See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/226657531

Seismic Hazard Analysis Using the Adaptive Kernel Density Estimation

Technique for Chennai City

Article in Pure and Applied Geophysics · February 2012

DOI: 10.1007/s00024-011-0264-8

CITATIONS READS

21 330

2 authors:

Chethanamba Ramanna G. R. Dodagoudar

7 PUBLICATIONS 30 CITATIONS
Indian Institute of Technology Madras
127 PUBLICATIONS 1,294 CITATIONS
SEE PROFILE
SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Geoenvironmental Engineering Collaboration Work View project

Performance based seismic analysis of bridge systems with rocking foundations View project

All content following this page was uploaded by Chethanamba Ramanna on 13 July 2017.

The user has requested enhancement of the downloaded file.

Pure Appl. Geophys.
Ó 2011 Springer Basel AG
DOI 10.1007/s00024-011-0264-8 Pure and Applied Geophysics

Seismic Hazard Analysis Using the Adaptive Kernel Density Estimation Technique
for Chennai City
C. K. RAMANNA1 and G. R. DODAGOUDAR1

Abstract—Conventional method of probabilistic seismic haz- designing earthquake resistant structures. The first
ard analysis (PSHA) using the Cornell–McGuire approach requires
identification of homogeneous source zones as the first step. This
and foremost step in PSHA using the conventional
criterion brings along many issues and, hence, several alternative technique of Cornell–McGuire approach (CORNELL,
methods to hazard estimation have come up in the last few years. 1968; MCGUIRE and ABRABASZ, 1990; REITER, 1991) is
Methods such as zoneless or zone-free methods, modelling of
the identification of the sources, viz., point, line
earth’s crust using numerical methods with finite element analysis,
have been proposed. Delineating a homogeneous source zone in (faults) and seismic area source zone, and charac-
regions of distributed seismicity and/or diffused seismicity is rather terizing them in terms of maximum magnitude and
a difficult task. In this study, the zone-free method using the the activity rate. In the case of delineation of seismic
adaptive kernel technique to hazard estimation is explored for
regions having distributed and diffused seismicity. Chennai city is source zone, it is required that the zone has uniform
in such a region with low to moderate seismicity so it has been used seismicity, i.e., the relative likelihood of small to
as a case study. The adaptive kernel technique is statistically large earthquakes [b value of the Gutenberg–Richter
superior to the fixed kernel technique primarily because the
bandwidth of the kernel is varied spatially depending on the clus-
(G–R) recurrence law] is uniform. This condition
tering or sparseness of the epicentres. Although the fixed kernel puts forward a greater challenge in terms of dividing
technique has proven to work well in general density estimation the area into zones, thereby introducing subjectivity
cases, it fails to perform in the case of multimodal and long tail
and hence more time and effort.
distributions. In such situations, the adaptive kernel technique
serves the purpose and is more relevant in earthquake engineering A brief compilation of the problems relating
as the activity rate probability density surface is multimodal in seismic area zoning can be summarized as follows:
nature. The peak ground acceleration (PGA) obtained from all the requires expert knowledge of geology, seismology
three approaches (i.e., the Cornell–McGuire approach, fixed kernel
and adaptive kernel techniques) for 10% probability of exceedance
and tectonics of the region (Reiter, 1991; KRAMER,
in 50 years is around 0.087 g. The uniform hazard spectra (UHS) 1996); the aerial extent to which G–R recurrence law
are also provided for different structural periods. is applicable is unknown as it is not a property of a
small region (ALLEN et al., 1965; KRINITZSKY, 1993) or
Key words: Distributed seismicity, diffused seismicity,
kernel density estimation, adaptive kernel density estimation, one particular fault (BAK, 1996); there would be an
probabilistic seismic hazard, PGA, uniform hazard spectra. abrupt change of seismicity at zonal boundary
(BENDER, 1986); homogeneous assumption of seis-
micity within the zone causes higher value of hazard
(BEAUVAL et al., 2006b); and difficulty in fitting the
1. Introduction G–R recurrence law for low to moderate seismicity
region due to scarcity of seismicity data (BEAUVAL and
Probabilistic Seismic Hazard Analysis (PSHA) SCOTTI, 2003). In order to overcome these drawbacks,
has been widely used to determine the peak ground several researchers have developed various alternative
acceleration or the uniform hazard spectra (UHS) for methods of PSHA of which the techniques of zone-
free or spatial smoothing can be found in the literature
(FRANKEL, 1995; WOO, 1996; STOCK and SMITH 2002a,
1 b; LAPAJNE et al., 2003; ZOLFAGHARI, 2009).
Department of Civil Engineering, Indian Institute of
Technology Madras, Chennai 600 036, India. E-mail: ce07d001@ The present study is the extension of the zone-free
smail.iitm.ac.in; [email protected] method to PSHA of WOO (1996) and the technique
 C. K. Ramanna, G. R. Dodagoudar Pure Appl. Geophys.

