyper- Alkhalifah, T., & Tsvankin, I. 1995. Velocity analysis for
bolic moveout equation without this correction factor transversely isotropic media. Geophysics, 60, 15501566.
was used. The field-data example confirmed this for Alkhalifah, T., Tsvankin, I., Larner, K., & Toldi, J. 1996.
maximum ODR up to two, but the presence of AVO Velocity analysis and imaging in transversely isotropic
media: methodology and a case study. The Leading Edge,
variation for offsets with maximum ODR between two
15, 371378.
and four caused this method to fail for ODR up to four.
Baker, G.A. Jr. 1975. Essentials of Pade Approximations.
For maximum ODR less than two, this reduction in bias Academic Press, Inc., New York.
goes together with largely accurate moveout correction, Baker, G.A. Jr., & Graves-Morris, P. 1981a. Pade Approxi-
even for strong anellipticity. For ODR up to four, how- mants, part I: Basic Theory. Addison-Wesley.
ever, the reduction in bias of goes together with signif- Baker, G.A. Jr., & Graves-Morris, P. 1981b. Pade Approx-
icant distortion in the moveout correction caused by the imants, part II: Extensions and Applications. Addison-
inaccuracy of the nonhyperbolic moveout equation. In Wesley.
this case the Alkhalifah-Tsvankin nonhyperbolic move- Coleman, T.F., & Li, Y. 1994. On the Convergence of Re-
out equation fails to accurately flatten the CMP gath- flective Newton Methods for Large-Scale Nonlinear Min-
ers. The rational interpolation method combines unbi- imization Subject to Bounds. Mathematical Program-
ming, 67, 189224.
ased estimates of and VN M O with accurate moveout
Coleman, T.F., & Li, Y. 1996. An Interior, Trust Region
correction in all cases. Approach for Nonlinear Minimization Subject to Bounds.
From synthetic tests and field data, we found that SIAM Journal on Optimization, 6, 418445.
straightforward application of the correction factor C = Dellinger, J., Muir, F., & Karrenback, M. 1993. Anelliptic
1.2, recommended by Grechka and Tsvankin (1998), approximations for TI media. J. Seis. Expl., 2, 2340.
typically leads to less than satisfactory results. For max- Fomel, S. 2004. On anelliptic approximations for qP veloci-
imum ODR=4 and . 0.1, the shortcomings in move- ties in VTI media. Geophysical Prospecting, 52, 247259.
out correction and estimated values of obtained using Fowler, P.J. 2003. Practical VTI approximations: a system-
the nonhyperbolic moveout equation approach without atic anatomy. Journal of Applied Geophysics, 54, 347
the correction factor are small and probably negligible 367.
Graves-Morris, P.R., & Hopkins, T.R. 1981. Reliable rational
from a practical point of view. Even though in this par-
interpolation. Num. Math, 36, 111128.
ticular case satisfactory results can be obtained with the Grechka, V., & Tsvankin, I. 1998. Feasibility of nonhyper-
Alkhalifah-Tsvankin nonhyperbolic moveout equation, bolic moveout inversion in transversely isotropic media.
straightforward application of the rational interpolation Geophysics, 63, 957969.
method combines unbiased parameter estimation with Helbig, K. 1983. Elliptical anisotropy - its significance and
accurate moveout correction, for arbitrary ODR ranges meaning. Geophysics, 48, 825832.
and arbitrary levels of anellipticity. Jones, L.E.A., & Wang, H.F. 1981. Ultrasonic velocities in
Cretaceous shales from the Williston Basin. Geophysics,
46, 288297.
Sayers, C.M. 1994. The elastic anisotropy of shales. J. Geo-
Acknowledgment
phys. Res., 99 (B1), 767774.
The authors thank Devon Energy and Mike Ammer- Schoenberg, M. A., & de Hoop, M.V. 2000. Approximate
man for permission to show these data and Scott Schap- dispersion relations for qP-qSV waves in transversely
isotropic media. Geophysics, 65, 919933.
per for generating the residual topography map shown
Stovas, A., & Ursin, B. 2004. New travel-time approxima-
in Figure 9c. The careful and critical reviews by Ken
tions for a transversely isotropic medium. J. Geophys.
Larner and Ilya Tsvankin are much appreciated, as well Eng., 1, 128133.
as the valuable discussions with Ilya Tsvankin, Ken
Larner, Vladimir Grechka, Pawan Dewangan, and Ivan
18 H. Douma, A. Calvert & E. Jenner
Thomsen, L. 1986. Weak elastic anisotropy. Geophysics, 51, expressions for the coefficients are given by
19541966.
Toldi, J., Alkhalifah, T., Berthet, P., Arnaud, J., Williamson, d1 = ((T0 T4 )X1 X2 X3 (T1 X1 (X2 X3 ) + . . .
P., & Conche, B. 1999. Case study of estimation of T3 X3 (X1 X2 ) + T2 X2 (X3 X1 )) . . .
anisotropy. The Leading Edge, 18, 588593.
((T1 T2 )(T0 T3 )X12 X22 + . . .
Tsvankin, I. 1996. P-wave signatures and notation for trans-
versely isotropic media: An overview. Geophysics, 61, ((T2 T0 )(T1 T3 )X12 + . . .
467483. (T0 T1 )(T2 T3 )X22 )X32 )X4 + . . .
Tsvankin, I. 2001. Seismic signatures and analysis of reflec-
tion data in anisotropic media. Elsevier Science Ltd. ((T0 T3 )X1 X2 (T1 X1 T2 X2 + . . .
Tsvankin, I., & Thomsen, L. 1994. Nonhyperbolic reflection T4 (X2 X1 )) ((T0 T2 )(T1 T4 )X12 . . .
moveout in anisotropic media. Geophysics, 59, 1290
(T0 T1 )(T2 T4 )X22 )X3 + . . .
1304.
Van der Baan, M. 2004. Processing of anisotropic data in (T3 T4 )(T1 X2 T2 X1 + . . .
the p domain: I-Geometric spreading and moveout T0 (X1 X2 ))X32 )X42 )/ . . .
corrections. Geophysics, 69, 719730.
Van der Baan, M., & Kendall, J. M. 2002. Estimating (X1 X2 X3 (T2 (T3 (X2 X3 )(X1 X4 ) . . .
anisotropy parameters and traveltimes in the -p domain. T4 (X1 X3 )(X2 X4)) + . . .
Geophysics, 67, 10761086.
T1 (T4 (X2 X3 )(X1 X4 ) . . .
Wookey, J., Van der Baan, M., Smit, D., & Kendall, J.-M.
2002. Tau-p domain VTI parameter inversions using T3 (X1 X3 )(X2 X4 ) + . . .
limited-offset data. Page F041 of: Extended Abstracts, T2 (X1 X2 )(X3 X4 )) + . . .
64th Meeting, European Association of Geoscientists and
Engineers. T3 T4 (X1 X2 )(X3 X4 ))X4 ) , (A2)
Zhang, F., & Uren, N. 2001. Approximate explicit ray veloc-
ity functions and travel times for P-waves in TI media. d2 = (T0 (X1 X2 )(X1 X3 )(X2 X3 ) + . . .
Pages 106109 of: Expanded abstracts of the 71 Ann. In- T3 X1 X2 (X1 X2 )(1 + d1 X3 ) + . . .
ternat. Mtg. Soc. of Expl. Geophys, San Antonio, Texas,
USA. X3 (T1 X2 (1 + d1 X1 )(X2 X3 ) + . . .
T2 X1 (1 + d1 X2 )(X3 X1 )))/ . . .
(X1 X2 X3 (T2 X2 (X1 X3 ) + . . .
APPENDIX A: [2/2] RATIONAL
T3 X3 (X2 X1 ) + T1 X1 (X3 X2 ))) , (A3)
INTERPOLATION FOR NONHYPERBOLIC
MOVEOUT IN A SINGLE HORIZONTAL n2 = T0 (X1 X2 ) + . . .
VTI LAYER T1 (1 + X1 (d1 + d2 X1 ))X2 . . .
Using the definition and normalization of the rational T2 X1 (1 + X2 (d1 + d2 X2 ))/ . . .
approximation outlined in the main text, we can write X1 X2 (X1 X2 ) , (A4)
the [2/2] rational approximation for squared traveltimes T0 T 1
T as a function of squared offset X as n1 = T 1 d1 (n2 d2 T1 )X1 . (A5)
X1
T0 + n 1 X + n 2 X 2 Note that only the expression for d1 is explicit in just
T (X) , (A1)
1 + d1 X + d2 X 2 Ti , Xi , and T0 , whereas d2 also depends on d1 , n2 on
d1,2 , and n1 on n2 and d1,2 . Calculating the coefficients
where T0 = t20 is the squared zero-offset two-way trav-
n1,2 and d1,2 using the above expressions and the in-
eltime, and n1,2 and d1,2 are the coefficients of the nu-
terpolation traveltimes ti and offsets xi , we can use the
merator and denominator of the rational approximant,
resulting values of n1,2 and d1,2 in equation (A1) to
respectively. Using four squared traveltimes Ti = t2i ,
evaluate interpolated traveltimes t for offsets x between
with i = 1, . . . , 4, and four accompanying squared off-
offsets xi .
sets Xi = x2i as interpolation points, we arrive at a
linear system of four equations with four unknowns, the
coefficients n1,2 and d1,2 . Since there are only four co-
efficients, we simply solve (using Mathematica) for the
coefficients in terms of Ti , Xi , and T0 . The resulting
CWP-499
ABSTRACT
First order internal multiples are a source of coherent noise in seismic images.
There are a number of techniques to estimate internal multiples in the data,
but few methods exist that estimate imaging artifacts caused by internal mul-
tiples. We propose a method to do this in which the artifacts are estimated
as part of the imaging process. Our technique is based on a hybrid of the
Lippmann-Schwinger scattering series and the generalized Bremmer coupling
series. Although we require knowledge of the velocity model this allows us to
estimate internal multiples without assumptions inherent to other methods.
Key words: internal multiple attenuation, Bremmer series, downward contin-
uation, imaging artifacts
d1 (s0 , r0 , t) = 1
D2 Q (0)Q,s0 (0)
4 t ,r0
Z
dz1 H(0, z)Q,r1 (z1 )Q,s1 (z1 )
0
(E1 E2 a)(z1 , s1 , r1 , t0 ) ,
(7)
Figure 2. Illustration of equation 13, the estimation of the
where
multiple at the depth z. The lines represent wavepaths rather
r+s than rays and do not connect as the imaging condition has
E1 : k(z, x) 7 (r s)k(z, 2
),
not been applied.
E2 : l(z, r, s) 7 (t)l(z, r, s).
We denote by a = 2c30 c the velocity contrast, in which
c0 denotes the smooth background velocity and c de-
notes the velocity contrast. Together the E1 and E2 op-
erators map the velocity contrast at depth to data at
depth. where M is the imaging operator discussed by Stolk &
de Hoop (Stolk & de Hoop, 2004a). We apply (11)-(12)
to (10), from which we obtain an image (first term on
3 INVERSE SCATTERING the right-hand side) minus artifacts (second term). We
compare our method with that of Weglein et al. (1997)
In inverse scattering the goal is to solve for Vb in terms
and ten Kroode (2002) in Appendix B of Malcolm & de
of the data d (cf. (6)). To this end, we assume that the
Hoop (2005).
contrast operator Vb can be written as a series
The removal of negative times in the construction
X
Vb = Vbm (d) , (8) in (11) is an important step in our algorithm. It can
mN also be computationally expensive to do this because
windowing in time requires returning to the time do-
where Vbm is of order m in the data. Substituting this main. It is also this step that, in practice, imposes a
equation into (6) leads to the following relation between limitation on the thickness of the multiple generating
the Vbm (d), layer from which we can estimate multiples. To be able
Dt2 M0 (Vbm U0 ) = Dt4 M0 (Vbm1 L0 (Vb1 U0 )) . (9) to attenuate internal multiples in a certain layer it is
necessary to be able to remove the primary from the
From this it follows that top of the layer without attenuating the reflection from
the bottom of the layer.
!
X
Dt2 M0 (Vb U0 ) = D Dt2 M0 (Vbm U ) . (10)
mN
midpoint (km)
5 10 15
0 (e)
d
(a)
?
(d) (b) (f )
a1 (z., ) d1 (z) d3 (z) a3 (z, .)
(c) (c)
depth (km)
(+)
5 d1 (z + z) d3 (z + z) .
(t)
d1 (z, s0 , r, ) d1 (z, s, r0 , )ds0 dr0 .
(13)
offset (km) Expression (13) is valid upon restricting d1 to t > 0 and
assumes a point source. If the point source assumption
0 0.5 1.0 1.5 is not satisfied, an estimate of the wavelet should be de-
0 convolved from the estimated multiple. Equation (13)
is illustrated in Figure 2. Note the similarity between
this expression and that used in the surface-related
multiple elimination (SRME) procedure of Fokkema &
van den Berg (1993) and Berkhout & Verschuur (1997).
An estimate of the artifacts in the image is also made
at the current depth in (f).
We then downward continue both the data and the
time (s)
time (s)
to generate the examples shown here uses a pseudo- 0
screen propagator with an implicit finite difference wide-
angle correction (Jin et al., 1998). In Figure 4 we show
d1 (z = 1.5), a single common-midpoint gather (cmp)
downward continued to the depth z = 1.5 km of the top 2
of the layer. The primary reflected from the top of the
layer is located around t = 0, the reflection from the
bottom of the layer at about t = 1 s and the first order 4
internal multiple at about t = 2 s.
We now estimate the multiples at depth using (13).
This requires restricting d1 to time t > 0. The proce-
dure removes the primary reflection from the current
depth (which theoretically arrives at t = 0), in this offset (km)
case 1.5 km, before doing the convolution. If this pro- -1.0 -0.5 0 0.5
cess is not done correctly and energy remains at t 0,
all subsequent primaries will be duplicated in the esti-
mated multiples section. In this model a simple time-
windowing procedure is sufficient, because the reflec-
-2
tions are far apart in time. In some situations, we find
a -p filter to be more effective. This is because we typ-
time (s)
-2 2
depth (km)
time (s)
0 4
2 6
4 8
is at zero time.)
To illustrate the ability of the method to estimate
multiples in more complicated velocity models, we add 4
a low-velocity lens to the model. The resulting veloc-
ity model is shown in Figure 7. The layer in this model
is the same as the layer in the previous model except 6
that it is 0.5 km deeper to allow more space for the
low-velocity lens. The lens is located in the center of
the model; it is circular with Gaussian velocity varia-
tions, a diameter of 600 m and a maximum contrast 8
of 2 km/s. The addition of the lens has a large influ-
ence on the recorded data. A shot record directly above
the lens is shown in Figure 8. Note the ringing, perhaps
caused by numerical dispersion in the data modeling,
that is particularly strong on the multiple. We use a Figure 6. On the top is the image with an artifact from a
double-square-root propagator that works in midpoint- first-order internal multiple at about 5.7 km depth. On the
offset coordinates rather than shot and receiver. To ac- bottom is the estimated artifact.
commodate this choice, we use a subset of the available
offsets so that each midpoint has the same number of
offsets. The data from a midpoint of 9.8 km are shown we have removed the effects of the lens, the example is
in Figure 9. The first arrival is highlighted in this fig- essentially the same as the flat case and the multiples
ure to show the triplications caused by the lens more are estimated accurately as is shown in Figure 11. Once
clearly. again, the multiple is relatively weak in the estimated
To estimate the multiple, we propagate the data image. This is because of the residual moveout on the
to 2 km, the top of the layer, and again show the cmp common image gather, which is shown in Figure 12.
at midpoint 9.8 km Figure 10 along with the estimated To illustrate the dependence of this method on the
multiples at this depth. Note that the caustic has been background velocity model, we perturb the velocity in
removed by the propagation through the lens and that this section to ascertain the influence of the velocity on
the multiple is accurately estimated. At this point, since the final result. In theory, from equations (14) and (15),
Multiple attenuation examples 25
2 6
depth (km)
time (s)
km/s
2
4 4
6 2
8 4
offset (km)
0 0.5 1.0 1.5
Figure 7. Velocity model, similar to the flat layered exam- 0
ple discussed previously, with the addition of a low-velocity
lens to demonstrate that the method works in laterally het-
erogeneous velocity models.
time (s)
receiver position (km)
0 5 10 15 20
0
1
2
time (s)
4
Figure 9. Common midpoint gather at 9.8 km, with only
the offsets used to compute the images shown later. Note the
triplications caused by the lens. Top: full gather. Bottom:
zoom of the primary reflection from the top layer.
depth (km)
time (s)
2 4
8
4
midpoint (km)
offset (km) 0 0.5 1.0
0 0.5 1.0 1.5 0
0
depth (km)
time (s)
4
2
8
4
depth (km)
depth (km)
5
6
midpoint (km)
ph (s/km) 0 0.5 1.0
0 0.05 0
0
depth (km)
depth (km)
5
6
2 2
depth (km)
depth (km)
4 4
6 6
8 8
2 2
depth (km)
4 depth (km) 4
6 6
8 8
Figure 14. In these images, a second lens has been added Figure 15. The lens was removed from the velocity model
beneath the layer to introduce a laterally varying velocity before generating these images. Because this perturbation is
perturbation. Top: Image with artifacts from internal mul- above the top of the layer, we expect this to have an impact
tiples. Bottom: Estimated artifacts from first-order internal on the estimated multiple. Note the change in the accuracy
multiples. of the estimate beneath the lens. Top: Image with artifacts
from first-order internal multiples. Bottom: Estimated arti-
facts from first-order internal multiples.
timated in a depth window from 2.5 to 4.2 km depth and
the -p filter was applied at 2.5 and 3.25 km. Thus we
expect to see the multiples from the top and bottom of This ringing does not appear in the estimated artifacts
the chalk layer. The estimated multiples are imaged at because these estimates are made using the primaries,
approximately the depth expected based on the velocity which arrive at earlier times and are thus less strongly
model. The multiple at about 4.5 km depth is imaged influenced by the ringing. To highlight the multiples as
at the depth expected for an internal multiple entirely much as possible, we have stacked only the smallest p-
within the chalk layer extending from approximately 3 values (i.e. the first few traces of the common image
to 4 km depth. The multiples at about 5.5 km are either gather) to make the image and we have low-passed fil-
peg-legs from the bottom of the chalk and the bottom of tered the image to remove as much of the high-frequency
layer 2 or internal multiples with both deeper scattering noise visible after the chalk layer as possible, without
points at the bottom of the layer 2. The artifacts caused damaging the imaged structure. In Figure 20 we com-
by these multiples in the image are not easy to see, how- pare our estimate to one made by ten Kroode (2005),
ever, as they are obscured by other ringing in the data. in which he estimates the multiples from the top of the
Multiple attenuation examples 29
2
2 6
depth (km)
4 chalk
depth (km)
layer 2
km/s
4
4 6
2
8
6
10
midpoint (km)
0 0.5 1.0
0
depth of the up-to-down reflection). The main compu-
tational cost of the algorithm comes from the propa-
2 gation of the data and the internal multiples. Because
two data sets are propagated (the data themselves and
depth (km)
-2 -2
time (s)
time (s)
0 0
2 2
4 4
Figure 18. Data downward continued to a depth of 2.5 km. Left: before the removal of the part of the data at t 0. Right:
after the removal of the part of the data at t 0 using a -p filter.
4 4
depth (km)
depth (km)
6 6
8 8
10 10
Figure 19. On the left is the image with an artifact from a first-order internal multiple at about 4.5 km depth. On the right
is the estimated artifact.
depth (km)
of interbed multiples. Pages 15271530 of: Expanded 6
Abstracts. Soc. Explor. Geophys.
Jin, Shengwen, Wu, Ru-Shan, & Peng, Chengbin. 1998.
Prestack depth migration using a hybrid pseudo-screen 8
propagator. SEG Technical Program Expanded Ab-
stracts, 17(1), 18191822.
Kelamis, P., Erickson, K., Burnstad, R., Clark, R., & Ver- 10
schuur, D. 2002. Data-driven internal multiple attenua-
tion - Applications and issues on land data. Soc. of Expl.
Geophys. Pages 20352038.
Lippmann, B. A. 1956. Rearrangement Collisions. Phys.
Rev., 102(1), 264268. Figure 20. This figure repeats the estimated artifacts from
Malcolm, A. E., & de Hoop, M. V. 2005. A method for inverse internal multiples shown in Figure 19, with arrows indicating
scattering based on the generalized Bremmer coupling locations at which multiples can be observed in the results
series. submitted to Inverse Problems, CWP-486. of ten Kroode (2005).
McMaken, H. 1986. On the convergence of the Bremmer
series for the Helmholtz equation in 2d. Wave Motion,
8, 277283.
Moses, H. E. 1956. Calculation of scattering potential from
reflection coefficients. Phys. Rev., 102, 559567.
Prosser, R. T. 1969. Formal solutions of inverse scattering
problems. Journal of Math. Phys., 10, 18191822.
Razavy, M. 1975. Determination of the wave velocity in an
inhomogeneous medium from reflection data. J. Acoust.
Soc. Am., 58, 956963.
Sava, Paul, & Guitton, Antoine. 2005. Multiple attenuation
in the image space. Geophysics, 70(1), V10V20.
Sloat, John. 1948. Identification of echo reflections. Geo-
physics, 13(1), 2735.
Stolk, C. C., & de Hoop, M. V. 2004a. Modeling of seismic
data in the downward continuation approach. submitted
to SIAM J. Appl. Math. CWP468P.
Stolk, C. C., & de Hoop, M. V. 2004b. Seismic inverse scat-
tering in the downward continuation approach. submitted
to SIAM J. Appl. Math. CWP469P.
ten Kroode, A. P. E. 2002. Prediction of internal multiples.
Wave Motion, 35, 315338.
ten Kroode, A. P. E. 2005. personal communication.
van Borselen, R. 2002. Data-driven interbed multiple re-
moval: Strategies and examples. In: Expanded Abstracts.
Society of Exploration Geophysicists.
van Stralen, M. J. N. 1997. Directional decomposition of
electromagnetic and acoustic wave-fields. Ph.D. thesis,
Delft University of Technology.
Verschuur, D. J., & Berkhout, A. 1997. Estimation of mul-
tiple scattering by iterative inversion, Part II: Practical
aspects and examples. Geophysics, 62(5), 15961611.
Weglein, A., Gasparotto, F. A., Carvalho, P. M., & Stolt,
R. H. 1997. An inverse-scattering series method for atten-
uating multiples in seismic reflection data. Geophysics,
62, 19751989.
32 A. Malcolm, M.V. de Hoop & H. Calandra
CWP-500P
ABSTRACT
Seismic interferometry yields the Greens function that accounts for wave prop-
agation between receivers by correlating the waves recorded at these receivers.
We present a derivation of this principle based on the method of stationary
phase. Although this derivation is applicable to simple media only, it provides
insight into the physical principle of seismic interferometry. In a homogeneous
medium with one horizontal reflector and without a free surface, the correla-
tion of the waves recorded at two receivers correctly gives both the direct wave
and the single-reflected waves. When more reflectors are present a product of
the single-reflected waves occurs in the cross correlation that leads to spurious
multiples when the waves are excited at the surface only. We give a heuris-
tic argument that these spurious multiples disappear when sources below the
reflectors are included. We also extend the derivation to a smoothly varying
inhomogeneous background medium.
Key words: siesmic interferometry, stationary phase, multiples
1 INTRODUCTION ble only for finite media that have a discrete frequency
spectrum. This requirement was relaxed in an alterna-
Traditionally, imaging techniques are based on the il-
tive derivation based on the representation theorem for
lumination of an object by a coherent source. In many
one-way wave propagation [Wapenaar et al., 2004], and
applications coherent sources are not available. Seismic
by using the general representation theorem [Weaver
interferometry is a technique in which the Greens func-
and Lobkis, 2004; Wapenaar, 2004]. Alternative, but
tion that describes the waves that propagate between
equivalent proofs of the emergence of the Greens func-
two receivers is extracted by computing the correlation
tion have been formulated using the principle of time-
of signals recorded at these two receivers. These signals
reversed imaging [Derode et al., 2003ab; Bakulin and
may have been excited by either coherent or incoher-
Calvert, 2004; Roux and Fink, 2003]. The relationship
ent sources. The advantages of this technique are that
between these approaches is shown by Wapenaar et al.
incoherent noise can be the source of the waves used
[2004] .
for imaging and that one can effectively use a wavefield
The reconstruction of the Greens function from
that is excited at one of the receivers, even though no
recordings of incoherent signals has been shown obser-
physical source exists at that location.
vationally using ultrasound [Weaver and Lbkis, 2001;
The first formulation of this technique is due to
Larose et al., 2004; Malcolm et al., 2004]. Seismic inter-
Claerbout [1968] , who used the phrase daylight imag-
ferometry has been used in helioseismology [Rickett and
ing because the daylight that we use in our vision also
Claerbout, 1999; Rickett and Claerbout, 2000], in explo-
provides an incoherent illumination of the objects that
ration seismology [Bakulin and Calvert, 2004; Schuster
we view. His derivation was applicable to layered media.
et al., 2004], in crustal seismology for the retrieval of
The emergence of the Greens function was subsequently
the surface wave Greens function [Campillo and Paul,
derived for general media using normal-mode theory
2003; Shapiro and Campillo, 2004; Shapiro et al., 2005],
[Lobkis and Weaver, 2001]. That derivation is applica-
34 R. Snieder, K. Wapenaar & K. Larner
and for extracting the response of buildings from an in- 2 ILLUMINATING THE SUBSURFACE
coherent excitation [Snieder, 2005]. FROM SOURCES AT THE SURFACE
The mechanism of seismic interferometry can be ex-
Consider the problem wherein many sources at the sur-
plained using the method of stationary phase [Snieder,
face z = 0 illuminate the subsurface. These sources can
2004a]. This is not surprising because the stationary
be either coherent or incoherent, and the may act either
phase approximation is the natural tool to account for
simultaneously or sequentially. The sources are placed
the destructive and constructive interference that forms
at locations rS = (x, y, 0) and have a source-time signal
the physical basis of seismic interferometry. The deriva-
SS (t) that corresponds in the frequency domain to the
tion of seismic interferometry with the principle of sta-
complex spectrum SS (). The Earth response that is
tionary phase is approximate, but, as shown by Roux
excited by these sources is recorded at two receivers at
et al. [2005] , the stationary phase integral can also be
locations rA = (xA , 0, zA ) and rB = (xB , 0, zB ), respec-
evaluated exactly using elliptical coordinates. This gives
tively. Without loss of generality, we align the x-axis of
exactly the same final result as the stationary phase ap-
the employed coordinate system with the horizontal sep-
proximation. The derivation of seismic interferometry
aration of the sources; hence in this coordinate system
based on stationary phase has also been used for waves
the y-coordinate of both sources vanishes.
in a waveguide [Sabra et al., 2005].
The source-time functions SS (t) may be impulsive,
The derivation of stationary phase is applicable for
but they might also correspond to functions with a more
simple media only where one can easily account for the
random character, as would be excited by, for example,
different rays that propagate through the media. In this
traffic noise in a land survey or wave noise at the sea-
sense the derivation based on stationary phase is less
surface during a marine survey. In the sequel we assume
general than derivations based on normal modes, rep-
that the source time functions for sources at rS and rS 0
resentation theorems, or time-reversed imaging. Despite
are uncorrelated when averaged over time and that the
this limitation, the derivation based on stationary phase
power spectrum of the source time functions is identical:
is useful because it sheds light on the physics that under-
Z Taver
lies seismic interferometry. The value of this derivation
is mostly didactic, but it also highlights sampling issues SS (t)SS 0 (t + )dt = SS 0 C( ) , (1)
0
and the generation of spurious multiples.
where Taver denotes the length of the time-averaging
Here we show that singly reflected waves that prop-
and C( ) the autocorrelation of the source time func-
agate between two receivers in the subsurface can cor-
tions. The autocorrelation is the Fourier transform of
rectly be reproduced by correlating the waves that have
the power spectrum. Since all sources are assumed to
been excited by uncorrelated sources at the surface and
have the same power spectrum, they have the same au-
are recorded at the two receivers. We derive this for the
tocorrelation as well.
simplest case for a homogeneous medium without a free
The source-time functions may have a differ-
surface and flat horizontal reflectors in the subsurface.
ent character in different imaging experiments. In
In section 2, we derive the general framework for il-
the virtual-source experiments of Bakulin and Calvert
lumination of the subsurface by incoherent sources, and
[2004] the shots do not overlap in time. The shots are
introduce the employed single-scattering model in sec-
recorded and processed one after the other, and the cross
tion 3. In section 4 we show how this leads to the re-
terms between the shots in expression (1) by definition
trieval of the direct wave that propagates between the
vanish. The situation is more complicated when truly
receivers, and, in section 5, we show that this proce-
incoherent sources illuminate the surface. Consider for
dure also correctly leads to the single-reflected wave that
example the seismic noise generated by waves at the
propagates between the receivers. The correlation of the
sea-surface. The seismic waves excited at different lo-
single-reflected waves leads to a contribution that is pro-
cations on average are uncorrelated. Snieder [Snieder,
portional to the square of the reflection coefficient. We
2004a] has shown that for random sources the ensemble
show in section 6 that this term is kinematically equiva-
average of expression (1) vanishes and that, in a single
lent to the direct wave that propagates between the re-
realization, the ratio of the cross term S 6= S 0 to the
ceivers. In section 7 we show a numerical example that
diagonal term S = S 0 is given by
illustrates the role of stationary phase in seismic inter-
r
ferometry. In section 8, we generalize the derivation to |cross terms| tcorr
the case of a layered medium with more than one reflec- , (2)
|diagonal terms| Nsources Taver
tor, and we show that the product of singly-reflected
waves from different reflectors gives a nonzero contribu- where Nsources denotes the number of sources and tcorr
tion to the cross-correlation. We refer to these terms as is the temporal width of the cross-correlation C( ). This
spurious multiples because these terms depend on the width is inversely proportional to the bandwidth f
product of reflection coefficients, just like real multiples of the sources. The product of the bandwidth and the
do. The spurious multiples, however, have arrival times averaging length Taver is equal to Nf reedom , the number
that differ from those of real multiples. of degrees of freedom in the data [Landau, 1967; Bucci
and Franceschetti, 1989]. Therefore the ratio of the cross
Spurious multiples in seismic interferometry of primaries 35
B L B
rB =(x B,0,zB ) rB =(x B,0,z B)
A LA
rA=(xA ,0, zA )
rA=(xA ,0, zA )
z=D
rRA rRB z=D
Figure 2. Definition of the geometric variables in the anal-
rA ysis of term 1.
rB
Since the sources are uncorrelated, as stated in expres-
Figure 1. The geometry of an imaging experiment with a sion (4), the cross-terms S 6= S 0 in this double sum
source at the surface and two receivers at rA and rB . The vanish; hence
mirror image of these receivers in the reflector is indicated X f ull
at the locations rA0 and rB 0 , respectively. CAB () = G (rA , rS )Gf ull (rB , rS ) |S()|2 .
S
(8)
terms to the diagonal terms in a single realization is When the sources are densely distributed along the sur-
equal to face, with n sources per unit surface area, theP sum over
sources can be replaced by an integration: S ( )
|cross terms| 1 R
p . (3) n ( )dx over the surface; this gives
|diagonal terms| Nsources Nf reedom Z
It follows from expression (2) that for random, in- CAB () = |S()|2 n Gf ull (rA , rS )Gf ull (rB , rS )dxdy ,
coherent, source-time functions the cross terms can be (9)
made arbitrarily small by increasing the length of the with x and y the coordinates of the source at the surface
time window over which the averaging is carried out. In as shown in figure 1.
the following we assume that the cross-terms S 6= S 0
can be ignored altogether. In that case, expression (1)
corresponds in the frequency domain to 3 A SINGLE-SCATTERING MODEL
2
SS ()SS 0 () = SS 0 |S()| . (4) In order to better understand the physics of seismic in-
This condition of uncorrelated sources was also used by terferometry, we illustrate this technique with a model
Wapenaar [2004] , who assumed that the sources are that consists of a single reflector with plane-wave re-
located in the deep subsurface. flection coefficient r that is embedded in a homoge-
We now consider correlation of the waves recorded neous medium. The Greens function in the homoge-
at two receivers for the special case of an acoustic neous medium is given by
medium. The waves recorded at the receivers A and B eikR
are given by G(R) = , (10)
4R
uA () = S Gf ull (rA , rS , )SS () ,
P
with the wavenumber given by k = /c, c the wave
(5) velocity, and R the distance of propagation. We pre-
uB () = S Gf ull (rB , rS , )SS () ,
P
sume that there is no free surface, so this model does
not include any multiply reflected waves. As shown in
with Gf ull the full Greens function which consists of the
figure 1, rRA denotes the reflection point of the wave
direct wave, primaries, and multiples. In the frequency
that propagates to rA . The full Greens function is the
domain the temporal correlation of the waves recorded
superposition of the direct wave and the single-reflected
is given by
wave:
CAB () = uA ()uB () , (6) Gf ull (rA , rS ) = G(|rA rS |)
where the asterisk denotes the complex conjugate. In- +r G(|rA rRA | + |rRA rS |) ,
serting equation (5) in the previous expression gives
X f ull Gf ull (rB , rS ) = G(|rB rS |)
CAB () = G (rA , rS )Gf ull (rB , rS 0 )SS ()SS 0 () . +r G(|rB rRB | + |rRB rS |) .
S,S 0 (11)
(7)
36 R. Snieder, K. Wapenaar & K. Larner
In this expression, we used that the reflected wave is given by the Greens function that accounts for the propagation
from the source to an image point of the receiver below the reflector. The image points of the receivers A and B are
indicated in figure 1 by r0A and r0B , respectively. As shown in that figure, for receiver A the total distance covered by
the reflected wave is |rA rRA | + |rRA rS |.
Inserting equation (11) into expression (9) gives an expression for the correlation, which consists of a sum of four
terms:
Z
CAB () = n |S()|2 G(|rA r|)G (|rB r|)dxdy
| {z }
T1
Z
+n |S()|2 r G(|rA r|)G (|rB rRB | + |rRB r|)dxdy
| {z }
T2
Z (12)
2
+n |S()| r G(|rA rRA | + |rRA r|)G (|rB r|)dxdy
| {z }
T3
Z
+n |S()|2 r2 G(|rA rRA | + |rRA r|)G (|rB rRB | + |rRB r|)dxdy .
| {z }
T4
Term 1 (T1) is the correlation of the direct waves that propagate to the two receivers; this term does not depend
on the reflection coefficient. Terms 2 and 3 are proportional to the reflection coefficient r, for this reason they can
be expected to account for the single-reflected waves in the Greens function that are extracted from the correlation.
Term 4 depends on r 2 . In the following, we analyze terms 1-4 in order to establish the connection between the
correlation and the Greens function for this simple wave propagation problem.
4 ANALYSIS OF TERM 1
The derivation shown in this section is similar to that in an earlier analysis [Snieder, 2004a]. Using the lengths L A
and LB , as defined in figure 2, and the Greens function (10), term 1 can be written as
Z
1 exp (ik (LA LB ))
T1 = dxdy . (13)
(4)2 LA LB
The integrand has an oscillatory character, but as we will see, the integrand has a stationary point. For this reason
we analyze this integral in the stationary phase approximation [Bleistein, 1984; Snieder, 2004b]. Referring to figure
2, the lengths LA,B are given by
q
LA,B = (x xA,B )2 + y 2 + zA,B
2
. (14)
The stationary point of the integrand follows by setting the partial x- and y-derivatives of L = L A LB equal
to zero. For the y-derivative this gives
L y y
0= = . (15)
y LA LB
This derivative vanishes for y = 0; hence the condition of stationarity with respect to y implies that the stationary
source point lies in the vertical plane of the receivers. The stationarity condition with respect to the x-coordinate
gives
L x xA x xB
0= = = sin A sin B , (16)
x LA LB
where the angles A and B are defined in figure 2. The phase thus is stationary when
A = B and y=0. (17)
The stationarity condition A = B is illustrated in figure 3; it implies that the stationary source point at the
surface is aligned with the line joining the two receivers. In these figures the receivers are at different depths.
Spurious multiples in seismic interferometry of primaries 37
Note that when the receivers are at the same depth z=0 rS
(zA = zB ) there is no stationary source position, except
LB
for sources infinitely far away. Any attenuation will sup-
rB
press the contribution of those sources.
Kinematically, expression (13) gives a contribution
at a lag-time that is equal to the time it takes for the R LA
rA
wave to propagate from receiver B to receiver A because
the wave that propagates along the path shown in fig-
ure 3 arrives at receiver A with a time delay |rA rB |/c z=D
compared to the wave that arrives at receiver B. It is
nontrivial that the evaluation of the integral in (13) Figure 3. Definition of the geometric variables for the sta-
tionary source position in term 1.
gives a contribution that is also dynamically equal to the
Greens function of the waves that propagate between
the receivers A and B. In the following we evaluate the Q
integral in the stationary phase approximation.
Evaluating the second derivatives of L = LA LB P P
while using (17) for the stationary point gives
rB Q
2L 2
zA z2 z2 1 z2 1
= 3
B
3
= A2
B
x2 LA LB LA LA L2B LB
(18)
1 1 rA
= cos 2 ,
LA LB
Figure 4. The relation between an element P P 0 along the
and
surface and the corresponding element QQ0 perpendicular to
2L L2A L2B 1 1 the receiver line.
= = . (19)
y 2 L3A L3B LA LB
In this example, and the following examples, the factor n |S()|2 of expression (12), this gives a total
2 L/xy = 0 at the stationary point, and the contribution that is equal to
two-dimensional stationary phase integral reduces to
n |S()|2 c G(R)
the product of two one-dimensional stationary phase T1 = . (22)
integrals over the x- and y-coordinates, respectively. 2 cos i
In the following, LA and LB are the path lengths This means that the contribution of term 1 to the
for the stationary source position as shown in figure 3. correlation is, in the frequency domain, proportional to
Note that in the geometry of figure 3, LA > LB so the Greens function of the direct wave that propagates
that L1 1
A LB < 0. Evaluating the integral (13) in the between the receivers. Note that this Greens function
stationary phase approximation thus gives is multiplied by the source density n at the surface; a
denser source distribution gives a stronger correlation
1 exp (ik (LA LB )) than does a less dense one. The Greens function is
T1 = also multiplied by the power spectrum |S()|2 of the
(4)2 LA LB
sources. This power spectrum can be measured, and
r one can correct for this term. In order to retrieve the
2 1
ei/4 Greens function from term 1, one needs to multiply
k
s
1 1 with i.
cos 2 (20) R Because of the employed Fourier transform,
LB LA f (t) = F () exp(it)d, this multiplication corre-
sponds to a differentiation in the time domain. This dif-
r
2 1 ferentiation corrects for the integration that is carried
ei/4 , out in the cross-correlation. This need to carry out the
k
r
1 1
differentiation was also noted in other formulations of
LB LA
seismic interferometry, e.g., [Lobkis and Weaver, 2001;
Using the relation k = /c, we can write this expression Snieder, 2004a; Weaver and Lobkis, 2004].
as The term cos in the denominator is an obliquity
c exp (ik (LA LB )) factor that corrects for the fact that the length element
T1 = , (21) QQ0 of figure 4 perpendicular to the ray corresponds to
8(i) cos LA L B
a line element P P 0 along the surface whose lengths that
The distance LA LB is equal to the receiver separation
are related by QQ0 = P P 0 / cos .
R shown in figure 3. With expression (10) and including
In the virtual-source experiment of Bakulin and
38 R. Snieder, K. Wapenaar & K. Larner
Calvert [2004], the sources at the surface were placed z=0 rS =(x,y,0)
in a line. In that case there is no integration over the y-
coordinate, and the latter terms in expression (20) that B L B
come from the y-integration are absent; in that case rB =(x B,0,z B)
r
in |S()|2 c 1 1
T 1line = G(R) . (23) L1
8 i cos LB LA rA=(xA ,0, zA )
Note the presence of the factor i and the term 1/ i.
L2 A
Correcting for these terms involves a Hilbert trans- z=D
form and a fractional derivative. These correction fac- rR=(x R,yR ,D)
tors are common in two-dimensional imaging experi-
ments [Yilmaz, 1987; Bleistein et al., 2001; Haney and Figure 5. Definition of the geometric variables in the anal-
Snieder, 2005]. Without these corrections the recon- ysis of term 3.
structed Greens function does not have the proper
phase and frequency dependence. More seriously, in con- the surface and the vertical, while vS is the velocity at
trast to equation (22), expression (23) depends explic- the intersection of these rays with the surface.
itly on the distances LA and LB . It turns out that when
the derivation leading to expression (23) is repeated us-
ing the Greens function in two dimensions, an expres-
5 ANALYSIS OF TERM 3
sion analogous to equation (22) is obtained. The pres-
ence of the fractional derivatives and the lengths LA The analysis of this term could be achieved by apply-
and LB is thus due to a mismatch between the dimen- ing the theory of the previous sections to receivers at
sionality of the physical space through which the waves the image points rA0 and rB0 of figure 1. Here we show
propagate (3D versus 2D) and the dimensionality of the explicitly that the cross-correlation correctly produces
source distribution (2D versus 1D). The derivation of the single-reflected waves. Using the lengths defined in
seismic interferometry by Roux and Fink [2003] is based figure 5 this term is given by
on wave propagation in three dimensions, while the em- Z
1 exp (ik (L1 + L2 LB ))
ployed sources are placed along a line. As shown by the T3 = dxdy . (25)
example of expression (23), this leads to a Greens func- (4)2 (L1 + L2 ) LB
tion that is kinematically correct, but whose amplitude Before we can analyze this expression we need the co-
and phase is not. ordinates of the reflection point rR because this deter-
A complex overburden located between the surface mines the lengths L1 and L2 . The condition that the
sources and the receivers in the subsurface acts as a reflection angle is equal to angle of incidence gives
diffusor of seismic waves because the waves radiated
(D zA ) x + DxA
from a single source arrive at a receiver from various di- xR = ,
2D zA
rections resulting from the scattering and multipathing (26)
that has occurred in the overburden. In the experiment (D zA ) y
of Bukulin and Calvert [2004] the shots were placed on a yR = .
2D zA
line at the surface, but the complex overburden in their
experiment created multipathing that provided a more Using this, the lengths L1 and L2 are given by
2
diffuse illumination of their receivers in the subsurface
D
that helps to focus the waves onto the virtual sources. L21 = (x xA )2
2D zA
The analysis of this section can be generalized for (27)
a heterogeneous medium in which the velocity is suf-
D
2
2 2
ficiently smooth to warrant the use of ray theory. We + y +D ,
2D zA
show in appendix A that term 1 is then given by
and
X n|S()|2 vS Gray (rA , rB ) 2
T1 = , (24)
2 cos i D zA
stat. points L22 = (x xA )2
2D zA
where Gray (rA , rB ) is the ray-geometric Greens func- (28)
2
tion for the waves recorded at rA that are generated by D zA 2 2
+ y + (D zA ) ,
a point source at rB . The summation in this expression 2D zA
is over all the stationary source points on the surface
while LB is given by expression (14).
z = 0. These points can be found by tracing rays from
The stationary points of the integral (25) follow
rA to rB and by extending these rays to the surface
z = 0. The angle is the angle between these rays at
Spurious multiples in seismic interferometry of primaries 39
xxA rS
z=0 z=0
rS =(x,y,0) LB
zA rB
2DzA rA L1 R1 L1
2 rA
DzA
L2
1
z=D L2 =R2
3 z=D
L2 rR
rA
Figure 7. Definition of the geometric variables for the sta-
Figure 6. The angles 1 , 2 , and 3 and their relation to tionary source position for the analysis of term 3.
the geometric variables for the reflected wave.
sin B ,
from the first partial derivatives of L = L1 + L2 LB : L
0== sin A sin B . (35)
L
D
2
y x
0= = The integrand thus is stationary when the source
y 2D zA L1
(29) position satisfies
2
D zA y y A = B and y=0. (36)
+ .
2D zA L2 LB
This condition is depicted in figure 7: waves radiated
Again, the stationary source position occurs for y = 0; from the stationary source position at the surface travel
it is located in the vertical plane of the receivers. The in a straight line from the source through receiver B via
condition for stationarity in the x-direction is a specular reflection to receiver A. Just as in the anal-
2
ysis of term 1, the time delay of this wave recorded at
L D x xA
0= = the two receivers is equal to the time it takes the wave
x 2D zA L1
to travel from receiver B via the reflector to receiver
D zA
2
x xA
x xB A. Thus the correlation is kinematically equal to the
+ . Greens function for the reflected waves. With the fol-
2D zA L2 LB
(30) lowing stationary phase evaluation of the integral (25),
we verify that the retrieved Greens function is also dy-
In order to interpret this last condition geometri- namically correct.
cally it is useful to relate the ratios in this expression From expression (27) we get at the stationary point
to the angle of incidence at the reflector. Referring to 2 L1
D
2 2
D
figure 6, the following identities hold: cos 1 = D/L1 , =
x2 2D zA L31
cos 2 = (DzA )/L2 , and cos 3 = (2DzA )/(L1 +L2 ).
Since these angles are all equal to the angle of incidence 2
D2 1
L1 L1
A of the reflected wave, we obtain: = = cos2 .
L1 + L 2L21 L1 (L1 + L2 )2
D D zA 2D zA (37)
cos A = = = . (31)
L1 L2 L1 + L 2 In the second identity we have used expression (34),
Also, since x xA = L1 sin 1 + L2 sin 2 , and since both while the last identity follows from equation (31). In a
angles are equal to A , similar way it follows that
x xA 2 L2 L2
sin A = . (32) = cos2 , (38)
L1 + L 2 x2 (L1 + L2 )2
Dividing this expression by the last identity of (31) gives and, using (18), we obtain for the curvature of LB ,
x xA
tan A = . (33) 2 LB 1
2D zA = cos2 . (39)
x2 LB
Finally, from (31), In the last expression, we used the stationary phase con-
D L1 D zA L2 dition = . Combining these results in the path dif-
= , = . (34)
2D zA L1 + L2 2D zA L1 + L 2 ference L = L1 + L2 LB gives
2L
Using expression (34) in (30), and using (32) to elim- 2 1 1
= cos . (40)
inate x xA , gives, with the relation (x xB )/LB = x2 L1 + L 2 LB
40 R. Snieder, K. Wapenaar & K. Larner
Differentiation of (27) gives z=0 rS =(x,y,0)
( 4
2 L1 D
(x xA )2
y 2 2D zA rB =(x B,0,z B)
) (41)
2
D
+ D2 /L31 . rA=(xA ,0, zA )
2D zA
A B
With expressions (33) and (34) this is equal to z=D
2 2 rRA rRB
2 L1
D 1 L1 ` 2
2
= tan + 1 . (42)
y L1 L1 L1 + L 2 Figure 8. Definition of the geometric variables in the anal-
ysis of term 4.
Using the identity D/L1 = cos this gives
2 L1 L1
= . (43) surface such that the direct wave from the source to re-
y 2 (L1 + L2 )2
ceiver A propagates along the same path as the wave
A similar analysis for L2 gives that travels from the source to the reflector, and ulti-
mately to receiver B.
2 L2 L2
= . (44) The Greens function in expression (47) is the
y 2 (L1 + L2 )2 causal Greens function, while its complex conjugate
This gives, for the curvature of L with respect to y, in equation (48) is the acausal Greens function. It is
known that seismic interferometry gives the superpo-
2L 1 1
= . (45) sition of the causal and the acausal Greens function
y 2 L1 + L 2 LB [Lobkis and Weaver, 2001; Derode et al., 2003a; Mal-
In these expressions it is understood that all lengths are colm et al., 2004; Derode et al., 2003b]. The causal
evaluated at the stationary point. Greens function can easily be retrieved from the cross-
The stationary phase evaluation of the integral correlation either by truncating the cross-correlation for
(25) can now be carried out. Keeping in mind that t < 0, or by averaging the cross-correlation for negative
(L1 + L2 )1 L1B < 0, and using the same steps as times and positive times.
in section 4, gives
1 exp(ik(L1 + L2 LB )) i/4 2
T3 = e 6 ANALYSIS OF TERM 4
(4)2 (L1 + L2 )LB
1 For the analysis of term 4, we carry out the stationary
2 1 1 phase analysis of the following integral:
.
k cos LB L1 + L 2 Z
1 exp (ikL)
(46) T4 = dxdy , (49)
As shown in figure 7, L1 LB = R1 , and L2 = R2 . (4)2 X
With the definition (10) for the Greens function this with
gives after taking the rn |S()|2 terms into account: L = |rA rRA | + |rRA rS |
n |S()|2 c G(R1 + R2 ) (50)
T3 = r . (47) |rB rRB | |rRB rS |
2 cos i
Note the resemblance with expression (22) for the con- and
tribution of term 1 that gives the direct wave that prop- X = (|rA rRA | + |rRA rS |)
agates between the receiver. Expression (47) shows that (51)
the contribution of term 3 leads to the singly-reflected (|rB rRB | + |rRB rS |)
wave that propagates from receiver B via the reflector
to receiver A. The same corrections must be applied to where all variables are defined in figure 8. The stationary
this term as to term 1 as discussed in section 4. point follows from setting the x- and y-derivatives of
The same analysis can be applied to term 2 of ex- the phase equal to zero. As in the previous sections,
pression (12), the final result is the complex conjugate the stationarity condition with respect to y leads to the
of expression (47) so that condition y = 0; this means again that the stationary
point lies in the vertical plane of the receivers. Using the
n |S()|2 c
G(R1 + R2 ) same steps that led to expression (35), one finds that the
T2 = r . (48)
2 cos i stationarity condition with respect to x is given by
The stationary point now lies at the location on the L
0= = sin A sin B , (52)
x
Spurious multiples in seismic interferometry of primaries 41
z=0 complex conjugate appears because the wave arrives at
rS receiver A before it hits receiver B.
rB
7 A NUMERICAL EXAMPLE
L1 rA R
We present a numerical example of the theory. For sim-
L 2B plicity we consider the theory in two dimensions. A re-
L2A flector with reflection coefficient r = 0.8 is located at
z=D
rR a depth 1500 m below the surface. This is not a small
reflection coefficient, but since there is only one reflec-
Figure 9. Definition of the geometric variables in the anal- tor and no free surface, this model does not generate
ysis of term 4 for the stationary source position. any multiple reflections, regardless of how large the re-
flection coefficient is. The wave velocity is c = 2000
m/s, and the receivers are located at rA = (0, 1000)
where the angles A and B are defined in figure 8. The m and rB = (300, 500) m, respectively. We used noise
point of stationary phase thus is defined by the condi- sources at the surface with a spacing x = 20 m., and a
tions Ricker wavelet with a dominant frequency of 50 Hz for
A = B and y=0. (53) the power spectrum |S()|2 of the noise.
