2-D Niblett-Bostick magnetotelluric inversion - MTNet
2-D Niblett-Bostick magnetotelluric inversion - MTNet
2-D Niblett-Bostick magnetotelluric inversion - MTNet
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
J. RODRÍGUEZ et al. 2-D <strong>Niblett</strong>-<strong>Bostick</strong><br />
( t)<br />
( t)<br />
σ = σ .<br />
(15)<br />
j w l ( j+<br />
l)<br />
l=<br />
−n<br />
of the data in such a way that the averaging functions<br />
resemble boxcar functions for the averages to have the usual<br />
meaning. The resulting models are not intended to produce<br />
responses that fit the data, but to address the non-uniqueness<br />
character of the inverse problem. Within the limitations of<br />
linearization, they can be called true averages. More rigorous<br />
approaches for computing averages have been developed for<br />
the 1-D <strong>magnetotelluric</strong> problem (e.g. Weidelt, 1992), but their<br />
extension to higher dimensions are far from trivial.<br />
Simulated averages<br />
The shortcut that we take consists of computing models<br />
that are averages of themselves everywhere with a given<br />
neighborhood, and that at the same time their responses fit<br />
the data at an optimal level. Consider that<br />
(15)<br />
The dynamics of the HANN in (14) transforms to<br />
(16)<br />
Following Zhang and Paulson (1997), we will refer to<br />
this as the regularized version of the HANN or RHANN.<br />
The updated conductivities sj<br />
Geologica Acta, 8(1), 15-30 (2010)<br />
19<br />
DOI: 10.1344/105.000001513<br />
(t+l) are obtained from averages<br />
of the original values sj (t) . The choice of the appropriate<br />
filter depends on the application. Zhang and Paulson (1997)<br />
found binomial filters useful to stabilize their solution.<br />
In our case, we use boxcar functions, for this leads to the<br />
simplest and most intuitive type of averages. The filter, with<br />
2n+1 uniform weights wk = (2n+1) -1 its performance in 1-D, and compare its results with<br />
existing methods of average estimations. In the following<br />
lines, we test the performance of equation (16) as it applies<br />
to the solution of equation (6). The elements of the matrix<br />
are computed using equation (7) which represents the 1-D<br />
version of the proposed approximation. The synthetic data<br />
for the tests, shown in Fig. 1A, correspond to apparent<br />
conductivity responses of the model shown in Fig. 1B.<br />
The tests are intended to demonstrate that the proposed<br />
approach for the computation of averages has the same<br />
properties as the simple formula for averages in Gómez-<br />
Treviño (1996). The formula is<br />
(18)<br />
n<br />
( t)<br />
( t)<br />
σ where represents the average of conductivity<br />
j = ∑ w lσ<br />
( j+<br />
l)<br />
.<br />
(15)<br />
1 2<br />
n<br />
l=<br />
−n<br />
between zi ( t)<br />
( t)<br />
= 0.707da(Ti),i = 1,2. The averages are assigned<br />
σ j = ∑ w lσ<br />