used to determine the earthquake occurrence model In nonparametric methods no prior assumption
of STOCK and SMITH (2002a, b) to obtain the expected regarding the distribution is made and the distribution
ground motion at a particular site. WOO (1996) has is determined from the data in hand. There are several
used the fixed kernel technique to determine the methods for nonparametric density estimation, start-
spatial activity rate around the site of interest by ing with the simplest histogram to more sophisticated
placing fixed kernels whose bandwidths are a func- methods such as kernels, splines and wavelets.
tion of magnitude on every epicentre. This form of The kernel density estimation technique consists
uncertainty in activity rate is integrated with other of placing a density curve of any form, such as tri-
forms of uncertainty such as earthquake magnitude angular, normal, quartic, uniform or Epanechnikov,
and distance. In the work of STOCK and SMITH (2002a, on each data/sample point and density at any point is
b) the earthquake occurrence model has been deter- calculated as the normalized sum effect of each of
mined using adaptive kernel density estimation these kernels. A multivariate fixed bandwidth kernel
technique to smoothen the activity rate irrespective of density function is of the form:
the magnitude. The activity rate in their work is  
1 X n
1
determined as the number of earthquakes above cut- f ðxÞ ¼ d K ðx  xi Þ
nh i¼1 h
off magnitude divided by the number of completeness
years for every cell in the grid formed around the site where h is the fixed bandwidth, xi is the sample/data
of interest. In the present study, the fixed kernels on points in d-dimensional space, x is the point at which
every epicentre is replaced with adaptive kernels to density is estimated and the kernel K(.) is a sym-
determine the spatial activity rate and integrated with metrical unimodal standardised density.
other forms of uncertainty in magnitude and distance The most important parameter in the kernel tech-
to determine the probability of exceedance of the nique is the choice of the correct bandwidth h rather
selected ground motion. This methodology of seismic than the type of the kernel (SILVERMAN, 1986; de SMITH
hazard analysis has been adopted for Chennai city et al., 2009). However, choosing larger bandwidth
and the results are provided in the form of UHS tends to smoothen the features (Fig. 1a), while smaller
for return periods of 72, 475, 975 and 2,475 years bandwidth causes the PDF to be spiky (Fig. 1b). Also,
indicating the 50, 10, 5 and 2% probability of when a very small value of h is used, the variance
exceedances in 50 years respectively. becomes larger; however, reducing the bias, while
choosing a larger value of h reduces the variance at the
expense of increased bias or error. Hence, an optimal
2. Nonparametric Density Estimation Technique bandwidth has to be chosen for which several methods
Using Fixed and Adaptive Kernels are available in the literature (SILVERMAN, 1986;
SCOTT, 1992; WAND and JONES, 1995) and are sum-
The probability distribution of a random variable marized by MUGDADI and AHMAD (2004), viz. the least-
X is given as square cross-validation, biased cross-validation, plug-
in bandwidth selection, smoothed cross-validation,
Zb
root-n bandwidth selection and the contrast methods.
Pða\X\bÞ ¼ f ð xÞdx for all a\ b
Emphasizing more on the bandwidth, there are tech-
a
niques wherein the bandwidth is varied spatially
where f(x) is the probability density function (PDF), known as variable kernel or adaptive kernel. In the
a and b are the selected lower and upper bounds. fixed kernel technique, the bandwidth remains the
There exist broadly two categories of density esti- same for each kernel, whereas in the adaptive kernel
mation—parametric and nonparametric methods. In technique, the bandwidth is varied on the principle
parametric methods a distribution is assumed and the that smaller bandwidth is more appropriate in regions
parameters of the distribution are estimated from the of high density since a large number of samples
data. Goodness of fit tests is carried out to check enables a more accurate estimation of density in these
the assumption made regarding the distribution type. regions. On the other hand, a larger bandwidth is more
 Seismic Hazard Analysis

(a) Final density estimate

Individual kernels
0.45
0.4
Density Estimate
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
4 6 8 10 12 14 16
X

(b) Final density estimate

Individual kernels
0.9
0.8
0.7
Density Estimate

0.6
0.5
0.4
0.3
0.2
0.1
0
6 7 8 9 10 11 12 13 14 15 16
X

Figure 1
a Large bandwidth kernels. b Small bandwidth kernels

appropriate in low density regions where few sam- SILVERMAN (1986) is adopted. It is based on the
ple points are available (BRUNSDON, 1995). Hence, square-root law of ABRAMSON (1982). Further, there
statistically the principle on which the adaptive exist two types of estimators for adaptive kernel
kernel technique is based is more appropriate but is density estimation techniques—balloon estimator
advantageous for smoothening long tail distributions and sample point estimator. In the balloon estimator,
where under smoothing in the tails is likely to cause the bandwidth of the density curves or kernels is
difficulties (SILVERMAN, 1986). This is because in the varied at the estimation point x, whereas in the
fixed kernel density estimator, the window width is sample point estimator, the bandwidth is varied at
fixed across the entire sample and, hence, a ten- the sample or data points xi. The balloon estimator
dency of spurious noise appears in the tails and if suffers from discontinuities, bias problems and
bandwidth is adjusted accordingly to deal with this integration to infinity (MITTAL and PARAGIOS, 2004),
problem; then the essential details are lost due to therefore the sample point estimator is adopted in
over smoothening in the main part. There are sev- this study.
eral techniques of varying the bandwidth, however, The PDF using the sample point kernel density
in the present study the procedure suggested by estimator in d-dimensional space is given as
 C. K. Ramanna, G. R. Dodagoudar Pure Appl. Geophys.

 
1X n
1 1 et al., 2008) and for an LNG plant at Taranto (PRIN-
f ðxÞ ¼ K ðx  x i Þ CIPIA, 2005). CHEN et al. (1998) have used the kernel
n i¼1 ½hðxi Þd hðxi Þ
smoothing technique to define influence area for each
The three step procedure of determining the earthquake to determine the maximum earthquake at
adaptive kernel density estimation is as follows: a site. The kernel techniques have also been used for
source site characterisation in mining induced seis-
1. Find a pilot estimate of f(x) that satisfies f(xi) [ 0
micity (STAINSLAW and SIKORA, 2008). A modified
for all i.
version of the WOO (1996) technique can be found in
2. Define local bandwidth factor ki by
the work of CHAN and GRUNTHAL (2008) known as a
ki ¼ ff ðxi Þ=gga hybrid method wherein the fixed kernel method was
where g is the geometric mean of f(xi) applied after forming source zones (based solely on
X geology) to arrive at a uniform seismic hazard map
log g ¼ n1 log f ðxi Þ for all of Europe. The first work on adaptive kernel
application to earthquake engineering was by STOCK
and a is the sensitivity parameter (0 B a B 1).
and SMITH (2002a, b).
3. Define the adaptive kernel estimate f(x) by The probability of ground motion parameter at a
Xn  1 1  given site, Y, exceeding a specified level, y, during a
f ðxÞ ¼ n1 hd kd
i K h ki ðx  xi Þ ð1Þ specified time, t is given by
i¼1
PðY [ yÞ ¼ 1  eky t  ky t
where h is the global bandwidth. When a is 0, the
estimate reduces to fixed kernel technique and a value where ky* is (mean annual rate of exceedance) the
closer to 1 causes the estimate to become more sen- average frequency during time period t at which the
sitive to the variation in the pilot estimate and also level of ground motion parameter, Y, exceeds level y
causes large variation in the bandwidth from point to at a given site. The parameter ky* incorporates the
point. A value of a = 0.5 has been proven to give uncertainty in time, size and location of future
good results (SILVERMAN, 1986; ABRAMSON, 1982). earthquakes and uncertainty in the level of ground
The above three step procedure has been used by motion they produce at the site. It is given as
STOCK and SMITH (2002a, b) to arrive at the earth- ZZ
XNS
quake occurrence model by smoothing the annual ky ¼ ti P½Y [ y jm; r fMi ðmÞfRi ðrÞ dm dr
activity rate for Australia and New Zealand. i¼1