The contributions of the sources at the surface to
This condition of stationary phase corresponds to the terms T1-T4 is shown in the left panel of figure 10,
the source position shown in figure 9. The stationary while the sum over all source positions is shown in the
source position launches two waves that, after reflec- right panel. The right panel panels shows four distinct
tion at the interface, both propagate along the line that arrivals. The arrivals T1 and T3 are causal, while the ar-
joins the receivers. Since both reflected waves are pro- rivals T2 and T4 are acausal. T1 is the strongest arrival,
portional to r, this contribution to the correlation is because it does not depend on the reflection coefficient.
proportional to r 2 . The correlation of the waves shown The arrivals T2 and T3 are weaker, because they are
in figure 9 is nonzero for a lag-time that is equal to the the singly reflected waves, while the arrival T4 is the
time it takes for the waves to propagate between the re- weakest because it varies as r 2 .
ceivers. Kinematically, this term can thus be expected Note that each of the arrivals in the right panel
to correspond to the Greens function of the direct wave of figure 10 corresponds to a stationary source point in
that propagates between the receivers. the left panel of that figure. The nonzero contribution
In order to carry out the stationary phase analy- of these arrivals is solely due to the stationary source
sis, the second derivatives of the phase is needed. These points, the sources placed at other locations give con-
derivatives follow from the expressions (37), (38), (43) tributions that interfere desctructively.
and (44). Using the lengths defined in figure 9, term 4 Figure 11 shows for term T3 a comparison between
is given in the stationary phase approximation by the exact waveform, computed with the 2D Greens
1 exp(ik(L1 + L2A L1 L2B )) i/4 2 function, that is shown with the solid line, and the term
T4 = e 3 obtained by summing the correlation over the sources
(4)2 (L1 + L2A )(L1 + L2B )
at the surface (shown with crosses). The waveform ob-
1 tained from seismic interferometry matches the exact
2 1 1
. waveform well. Note that the shown waveforms do not
k cos L1 + L2A L1 + L2B
look like a Ricker wavelet; as theory predicts they are
(54)
shifted over a phase angle equal to /4.
According to the geometry of figure 9, R = L2B L2A .
The left panel of figure 10 shows weak arrivals be-
With the definition (10) for the Greens function this
tween T1 and T4. These weak arrivals are due to end-
gives, after taking the r 2 n |S()|2 terms into account,
point contributions from the sum over the traces on
n |S()|2 c
the left panel of that figure, especially for the end-
G(R)
T4 = r2 . (55) points where the arrival time tends to a constant. In
2 cos i
the numerical example we tapered the contribution of
Apart from the r 2 -term and a complex conjugation traces near the endpoints of the source region. With-
of the Greens function, this term is similar to term 1 out this tapering these endpoint contributions are much
as given in expression (22). The r 2 -term arises because stronger. This is of importance for virtual source imag-
both of the waves that are reflected upward from the ing [Bakulin and Calvert, 2004] because a careless sum-
reflector are proportional to the reflection coefficient. mation over all source positions may lead to endpoint
The fact that this contribution is multiplied with r 2 contributions that could be confused with waves re-
is simply due to the fact that both interfering waves flected off reflectors in the subsurface.
that contribute to this term are proportional to r. The
42 R. Snieder, K. Wapenaar & K. Larner
-1.0
1 -1.0 1
0.8
T2 0.8
T2
0.6 0.6
-0.5 -0.5
0.4 T4 0.4
T4
0.2 0.2
00 00
0.2 0.2
0.4
0.5
T1 0.5
0.4
T1
0.6 0.6
t 0.8
t 0.8
T3 T3
1.013000 2000 1000 0 1000 2000 3000
1.0 11 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
-3000 -1500 0 1500 3000
x
Figure 10. Left panel, the contribution of sources at the surface to the terms T1-T4 as a function of the source position x. For
clarity only every fifth source position is shown. Right panel, the sum over all source positions at the surface.
12
x 10
than one reflector. The terms T2 and T3 in expression
10
(12) involve the cross term between the direct wave and
T3 the single-reflected waves. Since these terms are linear in
8
the reflection coefficients, one can retrieve the sum of all
6 the single reflected waves by summing the terms T2 and
4
T3 over the different reflectors. This means that in the
presence of more than one reflector, the cross terms T2
2
and T3 between the direct wave and the single-reflected
0 waves produce the full set of single reflections.
2
The term T4 in expression (12) contains the prod-
uct of the single reflected waves. This means that for
4
more
P than one reflector this term contains double sum
j,j 0 rj rj 0 ( ). This double sum can be split into the
6
8
0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79 0.8 0.81
terms j 0 = j and the terms j 6= j 0 :
0.72 0.74 0.76 0.78 0.80
t X
rj rj 0 ( ) =
X 2
rj ( ) +
X
rj rj 0 ( ) (56)
j,j 0 j j6=,j 0
Figure 11. Solid line, the exact arrival for term T3 com-
Analysis of the first term is identical to that of the term
puted with the 2D Greens function. The crosses indicate
the sum of the correlation for term T3 over all sources at T4 in section ??. This means that one can simply sum
the surface. This sum is tapered near the end of the source expression (55) over all reflectors in the subsurface.
region. The last contribution that needs to be accounted
for is the contribution of the second sum in the right
hand side of equation (56) to the term T4. We consider
8 THE CASE OF MORE THAN ONE two reflectors at depths D1 and D2 with reflection coef-
REFLECTOR ficients r1 and r2 , respectively. The derivation holds for
Up to this point the analysis has been based on the any pair of reflectors. The wave paths associated with
assumption of a single reflector in the subsurface. Sup- two different reflectors to term T4 is shown in figure 12.
pose there are more reflectors at depths Dj with re- The integrand in the term T4 of expression (12) now
flection coefficients rj . Assuming that the wave velocity contains a phase term exp ikL with
remains constant, the second term in each of equations L = L1
(A)
+ L2
(A)
L1
(B)
L2
(B)
, (57)
(11) needs to be replaced by a sum over all reflectors. In
expression (12), the term T1 contains the direct waves where these lengths are defined in figure 12. As before,
only; this term is not influenced by the presence of more the phase is stationary with respect to the y-coordinate
Spurious multiples in seismic interferometry of primaries 43
rS =(x,y,0) z=0 rS1
z=0
rB
rB =(x B,0,z B)
B (B) rA
rA=(xA ,0, zA ) L(B)
2 L1
(A)
L2 r1
z=D1 z=D1
A (A) r1
L1 r2
z=D2 z=D2
z=D1
The difference in the travel time of the waves that
propagate along the two trajectories shown in the solid
z=D2 lines is not equal to the time it takes to propagate be-
tween the receivers. The stationary phase contributions
Figure 13. The stationary source points rS1 and rS2 for the that correspond to these cross terms are nonzero. This
correlation of waves reflected from two different reflectors.
implies that the correlation of the waves reflected off dif-
The corresponding wave paths to the receivers are shown
ferent reflectors give stationary phase contributions that
with solid and dashed lines, respectively.
are proportional to r1 r2 . These contributions depend on
the product of reflection coefficients, hence the correla-
when y = 0. The condition that the phase is stationary tion of the single-reflected waves from different reflectors
with respect to x gives the condition gives spurious arrivals that dynamically are equivalent
L to peg-leg multiples that are reflected once from the free
0= = sin A sin B , (58) surface and twice from reflectors in the subsurface, be-
x
cause the peg-leg surface multiples are also proportional
where the angles A and B are defined in figure 12. This to r1 r2 .
follows from the derivation that led to the first term in This is a puzzling conclusion, since theory predicts
the right hand side of expression (35). The stationary that the full Greens function can be retrieved when the
phase condition for this term therefore gives sources are placed on a closed surface that surrounds
A = B , y=0. (59) the region of interest [Wapenaar et al., 2004; Wapenaar
et al., 2005]. The key point is that in the derivation of
The stationary phase condition (59) gives two sta- this paper the sources are placed at the upper surface
tionary source points rS1 and rS2 at the surface that are only. Let us consider what would happen if we also had
shown in figure 13. The wave paths shown in solid lines sources at a surface z = zm that is located below the
indicates the cross correlation of the waves that propa- reflectors as shown in figure 14.
gate along the following trajectories: rS1 reflector 1 The reflection response and transmission response
rB and rS1 reflector 2 rA , while the wave paths of the subsurface are not independent [Claerbout, 1968;
shown in dashed lines indicate the correlation of the Wapenaar et al., 2004]. This suggests that sources below
waves that propagate along the following trajectories: the reflectors are essential for the cancellation of the
rS2 reflector 1 rA and rS2 reflector 2 rB . cross-terms on singly-reflected waves, and we provide a
These wave paths act as the direct and exchange scat- heuristic argument that this is indeed the case.
tering events in quantum mechanics [Levy-Leblond and Consider the situation in figure 14 where sources
Balibur, 1990].
44 R. Snieder, K. Wapenaar & K. Larner
are present at the surface z = 0 above the reflectors, relevant because the primary reflections rather than the
and at the surface z = zm below the reflectors. The direct waves are used to image the subsurface.
points rS1 and rSm on these surfaces are the stationary Note that it is arbitrary how the cross-correlation
source points for the cross-terms that correspond to the is defined. When the cross-correlation is defined by
path indicated with solid lines and dashed lines, respec- CAB () = uA ()uB () instead of expression (6) the
tively. The waves excited at the surface z = 0 propagate roles of the causal and acausal Greens functions are
along the paths shown with solid lines, while the waves interchanged. Since in practice one extracts the causal
excited below the reflectors propagate along the paths Greens function from the correlation, this arbitrariness
shown in dashed lines. These wave paths coincide after in the definition of the correlation does not matter.
their first encounter with reflector 1, hence the contri- The four terms in expression (60) correspond to the
bution of waves radiated from the stationary point; rS1 waves that propagate along the four trajectories shown
and rSm to the cross-correlation is nonzero for the same in figure 15. The waves in the upper diagrams are the
delay time. The contribution of the waves excited at rS1 direct waves that propagate in opposite directions be-
is proportional to r1 r2 , while the contribution of the tween the two receivers. The waves in the bottom di-
waves excited at rSm is equal to r1 r2 , because the re- agrams are the single-reflected waves that propagate
flection coefficient of reflector 1 for a downward reflected in opposite directions between the receivers. These di-
waves is r1 rather than r1 . As shown in the examples agrams provide an illustration of why the correlation
in the previous sections, it does not matter how far the leads to the superposition of the causal and acausal
stationary point is removed from the surface. Therefore, Greens function.
the stationary points rS1 and rSm give contributions to In the derivation of this paper we did not explicitly
the cross-correlation that are equal, but have opposite account for the radiation pattern of the point source. It
sign. This means that the sum of the cross terms of the follows from the figure 15 (and figure A1 for a heteroge-
cross-correlation of these two stationary points gives a neous medium) that the paths that render the phase
vanishing contribution. of the correlation stationary correspond to rays that
In practical situations the sources are typically lo- propagate in the same direction to the two receivers.
cated at the surface z = 0 only; this is for example the This means that if the source does not radiate energy
case in the virtual-source experiment of Bakulin and isotropically, the two receivers are still illuminated with
Calvert [2004]. In that case the cross terms of waves re- the same source strength. Similarly, when the reflection
flected from different reflectors give a non-zero contribu- coefficient depends on the angle of incidence, the sta-
tion that is proportional to the product of reflection co- tionary phase approximation selects the reflection angle
efficients. These non-zero contributions are proportional at the angle of the reflected wave that propagates be-
to the product of reflection coefficients. Therefore, vir- tween the receivers A and B, as shown in figure 15.
tual source imaging introduces spurious multiples when When more reflectors are present, the contribution
the sources cannot be placed on a closed surface. of term 4 is proportional to rj rj 0 . When only sources
at the surface z = 0 are used, these cross terms lead
to spurious multiples that have the same strength as
9 DISCUSSION peg-leg multiples. These spurious multiples are not re-
moved by algorithms for the removal of surface-related
Inserting expression (22), (47), (48) and (55) into equa- multiples [Verschuur et al, 1992; van Borselen et al.,
tion (12) gives for one reflector the following total con- 1996; Dragoset and Jericevic, 1998] because kinemati-
tribution to the correlation: cally they do not correspond to peg-leg multiples.
n |S()|2 c G(R)
The analysis of this paper shows that the Greens
2 G(R)
CAB () = +r function is retrieved from the stationary phase con-
2 cos i i
(60) tribution from the integration (summation) over all
ff
sources. The sources far from the stationary point
G(R1 + R2 ) G(R1 + R2 )
+r +r . give an oscillatory contribution that averages to zero.
i i
When random noise is used as a source, these source
In the presence of more reflectors, one can sum the last in general are spread out over the surface. When
three terms over the different reflectors. The correlation man-made sources are used, however, one may limit
is a weighted average of the causal and acausal Greens these sources to the stationary phase region.
function for the direct wave and the singly-reflected
waves. Note that for the direct wave, the acausal Greens Acknowledgments: We appreciate the comments of
function is weighted by r 2 . This is because the contribu- John Stockwell, and Kurang Mehta.
tion of upgoing waves is r 2 times the contribution to the
Greens function from the downgoing waves. In practi-
cal applications of seismic interferometry in reflection
seismology, this contribution to the direct waves is not
Spurious multiples in seismic interferometry of primaries 45
Figure 15. The wave paths that correspond to the stationary contributions to the correlations for the causal direct wave from
term 1 (upper left panel), the acausal direct wave from term 4 (upper right panel), the acausal reflected wave from term 2
(lower left panel), the causal reflected wave from term 3 (lower left panel).
ABSTRACT
We compare two approaches for deriving the fact that the Greens function in
an arbitrary inhomogeneous open system can be obtained by cross-correlating
recordings of the wave field at two positions. One approach is based on physical
arguments, exploiting the principle of time reversal invariance of the acoustic
wave equation. The other approach is based on Rayleighs reciprocity theorem.
Using a unified notation for both approaches, we discuss their similarities and
differences.
Key words: seismic interferometry
20
10
c d
0
o x1 180
o
0 a A B b 10
I x3
20
30
C 0.67 0.68 0.69 0.7 0.71 0.72 0.73 0.74
0.68 0.7 0.72 0.74
S t
e f
o
90
Figure 1. Single diffractor (C) in a homogeneous model. Figure 3. Zoomed-in version of event c in Figure 2b. The
The receivers are at A and B. The integration is carried out solid line is the directly modelled wave field. The circles rep-
along the sources on the surface S. The main contributions resent the integration result of equation (10) (i.e., the sum
come from the stationary points a-d. The contributions from of the traces in Figure 2a). The dashed line represents the
stationary points e and f cancel. integration result of equation (11).
Roel Snieder
Center for Wave Phenomena, Colorado School of Mines, Golden CO 80401, email rsnieder@mines.edu
ABSTRACT
The motion of a building depends on the excitation, the coupling of the building
to the ground, and the mechanical properties of the building. We separate the
building response from the excitation and the ground coupling by deconvolving
the motion recorded at different levels in the building, and apply this to record-
ings of the motion in the Robert A. Millikan Library in Pasadena, California.
The waveforms obtained from deconvolution with the motion in the top floor
show a superposition of one upgoing and one downgoing wave. The waveforms
obtained by deconvolution with the motion in the basement can be formulated
either as a sum of upgoing and downgoing waves, or as a sum over normal modes.
Since these deconvolved waves for late time have a monochromatic character,
they are most easily analyzed with normal-mode theory. For this building we
estimate a shear velocity c = 322 m/s and a quality factor Q = 20. These values
explain both the propagating waves as well as the normal modes. We show for
this application of seismic interferometry that deconvolution of waveforms is
superior to correlation.
Key words: seismic interferometry, building response
N
E N
10
10
9
9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
B B
interval 1 interval 2
0 5 10 15 20 25 30 35 40 -1 0 1 2 3 4 5
time (s) time (s)
10
10
9
9
8
8
7
7
6
6
5
5
4
4
3 3
2 2
1 1
B B
Figure 5. The waveforms of figure 3 at the different floors Figure 6. The waveforms of figure 3 at the different floors
after deconvolution with the waves recorded at the top floor. after deconvolution with the waves recorded in the basement
using only part of the data of figure 3. The deconvolved waves
shown in thick lines are obtained by using only the data in
wave. There is little indication that these upgoing and interval 1 of figure 3, while the deconvolved waves shown in
downgoing waves are reflected within the building. The the thin lines are computed from the data in interval 2 of
figure 3.
reflection coefficients by the floors within the building
(Safak, 1999) therefore must be small. The reflection
coefficient for elastic waves by a floor in the building
depends on the product of the frequency and the mass
of the floor (Doyle, 1989; Safak, 1999). This means that
the absence of waves reflected off the floors in the build- 10
ing may be due to the relatively low frequencies in the 9
waveforms used in this study. In addition, the domi- 8
nant wavelength of the employed waves spans several 7
floors, this further suppresses reflections generated by 6
the individual floors because a medium with small-scale 5
variations can be treated as an effective medium that be- 4
haves like a homogeneous medium with properties that
3
are determined by the background velocity and the em-
2
bedded scatterers (Frisch, 1968; Keller & Karal, 1966;
1
Tatarskii & Gertsenshtein, 1963).
B
The deconvolved waveforms in the figures 4 and 5
are computed from the full waveforms shown in figure 3. -0.4 -0.2 0 0.2 0.4
It is, however, not necessary to use the full waveforms. time (s)
We have also deconvolved the signals using the time in-
Figure 7. The waveforms of figure 3 at the different floors
tervals 1 and 2 as shown in figure 3. Interval 1 straddles
after deconvolution with the waves recorded in the top floor
the S-wave arrival and is 4 s long, while interval 2 con- using only part of the data of figure 3. The deconvolved waves
tains the surface wave arrivals and has a duration of 25 shown in thick lines are obtained by using only the data in
s. Both intervals were padded with zeroes to a duration interval 1 of figure 3, while the deconvolved waves shown in
of 40 s. The signals deconvolved with the waves recorded the thin lines are computed from the data in interval 2 of
in the basement for each of the intervals are shown in figure 3.
figure 6. The thick line denotes the deconvolved wave-
forms from interval 1 while the thin line denotes the
deconvolved waves from interval 2. the S-wave ensures that sufficient low-frequency infor-
The similarity of the waves deconvolved over differ- mation is present to reproduce the resonance. Note also
ent time intervals is striking. Note how the deconvolved that the deconvolved waves from interval 2 are based on
waves from interval 1 display the resonance of the build- the surface wave signal. Nevertheless, these deconvolved
ing, despite the fact that these waves are based on the waves display the upward and downward propagating
impulsive S-wave arrival only. The broadband nature of waves early in the deconvolved signal.
Building response from seismic interferometry 59
The recorded waves in interval 2 are dominated by low-frequency surface waves. These waves visually mask the
higher frequency components in interval 2. The deconvolution equalizes the frequency content and therefore brings
out the high-frequency propagating waves in figure 6. Interval 1 is shorter than interval 2, and one might think that
interval 1 therefore contains less information than interval 2. Because of the impulsive character of the S-wave, the
waves in interval 1 have a larger bandwidth than the waves in interval 2. This larger bandwidth helps stabilize the
deconvolution. The similarity of the deconvolved waves for the intervals 1 and 2 shown in the the figures 6 and 7
implies that for the level of shaking used in this study the building responds linearly.
The waves deconvolved with the signal recorded at the top floor for interval 1 and interval 2 is shown in figure 7
with a thick and thin line, respectively. As in the preceding figure, these deconvolved waves are similar. This implies
that the S-wave and the surface wave both contain information about the upward and downward propagating waves
in the building. The deconvolution defined in equation (2) and the choice of are not optimized. A more careful
choice of the deconvolution algorithm could make the deconvolved waves from the intervals 1 and 2 even more similar.
The deconvolved waves behave in the same way as a hologram. A part of a hologram can be used to reconstruct
the image, albeit with a degraded resolution compared to the image of the full hologram (Lauterborn et al., 1995).
As shown in the figures 6 and 7, the deconvolved waves that are computed from different sub-intervals of the whole
signal lead to the same deconvolved waves.
+
With the wave number defined by
k = /c , (6)
60 R. Snieder
and for the attenuation model (3), this expression is in the frequency domain given by
n
n
u(z, ) = eik(2nH+z) e|k|(2nH+z)
P
n=0 S()R ()
o (7)
+ eik(2(n+1)Hz) e|k|(2(n+1)Hz) .
In this expression n counts the number of bounces off the base of the building. The first term denotes the upward
propagating waves, while the last term accounts for the downward propagating waves that have bounced n times in
the building.
The motion at height z deconvolved with the motion at the top floor is denoted by T (z, ), so that in the
frequency domain
u(z, )
T (z, ) . (8)
u(z = H, )
Similarly, the motion deconvolved with the motion at the bottom floor is denoted by B(z, ), hence
u(z, )
B(z, ) . (9)
u(z = 0, )
Let us first analyze T (). Inserting expression (7) in the numerator and denominator of the definition (8) gives
n o
n ik(2nH+z) |k|(2nH+z)
+ eik(2(n+1)Hz) e|k|(2(n+1)Hz)
P
n=0 S()R () e e
T (z, ) = . (10)
2
P n ik2(n+1)H e|k|2(n+1)H
n=0 S()R ()e
The difference with expression (15) is that the frequency-dependent denominator has disappeared. Note that since the
argument of each of the complex exponentials is positive, B(z, ) is a causal function. This deconvolved response is an
infinite sum of upgoing and downgoing attenuated waves. This sum differs from the sum of upgoing and downgoing
waves in the building, because B(z, ) does not depend on the ground coupling, whereas the original sum of upgoing
and downgoing waves (7) does depend on the ground coupling through the reflection coefficient R().
In expression (17) the reflection coefficient at the base of the building is equal to 1, because the wave that
has bounced n times off the base of the building is proportional to (1)n . There is a simple explanation for this
(Jon Sheiman, personal communication, 2004). The deconvolution of the motion of the basement with itself gives,
by definition, a bandpass-filtered delta function as shown in the bottom trace of figure 4. When the wave that has
reflected off the top of the building propagates downward, it must give a vanishing contribution at the base of the
building because the deconvolved wave at that level vanishes for t > 0. The motion at the base can only vanish
when an upward propagating wave is launched upward with the opposite polarity as the downward propagating wave
that strikes the basement. This corresponds to a reflection coefficient for the deconvolved waves that is equal to 1
rather than the reflection coefficient R() of the subsurface. It has been shown earlier that interferometric imaging
can be used to determine waveforms for the system with different boundary conditions than the physical boundary
conditions (Riley & Claerbout, 1976; Wapenaar et al., 2004). Riley and Claerbout (Riley & Claerbout, 1976) coined
the phrase Noahs deconvolution for this technique.
An alternative way to interpret B(z, ) is based on normal modes. Using the inverse Fourier transform, and
expression (6), the deconvolved response is in the time domain given by
Z i(tz/c) ||z/c
e e + ei(t(2Hz)/c) e||(2Hz)/c
B(z, t) = d . (18)
1 + e2iH/c e2||H
As shown in appendix A this integral can be solved by contour integration. The integrand has simple poles when
1 + e2iH/c e2||H = 0 , (19)
the location of the poles in the lower half-plane is shown in figure 8. For t > (2H z)/c, the contour must be closed
in the lower half plane, and as shown in appendix A the integral (18) can be written as a sum of damped normal
modes:
4c P m+1
B(z, t) = m=0 (1) exp( m t)
H
(20)
m (H z)
cos sin ( m t) ,
c
with
(m + 1/2) c
m = m = 0, 1, 2, (21)
H
It should be noted that these normal modes are not the normal modes of the building, because its normal modes
depend in general on the coupling to the ground. The normal modes in the sum (20) are independent of the reflection
coefficient R(), hence the normal modes in the deconvolved response depend on the properties of the building only.
This is consistent with the traveling wave formulation of expression (17), where the reflection coefficient for the
deconvolved wave is equal to 1 rather than the reflection coefficient R() of the subsurface.
Each term in the sum (20) is exponentially damping. The term with the fundamental mode (m = 0) has the
smallest damping. This means that for large times (t 2H/c) the fundamental mode dominates and hence
4c 0 (H z)
B(z, t) exp( 0 t) cos sin ( 0 t) , (22)
H c
with
c
0 = . (23)
2H
62 R. Snieder
The period that corresponds to this angular fre-
quency is given by
4H
T0 = . (24)
c
Note that this is the time needed to propagate up and
down the building twice. This period is determined by
the factor (1)m in expression (17). Because of this fac-
tor the wave changes polarity if it propagates up and
down the building once. If the wave travels up and down
the building twice and covers a distance 4H, the polar- ...... ......
ity changes twice and the reverberating wave reinforces
itself to form a resonance.
5 INTERPRETATION OF THE
DECONVOLVED WAVEFORMS
The theory of the previous section agrees with the de-
convolved waves in the figures 4 and 5. Let us first con-
sider the waves deconvolved with the waves at the top
Figure 8. The location of the poles in the complex -plane
floor as shown in figure 5. These deconvolved waves are and the contour integration that is used for t > (2H z)/c.
given by expression (12) that gives the superposition of
an acausal upgoing wave and a causal downgoing wave;
both waves are clearly visible in figure 5. Given the floor
spacing of 4.27 m (Clinton, 2004) these waves can be
used to estimate the shear velocity in the building. It
follows from eexpression (12) that if there is no attenu- 0.1
ation ( = 0), and if the data have infinite bandwidth,
that the deconvolution is in the time domain given by
a superposition of upward and downward propagating
time (s)
delta functions 0
ff
zH zH
T (z, t) = t + t+ .
c c
(25)
-0.1
(In deriving this result expression (6) is used.) The at-
tenuation and the finite bandwidth of the data cause
the broader pulses shown in figure 5. -40 -20 0 20 40
We measured the arrival time of the upward and distance (m)
downward propagating waves by picking the maximum Figure 9. The arrival times of the upgoing and downgo-
of these waves. These arrival times are shown in figure ing waves in figure 5. A negative distance/time corresponds
9. The distance is measured relative to the position of to the upgoing wave, a positive distance/time to the down-
the accelerometer at the top floor. For the upward prop- going wave. The travels times at the floors 1-3 are marked
agating wave this distance is given a negative value. For with solid circles. The solid line indicates the travel time pre-
the floors 4-10 the upward and downward propagating dicted for the shear velocity inferred from the normal mode
waves overlap. This may bias the travel time measure- measurements that give a velocity of 322 m/s.
ments. The travel times at these floors are indicated
with open squares. The travel time determined from the
the downward going wave has a consistently smaller am-
waves recorded in the basement may be biased by the
plitude than the upward propagating wave. The abso-
presence of the solid earth below the basement, these
lute value of the amplitude at different floors cannot
travel times are also indicated with open squares. De-
be compared with great accuracy, because the absolute
spite these reservations, the measurements in figure 9
amplitude is affected by the receiver coupling and other
display a fairly linear dependence of the travel time with
uncertainties. The ratio of the amplitude of the down-
distance, this indicates a constant shear velocity in the
going wave and the upgoing wave, however, does not
building.
depend on the receiver coupling. Figure 10 shows the
According to expression (12), the upward and
natural logarithm of the ratio of the downgoing wave
downward propagating waves both decay due to atten-
and the upgoing wave at each floor. The amplitude mea-
uation. This attenuation can be seen in figure 5 because
Building response from seismic interferometry 63
0
6
4
-0.1
2
0
-0.2
-2
-4
0 20 40 60 80 0 5 10 15 20
two-way distance (m) time (s)
Figure 10. The natural logarithm of the ratio of the ampli- Figure 11. The natural logarithm of the envelope of the
tudes of the upgoing and downgoing waves of figure 5 as a deconvolved waves in figure 4 after applying a bandpass filter
function of the two-way distance to the top of the building. with corner frequencies of 1 Hz and 3 Hz, respectively. For
The amplitude ratio for the floors 1-3 is indicated with solid clarity the floor number is added to each curve. The best-
circles. The log-amplitude ratio predicted by the attenuation fitting straight line to each curve is indicated with thick solid
of expression (28) is shown with the straight line. lines.
surements in the floors 4-10 and in the basement are of the resonance. Given that the height of the building
likely to be unreliable because of the interference of the measured to the basement is 47 m, this gives with ex-
upgoing and downgoing waves and the presence of the pression (24) a shear velocity of
solid earth below the basement, respectively. The am-
c = 322 m/s . (26)
plitude ratios at these levels are indicated with open
squares. The amplitude ratio for the floors 1-3 is indi- The travel time as a function of distance for this velocity
cated by solid circles and are most reliable. The two-way is indicated by the solid line in figure 9. The proximity of
distance is measured relative to the receiver at the top this travel time curve to the arrival times of the upward
floor. The scatter in the amplitude ratio is considerable and downward propagating waves, shows that the trav-
because the amplitude difference between the upgoing eling waves and the normal modes predict a shear veloc-
and downgoing waves is fairly small. In a taller build- ity that is similar. This provides a consistency check on
ing these amplitude differences would be larger, and the the analysis. A systematic difference between the veloc-
attenuation can be determined with greater accuracy. ity of the propagating waves and the normal modes can
According to expressions (17) and (20), the signals be due to dispersion caused by the internal structure in
deconvolved with the bottom floor can be seen either as the building, and to amplitude variations between floors
a superposition of upward and downward propagating that are ignored in expression (4) that forms the basis
waves, or as a sum of normal modes. The interpretation of the mathematical model of section 4.
in terms of propagating waves is most useful for the According to expression (22) the resonance decays
early part of the deconvolved waves in figure 4. In that with time due to anelastic attenuation. In order to quan-
figure, the upward and downward propagating waves tify the attenuation we bandpass filtered the decon-
are not as clear as in figure 5 for the waves deconvolved volved waves of figure 4 with a Butterworth filter with
with the signals at the top floor, because in figure 5 only cutoff frequencies of 1 and 3 Hz, respectively. This fil-
one upgoing wave and one downgoing wave are present, ter extracts the fundamental mode from the waveforms.
whereas according to expression (17) many upgoing and The natural logarithm of the envelope of the bandpass-
downgoing waves interfere with each other in figure 4. filtered waveforms is shown in figure 11. Since the res-
For this reason we analyze the waves deconvolved with onance is weak for the lowest floor, we used only the
the signal in the basement in figure 4 from the normal top 9 floors in the normal-mode analysis. We added the
mode point of view as formulated in expression (20). floor number to each curve in order to separate them
Since the fundamental mode is much stronger than the in the figure. Since only the slope depends on the at-
higher modes, we use the expressions (22) through (24) tenuation this offset does not affect the analysis. Note
in the following. that, apart from some fluctuations, the envelope of the
The amplitude spectrum of the deconvolved waves deconvolved waves decays with time. This contrasts the
of figure 4 averaged over all the floors has a pronounced original waveforms in figure 3 that do not decay with
peak at 1.72 Hz. This reflects the monochromatic nature time because the motion is continuously excited by the
64 R. Snieder
surface waves. The deconvolution extracts the decay of
the resonance with time, this makes it possible to mea-
sure the anelastic attenuation in the building. 10
Between 1.5 s and 14 s the logarithm of the envelope 9
decays linearly with time, this is consistent with the 8
exponential decay in expression (22). For later times 7
the resonance is of the same order of magnitude as the 6
ambient noise, and the exponential decay is not valid. 5
In order to determine the attenuation we fitted straight 4
lines to the curves for 1.5 s < t < 14 s. The least-squares
3
fit of the envelopes is shown by the solid lines in figure
2
11. The slopes are similar and the average slope is given
by 1
B
slope = 0.1321 0.0017 s1 . (27)
The error is determined by the standard deviation of -1 0 1 2 3 4 5
time (s)
the slope for the deconvolved waves at different floors.
According to the expressions (4) and (22) the slope is Figure 12. The waveforms of figure 3 at the different floors
equal to /2Q. For the resonant frequency of 1.72 Hz, after correlation with the waves recorded in the basement.
this gives
Q = 20.45 . (28) base of the building. Suppose that instead of the decon-
volution we had used the correlation, as is common in in-
This value of the attenuation can be compared with terferometric imaging (Lobkis & Weaver, 2001; Derode
the attenuation of the propagating waves shown in fig- et al., 2003; Snieder, 2004a; Wapenaar, 2004; Snieder,
ure 10. The propagating waves in figure 5 have a dom- 2005). In the frequency domain, the correlation of the
inant frequency of about f = 5Hz. The propagating waves recorded at height z with those in the basement
waves decay with distance as exp(f z/Qc). For the is given by
value of Q given above, and a velocity of 322 m/s, this
decay is shown by the solid line in figure 10. The vari- C(z, ) = u(z, )u (z = 0, ) . (29)
ability in the amplitude measurements in that figure is
When equation (7) is inserted in this expression, the re-
fairly large. For the lower three floors where the upgo-
sult contains the power spectrum |S()|2 of the excita-
ing and downgoing waves dont interfere, the attenua-
tion as well as products of the reflection coefficient R().
tion inferred from the resonance agrees will with ampli-
In contrast to this, the deconvolved waves of expressions
tude decay determined from the propagating waves as
(12) and (15) depends on neither of these quantities. It
indicated with the solid circles. The comparison of the
is instructive to consider the waveforms obtained from
attenuation from the normal modes and the propagat-
correlation with the signal in the basement as defined
ing waves provides a consistency check on the employed
by equation (29), these correlated waveforms are shown
model of wave propagation in the building.
in figure 12. This figure should be compared with figure
4 for the deconvolved waves. The deconvolved waves are
causal while the waveforms obtained by correlation are
6 DISCUSSION not. This is due to the fact that the waveforms com-
We have shown that the deconvolution of the motion puted by correlation depends on the power spectrum
recorded at different floors in the building is an effec- |S()|2 of the excitation. The multiplication with the
tive tool for extracting the building response. The de- power spectrum in the frequency domain corresponds in
convolution with respect to the signals recorded in the the time domain to a convolution with the autocorrela-
basement and the top floor provide complementary in- tion of the excitation. For the surface waves that excite
formation. The deconvolution with the signal recorded the building, this autocorrelation has a fairly long time
at the top floor gives a one upgoing and one downgoing duration. This leads to a-casual arrivals in the correlated
propagating wave that clearly are separated. The decon- waves of figure 12. Note that the waveforms computed
volution with the waveforms recorded in the basement from correlation show neither the upgoing and down-
provides information on the fundamental mode of the going waves nor the clear resonance of the deconvolved
building. waves of figure 4. This means that for this application
The deconvolved waves are independent of the ex- deconvolution is superior to correlation.
citation and of the ground coupling. This can be seen Expression (12) can be generalized for SH-waves
in expressions (12) and (15) that are independent of the in an arbitrary layered medium. In this case the de-
excitation S() and the reflection coefficient R() at the convolved waves T (z, ) are equal to the P11 -element
of the propagator matrix (Trampert et al., 1993). This
Building response from seismic interferometry 65
contrasts formulations of seismic interferometry based a nine-storey reinforced concrete building. Bull. Seismol.
on correlation where the Greens function is obtained Soc. Am., 68, 219229.
(Lobkis & Weaver, 2001; Derode et al., 2003; Snieder, Foutch, D.A., Luco, J.E., Trifunac, M.D., & Udwadia, M.E.
2004a; Wapenaar, 2004; Snieder, 2005). According to 1975. Full scale, three dimensional tests of structural de-
expression (7.43) of Aki and Richards (Aki & Richards, formation during forced excitation of a nine-storey rein-
forced concrete building. Pages 206215 of: Proceedings,
2002), the P11 -element of the propagator matrix for SH-
U.S. National Conference on Earthquake Engineering.
waves in a lossless homogeneous medium is given by Frisch, U. 1968. Wave Propagation in Random Media. Pages
1 ik(zH) 75198 of: Bharucha-Reid, A.T. (ed), Probabilistic meth-
P11 (z, H) = cos k(z H) = e + eik(Hz) . ods in applied mathematics. Academic Press.
2
(30) Hudson, D.E. 1962. Synchronized vibration generators for
Apart from terms that depend on the attenuation this dynamic tests of full-scale structures. Technical report,
expression is identical to equation (12). We can show Earthquake Engineering Research Laboratory, California
that this is also the case for a general layered medium Institute of Technology.
Keller, J.B., & Karal, F.C. 1966. Effective Dielectric Con-
that has internal reflections.
stant, Permeability, and Conductivity of a Random
The deconvolved waves can be used to estimate
Medium and the Velocity and Attenuation Coefficient of
the shear velocity and attenuation in the Millikan Coherent Waves. J. Math. Phys., 7, 661670.
Library. The waves deconvolved with the motion in Kuroiwa, J.H. 1967. Vibration Test of a Multistorey Build-
the top floor lead to clear upgoing and downgoing ing. Ph.D. thesis, California Institute of Technology.
waves. The velocity of propagation can be measured Lauterborn, W., Kurz, T., & Wiesenfeldt, M. 1995. Co-
from the arrival time of these waves, while the ratio herent Optics, Fundamentals and Applications. Berlin:
of the amplitude of the upgoing and downgoing waves Springer-Verlag.
constrains the attenuation. The waveforms obtained by Lobkis, O.I., & Weaver, R.L. 2001. On the emergence of the
deconvolution with the motion in the basement gives Greens function in the correlations of a diffuse field. J.
Acoust. Soc. Am., 110, 30113017.
the motion of the fundamental mode of the building.
Luco, J., Trifunac, M., & Wong, H. 1987. On the appar-
The frequency and temporal decay constrain the shear
ent change in dynamic behavior of a 9-story reinforced-
velocity and attenuation as well. As shown in figures concrete building. Bull. Seismol. Soc. Am., 77, 1961
9 and 10, these complementary pieces of information 1963.
are consistent. This shows that the deconvolution of Luco, J. E. 1986. Soil-structure interaction effects on forced
the motion in the building recorded at different levels vibration tests. Los Angeles, California: Technical Report
can successfully be used to eliminate the imprint of the Report 86-05, University of Southern California, Depart-
excitation and the ground coupling, and that the values ment of Civil Engineering.
of the shear velocity and attenuation from propagating Riley, D.C., & Claerbout, J.F. 1976. 2-D multiple reflections.
waves and from the fundamental mode are consistent. Geophysics, 41, 592620.
Safak, E. 1995. Detection and identification of soil-structure
interaction in buildings from vibration recordings. J.
Acknowledgments: We appreciate the insights
Struct. Eng., 121, 899906.
that we got from discussions with Jon Sheiman, Safak, E. 1999. Wave-propagation formulation of seismic re-
Rodney Calvert, Huub Douma, Ken Larner and Matt sponse of multistory buildings. J. Struct. Eng., 125, 426
Haney. 437.
Snieder, R. 2004a. Extracting the Greens function from the
correlation of coda waves: A derivation based on station-
REFERENCES ary phase. Phys. Rev. E, 69, 046610.
Snieder, R. 2004b. A Guided Tour of Mathematical Meth-
Aki, K., & Richards, P.G. 2002. Quantitative Seismology. ods for the Physical Sciences. 2nd edn. Cambridge, UK:
Second edn. Sausalito: Univ. Science Books. Cambridge Univ. Press.
Backus, M.M. 1959. Water reverberations their nature and Snieder, R. 2005. Spurious multiples in interferometric imag-
elimination. Geophysics, 24, 233261. ing of primaries. Geophysics, submitted.
Clinton, J.F. 2004. Modern digital seismology - Instrumen- Tatarskii, V.I., & Gertsenshtein, M.E. 1963. Propagation of
tation, and small amplitude studies in the engineering waves in a medium with strong fluctuations in refractive
world. Ph.D. thesis, California Institute of Technology. index. Soviet Phys. JETP, 17, 458463.
Derode, A., Larose, E., Campillo, M., & Fink, M. 2003. Trampert, J., Cara, M., & Frogneux, M. 1993. SH propaga-
How to estimate the Greens function for a heteroge- tor matrix and Qs estimates from borehole- and surface-
neous medium between two passive sensors? Application recorded earthquake data. Geophys. J. Int., 112, 290
to acoustic waves. Appl. Phys. Lett., 83, 30543056. 299.
Doyle, J.F. 1989. Wave propagation in structures: An FFT- Trifunac, M.D. 1972. Comparisons between ambient and
based spectral analysis methodology. New York: Springer. forced vibration experiments. Int. J. Earthquake Eng.
Foutch, D.A. 1976. A Study of the Vibrational Characteris- and Struct. Dyn., 1, 133150.
tics of Two Multistorey Buildings. Ph.D. thesis, Califor- Udwadia, F.E., & Trifunac, M.D. 1974. Time and ampli-
nia Institute of Technology. tude dependent response of structures. Int. J. Earth-
Foutch, D.A., & Jennings, P.C. 1978. Foundation response of quake Eng. and Struct. Dyn., 2, 359378.
66 R. Snieder
Wapenaar, K. 2004. Retrieving the elastodynamic Greens ets equals
function of an arbitrary inhomogeneous medium by cross
correlation. Phys. Rev. Lett., 93, 254301. cos (m (t z/c)) + cos ( m (t (2H z)/c))
Wapenaar, K., Thorbecke, J., & Dragonov, D. 2004. Rela-
tions between reflection and transmission responses of = 2 cos ( m H/c) cos ( m (H z)/c) cos( m t) (A6)
three-dimensional inhomogeneous media. Geophys. J.
Int., 156, 179194. 2 sin ( m H/c) cos ( m (H z)/c) sin( m t) .
Wapenaar, K., Fokkema, J., & Snieder, R. 2005. Retrieving
the Greens function by cross-correlation: a comparison According to expression (21), cos( m H/c) = 0 and
of approaches. Submitted to J. Acoust. Soc. Am. sin( m H/c) = (1)m , so that
cos ( m (t z/c)) + cos ( m (t (2H z)/c)) =
Roel Snieder
Center for Wave Phenomena, Colorado School of Mines, Golden CO 80401, email rsnieder@mines.edu
ABSTRACT
Coda waves are sensitive to changes in the subsurface because the strong scat-
tering that generates these waves causes coda waves to repeatedly sample a
limited region of space. Coda wave interferometry is a technique that exploits
this sensitivity to estimate weak changes in the medium from a comparison of
the coda waves before and after the perturbation. Here I present the general
theory of coda wave interferometry, and show how the time-shifted correlation
coefficient can be used to estimate the mean and variance of the travel time per-
turbation caused by the perturbation of the medium. This mean and average
are defined based on the intensity of the coda waves. I show how this general
theory can be used to estimate changes in the wave velocity, in the location of
scatterer positions, and in the source location.
Key words: coda wave interferometry
t + P = ds = 2 ds . (42)
P v + v P v v where and are the velocities of P -waves and S-
waves, respectively. Since < the perturbation in
With expression (41) this gives
the shear velocity dominates the perturbation in the
P -wave
velocity. For example, for a a Poisson medium
Z Z
1 v v 1
P = ds = ds , (43)
P v v v P v
( = 3):
where I used in the last identity that the relative velocity v
= 0.09 + 0.91 . (49)
perturbation is assumed to be constant. With expression v
(41) the travel time perturbation can be written as The theory of this section is valid when the rela-
tive velocity change is independent of location. In re-
v
P = t. (44) alistic situations this is not necessarily the case. The
v
theory can be extended for situations where the rela-
Note that the velocity v is not necessarily constant. Ac- tive velocity change depends on position. In that case
cording to expression (44), the travel time perturbation the mean travel time change is linearly related to the
depends on the arrival time of the wave only, but is in- velocity change:
dependent of the particular path. This means that in a Z
small time window the mean travel time perturbation h i = K(r, t)v(r)dV , (50)
is also given by
v where the kernel K(r, t) depends on the intensity-
tmax = h i = t. (45) weighted average over all scattering paths. This expres-
v
sion can be used as the basis of a standard linear inver-
Since the travel time perturbation is the same for all sion for the velocity change v(r) given the mean travel
trajectories, the variance of the travel time perturbation time change observed for different source-receiver pairs
vanishes and different time windows. The kernel K(r, t) has been
2 = 0 . (46) derived both for single-scattered waves [Pacheco and
Snieder, 2005a] as well as for strongly scattered waves
As shown in expression (36), the mean travel time per- [Pacheco and Snieder, 2005b].
turbation is equal to the shift time that gives the max-
imum of the time-shifted correlation coefficient, this
quantity can easily be retrieved from the data. Accord- 6.2 Random displacement of scatterers
ing to expression (45) the relative velocity change then As a second example we consider a perturbation that
follows from consists of uncorrelated movement of the scatterers.
v tmax This is of relevance for studying the motion of particles
= . (47)
v t in colloidal suspensions [Heckmeier and Maret, 1997]
and of bubbles in a turbulent fluid [Cowan et al., 2000;
This has been applied by Snieder et al. [2002] to
Page et al., 2000]. The theory presented here is equiva-
measure the velocity change in a granite sample with
lent to diffusing wave spectroscopy [Cowan at al., 2002;
Coda wave interferometry theory 75
Weitz and Pine, 1993], although the derivation is differ- 6.3 A displaced source position
ent.
The relative distance between earthquakes can be found
Let us consider scatterers that move independently
from the absolute location of the events, but in this ap-
in three dimensions and that have a root mean square
proach errors in the employed velocity model may lead
displacement between the two measurements of the
to large errors in the distance between the events [Pavlis,
waves that are used to study the motion of the scatter-
1992] . An alternative approach is the double-difference
ers. On average the path length for each scattered wave
method where the relative event location is computed
does not change, hence the mean perturbation of the
from the differential travel time of the direct arrivals
travel time vanishes:
[Shearer, 1997; Astiz and Shearer, 2000; Walhauser and
h i = 0 . (51) Ellsworth, 2002] . Coda wave interferometry can also be
used to determine the relative distance between earth-
Some scattering paths are longer, while other scattering
quakes, provided the source mechanisms of the events
paths are shorter, therefore the variance of the travel
are identical. This provides additional information on
time perturbation is nonzero.
the relative position between events that can be used in
We compute the variance of the travel time by using
addition to the constraints obtained from the double-
that for a wave that has scattered n times, the variance
difference method.
in the path length is given by [Snieder and Scales, 1998]:
Let us consider two nearby seismic sources with the
2L = 2n 1 cos 2 . same source mechanism. When the source position is
`
(52)
perturbed over the distance r, the distance from the
In this expression cos is the average of the cosine of source to the first scatterer along every scatterer path
the scattering angle over all paths in the employed time is perturbed. The distance between the scatterers is un-
window. The number of scatterers encountered is related perturbed, so we only need to account for the change
to the travel time by n = vt/l, where v is the wave in the distance to the first scatterer along every scatter-
velocity and l the scattering mean free path. Using this ing path. For this reason it does not matter if the wave
in expression (52) gives travels from the first scatterer along every path directly
2vt 2 to the receiver, or whether the wave visits many other
2L = , (53) scatterers first. The theory of this section therefore is
l
applicable both for single scattering as well as for mul-
where l is the transport mean free path. This is the tiple scattering.
distance of propagation over which the scattered wave For a scattering path P with take-off direction tP at
has lost all information about its direction of propaga- the source, the perturbation in the travel time is given
tion [Lagendijk and van Tiggelen, 1996; van Rossum and by
Nieuwenhuizen, 1999]. For a constant velocity the vari- `
ance in the path travel time is related to the variance tP r
P = . (56)
in the path length by = L /v, so that v
2t 2 For a scattering path where the wave leaves the source
2 = , (54) as a P -wave, v is the P -wave velocity at the source loca-
vl
tion, whereas for a path where the wave leaves the source
With expression (39) this means that the measured as an S-wave, v denotes the shear velocity. When the
maximum of the cross-correlation is related to the mean scatterers are distributed homogeneously, some paths
displacement of the scatterers by are longer when the source location is perturbed, while
2 2 t others are shorter. The resulting mean travel time per-
Rmax = 1 . (55) turbation vanishes [Snieder and Vrijlandt, 2005] . This
vl
can be shown by integrating the travel time perturba-
Given a measured value of the maximum of the tion (56) over all take-off directions.
cross-correlation, one can infer the mean scatterer dis- The variance of the travel time perturbation, how-
placement from this expression if the wave velocity ever, is nonzero, because some paths are longer when
and the transport mean free path is known. Snieder et the source location is perturbed while others are shorter.
al. [2002] show in a numerical experiment that the mean The variance of the travel time perturbation depends on
scatterer displacement can correctly be retrieved from the type of source (explosion, point force, double cou-
the coda waves. The scatterer displacement can be com- ple), as well as on the orientation of the perturbation
puted from several non-overlapping time windows of the of the source location relative to the source mechanism
coda, and they show how this redundancy can be used [Snieder and Vrijlandt, 2005]. An important application
to compute error bounds on the scatterer displacement is the location of aftershocks. In that case all events are
inferred from the coda waves. located in the plane of the fault-plane of the main shock.
Snieder and Vrijlandt [2005] show that in this case the
76 R. Snieder
variance of the travel time is given by in the maximum of the correlation coefficient, this bias
6 7
can be removed using the correction factor of expression
+ 8 (40).
8
2 = r2 . (57) The present theory can be applied to a constant
2 3
7 + 6 change in the velocity, to uncorrelated perturbations
6 in the locations of the scatterers, and to changes in
In this expression is the P -wave velocity, and de- the source position. For a change in the velocity (sec-
notes the shear-wave velocity. This expression depends tion 6.1), the means travel time perturbation is nonzero
on both velocities, because the path summation includes (equation (45)), but the variance of the travel time per-
waves that leave the source as a P -wave, as well as waves turbation is zero (expression (46)). For random per-
that leave the source as an S-wave. The P - and S-wave turbations in the scatterer location the mean travel
velocities in expression (57) are raised to high powers. time perturbation vanishes (expression (51)), but ac-
Since < , the terms with the shear velocity dominate. cording to equation (52) the variance of the travel time
Note that expression (57) depends on the distance r be- is nonzero and depends linearly on time. This con-
tween the events, but not on their relative orientation trasts the case of a perturbation in the source position
in the fault plane. The reason for this is that the av- where the mean travel time perturbation also vanishes
eraging (38) in this application involves an integration and where according to expression (57) the variance is
over all take-off directions. In this directional averaging nonzero and independent of time. In general, it may not
information on the direction of the event separation is be obvious how a medium is perturbed. As shown above,
lost. the three different perturbation leave a different imprint
Expression (57) can be used to estimate the source on the mean and the variance of the travel time per-
separation in the following way. For several time win- turbation. Since these quantities can be estimated for
dows in the coda the time-shifted correlation coefficient several independent windows of the coda waves using
can be computed from expression (7). Using expres- the time-shifted correlation coefficient, it is possible to
sion (39), the maximum of this function can be equated discriminate between these different perturbations using
to the variance of the travel time perturbation given the recorded coda waves.
by equation (57). The resulting expression can then be
solved for the event separation r. In the presence of sig-
nificant noise, the correction factor of equation (40) can REFERENCES
be used to eliminate the bias in the cross-correlation due
to noise. K. Aki. Theory of earthquake prediction with special ref-
erence to monitoring of the quality factor of lithosphere
Snieder and Vrijlandt [2005] applied this technique
by the coda method. Earthquake Res. Bull., 3:219230,
to events on the Hayward Fault, California, and showed 1985.
that the event separation obtained from the coda waves K. Aki and L.B. Chouet. Origin of coda waves: source, atten-
agrees with the event separation of the same events de- uation, and scattering effects. J. Geophys. Res., 80:3322
termined by Waldhauser and Ellsworth [2002] with the 3342, 1975.
double-difference method. L. Astiz and P.M. Shearer. Earthquake locations in the in-
Note that the estimation of the event separation ner Continental Borderland, offshore southern California.
can be carried out with a single station. An error esti- Bull. Seism. Soc. Am., 90:425449, 2000.
mate of the event separation can be obtained by com- O.M. Bucci and G Franceschetti. On the degrees of free-
paring the event separation inferred from different non- dom of scattered fields. IEEE Trans. on Antennas and
Propagation, 37:918926, 1989.
overlapping time windows in the coda.
M.L. Cowan, I.P. Jones, J.H. Page, and D.A. Weitz. Diffusing
acoustic wave spectroscopy. Phys. Rev. E, 65:066605
1/11, 2002.
