You are on page 1of 10

Bulletin of the Seismological Society of America, Vol. 101, No. 2, pp. 558567, April 2011, doi: 10.

1785/0120100182

A Comparison of Seismic Instrument Noise


Coherence Analysis Techniques
by A. T. Ringler, C. R. Hutt, J. R. Evans, and L. D. Sandoval*

Abstract The self-noise of a seismic instrument is a fundamental characteristic


used to evaluate the quality of the instrument. It is important to be able to measure
this self-noise robustly, to understand how differences among test configurations
affect the tests, and to understand how different processing techniques and isolation
methods (from nonseismic sources) can contribute to differences in results. We com-
pare two popular coherence methods used for calculating incoherent noise, which is
widely used as an estimate of instrument self-noise (incoherent noise and self-noise
are not strictly identical but in observatory practice are approximately equivalent;
Holcomb, 1989; Sleeman et al., 2006). Beyond directly comparing these two coher-
ence methods on similar models of seismometers, we compare how small changes in
test conditions can contribute to incoherent-noise estimates. These conditions include
timing errors, signal-to-noise ratio changes (ratios between background noise and
instrument incoherent noise), relative sensor locations, misalignment errors, proces-
sing techniques, and different configurations of sensor types.

Introduction
The quality of a seismogram containing event data is of- izing the incoherent noise of seismic instruments and digiti-
ten measured by the ratio of the event energy to the pre-event zers for many years (Hutt, 1990). A recently introduced
energy (Nolet, 2008), using a fixed time duration. In fact, three-instrument incoherent-noise calculation method (Slee-
many Earth studies sieve usable seismic event data by this en- man et al., 2006) shows great promise, in part because it
ergy ratio (e.g., Phillips and Stead, 2008). Seismic noise also reduces the number of a priori assumptions required. Many
limits our ability to determine earthquake focal mechanisms variables affect an instruments self-noise when estimated by
(Kafka and Weidner, 1979) and can disguise small-amplitude coherence analysis (e.g., variability in instrument quality,
phases in earthquake records (Banka and Crossley, 1999). differing site conditions, and differences in installation
These examples demonstrate that seismic noise is a funda- methods). Some of these can be falsely attributed to the instru-
mental limitation in interpreting and using seismic data in ments self-noise, when in fact they are simply the instru-
Earth studies. Differentiating between various noise sources ments response to different seismic and nonseismic noise
can be difficult because of the large number of variables con- sources. Additionally, it has been shown that an instruments
tributing to the noise seen in seismic records (atmospheric, incoherent-noise estimate can depend on the estimation tech-
cultural, installation, instrument, microseism, magnetic field, nique (Holcomb, 1990). To use either technique properly, one
electrical interference, site nonlinearity, and so forth). Many must understand what contributes to incoherent-noise
of these noise sources are always present in seismic records. estimates and how these contributions affect the resulting self-
However, reductions continue to be made in instrument noise noise estimate. Furthermore, if we wish to compare the
through improvements to instrument electronics, installation incoherent-noise estimates from different tests, we require
techniques, and reductions in sensitivities to various non- an accepted standard method for calculating incoherent noise
seismic sources (Wielandt, 2002). As instrument self-noise (Evans et al., 2010).
decreases, it becomes important that we be able to measure To understand variations arising from calculating instru-
and characterize it with greater accuracy, even in the presence ment incoherent noise, we review both the two-sensor
of comparable background signals from other sources. (Holcomb, 1989) and three-sensor (Sleeman et al., 2006)
The well-known two-sensor coherence analysis method incoherent-noise techniques; we include a brief discussion
of Holcomb (1989) has been a standard method for character- of other methods. By comparing incoherent-noise estimation
techniques for both real and synthetic noise data, we char-
*Working under contract with the Albuquerque Seismological acterize limitations in the techniques and suggest processing
Laboratory, Albuquerque, New Mexico 87198-2010. guidelines to produce robust, consistent incoherent-noise