where Ns is the number of potential earthquake

sources each of which has an average rate of
3. Kernel Technique for Seismic Hazard Analysis
threshold magnitude exceedance ti also known as
activity rate, P[Y [ y*|m,r] is obtained from the
The kernels have been used in various forms in
predictive relationship and fM(m) and fR(r) are the
seismic hazard estimation. The very first researcher to
PDF for magnitude and distance respectively. For
use kernels in a vague sense was BENDER (1986) to
simplicity rather than efficiency, dividing the possible
handle the abrupt change of seismicity rate at zonal
ranges of magnitude and distance into NM and NR
boundary where each point in source zone was
segments respectively, the total mean annual rate of
regarded as normally distributed. WOO (1996) used
exceedance ky* of the selected ground motion
the fixed kernel technique to replace the mean annual
parameter Y exceeding a particular value y* is given
rate of exceedance for magnitude M of G–R recur-
by
rence law tm with spatial activity rate t(m,x). A
detailed study of this technique can be found in the NS X
X NM X
NR    
works of MOLINA et al. (2001) and BEAUVAL et al. k y ¼ ti P Y [ y jmj ; rk P M ¼ mj
i¼1 j¼1 k¼1
(2006a). Further, this technique has been applied to
develop UHS for the Pyrenean region (SECANELL P½ R ¼ r k  ð2Þ
 Seismic Hazard Analysis

where NM is the range of magnitudes, NR is all the each earthquake event within the bin, the distance to
possible range of distances from site to source, the nearest epicentre is determined. The mean nearest
P[M = mj] and P[R = rk] are obtained from the PDF distance for each bin is obtained and through a least-
of magnitude and distance respectively. In the kernel squares fit between the magnitude and bandwidth, the
technique, the seismicity rate ti is replaced by the parameters c and d are obtained.
spatial activity rate density function and is given as The PSHA using the adaptive kernel technique
consists of two parts. Firstly, determine the activity
X
N
KðM; x  xi Þ
tðM; xÞ ¼ ð3Þ rate t(M,x) using the adaptive kernel technique and
i¼1
Ti secondly, integrate t(M,x) with other forms of
where N is the number of earthquake events, x is the uncertainty to determine ky* as per (Eq. 2). The latter
observation/estimation point, xi is the epicentral part is as per WOO (1996) methodology, whereas for
location and Ti is the effective return period evaluated the first part, we have used the three step procedure of
using the following expression: SILVERMAN (1986) for determining the spatial proba-
X bility density surface, i.e., the spatial activity rate
Ti ¼ time period pi density function t. In (Eq. 3), although K() is a
i
function of magnitude M and space x, the kernel size
where pi is the detection probability of the event in a is fixed for a particular magnitude since h is a func-
particular time period and a numerical value is tion of only magnitude (Eq. 5). In adaptive kernel
assigned to it based on the seismicity of the region. technique, h is a function of both magnitude and
VERE-JONES (1992) suggested the use of an space, i.e., h(M,xi) hence K() varies spatially for each
anisotropic kernel for earthquakes and is given as magnitude M.
 ! In the first step of determining the pilot estimate,
2 n
n  1 1 þ d cos2 / r the fixed kernel density estimation technique of WOO
KðM; rÞ ¼ 2 1þ
ph ðMÞ 1 þ ðd=2Þ hðMÞ (1996) has been used. The bandwidth for this esti-
ð4Þ mate is determined as per (Eq. 5). The local
bandwidth factor for each epicentre is then deter-
where n is the exponent of the power law or fractal mined as per Step 2 given in the earlier section and
scaling index taking value between 1.5 and 2, h is the the local bandwidth h(M,xi) is determined as the
bandwidth and is a function of magnitude, r is the product of global bandwidth h(M) and local band-
distance to the epicentre, i.e., (x - xi), the parameter width factor ki. In Step 3, the adaptive kernel density
/ is the angle subtended at r between the intersection is estimated as per (Eq. 1).
of the fault plane with the earth’s surface and the
epicentre location and d is the degree of anisotropy
taking value between 0 and 2. A value of zero indi- 4. Geological and Tectonic Setting of Chennai
cates isotropy and higher value signifies anisotropy.
Anisotropic kernel is useful when the activity rate Chennai city (13.0833°N, 80.2833°E) is located in
associated with a particular fault needs to be deter- the southern part of Peninsular India (PI) on the
mined. However, in this study, isotropic kernel was Coromandel Coast of Bay of Bengal. It is the fourth
used for Chennai city as it lies in the southern part of most populous metropolitan (with 6.5 million people
Peninsular India which is known for its distributed in a 1,189 km2 area) and the fifth most populous city
and diffused seismicity. The bandwidth of the kernel (with 4.34 million people in a 176 km2 area) in India
for magnitude M is determined as according to the 2001 census. It is one of the fastest
growing cities in the country with many industries
hðMÞ ¼ ceðdMÞ ð5Þ
emerging in various sectors such as information
where parameters c and d depend on the spatial dis- technology, telecom, and automobile. The seismic
tribution of earthquake epicentres. They are map drawn by the Bureau of Indian Standards (IS
calculated by forming various magnitude bins and for 1893: 2002) has shifted Chennai from Zone II (lower
 C. K. Ramanna, G. R. Dodagoudar Pure Appl. Geophys.