7 DISCUSSION M.L. Cowan, J.H. Page, and D.A. Weitz. Velocity fluctua-
tions in fluidized suspensions probed by ultrasonic corre-
The main result of the theory of this paper is that the lation spectroscopy. Phys. Rev. Lett., 85:453456, 2000.
time-shifted correlation coefficient of expression (7) can H. Douma and R. Snieder. Correcting for bias due to noise
be related to the mean and variance of the travel time in coda wave interferometry. Submitted to Geophys. J.
perturbation. According to expression (36), the mean Int., 2005.
travel time perturbation follows from the shift time that A. Gret, R. Snieder, R.C. Aster, and P.R. Kyle. Monitor-
gives the maximum of the time-shifted correlation coef- ing rapid temporal changes in a volcano with coda wave
ficient, and equation (39) relates the maximum of the interferometry. Geophys. Res. Lett., in press, 2005.
cross-correlation coefficient to the variance of the travel A. Gret, R. Snieder, and J. Scales. Time-lapse monitoring of
rock properties with coda wave interferometry. Submitted
time perturbation. This means that the mean and vari-
to J. Geophys. Res., 2005.
ance of the travel time perturbation can be obtained
R. Gret, A.and Snieder and U. Ozbay. Monitoring stress
from a comparison of the coda waves before and after change in an undergraound mining environment with
the perturbation. Additive random noise leads to a bias
Coda wave interferometry theory 77
coda wave interferometry. Submitted to the J. of Mining velocity. Science, 295:22532255, 2002.
Applications, 2004. R. Snieder and J.A. Scales. Time reversed imaging as a diag-
M. Heckmeier and G. Maret. Dark speckle imaging of nostic of wave and particle chaos. Phys. Rev. E, 58:5668
colloidal suspensions in multiple light scattering media. 5675, 1998.
Progr. Colloid. Polym. Sci., 104:1216, 1997. M.C.W. van Rossum and Th.M. Nieuwenhuizen. Multiple
A. Jin and K. Aki. Temporal changes in the coda Q be- scattering of classical waves: microscopy, mesoscopy and
fore the Tangshan earthquake of 1976 and the Haicheng diffusion. Rev. Mod. Phys., 71:313371, 1999.
earthquake of 1975. J. Geophys. Res., 91:665673, 1986. B.A. van Tiggelen. Localization of waves. In J.P. Fouque, edi-
A. Lagendijk and B.A. van Tiggelen. Resonant multiple scat- tor, Diffuse waves in complex media, pages 160. Kluwer,
tering of light. Phys. Rep., 270:143215, 1996. Dordrecht, 1999.
H.J. Landau. Necessary density conditions for sampling and F. Waldhauser and W.L. Ellsworth. Fault structure and me-
interpolation of certain entire functions. Acta Mathemat- chanics of the Hayward Fault, California, from double-
ica, 117:3752, 1967. difference earthquake locations. J. Geophys. Res.,
W. Lauterborn, T. Kurz, and M. Wiesenfeldt. Coherent Op- 107:10.1029/2000JB000084, 2002.
tics, Fundamentals and Applications. Springer-Verlag, D.A. Weitz and D.J. Pine. Diffusing wave spectroscopy. In
Berlin, 1995. W. Brown, editor, Dynamic light scattering, The method
S. Matsumoto, K. Obra, K. Yoshimoto, T. Saito, A. Ito, and and some applications, pages 652720. Clarendon Press,
A. Hasegawa. Temporal changes in P -wave scatterer dis- Oxford, 1993.
tribution associated with the M6.1 earthquake near Iwate
volcano, northeastern Japan. Geophys. J. Int., 145:48
58, 2001.
C. Pacheco and R. Snieder. Localizing time-lapse changes
with multiply scattered waves. Submitted to J. Acoust.
Soc. Am., 2005.
C. Pacheco and R. Snieder. Localizing time-lapse changes
with single scattered waves. Submitted to Geophys. J.
Int., 2005.
J.H. Page, M.L. Cowan, and D.A. Weitz. Diffusing acoustic
wave spectroscopy of fluidized suspensions. Physica B,
279:130133, 2000.
G. Pavlis. Appraising relative eartquake location errors. Bull.
Seism. Soc. Am., 82:836859, 1992.
G. Poupinet, W.L. Ellsworth, and J. Frechet. Monitoring ve-
locity variations in the crust using earthquake doublets:
an application to the Calaveras Fault, California. J. Geo-
phys. Res., 89:57195731, 1984.
A. Ratdomopurbo and G. Poupinet. Monitoring a temporal
change of seismic velocity in a volcano: Application to
the 1992 eruption of Mt. Merapi (Indonesia). Geophys.
Res. Lett., 22:775778, 1995.
P.M. Roberts, W. Scott Phillips, and M.C. Fehler. Develop-
ment of the active doublet method for measuring small
velocity and attenuation changes in solids. J. Acoust.
Soc. Am., 91:32913302, 1992.
P.M. Shearer. Improving local earthquake locations using the
L1 -norm and waveform cross-correlation: Application to
the Whittier Narrows California aftershock sequence. J.
Geophys. Res., 102:82698283, 1997.
M. Snieder and M. Vrijlandt. Constraining relative source
locations with coda wave interferometry: Theory and ap-
plication to earthquake doublets in the Hayward Fault,
California. J. Geophys. Res., in press, 2005.
R. Snieder. Imaging and averaging in complex media. In J.P.
Fouque, editor, Diffuse waves in complex media, pages
405454. Kluwer, Dordrecht, 1999.
R. Snieder. Coda wave interferometry and the equilibration
of energy in elastic media. Phys. Rev. E, 66:0466151,8,
2002.
R. Snieder. Extracting the Greens function from the cor-
relation of coda waves: A derivation based on stationary
phase. Phys. Rev. E, 69:046610, 2004.
R. Snieder, A. Gret, H. Douma, and J. Scales. Coda wave in-
terferometry for estimating nonlinear behavior in seismic
78 R. Snieder
CWP-504
ABSTRACT
Observing the azimuthally varying seismic attenuation in data that show az-
imuthal velocity anisotropy could contribute not only to the interpretation of
the subsurface symmetry systems, but also to the characterization of its phys-
ical parameters. In this paper, we estimate azimuthal variations of the P-wave
effective quality factor (Q) from field surface seismic data. We also provide an
interpretation of our results taking into account the NMO ellipse parameters
measured from the data, and potential bias caused by systematic noise. By
assuming that Q is frequency independent, and the medium at each particu-
lar azimuth the medium is laterally homogeneous, we use the spectral ratio
method and a regularized linear inversion scheme to estimate the quality factor
in azimuth-sectored data. The regularization parameters are chosen by a 2
criterion that is based on estimates of the variance in the field data. Tests on
synthetic data show that this regularized inversion provides robust estimates
of Q for signal-to-noise ratios lower than those observed in the data. Appli-
cation of this methodology to P-wave data from the East Decatur Field in
Texas yields non-negligible azimuthal variations in Q. The azimuthal signature
of attenuation appears to be consistent with the effective NMO ellipse from
the same interface. However, data residuals show non-random structures that
suggest a strong systematic component to the noise. We provide a brief analy-
sis of scattering-related absorption and of frequency imprints of source-receiver
arrays as possible sources of systematic noise.
Key words: attenuation, azimuthal anisotropy, regularized inverse solution
x
Taking the logarithm of equation (2) gives;
1
x
2 S1 (f ) A1 f
x
3 ln = ln (t1 t2 ) (3)
S2 (f ) A2 Q
Equation (3) shows that the quality factor Q can be
directly estimated from the slope of the ratio of the
spectra (in logarithmic scale) for any two traces, if the
Power
Offset
Signal
Bandwidth
Frequency
Figure 2. Generation of the data vector d (equation 8) from seismic traces using equation (6). Seismic data is sorted according
to the CMP number and offset (left hand side). The waveform of the reflection event of interest is windowed on each trace (red
dotted lines denote the time window). Then, for any two traces, we compute the ratio of the power spectra of the event (in
logarithmic scale) on a sample-by-sample basis, which yields the data points on the right hand side panel in the figure. Each
pair of traces provides the data di , separated by the dashed lines.
any assumptions on the frequency-independent inter- The relationship between the matrices in equation (6)
cept term. The form of the objective function in equa- and in equation (9) are described in Appendix A. Of
tion (6) falls into the category of direct regularization course, it would be simpler to just apply the General-
methods known as Tikhonov regularization (Hansen, ized Least Squares (GLS) solution (Hansen, 1998; Green
1998). and Silverman, 1994) and get independent estimates
The first term in equation (6) seeks to minimize the of attenuation for each trace pair. However, as we will
total misfit between the data (the spectral ratio) mea- see later on, if estimates of Q obtained independently
sured for each trace combination and data predicted by for each pair, the actual inverted values for two noise-
its corresponding mi . This term would be sufficient to contaminated, closely neighboring trace pairs may be
carry out the inversion if the model assumptions were different by a factor of three. So, if we chose the best
correct and the data were noise-free. Acknowledging the possible fit to the data (the GLS solution) we would vio-
existence of non-negligible variance in our data, the sec- late the homogeneity assumption made in equation (3).
ond term in equation (6) penalizes the norm of R times As mentioned earlier, it is important to provide
the model. This allows us to find a solution that does not meaningful model variances, such that the final results
provide the minimum misfit, but complies to the model can be properly interpreted. The variance in the model
constraints enforced by Rs. In our case, R is the sum of parameters can be directly estimated from the inverse
the first and second derivative operators. The regular- operator A [equation (8)] and the data covariance ma-
ized inverse operator related to the objective function trix (Hansen,1998; van Wijk et al., 2002). For our par-
in equation (6) can be found from its normal equations ticular case, we assume the data covariance matrix is a
followed by simple linear operations that reduce our in- diagonal matrix of the data variances d,i . To estimate
verse problem to a single linear system. These steps are the variance in the data (d,i ) we follow the approach
described in Appendix A. Our regularized inverse solu- described in van Wijk et al. (2002) and minimize the
tion has the same form as the standard Tikhonov regu- function
larized inverse:
g(i ) = ||i di ||2 + 2d,i ||Rd i ||2 (10)
m = A d, (7) with respect to i , for a chosen operator Rd . This gives
where A is the so-called regularized pseudo-inverse op- us an estimate i,d for a given d , a smoothed ver-
erator, sion of our data vector di according to the structure
imposed by the operator Rd . This type of fitting pro-
A = (AT A + 2 RT R)1 AT . (8) cedure is also know as nonparametric regression (Green
Substituting equation (8) into equation (7) gives and Silverman, 1994). If Rd is chosen so that it repre-
sents a certain structure present in the data, then the
m = (AT A + 2 RT R)1 AT d. (9) variance of i,d di yields an estimate of d,i (van
Wijk et al., 2002), the data variance with respect to the
Estimation of azimuthally varying attenuation 93
i=3
0.045 i=1
i=2 i=3 i=4 i=4
0.05 i=1 0.04 i=2
0.035
Log of Spectral Ratio
0
0.03
||i di ||2
0.025
0.05
0.02
0.1 0.015
0.01
0.15
0.005
0.2 0
0 0.05 0.1 0.15
Frequency ||Ri ||2
(a) (b)
4
x 10
10
8
d
i=3
2 i=1
i=4
i=2
0
0 500 1000 1500 2000 2500 3000
d
(c)
Figure 3. (a) Data from vectors di (equation 6) as in Figure 2(circles) and the smoothed data i (equation 10, crosses). di
were computed from noise-contaminated power spectra (Figure 1). The choice of d,i that gives i can be done by picking
d,i that yields the maximum curvature of L-curves in (b); for this value of d,i the data variance is approximately constant
(c). Each of the four curves in (b,c) pertain to the data vectors di in (a). The data in (a) with i = 1, 2, 3, 4 have respectively
increasing differences in offset.
structure predicted by Rd . In our particular application lem, the choice of the weighting parameter determines
we expect the data to have a linear dependence on fre- m . For our particular application we apply a combined
quency, hence we choose Rd to be a second derivative L-curve/2 criterion. If the L-curve criterion (Hansen,
operator. The procedure of choosing the optimum d 1998) alone is used, and the L-Curve varies smoothly,
will be described in the next section. there is a wide range of values that satisfy the crite-
Another separate issue is determining the solu- rion but would result in vastly different models. If an
tion m . Once the regularized inverse operator is es- estimate of the variance in the data is available, setting
tablished, we can apply it to our data and obtain a its corresponding 2 to some confidence level allows one
regularized model. As in any other regularization prob-
94 I. Vasconcelos & E. Jenner
to compute a misfit cut-off for the L-curve that gives the were then modeled with P-wave velocity V = 2.0km/s,
best (van Wijk et al., 2002). Q = 250 , a reflector depth of 1km and a Ricker source
There are important practical issues related to the wavelet. The data was then contaminated with Gaussian
inversion methodology discussed above. To use equa- random noise. Noise variances in the synthetic spectra
tion 3 in our inversion scheme, we need to make an ex- were chosen to be at least two orders of magnitude larger
tra assumption. Since at a particular azimuth our model than that of real data traces for times before the first
is supposed to consist of homogeneous layers with hori- arrival. Figure 1 shows noise-contaminated power spec-
zontal interfaces, while equation (3) is valid for homoge- tra of the reflected signal from four traces. Figure 1 also
neous half-spaces, the contrasts between layers are as- illustrates that as the traveltime increases (with increas-
sumed to be small and raypaths should be close to ver- ing offset), the frequency peak of the power spectrum
tical. In practice, this assumption is enforced by only shifts towards lower frequencies. This behavior should
taking into account only near-offset traces (i.e., offset- always be observed in data for media were the constant
to-depth ratio less than 1). Q assumption is valid, and the attenuation coefficient is
The forward operator Ai [equation (3)] and the proportional to frequency.
data vector di pertain to a windowed portion of the Once the synthetic power spectra were generated,
corresponding traces containing the reflection event of we computed the data vectors di [equation 6)] by taking
interest. The center of window is equal to to the travel- the natural logarithm of the ratio between the spectra of
time of the event predicted using the estimated moveout all possible trace pairs (Figure 2). Before we can find the
parameters for the corresponding CMP gather. Once the regularized inverse solution, it is necessary to estimate
window center positions are determined, the window- d,i (the variance in the data). After many iterations
ing is carried out with a Hanning-type taper (Priestley, over the value of d,i we produce the L-curves in Fig-
1981). Mateeva (2003) discussed some of the implica- ure 3b. The choice of d,i can be made by picking some
tions of windowing in the Q estimation process. Her point where the L-curve has the maximum curvature.
studies show that the window size imposes a bias on Another option for the choice of d,i comes from the
attenuation estimates, and this bias decreases with in- curves in Figure 3c that show that as d,i increases, the
creasing window size. Our window size was typically on data variance increases rapidly at first and then becomes
the order of 1.5 times the dominant wavelength of the approximately constant. A natural choice for the regu-
signal. Mateeva (2003) also shows that windowing data larization parameter would be the value d,i at which
with no tapering can severely distort estimates of Q, but the increase data variance is negligible. After establish-
the choice of the tapering function has little influence ing the optimum value for the regularization parameter,
on the results. we can predict the modified data vector (Figure 3a)
After windowing, we pad the time series with zeros and the total data variance. For the data with the high-
(for as many as ten times the length of the windowed est offset difference between the two traces (i = 4) the
series) to ensure a dense enough sampling of the power corresponding d is smallest (Figure 3). This happens
spectra. Next, we determine the signal frequency band- because all power spectra had their samples randomly
width by comparing the spectrum of our windowed time distorted with the same variance and the ratios that
series with the spectra from far-offset traces for time correspond to spectra with a higher attenuation differ-
windows positioned before the first arrival. The reason ence will show smaller variance. If the noise in our data
for this is that data before the first arrival should be is limited to random fluctuations in the power spectra,
representative of background noise. traces with larger offset differences will give more stable
estimates of Q. As to the issue of how reliable the es-
timates of the data variances (d,i ) are, we refer to the
3 SYNTHETIC DATA TESTING work of van Wijk et al. (2002).
With an estimate for the variance in the data, we
Since one of our assumptions about the data is that set a threshold for the maximum tolerable misfit for
each vertical plane can be treated as a horizontally lay- a certain confidence level, according to the 2 test for
ered, laterally homogeneous medium, the main focus goodness of fit (van Wijk et al., 2002). This threshold
of our synthetic data test is to assess the performance can then be used to decide the amount of regularization
of our inversion methodology on CMP-sorted, surface we can impose in our data. In practice, we wish to be
seismic data that would represent an azimuth sector. conservative about the model smoothness in such a way
Given that our main objective is to study the robust- we choose the value of (equation 9) that provides a
ness of the regularized inverse operator in the presence misfit under the confidence level threshold. Figure 4 il-
of noise, the simplest model for our synthetic example is lustrates the L-curve for R, with the corresponding mis-
that of a homogeneous, isotropic, attenuative layer. For fit threshold set by the 2 criterion. For the chosen we
such a model, we can predict the the power spectrum can find the regularized inverse solution (Figure 5a,c).
of the signal at any trace simply by using equation (1), Figure 5a,c shows that the regularized inverse solution
and having as input our medium parameters and the is closer to the true model than the standard general-
power spectrum of the source pulse. The synthetic data
Estimation of azimuthally varying attenuation 95
ized least squares (GLS) solution. We can also observe
that there is some structure in the regularized inverse
1.83 Lcurve
solutions. This structure is related to the ordering of the 68% confidence
data vectors di within d, which in turn is related to how 1.825
traces are sorted in the input data.
If the data variance estimation procedure provided 1.82
||Am d||2
perfect estimates and we decided to push the regular-
1.815
ization to higher confidence levels, the structure of the
regularized solution would be somewhat less smooth, de- 1.81
pending on where the misfit threshold (determined from
2 ) intersects the L-curve. The regularized inverse so- 1.805
GLS
True model Data
600 error 2nd order Regularized Fit
2nd order TGLS 0.05 Nonregularized fit
500
300 0.05
200
0.1
100
0 0.15
5 10 15 20 25 30 35 40 45
Frequency
trace pair
(b)
(a)
GLS Data
True model Regularized Fit
600 error 2nd order 0.05 Nonregularized fit
2nd order TGLS
Log of Spectral Ratio
500
0
400
Q
300 0.05
200
0.1
100
0.15
0
5 10 15 20 25 30 35 40 45
trace pair Frequency
(c) (d)
Figure 5. Inversion results for the synthetic data with different signal-to-noise ratios. The power spectra that generated data
for (a,b) are distorted with twice the variance as the spectra for the results in (c,d). (a) and (b) show the Q estimates from the
Tikhonov generalized least squares solution (TGLS, 2nd order stands for the use of a first and second derivative operators in
R, Appendix A) and the standard generalized least squares. The error bars represent the variance in the model parameters. (c)
and (d) are the models fitted to the data, where the regularized fits are related to the TGLS solutions and the non-regularized
fits pertain to the GLS solutions.
noise levels are likely smaller than those in synthetic ness of the power spectra (Figure 8) is expected to be
tests. less than the that in our synthetic examples. However,
Once the reflection event at a given azimuth bin although there is little random noise in the spectra, we
has been windowed, we can set up our data space d observe pronounced, repeatable, nonlinear structures in
[equation (8)] from the data vectors di computed for the spectral ratios (Figure 9b).
all trace pair combinations from offset-to-depth ratios The algorithm described above treats these struc-
up to 1.0. Next we estimate the variance in the data, tures in the fit residuals as random noise. Clearly, the
which, from a simple visual examination of the smooth- presence of non-random repeatable structures point to
Estimation of azimuthally varying attenuation 97
Crossline
5130 5180 5230 5280 5330 5380 5430
0
0.5
t(s)
1.0
1.5
2.0
Figure 6. Post-stack migrated image of the East Decatur field data at Inline 1350. The Barnett shale formation is located
approximately in the 1.0 s to 1.3 s interval. The overburden is composed of sand/shale sequences.
Inline
Inline
Crossline Crossline
(a) (b)
Figure 7. Parameters of P-wave NMO ellipses for the reflection from the bottom of the Barnett formation in the East Decatur
field. (a) The magnitude of azimuthal anisotropy (NMO ellipse eccentricities) in percent.(b) The orientation of the semi-major
axis of the NMO ellipses with respect to North. The two CMPs analyzed (CMP1 and CMP2) lie within the area of stronger
azimuthal anisotropy in the upper right-hand corner of the area, respectively with coordinates (Inline 1380, Crossline 5285)
and (Inline 1392, Crossline 5292). Increasing Crossline coordinates point to the North.
98 I. Vasconcelos & E. Jenner
x1 x
1
x2 x
2
x3 x
3
Amplitude
Power
0 10 20 30 40 50 60 70
1290 1295 1300 1305 1310
time(ms) Frequency(Hz)
(a) (b)
Figure 8. Widowed data from CMP1 for a 5o azimuth bin at an azimuth of 40o (a), and their corresponding power spectra
(b). Offsets are x1 = 1614 m, x2 = 1661 m and x3 = 1708 m. The inversion was carried out considering frequency samples from
25-50 Hz.
the fact that the assumptions do not account for all the ellipse, approximately oriented East-West (Figure 7). If
relevant physical phenomena behind the data, although the fractures in the shale are fluid-filled, a possible phys-
the model can represent a component of the data(Green ical interpretation is that waves traveling perpendicu-
and Silverman, 1994; Gu, 2002). In other words, we can lar to the most prominent set of fractures should have
take advantage of the fact that the regularization pa- smaller velocities and experience higher attenuation. It
rameters can be tuned to set how much we impose a is also interesting to observe that the magnitude of az-
model structure. Then, it is possible to fit a linear com- imuthal attenuation anisotropy is much larger than that
ponent of the spectral ratio by trying to account for the of azimuthal NMO velocity anisotropy. The interpreta-
fact that the data variances may be overestimated due tion of these results, however, has to include systematic
to the presence of correlated noise components. We do components, which is discussed next.
this in practice by setting our to a value that corre- If all or some of the nonlinear structure in the data
sponds to a misfit lower than the threshold set by the is related to physical phenomena, it is not clear whether
estimated d . We can then try to pick a value of such or not a linear fit to the data makes any physical sense.
that we get a model with reduced norm, at the same Here, we attempt to model nonlinear noise due to dif-
time being conservative towards the regularization as ferent physical phenomena, evaluate its influence on the
we try to keep the misfit and the regularization-induced data, and briefly discuss how it should be dealt with.
bias as small as possible. Figure 9 shows the estimated Correcting for possible sources of noise and employ-
model and corresponding data fit obtained following this ing more involved parameter fitting techniques such as
approach, before and after regularization. To find the semiparametric fitting and residuals analysis (Grey and
value of Q for a given azimuth, we chose to take the me- Silverman, 1998; Gu, 2002) would be preferable. These
dian of the regularized Q measurements for each trace procedures, however, are beyond the scope of this paper.
pairs to minimize the contribution of outliers. We attribute the nonlinearities in the spectral ra-
The effective quality factor for 5o azimuth bins with tios to two distinct physical phenomena. First, in our
o
20 azimuth increments are displayed for both superbins measurements we cannot separate the contribution of
in Figure 10. It is clear from Figure 10 that if our as- intrinsic absorption from that of layering-induced ab-
sumptions and estimates of the variance in the data are sorption. As it is discussed by Mateeva (2003), the at-
accurate, there is a non-negligible azimuthal signature tenuation caused by multiple scattering may influence
of attenuation. The direction of maximum attenuation the data in different ways, depending on the media prop-
(lowest Q) is close to the semi-minor axis of the NMO erties. Second, for the East Decatur field, the data were
Estimation of azimuthally varying attenuation 99
GLS
2nd order median
200 2nd order error 0.6
2nd order TGLS Data
180 Regularized Fit
0.4 Nonregularized fit
140 0.2
120
0
Q
100
0.2
80
60 0.4
40
0.6
20
0 0.8
5 10 15 20 25 30 35 40 45 50
trace pair Frequency
(a) (b)
Figure 9. Inverted parameter Q for a 5o azimuth bin from CMP1 at 40o azimuth (a), and the corresponding data fit (b). When
compared to the synthetic results in Figure 5, Q measurements in (a) are more scattered, and the fit residuals in (b) exhibit
non-random, repeatable structures.
90 250
120 60
200
400
150
150 30 350
100
300
50
250
180 0 200
Q
150
100
210 330 50
240 300 50
0 50 100 150 200 250 300 350
270
Azimuth (o )
(a) (b)
Figure 10. Estimates of the effective quality factor at different azimuths shown in a polar plot (a) and with error estimates
(b). The results from CMP1 are marked by the blue circles, and from from CMP2 by red triangles. The error bars in (b) are
computed as the average standard deviation of the model variances for each trace pair within a sector (Figure 9). The azimuth
is with respect to the North. The average orientation of the semi-major axis of the NMO ellipses is North-South.
100 I. Vasconcelos & E. Jenner
acquired using source and receiver arrays. Since the re- Offset(m)
sponse of each array component is simply stacked to 100 300 500 700
1.20
form the data, the signature of the source-receiver ar-
rays is frequency dependent (Appendix B).
For our particular dataset, it is also important to
1.25
find out if multiple scattering due to thin layering could
add a nonlinear component to the spectral ratios, and
if this component has a structure that correlates to
that observed in the data (Figure 9b). Using velocities 1.30
t(s)
for East Decatur with the reflectivity method (Muller, 1.35
1985). This method produces a complete wave-field with
all multiples but no intrinsic absorption. Figure 12a
shows that there is a clear offset-dependent nonlinear 1.40
frequency imprint of scattering-induced attenuation in
the spectral ratios. The data structure in Figure 12b is
somewhat different from that observed in field data (e.g. 1.45
Figure 9b), as field structures not only vary with offset,
but also differ from one azimuth bin to another.
Even though layering-induced absorption causes
1.50
nonlinear behavior in the spectral ratios, its influence
should be azimuthally independent, as long as the
medium is laterally homogeneous. This means that the Figure 11. Synthetic seismogram of East Decatur data. The
bias in the estimates of Q related to multiple scattering seismogram was computed using the reflectivity method with
should be the same at all azimuths. Also, in Figure 12 well log data as input. The event of interest in the seismo-
we can observe that the bias in the fitted slopes intro- gram is at approximately 1.28 s.
duced by thin layering is offset dependent. This implies
that if our data is not corrected for scattering absorp-
tion, the regularized solution may contain a structure respond to the orthogonal direction (Appendix B). Us-
related to the bias in each Q measurement, depending ing the results of Appendix B, we computed the fre-
on how regularized the solution is. If too much smooth- quency signatures of the source-receiver arrays for East
ing (regularization) is applied to uncorrected data, it Decatur data (Figure 13a,c). The biggest deviations of
is difficult to assess how the bias from scattering ab- the spectral ratios occur at the 45o N direction, while
sorption translates into errors in the estimated model they become negligible at the 135o N direction (Fig-
parameters. ure 13b,d). These azimuths coincide with the predicted
As discussed before, use of source and receiver ar- directions of the maximum and minimum contribution
rays in acquisition creates frequency signatures that of the source array, which confirms that the influence
may distort the spectral ratios. In Appendix B we show of the receiver array is relatively minor. Figure 13 also
how to obtain the frequency signatures of arrays for a suggests that the bias in the spectral ratios induced by
simple homogeneous, isotropic model. For East Decatur the source and receiver arrays varies with both offset
field data, receiver arrays are composed of eight sen- and azimuth. Furthermore, the frequency dependence
sors, each at a distance of two meters from the cen- of this bias within the signal bandwidth can be consid-
ter of the array with equal step in azimuth. Source ar- ered as approximately linear. Hence, there will be an
rays are made of a line of three vibrator trucks, with apparent azimuthal variation of attenuation due only
a 20 m separation between the trucks. The source line to the particular geometry of the source and receiver
is always oriented in the direction 45o N. Because the arrays.
spacing between the sources is considerably bigger than
that between receivers in their respective arrays, the
source array should produce a more substantial distor- 5 DISCUSSION AND CONCLUSION
tion in the power spectra. Also, there should be some
With the objective of estimating the azimuthally vary-
azimuthal dependence of the source-receiver transfer
ing attenuation, we proposed a methodology that as-
function because the sources are set in a line with a
sumes that at each azimuth the medium can be consid-
fixed orientation for any source-receiver azimuth. The ered, as horizontally layered and laterally homogeneous.
maximum contribution of the source array should be
The methodology includes a search for a pseudo-inverse
for the source-receiver azimuth that coincides with the
operator that can both fit the spectral ratio data and
source array orientation, and the minimum should cor- honor the model assumptions. In the regularization we
Estimation of azimuthally varying attenuation 101
0.2
11m
2
21m
Signal
bandwith
1
0.1
0.5
0.05
0
0 0.5
10 20 30 40 50 60
Frequency(Hz) Frequency
(a) (b)
Figure 12. (a) Power spectra for the data modeled with the reflectivity method for the event of interest in Figure 11 at offsets
11, 21, 31 and 41 m. (b) Data vectors di (circles) and their smoothed versions i (crosses), used to predict the variance in the
data. The data in (b) are computed for the same bandwidth as the spectral ratios for the field data in Figure 9.
use equally weighted first and second derivatives opera- pends on reflected traveltimes computed from moveout
tors to impose smoothness on the estimates of Q within parameters.
each azimuth bin. This methodology is also designed to In contrast to the synthetic data, the spectral ra-
provide reliable error estimates. The variances in the tios from the field gathers contained very little random
model parameters are computed from the regularized noise. Instead, we observed a strong nonlinear system-
pseudo-inverse operator and from the data variances atic component to the noise, which produces nonlinear
obtained with a nonparametric fitting technique. The structures in the spectral ratios. Because our method-
estimates for data variance also provide the basis for ology cannot separate systematic and random compo-
choosing the regularization weighting parameter. nents of the noise, the data variance may be overesti-
Noisy synthetic data were generated by distorting mated. In evaluating the linear component of the spec-
power spectra with uncorrelated Gaussian noise. For tral ratios, we chose to avoid smoothing the data too
this type of noise, estimates of Q from traces with larger much, keeping the misfit under the 68% confidence level
traveltime differences tend to be more stable because the threshold and the regularization-induced bias as small
data variances are smaller. This synthetic test shows as possible. In this context, our results for the inversion
that even for signal-to-noise ratios smaller than those for Q in 5o azimuth sectors from both superbins show a
typically observed in real data, the regularized inverse similar azimuthally variable signature. The direction of
solution is considerably closer to the true model than the maximum attenuation is approximately East-West,
the standard generalized least squares (GLS) solution. which coincides with the average orientation of the semi-
While the regularized solution is always smoother than minor axis of the P-wave NMO ellipse.Taking fluid-filled
the GLS solution, for data with higher noise levels, it fractures with a single predominant orientation as a
may contain structures that related to the particular or- model for the Barnett shale, maximum attenuation oc-
der of the input data. In our case, this order is associated curs in the direction perpendicular to the fracture strike.
to the sorting of the field data traces. An important observation is that the magnitude of the
In the East Decatur dataset, we applied our inver- azimuthal variations of attenuation is much larger than
sion methodology to two 9 9 CMP-sorted superbins that of NMO velocities. In the context of East Decatur
that lie within one of the areas of strongest azimuthal field, this suggests that attenuation measurements are
velocity anisotropy. Since most of the azimuthal aniso- much more sensitive to the anisotropy induced by the
tropy observed in the East Decatur is is believed to be presence of fluid-filled cracks.
caused by the fracturing in the Barnett shale forma- We considered two possible causes for systematic
tion, our studies focused on the reflection from its bot- noise in the spectral ratios: absorption effects due to
tom. The only pre-processing step applied to the data thin layering and frequency distortions caused by stack-
was source-receiver statics correction, which is necessary ing the response of source and receiver arrays. We ana-
because the position of the window around the signal de- lyzed the first case by computing spectral ratio data for
synthetic seismograms generated with the reflectivity
102 I. Vasconcelos & E. Jenner
0.4
0.6
0.8
2 =26o
2 =21o
1 o
2 =16
o
2 =11
1.2
0 10 20 30 40 50 60
Frequency(Hz)
(b)
(a)
0
Log of Spectral Ratio
0.2
0.4
0.6
0.8 o
2 =26
o
2 =21
1 o
2 =16
o
2 =11
1.2
0 10 20 30 40 50 60
Frequency(Hz)
(c) (d)
Figure 13. Transfer functions T , (a) and (c), of the source and receiver arrays combined as a function of the emergence angle
and frequency . (a) Source-receiver transfer function for the source-receiver azimuth that yields the maximum frequency
distortion (azimuth 45o N), and (c), the source-receiver azimuth that gives the smallest frequency distortion. The frequency
signatures in (a) and (c) translate into the signatures in the spectral ratios shown, respectively, in (b) and (d). Each curve on
the spectral ratio plots is computed between one of the signals with the from the legend and a signal with the 5 o emergence
angle. The same source and receiver array geometries used here were employed in the East Decatur field data acquisition. The
emergence angles are directly related to offsets.
method. This way can generate shot gathers accounting the trace offsets. For a horizontally layered, laterally ho-
only for the attenuation caused by multiple scattering mogeneous medium, attenuation effects due to multiple
for the reflectivity properties of the East Decatur field. scattering are independent of azimuth.
Spectral ratios from the modeled data showed nonlinear To examine the frequency signatures of the source
features that change from one trace pair to the other, and receiver arrays, we computed their corresponding
which suggests that the bias in slope estimation intro- transfer functions using the source and receiver array
duced by layering-induced absorption is a function of geometries from field data. The frequency distortions in-
Estimation of azimuthally varying attenuation 103
troduced by the receiver array are negligible compared REFERENCES
to those caused the source array. Also, since the sources
Arts, R. J., and Rasolofosaon, P. N. J., 1992, Approximation
are set up in a line with fixed orientation, its frequency of velocity and attenuation in general anisotropic rocks:
signatures are azimuthally dependent. These signatures 62nd Ann. Intl. Mtg. Soc. Expl. Geophys., Expanded Ab-
translate into the spectral ratios with a weak nonlinear stracts, 640-643.
dependence on frequency. The source and receiver ar- Cardona R., 2002, Fluid Substitution Theories and Multi-
rays introduce a negative bias in slope estimates from component Seismic Characterization of Fractured Reser-
spectral ratios, which is a function of both offset and az- voirs: Ph.D. thesis, Colorado School of Mines.
imuth. This generates an apparent azimuthally variable Clark, R. A., Carter, A. J., Nevill, P. C., and Benson, P. M.,
attenuation effect due only to the source and receiver 2001, Attenuation measurements from surface seismic -
array geometry, where attenuation is strongest at the azimuthal variation and time lapse case studies: 63rd
EAGE Conference and Technical Exhibition, Expanded
45o N direction and weakest at 135o N.
Abstracts, L-28.
By accounting for layering-induced absorption and Dasgupta, R., and Clark, R. A., 1998, Estimation of Q from
for the array transfer functions we could not completely surface seismic reflection data: Geophysics, 63, 2120-
reproduce the structures in the spectral ratios from field 2128.
data. Strictly, in attenuative media reflectivity and ge- Grechka, V., and Tsvankin, I., 1998, 3-D description of nor-
ometrical spreading should be frequency dependent be- mal moveout in anisotropic inhomogeneous media. Geo-
cause of velocity dispersion. These effects would be de- physics 63, 1079-1092.
pendent on incidence angle, which in turn would yield Grey, P. J., and Silverman, and B. W., 1994, Nonparametric
structures in the spectral ratios that should vary with regression and generalized linear models: Monographs on
offset. Source and receiver radiation patterns in lossy Statistics and Applied Probability, 58, Chapman & Hall.
Gu, C., 2002, Smoothing spline ANOVA models: Springer-
media also become frequency dependent, and their ef-
Verlag New York, Inc.
fects in spectral ratios should be corrected for. Gurevich, B., Zyrianov, V. B., and Lopatnikov, S. L., 1997,
In azimuthally anisotropic attenuative medium, in- Seismic attenuation in finely layered porous rocks: Effects
trinsic absorption will have a dependence in polar an- of fluid flow and scattering: Geophysics, 62, 319-324.
gle as well as in azimuth (Zhu, 2005b). For the results Hackert, C. L., Parra, J. O., Brown, R. L., and Collier, H.
presented in this paper we ignored the influence of the A., 2001, Characterization of dispersion, attenuation,
polar angle. However, the methodology presented here and anisotropy at the Buena Vista Hills field, California:
could be applied to trace clusters with different offset Geophysics, 66, 90-96.
ranges to obtain an estimate of the dependence of at- Hansen, P. C., 1998, Rank deficient and discrete ill-posed
tenuation with polar angle. With closed form expres- problems: numerical aspects of linear inversion: Society
for Industrial and Applied Mathematics.
sions for anisotropic attenuation in terms of anisotropic
Hicks, G. J., and Pratt, R. G., 2001, Reflection waveform
parameters as in Zhu and Tsvankin (2004, 2005b), the inversion using local descent methods: Estimating atten-
methodology described here could be applied to charac- uation and velocity over a gas-sand deposit: Geophysics,
terize anisotropic attenuation. Furthermore, our inver- 66, 598-612.
sion procedure can be applied to other types of datasets Hosten, B., Deschamps, M., and Tittman, B. R., 1987, In-
with redundant wave-field information (e.g. walkaway homogeneous wave generation and propagation in lossy
vertical seismic profiles, tomographic experiments), and anisotropic rocks: application to the characterization of
to time lapse studies. viscoelastic composite materials: J. Acoust. Soc. Am.,
82, 1763-1770.
Jenner, E., 2001, Azimuthal Anisotropy of 3-D Compres-
sional Wave Seismic Data, Weyburn Field, Saskatchewan,
6 ACKNOWLEDGMENTS Canada: Ph.D. thesis, Colorado School of Mines.
Lynn, H. B., Campagna, D., Simon K. M., and Beckham,
We greatly appreciate the input given by Alex Calvert
W. E., 1999, Relationship of P-wave seismic attributes,
(GX Tech., AXIS) throughout the course of the project.
azimuthal anisotropy, and commercial gas pay in 3-D P-
We are thankful to Roel Snieder (CWP) for his help wave multiazimuth data, Rulison Field, Piceance Basin,
on the issue of array transfer functions. I.V. is grateful Colorado: Geophysics 64, 1293-1311.
to Huub Douma, Ken Larner and CWP colleagues for Mateeva, A., 2003, Thin horizontal layering as a stratigraphic
insightful discussions. Our appreciation also goes to Ilya filter in absorption estimation and seismic deconvolution:
Tsvankin and Kasper van Wijk for reviewing this paper. Ph.D. thesis, Colorado School of Mines.
We would like to thank GX Technology (AXIS Imag. Muller, 1985, The reflectivity method: a tutorial: Journal of
Div.) and Devon Energy for giving permission to publish Geophysics 58, 153-174.
these results. Parra, J. O., Hackert, C. L., and Xu, P., 2002, Characteri-
zation of low Q zones at the Buena Vista Reservoir, Cal-
ifornia: Geophysics, 67, 1061-1070.
Prasad, M., and Nur, A., 2003, Velocity and attenuation
anisotropy in reservoir rocks: 73rd Ann. Intl. Mtg. Soc.
Expl. Geophys., Expanded Abstracts, 1652-1655.
104 I. Vasconcelos & E. Jenner
Quan, Y., and Harris, J. M., 1997, Seismic attenuation to- equal to the 1/Qi elements from all of the mi matrices,
mography using the frequency shift method: Geophysics, sich that:
62, 895-905. 0 1 1
Rathore, J. S., Fjaer, E., Holt R. M., and Renlie, L., 1995, Q1
Acoustic anisotropy of a synthetic sandstone with a con- B Q1 C
B 2 C
trolled crack geometry: Geophys. Prosp., 43, 805-829. s=B
B .
C.
C (A4)
Vasco, D. W., Peterson Jr., J. E., Majer, and E. L., 1996, @ . A
A simultaneous inversion of seismic traveltimes and am- 1
QN
plitudes for velocity and attenuation: Geophysics, 61,
1738-1757. To find the solutions that minimize equation (A1), we
van Wijk, K., Scales, J. A., Navidi, W., and Tenorio, L., 2002, must differentiate it with respect to all model vectors
On estimating errors in inverse calculations: Geophysical (m1 , m2 , ..., mN , s), and set the result to zero:
Journal Intl., 149, 624-632.
N
White, R. E., 1992, The accuracy of estimating Q from seis- X
mic data: Geophysics, 57, 1508-1511. ||Ai mi di ||2 + 2 ||Rs||2 = 0; (A5)
i=1
m i s
Zhang, C., and Ulrych, T. J., 2002, Estimation of quality
factors from CMP records: Geophysics, 67, 1542-1547. From which we obtain the following normal equations:
Zhu, Y., and Tsvankin, I., 2004, Plane-wave propagation
N N
and radiation patterns in attenuative TI media: CWP X X
Project Review, CWP-479, 125-142. ATi Ai mi + 2 RT Rs = ATi di ; (A6)
Zhu, Y., and Tsvankin, I., 2005a, Physical modeling and i=1 i=1
analysis of P-wave attenuation anisotropy in transversely Let A be a matrix that contains all of our Ai s
isotropic media: CWP Project Review. [equation (A2)] such that:
Zhu, Y., and Tsvankin, I., 2005b, Plane-wave attenuation
anisotropy in orthorhombic media: CWP Project Review. A = k,l , (A7)
here k,l are the elements of A with row k and column l.
From here on we will use k and l to denote respectively
APPENDIX A: THE TIKHONOV row and column indices of any matrix. Its only nonzero
GENERALIZED LEAST SQUARES elements are
SOLUTION
x+n(i1),2i1 = fx ti , 1 < x < n (A8)
Since our objective is to impose smoothness in the mea-
surements of Q from one trace pair to another, it is im- and
portant to incorporate the linear systems from all trace
x+n(i1),2i = 1, 1 < x < n (A9)
pairs into a single linear system that sets up the so-
called generalized least squares problem. where 1 < i < N and maps elements from A to the
Our model objective function is: corresponding it h trace combination. The index x scans
N
over frequency samples. The matrix A has N times n
(N n) rows and 2N columns, and contains the informa-
X
f (mi , s) = ||Ai mi di ||2 + 2 ||Rs||2 ; (A1)
i=1 tion from the forward operators of all trace pair combi-
nations. With A we can set up a linear system:
where the forward operators:
0 1 A = md (A10)
f1 ti 1
B f2 ti 1C where m contains all model matrices mi and d contains
B C
Ai = B
B . .CC, (A2) all data matrices di .
@ . .A Now, to go from the normal equation in equa-
fn ti 1 tion (A6) to the final form of the pseudo-inverse so-
lution, lets rewrite the matrices Ai , mi and di in the
for the it h trace pair. There are n frequency samples
following manner:
in the signal bandwidth and ti is traveltime difference
between the two traces. The model parameters matrices Ai = i A, (A11)
mi :
1 mi = i m, (A12)
mi = Qi , (A3)
a
di = i d, (A13)
where Qi is the Q measurement and a is the intercept of
the line fitted to the spectral ratio. The data di contain where
the values of the natural logarithm of the spectral ratio i = k,l , (A14)
that correspond to the frequency samples in Ai . The
vector s in equation (A1) contains all of the elements i = k,l , (A15)
Estimation of azimuthally varying attenuation 105
with nonzero elements x
p~
x+n(i1),x+n(i1) = 1; 1 < x < n , 1 < i < N (A16) p (~
p)
r
2i1,2i = 1; 1 < i < N (A17) ~c
y
both i and i are square matrices of dimensions N n
and 2N , respectively. Thus our new Ai s dimensions are
N n 2N , mi is now 2N 1 and di is N n 1. Also,
with the objective of designing the pseudo-inverse oper-
ator, we can simply replace s by m, and make sure R
only operates on the slope terms by setting his elements ~r
that correspond to intercept terms to zero. This way, we
can use equation (A12) and equation (A13) to rewrite z
equation (A6):
Figure A1. The array geometry used to derive the arrays
N N
X X transfer function. An array centered at ~c has n array com-
( ATi Ai i )m 2 T
+ R Rm = ( ATi i )d ; (A18) ponents with position denoted by p ~. The array records an
i=1 i=1 arrival from ~r that makes an emergence angle with the ver-
From equation (A11) and equation (A16), the sum of tical. The array element at p~ makes an angle of p with the
all Ai over N results in A. Likewise, summing i and x-axis, and ~r makes an angle of r .
i over N results in identity matrices. When then get:
AT Am + 2 RT Rm = AT d ; (A19) For a point force source in a homogeneous, isotropic
medium:
From which we obtain the pseudo-inverse solution:
eik|~r+d(~p)xi (~p)|
m = (AT A + 2 RT R)1 AT d. (A20) G(~r, p
~) = , (B2)
|~r + d(~p)xi (~ p)|
To provide an example of the regularization opera- k being wavenumber, and xi (~ p) is a unit vector with
tor R, let us consider: the same orientation as p~ ~c. For convenience, in equa-
tion (B2) we set the array center ~c to the origin of the
1 N1 0 N1 N1
0 1
0 ... 0
B 1 coordinate frame. Then putting equation (B2) in equa-
B N 0 1 N1 0 ... 1 C
0
N C
tion (B1) we get:
R=B .B . . . . .
. C C.
@ . . . . . .
. A Z
eik|~r+d(~p)xi (~p)|
N1 0 N1 0 ...1 N1
0 ~u(, ~r) = S() d~
p; (B3)
C |~r + d(~ p)xi (~ p)|
(A21)
This would be the regularization operator to be used if and, if we assume R >> d(~ p):
we choose to penalize the norm of the residuals from the
Z
S()
mean of the model parameters. All the zeros in equa- ~u(, ~r) = eik|~r+d(~p)xi (~p)| d~
p (B4)
R C
tion (A21) correspond to the intercept terms in m.
Now, we can write
|~r + d(~ p)|2 = R2 + 2d(~
p)xi (~ p)[~r.xi ] + d2 (~
p), (B5)
APPENDIX B: TRANSFER FUNCTIONS OF
with
SOURCE AND RECEIVER ARRAYS
~r.xi = R sin cos(p (~
p) r ). (B6)
Let us consider an array geometry according to Fig-
ure A1. From such geometry, we can define the quanti- Because R >> d(~ p) we can drop the term quadratic on
ties R = |~r ~c| and d(~p) = |~
p ~c|. If we simply stack d(~
p). Next, by taking the square root of equation (B5)
the responses of the array components, then the total and performing a Taylor expansion of the radical on the
displacement u~ of the array: right-hand side we get:
Z
u
~ (, ~r) = S() G(~r, p~)d~p (B1) |~r + d(~ p)|2 = R + d(~
p)xi (~ p) sin cos(p (~
p) r ). (B7)
C
Which, if replaced in equation (B4), gives:
where S() is the source function in the frequency do-
eikR
Z
main, and G(~r, p
~) is the Greens function of the medium ~u(, ~r) = S() eikd(~p) sin cos(p (~p)r ) d~
p (B8)
between ~r and p~. C is a closed contour (away from the R C
origin) over which array elements are placed. or
u
~ (, ~r) = S() G(~r, ~c) T (). (B9)
106 I. Vasconcelos & E. Jenner
Equation (B8) represents the total displacement ~u in
terms of the source function S(), the Greens function
at the center of the array G(~r, ~c) and the transfer func-
tion of the array:
Z
T () = eikd(~p) sin cos(p (~p)r ) d~
p, (B10)
C
or simply,
Z
T () = 1 + eikd(~p) sin cos(p (~p)r ) d~
p, (B11)
C
Z
Ts () = 1 + eikds sin cos(ps (p~s )r ) dp~s , (B14)
Cs
ABSTRACT
where U is the polarization vector, and k = k ikI is the wave vector (both vectors are complex). We consider
plane-wave propagation in orthorhombic media with orthorhombic attenuation, which means that the symmetry of
the imaginary part of the stiffness matrix is identical to that of the real part. Then it is convenient to choose a
Cartesian coordinate system aligned with the natural coordinate frame of the model, so that each coordinate plane
coincides with one of the symmetry planes.
Substituting the plane wave (2) into the wave equation yields the following Christoffel equation:
c11 k12 + c66 k22 + c55 k32 2
" # ( )
(c12 + c66 )k1 k2 (c13 + c55 )k1 k3 U1
(c12 + c66 )k1 k2 c66 k12 + c22 k22 + c44 k32 2 (c23 + c44 )k2 k3 U2 = 0, (3)
(c13 + c55 )k1 k3 (c23 + c44 )k2 k3 c55 k12 + c44 k22 + c33 k32 2 U3
Orthorhombic attenuation 109
where cij = cij + icIij is the complex stiffness coefficient. Following Carcione (2001) and Zhu and Tsvankin (2004),
the elements of the quality-factor matrix can be defined as (no index summation is implied)
cij
Qij I . (4)
cij
Assuming homogeneous wave propagation (k k kI ), the wave vector can be expressed through the unit vector n
in the slowness direction: k = n k. Then the Christoffel equation (3) becomes
Gij V 2 ik Uk = 0 ,
(5)
where V = is the complex phase velocity, and Gij are the elements of the complex Christoffel matrix:
k
G11 = c11 n21 + c66 n22 + c55 n23 ,
G22 = c66 n21 + c22 n22 + c44 n23 ,
G33 = c55 n21 + c44 n22 + c33 n23 ,
G12 = (c12 + c66 )n1 n2 ,
G13 = (c13 + c55 )n1 n3 ,
G23 = (c23 + c44 )n2 n3 . (6)
The components of the unit slowness (phase) vector n can be expressed through the polar phase angle and the
azimuthal phase angle : n1 = sin cos , n2 = sin sin , n3 = cos .
Note that although equation (5) has the same form as the Christoffel equation in nonattenuative (purely elastic)
orthorhombic media, the velocity, polarization and stiffnesses are complex. As a result, plane-wave propagation in
attenuative media is described by two coupled equations obtained by separating the real and imaginary parts of the
Christoffel equation.
Equation (7) has the same form as the Christoffel equation for VTI media with VTI attenuation discussed in detail
by Zhu and Tsvankin (2004, 2005a). The only difference between the two equations is that while for VTI media
c44 = c55 , that is generally not the case for orthorhombic symmetry. However, the stiffness c44 influences only
the SH-wave polarized perpendicular to the propagation plane (see below),while c55 contributes to the velocity and
attenuation of the in-plane polarized waves (P and SV). Therefore, the well-known equivalence between the Christoffel
equation in purely elastic VTI media and symmetry planes of orthorhombic media (e.g., Tsvankin, 1997, 2001) holds
for attenuative models with the same symmetries of the real and imaginary parts of stiffness tensor.
Since the Christoffel matrix for wave propagation in the [x1 , x3 ]-plane has four vanishing elements, equation (7)
splits into two separate equations, one for the SH-wave polarized in the x2 -direction (the displacement component
U2 ), and the other for the in-plane polarized P- and SV-waves (the components U1 and U3 ). The solutions for the
velocity and attenuation of all three modes can be obtained by simply adapting the results of Zhu and Tsvankin
(2004, 2005a) for VTI media.
Assuming homogeneous wave propagation (k = n k), the Christoffel equation for the SH-wave takes the form
c66 sin2 + c44 cos2 k2 2 = 0 .
(8)
Using the VTI result of Zhu and Tsvankin (2004, 2005a), the normalized attenuation coefficient of SH-waves in the
[x1 , x3 ]-plane can be obtained from equation (8) as
(2)
p
ASH = 1 + (Q44 (2) )2 Q44 (2) , (9)
where the superscript (2) stands for the x2 -axis orthogonal to the propagation plane (the same convention as in
110 Y. Zhu & I. Tsvankin
Tsvankin, 1997, 2001), and
(1 + 2 (2) ) sin2 + cos2
(2) .