558
A Comparison of Seismic Instrument Noise Coherence Analysis Techniques 559

estimates (Evans et al., 2010). Standardized processing Three-Sensor Method


methods and well-controlled test conditions allow one to
The three-sensor coherence analysis method of Sleeman
distinguish noise contributions caused by Earth-noise, instal-
et al. (2006) makes the same coherence assumption as the
lation methods, sensor variability, and so forth. In some
two-sensor method, but instead of correcting for instrument
cases, these noise sources raise an instruments apparent
response prior to calculating the noise, one computes and
incoherent noise by only a few decibels, but several small
uses a relative transfer function between instruments to cor-
contributions together can make a quiet sensor look quite
rect the cross-power spectra. Calculating this relative transfer
noisy (Wielandt, 2002). We hope that a better understanding
function requires a third colocated sensor, k, recorded in par-
of the variability in current analysis techniques will lead to
allel with the other instruments. The relative transfer function
improved understanding of noise contributions to seismic
between sensor i and sensor j is defined as hij  Pik =Pjk
data and ultimately to improved seismic data.
and allows one to remove errors in the noise estimates caused
by errors in a priori transfer functions. Instrument incoherent
Background
noise is now estimated by
The main assumption used in all coherence analysis
techniques is that the incoherent noise between different nii  Pii  Pij hij : (4)
instruments gives an approximation to the instruments
self-noise, making the cross power of the noise signals Comparing the noise in another domain (e.g., acceleration or
negligible (Wielandt, 2002). Our current understanding of displacement for a velocity seismometer) still requires that
instrument self-noise suggests this assumption is a good the noise be corrected to ground motion using an a priori
approximation in the absence of nonseismic sources. instrument response, but in this case the errors caused by
uncertainties in the transfer function do not contribute as
Two-Sensor Method strongly to the incoherent-noise estimate. Another advantage
of this method is that it introduces a third recording of
The two-sensor coherence analysis method of Holcomb ambient seismic noise, yielding a more robust estimate of the
(1989) is based on the assumption that two colocated coa- coherent portions of the total noise.
ligned seismometers have the same input seismic signal Xt.
One computes the power spectra of the time series from each Other Incoherent-Noise Methods
seismometer, Pii and Pjj (the power spectra from seismo-
meters i and j). One also computes the cross power, Pij , Before these methods of calculating incoherent noise
between the two time series. After correcting both the cross came into wide use, a method using two identical sensors
power and power spectra for instrument responses, the noise was applied (Peterson et al., 1980). This method assumes
of instrument i is estimated by the difference between its that the incoherent noise is distributed equally between the
power spectrum and the cross power: two sensors. Although this assumption is not always valid, it
does have the advantages of not introducing errors from the
nii  Pii  Pij : (1)
transfer function into the noise estimate and that it requires
The incoherent noise of instrument j is defined equivalently. only two sensors. (Differences between the true transfer
Because this noise estimate generally is complex (the cross functions of the two sensors still contribute errors.) Peterson
power is complex), one takes the modulus of nii to get a non- et al.s method differs slightly from the variant method of
negative real value for incoherent-noise power. Holcomb (1989), in which the incoherent-noise levels of
If the transfer functions of the two seismometers differ, the sensors are assumed equal.
this correction must be made before calculating nii , and a Sensor self-noise for electromagnetic seismometers
priori instrument-correction errors will contribute to the can be calculated by theoretical means (Rodgers, 1994).
instruments apparent self-noise (Holcomb, 1990). Holcomb Although Rogers method removes possible errors present in
(1989) presented a variation on his method for cases in which incoherent-noise estimates (e.g., local background noise and
the two instruments have equal self-noise so that nii  njj . atmospheric effects), it also removes elevated noise levels
In this case, the incoherent noise can be estimated by caused by possible variability in quality among sensors of a
1  2 given model. Variations between theoretical and incoherent-
nii  Pii ; (2) noise estimates are discussed in the companion paper by
2
Ringler and Hutt (2010).
where is the coherence, which is defined by
jPij j2 Methods
2  : (3)
Pii Pjj
In order to understand the limitations and characteristics
Again, this method requires that the power spectra be cor- of the two-sensor and three-sensor techniques, we compared
rected for instrument response unless the two instruments their performance with synthetic records. We generated
have identical transfer functions. a known pseudorandom signal in common between
560 A. T. Ringler, C. R. Hutt, J. R. Evans, and L. D. Sandoval