activity zone) to Zone III (higher activity zone) due to lineament have been observed. A numerous small
the importance the city is gaining as well as due to earthquakes of maximum Mw = 4.6 have been
the increased seismicity of the PI (earthquakes such observed around Bangalore (12.97°N 77.58°E) region.
as Latur 1993 Mw = 6.1, Jabalpur 1997 Mw = 5.8 Also, near Kanchipuram (12.833°N 79.75°E) a pre-
and Bhuj 2001 Mw = 7.7). instrumental earthquake of magnitude Mw = 5.05 in
Peninsular India is considered a Precambrian stable 1823 has been observed. The major earthquakes stated
continental region with the seismicity being low to above were located on or near the faults or fault
moderate. Some parts of PI are known for diffused junctions whereas other earthquakes are diffused in
seismicity (RAO, 1992; WALLING and MOHANTY, 2009) nature.
and some for distributed seismicity. CHANDRA (1977), For seismic hazard analysis an influence area of
in his study of seismicity of PI, formed a large seismic 300 km radius around Chennai is considered, of
zone comprising the entire Kerala, Tamil Nadu and which major earthquake zones such as Ongole,
parts of southern Karnataka, mainly because, ‘‘the Bangalore and regions closer to Coimbatore lie on the
irregular distribution of earthquake activity in the periphery, whereas the Pondicherry region and Kan-
region made it difficult to relate epicentres to specific chipuram lie closer to Chennai. Figure 2 shows the
structures’’, indicating distributed seismicity. Such broad tectonic settings of PI (RAO, 2000) along with
zonation pattern can be found in the works of GUPTA the influence area (300 km radius) around Chennai
et al. (1995) and GUPTA (2006). In the southern part of comprising the Southern Granulite Terrain (SGT)
PI very rare earthquakes of magnitude greater than 5.0 Craton, the Transition Zone (TZ), the Dharwar Cra-
have occurred, with a maximum historic event being ton, the Cuddapah Basin (CB) and the Eastern Ghats
Mw = 6.0 (intensity VII) which occurred on 8 Feb- (EG). The maximum observed seismicity and strain
ruary 1900 in Coimbatore at a focal depth of around rate per year computed by RAO (2000) for these
70 km (DASGUPTA et al., 2000; GUPTA, 2006). Earth- seismotectonic regions are summarized in Table 1.
quakes in these regions are due to intraplate stress with These regions are considered to be the low strain rate,
pre-existing weak zones. SUBRAHMANYA (1995) seismicity, and heat flow regions as compared to
observed a major compressional lineament known as other parts of the PI. From the brief review of the
the Mulki-Pulicat Axis (MPA) running from Mulki geological and seismotectonic settings of the study
(approximately 13°N) on the east coast to Pulicat Lake region, it can be concluded that the city lies in a
on the west coast (very close to Chennai city). This moderate seismicity region and there exist good
compression zone is the result of continuous spreading earthquake records for the region. The earthquakes
of the sea floor in the Indian Ocean. The author further are due to intraplate stress and the seismicity is dis-
concluded that a large number of small earthquakes tributed and diffused, hence delineation into area
occur south of MPA whereas major earthquakes can be source zones will bring in subjectivity. The seismicity
noticed only in the north of MPA as it acts as a major is low in most part of the influence area (in the off-
block with significant stress accumulation. Seismi- shore region) and hence the assumption of
cally active regions like Ongole (15.60°N 80.10°E) homogeneity will affect the hazard values. With these
Mw = 5.2 in 1959 and 1967 and Coimbatore (10.8°N points in mind, an attempt is made in this study to
76.8°E) Mw = 6.0 in 1900 are being observed on the explore the applicability of zone-free technique for
Gundlakamma Fault and Cauveri Fault, and the swarm hazard analysis, in specific, adaptive kernel technique
of NNW trending basement faults in the region (GSI, due to its statistical edge over the fixed kernel
2000) respectively. Also, a few major earthquakes technique.
around Pondicherry (11.93°N 79.83°E) with maxi-
mum Mw = 5.6 in 1867 with its epicentre close to the
junction of Tirukkavillur–Pondicherry Fault and 5. Seismic Hazard Analysis for Chennai
Crystalline-Sedimentary Contact Fault and the other
off coast with Mw = 5.5 in 2001 on the offshore The objective of the present study is to test the
extension of the fault-plane Pondicherry–Kambam performance of the adaptive kernel technique with
 Seismic Hazard Analysis

Figure 2
Broad tectonic units in Peninsular India along with significant earthquakes that occurred with Mw = 6.0 and above [modified after (RAO,
2000)]

the fixed kernel technique and the conventional Table 1

Cornell–McGuire methodology for low to moderate Maximum magnitude and strain rate per year for different
seismicity regions of India such as Chennai. The seismotectonic units within the influence area [modified after RAO,
seismological data required for the hazard analysis 2000]

