SUGGESTED PICKER PARAMETER CHANGES

James C. Pechmann

University of Utah, Dept. of Geology and Geophysics

July 7, 1998

Updated August 29, 1998

FURTHER SUGGESTIONS: OCTOBER 2, 2006.

Original 1998 Suggestions.


     This document describes some suggested changes to the standard
parameter settings used in the Earthworm version of Rex Allen's
P-Picker algorithm.  These changes have improved the picker's
performance at two Earthworm installations where they have been
implemented as of this writing:  the University of Utah Seismograph
Stations (UUSS) and the University of Washington.  I consider these
recommended parameter changes to be provisional because I am still
working on optimizing the parameters and because I have not yet done
a systematic comparison between the picker performance with and
without these parameter changes.  Nevertheless, I have been encouraged
to make these provisional recommendations available to other Earthworm
users who may wish to implement them.

     The standard parameter settings were originally developed by Rex
Allen at the U.S. Geological Survey office in Menlo Park, California.
These parameters have been in use for many years in Rex Allen P-Picker
installations operating on data from the Northern California Seismic
Network (NCSN), the Southern California Seismic Network (SCSN), and
elsewhere.  While the performance of the picker with these parameters
has apparently been satisfactory for some networks, such as the NCSN
and SCSN, that has not been the case for other networks including the
UUSS network.  The problem appears to be that the P-Picker algorithm
with the standard parameters does not do a very good job of picking
relatively low-amplitude, emergent first arrivals.  Consequently, for
networks with average station spacings significantly larger than those
of the California networks, the picker is unable to pick enough
arrival times from most of the smaller seismic events to enable them
to be located.

     I believe that the picker performance can be significantly
improved by changing just four of the program's parameters--five for
broadband stations--as detailed below.  These recommended changes are
based primarily on detailed analyses of the reasons for the picker's
failure to pick arrival times on several seismic records with small
but clear P-wave first arrivals.  I will first give a summary of the
recommended changes to the picker parameters and then give a brief
explanation about what these changes do and why I think they are
needed.  The reasons for the first three changes are fairly obvious
once it is understood what the program does.  The last two recommended
changes are not as straightforward, and are subject to revision
pending further work.


SUMMARY OF SUGGESTED CHANGES:

     The 2-letter parameter names given below are those used in the
original Earthworm version of the code.  The longer names are those
used in the 1997 and later versions of the code written by Will Kohler
of the USGS.

(1)  Change i8 = MinPeakSize from 60 to 20 counts

(2)  Change i9 = MinCodaLen from 3 to 0 seconds

(3)  FOR BROADBAND STATIONS ONLY
     Change c1 = RawDataFilt from 0.985 to 0.939 

(4)  Change c3 = StaFilt from 0.6 to 0.4

(5)  Change c4 = LtaFilt from 0.03 to 0.015


EXPLANATION OF SUGGESTED CHANGES:

(1)  Change i8 = MinPeakSize from 60 to 20 counts

    In one of the quality checks performed by the picker, the absolute
values of the amplitudes of the first three extrema (peaks or troughs)
are compared to a threshold value.  This value was formerly called i8
but is called MinPeakSize in more recent versions of the code.  At
least one of these extrema must be greater than or equal to i8 for the
pick to be considered valid.

    For UUSS data, and probably data from other networks as well, the
standard threshold amplitude of i8 = 60 counts used for the NCSN is
unnecessarily high.  I chose to reset this parameter to 20, which is
roughly twice the zero-to-peak noise level for quiet UUSS stations.


(2)  Change i9 = MinCodaLen from 3 to 0 seconds

     It is very important to change this parameter to a value less
than 2 sec, even though for some versions of Earthworm this change
will require recompiling some code.  Here's why.

     The signal duration is defined as the difference between the
P-wave arrival time and the time when the average absolute value of a
200-sample-long window drops below a threshold value.  Except for
"noisy" stations, this threshold is c7 = CodaTerm.  The USGS uses a
value of c7 = 49.14 counts for both historical and practical reasons.

     For some P-wave arrivals with good signal-to-noise ratio on the
UUSS network, the average absolute value over the first 200 samples
after the pick is less than 49.14 counts.  Arrival time picks for such
arrivals will be rejected after 2 sec if the sampling rate is 100 Hz
and the minimum signal duration, i9, is set to 3 sec (the standard
NCSN value).  Setting i9 = MinCodaLen = 0 sec solves this problem.

     This problem could also be solved by lowering the threshold
value, c7 = CodaTerm, perhaps to some value representative of the
background noise levels.  This solution is less desireable because it
would lower the maximum magnitude that could be determined from the
signal duration measurements which, in the interests of timeliness,
are currently limited to 144 sec.