instruments and added pseudorandom noise separately to able to accelerometers and other seismic instruments (with a
each instruments output, comparing results from the two few key differences noted in passing). For completeness, we
techniques. We varied the percentage of smoothing of the discuss how the incoherent-noise estimation methods apply
spectra used, the length and overlap of the subwindows used to seismic instruments other than broadband velocity seis-
to compute power spectra, and the signal-to-noise ratio (SNR) mometers (e.g., geophones and accelerometers) in separate
between the known coherent signal and the noise we added to sections.
each time series. The only appreciable difference we found To look at incoherent noise across a maximum range of
was from varying the SNR (Fig. 1). As the SNR is increased, frequencies, we used data recorded at 40 samples per second
the estimation error via the Sleeman et al. (2006) method (sps; some had to be up-sampled from 20-sps legacy records)
stayed approximately constant, while that of the Holcomb and at 1 sps, the latter to better resolve the longest periods.
(1989) method increased sharply (Table 1). That is, the All up-sampling was done by frequency domain zero pad-
Sleeman et al. method is able to remove a larger ambient Earth ding (Lyons, 2004). For 40-sps data, we used total record
signal (and other coherent sources) from the incoherent-noise lengths of 220 samples, an interval of 07:16:54.4 (hh:mm:
estimates than is the Holcomb method. ss.s) duration. For 1-sps data, we used 215 samples, over
Our comparison of the two methods via synthetic data an interval of 09:06:08.0. With these long intervals and
suggests that only modest benefits are gained by making dual sample rates, we are able to extend our frequency range
small changes in the processing techniques used. To conform while still providing adequate overlap between the Welch
to a set standard, we used computational parameters speci-
method subwindows. We also up-sample the 1-sps data to
fied by Evans et al. (2010) and described subsequently in this
8 sps by using the same frequency domain method. With
paper. However, the SNR of the data used for calculating the
this combination of 40-sps and 8-sps records, we can reliably
incoherent noise of an instrument is an important constraint
estimate instrument noise over the frequency band
on our ability to estimate the noise accurately, so, when
0.00120 Hz.
comparing the two methods using real seismic data, most
The 40-sps data are divided into subwindows of 216 sam-
of our attention is on variability in the SNR (as a function
ples, overlapping one another by 7=8ths of the subwindow
of frequency), differences in sensor models, and common
instrument errors. In the rest of this paper, where not other- duration; the 1-sps data have subwindows of 211 samples with
wise indicated, we estimate the power spectra by way of the the same degree of overlap. For each subwindow, we applied a
modified Welch method (Welch, 1967). We also restrict the von Hann cosine-bell taper to reduce spectral leakage (von
general discussion of this paper largely to broadband weak- Hann weighting is sometimes called hanning). To increase
motion velocity seismometers. Nevertheless, they are applic- the coherence and minimize the variance of the result, it is
important to use significant overlapping (thus many and
redundant subwindows) when computing power spectral
SelfNoise Synthetic Random Data, 40 dB SNR
densities (PSDs) for coherence analysis because this improves
30 the effectiveness of the ensemble-averaging step of the Welch
method (Stearns and David, 1988); we did not apply any other
20 smoothing algorithms. To reduce the variance of spectral
PSD of DATA1 Z estimates further, it is important to use long-duration records
Signal
Holcomb SelfNoise (Oppenheim and Schafer, 1975). Although there are many
10
Power (dB)

Sleeman SelfNoise refined approaches to estimating power spectra to remove bias


Noise
caused by a single taper (Prieto et al., 2007), we have chosen
0
to use the (older) Welch standard so that our results can be
compared more easily with previous studies and because,
10 for real seismic data, such nuances are minority contributors
to error in the result.
20 Using these conventions, we used both techniques to
compute the power and cross-power spectra needed for esti-
30 mating the incoherent noise. For reference, we compared our
102 101 100 results to the new low-noise model (NLNM) of Peterson
Frequency (Hz) (1993). Finally, we used noise estimates derived from 1-sps
data for frequencies below 0.0073 Hz and estimates derived
Figure 1. Comparison between the two-sensor and three-sensor from 40-sps data for frequencies from there up to 16 Hz
self-noise estimation methods for synthetic data with an SNR of (corner of most antialias filters). (Higher frequencies are
40 dB. (PSD, power spectral density.) Solid light-gray line, input
signal; thick dashed black line, input signal with noise; thin dashed
desired by operators of many regional seismic networks and
black line, Holcomb noise estimate; thick dashed dark-gray line, can be provided by 200-sps, as described by Evans
Sleeman noise estimate; dotted light-gray line, input noise. et al., 2010.)
A Comparison of Seismic Instrument Noise Coherence Analysis Techniques 561

Table 1
Relative Mean Square Error* in Self-Noise Estimates Using Synthetic Pseudorandom
Noise for Different SNRs
Method 2 dB SNR 0 dB SNR 6 dB SNR 12 dB SNR 20 dB SNR 40 dB SNR

Holcomb (1989) 0.195 0.28 2 4.25 18.1 112.8


Sleeman et al. (2006) 0.155 0.18 0.13 0.155 0.16 0.155

*This error measurement is the mean square error (dB2 ) over the 1100 Hz frequency band.

Results time period at a good inland vault. Specifically, we apply both


incoherent-noise estimates to a set of Streckeisen STS-2
Here we discuss differences between the two methods broadband seismometers (Fig. 2 and Fig. 3) in a test
and how both results change under known errors in the test performed in the Albuquerque Seismological Laboratory
conditions.
(ASL) underground vault. We low-pass filtered all the time-
series data recorded during the test and used this filtered trace
Processing Techniques to pick a visually quiet time segment; however, all analyses
Using the conventions discussed previously in our paper were performed on unfiltered data. This selection process
for calculating noise estimates minimizes the error caused by helps remove Earth seismic signals (earthquake signals and
estimating the power spectra. If we assume that our data are noise bursts) and large instrument transients that can elevate
Gaussian and white, then our power spectra estimates fall the apparent incoherent noise of the instrument. (Instrument
within the 99% pointwise confidence band (Otnes and transients are an important quality issue but are beyond the
Enochson, 1972). We also must account for error in the scope of this study.) At high frequencies, and even more so
methods used and error in the test processes. in the microseism frequency band (about 0.50.05 Hz), the
As indicated by the relative mean square error for our three-sensor method gives lower incoherent-noise estimates
synthetic data (Table 1), we see that both methods do an than does the Holcomb method. We attribute these effects
excellent job of estimating the noise at low signal-to-noise in part to relatively large coherent seismic noise at these
ratios. However, with large SNR (ambient noise significantly frequencies (Fig. 1) and to greater contamination by modest
larger than instrument noise), the two-sensor method fails to orientation errors between vertical components of the sensors.
give reliable noise estimates (Fig. 1). At very large SNRs At lower coherent-noise levels (< 0:025 Hz, 80 s), we see
( 70 dB), both methods develop noticeable errors in the that the methods give similar incoherent-noise estimates; thus,
instrument incoherent-noise estimates. we may compare results from the two methods directly at very
For very large SNRs, we may assume that Pii is much
larger than nii and that the contribution of the modulus of the IncoherentNoise 2009 103 00 00 duration AmpORD
transfer function is negligible. Then, for the two-sensor 120
PSD of TST1 00 Z
method, we see that the noise grows approximately as the Sleeman Noise
Power (dB) relative to 1(m/s2)2/Hz