has been compiled from sources like India Meteoro- Seismotectonic unit Maximum Strain rate ranges
logical Department (IMD), US Geological Survey magnitude
Moment Tensor and Broadband Source Parameter Cuddapah Basin (CB) 3.5 (1.2–1.4) 9 10-10
Search website (USGS), The National Earthquake Southern Granulite Terrain 6.0 (2.4–4.6) 9 10-10
Information Centre (NEIC), The Harvard Seismology (SGT)
Transition Zone (TZ) 5.0 (3.3–6.5) 9 10-10
Central Moment Tensor Project, The International Eastern Ghats (EG) 5.7 (8.0–11.0) 9 10-10
Seismological Centre (ISC), JAISWAL and SINHA Dharwar Craton (without 6.3 (4.0–6.1) 9 10-10
(2007), CHANDRA (1977), RAO and RAO (1984) and Koyna events)
RAO (2000). A total of 173 earthquakes of Mw C 3.5
were compiled for a circular area of 300 km radius
around Chennai from the year 1507 to 2008 A.D. The kernel techniques. In the present study only the his-
minimum magnitude considered for analysis is 4.0. toric and instrumental seismicity in the form of
Foreshocks and aftershocks were removed using distributed seismicity has been considered, however,
GARDNER and KNOPOFF (1974) dynamic windowing neotectonic data can be included to the catalogue
technique which resulted in 151 mainshocks. This which suggests the occurrence of earthquakes with a
Poissonian catalogue has been used for the hazard magnitude range beyond the historical catalogue.
analysis. The spatial distribution of epicentres around Also if major faults causing earthquakes exist, then
Chennai after removal of dependent events is shown anisotropic form of kernel can be used in determining
in Fig. 3. A magnitude range of Mw ± 0.49 is con- the activity rate (Eq. 4). However, no such major
sidered while determining the activity rate for every faults such as the San Andreas Fault in California
magnitude bin while performing the PSHA using the exist around Chennai.
 C. K. Ramanna, G. R. Dodagoudar Pure Appl. Geophys.

Figure 3
Poissonian earthquake epicenters in the influence area around Chennai city

Due to non availability of strong motion data for the attenuation coefficients for zero structural period
Chennai region, only one attenuation relationship as were c1 = 1.7816, c2 = 0.9205, c3 = -0.0673,
suggested by RAGHUKANTH and IYENGAR (2007) for the c4 = 0.0035 and r(ln e) = 0.3136, M is the moment
southern part of the PI has been used. This attenua- magnitude and R is the hypocentral distance. A
tion relationship was developed based on the hypocentral depth of 17 km for all epicentres was
seismological model. The functional form of the considered (Fig. 4) based on the seismicity data of
attenuation relationship is the PI. The parameters c and d for the influence area
were obtained (Fig. 5).
ln y ¼ c1 þ c2 ðM  6Þ þ c3 ðM  6Þ2  ln R  c4 R
The whole area of 360,000 km2 (600 9 600 km)
þ ln e
was divided into grids of 10 9 10 km size and the
In the above equation, y = (Sa/g) stands for the activity rate density for each magnitude range was
ratio of spectral acceleration (Sa) owing to gravity, calculated using the fixed kernel and adaptive kernel
 Seismic Hazard Analysis

01 and -5.92E-01 respectively. Similarly the maxi-

mum and minimum ratios obtained for Mw = 5.0
were 1.31 and -4.56E-01 respectively. It can be
observed that at regions of epicentre clustering, the
adaptive kernel yields higher activity rates and in
regions of sparsely located epicentres, the fixed ker-
nel yields higher activity rate. The spatial activity
rates so obtained are clubbed with the other forms of
uncertainty to determine the probability of exceed-
ance for the selected ground motion parameter
Figure 4
Variation of focal depth for earthquakes in Peninsular India (Eq. 2). The UHS obtained using the fixed and
adaptive kernel techniques are shown in Fig. 9.
The conventional Cornell–McGuire methodology
100
h = 2.53e
0.573M w of PSHA was carried out using CRISIS 2007 (ORDAZ
et al., 2007). Since the historical earthquake records
are more complete for large earthquakes as compared
to small earthquakes, fitting of the G–R recurrence
Bandwidth (km)

curve for such data results in the underestimation of

the mean annual rate of exceedance as the curve is
flatter. The incompleteness of small earthquakes is
because they go undetected due to physical and
demographic reasons. For this purpose the com-
pleteness analysis was carried out using the method
by Stepp (1974) considering a circular area of
10
3.0 3.5 4.0 4.5 5.0 5.5
300 km radius as one single source zone covering all
Magnitude (M w) the epicentres as shown in Fig. 3 and the complete-
ness periods so obtained are given in Table 2.
Figure 5
Magnitude and bandwidth variation for influence area around Figure 10 depicts the G–R recurrence law for the
Chennai city study area and the corresponding parameters are k
(M0 = 4.0) is 3.3, a is 3.8, b is 0.94, a (2.303a) is
techniques at these grid points (Eq. 3). Figure 6 8.7465 and b (2.303b) is 2.1601. The UHS obtained
shows the variation of bandwidth for Mw = 4.0 bin using all the three techniques are shown in Fig. 9.
for two epicentres, one located in a cluster and the The PGA values for 475 years return period from the
other located in isolation. The bandwidth obtained present study and from the results of other works are
from the fixed kernel using (Eq. 5) was 25 km, summarized in Table 3.
whereas the bandwidth in the adaptive kernel tech-
nique is varied spatially depending on the clustering
and sparseness of the epicentres. For example, the 6. Discussion and Conclusions
spatial activity rates from both the techniques for
Mw = 4.0 and Mw = 5.0 bins along with the spatial This paper has provided a method for estimation
distribution of epicentres are shown in Figs. 7a, b, c of spatial activity rate needed in probabilistic seismic
and 8a, b, c respectively. The difference in the hazard analysis using the adaptive kernel density
activity rate obtained from the fixed and adaptive estimation technique. This technique has the capa-
kernel techniques expressed as ratio of (adaptive– bility of considering explicitly the clustering and
fixed)/fixed is shown in Figs. 7d and 8d for Mw = 4.0 sparseness of the earthquake epicentres in the esti-
and Mw = 5.0 bins respectively. The maximum and mation of bandwidth. The method has mathematical
minimum ratios obtained for Mw = 4.0 were 6.01E- rigor and, hence, provides more realistic probability
 C. K. Ramanna, G. R. Dodagoudar Pure Appl. Geophys.