     The parameter i9 = MinCodaLen was unfortunately hardcoded to a
value of 3 sec in the picker software supplied with Earthworm versions
2.8 (May 1997) through 3.1 (April 1998).  If you are running one of
these versions of Earthworm, then to set i9 = MinCodaLen = 0 sec you
need to make the following change in the subroutine pick_ra.c and then
recompile:

                     Change the line

const int MinCodaLen = 3;      /* Must be an odd number */

                            to

const int MinCodaLen = 0;      /* Must be an odd number */


(3)  FOR BROADBAND STATIONS ONLY
     Change c1 = RawDataFilt from 0.985 to 0.939 

     The first thing that the picker algorithm does is to apply a
recursive highpass filter to the data to remove the DC offset.  If
c1 = 0.985 (the standard NCSN value) and if the sampling rate is
100 Hz, then the corner frequency of the filter is 0.24 Hz.  

     On broadband records, which for this purpose I define as records
from seismometers with free periods greater than one second, small
amplitude P-wave arrivals are often obscured by microseisms.  The
microseisms can be filtered out more effectively by setting
c1 = 0.939, which sets the corner frequency of the highpass filter to
1.0 Hz.  This change ought to improve the performance of the picker on
broadband records.  I haven't actually verified that it does improve
the picking, but I have verified that the change in c1 changes the
filtering as predicted.  For stations with short-period (1 Hz)
seismometers, the standard NCSN value of c1= 0.985 works just fine and
should probably be left alone.


(4)  Change c3 = StaFilt from 0.6 to 0.4
(5)  Change c4 = LtaFilt from 0.03 to 0.015

     The need for these changes is less obvious than for the other two,
but nonethless they are very important.  c3 and c4 control the time
constants for the short-term and long-term averages, respectively, of 
the characteristic function used in the identification of P-wave
first arrivals (see Allen, 1978, BSSA 68, 1521-1532).  The changes
above roughly double the time constants for these two averages, whose
ratio serves as the primary detector of P-wave arrivals in the picker
program.  Suppose we define NSTA and NLTA as the number of samples it
takes for the weight in the (recursively calculated) weighted
short-term and long-term averages, respectively, to drop below 10%.
Then the above changes increase NSTA from 3 to 5 samples and NLTA from
75 to 152 samples.

     The motivation for these changes is that the dominant period of the
background noise is 0.05 sec in the primary UUSS data source:  analog
telemetry data digitized at 100 Hz.  It seemed likely that the picker
performance would improve if the short-term average was computed over
a longer time window to better average out this noise.  But how much
longer?  If the short-term average is too long, then the picks will be
late relative to the onset of the arrival.  The value for c3 = 0.4
(NSTA = 5 samples) was based on the observation that the shortest 
zero-to-peak times for first arrivals recorded on the UUSS analog
telemetry network are ~ 0.04 sec.

     I increased the time constant for the long term average
proportionately to that of the short-term average on the assumption
that the ratio of these time constants had previously been optimized.
Experiments using different values of c4 while holding c3 fixed at
0.4 seemed to support this assumption.  However, I intend to do a
grid search on a set of test events to find the optimum values of
c3 and c4.  The optimum values for c3 and c4 depend on some complex
tradeoffs, the description of which is beyond the scope of this short
document.

FURTHER SUGGESTIONS: OCTOBER 2, 2006.


c6 = DeadSta .  For 24-bit digital telemetry stations, we set this
parameter to 1000000 instead of 1200--essentially turning off the dead
station check.  That's because this check is primarily useful for analog
telemetry stations, for which telemetry dropouts can cause large amplitude
noise.  We do not tend to have this problem with the digital telemetry
stations we record.  Therefore, we set c6 = DeadSta to a higher value for
digital telemetry stations in order to help keep this check from causing
the rejection of arrival time picks in the codas of prior seismic events.

c7 = CodaTerm .  We use gain-dependent values so that the coda termination
values all correspond to approximately the same ground velocity values.

c1 = RawDatafilt .  In my 1998 report on suggested picker parameter changes,
I recommended that for broadband stations this parameter be set to a value
that sets the corner frequency of the highpass filter to 1.0 Hz.  One point
that I did not make very clear in this report is that the correct value of
c1 for a given highpass filter corner frequency depends on the sampling
rate.  If we define the corner frequency of the filter, fc, as the frequency
at which the amplitude response is 1/sqrt(2) times it's maximum value, then
it can be shown that

fc = (1 / (pi T) arcsin [ (1 - c1) / sqrt(2(1 + c1**2)) ]
 
c1 = [sin(2 pi fc T) -1] / [(2 ( sin(pi fc T) )**2) -1]
 
where T = the sampling interval and pi = 3.14159.  For example, if fc = 1.0
Hz, then c1 = 0.939 for T = 0.01 sec and c1 = 0.854 for T = 0.025 sec.