130 Holcomb Noise


negative of the cross power: NLNM
140
nii Pij : (5)
150
On the other hand, for the three-sensor method, we have
160
Pij Pik
nii : (6)
Pjk 170

We also know that the cross power can be written as 180


Pij  X  nii X  njj  , where X is the power of the
190
signal and * denotes the complex conjugate. Using our equa-
tion for the cross power in terms of the signal, we see that the 200
noise estimate from both methods grows approximately as a
210
quadratic in signal power, giving a possible explanation to the 103 102 101 100 101
limiting behavior under large SNR in both methods. We note Frequency (Hz)
that such large SNRs do not appear to arise in practice.
Figure 2. Incoherent-noise estimates for real data from an
STS-2 seismometer as recorded in the ASL underground vault, using
General Comparison amplitude operating range diagram conventions (AmpORD) of
Evans et al. (2010). Dashed dark-gray line, PSD of signal and noise;
We first compare both methods (using real data) in a best dashed black line, Sleeman noise estimate; dashed light-gray line,
case scenario in which all instruments are of similar model Holcomb noise estimate; solid black line, NLNM (new low-noise
type, are well aligned, and the data are taken during a quiet model) of Peterson (1993).
562 A. T. Ringler, C. R. Hutt, J. R. Evans, and L. D. Sandoval

Self-Noise Test 1, Vertical Components Duration, AmpORD


120 low frequencies without further complication. The differences
Noise of TST1 00 in the incoherent-noise estimates between these two tests at
130
Power (dB) relative to 1(m/s2)2/Hz

Noise of TST3 00
Noise of TST2 00
medium to high frequencies indicate that they are not directly
140 PSD of TST1 00 comparable there.
PSD of TST3 00
PSD of TST2 00 Based on the methods used for modeling the instrument
150 NLNM incoherent noise, the two-sensor method will always yield
160 incoherent-noise estimates that are equal to or greater than
those of the three-sensor method. This inequality can be seen
170
by comparing the terms used to remove the coherent signal.
180 For the two-sensor method, we have
190 jPij j  jX2  Xnii  njj j (7)
200 because nii njj  0. Similarly, for the three-sensor method
210 we see that
103 102 101 100 101  
Pij Pik 
 
 P  jX  2Xnjj  njj j:
2 2
Frequency (Hz) (8)
jk
Figure 3. Noise estimates between three STS-2 vertical-
Because n2jj
need not equal zero, we see that the three-sensor
component seismometers using data recorded in the ASL under-
ground vault with the three-sensor method. Thick dashed lines, method tends to remove more from Pii when estimating
Sleeman noise estimates; thin dashed lines, PSDs of signal and the incoherent noise, thus yielding lower incoherent-noise
noise; solid black line, NLNM. estimates.

Angle offset: Sleeman


Power (dB) relative to 1(m/s2)2/Hz

120
Baseline Sleeman
140 1 degree Sleeman
NLNM
160

180

200

103 102 101 100 101


Frequency (Hz)

Angle offset: Holcomb


Power (dB) relative to 1(m/s2)2/Hz

120
Baseline Holcomb
140 1 degree Holcomb
NLNM
160

180

200

103 102 101 100 101


Frequency (Hz)

Figure 4. Incoherent-noise estimates for known orientation errors between sensors. Baseline estimates are the incoherent noise estimated
when the instruments were coaligned accurately. Both estimation methods exhibit raised incoherent-noise estimates with increased orienta-
tion errors but only in the microseism band. Top, Sleeman noise estimate; bottom, Holcomb noise estimate.
A Comparison of Seismic Instrument Noise Coherence Analysis Techniques 563

Coherence Differences From Timing Errors, 10,000 seconds


Sensor Alignment
In regions with high ambient Earth-noise, alignment 1

error between the sensors becomes a major contributor to


the estimated incoherent noise of the instrument (Fig. 4). 0.98
We also note increased incoherent-noise estimates for the

Coherence
vertical component of a seismometer in the microseism band,
0.96
where noise is large (Fig. 2), perhaps due to small deviations
of the sensors from being slightly off axis from vertical.