Q44
(1 + 2 (2) ) sin2 + cos2
Q66
For P- and SV-waves in the regime of homogeneous wave propagation, equation (7) reduces to
c11 sin2 + c55 cos2 k2 2
2
(c13 + c55 ) sin cos
2k U1
= 0. (10)
(c13 + c55 ) sin cos k2 c55 sin + c33 cos k 2
2 2
U3
The wavenumber obtained from equation (10) is described by an expression analogous to that in nonattenuative VTI
media (e.g., Tsvankin, 2001):
p
2 (c11 + c55 ) sin2 + (c33 + c55 ) cos2
k =
q 1/2
2
2 2 2
(c11 c55 ) sin (c33 c55 ) cos2 + 4 (c13 + c55 ) sin cos2 . (11)
(2)
The normalized attenuation coefficients AP,SV were derived from the complex part of equation (11) by Zhu and
(2)
Tsvankin (2004, 2005a). For example, the P-wave coefficients AP in the vertical and horizontal directions are given
by
1
p
(2)
AP ( = 0 ) = Q33 1 + 1/Q233 1 , (12)
2Q33
1
p
(2)
AP ( = 90 ) = Q11 1 + 1/Q211 1 . (13)
2Q11
The SV-wave attenuation coefficient in both the vertical and horizontal directions is
1
p
(2) (2)
ASV ( = 0 ) = ASV ( = 90 ) = Q55 1 + 1/Q255 1 . (14)
2Q55
For plane-wave propagation in the [x2 , x3 ]-plane (n1 = 0, n2 = sin , and n3 = cos ), the Christoffel equation (3)
gives
c66 k12 + c55 k32 2 0 0 U1
0 c22 k12 + c44 k32 2 (c23 + c44 )k1 k3 U2 = 0. (15)
0 (c23 + c44 )k1 k3 c44 k12 + c33 k32 2 U3
The SH-wave, which is polarized in the x1 -direction, is described by the element c66 k12 + c55 k32 2 of the
matrix in equation (15). It is clear from equations (7) and (15) that both the velocity and attenuation of the SH-wave
can be obtained from the corresponding equations for the [x1 , x3 ]-plane (or VTI media) by making the substitution
4 5 in the subscripts of the stiffnesses and elements Qij . For example, the normalized attenuation coefficient for
homogeneous SH-wave propagation can be adapted from equation (9):
(1)
p
ASH = 1 + (Q55 (1) )2 Q55 (1) , (16)
where
(1 + 2 (1) ) sin2 + cos2
(1) .
Q55
(1 + 2 (1) ) sin2 + cos2
Q66
The velocity and attenuation of P- and SV-waves in the [x2 , x3 ]-plane are described by (for homogeneous wave
propagation)
c22 sin2 + c44 cos2 k2 2 2
(c23 + c44 ) sin cos
2k U2
= 0. (17)
(c23 + c44 ) sin cos k2 c44 sin + c33 cos k 2
2 2
U3
The P- and SV-wave attenuation coefficients can be obtained from the equations for the [x 1 , x3 ]-plane using the
following substitutions in the subscripts: 1 2 and 5 4.
The same substitutions were used by Tsvankin (1997, 2001) in his extension of the VTI velocity equations to
the symmetry planes of orthorhombic media. The equivalence with vertical transverse isotropy is also valid for the
complex Christoffel equation in the [x1 , x2 ] symmetry plane.
Orthorhombic attenuation 111
4 ATTENUATION-ANISOTROPY PARAMETERS
The Thomsen-style notation for velocity anisotropy suggested by Tsvankin (1997, 2001) helps to simplify the analytic
description of a wide range of seismic signatures for orthorhombic media. Tsvankins parameters provided a basis
for developing efficient seismic inversion and processing methods operating with orthorhombic models (Grechka and
Tsvankin, 1999; Grechka et al., 1999; Bakulin et al., 2000). Here, we extend his approach to attenuative orthorhombic
media with the main goal of defining the parameter combinations that govern the directionally dependent attenuation
coefficient.
Since our notation is designed primarily for reflection data, we choose the P- and S-wave attenuation coefficients
in the vertical (x3 ) direction (AP 0 and AS0 ) as the reference isotropic quantities. The coefficient AS0 corresponds
to the S-wave polarized in the x1 -direction, which may be either the fast or slow shear mode depending on the
relationship between the stiffnesses c44 and c55 . According to equations (12) and (14), the approximate (accurate to
the second order in 1/Q) coefficients AP 0 and AS0 are given by
1
AP 0 , (18)
2Q33
1
AS0 . (19)
2Q55
To characterize the attenuation of waves propagating in the [x1 , x3 ]-plane, we define three attenuation-anisotropy
parameters analogous to the Thomsen-style parameters Q , Q , and Q introduced for VTI media with VTI attenu-
ation by Zhu and Tsvankin (2004, 2005a). The parameters (2) Q
(2)
and Q (the superscript (2) stands for the x2 -axis
perpendicular to the [x1 , x3 ]-plane) determine the fractional difference between the normalized attenuation coefficients
(2)
in the x1 - and x3 -directions for the P- and SH-waves, respectively. Another parameter, Q , is expressed through the
second derivative of the P-wave attenuation coefficient in the vertical direction and, therefore, governs the P-wave
attenuation for near-vertical propagation in the [x1 , x3 ]-plane.
Q33 Q11
(2) , (20)
Q
Q11
Q33 Q55 (c13 + c33 )2 Q33 Q13
c55 +2 c13 (c13 + c55 )
(2) Q55 (c33 c55 ) Q13
Q (21)
c33 (c33 c55 )
Q33 Q55 (2) Q33 Q13
4 g +2 (1 + 2 (2) 2g (2) ) , (22)
Q55 Q13
(2) Q44 Q66
Q , (23)
Q66
(2) c55
where equation (22) for Q is simplified by assuming that the ratio g (2) and the absolute value of Tsvankins
c33
(2)
velocity-anisotropy parameter are small. Since the Christoffel equation in the [x1 , x3 ]-plane has the same form as
in VTI media, equations (20)(23) are identical to the definitions of the corresponding VTI parameters. In contrast
to VTI models, however, the parameters of orthorhombic media with the subscripts 55 and 44 are generally
different, and one needs to use c55 and Q55 (not c44 or Q44 ) in equation (21) and Q44 (not Q55 ) in equation (23).
Similarly, we adapt the VTI definitions of Zhu and Tsvankin (2004, 2005a) to introduce three attenuation-
anisotropy parameters in the [x2 , x3 ]-plane:
Q33 Q22
(1) , (24)
Q
Q22
Q33 Q44 (c23 + c33 )2 Q33 Q23
c44 +2 c23 (c23 + c44 )
(1) Q44 (c33 c44 ) Q23
Q (25)
c33 (c33 c44 )
Q33 Q44 (1) Q33 Q23
4 g +2 (1 + 2 (1) 2g (1) ) , (26)
Q44 Q23
(1) Q55 Q66
Q . (27)
Q66
In equation (26), (1) is the velocity-anisotropy parameter defined in the [x2 , x3 ]-plane (Tsvankin, 1997, 2001), and
c44
g (1) . Since the attenuation coefficient is supposed to be positive (otherwise, the amplitude will increase with
c33
112 Y. Zhu & I. Tsvankin
distance), the diagonal components of the Q matrix have to be positive as well. This constraint implies the parameters
(1)
Q
, (2)
Q
(1)
, Q (2)
, and Q are always larger than 1.
The only component of the Q-matrix that is not involved in the definitions of the reference isotropic quantities
and the attenuation-anisotropy parameters in the vertical symmetry planes is Q12 . Following the approach of Tsvankin
(3)
(1997, 2001), we use Q12 to introduce one more anisotropy parameter, Q , which plays the role of the VTI parameter
Q in the [x1 , x2 ]-plane (x1 is treated as the symmetry axis of the equivalent VTI model):
(2) (2)
sin2 cos2 ,
ASV = AS0 1 + Q (35)
where
" #
(2) 1 (2) (2)
(2)
Q
(2)
Q
Q (2)
2(1 gQ ) + , (36)
gQ g (2)
Azim=0 Azim=20
3.6 3.6
Phase velocity (km/s)
3.2 3.2
3 3
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=45 Azim=70
3.6 3.6
Phase velocity (km/s)
3.4 3.4
3.2 3.2
3 3
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Figure 1. Influence of isotropic attenuation on the exact P-wave phase velocity computed from the Christoffel equation (5).
Each plot corresponds to a fixed azimuthal phase angle. The solid curves mark the velocity for a nonattenuative orthorhombic
model with the following Tsvankins (1997, 2001) parameters: VP 0 = 3 km/s, VS0 = 1.5 km/s, (1) = 0.25, (2) = 0.15,
(1) = 0.05, (2) = 0.1, (3) = 0.15, (1) = 0.28, and (2) = 0.15. The dashed curves are computed for a model with the same
velocity parameters and strong isotropic attenuation (Q33 = Q55 = 10; all attenuation-anisotropy parameters are set to zero).
anisotropic velocity model can be used in the processing of amplitude measurements and inversion for the attenuation-
anisotropy parameters.
Azim=0 Azim=20
3.8 3.8
3.6 3.6
Phase velocity (km/s)
3.2 3.2
3 3
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=45 Azim=70
3.8 3.8
3.6 3.6
Phase velocity (km/s)
3.4 3.4
3.2 3.2
3 3
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Figure 2. Influence of anisotropic attenuation on the exact P-wave phase velocity. The solid curves are the phase velocities for
the nonattenuative orthorhombic model from Figure 1. The dashed curves are computed for a model with the same velocity
(1) (2) (1) (2) (3)
parameters and strong orthorhombic attenuation: Q33 = Q55 = 10, Q = Q = 0.8, Q = Q = Q = 0.5, and
(1) (2)
Q = Q = 0.8.
(1)
Q () = Q sin2 + Q
(2)
cos2 . (40)
Evidently, the linearized P-wave attenuation coefficient in any vertical plane of orthorhombic media is described by
the VTI equation (Zhu and Tsvankin, 2004, 2005a) with the azimuthally varying parameters Q () and Q (). For
wave propagation in the [x1 , x3 ]-plane ( = 0 ), Q = (2)
Q
(2)
, Q = Q , and equation (38) reduces to equation (34).
(1) (1)
Similarly, for the [x2 , x3 ]-plane ( = 90 ), Q = Q , Q = Q , and equation (38) reduces to equation (37).
Remarkably, equations (38)(40) have exactly the same form as the linearized P-wave phase-velocity equa-
tions (1.107)(1.109) in Tsvankin (2001). This similarity is explained by the identical (orthorhombic) symmetry
imposed on both the real and imaginary parts of the stiffness matrix and the assumption of homogeneous wave prop-
(1)
agation. However, an important difference between the coefficient AP and phase velocity is that the parameters Q ,
(2) (3)
Q , and Q include a contribution of the velocity anisotropy, while the velocity function is practically independent
of attenuation. Also, as discussed below, the exact coefficient AP is influenced by the velocity-anisotropy parameters
even for fixed values of (1,2)
Q
(1,2,3)
and Q .
116 Y. Zhu & I. Tsvankin
6.3 Attenuation for VTI and HTI media
Transversely isotropic models with both vertical (VTI) and horizontal (HTI) symmetry axis can be treated as special
cases of orthorhombic media. For VTI media with VTI attenuation, all vertical planes are identical, and there is no
velocity or attenuation variation in the horizontal (isotropy) plane:
(1) = 0, (1)
Q
= 0,
(1) = 0, (1)
Q = 0,
(1) = 0, (1)
Q = 0.
AHTI (2)
cos2 sin2 cos2 + (2) cos4 + Q
(2)
sin2 cos2 sin4 .
P = AP 0 1 + Q Q
(42)
Azim=0 Azim=50
Q55= 4 Q55= 4
0.016 40 0.016 40
400 400
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Figure 3. Influence of the parameter AS0 = 1/(2Q55 ) (marked on the plot) on the normalized P-wave attenuation coefficient.
The velocity parameters correspond to an orthorhombic model formed by vertical cracks embedded in a VTI background
(Schoenberg and Helbig, 1997): VP 0 = 2.437 km/s, VS0 = 1.265 km/s, (1) = 0.329, (2) = 0.258, (1) = 0.083, (2) = 0.078,
(3) = 0.106, (1) = 0.182, and (2) = 0.0455. The P-wave vertical attenuation coefficient is AP 0 = 0.01 (Q33 = 50);
(1) (2)
each attenuation-anisotropy parameter is twice the corresponding velocity-anisotropy parameter: Q = 0.658, Q = 0.516,
(1) (2) (3) (1) (2)
Q = 0.166, Q = 0.156, Q = 0.212, Q = 0.364, and Q = 0.091.
the attenuation anisotropy is much more pronounced. This and other tests for a representative set of orthorhombic
models confirm that equation (38) adequately describes P-wave attenuation under the assumption of homogeneous
wave propagation.
To identify the source of errors in the weak-anisotropy approximation, we repeat the test in Figure 4 using a
purely isotropic velocity model (Figure 5). The approximate solution (dashed lines) in Figure 5 coincides with that
in Figure 4 because both models have identical attenuation-anisotropy parameters. The exact coefficient A P (solid
lines), however, is influenced by the velocity-anisotropy parameters in such a way that the error of the weak-anisotropy
approximation becomes much smaller when the velocity field is isotropic (Figure 5).
Hence, the accuracy of the approximation (38) is controlled primarily by the strength of the velocity anisotropy,
even if the magnitude of the attenuation anisotropy is much higher. This can be explained by the multiple linearizations
in the velocity-anisotropy parameters involved in deriving equations (A4), (A7), and (A8).
It should be emphasized that the influence of different subsets of the velocity-anisotropy parameters on the
attenuation coefficient varies with the azimuth . As illustrated in Figure 6, the contribution of the velocity-anisotropy
parameters defined in the [x1 , x3 ]-plane (the azimuth = 0 ) to the coefficient AP decreases away from that plane
and completely vanishes in the orthogonal direction. Note that according to the Christoffel equation (17), the P-wave
attenuation coefficient in the [x2 , x3 ]-plane ( = 90 ) is indeed fully independent of the velocity- and attenuation-
anisotropy parameters defined in the other two symmetry planes. Similarly, the maximum influence on A P of the
velocity-anisotropy parameters defined in the [x2 , x3 ]-plane is observed for azimuths close to 90 .
Azim=0 Azim=20
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=40 Azim=50
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=70 Azim=90
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Figure 4. Comparison of the exact coefficient AP (solid curves) with the linearized approximation (38) (dashed) for an
orthorhombic medium with orthorhombic attenuation. The model parameters are the same as in Figure 3 (Q 55 = 40).
Orthorhombic attenuation 119
Azim=0 Azim=20
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=40 Azim=50
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=70 Azim=90
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Figure 5. Comparison of the exact coefficient AP (solid curves) with the linearized approximation (38) (dashed) for a medium
with orthorhombic attenuation but a purely isotropic velocity function. The attenuation parameters are the same as in Figures 3
and 4, but the velocity VP 0 = 2.437 km/s is constant in all directions.
120 Y. Zhu & I. Tsvankin
Azim=0 Azim=20
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=40 Azim=50
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Azim=70 Azim=90
0.016 0.016
Normalized attenuation
Normalized attenuation
0.014 0.014
0.012 0.012
0.01 0.01
0 20 40 60 80 0 20 40 60 80
Polar angle (deg) Polar angle (deg)
Figure 6. Influence of the velocity anisotropy on the exact attenuation coefficient A P . The solid curves are computed for the
orthorhombic model with orthorhombic attenuation from Figures 3 and 4. The dashed curves are obtained by setting to zero
the velocity-anisotropy parameters (2) , (2) , and (2) defined in the [x1 , x3 ]-plane (azimuth = 0 ); all other model parameters
are unchanged.
Orthorhombic attenuation 121
The equivalence between the Christoffel equation in and six (instead of nine) attenuation-anisotropy param-
the symmetry planes of orthorhombic and VTI media, eters: AP 0 , (1,2)
Q
(1,2,3)
, and Q . Still, due to the non-
established previously for purely elastic media, holds negligible influence of the velocity anisotropy on AP ,
in the presence of orthorhombic attenuation. Therefore, inversion of attenuation measurements for orthorhom-
the symmetry-plane attenuation coefficients of all three bic media cannot be performed without anisotropic ve-
modes can be obtained by simply adapting the known locity analysis. Also, note that knowledge of the velocity
VTI equations. Also, the Thomsen-style VTI notation of field is required to correct for the difference between the
Zhu and Tsvankin (2004, 2005a) can be extended to or- phase attenuation coefficient studied here and the group
thorhombic media following the approach suggested by attenuation coefficient responsible for the amplitude de-
Tsvankin (1997, 2001) for velocity anisotropy. The set of cay along seismic rays (Zhu and Tsvankin, 2004).
attenuation-anisotropy parameters introduced here in-
cludes two reference (isotropic) P- and S-wave atten-
uation coefficients in the x3 direction, AP 0 and AS0 , 8 ACKNOWLEDGMENTS
and seven dimensionless anisotropy parameters, (1,2) Q
,
(1,2,3) (1,2) We are grateful to members of the A(nisotropy)-Team
Q , and Q .
of the Center for Wave Phenomena (CWP), Colorado
Adaptation of the linearized VTI equations leads
School of Mines (CSM), for helpful discussions, and to
to concise expressions for the symmetry-plane atten-
Ken Larner (CSM) for his review of the manuscript.
uation coefficients of P-, S1 -, and S2 -waves valid for
The support for this work was provided by the Consor-
weak attenuation and weak velocity and attenuation
tium Project on Seismic Inverse Methods for Complex
anisotropy. Furthermore, linearization of the Christof-
Structures at CWP and by the Chemical Sciences, Geo-
fel equation in the attenuation-anisotropy parameters
sciences and Biosciences Division, Office of Basic Energy
yields the approximate P-wave attenuation coefficient
Sciences, U.S. Department of Energy.
AP outside the symmetry planes as a simple function
of AP 0 , (1,2)
Q
, and Q (1,2,3)
. Interestingly, the linearized
coefficient AP expressed through the phase angles and
attenuation-anisotropy parameters has the same form REFERENCES
as the approximate P-wave phase-velocity function in Akbar, N., Dvorkin, J., and Nur, A., 1993, Relating P-wave
terms of Tsvankins velocity parameters. Also, as is the attenuation to permeability: Geophysics, 58, 2029.
case for velocity anisotropy, the approximate P-wave at- Bakulin, A., Grechka, V., and Tsvankin, I., 2000, Estima-
tenuation coefficient in each vertical plane of orthorhom- tion of fracture parameters from reflection seismic data.
bic media is given by the VTI equation with azimuthally Part II: Fractured models with orthorhombic symmetry:
varying parameters Q and Q . Geophysics, 65, 18031817.
Carcione, J. M., 2001, Wave fields in real media: wave propa-
This equivalence between the linearized equations
gation in anisotropic, anelastic, and porous media: Perg-
for attenuation and velocity anisotropy stems from
amon Press.
the identical (orthorhombic) symmetry of the real and Crampin, S., 1991, Effects of singularities on shear-wave
imaginary parts of the stiffness tensor and the assump- propagation in sedimentary basins: Geophysical Journal
tion of homogeneous wave propagation. Still, there are International, 107, 531543.
important differences between the treatment of velocity Grechka, V., Theophanis, S., and Tsvankin, I., 1999, Joint
and attenuation anisotropy. Our analysis shows that in inversion of P and PSwaves in orthorhombic media:
the absence of pronounced velocity dispersion the influ- Theory and a physical-modeling study: Geophysics, 64,
ence of attenuation (i.e., of the imaginary part of the 146161.
stiffness tensor) on velocity is practically negligible. In Grechka, V., and Tsvankin, I., 1999, 3-D moveout veloc-
ity analysis and parameter estimation for orthorhombic
contrast, the definitions of the attenuation-anisotropy
(1,2,3) media: Geophysics, 64, 820837.
parameters Q include the velocity-anisotropy pa-
Lynn, H. B., Campagna, D., Simon, K. M., and Beckham,
(1,2,3)
rameters . W. E., 1999, Relationship of P-wave seismic attributes,
Also, although the velocity anisotropy does not ex- azimuthal anisotropy, and commercial gas pay in 3-D P-
plicitly contribute to the linearized expressions for at- wave multiazimuth data, Rulison Field, Piceance Basin,
tenuation, the exact attenuation coefficient AP does Colorado: Geophysics, 64, 12931311.
vary with the velocity-anisotropy parameters even for Mavko, G. M., and Nur, A., 1979, Wave attenuation in par-
fixed values of Q (1,2,3)
. Moreover, the accuracy of the lin- tially saturated rocks: Geophysics, 44, 161178.
Pointer T., Liu, E., and Crampin, S., 1996, Seismic wave
earized equation for AP is controlled to a large degree by
propagation in media with interconnected cracks and
the strength of the velocity anisotropy. Numerical tests
pores: 66nd Annual International Meeting, Society of
demonstrate that the approximate AP remains close to Exploration Geophysicists, Expanded Abstracts, 1846
the exact value even for strongly attenuative media pro- 1849.
vided the velocity anisotropy is relatively weak. Schoenberg, M., and Helbig, K., 1997, Orthorhombic media:
Thus, the P-wave attenuation coefficient is primar- Modeling elastic wave behavior in a vertically fractured
ily governed by the orientation of the symmetry planes earth: Geophysics, 62, 19541974.
122 Y. Zhu & I. Tsvankin
Thomsen, L., 1986, Weak elastic anisotropy: Geophysics, 51,
19541966.
Tsvankin, I., 1997, Anisotropic parameters and P-wave veloc-
ity for orthorhombic media: Geophysics, 62, 12921309.
Tsvankin, I., 2001, Seismic signatures and analysis of reflec-
tion data in anisotropic media: Elsevier.
Zhu, Y., and Tsvankin, I., 2004, Plane-wave propagation and
radiation patterns in attenuative TI media: 74th Ann. In-
ternat. Mtg., Soc. Expl. Geophys., Expanded Abstracts,
139142.
Zhu, Y., and Tsvankin, I., 2005a, Plane-wave propagation in
attenuative TI media: Geophysics, under review.
Zhu, Y., and Tsvankin, I., 2005b, Physical modeling and
analysis of P-wave attenuation anisotropy in transversely
isotropic media: CWP Project Review (this volume).
(A3)
which is identical to the Christoffel equation for the reference nonattenuative medium.
The normalized attenuation coefficient A is obtained from the imaginary part of equation (A2):
1 Hu
A= 1+ , (A4)
2Q33 Hd
where
Hu = (1,6,5) C(1,6,5) (C(6,2,4) V 2 )(C(5,4,3) V 2 ) C(23,44)
2
ABSTRACT
The amplitudes and frequency content of seismic waves propagating through
anisotropic formations may be strongly distorted by directionally dependent
attenuation. Here, we present physical-modeling measurements of the P-wave
attenuation coefficient in a transversely isotropic phenolic sample.
Using the spectral-ratio method, we estimated the group (effective) attenua-
tion coefficient of P-waves transmitted through the sample for a wide range of
propagation angles (from 0 to 90 ) with the symmetry axis. Correction for the
difference between the group and phase angles was used to obtain the normal-
ized phase attenuation coefficient A that was inverted for the Thomsen-style
attenuation-anisotropy parameters Q and Q . Whereas the symmetry axes of
the angle-dependent coefficient A and of the velocity function have close ori-
entations, the magnitude of attenuation anisotropy far exceeds that of velocity
anisotropy. The quality factor Q increases more than tenfold from the symmetry
(slow) direction to the isotropy plane (fast direction).
The robustness of our results depends critically on several factors, such as the
availability of an accurate anisotropic velocity model and the adequacy of the
homogeneous concept of wave propagation. The methodology discussed here
can be extended to field measurements of anisotropic attenuation needed for
AVO (amplitude variation with offset) analysis and seismic fracture detection.
Key words: attenuation, attenuation anisotropy, transverse isotropy, physical
modeling
10.8 cm
medium with k(0)I = 0, the logarithm of the amplitude
ratio becomes
(1) (1)
U G
ln (0) = ln
k(1)I x(1) . (5)
U G(0)
Time (ms)
rival, obtained by filtering out the low (f < 5 kHz) and P
high (f > 750 kHz) frequencies, is shown in Figure 2b.
An aluminum block with negligibly small attenuation
served as the reference model. The spectrum of the refer-
ence trace acquired by a receiver located directly above
the source (Figure 3) was used to estimate the attenu-
ation coefficient by means of the spectral-ratio method
described above.
0.10
The parameters of the TTI velocity model needed
to process the attenuation measurements were obtained
by Dewangan et al. (2005) from reflection PP and
PS data (Figure 4). Tilted transverse isotropy is de-
scribed by the the P- and S-wave velocities in the sym-
metry direction (VP 0 and VS0 , respectively), Thomsen
anisotropy parameters and defined with respect to
the symmetry axis, the angle between the symmetry b) Receiver coordinate (cm)
axis and the vertical, and the thickness z of the sample. 20
The known values of = 70 and z = 10.8 cm were ac- 0
curately estimated from the reflection data, confirming
that the velocity-inversion algorithm is robust.
ANISOTROPY
4.1 Estimation of the attenuation-anisotropy
parameters
For each receiver position at the surface of the phenolic 400
sample, we divided the spectrum of the recorded trace
by that of the reference trace (Figure 3), as suggested
by equation (5). Records with a low signal-to-noise ratio
were excluded from the analysis. We use the frequency
band of 60110 kHz to estimate the attenuation coeffi- 600
cient. According to the spectral-ratio method described
above, the relevant elements Qij of the quality-factor
matrix are assumed to be constant in that frequency
band.
The normalized phase-attenuation coefficient A,
obtained after correcting for the difference between
group and phase attenuation (it does not exceed 6%),
exhibits a pronounced variation between the slow (0 ) Figure 2. (a) Raw transmission data excited by a P-wave
and fast (90 ) directions (Figure 5). The largest atten- transducer, and (b) the amplitude spectrum of the windowed
uation coefficient is observed along the symmetry axis first arrival. The solid line is the P-wave traveltime modeled
( = 0 ), where the P-wave phase velocity reaches its by Dewangan et al. (2005) using the inverted parameters
minimum value. Since the symmetry direction is or- from Figure 4. The time sampling interval is 2 s, and the
thogonal to the multiple thin layers bonded together width of the Gaussian window is 40 samples.
Physical modeling and analysis of P-wave attenuation anisotropy 129
a)
3
0.1
TTI parameters
0.05
Amplitude
2
0
0.05 1
0.1
0
0.075 0.08 0.085 0.09 0.095
VP0 V
Time (ms) S0
0
0.15
0.1
1
0.05
0 2
0 20 40 60 80 AP 0 Q Q
Phase angle (deg) Figure 7. Influence of errors in the velocity model on the
attenuation parameters. The error bars mark the standard
Figure 5. Measured normalized attenuation coefficient as a deviations in each parameter obtained by applying our algo-
function of the phase angle. rithm with 50 realizations of the input TTI velocity param-
eters. The standard deviations in the TTI parameters are
taken from Dewangan et al. (2005).
APPENDIX A: THOMSEN-STYLE
PARAMETERS FOR ATTENUATIVE TI
MEDIA
The attenuation coefficients in transversely isotropic
(TI) media with TI attenuation can be conveniently de-
scribed using the Thomsen-style notation of Zhu and
Tsvankin (2004, 2005). Instead of the five relevant com-
ponents Qij of the quality-factor matrix, they defined
two isotropic reference quantities (AP 0 and AS0 ) and
three dimensionless parameters (Q , Q , and Q ) de-
scribing attenuation anisotropy.
The reference parameters AP 0 and AS0 represent
CWP-507
Pawan Dewangan? , Ilya Tsvankin? , Mike Batzle , Kasper van Wijk , and Matt Haney??
? Center for Wave Phenomena, Department of Geophysics, Colorado School of Mines, Golden, CO 80401-1887, USA
Institutefor Experimental Geophysics, Department of Geophysics, Colorado School of Mines, Golden, CO 80401-1887, USA
?? Center for Wave Phenomena, Department of Geophysics, Colorado School of Mines, Golden, CO 80401-1887, USA
ABSTRACT
10.8 cm
Because of the deviation of the symmetry axis from both
the vertical and horizontal directions, the moveout of
PS-waves from horizontal reflectors becomes asymmet-
ric (i.e., the PS-wave traveltime does not stay the same
if the source and receiver are interchanged). As demon-
strated by Dewangan and Tsvankin (2004a), moveout- 60 cm
asymmetry attributes of PS-waves can help to esti-
mate all parameters of a horizontal TTI layer using Figure 1. Physical model representing a horizontal TTI
solely reflection data. The algorithm of Dewangan and layer. The coordinate system is chosen in such a way that
Tsvankin (2004a), based on a modification of the so- the symmetry axis is confined to the [x1 , x3 ]-plane and is
called PP+PS=SS method (Grechka and Tsvankin, dipping in the negative x1 -direction. To simulate a reflection
2002; Grechka and Dewangan, 2003), operates with survey, the sources and receivers were placed on top of the
long-offset PP and PS reflections acquired in the ver- sample in the symmetry-axis plane.
tical plane that contains the symmetry axis (hereafter
called the symmetry-axis plane). bic symmetry. One reflection survey was acquired using
Here, we show on physical-modeling data that the source and receiver transducers (flat-faced cylindrical
combination of PP and PS reflection traveltimes can piezoelectric ultrasonic contact transducers). To gener-
indeed constrain all parameters of a horizontal TTI ate P-waves, the source transducer was polarized ver-
layer. Multicomponent, multioffset reflection seismic tically; shear (SV) waves were excited by a horizontal
lines are acquired in the symmetry-axis plane of a phe- transducer.
nolic sample to record long-spread moveouts of PP- and Another reflection data set was generated by the
PS(PSV)-waves. Following the methodology of Dewan- source transducer and recorded by a scanning laser
gan and Tsvankin (2004a), the modified PP+PS=SS vibrometer that measures the absolute particle veloc-
method is used to compute pure SS-wave reflection trav- ity on the surface of the sample via the Doppler shift
eltimes and the asymmetry attributes of PS-waves. The (Nishizawa et al., 1997; Scales and van Wijk, 1999). The
moveout asymmetry information is then combined with records of multiple shots were stacked to improve data
the pure-mode (PP and SS) NMO velocities and zero- quality. The scanning head is programmed to move the
offset traveltimes to estimate the model parameters. beam after each measurement, so that dense arrays of
The accuracy of the inverted TTI model is verified by data can be recorded automatically. Therefore, data ac-
matching the measured PP and PS traveltimes and re- quisition with the laser vibrometer is much more effi-
producing the results of transmission experiments. cient compared with time-consuming transducer mea-
surements, where relatively large receivers have to be
moved manually.
2 EXPERIMENTAL SETUP
To simulate a TTI layer, we used XX-paper-based phe-
nolic composed of thin layers of paper bonded with phe- 3 SEISMIC REFLECTION EXPERIMENT
nolic resin. The effective medium due to this fine layer- To study the moveout asymmetry of PS-waves and test
ing is anisotropic, and phenolic itself is known to have the parameter-estimation methodology of Dewangan
either TI or orthorhombic symmetry (Isaac and Lawton, and Tsvankin (2004a), we acquired a multioffset 2D re-
1999; Grechka et al., 1999). The sample was prepared flection survey in the symmetry-axis plane, as described
by cutting a large block of commercially available phe- above. The inversion algorithm operates on common-
nolic into smaller blocks and pasting them together at midpoint (CMP) gathers of PP- and PS-waves and re-
an angle to form a TTI medium. To simulate steeply quires the offset-to-depth ratio to reach at least two.
dipping fractures similar to those identified by Angerer Since recording a CMP gather involves moving both the
et al. (2002) on field data, the tilt of the symmetry source and receiver transducers, it is cumbersome and
axis from the vertical was chosen to be 70 (Figure 1). prone to error in positioning. Instead, we decided to col-
The experiments were conducted in the Institute lect shot gathers (Figure 1) and interpret them as CMP
for Experimental Geophysics (IEG) at Colorado School gathers, which is valid for laterally homogeneous media.
of Mines (CSM). The measurements were made only To verify that the lateral heterogeneity of the sam-
in a vertical symmetry plane of the sample, where the ple is negligible, we recorded a constant-offset P-wave
velocities and polarizations are described by TI equa- section (Figure 2). The first arrival (the direct P-wave)
tions, even if the medium as a whole has orthorhom-
PS-wave inversion for tilted TI media 135
recorded with the laser vibrometer was 2 cm (Fig-
Trace no. ure 3b). Since the vibrometer measures the vertical ve-
10 20 locity at a point, it is possible to record closer to the
0 source but the signal quality deteriorates at far offsets.
The two data sets in Figure 3 are quite similar,
but identification of reflection events is hampered by
their interference with the ground roll. To suppress the
0.1 ground roll, we applied standard F-K dip filtering that
significantly improved the quality of the section (Fig-
ure 4). The first arrival is the direct P-wave travel-
Time (ms)
0
-20 -10 0 10 20 30
0
-20 -10 0 10 20 Offset (cm)
10 20
0
0.05 0.05
PP
Time (ms)
0.10 0.10
100
Frequency (kHz)
0.15 0.15
(a) (b)
200
Figure 3. Vertical component of the wavefield. (a) Data
recorded with the P-wave contact transducers; (b) densely
sampled data recorded with the laser vibrometer. The first
arrival is the direct P-wave; the PP-wave reflection from the
bottom of the block (solid line) arrives at a zero-offset time 300
of 0.064 ms.
0.05 0.05
Figure 5. Amplitude spectra of the P-wave traces showing
PP a decrease in frequency with offset.
Time (ms)
0.10 0.10
(a) (b)
0.05 0.05
Time (ms)
Time (ms)
PS PS
Figure 7. Horizontal component of the wavefield. Data Figure 8. Data from Figure 7 after application of F-K filter-
recorded with: (a) the P-wave source transducer and S-wave ing to suppress the ground roll. Negative offsets are displayed
receiver transducer; (b) the S-wave source transducer and the in reverse polarity to facilitate picking of PS traveltimes.
laser vibrometer as the receiver. The converted wave from the
bottom of the model has an asymmetric moveout curve with
the apex at 0.11 ms. The solid lines are the picked traveltimes traveltimes and interpolate/extrapolate them at near
of the PP, PS, and SS arrivals. offsets, we fitted a six-order polynomial to both PP and
PS moveouts using the least-squares method.
The idea of the PP+PS=SS method is to identify
the two data sets recorded using different experimen- two reciprocal PS-wave rays with the same reflection
tal setups. The moveout of the mode-converted PS- (conversion) point and combine their traveltimes with
and SP-waves is strongly asymmetric, with a substan- that of the PP-wave to compute the moveout of the
tial difference between the traveltimes for positive and SS-wave primary reflection. This can be accomplished
negative offsets. Since the model is laterally homoge- by matching the reflection slopes on common-receiver
neous, the PS-wave moveout asymmetry is caused en- gathers of PP- and PS-waves, as illustrated in Fig-
tirely by the oblique orientation of the symmetry axis. ure 9 (Grechka and Tsvankin, 2002). Since the PS rays
Note that the moveout of converted waves is symmetric recorded at points x(3) and x(4) have the same reflection
in any laterally homogeneous medium with a horizontal point as the PP reflection x(1) Rx(2) , the traveltime SS
symmetry plane, including TI models with a vertical of the SS-wave (not physically excited in the survey) is
(VTI) and horizontal (HTI) symmetry axis (Grechka determined from
and Tsvankin, 2000; Dewangan and Tsvankin, 2004a).
The traveltime picks of the PS-wave, marked by SS (x(3) , x(4) ) = tP S (x(1) , x(3) ) + tP S (x(2) , x(4) )
the solid line in Figure 8, were made using the laser vi- tP P (x(1) , x(2) ) , (1)
brometer dataset. Due to the moveout asymmetry, the
minimum PS-wave traveltime is recorded at an offset of where tP S and tP P are the traveltimes of the PS and
x=6 cm where the wavelet reverses its polarity. To fa- PP reflections, respectively. Note that application of
cilitate visual correlation of PS traveltimes, we removed this technique requires correlating PP and PS reflection
this polarity reversal from the sections in Figure 8. events and picking their traveltimes, although explicit
The PP-wave primary reflection can be identified velocity information is not needed.
even on the horizontal component around the zero-offset Alternatively, PP and PS arrivals with the same
time tP 0 = 0.064 ms. It may also be possible to ten- reflection point can be found by computing the time SS
tatively pick the SS-wave reflected arrival but, as ex- in equation (1) for each desired SS-wave shot-receiver
pected, in Figure 8 it is much weaker than the converted pair (x(3) , x(4) ) and a wide range of the coordinates (x(1)
modes. and x(2) ) (Grechka and Dewangan, 2003; Dewangan and
Tsvankin, 2004a):
min
3.3 Data processing SS (x(3) , x(4) ) = tP S (x(1) , x(3) )
x , x(2)
(1)
pPP ( x (1), x (2) ) = pPS ( x (1), x (3) ) pPP ( x (2), x (1) ) = pPS ( x (2), x (4) ) 0.17
t t
t PS ( x (2), x (4) )
t PS ( x (1), x (3) )
0.165
Time (ms)
x (1) x (2)
0.155
x (1) x (4) x (3) x (2)
0.15
P S S P
0.145
0.14
0 5 10 15
SS offset (cm)
R
d
tPS / t0
0.1
/z
0.6
PS
x
0.05
the SS-wave traveltime as a function of the SS-wave 0.8
offset xSS = |x(3) x(4) | (Figure 10). Then, conven- 0 1
tional hyperbolic velocity analysis was applied to the
0.05 1.2
constructed SS arrivals to estimate their stacking ve- 0 5 10 15 0 5 10 15
SS offset (cm) SS offset (cm)
locity (Vnmo,S 1780 m/s) and zero-offset traveltime
(tS0 0.149 ms).
Next, we followed the methodology of Dewangan Figure 11. Moveout-asymmetry attributes of the PS-wave
and Tsvankin (2004a) in computing the time and offset computed from equations (3) and (4). (a) The time asym-
asymmetry attributes of the PS-wave: metry tP S normalized by the zero-offset PS-wave time; (b)
the corresponding offset asymmetry xP S normalized by the
tP S (x(3) , x(4) ) = tP S (x(1) , x(3) ) tP S (x(2) , x(4) ) ; (3) layer thickness.
TTI parameters
parameters and the layer thickness z:
2
n o
m VP 0 , VS0 , , , , z , (6)
Time (ms)
is the direct P-wave followed by the relatively weak di- P
rect S-wave. The solid line marks the P-wave arrival
time computed for the inverted TTI model using equa-
tions (9) and (10). Evidently, the estimated model pa-
rameters accurately predict the P-wave velocity in the
transmission experiment that includes a wider range of
propagation directions compared to the reflection data
set. 0.10
It is interesting that the wavefront of the direct S-
wave arrival in Figure 13 exhibits a cusp (triplication)
at oblique angles with the symmetry axis. The existence
and size of the cusp is mostly governed by the magni-
tude of the anisotropic parameter (VP 0 /VS0 )2 ()
(Tsvankin, 2001; Thomsen, 2002) that reaches 1.24 for
our model. While SV-wave cusps in TI media are well
understood theoretically, their experimental observa- Figure 13. Transmitted wavefield excited by the P-wave
transducer and recorded by the laser vibrometer at the top
tions are rare (e.g., Slater et al., 1993).
of the model. The solid line is the P-wave traveltime modeled
To identify the cusp more clearly and measure the using the inverted parameters from Figure 12.
transmitted shear-wave traveltimes, we performed an-
other transmission experiment, this time with the S-
wave transducer as the source (Figure 14b). The trans- Receiver coordinate (cm) Receiver coordinate (cm)
0 10 0 10 20
mitted wavefield was also computed (Figure 14a) for the 0.04 0.04
Time (ms)
0.08 0.08
tudes, the agreement between the measured and mod-
eled wavefields in Figure 14 is excellent.
0.10 0.10
The spatial extent of the cusp in Figure 14 is signif-
icantly larger than that predicted by the group-velocity
surface (i.e., by ray-theory modeling). This is consis- 0.12 0.12
ABSTRACT
x(1) x(3) x(4) x(2) bottom of the overburden in such a way that they share
the same ray segments in the overburden (Figure 1).
Suppose the dipping event was excited at loca-
tion x(1) and recorded at location x(2) (or vice versa),
and the reflection from the bottom of the overburden
was acquired for a sufficiently wide range of source-
receiver offsets. Following the methodology of Grechka
and Tsvankin (2002b), we form a common-receiver
T R gather of the dipping event at point x(2) and deter-
mine the time slope on this gather at point x(1) . Since
the slowness vector is equal to the gradient of the trav-
elime surface, the estimated time slope coincides with
the ray parameter (horizontal slowness) of the reflection
Q x(1) T QR x(2) at the source location x(1) . We then use
the same algorithm to evaluate the time slopes of the
overburden reflections excited at x(1) and recorded at
Figure 1. 2D ray diagram of the layer-stripping algorithm
designed to find the interval traveltimes and source/receiver
different locations along the line. For a certain receiver
coordinates in the layer immediately above the dipping re- location x(3) , the time slope (ray parameter) of the over-
flector. Points T and R are located at the bottom of the burden reflection from x(1) to x(3) coincides with that
horizontally layered overburden; each layer in the overbur- of the dipping event,
den is assumed to have a horizontal symmetry plane. The
reflections from the dipping interface (x(1) T QR x(2) ) and
teff (x(1) , x(2) ) tovr (x(1) , x(3) )
(1)
= ; (2)
the bottom of the overburden (x(1) T x(3) ) share the same x x(1)
downgoing leg (x(1) T ). The upgoing leg of the dipping event the superscripts eff and ovr refer to the traveltimes
(R x(2) ) coincides with a leg of another overburden reflection, of the dipping event and the overburden reflection, re-
x(2) R x(4) . spectively. The identical ray parameters mean that the
reflections x(1) T QR x(2) and x(1) T x(3) share the same
given by the following Dix-type equation (Alkhalifah downgoing leg x(1) T in the overburden (Figure 1).
and Tsvankin, 1995; Tsvankin, 2001): Repeating this procedure at point x(2) and match-
ing the time slopes,
N
2 1 X (i) 2 (i)
Vnmo (N ) = Vnmo (p) t0 (p) , (1) teff (x(2) , x(1) ) tovr (x(2) , x(4) )
t0 (2)
= , (3)
i=1 x x(2)
where N is the number of layers, t0 is the zero-offset we find the overburden reflection x(2) R x(4) that has the
(i)
traveltime, t0 is the interval traveltime along the zero- same leg R x(2) as the dipping event x(1) T QR x(2) .
(i)
offset ray in layer i, and Vnmo is the interval NMO ve- Since the overburden is laterally homogeneous and
locity. The difference between equation (1) and the con- has a horizontal symmetry plane, the raypath of any
ventional Dix formula is in the meaning of the interval pure-mode wave reflected from the bottom of the over-
NMO velocity; it has to be computed for an imaginary burden is symmetric with respect to the reflection point.
reflector orthogonal to the slowness vector of the zero- Therefore, the interval traveltime in the layer above the
offset ray in each layer. Therefore, determination of the dipping reflector can be expressed as
interval NMO velocity in layer N from equation (1) in-
tint (T, R) = teff (x(1) , x(2) )
volves computing NMO velocities from non-existent dip-
ping interfaces in the overburden. While this operation 1 ovr (2) (4)
t (x , x ) + tovr (x(1) , x(3) ) . (4)
is trivial for isotropic media (if the ray parameter is 2
known), in the presence of anisotropy it requires an es-
The symmetry of the reflection raypath in the overbur-
timate of the relevant medium parameters in the whole
den also implies that the interval traveltime is obtained
overburden (Alkhalifah and Tsvankin, 1995).
for the source-receiver pair [T, R] with the horizontal
coordinates
2.1 Layer-stripping algorithm x(1) + x(3)
x(2) + x(4)
T = , R = . (5)
Here, we show how a variation of the PP+PS=SS 2 2
method (Grechka and Tsvankin, 2002b; Grechka and Essentially, the algorithm outlined above per-
Dewangan, 2003) can be used to obtain the exact inter- forms kinematic downward continuation of the wavefield
val traveltime-offset function in the layer immediately through a laterally homogeneous overburden. As in the
above the target dipping reflector. The idea is to com- PP+PS=SS method, this continuation procedure does
bine the reflections from the dipping interface and the not require knowledge of the velocity model and is not
146 P. Dewangan & I. Tsvankin
0.3
VTI overburden
0.4
VTI
Time (s)
0.5
sym 25 o
TTI
is
0.6
. ax
=
0.7
effective
0.8
=10o
Figure 2. Transversely isotropic model used to test the
2 1 0 1 2
layer-stripping algorithm for pure modes. The first layer has Offset (km)
a vertical symmetry axis (VTI medium) and the following pa-
rameters: the symmetry-direction P-wave velocity VP 0 = 2 Figure 3. CMP gathers of the PP-wave reflections from
km/s, the symmetry-direction S-wave velocity VS0 = 1 km/s, the bottom of the second layer (marked overburden) and
the thickness z = 0.25 km, and Thomsen anisotropy param- the dipping reflector (effective) for the model in Figure 2.
eters = 0.2 and = 0.1. The second layer is also VTI with The traveltimes were generated by an anisotropic ray-tracing
VP 0 = 4 km/s, VS0 = 2 km/s, z = 0.25 km, = 0.15, and code.
= 0.05; the third layer is dipping TTI with the symme-
try axis tilted at = 25 , the dip of the bottom = 10 , the top of the target layer is unknown, the goal of the
VP 0 = 4 km/s, VS0 = 2 km/s, z = 0.5 km, = 0.25, and
layer stripping is achieved by obtaining the horizontal
= 0.05. The Thomsen parameters in the TTI layer are
defined with respect to the symmetry axis (Dewangan and
source/receiver coordinates needed to construct the in-
Tsvankin, 2004a,b). terval traveltime function.
By repeating the above procedure to cover the
whole recorded range of the source-receiver offsets
restricted to isotropic media. Also, in contrast to the for the dipping event, we compute the interval PP-
Dix-type equation (1), our algorithm can be applied wave traveltime for a number of the corresponding
to long-offset data and can handle a laterally hetero- source/receiver pairs. These pairs do not necessarily
geneous target layer. form a common-midpoint (CMP), common-shot, or
common-receiver gather and need to be sorted to an-
alyze the interval traveltime in any desired configura-
2.2 Numerical example
tion. This represents a complication compared with the
The layer-stripping algorithm was tested on PP-wave PP+PS=SS method, in which the type of the output
reflection data from the layered TI model in Figure 2. gather can be specified in advance.
The traveltimes from the dipping reflector and the bot- To verify the accuracy of the layer stripping, we
tom of the overburden were computed by anisotropic computed the interval PP-wave traveltime in the dip-
ray tracing with a shot spacing of 25 m and a receiver ping layer and the corresponding source/receiver coor-
spacing of 100 m (Figure 3). To conform with the 2D dinates using ray tracing. The agreement between our
assumptions of the algorithm, the incidence plane co- method and the ray-tracing results is excellent (Fig-
incides with the dip plane of the reflector and contains ures 4 and 5). This and other synthetic tests we per-
the symmetry axis of the TTI layer. formed for a representative set of layered TI models
In our implementation of the layer stripping we confirm that the layer-stripping algorithm is exact and
follow the version of the PP+PS=SS method devel- can be applied for large source-receiver offsets.
oped by Grechka and Dewangan (2003) and Dewan-
gan and Tsvankin (2004a,b). For a given pair of points
[x(3) , x(4) ], this algorithm searches for the coordinates 3 LAYER STRIPPING FOR
x(1) and x(2) that minimize the interval traveltime in MODE-CONVERTED WAVES
equation (4). This procedure was shown to produce
the same results as the methodology based on reflec- 3.1 Layer-stripping algorithm
tion slopes discussed above (Dewangan and Tsvankin, In contrast to the generalized Dix equations, our layer-
2004a,b). Substitution of the estimated coordinates x(1) stripping algorithm can be easily adapted for mode-
and x(2) into equations (4) and (5) yields the interval converted (PS or SP) waves. Using the same model as-
traveltime in the target layer for the source and re- sumptions as those in the previous section, we consider
ceiver located at points T and R. Although the depth of
Layer stripping of PP and PS data 147
0.5
P P S S
R (km)
0.4
0.3 T R
0.2 S
P
Q
0.6 0.4 0.2 0 0.2
T (km)
Figure 6. 2D ray diagram of the layer-stripping algorithm
Figure 4. Source and receiver coordinates at the top of the
for PS-waves. The model is the same as that in Figure 1,
target layer obtained for the model in Figure 2 for a range of
with the overburden (the section above points T and R)
surface points [x(3) , x(4) ]. Here and in Figure 5 the triangles
composed of laterally homogeneous layers with a horizontal
are the output of the layer-stripping algorithm; the solid line
symmetry plane. The PS reflection from the dipping inter-
marks the results of ray tracing.
face (x(1) T QR x(2) ) and the PP reflection from the bottom
of the overburden (x(1) T x(3) ) share the same downgoing leg
(x(1) T ). The upgoing leg of the dipping PS event (R x(2) ) co-
0.34 incides with a leg of the overburden SS reflection x(2) R x(4) .
0.32
but now the downgoing leg represents a P-wave, while
the upgoing leg is an S-wave (Figure 6). By matching
0.3 the time slopes at point x(1) , we identify the overburden
Time (s)
Figure 5. Interval PP-wave traveltime in the target layer Our implementation of this layer-stripping algorithm
as a function of offset for the source/receiver pairs from Fig- for PS-waves is similar to that described above for PP-
ure 4. waves.
Figure 7 shows ray-traced CMP gathers of the tar-
the PS-wave converted at a dipping reflector overlaid get PS event and the pure-mode reflections from the
by a stack of horizontal layers with a horizontal sym- bottom of the overburden for the model in Figure 2.
metry plane (Figure 6). Since the upgoing leg of the Note the the PS-wave moveout is asymmetric (i.e., the
PS mode represents a shear wave, the algorithm has traveltime does not stay the same when the source and
to operate with both PP- and SS-waves reflected from receiver positions are interchanged) because of com-
the bottom of the overburden. In the absence of shear- bined influence of the reflector dip ( = 10 ) and the tilt
wave excitation, the needed SS traveltimes (tovr ) can of the symmetry axis ( = 25 ). This moveout asymme-
SS
be obtained by applying the PP+PS=SS method to the try, however, is handled by our layer-stripping method
PP and PS data reflection data (Grechka and Tsvankin, that relies only on the symmetry of the reflection ray-
2002b; Grechka and Dewangan, 2003). paths of the pure-mode reflections in the overburden.
As in the previous section, we consider the reflec- The layer-stripped interval PS traveltimes and the
tion raypath x(1) T QR x(2) from the dipping interface, corresponding source/receiver coordinates are close to
148 P. Dewangan & I. Tsvankin
0 0.4
PP 0.39
0.5
Time (s)
Time (s)
SS 0.38
1 PS
0.37
1.5
2 1 0 1 2
0.36
Offset (km) 0.8 0.6 0.4 0.2 0 0.2 0.4
Figure 7. CMP gathers of reflected waves for the model in
Offset / thickness
Figure 2 computed by anisotropic ray tracing. The PS-wave
Figure 9. Interval PS-wave traveltime in the target layer as a
is converted at the dipping interface, while the PP- and SS-
function of offset for the source/receiver pairs from Figure 8.
waves are reflected from the bottom of the overburden.
