Está en la página 1de 10

PII:

Computers & Geosciences Vol. 24, No. 2, pp. 173182, 1998 # 1998 Elsevier Science Ltd. All rights reserved Printed in Great Britain 0098-3004/98 $19.00 + 0.00 S0098-3004(97)00121-0

INVERTING THE PARAMETERS OF AN EARTHQUAKERUPTURED FAULT WITH A GENETIC ALGORITHM


NDEZ,2 and JOHN B. RUNDLE3 FERNA TING-TO YU, JOSE
Institute of Earth Sciences, Academia Sinica, PO Box 1-55 Nankang, Taipei, Taiwan; 2Instituto de ticas, Ciudad Universitaria, 28040 Astronom a Geodesia (CSIC-UCM), Facultad de Ciencias Matema Madrid, Spain, and 3Cooperative Institute for Research in Environmental Sciences, University of Colorado, Boulder, CO 80309, U.S.A. (e-mail: yutt@earth.sinica.edu.tw)
1

(Received 9 January 1997; revised 14 August 1997) AbstractNatural selection is the spirit of the genetic algorithm (GA): by keeping the good genes in the current generation, thereby producing better ospring during evolution. The crossover function ensures the heritage of good genes from parent to ospring. Meanwhile, the process of mutation creates a special gene, the character of which does not exist in the parent generation. A program based on genetic algorithms using C language is constructed to invert the parameters of an earthquake-ruptured fault. The verication and application of this code is shown to demonstrate its capabilities. It is determined that this code is able to nd the global extreme and can be used to solve more practical problems with constraints gathered from other sources. It is shown that GA is superior to other inverting schema in many aspects. This easy handling and yet powerful algorithm should have many suitable applications in the eld of geosciences. # 1998 Elsevier Science Ltd. All rights reserved Key Words: Genetic algorithm, Earthquake, Fault, Displacement eld, C.

INTRODUCTION

Seismologists have been relying on geodetic evidence to describe the ``power'' of earthquakes for years. The osets of faults caused by earthquakes occurring many years ago in fact still provide geophysicists with important information today. Ground movement associated with an earthquake is the only evidence that persists. Unlike a waveform instrument, it does not have to be operating when an earthquake ruptures. For this reason, the displacement eld measured by the geodetic method associated with an earthquake has provided important data to describe and also to constrain the parameters of earthquake faulting. Due to the complexity of the earthquake rupturing process, it is dicult to interpret the measured results. Therefore, a powerful inversion tool is needed for both the geodesist and seismologist to be able to interpret their geodetic measurements in terms of earthquake rupture mechanisms. Several techniques have been used to invert the earthquake rupture parameters of a coseismic displacement eld. One, the Monte Carlo method, used by Hudnut and others (1996) estimates the fault geometry and slip independence of seismological information, which minimizes the dierences between the observed and predicted horizontal and vertical displacements. Segall and Matthews (1988), as well as Sataka (1993), by optimizing some criterion of solution desirability over the class of sol173

ution that ts the data, used the method of least squares to analyze the undetermined linear systems. The quasi-Newton method with a nonlinear optim rnado ttir and ization algorithm was employed by A Segall (1994) by inverting the geodetic data to image the 1989 Loma Prieta earthquake. A trialand-error method was also used by Holdahl and Sauber (1994) to invert the coseismic slip in the 1964 Prince William Sound earthquake. All of these methods have inverted the geodetic eld data successfully. However, they are either problem-specic or require intense coding and abundant computing facilities. To process a general geodetic measurement with the least modication of the code for a personal computer or entry level workstation is the goal of this C code. The basic structure of this genetic algorithm (GA) code is modied from Denis Cormier as listed in Michalewicz (1992, appendix). The real-value GA is used instead of the traditional binary array method. With this alternative, the binary and real value conversion is eliminated in the code, thereby reducing signicantly the computing time.
THE SPIRIT OF THE GENETIC ALGORITHM