2
0.94

Timing Errors
0.92
To understand how small timing errors between sensors 0 Seconds
1 Seconds 6400
contribute to incoherent-noise estimates, we use modified 1 Seconds 640
0.9
test data with known timing errors introduced by time 103 102 101
domain shifts and subsequent frequency domain interpola- Frequency (Hz)
tion (the latter to match sampling times between the shifted
and unshifted time series). Small timing errors (< 0:1 s for Figure 6. Coherence variations caused by known timing
errors for subwindow lengths of 640 and 6400 samples using a
these broadband sensors) produce only minor changes in 10,000 second time window. Solid-gray line, 1-s timing error
estimated instrument incoherent noise, but errors greater than with 6400-s window; dashed-gray line, 1-s timing error with
0:5 s yield incoherent-noise estimates significantly above 640-s window.
true for frequencies higher than 0.0125 Hz (Fig. 5). We
conclude that timing errors smaller than about 1=10th of the Sensor Configurations
sample interval (e.g., 0.0025 s for 40 sps data and 0.0005 s
for 200 sps data) do not contribute importantly to incoherent- We have shown in our tests with synthetic data that the
noise estimates of seismic instruments at low frequencies. It two-sensor method is unable to accurately resolve instrument
is likely that these results differ from previous studies (e.g., incoherent noise at high SNR (Fig. 1). This result also is evi-
dent in the real-data self-noise estimates for our colocated
McDonald, 1994) because of the different SNRs prevailing in
STS-2 seismometers (Fig. 2). Although the three-sensor
those studies, as well as differing subwindow durations used
method does not suffer to the degree of the two-sensor
by the various authors (Fig. 6).
method, we still suggest selecting visually quiet time seg-
ments (no large earthquakes, Earth-noise pulses, or instru-
Comparison of Timing Errors
ment glitches) when selecting the data. (This can be done by
120
Baseline Sleeman
applying a low-pass filter, as described previously in this pa-
130
Baseline Holcomb per.) Although transient signals caused by the instrument it-
.125 seconds Sleeman
self must be taken into account when considering the overall
Power (dB) relative to 1(m/s ) /Hz

.125 seconds Holcomb


140 NLNM performance of any seismic instrument, we feel other tests are
2 2

150
more appropriate for such noise (e.g., popcorn noise detec-
tion, as described by Hutt et al., 2009, or electromagnetic-
160 interference-sensitivity testing, radio-frequency-interference
170
sensitivity testing, temperature-sensitivity testing, and
pressure-sensitivity testing).
180

190
Two-Sensor Equal-Noise Variant Method
200
Although, the three-sensor method can yield superior
210 3 2 1 0 1 estimates of an instruments incoherent noise at frequencies
10 10 10 10 10
Frequency (Hz) greater than about 0.0125 Hz, it is possible under appropriate
conditions to reliably estimate an instruments high-
Figure 5. Instrument incoherent-noise levels computed with frequency incoherent noise using only two sensors. For such
known timing errors introduced between the various sensors. Thin performance, one requires that both sensors have approxi-
dashed black line, Sleeman noise with no time offset; thick dashed mately equal expected incoherent-noise levels. Using this
dark-gray line, Holcomb noise with no time offset; thin dashed
light-gray line, Sleeman noise with 0.125 second offset; thick variation on the two-sensor method, we are able to estimate
dashed dark-gray line, Holcomb noise with 0.125 second offset; the instruments incoherent noise reliably at frequencies from
solid black line, NLNM. 0.0125 Hz up to the corner of the antialiasing filter (Fig. 7).
564 A. T. Ringler, C. R. Hutt, J. R. Evans, and L. D. Sandoval

IncoherentNoise 2009, 103:00;00 duration, 36,000 seconds


120 noise (above 0:1 Hz). In cases like these, long-period inco-
PSD of TST1 00 Z
Sleeman Noise herent noise is estimated more accurately simply because the
130 Holcomb Noise
2 ambient Earth signal is more coherent over larger distances at
Holcomb Noise
Power (dB) relative to 1(m/s2)2/Hz

140 NLNM lower frequencies. Local site conditions can drive up noise
estimates at all frequencies (Fig. 8).
150
The reader may infer from the foregoing information
160 that highly local linear and nonlinear elastic effects (e.g.,
rock fractures, voids such as vaults, and weathered rock or
170
soil) commonly make Earth signals incoherent even at
180 directly adjacent sensors (including well under 1 m separa-
190
tion). This caveat applies particularly to horizontal transla-
tional seismometers because they are highly sensitive to
200 tiny variations in tilt and such tilts are affected strongly by
210 site imperfections, even in the rock of very good vaults.
103 102 101 100 101 Overcoming such site-condition problems requires cositing
Frequency (Hz)
all sensors on a single rigid (e.g., granite) slab that is sup-
Figure 7. Coherence ( 2 ) two-sensor self-noise variant method ported at only three points, guaranteeing that tilt and other
compared to estimates of incoherent noise by the three-sensor and signal sources truly are identical at all sensors. This need also
two-sensor methods. Dashed dark-gray line, PSD of signal and becomes important at frequencies below about 0.05 Hz
noise; dashed black line, Sleeman noise estimate; dashed light-gray because of tilt induced by atmospheric pressure variations
line, Holcomb noise estimate; solid dark-gray line, Holcomb noise
estimate 2 ; solid black line, NLNM.
(e.g., wind).

