A better automatic body-wave picker with broad applicability.
Frédérick Massin
∗
and Alison Malcolm, Memorial University
SUMMARY
For robust earthquake analysis, we need efficient and reliable
automatic body-wave recognition methods. To do this, we
combine the advantages of standard methods in an innovative
and generalized approach. Using the component energy corre-
lation method, we demonstrate the mathematical and practical
advantages of the correlation operator and apply this operator
to the
¯
ST /
¯
LT and
¯
RP/
¯
LP methods. We also implement multi-
scale versions of these methods to reduce the dependence on
user-defined time-scale parameters. We compare our results
systematically to different methods, propose an optimal ap-
proach and demonstrate its reliability.
INTRODUCTION
Whether analyzing global tectonic events or micro-earthquakes
(natural or from human activity), we require estimates of body-
wave arrival times from seismic records. The constant sensitiv-
ity and precision of automatic body-wave recognition methods
are the basis of a robust earthquake analysis. Although body-
wave arrival detection is done by automatic systems, these sys-
tems need configuration and review by an operator. Our study
proposes a renewed generalized approach to reduce configura-
tion, and to increase the sensitivity and precision of automatic
systems resulting in a better earthquake analysis.
The evolution of automatic body-wave recognition methods
has followed two main directions. The methods in the first
branch are based on characteristic functions (C f ), which give
a proxy for the continuous probability of a body-wave arrival
from the continuous seismic signal, which is given as input.
Such methods are called uninformed because they are based on
continuous seismological data, i.e. they have an uninformed
prior. They often rely on variants of the Short Term - Long
Term averages ratio algorithm (
¯
ST /
¯
LT is used in Earthworm
and SeisComP3; Allen, 1982; Baer and Kradolfer, 1987; Lo-
max et al., 2012). In this method, the ratio of two running
means (one short and one long time scale) increases with am-
plitude changes. By definition, the long time scale defines the
lower bound on the period of detected body-waves with the
short time scale defining the period of best sensitivity. Similar
to
¯
ST /
¯
LT, other uninformed methods are based on an opera-
tor (e.g. ratio) by which multiple data streams (e.g. short and
long term running means) are combined into one stream, a pro-
cess we refer to as multiplexing. The resulting C f can be used
for probabilistic earthquake source scanning (Kao and Shan,
2004) or to extract body-wave information (arrival time, am-
plitude, frequency) for micro-earthquake location (in industrial
environment Chen, 2006; Wong et al., 2009) or further analy-
sis.
In addition to amplitude changes, body-wave arrivals are also
associated with frequency and polarization changes that may
not be detected by
¯
ST /
¯
LT. Informed methods can help to
detect these events. These methods aim to improve arrival
time data by increasing their time precision and the number
of detected arrivals, using pre-existing and potentially incom-
plete arrival time datasets from uninformed methods as prior
information. Significant work has been done with informed
methods (using autoregression, Wenner and match filters, see
Song et al., 2010,for a application to hydrofracture induced mi-
croseismic) and open-source standard libraries (obspy, Megies
et al., 2011) are now available to combine their advantages.
The advantages of informed methods include sub-sample ar-
rival time precision and sensitivity on data with signal to noise
ratio below one (with match filters, Chamberlain et al., 2014).
However, informed methods require pre-processing, advanced
configuration, and are difficult to use for real time applications
(as is necessary in industrial or hazard monitoring).
With further work, uninformed methods might be improved
for reliable and standard body-wave arrival analysis. Along
these lines, we combine the advantages of several C f from
uninformed methods to minimize the necessary configuration
parameters, and to improve the sensitivity and precision of au-
tomatic arrival time picking. We first investigate the interest-
ing advantages of the component energy correlation method
(Nagano et al., 1989), which we call C?
rms
. To get the same
advantages as C?
rms
, the correlation operator is then general-
ized for C f based on multiplexing. Finally, we follow pre-
existing work for multi-scaling with multiple C f to reduce the
dependence on time-scale parameters. We compare our results
with three C f (
¯
ST /
¯
LT, C?
rms
and
¯
RP/
¯
LP, Zahradník et al.,
2015) and propose an optimal approach.