REFERENCES
Alkhalifah, T., 1997, Velocity analysis using nonhyperbolic
moveout in transversely isotropic media: Geophysics, 62,
18391854.
Alkhalifah, T., and Tsvankin, I., 1995, Velocity analysis in
transversely isotropic media: Geophysics, 60, 15501566.
150 P. Dewangan & I. Tsvankin
CWP-509
ABSTRACT
Field records for small source-receiver offsets often contain intensive converted
PS-waves that cannot be generated in laterally homogeneous isotropic models.
Among the most likely physical reasons for this converted energy is the presence
of anisotropy on either side of the reflector. Here, we study the small-angle
reflection coefficients of the split converted PS1 - and PS2 -waves (RP S1 and
RP S2 ) for a horizontal interface separating two transversely isotropic media
with arbitrary orientations of the symmetry axis.
The normal-incidence reflection coefficients RP S1 (0) and RP S2 (0) vanish when
both halfspaces have a horizontal symmetry plane, which happens if the symme-
try axis is vertical or horizontal (i.e., if the medium is VTI or HTI). For a tilted
symmetry axis in either medium, however, the magnitude of the reflection coeffi-
cients can reach substantial values close to 0.1, even if the strength of anisotropy
is moderate. To study the influence of the orientation of the symmetry axis and
the anisotropy parameters, we develop concise weak-contrast, weak-anisotropy
approximations for the reflection coefficients and compare them with exact nu-
merical results.
In particular, the analytic solutions show that the contributions of the Thomsen
parameters and to the coefficients RP S1 (0) and RP S2 (0) are governed by
simple functions of the symmetry-axis tilt , which have the same form for
both halfspaces. If the symmetry-axis orientation and anisotropy parameters
do not change across the interface, the normal-incidence reflection coefficients
are insignificant, regardless of the strength of the velocity and density contrast.
The AVO (amplitude variation with offset) gradients of the PS-waves are mostly
influenced by the anisotropy of the incidence medium that causes shear-wave
splitting and determines the partitioning of energy between the PS1 and PS2
modes.
Because of their substantial amplitude, small-angle PS reflections in TI media
contain valuable information for anisotropic AVO inversion of multicomponent
data. Our analytic solutions provide a foundation for linear AVO-inversion algo-
rithms and can be used to guide nonlinear inversion based on the exact reflection
coefficients.
Key words: converted wave, transverse isotropy, TTI media, reflection coef-
ficient, AVO intercept, AVO gradient, azimuthal AVO
0.2
2=0.4
(0)
0.1
PS
=0.2
R
2
0
=0
2
0 20 40 60 80
2 ()
Figure 2. For an isotropic incidence halfspace overlying a
TTI reflecting halfspace, the PS-wave at normal incidence Figure 4. The normal-incidence PS-wave reflection coeffi-
is polarized (vector g P S ) in the symmetry-axis plane. For cient for an isotropic/TTI interface as a function of tilt 2
oblique incidence, we analyze the two components of the PS- of the symmetry axis and the parameter 2 . The solid lines
wave (PSV and PSH) separately. mark the exact RP S (0), the dashed lines represent the lin-
earized approximation (11). The parameters of the incidence
isotropic halfspace are VP,1 = 2.9 km/s, VS,1 = 1.5 km/s,
0.2 and 1 = 2 gm/cm3 . The parameters of the reflecting TTI
halfspace are VP 0,2 = 3.3 km/s, VS0,2 = 1.8 km/s, 2 =
2.2 gm/cm3 , 2 = 0.2, 2 = 0.11, and 2 = 30 .
0.15
from the vertical and horizontal directions. Therefore,
(0)
2 ()
If the azimuth of the symmetry axis is the same above
and below the reflector (1 = 2 ), the vertical plane
that contains both symmetry axes represents a plane
of symmetry for the whole model. In this case, the P-
Figure 5. Exact coefficient RP S (0) for an isotropic/TTI
wave at normal incidence excites only one (PS1 ) wave
interface as a function of the tilt 2 and the parameter 2 .
Except for 2 and 2 = 0.2, the model parameters are the
polarized in the symmetry-axis plane:
same as in Figure 4. g2
RP S1 (0) = { sin 21 [cos 21 (1 1 ) + 1 ]
4(1 + g)
+ sin 22 [cos 22 (2 2 ) + 2 ]} (13)
1.5
sin22.(1cos22) g2
1
dVP,1 ()
=
4(1 + g) VP 0,1 d =0
1 1 dVP,2 ()
+ , (14)
VP 0,2 d =0
3 AVO GRADIENTS
The AVO gradients of the split PS-waves can be com-
puted numerically by estimating the best-fit initial slope
of the exact reflection coefficient expressed as a function
of sin . In the linearized weak-anisotropy, weak-contrast
approximation, the gradient G is obtained explicitly as
the multiplier of sin [equation (10)]. The linearized ap-
proximations for the PS-wave AVO gradients are given
in Appendix B.
0.3
a) 2=0 ISO/TTI 0.2 b) ISO/TTI
ISO/ISO 2=90
0.2 ISO/ISO
0.1
2=60
PSH
RPSV
0.1
0
R
0 2=60
0.1 =0
2=90 2
0.1
0.2
0.2
20 0 20 20 0 20
() ()
0.25
c) d)
0.05
ISO/ISO
0.3 ISO/ISO
GPSV
GPSH
0
0.35
ISO/TTI
ISO/TTI 0.05
0.4
0 100 200 300 0 100 200 300
2 () 2 ()
Figure 8. Exact PS-wave reflection coefficients (top row) and AVO gradients (bottom row) for isotropic/isotropic and
isotropic/TTI interfaces. The velocity and density contrasts for both models are the same: = 0.2 gm/cm 3 , =
2.1 gm/cm3 , = 0.3 km/s, = 1.65 km/s, = 0.4 km/s, = 3.1 km/s, and g = 1.88. The anisotropy parameters of
the reflecting TTI halfspace are 2 = 0.3, 2 = 0.15, 2 = 0.11, and 2 = 60 .
rived in Appendix B for the special case of the inci- is governed only by and , while the SH-wave velocity
dence plane containing the symmetry axis of the upper is controlled by (Tsvankin, 2001).
halfspace. We also present numerical results for an arbi- As shown earlier, when the incidence medium is
trary orientation of the symmetry axis of the incidence isotropic, the gradient GP S1 contains both isotropic
medium. and anisotropic terms, while GP S2 is purely anisotropic.
In agreement with the linearized approxima- If the incidence halfspace is TTI with the symmetry
tions (B3) and (B4), the anisotropy parameters 1 and axis not confined to the incidence plane, then there
1 influence only GP S1 while 1 influences only GP S2 are no purely isotropic terms in either gradient, as
(Figures 9 and 10). To explain this result valid for an demonstrated for the special case of the HTI reflect-
arbitrary orientation of the symmetry axis of the inci- ing medium by equations (B5) and (B6). The influ-
dence halfspace, note that the PS1 -wave would be the ence of the anisotropy in the upper halfspace leads to
PSV mode if the symmetry axis were vertical, and the a substantial deviation of the gradients computed for
PS2 -wave would be the PSH (transversely polarized) a TTI/TTI interface from those for the corresponding
mode. The P- and SV-wave propagation in TI media isotropic model (compare Figures 11 and 8).
PS-wave AVO in TTI media 159
0 0
a) b)
0.05
0.05 2
PS1
PS1
0.1 2
0.1
G
G
1
1 2
0.15 0.15
0.2 0.2
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
a) b)
2
1
0.2 0.2
PS2
PS2
1
2
2
G
0.1 0.1
1
0 0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
0.5 0.5
a) b)
TTI/TTI TTI/TTI
GPS1
GPS2
0 0
ISO/ISO
ISO/ISO
0.5 0.5
0 100 200 300 0 100 200 300
1 () 1 ()
Figure 11. Exact AVO gradients of the (a) PS1 - and (b) PS2 -waves as functions of the azimuth 1 for a TTI/TTI in-
terface (black solid line) and isotropic/isotropic interface (black dashed line). The parameters of the isotropic model are
VP,1 = 2.9 km/s, VS,1 = 1.5 km/s, 1 = 2 gm/cm3 , VP,2 = 3.3 km/s, VS,2 = 1.8 km/s, and 2 = 2.2 gm/cm3 . For the
TTI/TTI interface, the parameters are VP 0,1 = 2.9 km/s, VS0,1 = 1.5 km/s, 1 = 2 gm/cm3 , 1 = 0.2, 1 = 0.1, 1 = 0.1, 1 =
60 , VP 0,2 = 3.3 km/s, VS0,2 = 1.8 km/s, 2 = 2.2 gm/cm3 , 2 = 0.3, 2 = 0.15, 2 = 0.11, 2 = 30 , and 2 = 30 .
reflecting TTI medium. The normal-incidence reflection gradient of the PS1 (PSV)-wave includes both purely
coefficient RP S (0) vanishes when the reflecting TI half- isotropic and anisotropic terms, while the PS2 (PSH)-
space is VTI or HTI, because in that case the model wave AVO gradient is entirely anisotropic. For an arbi-
as a whole has a horizontal symmetry plane. The coef- trary symmetry-axis orientation, however, neither AVO
ficient RP S (0) rapidly increases as the symmetry axis gradient contains purely isotropic terms since the con-
deviates from both the vertical and horizontal direc- tributions of the velocity and density contrasts are mul-
tions. Closed-form approximations and numerical mod- tiplied with functions of the symmetry-axis azimuth
eling show that RP S (0) can reach values close to 0.1 for 1 . The AVO gradients are mostly influenced by the
moderate values of the anisotropy parameters typical anisotropy in the incidence medium that causes shear-
for shale formations. When the tilt 2 of the symmetry wave splitting and pronounced azimuthal variation of
axis is relatively small, the coefficient RP S (0) is mostly the reflection coefficients of both PS-waves.
controlled by 2 , with the contribution of 2 becoming The linearized approximations developed here not
dominant for larger values of 2 . only provide physical insight into the behavior of the
When both halfspaces are anisotropic (TTI), a P- PS-wave reflection coefficients, but can be also used
wave at normal incidence excited two split PS-waves to quickly evaluate the PS-wave amplitudes for a wide
(PS1 and PS2 ) with the polarizations governed by range of TTI models. Potentially, these analytic expres-
the orientation of the symmetry axis in the incidence sions can help in the AVO inversion of PP and PS data
medium. The reflection coefficients of both PS-waves are for TTI media that can be implemented using the ap-
functions of the difference between the azimuths of the proach suggested by Jlek (2001) for anisotropic models
symmetry axes (2 1 ) and do not depend on either with a horizontal symmetry plane.
azimuth individually. The linearized normal-incidence
reflection coefficient goes to zero when the two halfs-
paces have the same orientation of the symmetry axis 5 ACKNOWLEDGMENTS
and the same parameters and , even though there
may be a jump in the velocities and densities across the We are grateful to members of the A(nisotropy)-Team
interface. of the Center for Wave Phenomena (CWP), Colorado
We also discussed the azimuthally varying AVO School of Mines and P. Jlek (BP Americas) for helpful
gradients of PS-waves responsible for the small-angle re- discussions. The support for this work was provided by
flection response. If the symmetry axis of the incidence the Consortium Project on Seismic Inverse Methods for
halfspace is confined to the vertical incidence plane Complex Structures at CWP and by the Chemical Sci-
(or if the incidence halfspace is isotropic), the AVO
PS-wave AVO in TTI media 161
ences, Geosciences and Biosciences Division, Office of Vavrycuk, V. 1999, Weak-contrast reflection/transmission
Basic Energy Sciences, U.S. Department of Energy. coefficients in weakly anisotropic elastic media: P-wave
incidence: Geophys. J. Int., 138, 553-562.
Vavrycuk, V., and Psenck, I. 1998, PP-wave reflection coef-
ficients in weakly anisotropic elastic media: Geophysics,
REFERENCES 63, 2129-2141.
Aki, K., and Richards, P. G. 1980, Quantitative Seismology:
Theory and methods: Vol.1. W. N. Freeman & Co., San
Francisco.
Alvarez, K., Donati, M., and Aldana, M. 1999, AVO analysis
for converted waves: 69th Annual Internat. Mtg., Soc.
Expl. Geophys., Expanded Abstracts, 876-879.
Artola, F. V. A., Leiderman, R., Fontoura, S. A. B., and
Silva, M. B. C. 2003, Zero-offset C-wave reflectivity in
horizontally layered media: 73rd Annual Internat. Mtg.,
Soc. Expl. Geophys., Expanded Abstracts, 761-764.
Banik, N. C. 1987, An effective anisotropy parameter in
transversely isotropic media: Geophysics, 52, 1654-1664.
Donati, S. M. 1998, Making AVO analysis for converted
waves a practical issue: 68th Annual Internat. Mtg., Soc.
Expl. Geophys., Expanded Abstracts, 2060-2063.
Jech, J., and Psenck, I. 1989, First-order perturbation
method for anisotropic media: Geophys. J. Int., 99, 369-
376.
Jlek, P. 2000, Converted PS-wave reflection coefficients in
anisotropic media: CWP Annual Project Review, CWP-
340, 275-299.
, 2001, Modeling and inversion of converted-wave reflec-
tion coefficients in anisotropic media: A tool for quanti-
tative AVO analysis: Ph.D. Thesis, Center for Wave Phe-
nomena, Colorado School of Mines, Golden, Colorado.
Larsen, A. J., Margrave, F. G., and Lu, H-X. 1999, AVO
analysis by simultaneous P-P and P-S weighted stacking
applied to 3C-3D seismic data: 69th Annual Internat.
Mtg., Soc. Expl. Geophys., Expanded Abstracts, 721-724.
Nefedkina, T., and Buzlukov, V. 1999, Seismic dynamic in-
version using multiwave AVO data: 69th Annual Inter-
nat. Mtg., Soc. Expl. Geophys., Expanded Abstracts,
888-891.
Psenck, I., and Vavrycuk, V. 1998, Weak contrast PP wave
displacement R/T coefficients in weakly anisotropic elas-
tic media: Pure Appl. Geophys., 151, 699-718.
Ruger, A. 1996, Reflection coefficients and azimuthal AVO
analysis in anisotropic media: Ph.D. Thesis, Center for
Wave Phenomena, Colorado School of Mines, Golden,
Colorado.
, 1997, P-wave reflection coefficients for transversely
isotropic models with vertical and horizontal axis of sym-
metry: Geophysics, 62, 713-722.
, 1998, Variation of P-wave reflectivity with offset and
azimuth in anisotropic media: Geophysics, 63, 935-947.
Shuey, R. T. 1985, A simplification of Zoeppritz-equations:
Geophysics, 50, 609-614.
Thomsen, L. 1993, Weak anisotropic reflections: In: Off-
set dependent reflectivity (Castagna and Backus, Eds.),
SEG, Tulsa.
, 2002, Understanding seismic anisotropy in exploration
and exploitation: Distinguished Instructor Short Course
(DISC), No. 5, SEG and EAGE.
Tsvankin, I. 1995, Seismic wavefields in layered isotropic me-
dia: Samizdat Press, Colorado School of Mines.
, 2001, Seismic signatures and analysis of reflection
data in anisotropic media: Elsevier Science.
162 J. Behura & I. Tsvankin
APPENDIX A: PERTURBATION APPROACH APPLIED TO THE
REFLECTION/TRANSMISSION PROBLEM
The approximate linearized reflection and transmission coefficients are derived by using an isotropic full space as
the reference medium. A horizontal planar interface (reflector) divides the full space into two halfspaces, which are
perturbed to obtain two weakly anisotropic media:
(I) (I)
aijkl = a0ijkl + aijkl , (A1)
(I) 0 (I)
= + , (A2)
(I)
|aijkl | << ka0ijkl k , (A3)
(I) 0
| | << . (A4)
(I)
In equation (A4), the index I (I = 1, 2) stands for the incidence and reflecting halfspaces, respectively, and a ijkl
and 0 are the density-normalized stiffness coefficients and density of the background isotropic medium. Since the
perturbations from the isotropic background in both halfspaces are small, the approximation involves both weak
anisotropy and weak elastic contrast between the halfspaces. Using these approximations, the polarization (p (N ) ) and
slowness vectors (g(N ) ) can be linearized as follows:
sin2 2 sin 22 2
GP SH = G P S2 =
g
g sin2 2 sin 22
+ {[2g + (1 + 2g) cos 22 ](2 2 ) 2 } . (B2)
4(1 + g)
The AVO gradients for a TTI/TTI interface are given here only for the incidence plane containing the symmetry
axis of the incidence TTI halfspace (1 = 0 ):
2 (2 + g) 2 sin2 2 sin2 2 2
GP S1 =
g 2g g
g
+ {4(1 + g)(1 1 ) + 4(1 + g) cos 41 (1 1 )
16(1 + g)
8 cos 21 1 + 32 2 + 2g(2 2 ) + 2 cos 22 (2 + 22 )
+4 cos 22 sin2 2 {2 + [2g + (1 + 2g) cos 22 ](2 2 )}
+(3 2g) cos 42 (2 2 )} , (B3)
sin2 2 sin 22 2
GP S2 =
g
2
g sin 2 sin 22
+ {[2g + (1 + 2g) cos 22 ](2 2 ) 2 } . (B4)
4(1 + g)
If the symmetry axis of the upper halfspace deviates from the incidence plane, fully linearized AVO gradients
can still be derived for the special case of the incidence HTI medium:
2 cos 1 (2 + g) cos 1 2 sin2 2 sin(2 1 ) sin 2 2
GP S1 =
g 2g g
g 2
+ {4 sin 2 cos(22 1 ) {[2g + (1 + 2g) cos 22 ](2 2 ) 2 }
16(1 + g)
+ cos 1 [81 + 161 + 32 2 + 2g(2 2 )
+2 cos 22 (2 + 22 ) (3 + 2g) cos 42 (2 2 )]} , (B5)
ABSTRACT
Recently, curvelets have been introduced in the field of applied harmonic anal-
ysis and shown to optimally sparsify smooth (C2 , i.e., twice continuously differ-
entiable) functions away from singularities along smooth curves. In addition, it
was shown that the curvelet representation of wave propagators is sparse. Since
the wavefronts in seismic data lie mainly along smooth surfaces (or curves in
two dimensions), and since the imaging operator belongs to the class of op-
erators that is sparsified by curvelets, curvelets are plausible candidates for
simultaneous sparse representation of both the seismic data and the imaging
operator. In this paper, we study the use of curvelets in pre-stack time migra-
tion, and show that simply translating, rotating and dilating curvelets according
to the pre-stack map time-migration equations we developed earlier, combined
with amplitude scaling, provides a reasonably accurate approximation to time-
migration. We demonstrate the principle in two dimensions but emphasize that
extension to three dimensions is possible using 3D equivalents of curvelets. We
treat time-migration in an attempt to learn the basic characteristics of seismic
imaging with curvelets, as a preparation for future imaging in heterogeneous
media with curvelets.
Key words: Pre-stack time-migration, common-offset, curvelets, map migra-
tion, translation, rotation, dilation
a)
b)
c)
d)
Figure 3. First column: curvelets in the spatial domain. Second column: associated spatial lattices, and spatial cells colored
according to the value of the coefficient. Third column: amplitude spectra and frequency-domain tilings. First row: a curvelet.
Second row: curvelet from Figure (a) with different translation indices. Third row: curvelet from Figure (a) with a different
angular index. Fourth row: curvelet from Figure (a) with a different (higher) scale index (here the translation indices and the
angular index are in fact also different, since they depend on the scale index).
the associated spatial lattice for each of the curvelets, shows a translated version of the curvelet in Figure 3a;
where the centers of the cells are the actual possible the spectral tile is the same, but the spatial tile has
locations of the centers of the curvelets in space. Here changed, i.e., indices j and l are held constant, but the
the spatial cells on the spatial lattice are colored ac- translation indices m1 and m2 are different. Figure 3c
cording to the coefficient of the curvelet (here always shows a rotated version of the curvelet in Figure 3a;
one); black equals one and white equals zero. Figure 3b now the spatial location is the same, but the spectral
170 H. Douma & M.V. de Hoop
Figure 4. Top row: a curvelet with a dominant frequency of about 30 Hz (left, shown in depth z = vt u /2, for consistency), the
normalized absolute value of the coefficient on the spatial lattice (middle), and its amplitude spectrum (right). Bottom row:
CO Kirchhoff migration of the curvelet in the top row. The middle panel in this row shows the coefficients on the spatial lattice
in the lower left quadrant of the leftmost panel (indicated with the dotted lines in the leftmost panel) for each of the numbered
wedges (labeled 1 to 4) in the spectrum (right). The Kirchhoff migration of a curvelet determines only part of the isochron,
and shows that a curvelet is not mapped onto one other curvelet, but rather several other curvelets. The resulting curvelets
are clustered together both in space and spectrum, at least for the constant media case shown here, indicating that a curvelet
remains curvelet-like after CO Kirchhoff time migration.
tile has moved within the same concentric squares, i.e., Curvelets remain curvelet-like when subjected
indices j, m1 , and m2 are the same, but index l has to our class of operators
changed. Notice how the spatial lattice changes as we
The action of operators belonging to the class of Fourier
change the angular index l. Finally, Figure 3d shows
integral operators that can be sparsely represented using
a dilated version of the curvelet shown in Figure 3a;
curvelets, which includes the CO time migration oper-
the spatial location is the same, but the spectral tile
ator, can be described in terms of propagation of sin-
has moved outward into a neighboring annulus (or sub-
gularities along a Hamiltonian flow. The remark in the
band), i.e. the scale index j is increased by one. Since
introduction thus applies: The action on a curvelet of a
the neighboring annulus is subdivided into more wedges,
particular scale can be approximated by flowing out the
the angular index l has also changed, although the di-
center of the curvelet in phase space in accordance with
rection of the curvelet is essentially the same. Similarly,
the Hamiltonian associated with the medium filtered for
since the larger scale has a finer spatial sampling asso-
this scale. This means that, in the appropriately filtered
ciated with it, the translation indices m1 and m2 have
media, curvelets remain fairly localized in both the spa-
also changed, while the curvelet location is essentially
tial domain and the spectral domain. Hence, the propa-
the same.
gated curvelet can be constructed by using neighboring
curvelets only, where neighboring is understood in the
context of phase-space; i.e., a neighboring curvelet is a
Common-offset pre-stack time migration with curvelets 171
Figure 5. Synthetic common-shot gather with cusped wavefront: original (a), reconstructed using only the 0.25 % largest
curvelet coefficients (b), and the difference (c). The reconstruction with 0.25 % of the curvelets is almost identical to the
original common-shot gather. In this example, using only 0.25 % of the curvelets results in about 30 times fewer curvelets than
input samples in the gather.
curvelet that is close in the spatial domain and has ori- they remain curvelet-like) after pre-stack time migra-
entation close to the orientation of the curvelet that is tion.
propagated along the central ray.
For homogeneous media, filtering is unnecessary,
and the above observation applies to the same medium Curvelets as building blocks of seismic data
for curvelets of all scales. To illustrate this, Figure 4
shows the result of CO Kirchhoff migration of a curvelet Seismic reflections in seismic data lie primarily along
[taken from Douma and de Hoop (2004)]. The top row smooth surfaces (or curves in two dimensions). Even
shows the input curvelet in space (the vertical axis was diffractions from discontinuities in the earths subsur-
converted to depth using z = vtu /2 for convenience), face, such as edges of geologic interfaces caused by fault-
and its associated amplitude spectrum. Again the coef- ing, lie along smooth surfaces. This is a direct conse-
ficient of the curvelet is shown in the middle panel, just quence of the wave-character of seismic data. As men-
as in Figure 3. The left-most panel of the bottom row tioned in the introduction, it is intuitive that curvelets
shows the CO Kirchhoff migrated curvelet. Notice how can be used to sparsely represent seismic data, since
the migrated curvelet is clearly localized in space and curvelets provide the sparsest representations of smooth
determines only part of the isochron, in sharp contrast (C 2 ) functions away from edges along piecewise C 2
to the whole isochron if a single sample (or a spike) curves (Candes & Donoho, 2004b). Throughout this
would be used as input to the migration. This confirms work, we simply adopt this intuition and illustrate this
that curvelets are indeed a more appropriate choice for with a simple synthetic example below.
building blocks of seismic data than are spikes (that Figure 5a shows part of a synthetic common-shot
are currently used to represent seismic data). The spec- gather, where the wavefront has a cusp. This data re-
trum of the migrated curvelet (bottom right) is clearly lates to a model with a syncline shaped reflector. Figure
localized after the migration, and overlies four wedges 5b shows the reconstructed gather where only the 0.25 %
in the curvelet tiling of the spectrum, indicating some largest curvelet coefficients were used. For the particu-
leakage into neighboring curvelets in the spectral do- lar example shown, this relates to a compression ratio of
main. The middle panel shows the coefficients for the about 30; i.e., we used 30 times less curvelets than there
spatial area in the lower left quadrant of the leftmost are sample values in the original gather, to reconstruct
figure (outlined by the dotted lines), for the wedges la- the data. From Figure 5c it is clear that the difference
beled 1 through 4. Indeed there is also some leakage between the original and reconstructed data is close to
to neighboring curvelets in space, but again this can be zero. This exemplifies that, using curvelets as building
considered small. This confirms that curvelets remain blocks of seismic data, the data can be sparsely repre-
localized in both the spatial and spectral domain (i.e., sented with curvelets, with much fewer curvelets than
the data has samples, and with essentially no residual,
even in areas where the wavefront has cusps.
172 H. Douma & M.V. de Hoop
Figure 6. Common offset (h = 1000 m) data (a) and migrated data (b) from a syncline shaped reflector embedded in a
constant velocity (v = 2000 m/s) medium, and demigrated and migrated line elements superposed on the data and migrated
data, respectively. The excellent agreement between the demigrated line elements and the seismic data (a), and the migrated line
elements and the migrated data (b), indicate the validity of the common-offset map time-demigration and migration equations,
respectively.
In our example, we have applied a hard thresh- quired to extract the local slopes from the data, such as
olding to the data; we simply did not use 99.75 % of local slant stacking in CDR (Zavalishin, 1981; Harlan &
the curvelets. At first sight one might think that there- Burridge, 1983; Sword, 1987; Riabinkin, 1991), stereoto-
fore the compression ratio should be 400. However, the mography (Billette & Lambare, 1998; Billette et al.,
curvelet transform is redundant, meaning that if all 2003), and parsimonious migration (Hua & McMechan,
curvelets are used to reconstruct the data, there are 2001; Hua & McMechan, 2003), or multidimensional
more curvelets than sample points in the data. Dif- prediction-error filters (Claerbout, 1992, p.186-201) and
ferent digital implementations of the curvelet trans- plane-wave destruction filters (Fomel, 2002; Claerbout,
form have different redundancies (Candes et al., 2005). 1992, p.93-97). Therefore, curvelets provide an appro-
(In this particular example, the apparent compression priate reparameterization of seismic data, that have the
ratio (400) and the associated implied redundancy of wave-character of the data built into them.
400/30 13 is so large only because a lot of zero-
padding was necessary to make the number of samples
in the gather both horizontally and vertically equal to 2D Common-offset map time migration
an equal power of 2; the actual redundancy of the used
transform is about 3.) The hard thresholding that we Douma & de Hoop (2005) present explicit expressions
used in our example, will in practice certainly not be for common-offset map time migration (i.e., migration
ideal to determine the threshold level, especially in a in a medium with constant velocity), that use only
practical situation where we have noise. Here, we re- the slope in a common-offset gather (and the velocity),
frain from any denoising issues, and focus on the imag- rather than the slope in a common-offset gather and the
ing with curvelets. Hence, we assume that an intelligent slope in a common-midpoint gather (and the velocity),
thresholding of the data has already determined the sig- such as the equations presented by Sword (1987, p.22).
nificant curvelet coefficients in the data. The expressions in three dimensions from Douma & de
We emphasize that by using curvelets as building
blocks of seismic data, the local slopes (or directions)
in the data are built into the data representation. Other
than a straightforward projection of the data onto the
curvelet frame (combined with an intelligent threshold-
ing procedure), no additional processing steps are re-
Common-offset pre-stack time migration with curvelets 173
Hoop (2005) simplify in two dimensions to x
2
vtu u
xm = x u , (2)
2 h
v"
2 # ( 2 )
u
u
u vtu u 2h
tm = t 1 t2u , (3)
2h v c
|u 1| |u + 1|
2pu tu
|u 1| + |u + 1|
p m = v"
u 2 # ( 2 )
(4)
t 1 vtu u 2h
u
t2u
2h v
in which z
u = u (pu , u , h)
v
Figure 7. The width of the curvelet in the direction perpen-
u 0 s 1
4
1 64 (p u h)
u
= tu @1 1
u A, (5) dicular to its main ridge determines wavelength c , while
2 2pu h 2u the vertical width of the curvelet determines the wavelength
relevant to seismic migration. It follows that c = cos u
with and hence = c cos u .
u = u (tu , pu , h)
4 2 positive). Figure 6b shows the migrated common-offset
2h 1 2h
t2u + 1 (vpu )2 . (6)
`
= 2
2 equivalent of the data shown in Figure 6a, with the mi-
v tu v
grated line elements, determined using xm , tm , and pm ,
In these expressions xu , tu , and pu = 12 tu /xu are drawn on top of the image. The migrated line elements
the midpoint location, two-way (unmigrated) travel- accurately follow the directions in the image, indicating
time, and the (unmigrated) slope in a CO section, re- that equations (2)-(4) accurately capture the kinematics
spectively, while xm , tm and pm = 21 tm /xm are their of CO time migration in a high-frequency approximate
migrated counterparts. Also, h denotes the half-offset sense. Note that the line elements from diffractions of
and v is the medium velocity. Equations (2)-(4) are ex- the edges of the syncline in Figure 6a, are all mapped to
plicit expressions that determine the migrated reflector the same location but with different orientations. This
coordinates (xm , tm , pm ) from the specular reflection co- is identical to building a (band-limited) delta-function
ordinates (xu , tu , pu ), given h and v. Equations (2)-(4) with plane waves from all directions; i.e., the Fourier
do not use the offset horizontal slowness ph = 21 tu /h, transform of a delta function has all directions.
such that in practice, only pu needs to be estimated, and Knowing that each curvelet has a (few) direction(s)
the slope in a common-midpoint gather can be ignored. associated with it, we aim to replace the line elements
It is possible to derive map migration equations that in Figure 6 with curvelets, and see to what extent mov-
use the offset horizontal slowness ph instead of the ve- ing curvelets around according to the map migration
locity (Sergey Fomel, personal communication). In this equations (2)-(4), gives us a good time migrated im-
way, map time-migration can be done without knowl- age. We are thus aiming to lift the applicability of map
edge of the medium velocity. This idea dates back to the migration beyond velocity model building, and show its
work of Ottolini (1983). In the context of pre-stack time- use for pre-stack (here CO) time imaging with curvelets.
migration with curvelets, this would require 3D equiva- This was also mentioned by Douma and de Hoop (2004).
lents of curvelets for 2D imaging since both pu and ph
would need to be known.
Figure 6a shows a common-offset gather (h = 1000 2.5-D Common-offset time migration with
m) from synthetic data generated from a syncline model curvelets
with constant velocity above the reflector. On top of the
reflections, line elements are drawn tangent to the re- As mentioned in the previous section, the CO map time-
flections. Each of these line elements determines a local migration equations (2)-(4) use the slope in the CO do-
slope, pu , while the center of the line determines the two- main only. This implies that the flow from a curvelet
way traveltime tu and the common-midpoint location in the data domain to the image domain is determined
xu . Using equations (2)-(4) and the velocity v = 2000 by one slope only. Therefore, the following scheme for
m/s, the migrated location (xm , zm ), with zm = vtm /2, time-imaging with curvelets emerges. Using curvelets as
and the local dip angle tan = vpm can be determined building blocks of seismic data, the directions (or local
( is the angle with the horizontal measured clockwise slopes) in the data follow from a straightforward projec-
174 H. Douma & M.V. de Hoop
Figure 8. CO Kirchhoff migration of the curvelet shown in Figure 4 (a), TRD transformation of this curvelet (b), and the
difference (c). The bottom row shows the real part of the spectrum for the CO Kirchhoff migrated curvelet (d), the TRD
transform (e), and the difference (f). The TRD transform gives a good approximation to the Kirchhoff migrated curvelet.
tion of the data onto the curvelet frame. Thresholding of ray from the receiver to the scattering point. The term
the curvelet coefficients then gives the curvelets associ- cos (y, xs , xr ) is usually referred to as the obliquity
ated with the main wavefronts in the data, and thus the factor. In 2D, for a constant background velocity and
directions associated with these wavefronts also. Subse- a CO acquisition geometry, equation (7) simplifies to
quently, these directions can be used in our map migra-
|k| = cos (xm , tm , xu , h), where h is the half offset,
tion equations (2)-(4) (together with an estimate of the v
xu the common midpoint location along the 2D acqui-
velocity), to determine the migrated location and orien- sition profile, and the angle (xm , tm , xu , h) is given by
tation of the curvelet. Hence, these equations determine
a translation and a rotation of the curvelet. 1 2 (xu + h xm )
(xm , tm , xu , h) = tan1
It is known (e.g., Bleistein et al. (2000), p.223 and 2 vtm
p.235) that after pre-stack migration, the length of the
ff
2 (xu h xm )
k-vector changes according to tan1 . (8)
vtm
|k| = y (y, xs , xr ) = cos (y, xs , xr ) , (7) A curvelet has dominant wavelength c in the direc-
v(y) tion orthogonal to the curvelet (see Figure 7). However,
where is the angular frequency, xs,r are the source the dominant wavelength in the vertical direction de-
and receiver locations, y is the output location in the termines the frequency relevant to migration. It follows
image, (y, xs , xr ) is the two-way traveltime from the that c = cos u , which gives
source location xs to the reflector at y to the receiver
= c cos u , (9)
position xr , v(y) is the velocity at output location y,
and (y, xs , xr ) is the half opening-angle between the where u is the phase-angle (measured clockwise posi-
ray from the source to the scattering point, and the tive with the vertical) of the curvelet in the data domain.
Common-offset pre-stack time migration with curvelets 175
Figure 9. Superposition of 8 curvelets (a), the associated amplitude spectrum (b), the CO Kirchhoff migrated result (c), and the
result from the amplitude corrected TRD transform (d). The amplitude corrected TRD transform gives a decent approximation
to the Kirchhoff migrated result, and the interference between different curvelets results in a reasonably smooth isochron. Since
we used only curvelets with k-vectors pointig to the left, only the left part of the isochron is constructed.
Combining equations (7) and (9) it follows that, after dynamics, often referred to as the true amplitude part
migration, the length of the k-vector can be written as of seismic imaging. For constant background velocities
c
q and a CO acquisition geometry, Bleistein et al. (2000)
|k| = cos (xm , tm , xu , h) 1 (v(y)pu )2 , (10) show that the equation for 2.5-D Kirchhoff inversion
v(y)
q for reflectivity is given by [their equation (6.3.25); note
where we used cos u = 1 (v(y)pu )2 . This means the missing division with the velocity v in the phase
function]
that the curvelet needs to be dilated (or stretched) in Z Z
the image domain with dilation factor D, given by 1
(y) = A(y, xu , ; h)ei[rs +rg ]/v . . .
1 2
D= q . (11) uS (xg , xs , ) dxu d, (12)
cos (xm , tm , xu , h) 1 (vpu )2
with
where we specialized to the constant background veloc-
4y3 rs + rg rs2 + rg2
`
ity case. Therefore, CO pre-stack time migration with A(y, xu , ; h) = ...
curvelets can be done by translating and rotating the (v rs rg )3/2
curvelets in the data domain according to the map cos (y, xu , h) ||ei 4 sgn ,
p
(13)
migration equations (2)-(4), and dilating the curvelets
with a stretch factor given by equation (11). In these where y = (y1 , y3 ) is the output location in the image
equations, the parameters xu , tu , and pu are the position domain, xu is the midpoint location along the 2D ac-
(x and t) and direction of the center of the curvelet re- quisition profile, rs,g are the distance from the output
spectively. Throughout this paper, we refer to the trans- location y to the source and receiver at xs = (xu h, 0)
lation, rotation and dilation of a curvelet as the TRD and xr = (xu + h, 0), respectively, and uS is the (singly)
transformation of a curvelet. scattered data. To find the amplitude correction for the
The TRD transformation of a curvelet provides the migration of a curvelet, we would therefore need to
kinematics of imaging with curvelets, but ignores the use a curvelet as scattered data uS , and evaluate the
176 H. Douma & M.V. de Hoop
Figure 10. A shallow dipping curvelet (a) and a steep dipping curvelet (d), their CO Kirchhoff migrated counterparts [(b)
and (e), respectively], and the TRD transformed counterparts [(c) and (f), respectively]. The steeper dipping curvelet is more
strongly bended towards the isochron than is the shallower dipping curvelet.
integrals in equation (12). A natural way to evaluate It is clear that for this particular curvelet, the TRD
the oscillatory integral in equation (12), is to use the transformation of the curvelet gives a reasonable ap-
method of stationary phase [e.g. Bleistein (1984, p.77- proximation to the Kirchhoff result. The maximum am-
81) and Bleistein et al. (2000, p.127-135)]. Such an anal- plitude of the difference between both methods is about
ysis should reveal the amplitude correction for the mi- 20% of the maximum amplitude in the Kirchhoff im-
grated curvelet. At the moment of writing this report, age. From the lines of equal phase in the spectra, we
we have not done such an analysis. Therefore, for the see that the curvelet is slightly bent due to the migra-
purpose of generating the results in this report, we sim- tion, whereas the TRD transformation does not take
ply scale the curvelet with A(y, xu , ; h) knowing that this bending into account (cf. Figures 8d and e). The
ultimately we want to do the stationary phase analysis. main difference in the spectrum occurs on the edges of
the support of the curvelet in the frequency domain,
where this bending is strongest. Overall, for this partic-
Numerical examples ular curvelet, the TRD transformation gives a satisfac-
tory approximate image when compared to the Kirch-
Figure 8a shows the 2.5-D CO Kirchhoff migrated hoff result.
curvelet from Figure 4 in more detail, while Figure 8d Even though Figure 8 shows good results for one
shows the real part of the associated spectrum. We plot (particular) curvelet, this does not show the interfer-
the real part of the spectrum instead of the amplitude ence between different curvelets. Figure 9a shows a su-
spectrum, to see the phase information and amplitude perposition of 8 curvelets, with the same central loca-
information combined in one plot. Figure 8b shows the tion in space, and the same scale index, but different
result of the TRD transformation of the same curvelet, directions (or angular indices). The vertical axis is here
where the translation and rotation is determined by the again converted from two-way traveltime to depth. Fig-
map migration equations (2)-(4), and the dilation is de- ure 9b shows the amplitude spectrum of all 8 curvelets
termined by the stretch factor given in equation (11). combined, and it is clear that we used curvelets that
Figure 8c shows the difference between the Kirchhoff mi- have leftward pointing k-vectors only. Figure 9c shows
grated curvelet and the TRD transformed curvelet. Be- the 2.5-D CO Kirchhoff migrated result, while Figure 9d
fore subtraction, both images were normalized to have shows the result of our TRD transformation combined
maximum amplitude one, such that the difference only with an amplitude scaling given by equation (13). As
shows relative amplitude differences between both im- expected, the Kirchhoff result gives the left part of the
ages. Figures 8e and f show the real part of the am- isochron only. Comparing Figures 9c and d, the TRD
plitude spectra of the images shown in Figures 8b and transformation combined with the amplitude correction
c. of equation (13) gives a reasonable approximation to the
Common-offset pre-stack time migration with curvelets 177
x the significant curvelet coefficients, we apply an inverse
curvelet transform, and transform the resulting curvelet
in the spatial domain according to our TRD transform.
This approach allows us to show the proof of concept
of imaging with curvelets using our TRD transform,
t isochron but does not provide an efficient way of such imaging
with curvelets. Ultimately one would want to calculate
the TRD transform in the curvelet frame, although an
Figure 11. The observed bending of curvelets, mostly for
approach that makes use of the finite support of the
the steeper dips in the data, could possibly be corrected for
curvelets in the spectral domain seems also worth in-
by bending the curvelet towards the isochron as depicted.
vestigating.
Even though the curvelets used in this paper are
Kirchhoff migrated result. The interference of the differ- 2D, they can be extended to higher dimensions. Fig-
ent curvelets (of the same scale) is good, and compares ure 12 shows an example of a 3D curvelet in both the
favorably to the Kirchhoff result, although the compar- spatial domain (a) and the spectral domain (b). In the
ison seems better for the shallower dipping part of the spatial domain, 3D equivalents of curvelets look like cir-
isochron. At the steeper parts of the isochron, the de- cular disks that are smooth along the disk and oscilla-
structive interference between different curvelets away tory orthogonal to the disk. Roughly speaking they are
from the isochron, is somewhat less effective and leaves smoothed circular pieces of a bandlimited plane wave in
the tails of the curvelets somewhat visible. 3D . Therefore, replacing the 2D map migration equa-
In order to see the differences between the results tions with their 3D equivalents, the TRD transform can
for the steeper dipping curvelets and the shallower ones, be extended to 3D, hence allowing 3D CO pre-stack
Figure 10 shows the comparison between the Kirchhoff time-imaging with curvelets. Decomposing a 3D CO vol-
result and our TRD transformation for both a shallow ume of data into 3D curvelets, the same procedure as
and a steep dipping curvelet. Here we did zero-offset outlined in this paper can be used to image 3D seismic
migration, to avoid the curvature of the isochron being data with 3D equivalents of curvelets. Of course, the
different at different locations on the isochron; i.e., a proper 3D amplitude correction should be used in this
semi-circle has constant curvature everywhere. Figure case.
10 shows that the steeper dipping curvelet is signifi- Even though in this paper we show results from
cantly bent towards the isochron, while the shallower migration only, modeling or demigration with curvelets
dipping curvelet is hardly bent at all. As a result, the works in the same way. In this case the TRD transform
TRD transform does better for the shallower dipping is defined according to the map demigration equations
curvelets than for the steeper dipping ones, as it does (Douma & de Hoop, 2005).
not include any such bending.
It remains to be seen how severe the error is if we ig-
nore this bending, and use our TRD transform on a syn- Conclusion
thetic data-set with many curvelets. In this case, there
would be interaction between curvelets from several dif- We have presented first examples of the use of curvelets
ferent scales, i.e., with different frequency content. We in CO pre-stack time migration. A simple numerical ex-
have not performed such a test yet, since this requires ample confirmed that curvelets can be used as building
the proper frequency weighting for curvelets of different blocks for seismic data that allow for a sparse repre-
scales. As we mentioned earlier, this frequency weight- sentation of such data. Since curvelets are roughly like
ing should follow from a stationary phase evaluation of pieces of bandlimited plane waves, the wave-character
equation (12) with a curvelet as scattered data uS . At of the seismic data, i.e., that it is bandlimited and that
the moment of writing this report we have not done such the recorded wavefronts have directions associated with
a calculation. In case future tests using a combination of them, can be built into the representation of the data.
the amplitude correction, obtained in this manner, and Therefore, in essence, curvelets are an appropriate repa-
the TRD transform would show that an omitted bend- rameterization of seismic data, that allow the wave-
ing correction in the TRD transform causes large differ- character of the data to be built into the representa-
ences with the Kirchhoff result, we could introduce an tion of the data. Knowing that the local directions (or
extra bending correction to bend the curvelets towards slopes) in the data (for a fixed scattering angle and az-
the isochron as depicted in Figure 11. imuth, or offset and azimuth in homogeneous media)
can be mapped one-to-one from the data to the image
In this paper, the TRD transform is calculated us- has a small range of k-vectors associated with it, rather than
only a single k direction, as has a plane wave.
ing a brute-force approach in the spatial domain. For
178 H. Douma & M.V. de Hoop
[b]
z kz
x kx
a) y b) ky
Figure 12. A 3D curvelet in space (a) and its associated amplitude spectrum (b). In the spatial domain, 3D equivalents of
curvelets look like circular disks that are smooth along the disk and oscillatory orthogonal to the disk.
version for unstacked data. Stanford Exploration Project where c denotes a curvelet with multi-index (and M
report SEP-37, 1-7.
some index-set), and f X . It is known [e.g., Hernandez
Hernandez, E., & Weiss, G. 1996. A first course on wavelets.
CRC Press, Inc.
and Weiss (1996, pp.334-336) and appendix B in this
Herrmann, F. 2003a. Optimal imaging with curvelets. In: paper] that if X is a Hilbert space, this is equivalent to
Expanded Abstracts. Soc. Expl. Geophys. requiring
Herrmann, Felix J. 2003b. Multifractional splines: applica- X
tion to seismic imaging. SPIE. Pages 240258.
kf k2X = |(f, c )|2 , f X . (A2)
M
Hua, B., & McMechan, G. A. 2001. Parsimonious 2D post-
stack Kirchhoff depth migration. Geophysics, 66, 1497 to hold. The latter expression implies that the energy
1503. of the signal f is conserved through the decomposition
Hua, B., & McMechan, G. A. 2003. Parsimonious 2D (A1); in other words, the curvelets c should be a parti-
prestack Kirchhoff depth migration. Geophysics, 68,
tion of unit energy. Condition (A2) is the definition of a
10431051.
Kleyn, A.H. 1977. On the migration of reflection time con-
(normalized) tight frame (see appendix B for a thorough
tour maps. Geophysical Prospecting, 25, 125140. explanation of tight frames). Hence, for the curvelets to
Mallat, S. G. 1998. A wavelet tour of signal processing. First satisfy the reconstruction formula (A1), we want them
edn. Academic Press, USA. to constitute a partition of unit energy.
Ottolini, R. 1983. Velocity independent seismic imaging. Curvelets can be thought of as 2D (anisotropic) ex-
Stanford Exploration Project report SEP-37, 1-7. tensions to wavelets. Just like wavelets are localized in
Riabinkin, L.A. 1991. Fundamentals of resolving power of one variable and its Fourier dual, curvelets are local-
controlled directional reception (CDR) of seismic waves. ized in two variables and their two Fourier duals. Such
Society of Exploration Geophysicists (Tulsa), Geophysics localization is of course understood within the limits im-
Reprint Series. Pages 3660.
posed by the Heisenberg uncertainty principle. Choosing
Smith, H. 1997. A Hardy space for Fourier integral operators.
J. Geom. Anal., 7.
the variables to be x and z, such localization is obtained
Smith, H. 1998. A parametrix construction for wave equa- through tiling of the spatial domain (x, z) and the spec-
tions with C 1,1 coefficients. Ann. Inst. Fourier, 48, 797 tral domain (kx , kz ). We treat the spectral localization
835. first, followed by the spatial one.
Stein, E. M. 1993. Harmonic Analysis: Real-variable meth- Let j,l (k) be a window (or tile) in the spectral do-
ods, orthogonality, and oscillatory integrals. Second edn. main, with j an index related to the radial (i.e., scale)
Princeton University Press. direction, and l an index related to the angular direc-
Strang, G., & Nguyen, T. 1997. Wavelets and filter banks. tion; i.e. the localization in the spectrum is dealt with in
Revised edn. Wellesley-Cambridge Press. polar coordinates r and . In order for the curvelets to
Sword, C. H. 1987. Tomographic determination of inter-
constitute a tight frame (see appendix B), the windows
val velocities from reflection seismic data: the method of
controlled directional reception. Ph.D. thesis, Stanford
j,l (k) must satisfy
Exploration Project, Stanford University.
XX
|j,l (k)|2 = 1, j J, l L, (A3)
Zavalishin, B.R. 1981. Perfection of methods for construct-
j l
ing seismic images using controlled directional reception.
Soviet Geology and Geophysics, 22, 98104. such that j,l (k) indeed is a partition of unit energy.
The window j,l (k) is constructed through multiplica-
tion of an angular window () and a radial window
180 H. Douma & M.V. de Hoop
w0 j = 0 j = 1 j=2 j=3
2 () 2 ( )
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 8 16 32
0 2 2 3 4 3 8 3
a) 2 4
3 3 b)
Figure A1. Example of a possible angular window () (a), and Lemarie-Meyer windows |w(2 j |k|)|2 for the radial windows
(b). The Lemarie-Meyer windows for j and j + 1 are equal at 2j+1 .
w(|k|), where k is the wave-vector. We treat the angu- Therefore, energy is indeed conserved in the angular di-
lar window first, followed by the radial window. rection using the window functions j,l (). Another way
Let () be an even, real valued, non-negative, an- of saying this, is that the windows j,l () satisfy the ad-
gular window function that is infinitely continuously missibility condition, i.e., that the energy is conserved.
differentiable (i.e. C ) and compactly supported on There are many windows (with different decay prop-
[, ). We choose our function to be 2-periodic, such erties) that satisfy this admissibility condition. Hence,
that () and ( + ) are exactly out of phase; Figure curvelets can be constructed using different angular win-
A1a shows a possible choice of such a window. Then the dows, much like in the construction of wavelets.
angular window satisfies The angular windowing provides the curvelets with
geometry. However, to make the curvelets multiscale, we
|()|2 + |( + )|2 = 1, [0, 2).
need to subdivide the frequencies in subbands. This is
Defining the windows j,l () = (2j l), with j 0 done using a radial window w. A possible choice for such
and l = 0, 1, 2, . . . , 2j 1, it follows that these windows a window is the Lemarie-Meyer window [see (Hernandez
cover the interval [/2j , ). The windows j,l ( + ) & Weiss, 1996, p.27-28) and Figure A1b]. Let w be a
then cover [ /2j , 0). Taking into account the 2 C real-valued, non-negative, function with support in-
periodicity of (), it follows that j,l () and j,l ( + ) cluded in [2/3, 8/3]. Then Meyer introduces a parti-
together cover [0, 2). It is important to note that each tion of unity (i.e. a partition of unit energy, such that
window j,l () has width 2/2j , and that the width of the sum of squares equals one), as
the window thus depends on the scale j. X 2
Due to the 2 periodicity of (), we have |w0 (r)|2 + j
w(2 r) = 1, r 0, (A6)
j0
j,l ( + ) = (2j ( + ) l)
with w0 a coarse scale C window that equals one on
= 2j l + 2 j [0, 2/3] and vanishes beyond 4/3.
= j,l+2j (). (A4) Let k denote the (angular) frequency vector, i.e.
k = (kx , kz ). Then, combining the angular and radial
Using this, it follows that for each j 0 windows into one window j,l (k) as
2j+1
X1
j
2X 1 2j+1
X1
j,l (k) = w 2j |k| bj/2c,l () + bj/2c,l ( + ) ,(A7)
|j,l ()|2 = |j,l ()|2 + |j,l ()|2
l=0 l=0 l=2j
j j and setting |0 (k)|2 = |w0 (|k|)|2 + |w(|k|)|2 , it follows
2X 1 2X 1
|j,l ()|2 + j,l+2j ()2
=
l=0 l=0
j
2X 1
|j,l ()|2 + |j,l ( + )|2
`
=
l=0
= 1. (A5)
Common-offset pre-stack time migration with curvelets 181
that where the approximation is used to indicate the omitted
2 P2bj/2c 1 2 constants (and where we ignore the differences in dimen-
|0 (k)| + l=0,j1 |j,l (k)| =
sions between width and length squared). This scaling
2 2 P2bj/2c 1 ` j 2
|w0 (|k|)| + |w(|k|)| + l=0,j1 w 2 |k| is referred to as the parabolic scaling. Note that the scal-
ing relation (A8) holds in the spatial domain, while in
bj/2c,l () + bj/2c,l ( + )2
the frequency domain we would have length width2 .
