Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
DOI 10.1007/s10950-012-9291-x
ORIGINAL ARTICLE
Received: 8 March 2011 / Accepted: 24 February 2012 / Published online: 21 March 2012
Springer Science+Business Media B.V. 2012
Abstract Despite technological advances in seismic instrumentation, the assessment of the intensity of an earthquake using an observational scale
as given, for example, by the modified Mercalli
intensity scale is highly useful for practical purposes. In order to link the qualitative numbers extracted from the acceleration record of an earthquake and other instrumental data such as peak
ground velocity, epicentral distance, and moment
magnitude on the one hand and the modified
Mercalli intensity scale on the other, simple statistical regression has been generally employed.
In this paper, we will employ three methods
of nonlinear regression, namely support vector
regression, multilayer perceptrons, and genetic
programming in order to find a functional dependence between the instrumental records and the
modified Mercalli intensity scale. The proposed
methods predict the intensity of an earthquake
1 Introduction
The macroseismic intensity is an essential parameter of earthquake ground motion that allows
a simple and understandable description of earthquake damage on the basis of observed effects at a
given place. It is measured, for example, using the
European Macroseismic Scale, the China Seismic
Intensity Scale, MercalliCancaniSieberg scale,
the Modified Mercalli Intensity (MMI) scale, or
the Japan Meteorological Agency (JMA) seismic
intensity scale (see e.g., Grnthal 2011).
The intensity scales quantify the effects of a
strong motion on the Earths surface, humans,
objects of nature, and man-made structures at a
given location based on detailed description of
indoor and outdoor effects that occur during the
490
shaking. For example, the modified Mercalli intensity (MMI) scale (Wood and Neumann 1931;
Richter 1958) measures the effects of an earthquake using 12 degrees, with 1 denoting not felt
and 12 as total destruction (take into account
that the US Geological Survey (USGS) no longer
assigns intensities higher than 10 and has not assigned 10 in many decades; see e.g., Stover and
Coffman (1993), Dengler and Dewey (1998); in
addition, the current practice of using the MMI
scale does neither follow the wording of the version by Wood and Neumann (1931) nor by Richter
(1958) (Grnthal 2011)). The values will differ
based on the distance to the earthquake, with
the highest intensities being usually around the
epicentral area, and based on subjective data that
are gathered from individuals who have felt the
quake at that location. Take into consideration
that all modern scales have 12 degrees with the
exception of the JMA scale which was upgraded
from 7 to 10 degrees in 1996 (see e.g., Musson et al.
2010).
The assessment of the intensity of a quake at
a given location used to be a slow process, as it
was usually performed by means of personalized
surveys; however, with the advent of macroseismic intensity internet surveys (a list of some of
these services can be found in Table 1), this is
not the fact anymore. One can see, for example, that when a strong shaking happens in the
URL
http://asc-india.org/menu/felt-india.htm
http://www.earthquakes.bgs.ac.uk/questionnaire/EqQuestIntro.html
http://www.zamg.ac.at/erdbeben/bebenbericht/index.php
http://www.emsc-csem.org/Earthquake/Contribute/testimonies.php
http://www.geonet.org.nz/earthquake/
http://www.haisentitoilterremoto.it/
http://www.seisme.prd.fr/english.php
http://earthquakescanada.nrcan.gc.ca/dyfi/
http://seismologie.oma.be/index.php?LANG=EN&CNT=BE&LEVEL=0
http://seisan.ingeominas.gov.co/RSNC/index.php
http://www.seismo.ethz.ch/eq/detected/eq_form/index_EN
http://earthquake.usgs.gov/earthquakes/dyfzi/
491
2 Previous studies
and
MMI =
and
MMI =
and
MMI =
In all the above equations, the PGA is expressed in centimeter per square second, while the
PGV is given in centimeter per second.