METHOD
Component energy correlation method
P-waves produce displacements primarily along the propaga-
tion direction, while S-waves produce displacements primarily
perpendicular to the propagation direction. Seismic noise by
contrast, is a stochastic signal (Wentzell, 1981) and as such,
exists on all polarizations. More specifically, the energy dissi-
pation through time of stochastic noise is correlated between
all polarization angles. P and S-wave arrivals temporarily de-
crease the cross-channel energy correlation, a measure we can
exploit to detect their arrivals. This component energy correla-
tion method was introduced by (Nagano et al., 1989„ C?
rms
);
the estimation of P and S wave arrival times and their uncer-
tainties is described in detail in (Zhizhin et al., 2006; Massin
et al., 2009). To account for body-waves incoming along the
bisector of the sensor orientation in down-hole acquisition en-
vironments, we can use a second instance of C?
rms
with corre-
sponding rotated components. C?
rms
is the product of R
ZE(t,T
C
)
and R
ZN(t,T
C
)
, the correlations between the energy dissipation
Page 2617
© 2016 SEG
SEG International Exposition and 86th Annual Meeting
Downloaded 07/04/17 to 134.153.188.68. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
automatic body-wave picker
of the vertical and the two horizontal components given by:
R
j,k(t,T
C
)
=
P
t+T
C
i=t
E
j
(i)E
k
(i)
P
t+T
C
i=t
E
2
j
(i)
P
t+T
C
i=t
E
2
k
(i)
. (1)
The energy dissipation E
k
(t) on k(t), a given component of
the seismic signal as a function of time t, is defined as a slid-
ing rms (root mean square,
P
t
i=t−T
k(i)
2
). C?
rms
is used as a
C f of a P-wave, C?
rms(E)
(function of R
EZ(t,T
C
)
) and C?
rms(N)
(function of R
NZ(t,T
C
)
) are used for S-waves. The correlation
coefficient can be used as a probability of arrival time estimate
between zero and one (one being worst). The correlation win-
dow (T
C
) defines the dominant period of the P-waves with best
sensitivity (Nagano et al., 1989). The energy dissipation win-
dow (T ) can be chosen arbitrarily as twice T
C
and defines the
lowest frequency of sensitivity (Zhizhin et al., 2006).
C?
rms
behaves as a multiplexor in which the three directional
channels of a seismometer are combined into one C f . By this
design, it is sensitive to polarization changes, a physical prop-
erty specific to body-waves.
Generalized correlation method
There are several C f based on the ratio operator. The
¯
ST /
¯
LT
uses the averaged absolute value of the signal (Allen, 1982)
or its envelope (Baer and Kradolfer, 1987). The
¯
RP/
¯
LP uses
the summed absolute values of the signal in the right and left
part of a moving time window (Chen, 2006; Wong et al., 2009;
Zahradník et al., 2015). In these C f , the ratio operator could be
substituted by the correlation operator with two advantages: 1)
the correlation outputs normalized values, whereas the
¯
RP/
¯
LP
method gives values dependent on the input data, 2) correlation
is a robust approximation of deconvolution.
An operator that estimates normalized probability allows the
use of a quantitative triggering threshold and the comparisons
of multiple properties that might have different characteristics.
The second advantage is based on the mathematical robustness
of the correlation operator. The deconvolution of g(t) from
f (t) is written in the frequency domain as:
F { f }
F {g}
=
F { f }F {g}
∗
F {g}F {g}
∗
, (2)
where F denotes the Fourier transform, and an
∗
indicates
the complex conjugate. Noting that Equation 2 is the Least-
Squares inverse of the convolution operator, F { f }F {g}
∗
is
the adjoint of the convolution. In the time domain, the de-
convolution of g from f is thus approximated by the cross-
correlation of g and f (F { f ? g} = F { f }F {g}
∗
)
The correlation produces a finite, normalized probability; al-
lowing us to combine several correlation-based C f for im-
proved sensitivity. For example, combining
¯
ST ?
¯
LT and C?
rms
will give a detector that is sensitive to both amplitude and po-
larization changes.
Multi-scaling
To adapt to multiple earthquake magnitudes and distances, we
use a multi-scale calculation of the C f following (Lomax et al.,
2012). The signal is first decomposed using high pass filters
which are then used for running mean and rms calculations.
