Fisika Bumi Bab 1
Fisika Bumi Bab 1
Fisika Bumi Bab 1
Introduction
Seismic tomography applied to the solid Earth is a non-linear process [Pavlis and Booker, 1983].
In general, solutions are obtained by linearization with respect to a reference Earth model [e.g.,
Aki and Lee, 1976; Nolet, 1978]. The tomographic images resulting from such linearized inversion
are dependent on the initial reference models and hypocentral locations [Michael, 1988; Vander
Hilst and Spakman, 1989; VanderHilst et al., 1991]. This dependence, in conjunction with
ambiguities intrinsic to the inversion of seismic data, is an issue that has to be addressed in any
application of seismic tomography [Lees and Shalev, 1992]. Most velocity models of the Earth's
interior are derived by first establishing a simple smooth model that explains some weighted
average of the observations. Subsequently, this initial model is modified until a sufficient degree
of coincidence between the bulk of the observations and predicted values is obtained. This
procedure is followed in seismic tomographic studies, with a one-dimensional (l-D) model usually
serving as the initial reference model for the three-dimensional (3-D) inversion. In many previous
studies the influence of the initial reference 1-D model on the results of the tomographic inversion
has not been fully appreciated. In the following we show that an inappropriate initial reference
model may not only affect the quality of the 3-D image by introducing artifacts, but it may also
influence the confidence calculations by underestimating the uncertainties of the results. To
overcome these problems, Kissling et al. [1984] and Kissling [1988] proposed a two-step
procedure to obtain 3-D tomographic results with minimal dependence on the reference model:
First, the travel time data are jointly inverted to obtain a 1-D tomographic solution, together with
revised hypocenter coordinates and station corrections. We call this new model the "minimum 1-
D model" [Kissling, 1988]. Second, the 3-D tomographic inversion is determined using the
minimum 1-D model as the starting model. In this study we present theoretical arguments for such
an approach and demonstrate its importance for 3-D tomographic results for a simple synthetic
test case mimicking San Andreas Fault structure in central California.
Coupled Hypocenter Velocity Model Problem
The arrival time of a seismic wave generated by an earthquake is a nonlinear function of the station
coordinates (s), the hypocentral parameters (h, including origin time and geographic coordinates),
and the velocity field (m). tob s = f( s,h,m ). (1) In general, neither the true hypocentral parameters
nor the velocity field are known. With arrival times and station 19,635 location being the only
measurable quantities, we cannot solve (1) directly. To proceed, we have to make an educated
guess of the unknown parameters. Using an a priori velocity model, we trace rays from a trial
source location to thereceivers and calculate theoretical arrival times (tcalc). The differences
between the observed and the calculated arrival time, the residual travel time (tres), can be
expanded as functions of the differences {A) between the estimated and the true hypocentral and
velocity parameters. To calculate suitable adjustments (corrections) to the hypocentral and model
parameters, we need to know the dependence of the observed travel times on all parameters. For
hypocenter parameters except the origin time this dependence is strongly nonlinear [e.g., Thurber,
1985], and for velocity parameters it is moderately nonlinear, even in a 1-D model [Pavlis and
Booker, 1983]. Applying a first-order Taylor series expansion to (1), we obtain a linear relationship
between the travel time residual and adjustments to the hypocentral (Ahk) and velocity (Ami)
parameters:
In matrix notation, the coupled hypocenter velocity model parameter relation can be written as
Neglecting the effect of (Mm) in equation (3) while locating the earthquakes, for example, has the
potential to introduce systematic errors into the estimated hypocenter locations [Thurber, 1992;
Eberhart-Phillips and Michael, 1993]. Similarly, neglecting of (Hh) in equation (3) may result in
biased velocity parameters [Michael, 1988; VanderHilst and pakman, 1989]. As we demonstrate
below, inclusion of both hypocenter and model parameters in the estimation procedure does not
necessarily guarantee successful recovery of either true distribution. Unless we have "guessed" the
correct hypocentral coordinates, tomographic imaging with local earthquake data demands the
updating of both hypocenter and velocity parameters. We concur with Thurber [1992] that this is
most reliably achieved by solving the coupled hypocenter-velocity model problem, rather than
alternating independent hypocenter and velocity adjustment steps. To reduce the computational
burden of solving the very large system of equations (3), Pavlis and Booker [1980] and Spencer
and Gubbins [1980] independently introduced an algorithm, to separate A into the two smaller
matrices, one containing the hypocenter location information, and one containing the model
parameter information. The reduced form of A pertaining to the velocity model may then be solved
separately to obtain the same solution as if the entire matrix were being inverted. Kissling [1988]
demonstrated the effectiveness of the parameter separation procedure for local earthquake
tomography in a test with data from Long Valley, California. (3) Parameter separation does not,
however, reduce the dependency of the solutions on the reference velocity model or the reference
hypocenters. Generally, the solution to (3) is taken as the least squares solution that minimizes a
weighted combination of the squared SAF Figure 1. Schematic velocity models of the San Andreas
Fault (SAF) in central California. a: Synthetic velocity model of SAF used to calculate synthetic,
yet realistic data sets for the Loma Prieta earthquake series (see text). Different shading denotes
different velocities (values are in kilometers per seconds). Note the lack of vertical velocity
gradients and the strong lateral change in velocity. b: Schematic 1-D model used to approximate
the unknown velocity structure for earthquake location and used as the reference model for 3-D
tomographic inversions. The two reference 1-D models (i.e., the refraction- based and the
minimum 1-D model) have identical layer thicknesses but different layer velocities. Different
shading denotes different velocities. For exact velocity values, see Figure 3.