Karim and Yamazaki (2002) performed a similar research relating PGA, PGV, PGD, and the
so-called spectrum intensity to the JMA seismic intensity scale. Other authors like Atkinson
and Sonley (2000), Sokolov (2002), Kaka and
Atkinson (2004) have developed empirical relationships between response spectra or Fourier acceleration spectra and modified Mercalli intensity.
Shabestari and Yamazaki (2001) developed some
expressions that related the JMA intensity scale
and the MMI scale. Tselentis and Danciu (2008)
derived empirical regression equations for MMI
and for various ground motion parameters such as
duration, cumulative absolute velocity, Housners
spectrum intensity, and total elastic input energy
index.
Note that most of the aforementioned relationships were performed using univariate and multivariate linear regression analysis showing large
scatter. It is the authors belief that those studies
would have been better using robust regression
methods instead of standard linear regression in
order to account for possible outliers, that is, some
observations that do not follow the pattern that
the other observations have and that usually occur when large measurement errors are present.
Our preference for robust regression methods is
based on the fact that, in particular, least squares
estimates for regression models are highly biased
by the outliers that dominate the recordings from
few earthquakes, since the regression might by
dragged towards them.
It seems, however, that the linear model chosen
in the previously mentioned papers comes just
from mathematical convenience. The relationship
that links the MMI scale and the instrumental data
is complicated enough that nonlinear methods
seem to be the best choice for a model. This is
the reason that motivated us to propose a new
approach in Section 5.
492
PGA, PGV, and PGD defined as the geometric mean of the two horizontal components
in the directions originally recorded. The first
parameter is specially important because according to Newtons second law, the PGA
multiplied by the mass provides the maximum
inertial force that affects the structure during
the shaking.
duration of the shaking;
spectral content (characteristic periods)
at some specific frequencies; according to
Sokolov (2002), the frequency range of the
amplitude spectrum |X( f )| of the acceleration
record that is of interest from an engineering
point of view lies between 0.3 and 14 Hz.
In fact, the frequency band 0.782.0 Hz is
representative for MMI greater than 8, while
the 3.06.0 Hz range represents best MMI
from 5 to 7 and the 7.08.0 Hz correlates best
with the lowest MMI.
magnitude (moment magnitude, local magnitude, surface wave magnitude, or body wave
magnitude);
epicentral distance;
amplitudes of acceleration, velocity, or displacement response spectra;
regional propagation path (geological conditions) and local soil conditions among others.
493
n
(8)
i K (x, xi ) + b ,
(7)
i=1
f x
f x
0,
E ( f (x) y) =
0
if | f (x) y| <
| f (x) y|
otherwise
(9)
f x
f x
E z
0,
494
(10)
yi f (xi ; , b ) i
(11)
min C
,b
n
i + i + ()
(12)
i=1
max
,
n
(i + i )
(13)
i=1
an independent dataset. One round of crossvalidation involves partitioning the dataset into
complementary subsets, performing the analysis
on one subset and validating the analysis on
the other subset. To reduce variability, multiple
rounds of cross-validation are performed using
different partitions, and the validation results are
averaged over the rounds.
It can be shown that Eq. 13 is a convex
quadratic optimization problem, and in consequence, it has a unique solution. Those Lagrange
multipliers i which are different from zero are
the so-called support vectors. The support vector
regression has the property that the solution is
sparse, that is, most of the i are zero.
The parameter b is found for a data point for
which 0 < i < C or 0 < i < C by solving the
equation:
b = yn
n
i K (xn , xi )
(14)
i=1
Fig. 3 Topology of a multilayer perceptron with a single output unit. This network has d inputs, p neurons in
the hidden layer, and a single output. w0ji represents the
weights between the j-th neuron of the hidden layer and
495
d
p
= h
w11 j g
w0ji xi + b 0j + b 11
j=1
(15)
i=1
the i-th input, while w11 j stands for the weights between the
output neuron and the j-th neuron of the hidden layer. b 0j
represents the bias weight of the j-th neuron of the hidden
layer and b 11 symbolizes the bias weight of the output layer
496
(16)
execution. The theory behind genetic programming is large. Here, just a brief review of its main
concepts will be given. The interested reader is
referred to Koza (1992) for an ample discussion
on the topic.