The genetic algorithm is dierent from other optimization and search procedures in four ways: (1) the GA works with a coding of the parameter set, not the parameters themselves; (2) the GA searches

174

ndez and J. B. Rundle T. T. Yu, J. Ferna

from a population of points, not a single point; (3) the GA uses payo (objective function) information, not derivatives in its auxiliary knowledge; and (4) the GA makes use of probabilistic transition rules, not deterministic ones (Goldberg, 1989). The genetic algorithm (like any other evolution program) for a particular problem must have the following ve components (Michalewicz, 1992): a genetic representation for potential solutions to the problem; a method to create an initial population of potential solutions; an evaluation function that plays the role of the environment, rating solutions in terms of their ``tness''; genetic operators that alter the composition of ospring; and values for various parameters that the genetic algorithm uses (population size, probabilities of applying genetic operators, etc.). The process begins by representing the model to be optimized by a real-value string. A simple genetic algorithm that yields good results in many practical problems is composed of three operators: reproduction, crossover and mutation. Reproduction is a process in which individual strings are copied on the basis of their objectivefunction values, f (called the tness function). Copying strings according to their tness values means that strings with a better value of tness have a higher probability of contributing one or more ospring in the next generation. This operator is an articial version of natural selection, a Darwinian survival of the ttest among strings. After reproduction, a simple crossover may proceed in two steps. First, members of the newly reproduced string in the mating pool are mated at random. Secondly, each pair of strings undergoes crossover as follows: an integer position k along the string is selected uniformly at random among 1 and the string length less one (1, l 1). Two new strings are created by swapping all characters between positions k + 1 and l inclusively. For example, strings A1 and A2 can be considered: A1 0 25 7 49j53 A2 3 7 15 25j7 Supposing in choosing a random number between 1 and 4, k = 4 is obtained (as indicated by the separator symbol v). The resulting crossover yields two new strings where the prime (') means the strings are part of the new generation: A1 H 0 25 7 49 7 A2 H 3 7 15 25 53

changing the chosen chromosome into a random number within the lower and upper bounds of such a parameter. By itself, mutation is a random walk through the string space. When used sparingly with reproduction and crossover, it insures against any premature loss of an important string structure (Goldberg, 1989).

DESCRIPTION OF THE CODE

The third process, mutation, is required because even though reproduction and crossover eectively search and recombine existing sub-strings, occasionally these processes may lose some potentially useful genetic material. In simple GA, mutation is the occasional (with small probability) random alteration of the values of a string position. In binary coding, this simply means changing a 1 to a 0 and viceversa. When used in real value coding, it means

This code consists of 11 sub-functions. Looping proceeds within an assigned maximum generation serving as the natural evolution. The elitist function ensures the best gene survives in the mating process. The crossover function produces the more suitable ospring, whereas mutation provides a chance to create a special ospring that bears a certain quality, which does not exist in the parent. The key parts of this code are the ``evaluate'' and ``Okada'' functions. The ``evaluate'' function denes the meaning of tting. In this instance, it is the sum of the absolute dierences between the three-dimensional computed and measured displacements. To minimize this tting index is the goal of evolution. Several ags are set to eliminate the unnecessary calculations which will then reduce the computation time. For example, if the chosen gene is not suitable for one point with a large margin, then this gene is not tested for the remaining points since one gene must provide a result that satises most of the points within a feasible range. A detailed description of each dened function is listed in Table 1. The code begins at an initial gene pool, and each gene set presents one possible solution to the inversion problem. The gene is assigned randomly within the upper and lower bounds of each parameter. Then each gene is brought in to evaluate its tness; for example, to nd a minimum value of function f(x, y, z)=x2xy+z and (10 x 10, 0 y 100, 50 z 50). Thus, to create a possible gene combination for (x, y, z) randomly are (7, 90, 40), (8, 7, 30), (0, 35, 40) F F F (xn, yn, zn), where n is the number of the population. The tness is calculated as function the f dened as, f1=639, f2= 22, f3=40, F F F . On the basis of the tness value, the gene is ranked. The lower the tness value, the greater the chances are for this gene to be mated. In the next generation, the gene is selected according to its ranking, then goes through the crossover and mutation processes, thereby becoming the ``Next Generation Genes''. The tness of these new genes is computed again. For example, if gene (1) and gene (3) are selected to perform the process of crossover, then the crosspoint is selected as k = 2. This also presents a fully random process in selecting mated genes and the cross-point. As a result, the new gene(1)' = (7, 90, 40), gene(2)' = (8, 7, 30) and gene(3)' = (0, 35, 40). If the second chromosome of gene(2)' is