However, the full two-sensor incoherent-noise method gives


better estimates at low frequencies (Fig. 7). High-Frequency Weak-Motion
The performance differences between incoherent-noise
estimate methods for high-frequency weak-motion sensors
Sensor Locations (e.g., geophones) is similar to that of broadband sensors.
A key assumption in incoherent-noise coherence tech- That is, elevated ambient noise increases the estimation
niques is that the instruments are recording identical ambient errors in the two-sensor method (Fig. 9). The increase in
noise signals. This assumption can be violated when the cultural noise at high frequencies also introduces errors in
sensors are not directly adjacent to one another and mounted all incoherent-noise methods (e.g., 60 Hz noise from
on a surface that deforms in the presence of ambient Earth electromagnetic interference and vibrations from people
noise, especially when estimating high-frequency incoherent- and machinery).
IncoherentNoise 2008, 249:00:00 duration, 46,234 seconds IncoherentNoise 2009, 171:18:29 duration, 2621 seconds
120 100
PSD of ANMX 00 Z PSD of TST2 10 Z
Sleeman Noise Sleeman Noise
130 110
Holcomb Noise Holcomb Noise
Power (dB) relative to 1(m/s ) /Hz

Power (dB) relative to 1(m/s ) /Hz

NLNM 120 NLNM


140
2 2

2 2

130
150
140
160 150

170 160
170
180
180
190
190
200
200
210 210
103 102 101 100 101 102 101 100 101 102
Frequency (Hz) Frequency (Hz)

Figure 8. The incoherent-noise estimate of a Geotech KS- Figure 9. The instrument incoherent-noise estimate of a Geo-
54000, as recorded in a borehole at ASL. Thin dashed black line, tech GS-13 recorded in the ASL underground vault. Thin dashed
PSD of signal and noise; thick dashed light-gray line, Sleeman noise black line, PSD; thick dashed light-gray line, Sleeman noise esti-
estimate; thick dashed dark-gray line, Holcomb noise estimate; so- mate; thick dashed dark-gray line, Holcomb noise estimate; solid
lid black line, NLNM. black line, NLNM.
A Comparison of Seismic Instrument Noise Coherence Analysis Techniques 565

IncoherentNoise 2009, 259:06:15 duration, 1311 seconds


100 vailing test conditions and to select an appropriate method
PSD of TST2 10 E
105 Sleeman Noise when estimating instrument incoherent noise.
Holcomb Noise
From well-controlled real and synthetic incoherent-noise
Power (dB) relative to 1(m/s ) /Hz