Genetic programming uses the following steps
to solve problems:
1. Generate an initial population of computer
programs
2. Iteratively perform the following sub-steps on
the population until the termination criteria is
satisfied:
a. Execute each program in the population
and assign it a fitness value according to
how well it solves the problem
b. Create a new population by executing the
the following evolutionary operators with
certain probability:
Reproduction: it selects an individual
from within the current population so
that it can have an offspring. There
are several forms of choosing which
individual deserves to breed including fitness proportionate selection,
rank selection, and tournament
selection.
Crossover: mimics sexual combination in nature, where two parents are
chosen and parts of their trees are
swapped in a form that each crossover
operation should result in a legal
structure.
Mutation: it causes random changes
in an individual before it is introduced into the subsequent population. During mutation it may happen
that all functions and terminals are
removed beneath an arbitrarily determined node and a new branch is
randomly created or a single node is
swapped for another.
3. The best computer program that appears in
any generation is designated as the result of
genetic programming.
One of the main uses of genetic programming
is to evolve relationships between variables: this
4 Earthquake data
Two earthquake datasets have been used in this
study. The first one is the USGSs Did You Feel
It? database (U.S. Geological Survey 2011) which
collects, by means of internet surveys, information
about how people actually experienced the earthquakes. The form of the questionnaire employed
in the DYFI database and the method for assignment of intensities are based on an algorithm
developed by Dengler and Dewey (1998).
In this dataset, one can find for a given earthquake a table of modified Mercalli intensities aggregated by city or postal code, number of responses for that region and epicentral distance,
and a representative latitude and longitude of the
surveyed region. In addition, it is possible to find
in the same database the depth of the earthquake
and the latitude and longitude of the epicenter.
The second employed database is the one of
the Center for Engineering Strong Motion Data
(CESMD - Center for Engineering Strong Motion
Data 2011). Here, one can find for some representative earthquakes the actual accelerograms of the
shakings measured at different stations. For each
station, one can find its code and name, its latitude
and longitude, and for the given earthquake, its
epicentral distance, magnitude, PGA, PGV, PDG,
and the amplitudes of acceleration response spectra for the 0.3, 1, and 3 s. However, one must
take into consideration that, usually, not all of the
above-mentioned data are available at the same
time in the CESMD database.
120
100
frequency
497
80
60
40
20
100
200
300
400
500
600
700
number of responses
498
MMI
Epicentral distance
PGA
PGV
Moment magnitude
Depth
MMI
Epicentral
distance
PGA
PGV
Moment
magnitude
Depth
1.00
0.32
0.79
0.73
0.26
0.14
0.32
1.00
0.52
0.01
0.66
0.23
0.79
0.52
1.00
0.72
0.01
0.30
0.73
0.01
0.72
1.00
0.55
0.15
0.26
0.66
0.01
0.55
1.00
0.17
0.14
0.23
0.30
0.15
0.17
1.00
499
7
6
5
4
3
2
2
0
50
100
150
200
250
300
350
400
10
Fig. 8 Plot of PGV vs. MMI for the training dataset. The
lines represents the relationship between these variables
shown, given by Eqs. 2, 4, 6, and 24
10
10
6
5
4
3
2
10
10
10
scale]
4.5
5.5
6.5
7.5
magnitude
Fig. 7 Plot of PGA vs. MMI for the training dataset. The
lines represents the relationship between these variables
shown, given by Eqs. 1, 3, 5, and 23
500
80
70
250
60
200
frequency
frequency
50
40
30
150
100
20
50
10
0
1
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
PGA (g)
120
300
100
250
200
frequency
frequency
80
60
40
150
100
20
50
0
0
50
100
150
200
250
300
350
400
0
0
10
15
20
25
30
35
PGV (cm/s)
40
501
140
120
frequency
100
80
60
40
20
0
4
4.5
5.5
6.5
7.5