Inverting fault parameters with a genetic algorithm Table 1. Denition of functions used in gas3d.c code Routine name Initialize Evaluate Keep_the_best Select Crossover Mutate Report Elistic Okada Xover Swap Function of the routine

175

initializes the values of genes within the variable bounds. Random values of genes of each genotype generated between the bound evaluate the tness of each chromosome with evaluated function keep tracking the best member of population standard proportional selection to the parent chromosome; ensure the best member survives for mating selects two parents that take part in the crossover execute a non-uniform mutation report progress of the simulation compare the performance between previous and this generation to insure the best gene goes to the mating process Okada elastic model to serve as the object function; each potential solution is brought into this routine to compute its corresponding tness performs crossover of the two selected parents a swap procedure that helps in swapping 2 variables

picked to be mutated, then gene(2)0 = (8, 75, 30), and the tness of this generation is f1' = 719, f2'=566, f3'=40. Comparing the best tness of this generation, 566, to that of the previous generation, 22, it is determined that the current generation is more suitable in the selection process. Accordingly, it is kept and ranked and chosen for going through another evolution. The evolution continues until either the tness of the generation becomes smaller than the pre-dened margin, or the number of maximum generation is reached. This code acts like a smart selector picking up suitable parameters to run the forward problem automatically. Therefore, the coding is straightforward and easy to modify and maintain. The ow chart of this code is shown in Figure 1. The Okada function is the newly dened function used to compute the three-dimensional displacements due to a uniform slip over a rectangular fault in an elastic half-space after Okada (1985).
CONTROLLING THE CODE

are the measured coseismic displacements at this point. To add another measured point, another line is added to the same format. The total points in this ``target.l'' must be equal to the npt value that is dened in the code. It is equal to 5 in the demonstration program (gas3d.c). The second le needed is called the ``gadata.txt''. This le provides the lower and upper bounds for inverted parameters. For the example shown above, three parameters need to be inverted. Therefore, the contents of this le is 4 100 10 5 0 50 100 X 50

The controlling variables of this code are the width of the fault (w), depth of the fault (hmin), length of the fault (length), dip angle of the fault (dip), slip on the strike direction (ustr), slip on the dip direction (udip) and the distance of the target point to the center of the fault in km (x, y) eight parameters, as shown in Figure 2. A combination of these eight parameters in Okada's formula, and the three-dimensional coseismic displacement eld on point (x, y) is calculated. As revealed, this is not a linear function. In fact, the result of this Okada function is not directly related to certain parameters. Both an adequate gene pool and evaluating generations are needed to nd the proper gene for the problem of inversion. Two les are needed to run this code. First, the ``target.dat'' le contains the measured data set with the data format being (int x, int y, oat Ux, oat Uy, oat Uz). (x, y) is the distance of the point away from the center of the fault. Ux, Uy, and Uz