2
= |w0 (|k|)|2 + |w(|k|)|2 + j1 w 2j |k| From the previous it follows that the volume of the
P `
P2bj/2c 1 angular wedge in the spectral domain is O 2j 2j/2 ,
l=0 bj/2c,l () + bj/2c,l ( + )
2 while the volume of the rectangle in the spatial do-
= |w0 (|k|)|2 + j0 w 2j |k| = 1
P `
, main is O 2 j/2
2 j
. Therefore, curvelets are really
where we used equations (A5) and (A6). Here, the no- Heisenberg tiles in phase-space, since their net volume
tation bjc denotes the integer part of j. From equation in phase-space is
(A8), it is clear that j,l (k) [together with 0 (k)] form
a partition of unity of the whole frequency plane. Note O 2j 2j/2 O 2j/2 2j = O(1) ,
that by construction, the j,l (k) have compact support,
which is in accordance with the Heisenberg uncertainty
i.e., curvelets are compactly supported in the spectral
principle.
domain. Note that there are many different radial win-
So far we have shown the tiling of the frequency
dows that would form such a partition of unity, just like
plane with curvelets. The windows j,l (k) allow us to
for the angular windows; i.e., many different variants of
rotate and dilate curvelets. To be able to translate
curvelets can be constructed by using different radial
curvelets in the spatial domain, the windows j,l (k) are
window functions.
multiplied with a local (i.e., over the spectral support
From the definition of j,l (k), it is clear that j,l (k)
of a curvelet) L2 orthonormal basis. Since all windows
is obtained through a rotation of j,0 (k). With this ob-
j,l (k) are simply rotated versions of j,0 (k), we first
servation, it becomes clear that the tiling j,l (k) of the
consider the window j,0 (k). The support of this win-
frequency plane (kx , kz ) can be depicted as seen in Fig-
dow is contained in the rectangle R = I1j I2j , with
ure 1a. Note that the definition of j,l (k) using bj/2c,l
in equation (A7), means that the number of angular I1j = {kx , tj kx tj + Lj }, I2j = {kz , |kz | lj /2} ,
windows only increases every other scale (see Figure 1a).
with Lj , and lj the smallest possible bounds, and tj
This splitting at every other scale is a consequence of
the largest possible bound, such that j,0 (k) = 0
the choice of dyadic subbands; with the subband filter-
outside this rectangle R. Here tj + Lj and lj deter-
ing being dyadic, the only way to have an integer num-
mine, respectively, the sampling intervals (or transla-
ber of angular windows per subband, is to increase the
tion steps) in x and z direction
in the spatial domain.
number of angular windows at every other scale only.
It is known that eiqx/L / 2L, with q Z, is an or-
Since the radial (or subband) windows w(2j |k|)
thonormal basis for L2 functions with period 2L; this
for j and j + 1 are equal at 2j+1 (see Figure A1b),
just follows from a Fourier series expansion for func-
the frequencies of j,l (k) are supported near the dyadic
tions with period 2L. This means that ei2m2 kz /lj / lj ,
p
coronae {2j |k| 2j+1 }, depicted in Figure 1a.
with m2 Z, is anporthonormal basis for L2 (I2j ). Also,
These dyadic coronae are directly related to the much
ei(m1 +1/2)kx /Lj / 2Lj with m1 Z, is an orthonormal
familiar dyadic subband filtering in wavelet theory. This
basis for L2 (I1j ) [see e.g. Hernandez and Weiss (1996,
decomposition of the frequency plane is the same as the
Chapter 1.4)]. Therefore, the sequence (uj,m )mZ2 [with
second dyadic decomposition, treated by (Stein, 1993)
m = (m1 , m2 )] defined as
in relation to oscillatory integrals of the second kind, or
Fourier Integral Operators. ei(m1 +1/2)kx /Lj ei2m2 kz /lj
Ignoring the splitting at every other scale, we see uj,m (kx , kz ) = p
2Lj lj
that the length of an angular wedge (i.e., in the radial
direction) is proportional to 2j+1 2j = 2j , while the is an orthonormal basis for L2 (I1j I2j ). Defining 1
width of the wedge is proportional to 2j 2j/2 = 2j/2 . and 2 by Lj = 1 2j and lj = 2 22bj/2c , we have
This
means that the volume of the angular wedge is uj,m (kx , kz ) =
O 2j 2j/2 . Therefore, in the spatial domain, the
2(bj/2c+j)/2 i(m1 +1/2)2j kx /1 im2 kz 2bj/2c /2
curvelet is supported near a rectangle of width pro- e e .
2 1 2
portional to 2j and length 2j/2 ; i.e., the width of
this rectangle is the inverse of the length of the angular For normalization purposes, the term 2(bj/2c+j)/2 is
wedge, while the length of the rectangle is the inverse rewritten as 23j/4 , i.e., bj/2c is replaced with j/2, giv-
of the width of the wedge. Therefore, curvelets obey the
anisotropic scaling relation
width length2 , (A8)
182 H. Douma & M.V. de Hoop
ing APPENDIX B: TIGHT FRAMES IN
HILBERT SPACES
uj,m (kx , kz ) =
23j/4 i(m1 +1/2)2j kx /1 im2 kz 2bj/2c /2 Let H be a Hilbert space, and let (, ) denote the inner
e e , product on H. A collection of elements {ej }jJ in H,
2 1 2
with J an index set, is called a frame, if there exist
Using this definition, and denoting by Rj,l the rota- constants A and B, 0 < A B < , such that
tion by angle j,l , the curvelet is then in the frequency X
domain defined as A kf k2H |(f, ej )|2 B kf k2H f H, (B1)
jJ
c (k) = 2j,l (k)uj,m Rj,l k , = (j, l, m), (A9)
where A and B are called frame bounds, and where
where c denotes the Fourier transform of the kkH denotes the norm on H induced by the inner
curvelet c L2 (R2 ). These are the fine scale product. If the frame bounds are equal (A = B), the
curvelets. The coarse scale curvelets are then de- frame is called tight. E.g., the four vectors = (0, 1),
1
fined as c0 (k) = 20 (k)um (k), with um (k) = 2 2 2 2
2 = (1, 0), 3 = , , and 4 = , ,
(20 )1 ei(m1 kx /0 +m2 kz /0 ) . Here, 0 is again some ap- 2 2 2 2
propriate constant determining the sampling rate (or form a tight frame forP R2 with frame bounds A = B = 2,
translations steps) in the spatial domain. Note that since it follows that 4i=1 |(f, i )|2 = 2 kf k2 . Here the
here 0 determines the sampling rate in both the x and frame bounds measure the redundancy of the frame in
z direction, since the coarse scale curvelets (or better R2 , i.e., four vectors in two dimensions have a redun-
wavelets) are isotropic. dancy of two.
Using the definition of curvelets given in equation It is known that the linear span of frame elements
(A9), it can be shown that curvelets form a (normalized) is dense in H [e.g., Hernandez and Weiss (1996, p.399)].
tight frame. From equation (A9) it is clear that in L2 This means, that any element f H can essentially be
we have written as a linear combination of the frame elements,
X 2 Z 2 and that the difference between this linear combination
f, c = (2) f(k) |j,l (k)| dk,
2 2
and f has a measure zero. If the frame elements were
m1 ,m2
also linearly independent, they would form a basis for
H. It is clear that the frame elements ej are not a basis,
where we used that uj,m Rj,l k is, by con-
mZ2 since adding the zero vector to {ej }jJ does not alter
struction, an orthonormal basis for L2 over the sup- the inequalities in (B1). Of course, adding any vector
2
port of j,l (k), i.e., uj,m Rj,l k = 1. Using that
to a basis, would destroy the linear independence of the
the j,l (k) constitute a partition of unity by equation basis, meaning it no longer would be a basis. When both
(A8), it then follows that frame bounds equal one (A = B = 1), the frame is called
2 a normalized tight frame. Each orthonormal basis is
P P2bj/2c1 P obviously a normalized tight frame, but the converse is
m1 ,m2 f , c(j,l,m1 ,m2 ) =
j0 l=0
not generally true, since the frame elements need not be
P2bj/2c1 R 2 linearly independent.
(2)2 j0 l=0 2
P
f (k) |j,l (k)| dk
2 hP Even though the frame elements do not need to be
R P2bj/2c1 i
= (2)2 f(k)
|j,l (k)|2 dk normal, it follows that they must satisfy kej k B,
j0 l=0
2 2 j J. To see this, let f = ek for some k J, and use
frame condition (B1) to see that kek k4 = |(ek , ek )|2
R
= (2)2 f(k) dk = (2)2 f .
|(ek , ej )|2 B kek k2 . Therefore we have kek k
2
P
jJ
This tells us that (c )M (with = (j, l, m1 , m2 ) a B, k J. If the frame is a normalized tight frame
multi-index) is a tight frame for L2 (R2 ). Here, M is (i.e., A = B = 1), we must have kek k 1. Hence, for
a multi-index set that has the appropriate ranges for a normalized tight frame, the linear independence of a
j, l, m1 and m2 . Finally, using the Parseval formula
basis is traded for the condition kek k 1, k J.
and the Plancherel formula for L2 (R2 ), i.e., f, c = If a normalized tight frame has kek k = 1, k
2 J, it follows that {ej }jJ is an orthonormal basis
(2)2 (f, c ) and f = (2)2 kf k22 , respectively, it fol-
for H. To see this, observe that P for fixed k J,
2
lows that for f L2 (R2 ) 1 = kek k4 = |(ek , ek )|2 jJ |(e k , e j )| 2
1.
P 2 2
X Therefore, we have jJ |(e k , e j )| = |(e ,
k k e )| +
|(f, c )|2 = kf k22 . (A10) P 2 2 P 2
M j6=k,jJ |(ek , ej )| = kek k + j6=k,jJ |(ek , ej )| =
P 2
1, and thus j6=k,jJ |(ek , ej )| = 0. This implies
This shows that (c )M is a normalized tight frame for (ek , ej ) = 0, k J and k 6= j. Since k is arbitrary,
L2 (R2 ), giving us the reconstruction formula (A1). and since kek k = 1, it follows that {ej }jJ is an or-
thonormal basis for H.
Common-offset pre-stack time migration with curvelets 183
It is known [e.g., Hernandez and Weiss (1996,
pp.334-336)] that for a Hilbert space H and a family
of elements {ej }jJ in H, the condition
X
kf k2H = |(f, ej )|2 f H, (B2)
jJ
ABSTRACT
with the distance between source and receiver and the For a real signal, S() equals the complex conjugate
velocity perturbation. The next section derives the S (),
shift in the location of the refocused pulse, followed by Z
1
a simulation test to validate the expression obtained. A S T R (t) = S ()eit dw.
2
Time reversed imaging 187
approximated up to second-order accuracy in (r/R) as
rr h r
|r rr | = R 1 cos( )+
R
r 1
r2
2
sin ( ) , (5)
2 R2
Array of where is the angle defined in Fig. 1. This term is
Source
R Receivers present in both the denominator and the numerator of
Eq. (4). In the denominator, we ignore the dependence
of |r rr | on (r/R) and approximate it to be equal to
R. This approximation results in an error of the order
(r/R), which can be ignored as R r. In the numera-
tor, the term |r rr | is multiplied with the wavenumber
k in the argument of an exponent. Ignoring the (r/R)
Figure 1. Model showing the source and an array of receivers and (r/R)2 terms in the numerator of Eq. (4) would re-
placed at a distance R for a limited aperture angle . sult in errors of the order of (r/) and (r 2 /R) respec-
tively, which may be significant. ( is the wavelength of
This shows that the time-reversed signal in the time the pulse.) We, therefore, express the term |rrr | in the
domain is equivalent to complex conjugation in the fre- exponent using Eq. (5). With this simplification, the de-
quency domain. Hence, the wave-field recorded by the nominator becomes a constant scaling factor except for
array of receivers shown in Fig. 1, when reversed in time, ||. Leaving out this scaling factor, the back-propagated
is equivalent to complex conjugation of Eq. (2): wave-field becomes the following integral over all the re-
ceivers.
exp[i(kR /4)]
u (R) = S (). (3)
8kR S ()
Z
If we back-propagate this wave-field with the same P (r, ) = exp i k0 r cos( )
||
velocity as the forward propagating velocity, we refocus
the waves at the original source location. When we back- 1 k0 r2 2 0
sin ( ) R(k k) d. (6)
propagate with a different velocity, the energy will refo- 2 R
cus at a location other than the original source location. By replacing the wavenumbers k = /c and k 0 = /c0 ,
Let us see what information this shift in the location we can also rewrite the back-propagated wave-field as
of the refocused pulse gives us about the velocity per- Z
turbation. The back-propagated wave-field radiated by S () r
h n
P (r, ) = exp i cos( )
each of the receivers is obtained by convolving this time- || c0
reversed signal and the Greens function. To compute
1 r2 1 1
the total back-propagated wave-field, we need to sum sin2 ( ) R d,
over all receivers that have recorded the forward propa- 2 Rc0 c0 c
gating wave. For the receivers that are densely sampled where c is the velocity for the forward propagation and
over the receiver array, we approximate the summation c0 is the velocity for back-propagation. To represent the
over all the receivers by integration over the aperture an- refocused energy in the time domain we integrate over
gle in order to simplify further calculations. Hence, in all frequencies (Snieder, 2004). The wave-field can be
the frequency domain, the back-propagated wave-field represented in the time domain in terms of a function f
at location r corresponds to as
Z Z r 1 1
h
P (r, ) = G(r, rr ())u (R)d P (r, t) = f t+ cos( ) R
c0 c0 c
" #
1 r2
Z
S () exp [i(k0 |r rr | kR)] sin2 ( ) d, (7)
= d, 2 Rc0
8
p
k0 |r rr | kR
(4) where f is defined as
Z
where goes from - to , rr is the receiver location and S () it
f (t) = e d.
k and k0 are the wavenumbers associated while forward
||
and back-propagation, respectively.
We are interested in the refocusing point close to
the original source location. This requires that R r The function f is a function of time and space
and hence the term |r rr | in the expression can be
188 K. Mehta, R. Snieder & K. Larner
that can be expressed in terms of x and z as follows: 25
r 1 r2
f t+ cos( ) sin2 ( ) Receivers
c0 2 Rc0 20
1 1
i
R
c0 c
h z x 1 15
= f t + 0 cos + 0 sin
c c 2c0 R z/ R
c 10
i
(x cos z sin )2 R ,
cc0
where 5
z = r cos ,
x = r sin , 0
c c0 c.
This function f can be approximated using a Taylors
5
series expansion up to second order in , where
15 10 5 0 5 10 15
x/
z x 1
= cos + sin Figure 2. Initial conditions of the simulation test model
c0 c0 2c0 R
c
showing a source pulse at the origin with an array of receivers
(x cos z sin )2 R . (8) forming an aperture at a distance R.
cc0
When we insert this representation into Eq. (7) and in-
tegrate over , we obtain an expression for the back-
propagated wave-field at an arbitrary location (x, z) c
lim z = R . (10)
close to the source location. We are interested in the 0 c
location of the refocused pulse which we define as given Let us consider this case of small , or equiva-
by the maximum value of this wave-field at time t = 0. lently just a single receiver (Fig. 8). Suppose we ex-
Since the medium is homogeneous and the acquisition cite the source pulse and record the signal at a dis-
geometry is symmetric with respect to x = 0, the peak is tance R using a single receiver. If this recorded signal
located on the X-axis; hence we set x = 0 in the wave- is back-propagated with same velocity, then the back-
field representation. This reduces the resultant wave- propagated wave-field refocuses at the original source
field representation to an expression involving terms in location. If, instead, the back-propagation velocity is
z and t only. We are interested in examining the shift different from the forward-propagation velocity (c0 6= c
in the location of the refocused pulse in the z-direction. , c 6= 0), the back-propagated waves travel over a dis-
This shift can be computed in two steps. The first step tance,
involves evaluating the wave-field at time t = 0. This
wave-field is then solved for its maximum as a function
of z. As a result, we get the shift in the location of the Rback = c0 t
refocused pulse, = (c + c)t
sin
R c
2 = ct + tc
z
=
c
. (9) c
sin 2
1 + 2 = R R . (11)
c
A stepwise derivation of this result is shown in Appendix Hence, the relative shift in the location of the refocused
A. pulse is R(c/c), which agrees with Eq. (10).
This expression for the shift in the location of the
refocused pulse holds for any source pulse that is sym-
metric at t = 0. When the forward and backward prop-
3 ILLUSTRATION USING NUMERICAL
agating velocities are the same (c = 0), the location
SIMULATION
of the refocused pulse is the same as that of the source
pulse (z = 0), as supported in Eq. (9). For a small Eq. (9) gives the expression for the location of the refo-
aperture angle, (sin )/ and (sin 2)/2 1; hence cused pulse. Let us compare the results obtained from
in the limit of 0, the shift of the refocused pulse this expression with a numerical simulation. The simu-
is a function of just the distance R and the velocity lation uses a simple finite-difference scheme for the 2-
perturbation c. D wave equation with absorbing boundary conditions
(Clayton et al., 1997). The geometry of the simulation
Time reversed imaging 189
25 5
20 Receivers
15
z/ R z/
10 0
5
5 5
15 10 5 0 5 10 15 5 0 5
x/ x/
Figure 3. Energy refocusing at the source location (along Figure 4. Detailed view of the refocused pulse obtained by
with some low amplitude noise caused by imperfect absorbing back-propagating the wave-field using the correct velocity.
boundary conditions) obtained after the wave-field recorded This shows the distortion in the shape of the refocused pulse.
by the receivers is back-propagated using the correct back-
propagating velocity.
5
z/R
0.05
5 0.05 0 0.05
5 0 5 c/c
x/ Figure 7. Relative shift in the location of the refocused pulse
Figure 6. Refocused Pulse obtained when the recorded as a function of relative perturbation in the back-propagating
wave-field is back-propagated using a higher velocity than velocity. The relative shift obtained from numerical simula-
that for forward-propagation. tion are denoted by diamonds for 90 aperture angle and by
circles for 5
1.25 A
Array of
G() Receivers
C
1.2 <>= <>= <>= <>= <>= <>= <@? <@? <@? <@? <@? <@? <@? <@? < 34 BCB
112 9: =A /0 >= ;Source B
A R
1.15
C
B
65 65 C
1.1 A B
C A
R
1.05
S
1.0
0 20 40 60 80 R(c/c)
/c)
c
Figure 10. Relative shift in the location of the refocused R(
pulse normalized by relative velocity perturbation as a func- A
C B
tion of angle
Figure 12. Shift in the location of the refocused pulse when
the wave-field is recorded by an array of receivers and back-
propagated using higher velocity (based on wave fronts)
R( c/c)
5 RELATION TO MIGRATION
Time-reversed imaging is mathematically the same as
Figure 11. Shift in the location of the refocused pulse when migration in the field of geophysics (Borcea et al., 2003).
the wave-field is recorded by a single receiver and back- The numerical simulation shows that if the recorded
propagated using higher velocity (based on wave fronts) wave-field is back-propagated (migrated) using the cor-
rect velocity, then the waves are imaged at the correct
nation can be given for the wave-front C-C; hence we position. This is demonstrated in Fig. 3 which shows
have a constructive interference at a distance R00 from the refocused pulse for a homogeneous medium with a
the source, yielding the refocusing. This distance R00 considerably large aperture.
can be referred to as apparent displacement. Hence, for The refocusing for a homogeneous medium is gov-
a larger aperture, the refocused pulse is shifted by a erned by the aperture of the receiver array. Changing
distance greater than R(c/c). This explanation is con- the shape of the aperture from circular (as in the simu-
sistent with Eqs. (9) and (10) and supports the results lations) to linear (as in migration) does not fundamen-
shown in Figs. 7 and 10. This, however, is just an ex- tally alter the refocusing. Reducing the aperture size,
treme case where we consider only two receivers placed however, leads to smearing of the refocused pulse, even
at the two ends. When we consider the whole receiver though it focuses at the correct position. This becomes
array, the wave-fronts coming from all the receivers will clear by comparing Fig. 13, showing refocusing with a
be displaced by a distance of R(c/c) in their direction smaller aperture, with Fig. 3. When a strong hetero-
of propagation resulting in constructive interference at geneity, characterized by velocity variations, is present
an effective distance [R(c/c)/ cos()] with the angle in the model, then even with the small aperture size, the
varying with the aperture of the receiver array. This dis- quality of refocusing improves as compared to that for a
homogeneous medium. Fig. 14 illustrates the improved
Time reversed imaging 193
by the receivers can be written in the frequency domain
25 as u = SG where S is the source spectrum and G is the
Greens function of the medium. The time window for
20 recording the wave-field is long enough to capture both
the direct arrival and the coda waves. Decomposing the
Greens function into the direct wave and the coda, this
15 can also be written as
z/ u = S(Go + Gcoda ).
10
With 20% velocity variations in our random medium
and a correlation length of the order of a tenth of the
5 wavelength of the source pulse, the amplitude of the
direct arrival is around twice that of the coda waves.
0 When the recorded wave-field is time-reversed and back-
propagated through the same medium, the resulting
5 wave-field can be written as,
15 10 5 0 5 10 15
u G = S (Go + Gcoda )(Go + Gcoda )
x/
= S (Go Go + Gcoda Gcoda + Gcoda Go
Figure 13. Refocused pulse gets smeared because of reduc-
tion in the aperture size formed by the array of receivers.
+Go Gcoda ), (14)
where u is the time-reversed wave-field (asterisk de-
notes the complex conjugate) and G is the Greens func-
25 tion of the same medium. The third term is the cross-
correlation of coda waves during forward-propagation
and the direct arrival during back-propagation. Simi-
20 larly, the fourth term is the cross-correlation of direct
arrival during forward-propagation and the coda waves
15 during back-propagation. Since the phase information
encoded in Go and Gcoda is unrelated, these terms do
z/ not contribute to the refocusing and we will refer to it as
10 focusing-noise. Hence, if the entire recorded wave-field
is back-propagated, the refocused wave-field is described
5 by the four terms in Eq. (14) out of which only the first
two terms represents refocusing.
Fig. 15 shows the refocused pulse when only the
0
coda waves are back-propagated. Even in the absence
of the direct arrival, the waves refocus in the pres-
5
ence of the random medium. A comparison of Figs. 14
15 10 5 0 5 10 15
and 15 show that refocusing is actually better when
x/ we back-propagate only the coda waves as opposed
Figure 14. Improved refocusing is obtained due to the pres- to back-propagating the entire wave-field. When only
ence of a random heterogeneity both while forward- and the coda waves are back-propagated, Eq. (14) reduces
back-propagation. This improvement is because of the scat- to u G = S (Gcoda Gcoda + Gcoda Go ). Hence, the re-
tering in the random medium. focused coda waves do not contain noise represented
by the fourth term in Eq. (14) when only the coda
focusing for a heterogeneous medium. The heteroge- waves are back-propagated. This explains the improve-
neous block, here, is random and characterized by Gaus- ment in focusing by muting the direct wave. If only
sian correlated velocity variations. The improved focus- the direct arrival is back-propagated, Eq. (14) becomes
ing is due to the scattering inside the random medium u G = S (Go Go +Go Gcoda ). The first term in the equa-
giving rise to the coda waves. For both forward- and tion indicates that waves refocus at the correct location
back-propagation, scattering allows energy, in the form with some noise described by the second term. However,
of coda waves, to arrive at and depart from a larger focusing deteriorates because it is the back-propagation
range of directions which would be absent for propaga- of the coda waves (represented by the second term in
tion in in homogeneous media. Hence, randomness in Eq. (14)) that improves the focusing when we back-
the medium, in effect enlarges the aperture angle. propagate through the same medium.
After forward propagation, the wave-field recorded In migration, the velocity model is characterized
by the best velocity estimates without knowledge of the
194 K. Mehta, R. Snieder & K. Larner
25 25
20 20
15 15
z/ z/
10 10
5 5
0 0
5 5
15 10 5 0 5 10 15 15 10 5 0 5 10 15
x/ x/
Figure 15. In the presence of heterogeneity both while Figure 17. Energy is scattered all around when only the
forward- and back-propagation, back-propagation of only the coda waves are back-propagated through a homogeneous
coda waves also gives better refocusing. medium.
APPENDIX A:
The wave-field in the time domain is given by Eq. (7).
REFERENCES When the function f is represented with a second-order
Taylor series in (r/R) and integrated over all the re-
Arfken, G. B., and Weber, H. J., Mathematical Methods for ceivers, we get the following expression for the back-
Physicists (Academic Press, San Diego, CA, 1995), 4th. propagated wave-field.
ed.
1
Bakulin, A., and Calvert, R., Virtual source: new method P (r, t) = 2f (t) + f(t)I1 + f(t)I2 , (A1)
for imaging and 4D below complex overburden, Ex- 2
panded abstracts of the 2004 SEG-meeting, 24772480
(2004).
196 K. Mehta, R. Snieder & K. Larner
where I1 and I2 are functions of space. As a result, the The term f(0) drops in the expression for z, which
wave-field is the sum of three terms. We need to calcu- indicates that the shift in the location of the refocused
late this wave-field at time t = 0. Also, we know that pulse is independent of frequency. For small values of
because of the symmetry of the receiver positions with velocity perturbation (c/c), we can ignore the second
respect to x = 0, the peak location will occur at x = 0. term in the denominator; and hence,
Hence, we can set x = 0 in the expression for the wave- sin
R c
2
field. In the second and third terms of the wave-field, z
=
c
. (A3)
sin 2
the integrals I1 and I2 are given by 1 + 2
z2
Z
z c
2
I1 = cos sin + R d
c0 2c0 R cc0
h
2z c z2 sin 2
i
= sin + 2R + ,
c0 cc0 2Rc0 2
2
z2
Z
z c
2
I2 = cos sin + R d
c0 2c0 R cc0
2
n
z sin 2 c sin 2
h io
= +
c02 2 c 2
2
c z c
+4R sin 2R2 .
cc0 c0 cc0
In the integral I2 , we are considering only the terms up
to second order in z. When we insert these values of I1
and I2 in Eq. (A1) and set t = 0, we get the wave-field
as a quadratic function of z as follows:
P (z, t) = A + Bz + Dz 2 , (A2)
where
2
c c
A = 2f (0) + 2Rf(0) R2 f(0) ,
cc0 cc0
2 sin c
h i
B = f(0) + f(0)R ,
c0 cc0
sin 2
n h io h i
D = f(0) 1 + f(0) K1 ,
2Rc 0 2 2c 02
and
sin 2 c sin 2
h i h i
K1 = 1+ 1 .
2 c 2
As the source pulse is symmetric with respect to
the origin, its first derivative vanishes at time t = 0;
hence f(0) = 0. The maximum value of the wave-field
P (z, t) at time t = 0 can be found by equating its first
z-derivative to zero. When solved for z, this gives the
shift in the location of the refocused pulse. Hence, this
shift is z = B/2D where the values of B and D at
time t = 0 are given by,
2 sin c
h i
B = f(0) R ,
c0 cc0
h i
D = f(0) K1 .
2c02
Given the values of B and D, we can compute z as
follows:
sin
c
2
R c
z = sin 2
c
sin 2
.
1+ 2
c
1 2
CWP-512
ABSTRACT
Previous studies of the sensitivity of depth migration to smoothing of the
migration-velocity model have treated smoothing of an initially correct model.
Aside from the relatively small amount of smoothing that is needed for imaging
with Kirchhoff migration and that does no harm to imaging with finite-difference
migration, smoothing of the model changes the model from the true one, so
those studies have shown the less smoothing the better. Because we never know
the subsurface velocity function with perfect accuracy, imaging is always com-
promised to some extent by error in the migration-velocity model. Given that
reality, perhaps some amount of smoothing of the inevitably erroneous velocity
model could improve quality of the migrated image.
We have performed a number of tests of imaging with erroneous velocity mod-
els for a simple synthetic 2D model of reflectors beneath salt. The salt layer
has a chirp-shape boundary so that we could assess imaging quality as a func-
tion of lateral wavelength of velocity variation in the overburden. Errors that
we introduce into the velocity model include lateral and vertical shifting of
the chirp-shape (usually top-of-salt) boundary, and erroneous amplitude of the
chirp, including random errors in the chirp shape. Primarily with poststack
migration of modeled exploding-reflector, we assess sub-salt image quality for
migrations with many different smoothings of erroneous velocity models. We
find that, depending on the type and size of error in the shape of the top-of-salt
boundary, as well as the lateral wavelength of the chirp, smoothing of the er-
roneous velocity model before migration can benefit image quality, sometimes
substantially. The form of error that can most benefit from smoothing is error
in the shape, as opposed to position, of the salt boundary. This observation,
based on numerous tests with exploding-reflector data, is supported by a small
number of tests with smoothing of the erroneous velocity model in prestack
migration.
Key words: velocity smoothing, depth migration
M1 M2
M3 M4
M5 M6
Figure 1. Velocity models M1-M6 used to generate the synthetic data. Lateral position is denoted by x, and depth by z.
2500
M4 4500 1000 200 X
2000
M5 2500 500 50 X
1500
M6 2500 500 100 X
1000
500
Table 1. Chirp parameters for the generic velocity model
0 2 4 6 8 10 12 14 16 18 20
used in the study. The parameter h is the ampliitude of the
x (km)
chirp, i.e., the peak departure of salt-boundary depth from
Figure 2. Range of spatial wavelengths for the two chirps its average value.
used in our study. The solid line shows wavelength as a
function of lateral position for models M1, M2, and M5; the
dashed line shows wavelength for models M3, M4, and M6. 3 ZERO-OFFSET VERSUS
EXPLODING-REFLECTOR DATA
nation of pertinent values for these many parameters. Ultimately one might prefer that a study of the sensitiv-
Moreover, none of the six models in Figure 1 exhibits ity of sub-salt imaging to errors in the velocity model
the large structural size of the salt in, for example, the be done with 3D prestack depth migration applied to
Sigsbee2 model. So, the study is a mere start. modeled 3D data. The cost of such a study is clearly
When we consider the different forms of error in prohibitive, certainly in this decade. Among the many
the velocity models and differing amounts of smoothing other complexities of such a study would be the issue of
of those erroneous models, the number of tests to per- how to define a useful 3D extension of the chirp model.
form could further multiply greatly. Errors could arise Our study therefore is strictly limited to 2D. Even
in all of the parameters (except for the sub-salt reflector 2D prestack depth migration imposes too large a com-
description) listed above. For example, the chirp-shape putational cost for other than a small number of to-
top of salt used for the migration-velocity model could ken comparison tests with the generic model. In order
be shifted laterally or vertically from the true position, to do enough comparisons, we limited the study pri-
or have erroneous amplitude. marily to imaging with poststack migration. The sim-
Our tests with the generic model all involve the plification doesnt stop here, however. We can envision
six models in Figure 1. All six have velocities of 2000 three different forms of input to poststack migration:
m/s, 4500 m/s, and 3500 m/s for the top layer, salt (1) modeled data from many source-to-receiver offsets
layer, and half-space, respectively. For all six models, the that have been stacked, (2) zero-offset (ZO) data ex-
average depth of the top and bottom of the salt is 1000 tracted from normal-moveout-corrected and unstacked
m and 1900 m, respectively. Models M1, M2, and M5 modeled data, and (3) exploding-reflector (ER) modeled
have the same range of spatial wavelength for the chirp, data. We did tests with all three forms, but the largest
and Models M3, M4, and M6 have a higher range of number with exploding-reflector data.
spatial wavelength. The variation of spatial wavelength The choice of exploding-reflector data may seem
with horizontal location is shown in Figure 2. Other puzzling. One reason for this choice is that generation of
differences among the six models are in the amplitude ER data is least computationally costly. We use a finite-
of the chirp shape. Table 1 summarizes the parameter difference code, second-order in time, fourth-order in
values for the chirp in each model. space, for the modeling; the cost of modeling zero-offset
Another important parameter for the tests would data would be essentially the same as that of modeling
be the range of frequencies contained in the seismic a full prestack data set. But there is another reason for
wavelet used in the wavefield modeling. All of our tests choosing exploding-reflector data over extracted zero-
involve just one choice of input wavelet a Ricker offset data.
wavelet, with dominant frequency of 15 Hz. Seismic data that result from ER modeling are
largely equivalent to, but differ in important respects
from, either ZO or common-midpoint (CMP) stacked
data, particularly in the presence of strong lateral veloc-
ity variation (Kjartasson & Rocca, 1979) and (Spetzler
& Snieder, 2001). Moreover, the pattern of multiples and
Velocity smoothing before depth migration 201
the relative amplitudes of the primaries and multiples Using equation (12) of Spetzler and Snieder (2004),
differ among the three forms of data. we calculated approximate focal depths (depths at
So, why do we consider a study using ER data to which caustics and triplications start to appear) for the
be useful? Because poststack migration is based on the six velocity models used in the study. Although only
exploding-reflector assumption, such migration of zero- roughly approximate because that equation assumes
offset data would be erroneous even if the migration ve- 1D lateral slowness variations and point sources, these
locity were correct. In contrast, because poststack (i.e., computed focal depths give a measure of the relative
zero-offset) migration and exploding-reflector modeling complexity of the various models and of wavefields in
of primaries are essentially exact inverses of one another, them. For our generic models, this complexity, which
we can count on accurate migration of ER primaries depends on the depth and lateral variation of the veloc-
when we use the correct migration velocities. Therefore, ity anomaly, is controlled mainly by the geometry of the
sensitivity of imaging to errors in velocity, including chirp. Figure 5 shows the focal depths below the surface
smoothing of erroneous velocity, is best isolated when as a function of lateral position x for the six chirp mod-
we apply poststack migration to ER data. els. Model M2, for example, has caustics that appear
That ER and ZO data differ from one another, as at the shallowest depth, whereas caustics arise deeper
do the results of poststack migration applied to these for the models with milder chirp shape, and for chirp
two forms of data, is exhibited in Figures 3 and 4. The shape at the base rather than the top of the salt. The
differences between the ER and ZO sections for models relatively poor image, in Figure 4, of the migrated zero-
M1, M2, and M4, in Figure 3, are striking. Particularly offset data for Model 2, as compared with the migrated
for models with larger chirp amplitude and in regions images for models M1 and M4, suggests dependence of
of the model with smaller chirp wavelength, sub-salt image quality on model complexity. (Again, the imaging
reflections show more numerous triplications character- problem arises because zero-offset migration is based on
istic of caustics and multi-pathing in the overburden. the exploding-reflector idea.) Also, as seen in Figure 5,
The exploding-reflector sections exhibit less loss of am- the levels of complexity of models M1 and M4 are equiv-
plitude with time and more complete expression of the alent. Consistent with this observation, the image qual-
diffractions than do the zero-offset sections. Since mi- ity of the depth-migrated zero-offset sections for these
gration aims to collapse diffractions, the distorted and models is comparable.
incomplete diffractions in the ZO data will be poorly When errors are present in the velocity model, t
collapsed in poststack-migrated results. The stronger he degradation of migrated ER data will differ from
amplitudes at late time in the ER data result from the that in (1) poststack-migrated ZO data, (2) prestack-
weaker geometric spreading from sources that are, in ef- migrated ZO data, and (3) prestack-migrated full-offset
fect, placed on the exploding reflectors than the spread- data. Even with the correct velocity model the quality of
ing from the surface line sources for the ZO data. Fi- imaging can be compromised in each of these treatments
nally, as expected, the timing and amplitudes of multi- of data. Weve already seen in Figure 4 that, because
ples in the exploding-reflector sections differ from those poststack migration is founded on the ER assumption,
in the counterpart zero-offset sections. The multiples poststack-migrated ZO data is erroneous even when the
in these data are internal ones. Surface multiples are correct velocity model is used for migration. Prestack-
largely absent because absorbing boundary conditions migrated ZO data do not suffer from that shortcoming,
(Clayton & Engquist, 1977) were used for all bound- but can exhibit image distortion and artifacts arising
aries. from insufficient pre-migration muting of wide-angle re-
The results of poststack (i.e., zero-offset) depth mi- flections and refractions prior to the migration, limited
gration of the exploding-reflector and zero-offset sec- aperture for the (shot-record) migration, and variations
tions using the true velocity for the migration are shown in wavefield illumination beneath the salt. Prestack-
in Figure 4, again for models M1, M2, and M4. We migrated, full-offset data can also be distorted because
used an f-x domain, finite-difference depth-migration al- of variable illumination, insufficient muting, and limited
gorithm (Claerbout, 1985) for both migrations. For all migration aperture. These data, however, have the ad-
models, depth migration of the exploding-reflector data vantage that the worst of these problems are mitigated
yields high-quality imaging of the primaries, with arti- to some extent by destructive interference of offset-
facts related to migration of the multiples. In contrast, dependent distortions and artifacts after stacking the
depth migration of the zero-offset sections results in de- migrated data for all offsets.
graded imaging of the sub-salt reflectors, especially in
regions of the model where the chirp has smaller wave-
length. Since these migrations were performed using the 4 LENGTH SCALES FOR VELOCITY
correct velocity model, the compromised migrated ZO SMOOTHING
data offer a poor starting point for study of sub-salt
imaging when smoothed, erroneous velocity models are For finite-difference migration, if we knew the velocity
used for the migration. model perfectly we would have no need to smooth the
velocities. Smoothing could only alter the model from
202 K. Larner & C. Pacheco
M1
x (km) x (km)
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
0 0
1 1
2 2
time (s)
time (s)
3 3
4 4
M2
x (km) x (km)
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
0 0
1 1
2 2
time (s)
time (s)
3 3
4 4
M4
x (km) x (km)
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
0 0
1 1
2 2
time (s)
time (s)
3 3
4 4
Figure 3. Exploding-reflector sections (left) and zero-offset sections (right) for models M1, M2, and M4.
the true one, resulting in erroneous migration the Before addressing smoothing of erroneous velocity
more smoothing the poorer the image. Versteeg (1993), models, let us see how much smoothing of our correct
however, showed that there is little harm done with an generic models is acceptable for the depth migration.
amount of spatial smoothing that is small in relation to Given the range of complexities suggested for the models
the wavelengths in the signal and fineness of detail in in Figure 5, we expect that the appropriate amount of
the velocity model. Even for Kirchhoff migration, Gray smoothing will differ from one model to another.
(2000) pointed out that smoothing too little is better We smooth the velocity functions using a two-
than smoothing too much. dimensional Gaussian-shaped operator similar to the
Velocity smoothing before depth migration 203
M1
x (km) x (km)
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
M2
x (km) x (km)
0 2 4 6 8 10 12 14 16 18 20
0 5 10 15 20 0
0
2000
2000
depth (m)
depth (m)
4000 4000
6000 6000
M4
x (km) x (km)
0 2 4 6 8 10 12 14 16 18 20
0 5 10 15 20 0
0
2000
2000
depth (m)
depth (m)
4000 4000
6000 6000
Figure 4. Finite-difference, zero-offset depth migrations of exploding-reflector data (left) and zero-offset data for models M1,
M2, and M4 using the correct velocity model.
one described in Vertsteeg (1993). We first convert the Figure 1, here we focus attention primarily on models
velocities to slownesses and then convolve the slow- M2 and M4, the ones with the largest amplitude chirp
nesses with a two-dimensional Gaussian filter of the shape for the top of salt. The conclusions drawn from
form exp[(x2 + z 2 )/(a/2)2 ]. With this definition, a is tests for these two models have general counterparts
the diameter at which the amplitude of the Gaussian from those for the other models. Figures 6 and 7 show
operator has decreased to e1 of its peak value. depth-migrated exploding-reflector sections for models
Although we did tests with all six models shown in M2 and M4 after applying different amounts of smooth-
204 K. Larner & C. Pacheco
6000
M1 horizontal ones. Conversely, the horizontal reflectors at
M2
M3 the left are better imaged than are the dipping ones
M4
5000
M5 there.
M6
Another aspect of smoothing, seen in Figure 8, are
4000
gaps in the imaged bottom of the salt for model M2
(m)
3000
smoothed velocity model does not have the detail nec-
z
true a = 80 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a = 160 m a = 240 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a = 320 m a = 400 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
Figure 6. Depth-migrated exploding-reflector sections for model M2 using the correct velocity model and after different amounts
of smoothing have been applied to that correct model. The diameter of the two-dimensional Gaussian smoothing operator is
denoted by a.
Some types error in the velocity model have rela- defocusing the image. Pon and Lines (2004) and Paf-
tively small influence on the migrated image. For ex- fenholz et al. (2001) obtained similar results with their
ample, vertical shifts of the salt boundary and constant modeled data sets. Smoothing of these erroneous veloc-
error in the velocity of the sediments surrounding the ity models has much the same influence on the quality of
salt body have relatively little influence on sub-salt im- imaging as does smoothing of the correct velocity mod-
age quality. These two types of error primarily cause els. Increasing the degree of smoothing in this case only
error in reflector depth without severely distorting or further degrades image quality.
206 K. Larner & C. Pacheco
true a = 80 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a = 160 m a = 240 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a = 320 m a = 400 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
In contrast, consistent with the results of Paffenholz As we shall see, where error in the velocity model
et al., we find that lateral shift of the chirp model can causes substantial degradation of the migrated image.
cause significant degradation of sub-salt imaging, as do smoothing the erroneous velocity model can improve
errors in the geometry of the chirp boundary. In general, the quality of the migrated image. Where it causes too
a given amount of lateral shift of the boundary causes severe degradation, once again no smoothing can help.
larger degradation of the sub-salt image than does a
comparable vertical shift.
Velocity smoothing before depth migration 207
x (km)
0 1 2 3 4 5 6
0
500
depth (m)
1000
Figure 8. Close view of the top and bottom of salt in the migrated image of model M2 for velocity smoothing with a=160 m.
Note the gaps in the image of the base of salt, in part the result of mistreatment of variable illumination at the salt base.
5.1 Lateral shift of the chirp image quality is already so poor that it would take a
large amount of smoothing to make it much worse.
A constant lateral shift of the entire salt boundary is an
unrealistic error in practice. Applied to our chirp-shape
top-of-salt boundary, however, a constant shift allows us 5.2 Exaggerated chirp amplitude
to study the influence of such an error systematically as
a function of lateral velocity variation in the overburden. The type of velocity error we consider next is that
Figure 9 shows the depth-migrated sections for caused by incorrect amplitude of the chirp-shape top
model M4 using the correct velocity model and that of salt. Figures 11 and 12 show depth-migrated images
model laterally shifted by different amounts, from 20 to of exploding-reflector data for models M2 and M4, each
100 m. Note that even a shift of only 40 m (trace spacing with the amplitude of the chirp-shape top of salt too
in the exploding-reflector section is 20 m) introduces dis- large by 50 m, and with various amounts of smoothing.
tortions in the depth-migrated image. As expected, im- The images for the erroneous models are significantly
age degradation is larger beneath the small-wavelength degraded from those generated with the correct migra-
region of the chirp and deeper in the section. Increas- tion velocity, more so for model M2 (with short spa-
ing the amount of lateral shift leads to worsened image tial wavelength of chirp) than for model M4. Smooth-
quality beneath the longer-wavelength portions of the ing of the erroneous velocities for model M2 results in
chirp. We next smooth the shifted velocity models and improved image quality for smoothing diameter up to
assess the quality of sub-salt imaging in the resulting about 320 m, but degrades the quality for larger smooth-
migrations. Figure 10 shows migrations with smoothed ing. The degree of improvement varies across the model.
versions of the model that was erroneously shifted later- For model M4, improvement in imaging for some targets
ally by 40 m. Smoothing with a of 160 to 240 m improves is best with a as large as 640 m. Note also that the re-
the image quality for many of the sub-salt targets, but gion of largest improvement generally progresses toward
smoothing by larger amounts results in reduced qual- the left (i.e., toward longer spatial wavelength of chirp)
ity of the migrated section. Smoothing of velocity mod- as the amount of smoothing increases.
els with larger lateral shifts (not shown here) produced For this 50-m error in chirp amplitude, we per-
similar results although the improvement introduced by formed migrations of models M1 through M4 (the four
smoothing becomes harder to recognize for larger shifts. models with chirp-shape top of salt) and smoothing
The data quality is so much compromised for large lat- ranging from a = 80 m to 1240 m in increments of 40
eral shift that smoothing can have little influence. The m. From subjective visual impressions of migrated sec-
208 K. Larner & C. Pacheco
true shift=20 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
shift=40 shift=60 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
shift=80 shift=100 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
Figure 9. Depth-migrated exploding-reflector sections for model M4 using the correct velocity and different amounts of lateral
shift of that model.
tions for the different amounts of smoothing, we made correlation between the focal depths in Figure 5 (a mea-
rough estimates of smoothing diameter that yields the sure of model complexity) and the optimum smooth-
best imaging of the different target reflectors across the ing diameters in Figure 13 for this particular form of
model. Figure 13 shows the estimated optimum smooth- velocity error: generally the shallower the focal depth
ing diameters for the four models, as a function of lateral (i.e., the greater the model complexity), the smaller the
position x. Comparison of the curves (linearly interpo- smoothing diameter that is best.
lated between the subjectively inferred values) in this Versteeg (1993) showed that the lower the complex-
figure with the curves in Figure 5 shows some degree of ity of the velocity field, the larger the amount of smooth-
Velocity smoothing before depth migration 209
true shifted
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a = 80 m a = 160 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a = 240 m a = 320 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
Figure 10. Depth-migrated exploding-reflector sections for model M4 using the correct velocity model, for that model laterally
shifted by 40 m, and for various degrees of smoothing applied to the laterally shifted model.
ing of the velocity model that is acceptable for imaging. lation between model complexity and optimum amount
Here, we find a counterpart result: the lower the com- of smoothing.
plexity of the velocity field, the larger the amount of
smoothing that yields the best imaging when the ini-
tial model is in error. In tests with larger error in chirp 5.3 Random perturbation of the chirp
amplitude (100-m too large), we found a similar corre- The final type of error in the migration-velocity model
that we show is a random perturbation of the chirp-
210 K. Larner & C. Pacheco
true erroneous
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=160 m a=320 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=480 m a=640 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
Figure 11. Depth-migrated exploding-reflector sections for model M2 with true velocity, for an erroneous velocity model caused
by chirp amplitude exaggerated by 50 m (a 50 percent exaggeration), and for smoothed versions of the erroneous velocity model.
shape top of the salt. We distorted the shape by adding a velocity error. For these tests, the standard deviation of
laterally bandlimited, Gaussian-distributed depth error the random depths prior to bandlimiting is = 50 m.
to the chirp. For the test results shown in Figures 14, 15, Figures 14 and 15 show migrated sections using
and 16, the correlation length l = 100 m, where l is the the true velocity model, a model with random error
lag at which the autocorrelation of the random depth added to the depth of the chirp, and variously smoothed
variation decreases by a factor e1 of its peak value. The versions of the erroneous velocity function for models
correlation length is a measure of lateral scale length of M2 and M4. For both models, migration with the erro-
Velocity smoothing before depth migration 211
true erroneous
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=160 m a=320 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=480 m a=640 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
Figure 12. Same as Figure 11 but for model M4. The 50-m increase in chirp amplitude represents a 25 percent exaggeration
of the amplitude for this model.
neous velocity function yields severely distorted imaging results for model M4, shown in Figure 15. As smooth-
of the sub-salt reflections. For model M2 (Figure 14), ing increases from a = 160 m to 640 m, targets that
which has the shorter-wavelength chirp-shape top of are best imaged again are those beneath progressively
salt, smoothing of the erroneous velocity model yields, longer-wavelength portions of the chirp.
at best, marginal improvement, with optimal smoothing Model M3 has chirp shape spanning the same range
varying from a = 160 m to at least 640 m for reflectors of wavelengths as those in model M4, but with smaller-
from right to left. The improvement brought about by amplitude chirp. Compared to results for similarly con-
smoothing the erroneous model is more evident in the sidered models M2 and M4, migration for data from
212 K. Larner & C. Pacheco
1400
M1
M2
M3
1200 M4
(m)
1000
optimum
800
a
600
400
200
2 4 6 8 10 12 14 16 18 20
x (km)
Figure 13. Optimum smoothing diameters for migration-velocity error caused by chirp amplitude that is 50-m too large, for
models M1-M4.
model M3 (Figure 16) exhibits more distortion of sub- exploding-reflector data cannot be acquired in the field
salt events when the randomly erroneous velocity model nor can they be obtained from field data. They are a fic-
is used. Migration with variously smoothed versions of tion. Next we show results for one example of a similar
the erroneous model, however, yields substantial im- study of the influence of smoothing, but with prestack
provement in sub-salt images. The pattern of improved migration performed on synthetic multi-offset data.
quality of imaging for various smoothings of the erro- Using finite-difference code, we modeled shot
neous velocity model is similar to that seen in Figures 14 records for a simulated 2D survey across the top of
and 15, but the improvement here is dramatic. In par- model M4, each shot having 500 channels, with 10-m
ticular, much of the sub-salt region is best imaged with group interval, and 80-m shot spacing. Migration was
smoothing using a = 320 and 640 m, with a = 640 m performed with a shot-record f -x domain algorithm
yielding the best result for lateral positions x < 10 km. (Claerbout, 1985). We then sorted the migrated data
Even broader smoothing might have resulted in a better into common-image gathers and stacked the gathers.
image yet. We found similar behavior for other choices of Figure 17 shows the prestack-migrated image for model
correlation distance and amplitude of the random per- M4 using the correct velocity model. The quality of
turbations. Smoothing of the migration-velocity model imaging for the target reflectors is excellent, superior
can help in imaging. to that obtained in poststack migration of exploding-
reflector data for model M4 (Figure 4), primarily be-
cause imaged multiples are much weaker in the prestack
6 SMOOTHING OF AN ERRONEOUS result.
VELOCITY MODEL: PRESTACK For this test, we generated the erroneous velocity
MIGRATION model by exaggerating the amplitude of the top-of-salt
chirp in model M4 by 100 m, a 50 percent increase from
In all of the above tests, we have used poststack- the true chirp amplitude. Prestack migration with this
migrated exploding-reflector data to assess the influ- erroneous velocity model is shown in the upper left of
ence of smoothing of erroneous velocity models on sub- Figure 18. With this level of velocity error, the shape
salt image quality. Such data and migration require so of the horizontal bottom of the salt is greatly distorted
relatively little computation that we could perform a toward the left of the section, and toward the right the
large number of tests. Despite being useful for our study, bottom of the salt is virtually not imaged. Similar obser-
Velocity smoothing before depth migration 213
true erroneous
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=160 m a=320 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=480 m a=640 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
Figure 14. Depth-migrated exploding-reflector sections for model M2 with the true velocity, for an erroneous velocity model
with random error (l = 100 m, = 50 m), and for variously smoothed versions of the erroneous velocity model.
vations hold for the sub-salt reflectors. The remainder subs-salt section (beneath the shorter-wavelength por-
of the figure shows the results of prestack migration us- tion of the chirp), and a = 720 m yielding the best imag-
ing the erroneous velocity model smoothed with what ing beneath the longer-wavelength portion. The best
we might consider to be large amounts of smoothing: images show distortion of the shapes of the target re-
a = 240, 480, and 720 m. Velocity smoothing clearly im- flectors, but these reflectors nevertheless are far better
proves the quality of the migrated images, with a = 480 imaged than when no smoothing is applied to the erro-
m yielding the best imaging of the right portion of the neous velocity model.