( j+<br />
l)<br />
.<br />
(15) to the mean geometrical depth z = √ z z . Any two data<br />
1 2<br />
l=<br />
−n<br />
points can be used in the formula regardless of how far<br />
1 1 ⎧ N n<br />
⎫<br />
( t+<br />
1)<br />
( t)<br />
apart they are in the sounding curve. If the points are<br />
i = + sgn⎨<br />
∑ Tij[ ∑ wlσ<br />
( j+<br />
l)<br />
] + I i ⎬.<br />
(16)<br />
2 2 ⎩ j≠i1<br />
= 1 1 l=<br />
−n<br />
⎧ N n<br />
⎭ ⎫<br />
( t+<br />
1)<br />
( t)<br />
contiguous, the windows are narrow and the average<br />
σ i = + sgn⎨<br />
∑ Tij[ ∑ wlσ<br />
( j+<br />
l)<br />
] + I i ⎬.<br />
(16)<br />
2 2<br />
model, the average of the real earth, has the highest<br />
⎩ j≠i<br />
= 1 l=<br />
−n<br />
⎭<br />
possible resolution. If, on the other hand, the chosen<br />
t+<br />
1)<br />
points are wide apart, the windows are themselves wide,<br />
t )<br />
for z and z tend to separate. As the windows widen, the<br />
1 2<br />
− n,..., n<br />
models tend to flatten, as they should. This intuitively<br />
1<br />
= , k = −n,...,<br />
n<br />
appealing feature of spatial averages is built into equation<br />
2n<br />
+ 1<br />
(18) in spite of its simplicity. Figures 1C and 1D show<br />
N<br />
n<br />
1 1<br />
( t)<br />
( t)<br />
+ sgn{<br />
this feature developing when considering averages taken<br />
∑ [( 1−<br />
b ) Tijσ<br />
j + β ∑ Tijwlσ<br />
( j+<br />
l)<br />
] + I i}.<br />
(17)<br />
2 2<br />
N<br />
n<br />
from data values 1 and 5 periods apart, respectively, for<br />
( t+<br />
1)<br />
j≠i<br />
= 1 1<br />
l=<br />
−n<br />
( t)<br />
( t)<br />
σ i = + sgn{ ∑ [( 1−<br />
b ) Tijσ<br />
j + β ∑ Tijwlσ<br />
( j+<br />
l)<br />
] + I i}.<br />
(17)<br />
2 2<br />
the synthetic data presented in Fig. 1A.<br />
j≠i<br />
= 1<br />
l=<br />
−n<br />
σ<br />
, k = -n,..., n (square<br />
a ( T2<br />
) δ a ( T2<br />
) −σ<br />
a ( T1<br />
) δ a ( T1<br />
)<br />
σ ( z1,<br />
z2<br />
) >= window), can be of different widths. , We introduce (18) a further A corresponding sequence of full filtered models (β =1)<br />
parameter β δ ato<br />
( Taccommodate<br />
2 ) σ−<br />
δ ( T )<br />
a ( Ta<br />
2 ) 1 δ a ( Ta<br />
2traditional<br />
) −σ<br />
a ( T1<br />
) trade-off δ a ( T1<br />
)<br />
< σ ( z<br />
between is presented in Figs. 1E and 1F when the RHANN algorithm<br />
1,<br />
z2<br />
) >=<br />
, (18)<br />
two contrasting features. δWe<br />
modified the dynamics to is applied to the same set of data. It can be observed that<br />
a ( T2<br />
) −δ<br />
a ( T1<br />
)<br />
the behavior of the averages follows that of the averages<br />
i ), ( i = 1,<br />
2 1, 2 ) ><br />
shown in Figs. 1C and 1D. There are some differences<br />
that can be traced back to basic differences between the<br />
two approaches. One is that equation (18) is based on the<br />
(17) original <strong>Niblett</strong>-<strong>Bostick</strong> approximation that assumes a<br />