NLNM
110
tests, we have shown how test conditions and methods con-
2 2

115 tribute to the overall incoherent-noise estimate for a seismic


120
instrument. An obvious shortcoming of this study, because all
of the tests were done either in ASL boreholes or its under-
125 ground vault, is that we have not constrained the contributions
130 of local site conditions, nonseismic sources, or installation
techniques (e.g., shielding from thermal convection). These
135
sources can contribute to the overall estimates of sensor inco-
140 herent noise because of changes in overall background noise
145 and instrument sensitivity to nonseismic sources and will
require additional work to be differentiated better. We note
150
101 100 101 102 that the variability in incoherent noise between like models
Frequency (Hz) of sensor at low frequencies is high enough that the same
set of instruments should be tested at multiple sites to under-
Figure 10. The incoherent noise estimates of an Episensor ES- stand how low-frequency incoherent-noise estimates vary
T strong-motion accelerometer. Thin dashed black line, PSD of sig- with location; unfortunately the logistic complications of such
nal and noise; thick dashed light-gray line. Sleeman noise estimate;
thick dashed dark-gray line, Holcomb noise estimate; solid black a study are prohibitive. We suspect that, at frequencies greater
line, NLNM. than about 0.01 Hz, there will be little to no change in instru-
ment incoherent-noise estimates.
Strong-Motion Accelerometers
Both noise-estimation techniques yield reliable, gener- Guidelines for Incoherent-Noise Estimates
ally similar estimates of incoherent noise for strong-motion Based on our analysis of incoherent-noise estimation
accelerometers, though low-noise accelerometers sometimes methods and data-gathering methods, we suggest guidelines
detect the microseismic peak (Fig. 10). This robustness can for estimating the incoherent noise of seismometers and, by
be attributed to the fact that the sensitivity of these instru- inference, of recorders; these suggestions are consistent with
ments is lower than that of weak-motion sensors. In fact, and enumerated by Hutt et al. (2009) and Evans et al. (2010).
the total power spectrum commonly matches the instrument All data for incoherent-noise estimates should be
incoherent noise rather closely outside the microseism band, obtained in the quietest possible location having the most
allowing one to use a single-sensor noise-estimation method stable possible temperature level. Most instruments are quite
(that is, by assuming that the PSD of an instruments total susceptible to temperature variation, but this source is not
output is entirely instrument noise during quiet time periods). inevitably a significant part of the sensors estimated incoher-
Within the microseism band, particularly in winter months, ent noise if the instruments are well-installed and properly
the two-sensor and three-sensor methods are more accurate protected during tests. Because different instruments will
(lower) than the PSD from a single accelerometer. respond differently to varying temperature and other nonseis-
mic sources, thus assuming the characteristics of incoherent
noise, it is important to isolate the instruments from these
Discussion noise sources so they do not elevate the incoherent-noise
estimates that were intended to reflect self-noise alone. To
By understanding the variability in incoherent-noise maximize the coherence of ambient Earth-noise measured
estimation techniques, one can understand and predict how during such tests, we suggest that the seismometers be
small changes in test conditions and instrument quality located as close as possible to one another and upon a three-
influence the estimated incoherent noise. Characterizing point-supported rigid slab (e.g., granite) that minimizes tilt
incoherent-noise variability between sensor designs and variations between adjacent sensors. When possible, three
individual sensors of the same design is important to the sensors with similar expected incoherent-noise levels should
design process, for objective comparison of sensors for be used in a given test run. Although it is possible that
manufacturers and users, and for understanding the physical instruments of the same model could have coherent non-
limits of detectable seismic signals (Laske, 2004). Although, Earth-noise (e.g., respond to temperature or tilt coherently;
the three-sensor method often is preferable, particularly for Wielandt, 2002), using instruments with significantly
estimating the incoherent noise for broadband weak-motion different self-noise levels can artificially elevate estimated
instruments, we have seen that the two-sensor method can incoherent-noise estimates of at least the quieter sensors. If
give comparable results under certain conditions (Fig. 8). three of the same model sensors are not available, then three
In any case, it is important to understand and record the pre- sensors of comparable anticipated quality should be used.
566 A. T. Ringler, C. R. Hutt, J. R. Evans, and L. D. Sandoval