214 K. Larner & C. Pacheco
true erroneous
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=160 m a=320 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=480 m a=640 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
true erroneous
x (km) x (km)
0 2 4 6 8 10 12 14 16 18 20
0 0 5 10 15 20
0
2000
2000
depth (m)
depth (m)
4000 4000
6000 6000
a=160 m a=320 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
a=480 m a=640 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
6000 6000
model give clear indication that, for some types of ve- as well as on the lateral wavelength of the true velocity
locity error, smoothing of velocities for migration veloc- structure in the overburden.
ity model can improve image quality, sometimes signif- The optimal choice for smoothing to address imag-
icantly. The amount of smoothing that is optimum in ing degradation caused by use of an erroneous veloc-
the sense that it gives imaging superior to that when ity model can be considerably larger than either (1)
either less or more smoothing is used depends on the that needed to overcome shortcomings of ray tracing
size and type of error in the migration-velocity model for Kirchhoff migration or (2) the amount of smoothing
that would be acceptable for any migration algorithm
216 K. Larner & C. Pacheco
x (km)
0 5 10 15 20
0
2000
depth (m)
4000
Figure 17. Prestack migrated image for model M4 using the correct velocity model. Compare with Figure 4.
when the initial migration-velocity model was perfectly roneous velocity models for use in prestack migration,
accurate. Because the velocity function for migration is use of smoothed velocities generally helped to improve
never fully accurate in practice, some degree of smooth- the quality of images greatly so for the one test with
ing is always appropriate. Moreover, although it will be prestack migration shown here. This is consistent with
difficult in practice to pin down an optimal spatial ex- what we found (although to a lesser extent) for post-
tent of smoothing, that amount can well be larger than stack migration of exploding-reflector data. The con-
is often used in practice even considerably larger than sistency is comforting given that most of our tests were
the spatial size of the errors in the migration-velocity with exploding-reflector data, which cannot be obtained
model. from field data. Supporting the results from the post-
Use of the generic chirp-shape salt boundary al- stack migrations of exploding-reflector data, the amount
lowed us to do simple tests, e.g., velocity error mod- of smoothing that is best can be considerably larger
eled as lateral shift of the boundary, in a systematic than might have been suspected from the spatial size of
effort to gain an idea of the relationship between op- errors and detail in the velocity model.
timum amount of smoothing and scale of the velocity A general ranking of the influence of the different
variations. For even this simple model, we have seen types of velocity error on image quality is as follows.
dramatic differences between zero-offset data modeled Constant vertical shift of the top of salt or constant er-
based on the exploding-reflector assumption and those ror in velocity of the overburden causes relatively little
modeled with wavefields generated by individuals shots. degradation. Smoothing of the velocity model will not
A significant manifestation of the difference arises from improve imaging for these types of error any more than
variations in the spatial distribution of subsurface illu- it would if the velocity model were perfectly accurate.
mination. These differences in the two forms of modeled Lateral shift of the top of salt causes image distortion
data in turn give rise to marked difference in image qual- that can be not only large, but such that imaging is not
ity when the data are depth-migrated with a poststack amenable to improvement by velocity smoothing. Error
algorithm (which is based on the exploding-reflector) in amplitude of the chirp-shape top of salt, including
using a migration-velocity function that is known per- random perturbation of the salt shape, can also cause
fectly. large distortions in the sub-salt image, but the imaging
Although we did only a few tests of smoothing er- can be substantially improved through use of smoothed
Velocity smoothing before depth migration 217
erroneous a = 240 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
a=480 m a=720 m
x (km) x (km)
0 5 10 15 20 0 5 10 15 20
0 0
2000 2000
depth (m)
depth (m)
4000 4000
Figure 18. Prestack-migrated image for model M4 using the erroneous velocity model with chirp amplitude exaggerated by
100 m, and for the erroneous velocity model smoothed with operator diameter a=240, 480, and 720 m.
ABSTRACT
We report on the first direct observation of a migrating fluid pulse inside a fault
zone that, based on previous evidence, is suspected to be a conduit for fluids
ascending from depth. We find that areas of high fault-plane reflectivity from a
fault at the South Eugene Island Block 330 field, offshore Louisiana, systemat-
ically moved up the fault 1 km between 1985 and 1992. The updip movement
can be explained by the presence of a high pressure fluid pulse ascending a verti-
cally permeable fault zone. These fault burps play a central role in hydrocarbon
migration.
Key words: hydrocarbon migration, fault zones, seismic reflectivity
2.3 A 600 m B
Hsand
1992 Hsand
1985
Afault 0.5 s
Bfault
North
1.2 1 km
North
twoway time (s)
A10ST
North North
low 1985 1992
relative
reflectivity
Figure 1. (A) Map of the two-way reflection time (TWT) to the B-fault plane. Using the approximation 100 m depth 0.1 s
TWT (Losh et al., 1999), this map shows the structure of the B-fault, most notably the down-to-the-SW dip. (B) Overlay of
the dip-filtered seismic reflection image (black & white wiggle plot) on top of the original image (red & blue variable density
plot) along the dashed line in (A). The dip-filtering highlights the fault-plane reflections from the B-fault and a nearby fault
known as the A-fault. The H-sand is shown to indicate throw. (C) Map of the B-fault, as in (A), but with reflectivity from the
fault-plane in 1985 plotted instead of TWT. The area of highest reflectivity is circled in gold. (D) Map of the B-fault reflectivity,
as in (C), but from 1992. The data extend over a slightly larger area than in (C); however, the spatial perspective is identical.
The area of highest reflectivity, circled in gold, is shifted roughly 1 km NE in the updip direction relative to its location in 1985,
as is expected for a fluid pulse ascending the B-fault. This movement is depicted by the arrow in (A). Also shown is the location
of the A10ST well intersection, where exceptionally high fluid pressures were encountered while drilling into the B-fault zone
in 1993 (Losh et al., 1999).
330 contain information about the distribution of fluid available data is due to a multifaceted drilling project
pressures across faults (Haney et al., 2004). conducted by the Global Basins Research Network
(GBRN) in the 1990s (Anderson et al., 1995). During
late 1993, GBRN intentionally drilled into and success-
FLUID MIGRATION AS SEEN IN TIME fully cored some of the growth faults at South Eugene
LAPSE SEISMIC DATA Island. Normal faults that transect the field separate up-
thrown sediments saturated by highly-pressurized fluids
The South Eugene Island field is an ideal location for from normally-pressurized dowthrown sediments; the
this study. Multiple vintages of seismic reflection sur- faults are typically at the same pressure as the upthrown
veys can be interpreted in the context of abundant sediments (Losh et al., 1999). However, exceptionally
fluid pressure, geochemical, and other data. Much of the pressurized fluid was encountered in one penetration of
Observation of a fault burp 221
a growth fault, the B-fault, in the A10ST well (Anderson cross correlation
et al., 1995; Losh et al., 1999). Losh et al. (1999) stated
that the isolated pocket of anomalously high fluid pres-
sure in the A10ST well may represent a spatially limited
pulse of anomalously pressured fluid.
NS lag (m)
To test the hypothesis of a moving fluid pulse, or
fault burp, we isolate the fault-plane reflections from the
B-fault in images derived from seismic surveys taken in
1985 and 1992 and look for indications of movement.
First, we pick the fault plane in the 3D seismic re-
flection images (Figure 1A). We then proceed by dip-
filtering the seismic reflection images in the direction of
the B-fault (Figure 1B). This dip-filtering step serves to
accentuate the fault-plane reflections while simultane-
ously attenuating the reflections from the sedimentary
layers. The final step is to extract the amplitude of the EW lag (m)
fault-plane reflection along the B-fault in a small time-
window around the picked fault-plane. In Figures 1C Figure 2. Spatial cross correlation of the 1985 and 1992 B-
fault amplitude maps shown in Figures 1C and 1D. There is
and 1D, we show reflectivity as a function of position
a maximum in the cross correlation when the spatial lag is 1
on the fault-plane for both datasets. Patches of high re-
km to the NE (the up-dip direction of the A-fault). At the
flectivity, or bright spots, are known to be associated maximum, the correlation coefficient is 0.78, which is close to
with the presence of fluids (Dobrin, 1976). the value 0.80 for a horizontal reflector known as the O-sand
The most striking pattern in the fault reflectivity at zero spatial lag (see Figure 3). The correlation shown here
maps of Figures 1C and 1D is the northeast movement can be seen visually in Figures 1C and 1D.
of the highest reflectivity areas between 1985 and 1992.
This movement, in the up-dip direction, is to be ex-
pected for a fluid pulse ascending the B-fault. From 1992; Revil & Cathles, 2002). We note that models by
the reflectivity maps at the B-fault, we estimate the Finkbeiner et al. (2001) and Rice (1992) are not the
movement of the fluid pulse to be on the order of 1 km only ones of relevance to the behavior we observe along
between 1985 and 1992, for an average velocity of the B-fault at South Eugene Island. Currently, a vug-
140 m/yr. In the next section, we describe how we es- wave model is being developed (Losh et al., 2005) that
timated this velocity for the pulse. Such fast fluid flow describes a fluid pulse which is mechanically coupled
up a growth fault at South Eugene Island is in line with with the shear strain in the fault. This model has the
the study by Finkbeiner et al. (2001). In their study, potential to bridge the link between what we have ob-
Finkbeiner et al. (2001) analyzed in situ stress data served at South Eugene Island and the phenomenon of
and pressure measurements and showed that overpres- slow or silent earthquakes that show sudden aseis-
sure in some of the deep reservoirs abutting the B-fault mic slip (Cervelli et al., 2002).
could induce frictional failure on the fault plane. The
fast movement of fluids up a permeable fault also agrees
with a nonlinear permeability model first discussed by QUANTIFYING THE PULSE MOVEMENT
Rice (Rice, 1992). In this model, the fluid pulse is a In Figure 2, we plot the spatial cross correlation of the
shock wave that moves with velocity v given by amplitude maps from the B-fault. There is a local max-
ko (1 o )g(g f ) sin imum at 1 km spatial lag in the NE direction. This
v =
f o corresponds to what is seen visually in Figures 1C and
exp(f / ) exp(i / )
1D. The value of the correlation at the maximum is
, (1) 0.78, which is close to the correlation value of 0.80 com-
i f
puted for a horizontal reflector known as the O-sand at
where ko is the permeability at zero effective stress, o zero-lag (see Figure 3). The zero-lag correlation for the
is the porosity at zero effective stress, g is gravitational B-fault amplitude maps, as shown in Figure 2, is 0.45.
acceleration, g is the density of the rock grain, f is the The fact that the correlation is as high as 0.80 for the
density of the fluid, is the dip of the fault, f is the dy- O-sand in its zero-lag position supports the similarity
namic viscosity of the fluid, is a parameter describing of the two surveys and their processing schemes. In an-
the linear dependence of porosity on effective stress (El- other study (Haney et al., 2005), we tested how the pres-
liott, 1999), i is the initial effective stress state ahead ence of noise, a lack of DMO processing, and application
of the pulse, f is the final effective stress state behind of AGC degraded migrated amplitudes. The application
the pulse, and is a parameter describing the exponen- of AGC most severely altered the relative amplitudes.
tial dependence of permeability on effective stress (Rice, Though an AGC appears to have been applied to the
222 M. Haney, R. Snieder, J. Sheiman & S. Losh
dip of the fault-plane reflection. The time dip is given
1985 1992 by
2sin
600 m 600 m p= , (3)
v
where is the dip of the fault-plane and v is the true
N N
subsurface velocity, not necessarily equal to the migra-
tion velocity vm . Given the true subsurface velocity,
the amount of lateral movement applied by migration
should be
ptin v 2
x = . (4)
4
The difference in the amount of lateral movement ap-
plied by migration, xm , and the amount of lateral
Figure 3. A comparison of the O-sand amplitude in the movement that should be applied, x, is the error in
1985 and 1992 surveys. The O-sand is a gently dipping sed- the lateral placement of the input point xe due to an
imentary layer boundary at 2.3 s TWT. Hence, the O- incorrect migration velocity
sand abuts the B-fault in the lower part of the B-fault plane 2
ptin (vm v2 )
displayed in Figure 1A. The consistency of the amplitudes xe = xm x = . (5)
in these two images, compared to the changing amplitudes
4
along the B-fault shown in Figures 1C and 1D, supports the Denoting the depth of the time sample in the stacked
interpretation of relative amplitude changes in the two sur- section as z, tin is given by
veys. Though we only show the O-sand here, this consistency 2z
holds for the other sedimentary layer boundaries as well. tin = . (6)
vcos
Therefore, using equation (3) and (6), equation (5) may
1985 data, it obviously did not damage the amplitudes be written as
from the O-sand enough to bring its correlation below
vm 2
0.80. We have also checked the K-sand and found good xe = ztan 1 . (7)
v
amplitude agreement in its zero-lag position. Both the
O- and K-sands are in the same depth interval as the To estimate the amount of error in lateral positioning
B-fault plane. due to over- or under-migration at the B-fault, we use
The similarity of the amplitudes in the 1985 and the approximate depth z and dip of the fault given
1992 surveys at the O-sand support the hypothesis that by 2 km and 45 , respectively. From equation (7), a
the 1 km of movement seen at the B-fault is an actual 5% error in migration velocity results in a lateral po-
change in reflectivity; however, we need to show that sitioning error of 200 m, less than the 700 m of lat-
the 1 km of movement in the seismic image is not the eral movement we observe at the B-fault in the 1985
result of over- or under-migration, given typical errors and 1992 seismic surveys. Given that the geology above
in migration velocities. Note that the 1 km of movement the B-fault is essentially a vertically-varying, compact-
up the B-fault, which dips at 45 , is equivalent to 700 ing stack of sediments and that a considerable number
m of lateral movement. To explore the issue of over- or of wells with wireline data passed through this section,
under-migration, we first make the approximation that we feel that a 5% error in migration velocity is a con-
the shooting lines in the 1985 and 1992 seismic sur- servative estimate of the error. In fact, a 700 m error in
veys were in the dip direction. This means the action lateral positioning requires a 17.5% error in migration
of 3D migration on the reflection from the B-fault is velocity, much too large for a target in a 1D geologic
purely a vertical and lateral movement in a single ver- setting above salt. Furthermore, the movement of the
tical plane. We then use the expression for the lateral pulse observed in the 1985 and 1992 data, if an artifact
movement applied to a time sample in a stacked sec- of incorrect migration velocities, would require that the
tion by a constant-velocity 2D migration, denoted xm errors in the two data sets be negatively correlated. For
(Yilmaz, 1987) example, if the 1985 data had a 5% error in migra-
tion velocity (under-migration) and the 1992 data had
2
ptin vm a +5% error in migration velocity (over-migration), the
xm = . (2)
4 residual mis-positioning between the two surveys would
In equation (2), tin is the two-way time of the input still only be 400 m. If both surveys had a +5% error in
point, vm is the migration velocity, and p is the time migration velocity, there would be no apparent lateral
movement. We therefore conclude that 1 km of move-
ment up the B-fault is significant enough to stand out
in the presence of typical positioning errors due to in-
correct migration velocities.
Observation of a fault burp 223
CONCLUSION in Eugene Island 330 Field, Gulf of Mexico. Lamont-
Doherty Earth Observatory: Global Basins Research Net-
At a location on the B-fault plane at the South Eugene work.
Island Block 330 field, where anomalously high pore Cervelli, P., Segall, P., Johnson, K., Lisowski, M., & Miklius,
pressures have been reported in the fault zone (Ander- A. 2002. Sudden aseismic fault slip on the south flank of
son et al., 1995; Losh et al., 1999), areas of high reflec- Kilauea volcano. Nature, 415, 10141018.
tivity appear to move up the fault-plane, as would be ex- Chaytor, J. D., Goldfinger, C., Huftile, G. J., Romsos,
pected for an ascending, pressure-driven fluid pulse. The C. G., & Meiner, M. 2004. Submerged Shorelines of Pil-
seismic waves are able to sense these different pressure grim Banks and the Northern Channel Islands, Southern
domains due to the effective stress controlling the elastic California Continental Borderland, as Vertical Tectonic
Strain Markers. In: EOS Trans. AGU, 85(47), Fall Meet.
properties of the soft Plio-Pleistocene sediments (Haney
Suppl., Abstract T13C-1398. AGU.
et al., 2005). This observation gives additional credence Dobrin, M. B. 1976. Introduction to geophysical prospecting.
to the dynamic fluid injection hypothesis (Whelan et al., New York: McGraw-Hill.
2001) and points to natural flow processes in sedimen- Elliott, D. A. 1999. Hydrofracture Permeability Response
tary basins that occur on production time scales. and Maximum Previous Consolidation Stress Estima-
The movement seen in the 1985 and 1992 seis- tions for Faulted and Micro-Faulted Silty-Shales Taken
mic data warrants more extensive study of the faults from the Eugene Island Block 330 Field Pathfinder Well
at South Eugene Island. One possibility in the future in the Gulf of Mexico. MSc Thesis, University of Cali-
is to go to the fault scarps on the ocean floor with a fornia, San Diego.
small submersible and take samples of the fluids com- Finkbeiner, T., Zoback, M. D., Stump, B., & Flemings, P.
2001. Stress, pore pressure and dynamically-constrained
ing directly out of the fault scarps, since Anderson et
hydrocarbon column heights in the South Eugene Island
al. (1994) report oil seeps that correlate with their loca- 330 Field, Gulf of Mexico. AAPG Bulletin, 85, 1007
tion. The shallow water depth at South Eugene Island 1031.
(< 100 m) makes such an investigation using a sub- Haney, M., Sheiman, J., Snieder, R., Naruk, S., Busch, J.,
mersible possible. Submersibles have proven to be useful & Wilkins, S. 2004. Fault-plane reflections as a diag-
in sub-sea geological studies, as shown by Chaytor et al. nostic of pressure differences in reservoirs: a case study
(2004) in the Pacific ocean and Roberts (2001) in the from South Eugene Island Block 330. Pages O07 of:
Gulf of Mexico. Other data sets to incorporate into fu- Engelder, T., Konstanty, J., & Grauls, D. (eds), Proceed-
ture fault studies are prestack data, especially for doing ings of the EAGE special session on Fault and Top Seals.
AVO analysis on a fault-plane reflection, and microseis- EAGE.
Haney, M., Snieder, R., & Ampuero, J. 2005. Spectral ele-
micity data. Without question, future studies on faults
ment modeling of fault-plane reflections arising from fluid
should incorporate time-lapse measurements to capture pressure distributions. in: CWP Project Review.
the essence of faults as dynamic and quickly-changing Holland, D. S., Leedy, J. B., & Lammlein, D. R. 1990. Eugene
zones. Island Block 330 field-USA, offshore Louisiana. Pages
103143 of: Beaumont, E. A., & Foster, N. H. (eds),
Structural traps III, tectonic fold and fault traps: AAPG
ACKNOWLEDGMENTS Treatise of Petroleum Geology Atlas of Oil and Gas.
AAPG.
Supported by a research grant from Shell International Hooper, E. C. D. 1991. Fluid migration along growth faults
Exploration and Production, whom also provided the in compacting sediments. J. of Pet. Geology, 14, 161
1992 seismic data. The 1985 data, originally acquired by 180.
Pennzoil, is currently in the public domain. We thank Losh, S., Eglinton, L., Schoell, M., & Wood, J. 1999. Vertical
and Lateral Fluid Flow Related to a Large Growth Fault.
R. Anderson, W. He, S. Naruk, J. Busch, and S. Wilkins
AAPG Bulletin, 83, 244276.
for help and discussions.
Losh, S., Phipps-Morgan, J., & Haney, M. 2005. High-
Pressure Fluid Pulses in an Active Fault, Offshore
Louisiana Gulf of Mexico. In: AGU Chapman Confer-
REFERENCES ence on radiated energy and the physics of earthquake
faulting.
Anderson, R. N., He, W., Hobart, M. A., Wilkinson, C. R., &
Miller, S. A., Collettini, C., Chiaraluce, L., Cocco, M.,
Roice, H. R. 1991. Active fluid flow in the Eugene Island
Barchi, M., & Kaus, B. J. P. 2004. Aftershocks driven
area, offshore Louisiana. The Leading Edge, 10, 1217.
by a high-pressure CO2 source at depth. Nature, 427,
Anderson, R. N., Flemings, P., Losh, S., Austin, J., & Wood-
724727.
hams, R. 1994. Gulf of Mexico growth fault drilled, seen Nur, A., & Walder, J. 1992. Hydraulic Pulses in the Earths
as oil, gas migration pathway. Oil & Gas Journal, 92, Crust. Pages 461474 of: Evans, B., & Wong, T.-F. (eds),
97103. Fault Mechanics and Transport Properties of Rocks. San
Anderson, R. N., Billeaud, L. B., Flemings, P. B., Losh, Diego: Academic Press.
S., & Whelan, J. K. 1995. Results of the Pathfinder Revil, A., & Cathles, L. M. 2002. Fluid transport by soli-
Drilling Program into a Major Growth Fault, Part of tary waves along growing faults: A field example from
the GBRN/DOE Dynamic Enhanced Recovery Project the South Eugene Island Basin, Gulf of Mexico. Earth
224 M. Haney, R. Snieder, J. Sheiman & S. Losh
and Planetary Science Letters, 202, 321335.
Rice, J. R. 1992. Fault Stress States, Pore Pressure Distribu-
tions, and the Weakness of the San Andreas Fault. Pages
475503 of: Evans, B., & Wong, T.-F. (eds), Fault Me-
chanics and Transport Properties of Rocks. San Diego:
Academic Press.
Roberts, H. H. 2001. Fluid and gas expulsion o the northern
Gulf of Mexico continental slope: Mud-prone to mineral-
prone responses. Pages 131143 of: Paull, C. K., & Dil-
lon, W. P. (eds), Natural Gas Hydrates: Occurence, dis-
tribution, and dynamics. Geophysical Monograph 124.
San Francisco: AGU.
Sibson, R. H. 1990. Conditions for fault-valve behavior.
Pages 1528 of: Knipe, R. J., & Rutter, E. H. (eds),
Deformation Mechanisms, Rheology and Tectonics. Ge-
ological Society Special Publication, no. 54. Oxford:
Alden Press.
Whelan, J. K., Eglinton, L., Kennicutt, II, M. C., & Qian,
Y. 2001. Short-time-scale (year) variations of petroleum
fluids from the U. S. Gulf Coast. Geochim. Cosmochim.
Acta, 65, 35293555.
Yilmaz, O. 1987. Seismic Data Processing. Tulsa: Society of
Exploration Geophysicists.
CWP-514
ABSTRACT
0.4
1
0.3
porosity
2
0.2
0.1
100 m A20ST A
0
0 1000 2000 3000
Z Hydrostatic depth (m)
B
F Figure 2. Porosity versus depth at South Eugene Island.
The thick, solid line is the best-fit normal compaction trend
using Athys Law (Athy, 1930). The faint solid lines are
Overpressured
density-derived porosity values from 11 wells at South Eu-
gene Island. To obtain the porosity, we assume that the solid
grains have a density of 2650 kg/m3 and the fluid has a den-
sity of 1000 kg/m3 , as in Revil and Cathles (2002). There
N is a clear break from the shallow, exponentially decreasing
porosity trend at a depth of 1800 m, at which point the
Figure 1. Regional map (top), from Anderson et al. (1995), porosity remains constant with increasing depth, as shown
and cartoon depth section (bottom) of the subsurface at by the flat dashed line. The two circles are density-derived
South Eugene Island. The four main faults discussed in this porosities from the upthrown block to the north of the mini-
thesis are shown in the bottom panel as the A, B, F, and Z basin at South Eugene Island. The dashed lines connecting
faults. Throw across the faults is depicted by the layer run- the circles to the main compaction trend are the interpreted
ning from left to right. Most of the wells at South Eugene porosity histories of the samples. They show a period of un-
Island were drilled into the shallow, hydrostatic section; the dercompaction, depicted as a horizontal line deviating from
A20ST well was unusual in that it was continued through the normal compaction trend, followed by a vertical unload-
the A-fault system and into the deep overpressured compart- ing path due to a late-stage pore-pressure increase.
ment.
2200
1 between porosity and density described earlier. After
finding this slope, we can construct the unloading path
for the density from equation (6) and the slope
2000
u (d ) = 0.04 (d max ) + s
0.47 (s f ) e0.0003max . (8)
1800
0 1000 2000
P (psi)
3000 4000 This expression contains an extra parameter max that
v
refers to the value of the effective stress when the sample
Figure 4. Density versus effective stress at South Eugene began to be unloaded. We do not know max for samples
Island. The thick solid line is the same normal compaction 1 and 2, but we do know that max must lie on the
trend shown in Figure 2, except transformed into density and main compaction trend. Hence, we can construct linear
effective stress. The faint solid lines are also the same as in
unloading paths for the density, as shown by the dashed
Figure 2, except that they are now limited to the hydrostatic
lines in Figure 4. With these unloading paths, we can
depths down to 1800 m. The circles represent two pressure
measurements, labeled 1 and 2, which were made in the over- then find the value for the maximum past effective stress
pressured upthrown block where a density log also existed. max . It is worth mentioning that the maximum past
For each pressure measurement, we plot the data point twice effective stress for sample 1 comes out to be 1500 psi
- one where it should lie on the normal compaction curve were by our approach of using Elliotts experimental results.
it to have been normally pressured, and the other where it In an independent measurement, Stump et al. (2002)
actually does plot because of severe overpressure. Note that performed uniaxial strain tests on a core sample taken
sample 1 is from a greater depth than sample 2. from the same location as sample 1 to find the maximum
past effective stress. Stump et al. (2002) report a value
ditions. From these relationships and equation (3), we of 1248 psi for this sample, close to our estimate of
obtain the normal compaction curve for density 1500 psi; visually, the discrepancy lies within the error
bars of the normal compaction curves fit to the density
c (d ) = s 0.47 (s f ) e0.0003d , (6) log data.
where s and f are the densities of the solid and fluid With the estimate of the maximum past effective
components, taken as 2650 kg/m3 and 1000 kg/m3 re- stress, we can also return to Figure 2 and find the depth
spectively, and d is in psi. We plot this normal com- at which samples 1 and 2 left the normal compaction
paction curve in Figure 4 together with the density mea- trend, since in the hydrostatic zone the depth is a lin-
surements. Also, in Figure 4, we show as circles two early scaled version of the effective stress. These depths
data points obtained from RFT pressure measurements correspond to a slightly lower porosity than that of sam-
and density log measurements in the overpressured up- ples 1 and 2. We interpret this as being the result of a
thrown block. We show the circles in two locations - late stage porosity increase and represent it as a vertical
one on the normal compaction trend where they would unloading path for samples 1 and 2 in Figure 2.
plot if the measurements were at hydrostatically pres-
sured locations, and the other where they actually plot
because of severe overpressures being present in the up- 3 SONIC VELOCITY VERSUS VERTICAL
thrown block. EFFECTIVE STRESS
At this point, we dont know exactly how the sam- For the purposes of modeling faults and to make infer-
ples taken in the upthrown block came to be off the ences about the distribution of pore pressure from seis-
normal compaction trend. Using a laboratory measure- mic interval velocity inversions, accurate pore-pressure-
ment of the unloading coefficient by Elliot (1999) on a versus-velocity relationships are critical (Dutta, 1997).
core sample taken near the locations of samples 1 and 2, In general, sonic velocity has a normal compaction curve
the path that these samples took to their present loca- and unloading paths as a function of effective stress that
tions can be estimated. Elliot (1999) characterized the are similar to those we just described for the density well
unloading, or elastic swelling, for the porosity of the core log data. To obtain these relationships, we proceed as
samples to be for the density logs: 1) We take 12 shallow wells to make
u (d ) = 0 (1 d ) , (7) up a data set of sonic velocity versus effective stress. 2)
We select the depth range with hydrostatic pressures
where 0 and characterize the deviation of the unload-
and plot the sonic velocity versus effective stress. 3) We
ing path from the normal compaction trend. Note the
fit this with a power law relation for the normal com-
superscript u, in contrast to equation (3), indicating the
230 M. Haney, R. Snieder & J-P. Ampuero
3000 found for vs on effective stress is essentially a down-
shifted version of the vp curve. An additional piece of
1
2 supporting evidence comes from the only vs data avail-
able at South Eugene Island, a shear log from the A20ST
sonic velocity (m/s)
2500
well, where samples 1 and 2 were taken. There, the ra-
2
tio of vp /vs from the sonic and shear logs falls between
1 3 to 3.5 in the overpressured upthrown block. Inserting
2000 the values for vp at samples 1 and 2 into equation (11)
to get vs and finding the corresponding ratio of vp /vs ,
we get vp /vs = 3.48 at sample 1 and vp /vs = 2.96 at
sample 2, within the range of the ratios observed in the
1500
0 1000 2000 3000 4000 sonic and shear logs.
P (psi) To summarize, we have established two empirical
v
Figure 5. Sonic velocity versus effective stress at South Eu- relationships between each of three basic rock proper-
gene Island. The thick solid line represents the normal com- ties and pore pressure at the South Eugene Island field.
paction curve fitted to the shallow well data, shown in the Most important for subsequent numerical modeling of
faint solid lines. We also plot samples 1 and 2 both where wave propagation, we have found relationships for the
they should fall on the normal compaction trend, were they density and the sonic velocity vp on both the nor-
to be normally pressured, and where they actually plot due mal compaction and unloading paths. Without shallow
to the severe overpressure where they were obtained. Using information on the shear velocity vs , we must make
the estimate for past maximum effective stress from the den-
the assumption that it is a down-shifted version of the
sity plot and the Bowers-type relation (Bowers, 1995) shown
vp (d ) relationship. From looking at the density-derived
in equation (10), we are able to construct the velocity un-
loading curves, shown as dashed lines. porosity-versus-depth-relationship, we are able to con-
clude that the deep, overpressured sediments below the
JD-sand are predominately overpressured because of
paction trend. 4) We then look at where the two samples compaction disequilibrium, since their porosity did not
from the overpressured upthrown block lie and construct change appreciably with depth. In contrast, both com-
unloading curves using the estimate for the maximum paction disequilibrium and unloading have contributed
past effective stress that we obtained in the previous to the current overpressured state of the sediments on
section on density. The wells we use for characterizing the upthrown side. The latter conclusion is in agreement
the sonic velocity come from A20ST, A14OH, A23, A6, with a previous study by Hart et al. (1995) on porosity
B10, B1, B2, B7, B8, A1, B14, and B20. and pressure at South Eugene Island. We use the em-
In Figure 5, we plot the normal compaction trend pirical relationships between the elastic parameters and
for sonic velocity as a thick solid line described by the fluid pressure to simulate fault-plane reflections from
power law equation (Bowers, 1995) different pressure distributions in the subsurface.
vpc (d ) = 1500 + 2.3 d0.77 , (9)
where vp is in m/s and d is in psi. Note again the 4 THE SPECTRAL ELEMENT METHOD
superscript c for the normal compaction relation. We
also construct the unloading curve for vp following the Numerical modeling of wave propagation in the Earth
relationship first suggested by Bowers (1995) can be based on either the weak (Zienkiewicz & Tay-
" 1/6.2 #0.77 lor, 2000) or strong (Boore, 1970) forms of the elasto-
u d dynamic equations of motion. By weak and strong, we
vp (d ) = 1500 + 2.3 max , (10)
max mean the integrated or differential forms of the equa-
tions of motion. The spectral element method (SEM),
where d and max are in psi and vp is again in m/s. though based on the weak form, combines favorable
To model elastic waves, one other parameter is aspects of both strong and weak formulations. For
needed in addition to and VP ; for instance, a seismol- instance, SEM naturally handles general geometries
ogist would naturally want the shear velocity. In the and exotic boundary conditions. In the finite-difference
absence of information on the shear wave velocity vs method (based on the strong form), it is notoriously
and pressure in the shallow, hydrostatic sediments, we difficult to implement a linear-slip boundary condition
assume that (Coates & Schoenberg, 1995) or any general boundary
vs (d ) = vp (d ) 1500, (11) condition for that matter (Boore, 1970; Kelly et al.,
1976). On the other hand, SEM does not require the
where this relationship holds on both the normal com- inversion of a large matrix, a property usually identified
paction curve and the unloading path. The data pre- with finite-difference methods. Formally, this last prop-
sented by Zimmer et al. (2002) for unconsolidated sands erty of SEM means that its mass matrix is diagonal
supports this assumption, in that the dependence they and its computational cost is relatively small. Note that
Fluid pressure and fault reflectivity 231
receiver array
source array
(3.0-GHz) cluster. All of the processing has been per-
formed on a workstation using the Seismic Un*x pack-
age (Stockwell, 1997). Figure 6 depicts the geometry of
2 km
the model. The normal fault we model has a vertical
throw of 20 m, a value characteristic of a small fault.
The model shown in Figure 6 has been previously stud-
ied by Townsend et al. (1998) in order to assess changes
4 km
in seismic attributes caused by faults disrupting the lat-
eral continuity of events.
1 We mesh the interior of the computational grid
2
shown in Figure 6 using a freely available mesh pro-
7
3
Layer # 4
8 gram developed by INRIA, called EMC2. The program
9 Layer #
5
6
10 can be downloaded at:
11
http://www-
12
rocq.inria.fr/gamma/cdrom/www/emc2/eng.htm.
Figure 6. The entire numerical model with a zoom-in of the For the examples in this paper, we use a semi-structured
normal fault. The zoom area is shown on the entire numerical mesh since the fault geometries modeled are not overly
model with a dashed rectangle. The lengths of the source complex. A semi-structured mesh is desirable when pos-
and receiver arrays are shown by extended arrows. Eleven sible since the accuracy of SEM depends on the Jacobian
sources evenly spaced over 2000 m (200-m shot interval) and of the transformation between a generally shaped ele-
241 receivers over 3000 m (12.5-m receiver interval). In the ment and a standard rectangular element over which the
zoom, the layers are labeled with numbers 1-12 corresponding integration is performed. Though the mesh has struc-
to the material properties for models listed in subsequent ture, it honors the slanted boundaries of the fault. Af-
tables. For models with a slipping fault, the portion of the
ter initial construction of the mesh, the quadrangle el-
fault plane that slips is shown by a thicker line in the zoom.
ements are regularized so that their shapes mimic rect-
angles as closely as possible. At the fault, the possibil-
SEM does this in a way similar to mass-lumping (Karni- ity exists for the fault to experience linear-slip (Schoen-
adakis & Sherwin, 1999), which has been used to diag- berg, 1980). With SEM, we can incorporate this aspect
onalize finite-element schemes. SEM has the additional by using a split-node (Andrews, 1999). In Appendix A,
property of spectral convergence, meaning that, as the we show how to incorporate a linear-slip interface into a
polynomial order of the basis functions is increased, the finite-element algorithm with a split-node. We also show
numerical error goes down exponentially (Karniadakis that the presence of a linear-slip interface modifies the
& Sherwin, 1999). numerical stability criterion, a fact that is also true for
The term spectral element indicates that SEM is SEM.
a mixture of finite-element and spectral methods (Ko- Since the SEM code is elastic, both primary and
matitsch & Vilotte, 1998). As a result, there are two converted waves show up on the vertical component
parameters relevant to the mesh in SEM: the size of of the displacement seismograms. We mute the con-
the elements and polynomial degree (n 1, where n is verted waves in order to proceed with conventional P -
the number of zero crossings of the basis functions used wave time-processing. We subtract off the direct waves
within each element). Komatitsch and Tromp (2003) re- by running a homogeneous subsurface simulation with
fer to these parameters when they speak of the global the elastic properties of layer 1 of Figure 6. After this
mesh and the local mesh. Concerning the local mesh, step, we perform a geometrical-spreading correction,
there is a known trade-off between accuracy and numer- NMO, DMO, and stack to simulate zero-offset data.
ical cost (Seriani & Priolo, 1994), which suggests that With the simulated zero-offset section, we proceed with
polynomial degrees no higher than 10 should be used a constant-velocity migration using the velocity of the
within the elements. For the numerical examples in this overburden (layer 1). Hence, a source of error in this
paper, we use a polynomial degree of eight. simulation originates from the slight undermigration of
the deepest reflectors and the fault-plane reflection. We
chose to migrate with constant velocity since we have
5 MODELING OF A FAULT interpreted time-migrated seismic sections in the Gulf
of Mexico (Haney et al., 2004) and wanted the SEM
As an example of the ability of SEM to model seis- modeling to mimmick the data as closely as possible.
mic scale structures, we discuss a complete modeling With this full suite of forward modeling and process-
and processing sequence for a simple fault model. The ing capabilities, we apply SEM to study various normal
SEM forward modeling has been run in serial (one fault models. Before going into the details of the mod-
node for each shot) on a 32-processor pentium IV Xeon
232 M. Haney, R. Snieder & J-P. Ampuero
2h h 0 h 2h
x (midpoint) 1
p
0.8
amplitude of K
p
st
0.6
0.4
0.2
0
t(x) 0 2 4 6 8 10
frequency (s1)
eling, though, we present the type of dip-filter we use to Since the time shifts in equation (15) are linear in m,
isolate the migrated fault-plane reflections. the series can be evaluated exactly.
Using the geometric series, equation (15) can be
written as
6 DIP FILTERING 1 ei(n+1)(ppst )h
1
K() = +
Using an array-based approach, we show in this section 2n + 1 1 ei(ppst )h
1 ei(n+1)(ppst )h
the form of the dip filter that we apply to migrated im-
1 . (16)
ages to accentuate the fault-plane reflections. Figure 7 1 ei(ppst )h
shows a dipping event in a post-stack seismic image. The
Putting the first two terms in the brackets of equa-
function of the dip filter is to stack along the dashed line
tion (16) under a common denominator and simplify-
and place the stack result at the intersection of the solid
ing further yields a filter similar to that obtained for
and dashed lines; where the coordinate system is cho-
convolution with a boxcar, or rectangular window (Op-
sen such that the midpoint is equal to 0. Note that, in
penheim & Schafer, 1975)
Figure 7, the dip filter emphasizes a direction different
from the dip of the event. This difference induces a time 1 sin[(n + 1)(p pst )h/2]
K() = 2
shift at the m-th input trace (shown as a vertical arrow). 2n + 1 sin[(p pst )h/2]
Suppose that the true dip, the dip of the event, is p (di- cos[n(p pst )h/2] 1] . (17)
mension s m1 ) and the dip of the stacking curve is pst .
The time shift at the m-th input trace is thus This filter is real because of the symmetric summation
about its output point and can be better understood by
tm = (p pst ) mh, (12) making the substitution t = (ppst)h in equation (17)
where h is the midpoint spacing and m is the discrete 1
sin[(n + 1)t/2]
variable running over midpoint location. Assuming that K() = 2
2n + 1 sin[t/2]
the dipping event has a constant waveform f (t), the
result of the summation, g(t), over the stacking curve cos[nt/2] 1] . (18)
can be written as The amplitude spectrum of the dip filter is shown in
1 Figure 8 for certain values of n and t.
g(t) = [f (t tn ) + + f (t t1 )+
2n + 1 A dip filter in terms of and k is needed to enhance
f (t) + f (t t1 ) + + f (t tn )] , (13) fault-plane reflections on a migrated time section. The
filter in equation (17) is only in terms of . To get the k
where 2n+1 is the length of the dip filter in terms of the dependence, we exploit the fact that p = k/ (Hatton
number of traces. In the example shown in Figure 7, n = et al., 1986). This means that linear features with a
2. When we again take the Fourier transform over time dip p in the t-x domain get mapped into linear features
and move to the -x domain, equation (13) becomes with the negative dip in the -k domain. Substituting
G() = F ()K(), (14)
Fluid pressure and fault reflectivity 233
p = k/ into equation (17) gives 2.0
midpoint (km)
2.5 3.0 3.5 4.0
x10 5
2.0
midpoint (km)
2.5 3.0 3.5 4.0
x10 5
0 0
1 sin[(n + 1)(pst + k)h/2]
K(, k) = 2 0.2 0.2
(km)
This is the form of a dip filter that corresponds to stack- 0.8 0.8
ing 2n + 1 traces centered about the output point along 1.0 1.0
a dip pst .
1.2 1.2
1.4 1.4
x10 5 x10 5
The dip filter operation discussed in the previous sec- 0.2 0.2
simulated reflection images for the two of the models 1.2 1.2
sions that highlight the fault-plane reflection. The dip x10 5 x10 5
velocity (m/s)
0.06 3000
2500
0.03
2000
amplitude
0
1940 1960 1980 2000 2020 2040
0.03
2400
density (kg/m3)
0.06
2200
950 1000 1050 1100 1150
depth (m)
0.2
2000
1940 1960 1980 2000 2020 2040
SSTVD (m)
0.1
Figure 11. Velocity and density logs in the A20ST well at
amplitude
0
the South Eugene Island field showing the interface between
the JD-sand and its lower bounding shale at 1975 m sub-sea
true vertical depth (SSTVD). The log information at this
0.1
contact is used to construct the sand/shale model in Table 1.
These logs also cut through a growth fault zone known as the
0.2
B-fault at around 2020 m SSTVD.
950 1000 1050 1100 1150
depth (m)
Figure 10. The reflected waves from Model 2 (top), a pore plane of the noise-free dip-filtered image with the ampli-
pressure contrast, and Slip-Model A embedded in Model 1 tude extracted after 20% Gaussian noise was added to
(bottom), a slipping fault, are shown in dashed lines. The re- the original image before dip filtering. As can be seen,
flected waveforms for these two models are different in shape. the dip-filtering step is robust in the presence of random
The reflected waveform for Model 2 should be the wavelet,
noise, giving roughly the 1/ n attenuation of noise ex-
which is plotted in the top panel as a solid line for compari-
ploited in stacking. The processing is such that the am-
son with the numerical result. The waveform for Slip-Model
A embedded in Model 1 should be the the derivative of the plitude of the reflected waves is an indication of the
wavelet, which is plotted in the bottom panel as a solid line reflection coefficient at the boundary giving rise to the
for comparison. The agreement between the numerical re- reflected wave, at least within any changes induced by
sults for the fault-plane reflections and the expected wave- amplitude variations with offset (AVO) in the stack. The
forms validates the sequence of modeling, processing, and pattern of the amplitude map reflects the spatially vary-
dip-filtering used here. ing presence of juxtaposition contacts for this model, as
the amplitude wildly oscillates up and down. The sec-
ond scenario, shown in the lower-left panel of Figure 12,
Layer Thickness (m) (kg/m3 ) vp (m/s) vs (m/s) compares the extracted amplitude on the fault-plane
of the noise-free dip-filtered image with automatic gain
1 900 2240 2600 1100 control (AGC) applied to the amplitude extracted after
2 50 2280 2750 1250
20% Gaussian noise was added to the original image be-
3 30 2240 2600 1100
4 50 2280 2750 1250
fore dip filtering and AGC. The AGC operates in a time
5 30 2240 2600 1100 window of 200 ms. In this panel, the true-amplitude de-
6 90 2280 2750 1250 grading nature of the AGC is clear (the absolute value
7 50 2280 2750 1250 of the amplitude is completely different from the result
8 30 2240 2600 1100 without AGC); however, along the fault some of the
9 50 2280 2750 1250 same qualitative patterns are present. The same can be
10 30 2240 2600 1100 said for the third scenario in the upper-right panel of
11 90 2280 2750 1250 Figure 12. In this panel, everything is the same as in
12 850 2240 2600 1100 the upper-left panel, except that DMO has been omit-
ted in the processing sequence. The omission of DMO
Table 1. Model 1 for the SEM modeling. The throw between serves to dampen the fault-plane reflections as a result
the upthrown (layers 2 through 6) and downthrown (layers 7
of mis-stacking. The lower-right panel of Figure 12 is
through 11) sediments is 20 m. The sediments on both sides
of the fault are at hydrostatic pore pressure. The geometry
the same as in the lower-left panel, except that DMO
of the model is given in Figure 6. The values are taken from has again been omitted in the processing sequence. This
the JD-sand and its lower bounding shale at a depth of 2 km. is the most extreme and problematic of the processing
The pore pressure is 2800 psi at that location. sequences, with AGC and without DMO; the patterns
on the fault-plane seen in the upper-left panel are barely
recognizable.
3 3
x 10 x 10
7.5 7.5
Fluid pressure and fault reflectivity 235
amplitude
amplitude
5 5
amplitude
amplitude
5 5
7 50 2280 2750 1250
2.5 2.5
8 30 2240 2600 1100
9 50 2280 2750 1250
0 0 10 30 2240 2600 1100
800 900 1000 1100
depth (m)
1200 1300 800 900 1000 1100
depth (m)
1200 1300 11 90 2280 2750 1250
12 850 2240 2600 1100
Figure 12. Amplitude along fault-plane for Model 1 shown Table 2. Model 2 for the SEM modeling. The geometry of
in Table 1, the juxtaposition contact model, with four differ- the model is given in Figure 6. There is 20 m of throw be-
ent processing scenarios. The amplitude is plotted as a func- tween the upthrown (layers 2 through 6) and downthrown
tion of depth on the fault-plane, and the extent of the fault- (layers 7 through 11) sediments, in addition to 600 psi of
plane is shown by a horizontal line in the bottom portion of pore pressure difference. The difference in pore pressures is
each plot. The processing scenarios are: 1) upper-left panel: the result of the upthrown sediments being overpressured.
amplitude extracted from the noise-free image (dashed) and The densities and velocities of layers 2 through 6 are rel-
from the image with 20% additive Gaussian noise (solid), atively lower than the normally compacted, hydrostatically
2) lower-left panel: amplitude extracted from the noise-free pressured layers 7-11 on the downthrown side. To relate the
image with AGC applied (dashed) and from the image with velocities and densities to pressure, the mechanism of over-
20% Gaussian noise added in before applying AGC (solid), pressure is taken to be purely compaction disequilibrium.
3) upper-right panel: amplitude extracted from the noise-
free image without DMO included in the processing sequence
(dashed) and from the image with 20% additive Gaussian
noise, but no DMO (solid), and 4) lower-right panel: ampli- Layer Thickness (m) (kg/m3 ) VP (m/s) VS (m/s)
tude extracted from the noise-free image with AGC applied
but no DMO (dashed) and from the image with 20% Gaus- 1 900 2240 2600 1100
sian noise added before applying AGC but no DMO (solid). 2 50 2260 2660 1160
There is no slip at the fault. 3 30 2225 2490 990
4 50 2260 2660 1160
5 30 2225 2490 990
9 AMPLITUDE OF WAVES REFLECTED 6 90 2260 2660 1160
FROM A PORE-PRESSURE CONTRAST 7 50 2280 2750 1250
ACROSS A FAULT 8 30 2240 2600 1100
9 50 2280 2750 1250
In Tables 2, 3, and 4 are three different models that in- 10 30 2240 2600 1100
corporate pore pressure contrasts of 600 psi, 300 psi, and 11 90 2280 2750 1250
150 psi, respectively, across the fault. The aim of using 12 850 2240 2600 1100
these models is to see how small a pore pressure contrast
can be and still be detectable in the fault-plane reflec- Table 3. Model 3 for the SEM modeling. The geometry of
tivity. These figures are relevant to studying the fault- the model is given in Figure 6. Same as Table 2 except that
plane reflectivity at South Eugene Island since pressure the pore pressure difference is 300 psi.
measurements taken near the large, minibasin-bounding
growth fault, known as the A-fault, show a 780 psi in-
crease in pore pressure over 18 m in going from the duction assuming that the mechanism of overpressure
hydrostatically pressured downthrown sediments to the is pure undercompaction.
overpressured upthrown sediments (Losh et al., 1999). In Figure 13 are plotted the maximum reflected am-
With the pore-pressure contrasts in the three models plitude from Model 2 (P = 600 psi) within a small
being 600 psi, 300 psi, and 150 psi, the effective stress window (100 ms) of the fault-plane for four different
values can be computed using a hydrostatic pressure of processing scenarios. These different processing scenar-
2800 psi (the same as for the juxtaposition model cov- ios are the same as in the previous section on juxtapo-
ered in the previous section). The juxtaposition model sition contrasts. The scenarios include the presence and
can be thought of as having a 0 psi pore-pressure dif- absence of 20% additive Gaussian noise with the full
ference across the fault. The values of the densities and processing sequence (upper-left panel of Figure 13), the
velocities used in the three models are consistent with application of AGC (lower-left panel of Figure 13), the
the effective stress relationships derived in the intro- omission of DMO in the processing sequence (upper-
236 M. Haney, R. Snieder & J-P. Ampuero
0.04
(kg/m3 )
0.02
Layer Thickness (m) VP (m/s) VS (m/s)
0.03
amplitude
1 900 2240 2600 1100
amplitude
0.02 0.01
2 50 2270 2705 1205
3 30 2232 2545 1045 0.01
5 30 2232 2545 1045 800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300
depth (m) depth (m)
6 90 2270 2705 1205 3
x 10
3
x 10
7 50 2280 2750 1250 16
7.5
8 30 2240 2600 1100 12
9 50 2280 2750 1250
amplitude
amplitude
5
10 30 2240 2600 1100 8
800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300
Table 4. Model 4 for the SEM modeling. The geometry of depth (m) depth (m)
the model is given in Figure 6. Same as Table 2 except that Figure 14. Amplitude along fault-plane for Models (clock-
the pore-pressure difference is 150 psi. wise from the upper left) 2 (upper-left), 3 (upper-right), 4
(lower-left), and 1 (lower-right). The amplitude is plotted
as a function of depth on the fault-plane and the extent of
0.04 0.04 the fault-plane is shown with a horizontal line in the bot-
tom portion of each plot. These models represent the sit-
0.03 0.03
uations of: 600-psi pore-pressure difference across the fault
amplitude
amplitude
0.02 0.02
(Model 2, upper-left), 300-psi pore-pressure difference across
0.01 0.01
the fault (Model 3, upper-right), 150-psi pore-pressure dif-
ference across the fault (Model 4, lower-left), and 0 psi pore
0 0
pressure difference across the fault; in other words, the jux-
800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300
depth (m) depth (m) taposition contact model (Model 1, lower-right), There is no
20 20
slip at the fault.
15 15
amplitude
amplitude
amplitude
amplitude
amplitude
amplitude
0.08 0.08 0.08
0.04
0.04 0.04 0.04
0 0 0 0
800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300
depth (m) depth (m) depth (m) depth (m)
40 40
0.09 0.04
30 30
0.06
amplitude
amplitude
amplitude
amplitude
20 20 0.02
0.03
10 10
0 0 0 0
800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300
800 900 1000 1100 1200 1300 800 900 1000 1100 1200 1300 depth (m) depth (m)
depth (m) depth (m)
Figure 15. Amplitude along fault-plane for Slip-Model A Figure 16. Amplitude along fault-plane for Slip-Models A
(see Table 5) at a fault embedded in Model 2 (see Table 2) for (upper-left), B (upper-right), C (lower-left), and D (lower-
four different processing scenarios. The amplitude is plotted right) shown in Table 15 at a fault embedded in Model 2
as a function of depth on the fault-plane and the extent of the having a pore pressure contrast of 600 psi across the fault.
fault-plane is shown with a long horizontal line in the bottom The amplitude is plotted as a function of depth on the fault-
portion of each plot. The sub-portion of the fault that slips plane and the extent of the fault-plane is shown with a long
is shown beneath this line with a shorter horizontal line. The horizontal line in the bottom portion of each plot. The sub-
processing scenarios are the same as those described in the portion of the fault that slips is shown beneath this line with
caption for Figure 12. The Slip-Model uses slip at the fault a shorter horizontal line. The maximum amplitude occurs
to represent weakness caused by elevated pore pressure. near the center of the slipping portion of the fault in each slip-
model. The amplitude map also resembles a triangle since it
is in essence the convolution of two boxcar functions: the slip-
Figure 6. In the next section, we present a weak scat- ping portion of the fault-plane and the dip filter. In moving
tering model to relate the value of the compliances to from Slip-Model A to D, the compliance of the fault slip be-
pore pressures in a thin fault zone. For now, these values comes smaller and, as a result, the reflection magnitude scales
in the same proportion, as predicted by equation (B6) in Ap-
are simply parameters describing the degree of bonding
pendix B for a weakly slipping interface. The Slip-Models use
between the surfaces on either side of the linear-slip in-
slip at the fault to represent weakness due to elevated pore
terface. pressure.