boxcar function as the kernel of equation (5). The RHANN<br />
When β =1 Hwe<br />
( xget<br />
, 0)<br />
the full filtered solution. On the other hand, algorithm uses the kernel as indicated in equation (5) which<br />
x 2<br />
| when σ a | = ωµ β =0 0 | we return to | , the original unfiltered solution (A.1) given includes a small negative sidelobe (Gómez-Treviño, 1987b)<br />
E y ( x,<br />
0)<br />
by equation (14). In this way, by varying this parameter we and extends to infinity. This is a somewhat less restrictive<br />
can gradually control the effect of the averaging process. approximation than the rather simpler boxcar function. The<br />
δE<br />
( x,<br />
0)<br />
H ( x,<br />
0)<br />
other difference originates in the fact that the windows in<br />
y δ x<br />
| σ a | = −2<br />
| σ a | Re[ − ]. (A.2)<br />
each approach are necessarily different, because in the first<br />
E ( x,<br />
0)<br />
H ( x,<br />
0)<br />
APPLICATIONS y<br />
x<br />
case they cannot be uniform, for they are determined by the<br />
data, and in the second case they are uniform for all depths.<br />
1-D averages<br />
The overall effect is a somewhat better performance for<br />
the approximation given by equation (5) judging by the<br />
Before proceeding to the application in 2-D of the above resemblance of the average models to the original model<br />
numerical averaging technique, it is convenient to illustrate from which the data were drawn.<br />
2<br />
∇ E y − iωµ<br />
0σ<br />
( x,<br />
z)<br />
E y = 0,<br />
(A.3)<br />
z<br />
− 2i δ<br />
E y ( x,<br />
z)<br />
= E(<br />
0)<br />
e ,<br />
(A.4)<br />
2 2<br />
δ = .<br />
(A.5)<br />
ωµ 0σ<br />
z z<br />
= 0 . 707δ<br />
a ( Ti ), i = 1,<br />
2<br />
1 2 z z<br />
n<br />
( t)<br />
( t)<br />
σ j = ∑ w lσ<br />
( j+<br />
l)<br />
.<br />
(15)<br />
l=<br />
−n<br />
1 1 ⎧ N n<br />
⎫<br />
( t+<br />
1)<br />
( t)<br />
σ i = + sgn⎨<br />
∑ Tij[ ∑ wlσ<br />
( j+<br />
l)<br />
] + I i ⎬.<br />
(16)<br />
2 2 ⎩ j≠i<br />
= 1 l=<br />
−n<br />
⎭<br />
( t+<br />
1)<br />
σ i<br />
(t )<br />
σ j<br />
1<br />
wk = , k = −n,...,<br />
n<br />
2n<br />
+ 1<br />
N<br />
n<br />
( t+<br />
1)<br />
1 1<br />
( t)<br />
( t)<br />
σ i = + sgn{ ∑ [( 1−<br />
b ) Tijσ<br />
j + β ∑ Tijwlσ<br />
( j+<br />
l)<br />
] + I i}.<br />
(17)<br />
2 2 j≠i<br />
= 1<br />
l=<br />
−n<br />
σ a ( T2<br />
) δ a ( T2<br />
) −σ<br />
a ( T1<br />
) δ a ( T1<br />
)<br />
< σ ( z1,<br />
z2<br />
) >=<br />
, (18)<br />
NEXES<br />
δ a ( T2<br />
) −δ<br />
a ( T1<br />
)<br />
< ( 1, 2 ) ><br />
H x ( x,<br />
0)<br />
2<br />
| σ a | = ωµ 0 | | ,<br />
(A.1)<br />
E y ( x,<br />
0)<br />
x , z)<br />
δE<br />
y ( x,<br />
0)<br />
δH<br />
x ( x,<br />
0)<br />
δ | σ a | = −2<br />
| σ a | Re[ − ]. (A.2)<br />
E y ( x,<br />
0)<br />
H x ( x,<br />
0)<br />
( x , z)<br />
x ( x,<br />
z)<br />
y ( x,<br />
z)<br />
2<br />
∇ E y − iωµ<br />
0σ<br />
( x,<br />
z)<br />
E y = 0,<br />
(A.3)<br />
z<br />
− 2i δ<br />