moment magnitude
Xk min(Xk )
1
max(Xk ) min(Xk )
(17)
zk =
Xk mean(Xk )
std(Xk )
(18)
and employing the means and standard deviations that can be found in Table 3. Even though
this procedure is inspired in the normalization of
Gaussian random variables, it is applicable to any
kind of distribution, since the idea is to reduce
the spread in the data. Both methods are popular in nonlinear regression for making the input
variables rather small in order to improve the
numerical stability of the employed algorithms,
regardless of the distribution of the data. In other
words, this process tends to make the training
502
Table 3 Mean, standard deviation, minimum, and maximum of each variable of the dataset
Variable
Mean
Standard deviation
Minimum
Maximum
MMI
x1 = epicentral distance (km)
x2 = PGA (g)
x3 = PGV (cm/s)
x4 = moment magnitude
3.6425
87.4656
0.0460
3.4752
5.2261
0.9669
102.9749
0.0619
5.4888
0.9574
2.0
1.6
0.002
0.080
4.0
7.5
393.7
0.588
62.88
7.2
The mean and standard deviation correspond to the training set and the minimum and maximum values correspond to the
whole database
process better behaved by improving the numerical condition of the underlying optimization algorithms employed and ensuring that various default
values involved in initialization and termination of
the algorithms are appropriate (see for instance,
Sarle 2002).
We performed the nonlinear regressions with
the algorithms described in Section 3 in the following way:
503
(outputNN+1)(max(MMI)min(MMI))
2
+ min(MMI)
(20)
In Appendix 2.3, we have included the MATLAB code that retrieves the estimated MMI from
the epicentral distance, PGA, PGV, and moment
magnitude using the symbolic regression (19) and
Eq. 20.
Take into account that the form of Eq. 19
should not be used outside of the present
study, inasmuch as the functional form and the
coefficients of this equation depend on the employed database. If this methodology is used with
another dataset, most probably the algorithm will
converge to a different functional form.
(21)
(19)
(22)
504
(23)
and
Here, the PGA must be expressed in centimeter per square second, while the PGV must
be given in centimeter per second. These relationships have been plotted in Figs. 7 and 8,
respectively.
5.2 A note on the usage of the regressions
Note that according to Fig. 10, the MMI in our
database is bounded between 2 and 6.5, and in
consequence, these should be constraints of the
values produced by the regression methods described above. Take into account that the minimum and the maximum values of each variable
should be set as guidelines on the interpolative
power of the presented results, inasmuch as it is
not advisable to extrapolate with them. If that is
6 Analysis of results
Although input MMI levels are discrete natural
numbers, the output is in the form of continuous
real numbers. We define a successful prediction
when the estimation is within 0.5 the MMI level
reported by the DYFI database. In this sense, we
evaluate the performance of the algorithm: the
first part, the loss function (8) with = 0.5, was
evaluated on the testing set (a set of data which
was not employed in the training phase of the
algorithm). For comparison reasons, the quadratic
mean error function (16) is calculated as well.
These numbers, together with the coefficients of
correlation of the predicted MMI vs. the actual
MMI and the percentage of misclassification on
the testing set are shown in Table 4.
In comparison to Eq. 23, which is only dependent of the PGA, the inclusion of more information in the MMI assessment is beneficial
for its prediction. The best nonlinear regression
method seems to be the one produced by MLP,
followed closely by the genetic programming, and
the SVR. In general, the MMI estimation shows
a good agreement with the reported intensity, the
-insensitive loss function, the mean square error,
and the missclassification error obtained with the
nonlinear algorithms are lower in comparison to
the values obtained with the linear regression (23).