Figure 6
Global and local bandwidths for Mw = 4.0 bin

0 100 200 300 400 500 600

600 600

500 500
1.6E-006
1.5E-006
1.4E-006
400 400 1.3E-006
Distance (km)

1.2E-006
1.1E-006
1E-006
9E-007
300 Chennai City 300 8E-007
7E-007
6E-007
5E-007
4E-007
200 200 3E-007
2E-007
1E-007
0
100 100

0 0
0 100 200 300 400 500 600
Distance (km)

(a) (b)

0.50
1E-006 0.40
9E-007 0.30
8E-007 0.20
7E-007 0.10
6E-007 -0.00

5E-007 -0.10

4E-007 -0.20
-0.30
3E-007
-0.40
2E-007
-0.50
1E-007
0

(d)
(c)
Figure 7
a Spatial distribution of epicenters for Mw = 4.0 bin. b Spatial activity probability surface for Mw = 4.0 bin from adaptive kernel technique.
c Spatial activity probability surface for Mw = 4.0 bin from fixed kernel technique. d Relative error in spatial activity probability surface for
Mw = 4.0 bin
 Seismic Hazard Analysis

0 100 200 300 400 500 600

600 600

1.9E-006
500 500 1.8E-006
1.7E-006
1.6E-006
1.5E-006
1.4E-006
Distance (km)

400 400 1.3E-006

1.2E-006
1.1E-006
1E-006
9E-007
300 Chennai City 300 8E-007
7E-007
6E-007
5E-007
4E-007
200 200 3E-007
2E-007
1E-007
0
100 100

0 0
0 100 200 300 400 500 600
Distance (km)
(a) (b)

8E-007 1.20
7.5E-007 1.10
7E-007 1.00
6.5E-007 0.90
6E-007 0.80
5.5E-007 0.70
5E-007 0.60
4.5E-007 0.50
4E-007 0.40
3.5E-007 0.30
3E-007 0.20
2.5E-007 0.10
2E-007 -0.00
1.5E-007 -0.10
-0.20
1E-007
-0.30
5E-008
-0.40
0

(c) (d)
Figure 8
a Spatial distribution of epicenters for Mw = 5.0 bin. b Spatial activity probability surface for Mw = 5.0 bin from adaptive kernel technique.
c Spatial activity probability surface for Mw = 5.0 bin from fixed kernel technique. d Relative error in spatial activity probability surface for
Mw = 5.0 bin

surfaces. Using the adaptive kernel density estimation be compared, N is the number of summed spatial
technique, a methodology is suggested for seismic areas and xi is the location of each spatial area under
hazard analysis. As a case study example, Chennai is the density curve. To determine the ISE score
considered for probabilistic seismic hazard analysis. between the fixed and adaptive kernel estimators, the
However, it was observed that the final hazard value estimated density values were compared at every grid
in the form of UHS was not influenced by the adap- location. The ISE score so obtained for every mag-
tive kernel technique. The reason for this is further nitude for the influence area is shown in Fig. 11.
investigated in the form of the integrated squared From the figure it can be seen that the ISE score is
error score. almost negligible indicating on an average there is no
The integrated squared error (ISE) score is com- difference among the two estimators especially for
monly used when the difference between two models higher magnitudes. This is true because for higher
are to be estimated and is given as magnitudes, the epicentres are very few and sparsely
X
N located resulting in a unimodal density estimate,
ISE ¼ ðgA ðxi Þ  gB ðxi ÞÞ2 hence both estimates result in the same value.
i¼1 Earthquakes of magnitudes 4.3, 4.4 and 4.5 are more
where gA(xi) and gB(xi) are the two-dimensional in Chennai region and show clustering; hence, the
density functions corresponding to any two models to score is high for them. This score or difference in
 C. K. Ramanna, G. R. Dodagoudar Pure Appl. Geophys.

(a) 0.14 (c) 0.35

Adaptive Kernel Adaptive Kernel
Fixed Kernel Fixed Kernel
0.12 0.3
Cornell-McGuire Cornell-McGuire

Spectral Acceleration [g]

Spectral Acceleration [g]

0.1 0.25

0.08 0.2

0.06 0.15

0.04 0.1

0.02 0.05

0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
Structural Period [s] Structural Period [s]

(b) 0.3 (d) 0.45

Adaptive Kernel Adaptive Kernel
Fixed Kernel 0.4 Fixed Kernel
0.25 Cornell-McGuire Cornell-McGuire
0.35
Spectral Acceleration [g]

Spectral Acceleration [g]

0.2 0.3

0.25
0.15
0.2

0.1 0.15

0.1
0.05
0.05

0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
Structural Period [s] Structural Period [s]

Figure 9
a UHS for 72 years return period. b UHS for 475 years return period. c UHS for 975 years return period. d UHS for 2475 years return period

Table 2
m)

10
log m = 3.8 - 0.94M w
Mean Annual Rate of Exceedance (

Catalogue parameters

Magnitude range Completeness Completeness

year period (years) 1

3.5–3.99 1968 40
4.0–4.49 1968 40
4.5–4.99 1958 50 0.1
[5.0 1800 209

density can be attributed to the multimodal nature of

0.01
the distributions. However, the hazard values esti- 3.6 3.8 4.0 4.2 4.4 4.6 4.8 5.0 5.2 5.4

mated from both the techniques are same. This is Magnitude (M w )

because, firstly, the hazard value is less influenced by Figure 10

lower magnitudes; secondly, the density difference G–R recurrence law
 Seismic Hazard Analysis

Table 3 very much valid due to the inference given earlier

PGA values obtained by various researchers for 475-year return (ISE score). However, the peak spectral acceleration
period values obtained from the Cornell–McGuire approach
Approach/reference PGA (g) differed from the kernel techniques. For a lower
return period of 72 years, the Cornell–McGuire
JAISWAL and SINHA (2007) (Frankel method 0.080 approach yielded higher value but as the return period
of zoneless approach)
VIPIN et al. (2009) 0.10 increased, the approach resulted in lower peak values.
IS 1893: 2002 0.080 Since the adaptive kernel technique is statistically
Kernel approach (present study) 0.087 superior to the fixed kernel technique for multimodal
Adaptive Kernel approach (present study) 0.087
Cornell–McGuire approach (present study) 0.083
density estimation, the method needs to be applied
and tested for high seismicity regions and also for
very low seismicity regions. This would help in
overcoming the drawbacks of the fixed kernel density
0.00010
estimation technique to hazard estimation for high
0.00008 seismicity regions. However, it can be concluded that
for moderate seismicity regions, the Cornell-McGuire
ISE Score

0.00006
approach and both the kernel smoothing techniques
0.00004
yield the same results.