The amplitude of the fault-plane reflection from
Slip-Model A embedded in Model 2 is displayed in Fig-
ure 15 for the four processing scenarios outlined earlier about the same proportion. This occurs because the re-
in the sections on juxtaposition contacts and pore pres- flection coefficient is proportional to the compliance for
sure difference across the fault: 20% noise (upper-left a relatively weakly slipping interface (see Appendix B,
panel), AGC (lower-left panel), no DMO (upper-right equation (B6)). For the the smallest compliance, Slip-
panel), and AGC and no DMO (lower-right panel). The Model D, the magnitude of the reflection is on the same
fault-plane extends over the long horizontal line at the order as that of the reflection caused by the P present
bottom of the plots in Figure 15. The shorter horizon- in Model 2 (shown in the upper-left plot of Figure 13).
tal line signifies the part of the fault plane that slips. Hence, this would seem to be a lower limit for compli-
Note that the location of the maximum reflectivity oc- ance for slipping behavior to dominate any P across
curs in the part of the fault that is slipping. Also, since the fault.
the reflection from Slip-Model A, embedded in Model 2,
is relatively large compared to the fault-plane reflection
for Model 2 alone (as shown in the upper-left plot of Fig- 11 RELATING LINEAR-SLIP TO A
ure 13 - an amplitude of 0.14 versus 0.04), the sources PRESSURIZED FAULT
of noise and error in the processing do not degrade the
amplitude map severely in any of the scenarios. This The values for the compliances used in the numerical
highlights the efficiency of Slip-Model A in reflecting modeling and shown in Table 5 do not, up to this point,
energy compared to the purely P contrast of Model 2. have any physical meaning in terms of the pressure lo-
The amplitude maps for Slip-Models A, B, C, and cally at the fault. In this section, we relate a slipping
D are shown in Figure 16. As the normal and shear interface to an effective-layer model that demonstrates
compliances are decreased in going from Slip-Model A much of the same wave-scattering behavior. With an
to Slip-Model D, the amplitude of the fault-plane reflec- effective layer described in terms of its thickness, den-
tion attributable to a slipping interface redcues in the sity, and velocity, the empirical relationships between
238 M. Haney, R. Snieder & J-P. Ampuero
Table 5. Four different slip interfaces described in terms of their normal and tangential compliances. These have been used in
the SEM modeling.
effective stress and density, and effective stress and ve- fraction of a wavelength fits in the layer. Inserting the
locity, derived in the introduction can give the compli- Taylor series approximation into equation (20) gives
ance values a physical meaning. The derivation here is 2ih
for normally incident P -waves, but can be extended to Rtot RP P . (23)
L
P -waves incident at an angle. We focus on normally in-
cident P -waves since the seismic imaging discussed in This expression shows that the total reflection from
previous sections utilized P P -scattered waves at small a thin weak layer is proportional to the derivative of
incidence angles. the incident wave. Widess has discussed this fact in
The effective-layer model begins from a weak scat- a famous paper on vertical seismic resolution (Widess,
tering assumption. For a thin layer, if the interface re- 1973).
flection coefficients at the upper and lower boundary are From equation (B6) in Appendix B, the P P reflec-
small, then the entire series of reverberations (Aki & tion coefficient for a weakly-slipping interface between
Richards, 1980) within the layer can be neglected. The two identical media (the host medium) is
total reflection coefficient from the thin layer can thus s iN
RP P , (24)
be approximated simply as the sum of the reflections 2
off the upper and lower interface. For the case when the where is the frequency, N is the normal compliance,
thin layer is sandwiched between two identical media, is the density of the host medium, is the P -wave
2ih
velocity of the host medium, and the superscript s indi-
Rtot RP P RP P exp , (20) cates that this is the reflection coefficient for a slipping
L
interface. This equation comes with its own assump-
where RP P is the P -wave to P -wave (P P ) reflection tion, namely that the dimensionless normal compliance
coefficient at the upper interface (the reflection at the is much smaller than 1, N 1. Equating this ex-
lower interface is RP P ), is the frequency, h is the pression to equation (23) gives
thickness of the thin layer, and L is the P -wave veloc-
iN 2ih
ity in the thin layer. In equation (20), we have assumed = RP P . (25)
that the impedance difference between the thin layer 2 L
and the host medium is small enough that the trans- By canceling common factors and solving this for N ,
mission coefficients in moving from the host medium the normal compliance, we get
into the thin layer and vice versa are close to 1. This 4h
is consistent with the weak-scattering assumption made N = RP P . (26)
L
above.
In the weak scattering approximation, we can substi-
The next approximation relies on the layer being
tute a weak-scattering approximation for the interface
sufficiently thin. If, for the argument of the exponential
reflection coefficient RP P . This can be obtained from
term in equation (20),
equation (B6) in Appendix B
2h
1, (21)
1
L RP P = + , (27)
2
then the exponential can be expanded to first order in
a Taylor series where = L , = (L + )/2, = L , and
= (L + )/2. Rewriting equation (27) in terms of the
2ih 2ih properties of the layer and host medium,
exp 1+ . (22)
L L L L
RP P = + . (28)
Note that the condition in equation (21) states that L + L +
1 4h/L , where L is the wavelength of the wave in Substituting into equation (26) for RP P yields
the thin layer. Hence, the condition means that only a
4h L L
N = + . (29)
L + L + L
Fluid pressure and fault reflectivity 239
6 6
amplitude
amplitude
a locally pressurized fault, since velocity and density 0 0
amplitude
amplitude
h
N = . (30) 2
5
L 2L
4
Equation (30) is somewhat different than equation (29), 10
800 900 1000 1100 1200 800 900 1000 1100 1200
most notably in that equation (30) contains material depth (m) depth (m)
properties of the layer only and not those of the host Figure 17. A test of the effective layer for a linear-slip
medium. The source of the difference between these fracture for a normally-incident P -wave. These traces are
two expressions is that the result of Schoenberg (1980) computed with a finite-element implementation of normally-
holds as h 0 and L 2L 0; that is, it is a strong- incident P -waves. The upper-left panel shows a transmitted
scattering approximation. The approximation that we plane wave (traveling to the right) and a reflected plane wave
(traveling to the left) from a horizontal thin bed of 10-m
have made in equation (29) is a weak-scattering one.
thickness with density and velocity values from Table 9. The
A weak-scattering approximation should be more ap- background medium has alpha = 2675 m/s and = 2260
propriate for an overpressurized fault since, according kg/m3 (the average values of the hydrostatic sand/shale lay-
to the pore pressure relationships described in the in- ers shown in Table 1). The upper-right plot shows the same
troduction, L 2L 0 cannot happen for any value of picture for the equivalent linear-slip interface with compli-
the pore pressure. Schoenbergs analogous conditions for ance calculated from equation (29). In the lower-left panel,
shear waves may be more realistic, since the shear veloc- a zoom in of the reflected waves for the thin bed (solid)
ity goes to nearly zero at zero effective stress (Zimmer and the equivalent linear-slip interface (dashed) are shown.
et al., 2002). The reflection from the equivalent linear-slip interface, equa-
Using a finite-element implementation of normally- tion (29), is in error by 15%. The lower-right panel shows
the comparison between the reflection from the thin bed
incident P -waves, we can test the above approximation
(solid) and the equivalent linear-slip interface (dashed) from
by comparing the reflection from a weakly slipping frac- Schoenberg (1980). The reflection from Schoenbergs equiv-
ture with the reflection from its effective weak thin layer. alent linear-slip interface, equation (30), is in error by 135%
We use a 10-m thick layer with L = 2077 m/s and L likely because it is a strong scattering approximation.
= 2124 kg/m3 in a background medium of = 2675
m/s and = 2260 kg/m3 . The layer thickness, L , and
L are taken from the results in Table 9 described be- equivalently the lithostatic stress v ) so that the effec-
low. The background medium parameters are the aver- tive stress varies only with pore pressure (d = v p),
age values of the sand/shale sequence in Table 1. The the compliance of a fracture can be put in terms of the
normal compliance N predicted by our weak-scattering thickness of the layer and the pore pressure
model, equation (29), for these parameters is 5 1010 4h
uL (P ) uL (P )
m Pa1 . For the Schoenberg model, equation (30), the N (h, P ) = u u + u .(31)
L (P ) + L (P ) + L (P )
normal compliance comes out to be 1.1 109 m Pa1 .
The source frequency in the simulation is peaked at 20 Note that we use the unloading relationships for L and
Hz, just as it was in the SEM modeling shown earlier L . This is because a locally pressurized fault likely re-
in this chapter. The result of the finite-element simula- sults from fluid migrating up the fault. Hence, the cor-
tion is shown by the traces in Figure 17. We find that rect curve to use is the unloading curve (Revil & Cath-
our weak-scattering effective-layer model is in much bet- les, 2002). This curve requires another parameter, the
ter agreement than Schoenbergs model (Schoenberg, maximum past effective stress max that the fault rock
1980). Our model is in error by 15% while the Schoen- experienced before being unloaded to its current state.
berg model is in error by 136%. This supports the weak- With this information, the slip models used in the pre-
scattering model for equivalent thin layers and weakly vious modeling section can be related to an effective
slipping interfaces. layer described by four parameters: fault zone of thick-
Using the effective-stress relationships we estimated ness h, fixed depth z, excess fluid pressure P Ph , and
at South Eugene Island for L and L in the case of un- maximum past effective stress max . Figure 18 shows
loading, uL (d ) and uL (d ), and fixing the depth (or, how, given a maximum past effective stress max and
fault zone thickness h, one can construct equivalent
240 M. Haney, R. Snieder & J-P. Ampuero
compliances for the velocity and density values traced
out along the unloading curve as a function of effective Slip-Model (kg/m3 ) vp (m/s) e (psi) p (psi)
stress. To convert from effective stress to pore pressure, A 2204 2019 11 5490
knowledge of the depth z or lithostatic level V in nec- B 2207 2175 88 5412
essary. C 2216 2293 322 5178
From this weak equivalent layer, we find that the D 2269 2472 1655 3845
pore-pressure value corresponding to a linear-slip inter-
face is not unique - it depends on three other parame- Table 6. Four different slip-interfaces described in terms of
ters besides the pressure. In Tables 6 through 9, we fix their effective layer parameters assuming a thickness of 10
two of these parameters - the depth of the fault zone m. The estimates are made under the assumption that the
(1850 m) and the thickness of the fault zone (10 m). fault rock began its unloading path after reaching a maxi-
mum vertical effective stress of 2800 psi. velocity and pore
The tables show the values of the effective thin layer
pressure derived in the introduction are used to relate these
for each Slip-Model, A through D, while varying max compliances to vertical effective stress e and pore pressure
between the different tables: max = 2800 psi in Ta- p at South Eugene Island. Also, for the four tables, the pore-
ble 6, max = 2400 psi in Table 7, max = 2000 psi pressure estimate assumes a depth of 1850 m, where the over-
in Table 8, and max = 1600 psi in Table 9. They also burden stress is 5500 psi and the hydrostatic pressure is 2800
show the effective stress corresponding to those values psi.
of velocity and density along the unloading path, and
use the depth to convert to pore pressure. We do not
show results for the variation with thickness h since it
is mundane, as seen in equation (29). The variation with Slip-Model (kg/m3 ) vp (m/s) e (psi) p (psi)
depth z is also fairly unimportant since it changes only A 2178 2037 31 5469
the value of the pore pressure for a given effective stress. B 2186 2192 238 5262
As seen in Tables 6 through 9, certain values of the com- C 2208 2301 775 4725
pliance cannot even exist for some values of max since
the compliance does not fall in the range of possible Table 7. Four different slip-interfaces described in terms of
compliances based on the values of density and velocity. their effective layer parameters assuming a thickness of 10 m.
This is shown in Figure 18, where in the bottom panel The estimates are made under the assumption that the fault
the possible compliance values corresponding to the un- rock began its unloading path after reaching a maximum
loading path in the upper figures does not extend all vertical effective stress of 2400 psi. There is no effective-layer
the way to zero. Hence, some small compliance values model for Slip-Model D along this unloading path.
cannot be modeled for those choices of h, z, and max .
In order to get smaller compliance values, the thickness
of the fault would need to be made smaller because the 12 CONCLUSION
compliance, from equation (29), scales with h. From the We have presented a complete numerical modeling ex-
Tables 6 through 9, we find that the earlier that the rock periment by utilizing an SEM implementation of the 2D
in the fault zone began to be unloaded, i.e., the smaller elastic wave equation and processing the resulting wave-
the max , the larger the compliance can be for smaller forms into their time-migrated images. We derived a
amounts of pore pressure. Hence, the more efficient the simple dip filter and used it to isolate fault-plane reflec-
pressurized fault zone can be in scattering seismic waves tions. We then exploited the relationships between the
for a small amount of pore pressure increase. elastic parameters, density and velocity, to create phys-
In the A10ST well at South Eugene Island, the ef- ically meaningful models of sealing faults that maintain
fective stress in the B-fault zone was measured at as a P of up to 600 psi. For these P models, we as-
low as 166 psi (Losh et al., 1999). In fact, Losh et al.
(1999) state that ... the fault itself ... is at significantly
lower vertical effective stress than the downthrown sed- Slip-Model (kg/m3 ) vp (m/s) e (psi) p (psi)
iments. In other words, the fault itself is overpres-
A 2149 2058 109 5391
sured. Hence, some of the low values for effective stress
B 2172 2203 700 4800
shown in Tables 6 through 9, while unusual, are entirely
C 2218 2292 1829 3671
possible for the growth faults at South Eugene Island.
Anomalously low effective stresses of 575 psi and 807
Table 8. Four different slip-interfaces described in terms of
psi were also measured in the nearby A20ST well as it their effective layer parameters assuming a thickness of 10 m.
crossed the A-fault system (Losh et al., 1999). The estimates are made under the assumption that the fault
rock began its unloading path after reaching a maximum
vertical effective stress of 2000 psi. There is no effective-layer
model for Slip-Model D along this unloading path.
Fluid pressure and fault reflectivity 241
sumed that the overpressure mechanism is purely due
Slip-Model (kg/m3 ) vp (m/s) e (psi) p (psi) to undercompaction. In the course of this modeling, we
A 2124 2077 453 5047 found that the minimum P necessary to give rise to
substantial fault-plane reflections is on the order of 150
Table 9. Four different slip-interfaces described in terms of psi. Taking advantage of the SEM modeling codes abil-
their effective layer parameters assuming a thickness of 10 m. ity to accommodate linear-slip interfaces, we selected
The estimates are made under the assumption that the fault four different values of the normal and shear compli-
rock began its unloading path after reaching a maximum ver- ances for the fault interface. We find that the slipping
tical effective stress of 1600 psi. There are no effective layer interfaces are more efficient at scattering seismic energy
models for Slip-Models B, C, and D along this unloading than are the pore pressure differences across the fault
path. for compliance values above 1010 m Pa1 . Unsat-
isfied with the lack of physical insight into the mean-
ing of a compliance, we derived, from a weak-scattering
model, an equivalent thin, weak layer that gives virtu-
2400
ally the same reflection as a linear-slip interface under
certain conditions. We used this equivalent layer model
density (kg/m3)
ACKNOWLEDGMENTS
2000
Jon Sheiman provided valuable input and thanks go out
to him. Andrey Bakulin first brought the idea of a weak
scattering model of a fracture to the authors attention.
1500
0 1000 2000 3000 4000 Shell International Exploration and Production has pro-
effective stress (psi)
x 10
10 vided funding for this research.
REFERENCES
compliance (m Pa1)
where h is the grid spacing. After applying the weak formulation to the 1D wave equation and integrating in space,
the resulting matrix equation to be solved in time is
= c2 K~
M ~u u, (A2)
2
where M and K are the mass and stiffness matrices, c = / is the squared wavespeed (assumed to be constant
here), and u is the (out of plane) displacement field (u = uy ). For the linear basis functions used here, the entries of
these matrices can be calculated exactly. For the (consistent) mass matrix, MJ,K , the non-zero entries are
ZL
K J dz = 2h/3 for J = K (A3)
0
ZL
K J dz = h/6 for J = K + 1, J = K 1. (A4)
0
where h is the spatial grid spacing (assumed to be constant). This means that M is symmetric and tridiagonal.
Similarly, the stiffness matrix, SJ,K , is also symmetric tridiagonal and the non-zero entries are
ZL
K J
dz = 2/h for J = K (A5)
z z
0
ZL
K J
dz = 1/h for J = K + 1, J = K 1. (A6)
z z
0
For a single split node, the matrices M and K have some structure. In fact, the left hand side of equation (A2)
is
.. ..
2 3
2 3
. . ..
6
6 .
7 .
6 ..
76 7
6 2h/3 h/6 76
76 um2 7
7
h/6 2h/3 h/6 um1
6 76 7
6 76 7
h/6 h/3 0 um
6 76 7
(A7)
6 76 7
0 h/3 h/6 um+1
6 76 7
6 76 7
h/6 2h/3 h/6 um+2
6 76 7
6 76 7
.. 7 6 um+3
6 76 7
6 7
6 h/6 2h/3 . 74 ..
5
4 5
.. .. .
. .
244 M. Haney, R. Snieder & J-P. Ampuero
where the spatial location on the grid is indicated by the subscripts m 1, m, etc. In equation (A7), the split node
is located between the gridpoints m and m + 1. Similarly, the right hand side of equation (A2) is
.. ..
2 3
2 3
. . ..
6
6 .
7 .
6 ..
76 7
6 2/h 1/h 76 u
7 6 m2 7
7
1/h 2/h 1/h 7 6 um1 7
6 76 7
6
1/h 1/h 0 7 6 um 7
6 76 7
c2 6 (A8)
6
0 1/h 1/h 7 6 um+1 7
76 7
6
1/h 2/h 1/h 7 6 um+2 7
6 76 7
6
. . 7 6 um+3 7
6 76 7
6
6 1/h 2/h . 74 ..
5
4 5
.. .. .
. .
We use the expressions in equations (A7) and (A8) to find the stability criterion for a split node. To further
simplify the calculation, we lump the mass matrix shown in equation (A7) so that the left hand side of equation (A2)
is instead
.. ..
2 3
2 3
. . ..
6
6 .
7 .
6 .. h 0
76 7
7 6 u 7
6 7 6 m2 7
0 h 0 7 6 um1 7
6 76 7
6
0 h/2 0 7 6 um 7
6 76 7
(A9)
6
0 h/2 0 7 6 um+1 7
6 76 7
6
0 h 0 7 6 um+2 7
6 76 7
6
. . 7 6 um+3 7
6 76 7
6
6 0 h . 74 ..
5
4 5
.. .. .
. .
The lumping is accomplished by summing the rows and placing the result on the main diagonal. This essentially
reduces the finite-element scheme into a finite-difference scheme. The difference is that care has been taken with
the inclusion of a split-node and, as seen in the next paragraph, the boundary condition at the split-node can be
accounted for in an entirely straightforward manner. This is in contrast to the fictitious lines in Alterman and
Lowenthal (1970) and Kelly et al. (1976).
Setting the m-th rows of the matrix/vector multiplications depicted in equations (A8) and (A9) equal to each
other, in accordance with the equation of motion, equation (A2), gives the spatially discretized equation of motion
at the split node
h c2 c2
um = um1 um . (A10)
2 h h
Note that there is currently no communication across the split node, between the gridpoint at m and at m + 1.
Interaction between these gridpoints is accomplished by including the traction yz between the nodes with the slip
law
um+1 um = yz , (A11)
where the coupling is described by the free parameter (the compliance). Note that = 0 is a welded contact and
= is a stress-free boundary. Using the slip law and Hookes Law u/z = yz , the traction at the split node
can be included into equation (A10)
h c2 c2 c2
um = um1 um + (um+1 um ). (A12)
2 h h
A time integration scheme must be chosen for equation (A12). An explicit scheme is obtained by evaluating all
the u-terms on the right hand side at the previous timestep, denoted n. For notational purposes, we show u m1
evaluated at timestep n by um1,n . For the left hand side, we choose the second difference approximation to um
centered at timestep n
um,n+1 2um,n + um,n1
um = , (A13)
t2
Fluid pressure and fault reflectivity 245
where t is the time spacing. The algorithm outlined here is to be solved for um,n+1 in equation (A13) by knowing
the displacements at timesteps n and n 1. Such a multistep algorithm can be seen as equivalent to a single-step
algorithm with prediction.
Inserting equation (A13) into equation (A12) results in the complete finite-element stencil
h h h c2 c2 c2
2
um,n+1 2
um,n + 2
um,n1 = um1,n um,n + (um+1,n um,n ). (A14)
2t t 2t h h
To investigate the stability of this scheme, we study the exponential growth or decay of the odd eigenmodes of u.
These eigenmodes are antisymmetric with respect to the slipping interface. We chose the odd eigenmodes since the
symmetric modes have no slip. If there wasnt a slipping interface and the medium was homogeneous, harmonic
functions would be inserted for both the spatial and temporal variation of u; for the case of slipping interface, care
has to be taken in treating the spatial variation since the medium is not homogeneous. This extra care amounts to
determining an additional free parameter at the interface, as demonstrated shortly.
An odd eigenmode for a single linear-slip interface located at the depth zs is
u(z, t) = U (z)eit
U (z) = sgn(z zs ) sin(k|z zs | + ), (A15)
where U (z) is the spatial variation of the odd eigenmode, is a free parameter determined from the linear-slip
boundary condition of equation (A11), k is the wavenumber, and is the frequency. As stated above, the slip
interface is between the m and m + 1 gridpoints, which are both located at the depth zs . To be precise, the m
gridpoint is obtained in the limit z zs for values z < zs . Similarly, the m + 1 gridpoint is obtained in the limit
z zs for values z > zs . Mathematically, this is the difference between a left-hand and right-hand limit. It matters
in the context of equation (A15) since the signum function is discontinuous at z = zs .
Taking the proper limits, the spatial components of an odd eigenmode at the m and m + 1 gridpoints are
Um = sin ,
Um+1 = sin . (A16)
From these equations, it can be seen that drives the magnitude of the slip discontinuity. We now use the relationships
in equation (A16) to determine from the linear-slip boundary condition of equation (A11). For this, we need the
shear stress component yz , which by Hookes Law is equal to u/z. In other words, we need the value of the
derivative of u at zs . From equation (A15)
U
= k cos(k|z zs | + ). (A17)
z
Therefore, at zs , yz = k cos . Inserting this expression for yz and the relationships in equation (A16) into the
boundary condition of equation (A11) yields
k
tan = . (A18)
2
From this equation, sin and cos can be found
k
sin = p
2 2 k 2 + 4
2
cos = p . (A19)
2 k 2 + 4
2
Since it does not matter, we take the postive sign in equation (A19).
We are now in a position to write down the discrete forms of the antisymmetric eigenmodes at the gridpoints
m 1, m, and m + 1 using equation (A15)
um+1,n = sin eint
um,n = sin eint
um1,n = sin(kh + ) = (sin kh cos + cos kh sin ) eint , (A20)
where the time t is discretized by timestep n and time spacing t. Using equation (A19) and the substitution
n = eint , (A21)
246 M. Haney, R. Snieder & J-P. Ampuero
equation (A20) can be written finally as
k
um+1,n = p n
2 k 2 + 4
2
k
um,n = p n
2 2 k 2 + 4
2 sin kh + k cos kh n
um1,n = p . (A22)
2 2 k 2 + 4
In equation (A22), the variation with timestep n is trivial, requiring only a change in the exponent of . Taking the
information in equation (A20), using it in equation (A14), and organizing terms yields
hk 2 c2
0 = + (k 2 sin kh k cos kh)
2t2 h
hk 2 hk n1
+ 2c k + . (A23)
t2 2t2
In order to have non-trivial solutions, the terms inside the brackets in equation (A23) must equal zero.
Dividing equation (A23) by the coefficient of the 2 term and setting the terms inside the brackets to zero gives
t2 c2 2t2 c2
0 = 2 2[1 2
(k k cos kh 2 sin kh) ] + 1. (A24)
h k h
For the numerical scheme to be stable, the two roots of equation (A24) should have magnitudes less than or equal
to 1. This means that solutions are not exponentially growing in time. Lets find the conditions such that this is the
case.
Note that equation (A24) has the general form
0 = 2 v + 1, (A25)
where the coefficient v is a complex number. Denote the two (possibly complex) roots of equation (A25) 1 and 2 .
From the quadratic formula, it is known that for any v
1 + 2 = v
1 2 = 1. (A26)
These equalities are a particular example of what are known as Vietes formulas and are analogous to the trace of a
matrix being equal to the sum of its eigenvalues and the determinant being equal to the product of its eigenvalues.
With help of the triangle inequality, the two equalities in equation (A26) can be recast in terms of the magnitudes of
1 and 2 , shown as |1 | and |2 |. For the first equality of equation (A26)
|1 | + |2 | |1 + 2 | = |v|. (A27)
And, for the second equality
|1 | |2 | = |1 2 | = |1| = 1. (A28)
This is necessary since it is the magnitude of the roots of equation (A24) that determine the stability of the algorithm.
Since we need both |1 | 1 and |2 | 1 for stability, equation (A28) tells us that
|1 | = 1 and |2 | = 1. (A29)
This makes sense because we are solving a wave equation whose solutions should not decay with time. Using this
information together with equation (A27) gives
2 = |1 | + |2 | |v| (A30)
or in other words,
v
| | 1. (A31)
2
Applying this stability criterion, valid for the general quadratic form of equation (A25), to the expression de-
scribing the roots in equation (A24) gives
t2 c2 2t2 c2
|1 (k k cos kh 2 sin kh) | 1. (A32)
h2 k h
Fluid pressure and fault reflectivity 247
From this expression, note that in the limit of low frequencies kh 0 (which corresponds to waves that are well
sampled by the grid), this stability criterion is always satisfied. Typically we look for instabilities to occur for waves
that are near the spatial Nyquist frequency (kh = ). For this frequency, equation (A32) becomes
2t2 c2 2t2 c2
|1 2
| 1. (A33)
h h
The term inside the absolute value on the left hand side of equation (A33) is real-valued; therefore, the inequality
may be rewritten as two simultaneous inequalities
2t2 c2 2t2 c2
1 1 2
1. (A34)
h h
Since the terms being subtracted from 1 in the middle of equation (A34) are both positive, the right hand inequality
is always satisfied and may be ignored. The other inequality is meaningful
2t2 c2 2t2 c2
1 1
h2 h
2t2 c2 2t2 c2
2
+ 2
h h
t2 c2
h
1+ 1
h2
s
ct h
1+ 1
h
h
ct p (A35)
1 + (h/)
This stability criterion has the same general form as that for a regular node that is not split (ct h). However, the
compliance of the split node does enter into the stability criterion uniquely in the inverse square root appearing on
the right hand side. For a free surface, = and the factor with the compliance vanishes and the stability criterion
can be met. As the compliance approaches zero, though, the denominator becomes large and the stability criterion
becomes harder to satisfy. For given values of c, h, , and t, there is a lower value for compliance min past which
the algorithm becomes unstable at the split node. This precludes the possibility of a welded interface ( = 0), though
for certain values of c, h, , and t, min may be so small that an effectively welded interface can be realized.
20 40 60 80
Figure A1. The linearized reflection coefficient as a function of angle (thick, dashed line) and the exact reflection coefficient
for a welded interface (kT , kN = 0).
Table A1. The earth model used to compute the reflection coefficients.
where now two compliances, the normal (kN ) and shear (kT ) compliance, describe the linear-slip interface. The
linearized P P -reflection coefficient at a planar interface between two isotropic halfspaces that results is:
" 2 # 2
1 1
RP P 2 sin2 + sec2 4 sin2
2 2
" 2 #
1
+i 2 sin2 sec kN
2
3
2i cos sin2 kT , (B6)
where , , and are the changes in P -velocity, S-velocity, and density of the two halfspaces, , , and
are the average P -velocity, S-velocity, and density of the two halfspaces, is the angular frequency of the incident
wave, kT and kN are the shear and normal compliances of the interface, and is the reflection/incidence angle. The
approximation in equation (B6) holds for small relative changes in the medium parameters (/, /, / 1),
small dimensionless shear compliance (kT 1), small dimensionless normal compliance (kN sec 1), and
small angles of incidence. Note that the condition for a small dimensionless normal compliance has an angular
dependence. For a welded interface (kT , kN = 0), the imaginary terms in equation (B6) vanish and the linearized
P P -reflection coefficient agrees with equation 5.44 of Aki and Richerds (1980). Again, we omit the lengthy derivation
of equation (B6), but it follows from Wang (1999).
In Figures 1 through 3, we compare the linearized P P -reflection coefficient to the exact P P -reflection coefficient
for the parameters shown in Table II while varying the amount of slip (the value of the compliance) at the interface.
Since the reflection coefficient for an interface with slip is frequency dependent, we set the frequency of the incident
wave to 10 Hz for all the examples. The elastic contrasts resulting from the parameters listed in Table II are suffi-
ciently small to expect agreement between the linearized approximation and the exact reflection coefficient at small
incidence/reflection angles. The formulas for the exact P P -reflection coefficient, including possible slip, can be found
in Chaisri and Krebes (2000).
Shown in Figure (A1) is the magnitude of the exact P P -reflection coefficient as a function of angle for a welded
interface (with elastic contrasts from Table II) and its linearized approximation. Figure (A1) is the typical case
250 M. Haney, R. Snieder & J-P. Ampuero
PP reflection coefficient magnitude
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
20 40 60 80
Figure A2. The linearized reflection coefficient as a function of angle (thick, dashed line) for a welded interface (k T , kN = 0)
and the exact reflection coefficient for an interface with the same elastic contrasts, but with some slip (k T = 5.0 1010 m/Pa
and kN = 2.5 1010 m/Pa). By not including slip in linearized reflection coefficient, it underestimates the magnitude of the
exact reflection coefficient.
20 40 60 80
Figure A3. The linearized reflection coefficient as a function of angle (thick, dashed line) for an interface with the elastic
contrasts from Table I and having linear-slip behavior (kT = 5.0 1010 m/Pa and kN = 2.5 1010 m/Pa). The expression
for the linearized reflection coefficient including slip is taken from equation (B6). The exact reflection coefficient for the same
interface is also shown. By including slip in the linearized reflection coefficient, the approximation agrees well with the exact
solution up to about 45 .
assumed in AVO studies, that of a welded interface, and the approximation is seen to be in good agreement for angles
out to about 50 . In Figure (A2), we plot the same linearized reflection coefficient as in Figure (A1), assuming a
welded interface; however, for the exact reflection coefficient, we include a small amount of slip by using non-zero
values for the compliances: kT = 5.0 1010 m/Pa and kN = 2.5 1010 m/Pa. By small, we mean that the
dimensionless compliances, kT and kN sec, are much less than one. Without the contribution of slip at the
interface, the linearized reflection coefficient underestimates the exact reflection coefficient.
Using equation (B6), the magnitude of the linearized reflection coefficient for an interface with slip can be
compared to the exact reflection coefficient. The comparison is shown in Figure (A3). The inclusion of slip in the
expression for the linearized reflection coefficient has resulted in better agreement with the exact reflection coefficient
(good agreement up to 45 ). Qualitatively, from looking at Figure (A2), the presence of non-weldedness at an interface
causes the AVO intercept to increase (magnitude at zero-offset is greater) and the AVO gradient to increase (the
curve steepens with increasing angle) from their values for the case of no slip. These linearized reflection coefficients
for an interface with slip give some insight into the influence of the various elastic contrasts and compliances that is
unavailable from the complicated exact expressions. Also, the linearized formulas should be useful in a least-squares,
linear inversion of AVO for rock and fracture properties.
CWP-515
ABSTRACT
Figure 4. Exact NMO ellipses for an effective orthorhombic Figure 5. Exact NMO ellipses for two vertical, orthogonal
model formed by two vertical, orthogonal sets of rotationally sets of microcorrugated fractures. The strike azimuth of the
invariant fractures. The semi-major axes (black arrows) of dominant fracture set is 90 . The parameters are VP b =2
the P- and S1 -wave NMO ellipses are parallel to the strike of km/s, VSb =1 km/s, N1 = 0.25, V1 = H1 = 0.12,
the dominant fractured set (azimuth 90 ). The parameters N V1 = N H1 = 0.17, and V H1 = 0.12. Each weak-
are VP b = 2 km/s, VSb = 1 km/s, N1 = 0.25, and V1 = ness of the second fracture set is equal to one-third of the
H1 = 0.12. Each weakness of the second fracture set is corresponding weakness of the first set.
equal to one-third of the corresponding weakness of the first
set. The radius of the external circle corresponds to 2 km/s.
(km /s)
(km /s)
(deg.)
(deg.)
0.1 2 70 0.1 2 20
S0
S0
V
V
0.05 1.9 68 0.05 1.9 18
0 66 0 16
1.8 1.8
0.1 0.2 0.3 0.4 3.6 3.8 4 4.2 4.4 0.8 0.9 1 1.1 1.2 0.1 0.2 0.3 0.4 3.6 3.8 4 4.2 4.4 0.8 0.9 1 1.1 1.2
V (km /s) z (km) VP0 (km /s) z (km)
P0
Fig. 2: Inverted parameters (dots) of a horizontal TTI layer Fig. 3: Inversion results for a model with the same parameters
obtained from 2D PP and PS data in the symmetry-axis plane. as those in Figure 2 except for the tilt = 20 .
The correct model parameters (VP 0 = 4 km/s, VS0 = 2 km/s,
= 0.25, = 0.1, = 70 , z = 1 km) are marked by the
crosses. The dashed line on the [, ] plot corresponds to the
correct value of the difference (). The input data were con- velocity variations or non-horizontal reflectors (e.g.,
taminated by Gaussian noise with the standard deviations of Tsvankin and Grechka, 2000). In active tectonic areas,
2% for the NMO velocities, 0.5% for the zero-offset traveltimes, originally horizontal TI layers are often rotated in such
and 2% for the PS-wave asymmetry attributes. a way that the symmetry axis remains orthogonal to
the layer boundaries. Hence, here we assume that the
tilt of the symmetry axis is equal to the dip of the
SS arrivals can be used to estimate the stacking velocity reflector. The approximate asymmetry attributes in the
(Vnmo,S ) and zero-offset traveltime (tS0 ) of the SS-waves symmetry-axis (dip) plane are convenient to express
that are not physically excited in the survey. Here, we through the SS-wave offset xSS . The leading terms in
supplement the pure-mode signatures in parameter esti- xSS , however, depend just on parameters that can be
mation with the PS-wave asymmetry attributes obtained obtained from the pure (PP and SS) modes, such as
from the PP+PS=SS method to form the data vector d the anisotropic coefficient ( )/(1 + 2), where
[Vnmo,P , tP 0 , Vnmo,S , tS0 , tP S (p), x0 ], which is governed VP20 /VS02
( ) (Grechka and Dewangan, 2003).
by the model vector m [VP 0 , VS0 , , , , z]. To esti- Therefore, we have to rely on long-offset PS data in
mate the vector m, we applied nonlinear inversion (the constraining the medium parameters.
Gauss-Newton method) based on exact equations for all
components of the data vector.
Parameter estimation
Figure 2 displays the results of inverting noise-
contaminated input data for a model with a large tilt . Processing the pure-mode reflections in the symmetry-
Such models can be considered typical for dipping frac- axis plane yields their NMO velocities, zero-offset
ture sets because the fracture planes seldom deviate far traveltimes and time slopes (horizontal slownesses p)
from the vertical. The inversion results are unbiased, and on the zero-offset section. The data vector (not includ-
the noise is not amplified by the parameter-estimation ing the asymmetry attributes) then has six elements
procedure. The standard deviations are close to 0.02 for (Vnmo,P , tP 0 , pP 0 , Vnmo,S , tS0 , and pS0 ) but only five
and , 1% for VP 0 , 2% for VS0 and z, and 1 for . The of them are independent because of the constraint
maximum offset-to-depth ratio for PP and PS data in all pS0 /pP 0 = tS0 /tP 0 . They are governed by the same
the tests is close to two. The best-constrained parameter model vector as that for a horizontal layer, with the tilt
combination is the difference ( ), which controls the () equal to the reflector dip and the depth z replaced by
asymmetry in the slowness domain [equation (5)] and has the normal distance zd between the common midpoint
a strong influence on the NMO velocity of the constructed and the reflector. Since the model vector includes six
SS-waves. It should be emphasized that the parameter es- parameters, the 2D inversion cannot be carried out
timation is feasible only if the asymmetry information of without additional information, such as the asymmetry
the PS-wave is included in the inversion algorithm. attributes of the PS-wave.
Models with mild tilts are not plausible if the anisotropy We propose the following algorithm to invert the 2D mul-
is caused by dipping fractures, but they may be ade- ticomponent data in the symmetry-axis plane. For each
quate for effective TTI media formed by progradational value of the tilt from 0 to 90 , we find the model vec-
sequences. For a tilt of 20 (Figure 3), the scatter in the tor from the pure-mode data. The range of plausible tilts
inversion results is slightly higher than that for large tilts, can be restricted by putting reasonable constraints on the
but the standard deviations are less than 0.03 for and parameter (0 1). Taking into account errors in
, 3% for VP 0 , VS0 and z, and 2 for (Figure 3). As ex- the measured quantities, we compute the range of models
pected, the parameter estimation breaks down when the that fit the pure-mode data within the noise level for each
model approaches VTI, and the tilt becomes smaller plausible value of tilt. Then, for all models obtained at
than 10 . the previous step, we compute the asymmetry attributes
tP S and xP S from the exact equations (3) and (4) and
minimize the misfit function for tP S and xP S over the
Model 2: Dipping TTI layer full range of offsets.
Another possible reason for the PS-wave moveout asym- The inversion for large tilts, which are typical for shale
metry is lateral heterogeneity in the form of lateral layers in fold-and-thrust belts, produces unbiased results
PS-wave moveout asymmetry
2.2 64
0.2 was applied to the inversion of multicomponent data
0.15 2.1 62 acquired over horizontal or dipping TI layers with a
(km /s)
tilted symmetry axis.
(deg.)
0.1 2 60
S0
The weak-anisotropy approximation helped to obtain con-
V
0.05 1.9 58
0
cise expressions for the asymmetry attributes of PSV-
0.05 0.15 0.25 0.35 0.45
1.8
3.8 3.9 4 4.1 4.2
56
0.8 0.9 1 1.1 1.2
waves in terms of the tilt of the symmetry axis and
VP0 (km /s) zd (km) Thomsens anisotropic parameters. We combined the fac-
Fig. 4: Inverted parameters (dots) of a dipping TTI layer with tors tP S and xP S with the NMO velocities, zero-offset
= 60 obtained from 2D PP and PS data in the symmetry- traveltimes and time slopes of PP-waves and the con-
axis plane. The correct model parameters are marked by the structed SS-waves in a nonlinear parameter-estimation al-
crosses. The dashed lines on the [, ], [VP 0 , VS0 ], and [ , zd ] gorithm. Although it is desirable to have a wide range of
plots correspond to the correct values of , VP 0 /VS0 , and source-receiver azimuths, the parameter estimation for a
sin /zd , respectively. The input data were contaminated by wide range of can be performed using just 2D data in
Gaussian noise with the same standard deviations as in the
caption of Figure 2.
the symmetry-axis plane. The orientation of this plane,
however, has to be determined beforehand from either
2.2 29 the pure-mode NMO ellipses (using wide-azimuth data)
0.2
2.1 27
or from the PS-wave polarization at small offsets.
0.15
(km /s)
0.1 2 25
0.05 1.9 23 if the symmetry axis deviates by at least 10 from the ver-
0
1.8 21 tical (VTI) and horizontal (HTI) directions. In contrast,
0.05 0.15 0.25 0.35 0.45 3.8 3.9 4 4.1 4.2 0.8 0.9 1 1.1 1.2
VP0 (km /s) zd (km) the parameters of a dipping TTI layer with the symmetry
axis orthogonal to the layers bottom are constrained only
Fig. 5: Inversion results for a model with the same parameters
as those in Figure 4 except for the tilt = 25 . The standard if > 25 , and the accuracy of the inverted parameter
deviations in the asymmetry attributes are increased to 6%. is marginal even for large (i.e., for steep dips).
References
(Figure 4). If we assume that the error in the asymme-
try attributes is 2%, the parameters VP 0 , VS0 , and zd
are well constrained (the standard deviation is less than Angerer, E., Horne, S. A., Gaiser, J. E., Walters, R.,
1%), and the standard deviation in is only 1 (Figure 4). Bagala, S. and Vetri, L., 2002, Characterization
However, the inverted anisotropic parameters and ex- of dipping fractures using PS mode-converted data:
hibit more scatter, with the standard deviation reaching 72th Ann. Internat. Mtg., Soc. of Expl. Geophys.,
0.06 and 0.04, respectively. Also, the estimates of and Expanded Abstracts, 10101013.
degrade rapidly as the error in the asymmetry attributes
increases, while the deviations in VP 0 , VS0 , and zd remain Grechka, V., and Dewangan, P., 2003, Generation and
small. It is clear from Figure 4 that the best-constrained processing of pseudo shear-wave data: Theory and
parameter combinations are , sin /zd , and VP 0 /VS0 . In case study: Geophysics, 68, 18071816.
principle, and for large tilts can be obtained with suf-
Grechka, V., and Tsvankin, I., 2002, PP+PS=SS: Geo-
ficient accuracy from wide-azimuth PP and SS data, as
physics, 67, 19611971.
demonstrated by Grechka et al. (2002).
If the tilt for the model from Figure 4 is reduced from Grechka, V., Pech, A., and Tsvankin, I., 2002, Multi-
60 to 25 (Figure 5), the uncertainty in tP S and xP S component stacking-velocity tomography for trans-
should increase because of the smaller magnitude of the versely isotropic media: Geophysics, 67, 15641574.
moveout asymmetry. If the error in the asymmetry at- Thomsen, L., 1999, Converted-wave reflection seismol-
tributes is set to 6%, the standard deviation of the tilt ogy over inhomogeneous, anisotropic media: Geo-
is almost the same (1 ) as that in Figure 4. Despite the physics, 64, 678690.
high resolution in , the standard deviations in VP 0 , VS0 ,
and zd increase to about 4%; the deviations in and Tsvankin, I., 2001, Seismic signatures and analysis of
are also substantial (0.08 and 0.05, respectively). The in- reflection data in anisotropic media: Elsevier Science
version breaks down for quasi-VTI models with mild tilts Publ. Co., Inc.
< 15 20 .
Tsvankin, I., and Grechka, V., 2000, Dip moveout of
converted waves and parameter estimation in trans-
Conclusions versely isotropic media: Geophys. Prosp., 48, 257
292.
We modified the PP+PS=SS method of Grechka and
Tsvankin (2002) to supplement the computed quasi-SS Tsvankin, I., and Grechka, V., 2002, 3D description
data with the time (tP S ) and offset (xP S ) asymmetry and inversion of reflection moveout of PS-waves in
attributes of the converted waves. The new algorithm anisotropic media: Geophys. Prosp., 50, 301316.
Main Menu
1800
subtract from eq. (8) the unperturbed eective wavenum-
ber kef f and divide by frequency , we obtain the eective
slowness perturbation
Fig. 1: Representation of a 2D random slowness model with a
Gaussian autocorrelation function. The correlation distance a
kef f
is 40 m. sef f = s0 . (9)
k0
where k0 =s0 is the wavenumber in the homogeneous Traveltime change of the Diuse Waveeld
reference medium, (r) is the normalized random veloc-
ity uctuation, and (r) is the waveeld propagating In order to localize changes in the scattering medium,
through the random medium. Eq. (3) can be formulated we derive an expression relating the mean traveltime
as a random integral equation (Frisch, 1965). We can ob- change of the diuse waveeld to the eective slowness
tain a solution to the mean eld radiated by a point perturbation in the medium as a function of position.
source in a random medium, which satises the equation With this expression we predict the temporal change
(Frisch, 1965) in the traveltimes of the multiply-scattered waves after
introducing a given slowness perturbation. Snieder et
(2 + kef f 2 )(r, w) = (r), (4) al. (2002) introduced coda wave interferometry whereby
multiply scattered waves are used to detect temporal
changes in the medium by using the scattering medium
kef f 2 = k0 2 k0 4 G0 (r, w)(r)d2 r, (5) as an interferometer. They found that for a small per-
A
turbation in the velocity, estimates of this perturbation
can be obtained from multiply scattered waves by a
where G0 (r, ) is the free-space Greens function and (r) time-windowed crosscorrelation. The assumption used
is the autocorrelation function of the velocity uctua- is that when the temporal change is weak, the inuence
tions. Eq. (4) is the Helmholtz equation with an eective of this change on the geometrical spreading and on the
wavenumber kef f . From Eq. (5) we can obtain the eec- scattering coecient can be ignored, and the dominant
tive slowness, i.e., the slowness with which the coherent change on the waveeld arises from the change in the
pulse propagates in the multiple-scattering medium. traveltime of the wave that travels along each path.
For small and homogeneous perturbations in the eective
Perturbations of the Eective Slowness slowness, the time-windowed crosscorrelation of the un-
perturbed and perturbed waveeld provides an estimate
We calculate the slowness eld as the inverse of the uc- of the time shift between the two waveforms. With the
tuating velocity eld v dened in eq. (1). This slowness traveltime change , we can estimate the eective slow-
eld represents the unperturbed slowness (see Figure 1). ness perturbation with the equation (Snieder et al., 2002)
Let us introduce a constant slowness perturbation in our
slowness eld, creating the perturbed slowness eld, sef f (t, T )
= , (10)
sef f t
sp = sunp + s0 , (6) where (t, T ) is the time shift of the time-windowed cross-
where sunp is the unperturbed slowness eld, sp is the per- correlation centered at time t with window size T . The
turbed eld, and s0 is the mean slowness perturbation. traveltime change is weighted by the energy of the wave
The eective slowness perturbation is proportional to the along each multiple-scattering path with total traveltime
mean slowness perturbation. The constant of proportion- t from the source to the receiver.
ality, which can be derived from the eective medium For localized slowness perturbations, eq. (10) is no longer
theory, depends on the specic parameters of the random valid and we need a new expression describing the trav-
slowness uctuations. eltime change of the multiply-scattered waves caused by
Examples
Having derived the relation between the mean traveltime
change of the diuse waveeld and the eective slowness
perturbation, we test the theory with nite-dierence
simulations of multiply scattered waves in a medium
with random velocity uctuations as dened in eq. (2)
before and after introducing a localized eective slowness 0 0.5 1 1.5 2
perturbation. t (s)
We consider a localized perturbation within an area in Fig. 3: Finite-dierence synthetic seismograms before (solid)
the shape of a ring, centered on the source location (see and after (dashed) introducing the slowness perturbation in
Figure 2). The strength of the eective slowness per- Figure 2 for ve dierent receivers located 2000-m away from
turbation sef f /sef f is constant within the ring, and a the source.
taper is applied outwards from the ring to avoid gener-
ating diractions from the borders. The strength of the
eective slowness perturbation is sef f /sef f =0.016. We
calculate the perturbed and unperturbed nite-dierence
lag (s)
0.012
0.01
0.01
sentation of the multiply scattered waveeld as a diusion
0.008
0.006 0.008
process. For the example presented here, we show that
0.004 0.006 our theory accurately predicts the mean traveltime
0.002
0 0.5 1 1.5 2 2.5 0.004
0.5 1 1.5 2 2.5
change of the diuse waveeld in a multiple-scattering
t (s) t (s)
medium. In order to detect a localized perturbation in
Fig. 4: Theoretical mean traveltime change (solid) versus trav- the multiple-scattering medium, we need access to many
eltime change for one receiver (asterisks) located 2000 m (left) realizations of the diuse waveeld propagating through
and 4000 m (right) away from the source. the medium characterized by strong and random velocity
0.02 0.016
uctuations. For the example shown here, we obtain an
0.018 estimate of the mean traveltime change by averaging
0.014
0.016 over dierent receiver locations. In general, many
0.014 0.012 realizations of the experiment are not available. In that
0.012
case, one can average over dierent source-receiver pairs.
lag (s)
lag (s)
0.01
0.01
0.008 0.008
Even tough we have tested the technique only for 2D
0.006 acoustic waves, the theory is valid for 3D scalar waves.
0.006
0.004 Further work needs to be done to adapt the technique
0.002
0 0.5 1 1.5 2 2.5
0.004
0.5 1 1.5 2 2.5 for the more complicated and realistic 3D elastic wave
t (s) t (s)
propagation. Application of this technique in geophysics,
Fig. 5: Theoretical mean (solid) versus estimated averaged (as- non-destructive testing, time-lapse monitoring, and other
terisks) traveltime change for a receiver located 2000 m (left) situations where elastic waves are used, require the
and 4000 m (right) away from the source. extension of the present formulation to include mode
conversions of the elastic waves propagating through
synthetic seismograms for arrays of 100 receivers around a medium characterized by random uctuations of its
the source for two source-receiver separations: 2000 and elastic parameters.
4000 m. Figure 3 shows the calculated waveeld for ve
receivers located 2000-m away from the source. Note the
strength of the multiple scattered waves after the highly References
attenuated ballistic arrival.
Figure 4 shows the traveltime change estimated with Frankel, A., and Clayton, R., 1984, A nite-dierence
coda wave interferometry for one receiver located 2000 simulation of wave propagation in two dimensional ran-
m and another located 4000 m away from the source ver- dom media: Bull. Seism. Soc. Am., 74, no. 6.
sus the predicted mean traveltime change calculated with
eq. (12). Note the uctuations of the estimated trav- Frisch, U., 1965, Wave propagtion in random media: In-
eltime change, for one receiver, around the theoretical stitute d Astrophysique.
mean traveltime change of the diuse waveeld. These
uctuations arise because the traveltime change for one Paaschens, J., 1997, Solution of the time dependent boltz-
realization of the multiply scattered waveeld diers from man equation: Phys. Rev. E, pages 11351141.
the mean traveltime change of the diuse waveeld.
Sheng, P., 1995, Wave propagation and mesoscopic phe-
To obtain a better estimate of the mean traveltime change nomena: , Academic Press.
from the synthetic seismograms, we measure the travel-
time change for all 100 receivers in the circular array, and Snieder, R., Gret, A., Douma, H., and Scales, J., 2002,
average the resulting 100 traveltime-change curves to ob- Coda wave interferometry for estimating nonlinear be-
tain the averaged traveltime change of the multiply scat- havoir in seismic velocity: Science, 295, no. 22, 2253
tered waveeld. Because the random medium is statisti- 2255.
cally homogeneous and the slowness perturbation is sym-
metric around the source, the 100 seismograms recorded Snieder, R., 2002, Coda wave interferometry and the equi-
in the circular array around the source are equivalent to libration of energy in elastic media: Phys. Rev. E, 66.
100 realizations of the multiply scattered waveeld. The
result of the averaging of the traveltime change for the re- Tatarski, V., 1963, Propagation of waves in a medium
ceivers located 2000 m and 4000 m away from the source with strong uctuation of the refractive index: Soviet
is shown in Figure 5. After averaging, the estimated mean Physics Jetp, pages 458463.
traveltime change shows better agreement with the pre-
dicted mean traveltime change. Tourin, A., and Fink, M., 2000, Multiple scattering of
sound: Waves in Random Media, pages R31R60.
Conclusions