Figure 15 illustrates how well the predicted MMI
Table 4 Performance of the different nonlinear regression algorithms on the testing dataset
Algorithm
-insensitive
loss function
Mean square
error
Correlation of predicted
vs. actual MMI
Percentage of
misclassification (%)
SVR
MLP
Genetic programming
Ordinary least squares regression (21)
Robust regression (22)
Equation 23
Equation 24
0.037
0.035
0.033
0.046
0.046
0.067
0.097
0.146
0.141
0.139
0.169
0.170
0.224
0.278
0.928
0.928
0.929
0.913
0.914
0.883
0.853
19.91
17.06
19.43
22.75
21.33
23.22
31.28
505
Regression: R=0.9280
Ideal fit
Data fit
(Actual,Predicted)
Predicted MMI
7 Conclusions
Actual MMI
In this paper, we presented three nonlinear regression methods, namely support vector regression,
multilayer perceptrons, and genetic programming
to model the relationship between the modified
Mercalli intensity scale and the earthquake moment magnitude, epicentral distance, PGA, and
PGV measured at the closest stations to the MMI
reading. In general, the MMI estimation shows
a good agreement with the reported intensity.
The best results were obtained by the multilayer
perceptron.
As seen from the results, nonlinear regression
should be applied in order to find a relationship
between MMI and instrumental information instead of the linear regressions that are popular in
this class of studies. Our numerical experiments
have shown for example that all of the nonlinear
regression algorithms employed perform better
than the linear regressions (21) and (22).
506
4.4895
405.5060
494.5590
45.3807
7.2997
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
10.7370
131.7847
83.2952
494.5590
277.7219
494.5590
494.5590
494.5590
494.5590
388.3332
494.5590
494.5590
494.5590
494.5590
249.7611
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
274.6066
494.5590
494.5590
494.5590
494.5590
494.5590
z1
0.6223
0.6048
0.1701
0.1353
0.0990
0.0249
0.6436
0.3647
0.2665
0.2675
0.2684
0.3014
0.6561
0.6358
0.3140
0.3875
1.8688
1.8978
2.1534
0.7113
0.6348
0.6232
0.5893
0.5777
0.3328
0.0646
0.2065
0.2539
0.7936
0.6813
0.6736
0.6600
0.6048
0.5932
0.5341
0.5003
0.7791
0.7752
0.7462
0.7239
0.6639
0.6329
0.3366
z2
1.5550
0.6038
0.6708
0.6178
1.3207
0.5704
0.4700
0.5536
0.6206
0.5871
0.6373
0.4198
0.1520
0.0823
0.5202
0.5536
0.6541
0.5202
0.6708
0.6346
0.0656
0.2664
0.3835
0.1492
0.0990
0.1353
0.2691
0.2859
1.1701
0.3361
0.4532
0.5536
0.6206
0.5704
0.6875
0.6541
0.4867
2.9440
0.1185
0.1353
0.6541
0.5704
0.5704
z3
2.7381
0.5537
0.5894
0.7897
2.8716
0.5443
0.4973
0.5462
0.5274
0.5199
0.5518
0.5462
0.4165
0.2211
0.4936
0.5593
0.4823
0.1948
0.3451
0.0382
0.3527
0.3395
0.0783
0.3132
0.2794
0.4372
0.3959
0.4729
0.0182
0.4729
0.4071
0.5199
0.5537
0.5537
0.6044
0.5744
0.4898
0.8085
0.3771
0.4259
0.6044
0.5086
0.5499
z4
2.0587
0.2432
0.2432
1.9541
1.9541
0.1386
1.0803
0.3478
0.3478
0.3478
0.3478
0.3478
0.8710
0.8710
0.1753
0.1753
1.3263
1.3263
1.3263
0.3478
0.3478
0.3478
0.3478
0.3478
0.3478
0.3478
0.3478
0.3478
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