Our formulation of multi-scaled C f (
M
C f ) is the product of a
range of C f based on a time scale T as defined in Equation 3:
M
C f = C f
(t,T
0
)
.C f
(t,T
1
)
...C f
(t,T
n
)
. (3)
Each C f
(t,T)
is calculated with the running mean (
P
t
i=t−T
|k(i)|
T
)
or rms (
P
t
i=t−T
k
2
(i)) for a frequency band [
1
T
n
,
1
T
n+1
] using the
output of a high pass filter with corner frequency
1
T
n+1
. We take
advantage of the running mean filter capabilities in the time
domain for reducing low frequency signals while retaining a
sharp step response (Smith, 1997).
Figure 1: Artificial data tests. Gray: artificial noisy data added
to three perturbations each simulating one property of P and
S waves. Dark (light) gray: vertical (horizontal) component
of ground displacements. Ns.Z : noise only (vertical com-
ponent of a 3-dimensional random motion) ; F.Z : frequency
changes (vertical component); A.(ZEN): amplitude changes
(three components) ; P.(ZEN): vertically polarized motion
added at 10/3 s and horizontally polarized motion added at 5s
(three components). Characteristic functions are represented
by overlying colored lines.
Using correlation and multi-scaling, we can combine meth-
ods and frequency bands for a robust general detector of body-
wave arrivals. The threshold over which a body-wave arrival
might be considered is the only required parameter. In the next
Page 2618
© 2016 SEG
SEG International Exposition and 86th Annual Meeting
Downloaded 07/04/17 to 134.153.188.68. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
automatic body-wave picker
section, we show the response of C?
rms
, generalized correla-
tion methods (
¯
ST ?
¯
LT and
¯
LP?
¯
RP) and multi-scaling with ar-
tificial and real data.
APPLICATIONS
Artificial tests
We use what we call artificial data (because it does not aim
to simulate a realistic Earth response) to study individually
three properties of body-waves (changes in frequency, ampli-
tude, and polarization) in the presence of seismic noise. Our
first key result illustrates that C?
rms
has the best sensitivity of
the standard approaches, characterized by narrower probabil-
ity peaks at the wave arrival time. Our second key result shows
that correlation-based methods have the best sensitivity to pure
frequency changes. Amongst multi-scales approaches, our last
key result is that
M
C?
rms
has a similar response to C?
rms
.
The response of
¯
ST /
¯
LT and
¯
RP/
¯
LP, shown by test A in Fig-
ure 1 to all three properties (described in the caption) have sim-
ilar precision and sensitivity. Using test A and B, we compare
the ratio methods based on running mean to the same meth-
ods based on rms. Although different in details, the responses
of ST /LT
rms
and RP/LP
rms
in test B are globally similar to
¯
ST /
¯
LT and
¯
RP/
¯
LP. The probability peaks of RP/LP
rms
are
clearer that the other ratio methods. The onset of ST /LT meth-
ods are consistent with the location of property changes as is
the local maximum for RP/LP methods. Using test B and C,
we compare the ratio and the correlation methods using rms
(C f
rms
). The correlation methods have a lower noise level and
an improved sensitivity to pure frequency changes, however
there is a loss of precision in LP?RP
rms
. C?
rms
has the most
impulsive response to amplitude and polarization changes com-
pared to all tested correlation or ratio methods.
The comparison of multi-scale implementations in test D is
based on correlation methods and rms (
M
C f
rms
). The multiple
inaccurate peaks of
M
LP?RP
rms
make it overall the least reli-
able of the tested methods. The noise level and the peak nar-
rowness of
M
ST ?LT
rms
and
M
C?
rms
are respectively the best of
all implementations of these two methods.
M
ST ?LT
rms
has the
advantage of being sensitive to all tested properties of body-
waves, while
M
C?
rms
is the most sensitive to amplitude and po-
larization changes amongst the tested multi-scale approaches.
Data test
The use of real data allows visual comparison of manual pick-
ing data with automatic methods. We use a set of micro-earthquakes
over a range of magnitudes and distances (Figure 2, 3, 4 cho-
sen to be as consistent as possible with properties of industrial
experiments). At first glance, this test shows good agreement
of the C f onsets with manual picks. However the data signal
to noise ratio affects the response of C f to body-wave arrivals,
particularly for
M
ST ?LT
rms
.
The consistency between the onsets of tested C f and the man-
ual P-wave arrival picks is significant through the entire dis-
tance range (Figure 2 and Figure 3). Before the arrival, the
C f noise level are stable enough to distinguish a clear onset.
Figure 2: Real data tests for P-waves (S-waves in Figure 3).