E y ( x,<br />
z)<br />
= E(<br />
0)<br />
e ,<br />
(A.4)<br />
2 2<br />
δ = .<br />
(A.5)<br />
ωµ 0σ<br />
( x , z)<br />
z z σ<br />
zi = 0 . 707δ<br />
a ( Ti<br />
), i = 1,<br />
2<br />
1 2 z z<br />
1 1 ⎧ N n<br />
⎫<br />
( t+<br />
1)<br />
( t)<br />
σ i = + sgn⎨<br />
∑ Tij[ ∑ wlσ<br />
( j+<br />
l)<br />
] + I i ⎬.<br />
(16)<br />
2 2 ⎩ j≠i<br />
= 1 l=<br />
−n<br />
⎭<br />
( t+<br />
1)<br />
σ i<br />
(t )<br />
σ j<br />
1<br />
wk = , k = −n,...,<br />
n<br />
2n<br />
+ 1<br />
N<br />
n<br />
( t+<br />
1)<br />
1 1<br />
( t)<br />
( t)<br />
σ i = + sgn{ ∑ [( 1−<br />
b ) Tijσ<br />
j + β ∑ Tijwlσ<br />
( j+<br />
l)<br />
] + I i}.<br />
(17)<br />
2 2 j≠i<br />
= 1<br />
l=<br />
−n<br />
σ a ( T2<br />
) δ a ( T2<br />
) −σ<br />
a ( T1<br />
) δ a ( T1<br />
)<br />
< σ ( z1,<br />
z2<br />
) >=<br />
,<br />
δ a ( T2<br />
) −δ<br />
a ( T1<br />
)<br />
(18)<br />
< ( 1, 2 ) ><br />
n<br />
( t)<br />
( t)<br />
σ j = ∑ w lσ<br />
( j+<br />
l)<br />
.<br />
(15)<br />
l=<br />
−n<br />
( 1)<br />
1 1 ⎧ N n<br />
⎫<br />
t+<br />
( t)<br />
i = + sgn⎨<br />
∑ Tij[ ∑ wlσ<br />
( j+<br />
l)<br />
] + I i ⎬.<br />
(16)<br />
2 2 ⎩ j≠i<br />
= 1 l=<br />
−n<br />
⎭<br />
− n,..., n<br />
N<br />
n<br />
1 1<br />
( t)<br />
( t)<br />
+ sgn{ ∑ [( 1−<br />
b ) Tijσ<br />
j + β ∑ Tijwlσ<br />
( j+<br />
l)<br />
] + I i}.<br />
(17)<br />
2 2 j≠i<br />
= 1<br />
l=<br />
−n<br />
σ a ( T2<br />
) δ a ( T2<br />
) −σ<br />
a ( T1<br />
) δ a ( T1<br />
)<br />
σ ( z1,<br />
z2<br />
) >= , (18)<br />
δ a ( T2<br />
) −δ<br />
a ( T1<br />
)<br />
z =<br />
3<br />
i ), i = 1,<br />
2<br />
ANNEXES<br />
H x ( x,<br />
0)<br />
2<br />
| σ a | = ωµ 0 | | ,<br />
(A.1)<br />
E y ( x,<br />
0)<br />
σ ( x,<br />
z)<br />
H x ( x,<br />
0)<br />
2<br />
| σ a | = ωµ 0 | | ,<br />
(A.1)<br />
E y ( x,<br />
0)<br />
δE<br />
y ( x,<br />
0)<br />
δH<br />
x ( x,<br />
0)<br />
δ | σ a | = −2<br />
| σ a | Re[ − ]. (A.2)<br />
E y ( x,<br />
0)<br />
H x ( x,<br />
0)<br />
δE<br />
y ( x,<br />
0)<br />
δH<br />
x ( x,<br />
0)<br />
| σ a | = −2<br />
| σ a | Re[ − ]. (A.2)<br />
E y ( x,<br />
z)<br />
E y ( x,<br />
0)<br />
H x ( x,<br />
0)<br />
H x ( x,<br />
z)<br />
E y ( x,<br />
z)<br />
z z σ<br />
zi = 0 . 707δ<br />
a ( Ti<br />
), i = 1,<br />
2<br />
1 2 z z z =<br />
ANNEXES<br />
H x ( x,<br />
0)<br />
2<br />
| σ a | = ωµ 0 | | ,<br />
(A.1)<br />
E y ( x,<br />
0)<br />
σ ( x,<br />
z)<br />
δE<br />
y ( x,<br />
0)<br />
δH<br />
x ( x,<br />
0)<br />
δ | σ a | = −2<br />
| σ a | Re[ − ]. (A.2)<br />
E y ( x,<br />
0)<br />
H x ( x,<br />
0)<br />
E y ( x,<br />
z)<br />
H x ( x,<br />
z)<br />
E y ( x,<br />
z)<br />
2<br />
∇ E y − iωµ<br />
0σ<br />
( x,<br />
z)<br />
E y = 0,<br />
(A.3)<br />
z<br />
− 2i δ<br />
E y ( x,<br />
z)<br />
= E(<br />
0)<br />
e ,<br />
(A.4)<br />
2 2<br />
δ = .<br />
(A.5)<br />
ωµ 0σ<br />
H x ( x,<br />
z)<br />
∇<br />
× E = −iωµ<br />
0H<br />
n<br />
∑