1.0803
0.5571
SV
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
494.5590
494.5590
116.8932
494.5590
494.5590
126.2576
489.1067
494.5590
24.8224
242.1119
494.5590
494.5590
209.7443
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
10.5189
494.5590
494.5590
23.3037
7.0743
494.5590
494.5590
494.5590
494.5590
494.5590
178.8668
494.5590
123.2760
10.9674
494.5590
66.9874
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
507
z1
0.7045
0.6358
0.6329
0.6116
0.8890
0.7433
0.6978
0.6600
0.6552
0.6436
0.5738
0.5012
0.4916
0.4402
0.3831
0.3802
0.3260
0.3202
0.2104
0.8323
0.7162
0.2602
0.0326
0.0206
0.8139
0.7045
0.3153
0.8101
0.7384
0.7239
0.6900
0.6852
0.0768
0.7839
0.7588
0.5361
0.4548
0.3231
0.4490
0.8120
0.7278
0.7084
0.6949
0.6949
0.6910
0.6871
0.6803
z2
1.5717
0.0488
0.0656
0.3668
0.7043
1.4713
0.4672
1.3040
4.0318
1.9567
1.1199
0.8187
1.1032
0.5676
0.0656
0.0851
0.3361
0.3696
0.2691
1.6052
0.0321
0.6373
0.6206
0.6038
1.0864
0.6206
0.6541
2.5424
2.6763
2.0069
0.9860
2.0403
0.5704
0.0014
0.1492
0.4365
0.6206
4.7347
0.5704
3.9649
0.0321
0.1158
1.9064
2.1240
0.4505
0.0321
0.1185
z3
0.3463
0.1171
0.0757
0.0532
0.6044
1.6597
0.7503
0.5981
3.7020
0.4553
0.6807
0.2531
1.0509
0.0551
0.0720
0.0062
0.3959
0.3921
0.2380
0.4515
0.4616
0.5894
0.5744
0.5669
0.2298
0.5631
0.6063
0.7390
0.9006
0.7052
0.0250
0.2016
0.5631
0.3357
0.3827
0.5481
0.6025
4.8181
0.5988
0.6751
0.4691
0.4015
0.3407
0.3839
0.1516
0.2907
0.2474
z4
0.3846
0.3846
0.3846
0.3846
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.7664
0.7664
0.7664
0.7664
0.7664
1.2895
1.2895
0.8710
0.5571
0.5571
0.5571
0.5571
0.5571
0.5571
1.2895
1.2895
1.1849
1.1849
1.3263
0.9756
0.8710
0.8710
0.8710
0.8710
0.8710
0.8710
0.8710
0.8710
508
SV
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
0.0167
12.2687
159.9334
289.0971
494.5590
494.5590
494.5590
494.5590
415.6563
494.5590
494.5590
494.5590
494.5590
427.5960
494.5590
494.5590
10.2066
494.5590
494.5590
392.9353
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
494.5590
331.6516
494.5590
494.5590
494.5590
494.5590
z1
0.6542
0.6223
0.6155
0.5913
0.5680
0.5554
0.2534
0.8823
2.2667
2.4081
2.4894
2.5833
2.5852
2.6346
0.5240
0.2781
1.2656
1.6171
0.6252
0.1106
0.1116
0.1164
0.4292
0.6170
0.7438
1.2540
0.5554
0.4296
0.2830
0.2727
0.4718
0.7171
0.5787
0.2249
0.2597
0.5496
0.7849
0.7646
0.7626
0.4170
0.2844
0.2350
0.2147
0.7462
0.7926
0.2253
0.1585
z2
0.3528
0.3696
0.6038
0.4867
0.3863
0.2691
2.3918
0.0823
0.4867
0.6206
0.6206
0.6206
0.6206
0.6038
0.6038
0.5704
0.6708
0.6373
0.6513
0.4198
0.5871
0.4867
0.6206
0.4030
0.6373
0.6373
0.6875
0.6373
0.6206
0.5704
0.5704
0.4839
0.9191
0.2859
0.5536
0.6373
0.3026
0.4672
0.3361
0.6206
0.4532
0.6373
0.6206
0.2859
1.2203
0.7684
0.1492
z3
0.4823
0.4466
0.5932
0.5481
0.4992
0.4278
5.5809
0.9720
0.1039
0.0032
0.0344
0.1434
0.0577
0.0551
0.5781
0.4259
0.5781
0.5199
0.4121
0.3489
0.4842
0.4391
0.5499
0.3827
0.4616
0.4710
0.6082
0.5932
0.5650
0.5687
0.5293
0.0595
0.2662
0.4992
0.5255
0.5744
0.3733
0.0633
0.4541
0.5875
0.4691