0.00002

Acknowledgments
4.0 4.5 5.0 5.5 6.0
Magnitude (M w ) Authors would like to express their sincere thanks to
Figure 11
the India Meteorological Department (IMD), New
Magnitude and ISE score variation Delhi for sharing the earthquake data relevant to the
study presented in the paper. The authors also wish to
thank Dr. Gordon Woo for sharing the KERFRACT
due to clustering of epicentres is located far away
program for carrying out a part of the work presented
from Chennai, thereby having lesser influence on the
in the paper. Comments from anonymous reviewers
final hazard value.
are highly appreciated and they greatly improved the
MOLINA et al. (2001) and BEAUVAL et al. (2006a)
clarity of the paper.
have observed that the fixed kernel method adopted
by WOO (1996) to PSHA yields compatible results (in
terms of PGA) when compared to Cornell-McGuire REFERENCES
approach for low to moderate seismicity regions
ABRAMSON, I. (1982), On bandwidth variation in kernel estimates—
when the background seismicity was added. How- a square root law, The Annals of Statistics 10, 1217–1223.
ever, they also observed that for high seismicity ALLEN, C.R., AMAND, P.ST., RICHTER, C.F., and NORDQUIST, J. M.
regions, the fixed kernel method yielded lower hazard (1965), Relationship between seismicity and geologic structure
in the southern California region, Bulletin of the Seismological
values as compared to the Cornell-McGuire Society of America 55(4), 753–797.
approach. Hence, in line with the above observations, BAK, P., How Nature Works, (New York: Springer-Verlag, 1996).
the results obtained from the Cornell–McGuire BEAUVAL, C., and SCOTTI, O. (2003), Mapping b-values in France
using two different magnitude ranges: possible non power-law
approach and fixed kernel technique in the form of
behaviour, Geophysical Research Letter 30(17), 4.
PGA (acceleration value at zero structural period BEAUVAL, C., SCOTTI, O., and BONILLA, F. (2006a), The role of
from Fig. 9) were compatible for Chennai city as it seismicity models in probabilistic seismic hazard estimation:
lies in the low to moderate seismicity region. It is Comparison of a zoning and a smoothing approach. Geophysical
Journal International 165, 584–595.
interesting to note that the adaptive kernel method- BEAUVAL, C., HAINZL, S., and SCHERBAUM, F. (2006b), The impact of
ology to PSHA also yielded a similar result which is the spatial uniform distribution of seismicity on probabilistic
 C. K. Ramanna, G. R. Dodagoudar Pure Appl. Geophys.