Gray: data recorded from M2 to M3 earthquakes with increas-
ing distance from 3 to 60 km from bottom to top. Data are
normalized by their respective maximum absolute values. The
manual P-wave arrival time is centered at 10 s. Various in-
strument types and signal to noise ratios are shown. Dark
(light) gray: vertical (horizontal) component of ground dis-
placements. Characteristic functions are represented by over-
lying colored lines with an exponential exaggeration of 0.5
(low values are the most exaggerated).
Figure 3: Identical to Figure 2 for S-waves.
Page 2619
© 2016 SEG
SEG International Exposition and 86th Annual Meeting
Downloaded 07/04/17 to 134.153.188.68. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
automatic body-wave picker
It is important to note that the onset of local maximum in the
C f peak is correlated with the body-wave arrival, not the maxi-
mum itself. These observations are clearer for
M
C?
rms
than for
M
ST ?LT
rms
. For
M
C?
rms
, the maximum level of noise is be-
low 0.2, which could be used as a general threshold to detect
probability peaks associated with body-wave arrivals.
The data noise level is the most important factor affecting the
results, but it is not explained by body wave attenuation as
function of distance. The worst results are not obtained with
data from further distances but rather with exceptionally noisy
data in the mid-distance range. The most important effect of
distance is the appearance of multiple sub-peaks in the vicinity
of the local maximum, reducing the precision of the P-wave
arrival pick.
Figure 4: C f response as a function of magnitude and distance
with associated averaged probabilities and colormap. Dark:
low C f values and signal to noise ratio (SNR). White: low C f
values and high SNR. Blue (green): high C f at P-wave (S)
arrival compare to noise level. Red cells: low C f at arrival.
DISCUSSION
We investigated several C f formulations in order to find the
most reliable way to estimate body-wave arrival times auto-
matically. The methods based on rms calculations, polariza-
tion changes, and correlation showed the best sensitivities to
body-wave arrivals and they can all be implemented in a multi-
scale approach for application to a wide range of distances and
magnitudes.
The worst responses are obtained with noisy data independent
of distance. At constant magnitude below 60 km, data quality
from station settings and site effects seem to impact the results
more than attenuation or source effects. Data quality appear
as a major limitation in our study.
M
ST ?LT
rms
appears to be
less reliable than
M
C?
rms
. The onsets of
M
ST ?LT
rms
are sig-
nificantly less consistent with manual picks and its probability
peaks have significantly smaller amplitudes than
M
C?
rms
. In
the absence of three-component data,
M
C?
rms
cannot be ap-
plied and the sensitivity and precision of
M
ST ?LT
rms
are fun-
damental limitations.
The consistency of ratio methods using running mean or rms,
and the advantages of RP/LP
rms
have already been shown for
industrial applications (Chen, 2006; Wong et al., 2009) or tec-
tonic events (Zahradník et al., 2015). In our broader study, the
best responses are obtained with the methods based on rms cal-
culations, polarization changes and correlation. This is consis-
tent with the results of previous authors (Nagano et al., 1989;
Zhizhin et al., 2006; Massin et al., 2009) but our study goes
further.
M
C?
rms
is our best multi-scale implementation and
can estimate body-wave arrivals with broadband data without
pre-filtering. This method is based on recursive algorithms
and does not require high-order statistics. The optimal use of
M
C?
rms
requires three-dimensional data and spreads over all
tested distances and magnitudes (M1 to M4, Figure 4).
The arrival time of a body-wave could be approximated with
the onset of the probability peak from a reliable C f . To get
further, our method has to be coupled with an onset recogni-
tion algorithm. Only local maximums over the threshold of
0.2 (as defined for
M
C?
rms
) should be considered. Theoret-
ically, the minimal time span allowed between two peaks is
dictated by the period of the first detected signal. In practice
multiple peaks between subsequent P and S arrivals might be
an issue for earthquake location (there are no such limitations
with shift and stack methods, Kao and Shan, 2004; Zahradník
et al., 2015). Time derivative and Kurtosis have been shown as
reliable onset indicators for seismic wave detection (Baer and
Kradolfer, 1987; Lomax et al., 2012); their maximum in the
vicinity of the probability peak should give best estimates of
arrival times.