Test data should be acquired for at least 24 hours so as to be References


likely to yield at least one 10-hr sample with a few or no
Banka, D., and D. Crossley (1999). Noise levels of superconducting
detectable transient Earth signals or instrument pulses. To gravimeters at seismic frequencies, Geophys. J. Int. 139, 8797.
maximize the meaningful bandwidth of the incoherent-noise Evans, J. R., F. Followill, C. R. Hutt, R. P. Kromer, R. L. Nigbor,
estimate, the data should be recorded simultaneously at A. T. Ringler, J. M. Steim, and E. Wielandt (2010). Method for
40 sps and 1 sps (and perhaps also at 200 sps), and the calculating self-noise spectra and operating ranges for seismo-
graphic inertial sensors and recorders, Seismol. Res. Lett. 81, no. 4,
estimated noise spectra from these time series should be
640646.
merged in the frequency domain at some frequency that is Holcomb, L. G. (1989). A direct method for calculating instrument noise
well-measured by both spectra. levels in side-by-side seismometer evaluations, U.S. Geol. Surv. Open
If three sensors of comparable quality are unavailable File Rep., 89-214, 34 pp.
but two are available, we recommend using the two-sensor Holcomb, L. G. (1990). A numerical study of some potential sources of error
in side-by-side seismometer evaluation, U.S. Geol. Surv. Open File
incoherent-noise method with the sampling and processing Rep. 90-406, 41 pp.
techniques as described previously in this paper. If both sen- Hutt, C. R. (1990). Standards for seismometer testing: A progress
sors used have similar expected incoherent-noise levels, then report, U. S. Geol. Surv. Open-File Rept. 20091055, 23 pp.
the 2 incoherent-noise technique for frequencies greater Hutt, C. R., R. L. Nigbor, and J. R. Evans (Editors), (2009). Proceedings of
the guidelines for seismometer testing workshop, Albuquerque, New
than 0.025 Hz yields similar estimates to those of the three-
Mexico, 910 May 2005 (GST2), U.S. Geol. Surv. Open File Rep.
sensor method when the minimum SNR of the test data is 2009-1055 48 pp.
below approximately 10 dB (Holcomb, 1989). At frequen- Kafka, A. L., and D. J. Weidner (1979). The focal mechanisms and depths of
cies lower than 0.025 Hz, Holcombs two-sensor method will small earthquakes as determined from Rayleigh wave radiation pat-
yield incoherent-noise estimates similar to those of the three- terns, Bull. Seismol. Soc. Am., 69, no. 5, 13791390.
Laske, G. (2004). The needs for/of low-frequency seismology, IRIS broad-
sensor method as long as the SNR is not exceptionally high. band instrumentation workshop, Incorporated Research Institutions
If these guidelines are followed, one can compare self-noise for Seismology, http://www.iris.edu/stations/seisWorkshop04/ (last ac-
estimates of various sensors yielding similar results unen- cessed June 2010).
cumbered by the processing methods per se. Lyons, R. G. (2004). Understanding Digital Signal Processing, Second Ed.,
Prentice-Hall, Englewood Cliffs, New Jersey, 688 pp.
We demonstrate that errors from a number of sources are
McDonald, T. S. (1994). Modified noise power ratio testing of high
present in incoherent-noise estimates. At present, however, resolution digitizers, Sandia National Laboratory Report, SAND94-
we have not introduced an estimate of these errors because 0221, 37 pp.
(1) errors at long periods (frequencies below 0:025 Hz) are Nolet, G. (2008). A Breviary of Seismic Tomography Imaging the Interior of
within 1 dB between the two methods evaluated, (2) our the Earth and Sun, Cambridge University Press, Cambridge, England,
344 pp.
methods for estimating the power spectra are at the 99% con- Oppenheim, A. V., and R. W. Schafer (1975). Digital Signal Processing,
fidence level, and (3) variations in incoherent-noise estimates Prentice-Hall, Englewood Cliffs, New Jersey, 585 pp.
for different data sets (if all were obtained when visually Otnes, R. K., and L. Enochson (1972). Digital Time Series Analysis, Wiley,
quiet) have standard deviations of only a few decibels; that New York, 467 pp.
Peterson, J. (1993). Observations and modeling of seismic background
is, the quantifiable error is small. On the other hand, errors in
noise, U.S. Geol. Surv. Open File Rep. 93-322, 94 pp.
estimating incoherent noise using differing sensors or instal- Peterson, J., C. R. Hutt, and L. G. Holcomb (1980). Test and calibration
lation methods contribute more to the variability in such of the seismic research observatory, U.S. Geol. Surv. Open File
estimates; because such errors are not easily quantified, we Rep. 80-187, 89 pp.
recommend that the comparable model(s) of sensors be used Phillips, W. S., and R. J. Stead (2008). Attenuation of Lg in the western U.S.
using USArray, J. Geophys. Res. 35, 5.
and their installation methods be well documented and as Priesto, G. A., R. L. Parker, D. J. Thomson, F. L. Graham, and R. L. Vernon
close to the guidelines as possible. (2007). Reducing the bias of multitaper spectrum estimates, Geophys.
J. Int. 171, 12691281.
Rodgers, P. W. (1994). Self-noise spectra for 34 common electromagnetic
Data and Resources seismometer/preamplifier pairs, Bull. Seismol. Soc. Am. 84, no. 1,
222228.
Ringler, A. T., and C. R. Hutt (2010). Self-noise models of seismic instru-
All data used in this study were recorded at ASL and are
ments, Seismol. Res. Lett. 81, no. 6, 972983.
available from us upon request. Sleeman, R., A. van Wettum, and J. Trampert (2006). Three-channel
correlation analysis: A new technique to measure instrumental noise
of digitizers and seismic sensors, Bull. Seismol. Soc. Am. 96, no. 1,
Acknowledgments 258271.
Stearns, S. D., and R. A. David (1988). Signal Processing Algorithms,
We thank D. Anderson for help with running the coherence azimuth Prentice-Hall, Rochelle Park, New Jersey, 349 pp.
software that we used to align the sensors under test. We also thank D. Hart Welch, P. D. (1967). The use of fast Fourier transform for the estimation of
for suggestions on the processing techniques and L. G. Holcomb for many power spectra: A method based on time averaging over short, modified
useful conversations about our methods and the two-sensor technique in par- periodograms, IEEE Trans. Audio Electroacoust. AU-15, 7073.
ticular. We also thank L. G. Holcomb and R. Sleeman for very helpful reviews Wielandt, E. (2002). Seismic sensors and their calibration, in New Manual
of the manuscript. Finally, we thank an anonymous reviewer for a very of Seismological Observatory Practices, GeoForschungsZentrum,
thoughtful and thorough review that substantially improved the manuscript. Potsdam, Germany, 49 pp.
A Comparison of Seismic Instrument Noise Coherence Analysis Techniques 567

U.S. Geological Survey Honeywell Technology Solutions Incorporated


Albuquerque Seismological Laboratory P. O. Box 82010
P. O. Box 82010 Albuquerque, New Mexico 87198-2010
Albuquerque, New Mexico 87198-2010 (L.D.S.)
(A.T.R., C.R.H.)

U.S. Geological Survey


345 Middlefield Road, MS-977
Menlo Park, California 94025
(J.R.E.) Manuscript received 29 June 2010

You might also like