A
B
error (eYe) and a measure of the velocity model change, typically its Euclidean norm
(mTm)[Spakmart and Nolet, 1988]. Under the assumption of normally distributed errors and
model perturbations, the least squares formulation results in the most likely solution that belongs
to the same family of solutions as the initial reference-model. By seeking a solution in the
neighbourhood of the initial reference model, defects in the reference model may lead to artifacts
in the 3-D tomographic images, particularly when the solution is not refined through model
updating and iteration. Because the parameter space commonly contains several thousand
unknowns, simple iteration schemes may be incapable of avoiding local minima and thus may not
converge to the global minimum. Furthermore, the size of A and/or the application of parameter
separation may make it difficult, if not impossible, to compute the normal diagnostics of model
performance, such as the model and data resolution matrices and model covariance matrix. Thus
potential trade-off between hypocentral parameters and velocities are difficult to identify.
Step 2. Establishing the Geometry and the Velocity Intervals of Potential 1-D Model(s)
Select about 500 of the best events in the data (i.e., those with the most high-quality P arrivals)
that cover the entire area under consideration. Relocate them with routine VELEST using a
damping coefficient of 0.01 for the hypocentral parameters and the station delays and 0.1 for the
velocity parameters. Invert for hypocenters every iteration and for station delays and velocity
parameters every second iteration. Repeat this procedure several times with new (updated)
velocities in the reference 1-D model, with perhaps the new station delays, and with new
hypocenter locations. Repeat the procedure also for reduced number of layers where possible by
combining adjacent layers with similar velocities. Unless clearly indicated by the data, in most
cases it is preferable to avoid low velocity layers, as they normally introduce instabilities. Our
experience suggests that shot or blast data should not be included in the 1-D model inversion.
Rather, such data should be used to set the near-surface velocities, and to test the performance of
the resulting minimum 1-D model when used for locating hypocenters. This countrintuitive
suggestion may be understood by considering that ray paths with both endpoints near the surface
sample, on average, a much more heterogeneous part of the Earth than do ray paths from events in
the seismogenic crust. The goal of this trial and error approach is to establish reasonable geometry
of the crustal model and corresponding intervals for the velocity parameters and station delays. In
addition, this approach provides valuable knowledge about the quality of the data. Procede to the
next step when (1 the earthquake locations, station delays, and velocity values do not vary
significantly in subsequent runs; (2 the total RMS value of all events shows a significant reduction
with respect to the first routine earthquake locations; and (3 the calculated 1-D velocity model and
the set of station corrections make some geological sense (e.g., stations with negative travel time
residuals should lie in local high-velocity areas with respect to the reference station, etc.) and do
not violate a priori information. If all these requirements are satisfied, the result may be called the
"updated a priori 1-D model with corresponding station residuals".
If several subsets of 500 events were extracted, test the dependence of your minimum 1-D model
on specific data. Find the 1-D model and station residuals that will best fit the results from all
subsets, mix data from different subsets, and repeat step 4. If the results are unsatisfactory, evaluate
the best 1-D model by the procedure described in step 6.