CONCLUSION
We combined the advantages of correlation and multi-scaling
with the detection of amplitude and polarization change into
a single C f . We obtained a body-wave recognition tool with
broad applicability from industrial surveys to earthquake mon-
itoring. We defined an optimal calculation scheme and the re-
lated threshold that can then be used for event detection with
our approach. We discussed its limitations, which are related
to data quality and to the need for 3-dimensional seismic records.
ACKNOWLEDGMENTS
The authors thank the Hibernia Management and Development
Corporation for funding this work.
Page 2620
© 2016 SEG
SEG International Exposition and 86th Annual Meeting
Downloaded 07/04/17 to 134.153.188.68. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
EDITED REFERENCES
Note: This reference list is a copyedited version of the reference list submitted by the author. Reference lists for the 2016
SEG Technical Program Expanded Abstracts have been copyedited so that references provided with the online
metadata for each paper will achieve a high degree of linking to cited sources that appear on the Web.
REFERENCES
Allen, R. V., 1982, Automatic phase pickers: Their present use and future prospects: Bulletin of the
Seismological Society of America, 72, S225–S242.
Baer, M., and U. Kradolfer, 1987, An automatic phase picker for local and teleseismic events: Bulletin of
the Geological Society of America, 77, 1437–1445.
Chamberlain, C. J., D. R. Shelly, J. Townend, and T. A. Stern, 2014, Low-frequency earthquakes reveal
punctuated slow slip on the deep extent of the Alpine Fault, New Zealand: Geochemistry,
Geophysics, Geosystems, 15, 2984–2999,
http://dx.doi.org/.1002/2014GC005436.
Chen, Z., 2006, A multi-window algorithm for real-time automatic detection and picking of P-phases of
microseismic events: CREWES Research Report, 18.
Kao, H., and S.-J. Shan, 2004, The Source-Scanning Algorithm: Mapping the distribution of seismic
sources in time and space: Geophysical Journal International, 157, 589–594,
http://dx.doi.org/10.1111/j.1365-246X.2004.02276.x.
Lomax, A., C. Satriano, and M. Vassallo, 2012, Automatic picker developments and optimization:
FilterPicker — A robust, broadband picker for real-time seismic monitoring and earthquake early
warning: Seismological Research Letters, 83, 531–540,
http://dx.doi.org/10.1785/gssrl.83.3.531.
Massin, F., V. Ferrazzini, P. Bachèlery, and Z. Duputel, 2009, A real time process for detection,
clustering, and relocation of volcano-tectonic events at Piton de La Fournaise volcano: Presented
at the American Geophysical Union, Fall Meeting.
Megies, T., M. Beyreuther, and R. Barsch, 2011, ObsPy — What can it do for data centers and
observatories? Annals of Geophysics, 54, 47–58.
Nagano, K., H. Niitsuma, and N. Chubachi, 1989, Automatic algorithm for triaxial hodogram source
location in downhole acoustic emission measurement: Geophysics, 54, 508–513,
http://dx.doi.org/10.1190/1.1442677.
Smith, S. W., 1997, Moving average filters, in The scientist and engineer’s guide to digital signal
processing, 277–284.
Song, F., H. S. Kuleli, M. N. Toksöz, E. Ay, and H. Zhang, 2010, An improved method for hydrofracture-
induced microseismic event detection and phase picking: Geophysics, 75, no. 6, A47–A52,
http://dx.doi.org/10.1190/1.3484716.
Wentzell, A. D., 1981, A course in the theory of stochastic processes: McGraw-Hill International.
Advanced book program: McGraw-Hill.
Wong, J., L. Han, J. Bancroft, and R. Stewart, 2009, Automatic time-picking of first arrivals on noisy
microseismic data: Presented at the CSEG Microseismic Workshop, CREWS, University of
Calgary.
Zahradník, J., J. Janský, and V. Plicka, 2015, Analysis of the source scanning algorithm with a new P-
wave picker: Journal of Seismology, 19, 423–441.
Zhizhin, M. N., D. Rouland, J. Bonnin, A. D. Gvishiani, and A. Burtsev, 2006, Rapid estimation of
earthquake source parameters from pattern analysis of waveforms recorded at a single three-
component Broadband Station, Port Vila, Vanuatu: Bulletin of the Seismological Society of
America, 96, 2329–2347.
Page 2621
© 2016 SEG
SEG International Exposition and 86th Annual Meeting
Downloaded 07/04/17 to 134.153.188.68. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/