0.5838
0.6007
0.5274
0.1742
0.5022
0.2662
z4
0.8710
0.8710
0.8710
0.8710
0.8710
0.8710
2.0587
2.0587
2.0587
2.0587
2.0587
2.0587
2.0587
2.0587
0.7664
0.4892
0.4892
0.4892
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
0.1753
1.1849
1.1849
0.3478
0.1753
0.1753
0.3478
0.3478
0.3478
0.3478
1.0803
1.0803
1.0803
1.0803
0.6617
0.7664
0.4525
0.8710
1.0803
0.5571
1.3263
0.6985
SV
138
139
140
141
142
143
144
145
146
147
494.5590
494.5590
494.5590
12.4008
116.9598
494.5590
494.5590
494.5590
494.5590
494.5590
509
z1
0.6861
0.6193
0.5613
2.6094
2.6859
2.9134
0.4625
0.2514
0.8658
0.6484
z2
0.1687
0.1185
0.6038
0.5704
0.6206
0.6541
0.5704
0.2859
0.6541
0.5536
z3
0.3771
0.4673
0.5706
0.8592
0.2888
0.1114
0.5481
0.2324
0.5255
0.5593
z4
0.8710
0.8710
0.8710
2.0587
2.0587
2.0587
0.7664
0.1753
0.1753
1.1849
510
References
Atkinson GM, Kaka SI (2007) Relationships between felt
intensity and instrumental ground motion in the central United States and California. Bull Seismol Soc
Am 97(2):497510
Atkinson GM, Sonley E (2000) Empirical relationships between modified Mercalli intensity and response spectra. Bull Seismol Soc Am 90(2):537544
Bishop CM (2007) Pattern recognition and machine learning. Springer, NY
Bommer JJ, Stafford PJ, Alarcn JE, Akkar S (2007) The
influence of magnitude range on empirical groundmotion prediction. Bull Seismol Soc Am 97(6):2152
2170
Center for Engineering Strong Motion Data (2011) Internet data reports. http://www.strongmotioncenter.org/.
Accessed 15 Jan 2011
Chang CC, Lin CJ (2011) LIBSVM: a library for support
vector machines. Software available at http://www.
csie.ntu.edu.tw/cjlin/libsvm. Accessed 15 Jan 2011
Davenport P (2004) Neural network analysis of seismic
intensity from instrumental records. In: Proceedings of
511
Tung ATY, Wong FS, Dong W (1993) A neural networks
based MMI attenuation model. In: National earthquake conference: earthquake hazard reduction in the
central and eastern United States: a time for examination and action. Memphis, Tennessee, US
US Geological Survey (2011) Did you feel it? database.
http://earthquake.usgs.gov/earthquakes/dyfi/.
Accessed 15 Jan 2011
Wald D, Quitoriano V, Dengler L, Dewey JM (1999a) Utilization of the internet for rapid community intensity
maps. Seismol Res Lett 70(6):680697
Wald DJ, Quitoriano V, Heaton TH, Kanamori H (1999b)
Relationships between peak ground acceleration,
peak ground velocity, and modified Mercalli intensity
in California. Earthq Spectra 15(3):557564
Wald DJ, Quitoriano V, Dewey J (2006) USGS did you
feel it? community internet intensity maps: macroseismic data collection via the internet. In: Proceedings of the first European conference on earthquake
engineering and seismology. Geneva, Switzerland
Wood HO, Neumann F (1931) Modified Mercalli intensity
scale of 1931. Bull Seismol Soc Am 21:277283
Worden CB, Gerstenberger MC, Rhoades DA, Wald DJ
(2012) Probabilistic relationships between groundmotion parameters and modified Mercalli intensity in
California. Bull Seismol Soc Am 102:204221