Each line represents one target parameter, and each parameter is associated with a lower and upper bound. In the code itself, most of the control parameters are listed in the beginning of the code. They are population size, maximum number of generations, number of problem variables, probability of crossover and probability of mutation. Another controlling variable is the number of target points that need to be calculated. This is dened as npt which is located in the function ``evaluate''. The population size and maximum number of generations are the key numbers that control the performance of the code. The greater the population size, the greater are the chances for the code to nd a suitable gene; nevertheless, more computation time is required as the population size grows. The larger the evaluated generation number, the greater are the chances that better ospring are produced, but again a longer period of time is required. The number of the problem variable corresponds to the program itself. It is important not to change this value, unless the code is fully understood. Probability of crossover is usually assigned a larger value, even greater than 70%, to ensure that better parents are chosen to generate the ospring. Meanwhile, the probability of mutation is given a low value of less than 10%, which is large enough

176

ndez and J. B. Rundle T. T. Yu, J. Ferna

Figure 1. Flow chart of major subroutines in GA code.

Inverting fault parameters with a genetic algorithm

177

Figure 2. Controlled parameters for ruptured fault of inverting process. Hmin is depth of fault; width is width of fault plane; length is ruptured length of fault; theta is dipping angle of fault; Ustr is slip along direction of strike of fault; and Udip is slip on direction of dipping of fault.

some of the prior geophysical constraints can reduce the computation time. One pure strikeslip, dipslip and an obliqueslip event were tested to determine the suitable variable of the 5 measured points. The results of these three examples are shown in Tables 24. For a population size of 500 with 3000 generations, this code only took 15 min on a Pentium 90 PC for the results to be inverted within a 10% bias. This code has been successfully tested with Microsoft C++ on the Microsoft Windows-95 system and IBM RS-6000 with AIX C without requiring any modications.

VERIFICATION AND APPLICATION OF THE CODE

to create the mutation of genes but not so great as to transform good genes into bad ones. In the demonstration, 5 parameters were inverted: W, Hmin, length, rake and slip. The center location, strike and dip of the ruptured fault can be found from the focal mechanism of the earthquake. The performance of this GA code drops dramatically when the variable is greater than 5. The use of

The code was rst tested to nd the global maximum of a function z = sin[x+sin(y)]; 3<x<3; 3<y<3. This function is shown in Figure 3, and with the modication of the objective function in the code, the maximum of z=1.0 is found when x=0.720 and y=2.124. This data set was derived by using 50 populations, 2000 generations, 0.7 as the probability of crossover and 0.15 for the mutation process. It is determined that the code is capable of nding a global extreme. This code was then used to nd the rupture parameters of the

Table 2. Target parameters and output results for pure strikeslip fault, rake = 08. Case I: pure strike-slip fault. Source parameters: W = 10 km, Hmin = 0 km, length = 40 km, rake = 08, slip = 1000 cm X 2 20 1 50 0 Y 3 1 20 50 20 Ux 743.42 413.04 31.63 4.12 31.65 Uy 9.37 241.47 1.74 4.14 0 Uz 1.32 160.11 0.48 1.68 0

Target.dat used for inverting pure strike-slip fault.

Inverted results from gas3d code. Var(0) = 11 Var(1) = 0 Var(2) = 40 Var(3) = 0 Var(4) = 914 cux = 688.135726 ux = 743.420000 cux = 378.312765 ux = 413.040000 cux = 34.174044 ux = 31.630000 cux = 4.224150 ux = 4.120000 cux = 31.585132 ux = 31.650000 0 0 10 0 500 cuy = 9.855973 uy = 9.370000 cuy = 218.300552 uy = 241.470000 cuy = 1.825958 uy = 1.740000 cuy = 4.243457 uy = 4.140000 cuy = 0.000000 uy = 0.000000 20 10 60 90 2000 cuz = 1.253523 uz = 1.320000 cuz = 147.462908 uz = 160.110000 cuz = 0.460635 uz = 0.480000 cuz = 1.691676 uz = 1.680000 cuz = 0.000000 uz = 0.000000 for point (2, 3), computed measurements For point (20, 1) for point (1, 20) for point (50, 50) for point(0, 20)

Gadata.txt le used for inverting strike-slip fault.