seismic-hazard estimation, Bulletin of the Seismological Society MCGUIRE, R.K., and ABRABASZ, W.J. (1990), An introduction to
of America 96(6), 2465–2471. probabilistic seismic hazard analysis, In S.H. Ward. ed. Geo-
BENDER, B. (1986), Modelling source zone boundary uncertainty in technical and Environmental Geophysics, Society of Exploration
seismic hazard analysis, Bulletin of the Seismological Society of Geophysicists 1, pp. 333–353.
America 76(2), 329–341. MITTAL, A., and PARAGIOS, N. (2004), Motion-based background
BRUNSDON, C. (1995), Estimating probability surfaces for geo- subtraction using adaptive kernel Density Estimation, 2004,
graphical point data: an adaptive kernel algorithm, Computers IEEE Computer Society Conference on Computer Vision and
and Geosciences 21(7), 877–894. Pattern Recognition (CVPR’04) 2, 302–309.
CHAN, C. H., and GRUNTHAL, G. (2008), Developing new approa- MOLINA, S., LINDHOLM, C. D., and BUNGUM, H. (2001), Probabilistic
ches to earthquake hazard assessment and forecasting, Network seismic hazard analysis: Zoning free versus zoning methodology,
for Research Infrastructures for European Seismology EC Project Bollettino Di Geofisica Teorica Ed Applicata 42(1–2), 19–39.
number: 026130. MUGDADI, A.R., and AHMAD, I.A. (2004), A bandwidth selection of
CHANDRA, U. (1977), Earthquakes of peninsular India—A seismo- kernel density estimation of functions of random variables,
tectonic study, Bulletin of the Seismological Society of America Computational Statistics and Data Analysis 47, 49–62.
67(5), 1387–1413. ORDAZ, M., AUGILAR, A., and ARBOLEDA, J. (2007), CRISIS2007 Ver
CHEN, Y., LIU, J., CHEN, L., CHEN, Q., and CHAN, L.S. (1998), 1.1: Program for computing seismic hazard. Instituto de Inge-
Global seismic hazard assessment based on area source model nieria UNAM, Mexico.
and seismicity data, Natural Hazards 17, 251–267. Principia (2005), Seismic hazard evaluation, LNG plant at Taranto
CORNELL, C.A. (1968), Engineering seismic risk analysis, Bulletin (Italy), Report No.: 673 Rev. 1, Project No. P-373.
of the Seismological Society of America 58, 1583–1606. RAGHUKANTH S.T.G., and IYENGAR R.N. (2007), Estimation of
DASGUPTA, S., PANDE, P., GANGULY, D., IQBAL, Z., SANYAL, K., seismic spectral acceleration in peninsular India, Journal of
VENKATRAMAN, N.V., DASGUPTA, S., SURAL, B., HARENDRANATH, Earth System Science 116, 199–214.
L., MAZUMDAR, K., SANYAL, S., ROY, A., DAS, L. K., MISRA, P.S., RAO, R.B. (1992), Seismicity and geodynamics of the low-to high-
and GUPTA, H. (2000), Seismotectonic Atlas of India and Its grade transition zone of peninsular India, Tectonophysics 201,
Environs, Geological Survey of India Spec. Publ. No. 59, 175–185.
Kolkata, India. RAO, R.B. (2000), Historical seismicity and deformation rates in
de SMITH, M.J., GOODCHILD, M.F., and LONGLEY, P.A., Geospatial the Indian peninsular shield, Journal of Seismology 4, 247–258.
Analysis—A Comprehensive Guide. 3rd edn. (Leicester: The RAO, R.B., and RAO, S.P. (1984), Historical seismicity of peninsular
Winchelsea Press, 2009). India, Bulletin of the Seismological Society of America 74(6),
FRANKEL, A. (1995). Mapping seismic hazard in the central and 2519–2533.
eastern United States, Seismological Research Letters, 66, 8–21. REITER, L., Earthquake Hazard Analysis: Issues and Insights (New
GARDNER, J.K., and KNOPOFF, L. (1974), Is the sequence of earth- York: Columbia University Press, 1991).
quakes in southern California, with aftershocks removed, SCOTT, W.D., Multivariate Density Estimation (New York: John
Poissonian? Bulletin of the Seismological Society of America Wiley and Sons, 1992).
64(5), 1363–1367. SECANELL, R., BERTIL, D., MARTIN, C., GOULA, X., SUSAGNA, T.,
GUPTA, I.D. (2006), Delineation of probable seismic sources in TAPIA, M., DOMINIQUE, P., CARBON, D., and FLETA, J. (2008),
India and neighbourhood by a comprehensive analysis of seis- Probabilistic seismic hazard assessment of the Pyrenean region,
motectonic characteristics of the region, Soil Dynamics and Journal of Seismology 12, 323–341.
Earthquake Engineering 26, 766–790. SILVERMAN, B.W., Density estimation for statistics and data anal-
GUPTA, I.D., TODOROVSKA, M.I., GUPTA, V.K., LEE, V.W., and ysis, Monographs on Statistics and Applied Probability (London:
TRIFUNAC, M.D. (1995), Selected topicsin probabilistic seismic Chapman and Hall, 1986).
hazard analysis, Report No. CE95-08, University of Southern STAINSLAW, L., and Sikora, O.B. (2008), Seismic hazard assessment
California, Los Angeles, 304. under complex source size distribution of mining—induced
Indian Standards (2002), IS 1893: 2002. Indian standard criteria seismicity, Tectonophysics 456, 28–37.
for earthquake resistant design of structures, Part 1—General STEPP, J.C. (1974), Analysis of completeness of the earthquake
provisions and buildings, New Delhi, Bureau of Indian sample in the Puget sound area. In: S.T. Harding, ed., NOAA
Standards. Technical Report ERL 267-ESL 30, Contributions to Seismic
JAISWAL, K., and SINHA, R. (2007), Probabilistic seismic—hazard Zoning 16–28.
estimation for peninsular India, Bulletin of the Seismological STOCK, C., and SMITH, E.G.C. (2002a), Adaptive kernel estimation
Society of America 97(1B), 318–330. and continuous probability representation of historical earth-
KRAMER, S.L., Geotechnical Earthquake Engineering (Englewood quake catalogs, Bulletin of the Seismological Society of America
Cliffs, New Jersey: Prentice Hall, 1996). 92(3), 904–912.
KRINITZSKY, E.L. (1993). Earthquake probability in engineering— STOCK, C., and SMITH, E.G.C. (2002b), Comparison of seismicity
Part 2: Earthquake recurrence and limitations of Gutenberg- models generated by different kernel estimations, Bulletin of the
Richter b-values for the engineering of critical structures The Seismological Society of America 92(3), 913–922.
Third Richard H. Jahns Distinguished Lecture in Engineering SUBRAHMANYA, K.R. (1995), Active intraplate deformation in south
Geology, Engineering Geology 36, 1–52. India, Tectonophysics 262, 231–241.
LAPAJNE, J., MOTNIKAR, B. S., and ZUPANCIC, P. (2003), Probabilistic VERE-JONES, D. (1992), Statistical methods for the description and
seismic hazard assessment methodology for distributed seismic- display of earthquake catalogs. In: A.T. WALDEN, and P. GUT-
ity, Bulletin of the Seismological Society of America 93(6), TORP, eds. Statistics in the Environmental and Earth Sciences.
2502–2515. (London: Arnold Publishers 1992), 220–246.
 Seismic Hazard Analysis

VIPIN, K.S., ANBAZHAGAN, P., and SITHARAM, T.G. (2009), Estima- WOO, G. (1996), Kernel estimation methods for seismic hazard
tion of peak ground acceleration and spectral acceleration for area source modeling, Bulletin of the Seismological Society of
south India with local site effects: probabilistic approach, Nat- America 86(2), 353–362.
ural Hazards and Earth System Sciences 9, 865–878. ZOLFAGHARI, M.R. (2009), Use of raster-based data layers to
WALLING, Y.M., and MOHANTY, K.W. (2009), An overview on the model spatial variation of seismotectonic data in probabilistic
seismic zonation and microzonation studies in India, Earth-Sci- seismic hazard assessment, Computers and Geosciences 35,
ence Reviews 96, 67–91. 1460–1469.
WAND, M.P., and JONES, M.C. (1995), Kernel Smoothing, Mono-
graphs on Statistics and Applied Probability (London: Chapman
and Hall).

(Received April 27, 2010, revised November 23, 2010, accepted December 20, 2010)

View publication stats