178

ndez and J. B. Rundle T. T. Yu, J. Ferna

Table 3. Target parameters and output results for oblique-slip fault, rake = 458. Case II: oblique-slip fault. Source parameters: W = 10 km, Hmin = 0 km, length = 40 km, rake = 458, slip = 1000 cm X 2 20 1 50 0 Y 3 1 20 50 20 Ux 525.96 248.82 23.34 5.54 22.38 Uy 111.39 231.95 59.59 6.31 58.42 Uz 351.63 114.88 39.65 3.89 39.34

Target.dat used for inverting oblique fault.

Inverted results from gas3d code. Var(0) = 10 Var(1) = 0 Var(2) = 49 Var(3) = 46 Var(4) = 912 cux = 473.971245 ux = 525.960000 cux = 231.164301 ux = 248.820000 cux = 25.038494 ux = 23.340000 cux = 5.952934 ux = 5.540000 cux = 24.372504 ux = 22.380000 cuy = 101.485798 uy = 11.390000 cuy = 214.746382 uy= 231.950000 cuy = 58.447940 uy = 59.590000 cuy = 6.979694 uy = 6.310000 cuy = 57.586003 uy = 58.420000 cuz = 330.010441 uz = 351.630000 cuz = 109.938258 uz = 114.880000 cuz = 39.023220 uz = 39.650000 cuz = 4.361722 uz = 3.890000 cuz = 38.855316 uz = 39.340000

1983, Borah Peak, Idaho earthquake with the vertical coseismic displacements used by Ward and Barrientos (1986) and Stein and Barrientos (1985), and is shown to be a more practical approach. With the inversion of the 30 listed items of data, it was determined that dip = 478, U (slip on the fault) = 2.0 m, W (Width of the fault) = 18 km, and depth of the fault is 0 km, which means the fault ruptured to the surface. The results of model

and data used are shown in Figure 4. These results are consistent with those from trial-and-error like approaches. The power of using natural-selection algorithms in geophysics can be demonstrated in the application of inverting the composite parameters of a rupture fault plane, as a single rectangular plane is not precise enough to explain most of the observed geodetic data. The geodetic dislocation model of

Table 4. Target parameters and output results for pure dip-slip fault, rake = 908. Case III: pure dip-slip fault. Source parameters: W = 10 km, Hmin = 0 km, length = 40 km, rake = 458, slip = 1000 cm X 2 20 1 50 0 Y 3 1 20 50 20 Ux 0.40 61.15 1.39 3.71 0. Uy 148.15 86.56 82.54 4.78 82.62 Uz 498.60 322.57 55.59 3.82 55.64

Target.dat used for inverting dip-slip fault.

Inverted results from gas3d code. Var(0) = 10 Var(1) = 0 Var(2) = 40 Var(3) = 90 Var(4) = 997 cux = 0.397364 ux = 0.400000 cux = 60.901086 ux = 61.150000 cux = 1.379846 ux = 1.390000 cux = 3.697617 ux = 3.710000 cux =0.000000 ux =0.000000 cuy = 147.561998 uy = 148.150000 cuy = 86.213788 uy= 86.560000 cuy = 82.207849 uy = 82.540000 cuy = 4.764690 uy = 4.780000 cuy = 82.5700 uy = 82.620000 cuz = 496.602815 uz = 498.600000 cuz = 321.281136 uz = 322.570000 cuz = 55.365323 uz = 55.590000 cuz = 3.808255 uz = 3.820000 cuz = 54.9552 uz = 55.640000

Inverting fault parameters with a genetic algorithm

179

Figure 3. Three-dimensional plot of function z=sin[x+sin(y)], 3<x<3, 3<y<3.

the Northridge earthquake proposed by Wald, Heaton and Hudnut (1996) is shown in Figure 5, as determined by inverting the 46 global positioning system (GPS) stations data gathered by Hudnut

and others (1996). Their inverted model shows that the slip on the fault plane is not uniform. Thus in this study, the authors divided the plane into 6 6 sub-blocks, and the location, dip angle, size of these

Figure 4. Vertical component leveling data (*) used for inverting rupture parameters of Borah Peak Earthquake, and result of inverted mode (solid curve). Thick line represents ruptured fault plane, and arrows show direction of rupture.

180

ndez and J. B. Rundle T. T. Yu, J. Ferna

blocks were determined by either geological or seismic constraints, and only to determine the amount of slip and rake of slip on each sub-block. The total responses due to these 36 sub-blocks were summed at each GPS location and were compared with the measured data as the tting variable. By minimizing the dierence of this variable to 10% of GPS measurement, the inverted results are shown in Figure 6. The pattern and maximum slip on this dislocation prole is similar to the results of Wald.
RESULTS AND DISCUSSION

Figure 5. Prole of coseismic slip inverted by GPS data of Northridge earthquake. Depth is along down-dip direction, and horizontal distance is along strike. Contour slip is in m. Five-pointed star in bottom right is the hypocenter of the mainshock. (After Wald, Heaton and Hudnut, 1996).

The most important issues in the problems of inversion concern accuracy and eciency. The quality of the input data and the capability of the forward model control accuracy. On the other hand, the inverse method, coding technique and the choice of model parameters control eciency. The most popular inversion processes in geophysical applications are linear least squares, and trial-and-error and Monte Carlo methods. The linear least squares method has been established as a standard method for inverse problems because it can show the statistical results of the inversion. However, there must

Figure 6. Coseismic slip prole of Northridge earthquake also inverted from same GPS data set by GA. 6 6 sub-blocks were divided by being assigned various slips and rakes. Sums of responses due to these 36 slips at the surface were compared with GPS data serving as a constraint of the program. Deployment of prole is the same as in Figure 5. Size and direction of each arrow denotes slip in that sub-block.

Inverting fault parameters with a genetic algorithm

181

be a well-posed problem for this method to be used; the data set to be inverted must have dimensions greater than or equal to the model itself. Other disadvantages include the diculty and poor eciency of dealing with large matrices, such as those 10 10 or larger. Trial-and-error methods are straightforward to code but are also the least ecient method for inversion. To guarantee that a possible solution will be reached, a data set with a large range of values is needed. As a result, this method is only feasible for simple models with few points to invert. The eciency of Monte Carlo methods depends on the probability distribution of the exact solution with respect to all the possible solutions. Recall that for an earthquake source image, eight parameters are needed for a simple elastic uniform planar-slip homogeneous model. With the entire data range of these parameters included, the probability of nding the exact solution at a single shot is between 1/1030 to 1/1040, depending on the desired accuracy. GA demonstrates a performance in dealing with multi-point data inversion superior to trial-and-error because it uses a constrained random-walk technique. Bias of the inversion results comes from the assumption made in using Okada's formula to compute the coseismic displacements, not from the code itself. The formula of Okada has been used to compute the displacement eld on many occasions, and the accuracy of this method has also been thoroughly tested. The bias lies in the assumptions made. The question is whether the assumption of a uniform slip over a rectangular fault plane is reasonable. It may also be asked if the use of elastic half-space material can truly reect the displacement eld induced by an earthquake. Yu (1995) made the assumption that using a uniform slip over a rectangular fault plane aects only the near-eld displacement, which is located right above the ruptured fault plane. Accordingly, for the detailed displacements in the near eld, the composite rupture models of Yu (1995), Hudnut and others (1996), and many other authors, have to be considered. The layered half-space model has been used in many displacement models, but it has little eect on the elastic response (coseismic) of surface displacement due to slip at the fault. The nite-element model by Cohen (1980) and others is powerful and has revealed many valuable results. However, the intense long times and lengthy preparation process makes it less than ideal in this kind of inversion computation. As this code computes only the coseismic displacement, the viscoelastic and gravitational eects due to the property of crust can be ignored. In this example, the GA has demonstrated its superiority in inverting (or optimizing) a complex data set to nd a simple model. The process of natural selection in the GA provides an ecient and straightforward code for dealing with such a

complex problem. Unlike other techniques, the GA neither totally relies on prior constraints (manually provided by the programmer) nor runs a trial-anderror scheme. Best of all, several powerful publicdomain GA codes are available, requiring only minor revision for the denition of a ``tness'' function. Another advantage in using the GA is that the selection process rejects automatically those data with a large bias. As a result, only a good quality data set should be used to perform inversion. The inversion results are more accurate and converge faster than with methods. This code has been tested under UNIX and Windows 95 operating systems. The source code is publicly available by anonymous FTP from the server FTP.IAMG.ORG.

Acknowledgments The research of T. T. Y. was partially supported by the Institute of Earth Sciences, Academia Sinica. Republic of China. J. B. R. and T. T. Y. were also supported by NASA grant NAG5-2353 at the University of Colorado. The research of J. F. was supported with funds from project AMB96-0498-C04-04 of the European Union in the framework of the Environment Program. All the authors are grateful to the editor, Dr. Bonham-Carter, and two anonymous reviewers who gave constructive suggestions for improving the quality of this manuscript.

REFERENCES rnado ttir, T. and Segall, P. (1994) The 1989 Loma Prieta A earthquake imaged from inversion of geodetic data. Journal Geophysical Research 99(B11), 2183521855. Cohen, S. C. (1980) Postseismic viscoelastic surface deformation and stress 1. Theoretical consideration, displacement, and strain calculations. Journal Geophysical Research 85(B6), 31313250. Goldberg, D. E. (1989) Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Publishing Co., Reading, Massachusetts, 412 pp. Holdahl, S. R. and Sauber, J. (1994) Coseismic slip in the 1964 Prince William Sound earthquake: a new geodetic inversion. PAGEOPH 142(1), 5582. Hudnut, K. W., Shen, Z., Murry, M., McClusky, S., King, R., Herring, T., Hager, B., Feng, Y., Donnellan, A. and Bock, Y. (1996) Co-seismic displacement of the 1994 Northridge, California, earthquake. Bulletin Seismological Society America 86(1B), S19S36. Michalewicz, Z. (1992) Genetic Algorithm + Data Structures = Evolution Programs. Springer Verlag, Berlin, 340 pp. Okada, Y. (1985) Surface deformation due to shear and tensile faults in a half-space. Bulletin Seismological Society America 75(4), 11351154. Sataka, K. (1993) Depth distribution of coseismic slip along the Nankai Trough, Japan, from joint inversion of geodetic and tsunami data. Journal Geophysical Research 98(B3), 45534565. Segall, P. and Matthews, M. V. (1988) Displacement calculations from geodetic data and the testing of geophysical deformation models. Journal Geophysical Research 93(B12), 1495414966. Stein, R. S. and Barrientos, S. E. (1985) Planar high-angle faulting in the Basin and Range: geodetic analysis of the 1983 Borah Peak, Idaho, earthquake. Journal Geophysical Research 90(B13), 1135511366.

182

ndez and J. B. Rundle T. T. Yu, J. Ferna vations of the 1983, Borah Peak, Idaho, earthquake. Journal Geophysical Research 91(B5), 49094919. Yu, T. T. (1995) Crustal deformation due to dipping fault in elastic gravitational layer overlaying viscoelastic gravitational half space: models and applications. Ph.D. Dissertation, University of Colorado, Boulder, Colorado, 174 pp.

Wald, D. J., Heaton, T. H. and Hudnut, K. W. (1996) The slip history of the 1994 Northridge, California, earthquake determined from strong-motion, teleseismic, GPS, and leveling data. Bulletin Seismological Society America 86(1B), S49S70. Ward, S. N. and Barrientos, S. E. (1986) An inversion for slip distribution and fault shape from geodetic obser-

También podría gustarte