GEOPHYSICS, VOL. 71, NO. 4 JULY-AUGUST 2006 ; P. SI177SI187, 14 FIGS., 2 TABLES. 10.1190/1.

2209748

Imaging passive seismic data

Brad Artman1

INTRODUCTION ABSTRACT
Imaging passive seismic data is the process of synthesizing the wealth of subsurface information available from reection seismic experiments by recording ambient sound using an array of geophones distributed at the surface. Crosscorrelating the traces of such a passive experiment can synthesize data that are identical to actively collected reection seismic data. With a correlation-based imaging condition, waveequation shot-prole depth migration can use raw transmission waveelds as input for producing a subsurface image. Migration is even more important for passively acquired data than for active data because with passive data, the source waveelds are likely to be weak compared with background and instrument noise a condition that leads to a low signalto-noise ratio. Fourier analysis of correlating long eld records shows that aliasing of the waveelds from distinct shots is unavoidable. Although this reduces the order of computations for correlation by the length of the original trace, the aliasing produces an output volume that may not be substantially more useful than the raw data because of the introduction of crosstalk between multiple sources. Direct migration of raw eld data still can produce an accurate image, even when the transmission waveelds from individual sources are not separated. To illustrate direct migration, I use images from a shallow passive seismic investigation targeting a buried hollow pipe and the water-table reection. These images show a strong anomaly at the 1-m depth of the pipe and faint events that could be the water table at a depth of around 3 m. The images are not clear enough to be irrefutable. I identify deciencies in survey design and execution to aid future efforts. Passive seismic imaging is an example of waveeld interferometric imaging. In this case, the goal is to produce subsurface structural images by recording the ambient noise eld of the earth using surface arrays of seismometers or geophones. The images produced with this technique are directly analogous to those produced with conventional reection seismic data. In the exploration seismic community, the words imaging and migration often are used synonymously. Likewise, this paper presents the processing of passive seismic data as a migration operation. Using subsurface sources to image the subsurface was introduced by Claerbout 1968 . That work provided a 1D proof that autocorrelation of time series collected on the surface of the earth can produce the equivalent of a zero-offset time section. Subsequently, Zhang 1989 used plane-wave decomposition to prove the result in three dimensions over a homogeneous medium. Derode et al. 2003 developed the Greens function of a heterogeneous medium with acoustic waves via correlation and validated the theory with an ultrasonic experiment. Wapenaar et al. 2004 used one-way reciprocity to prove that crosscorrelation of traces of the transmission response of an arbitrary medium synthesizes the complete reection response, in the form of shot gathers, collected in a conventional active-source experiment. Schuster et al. 2004 showed that the Kirchhoff migration kernel one uses to image correlated gathers is identical to the migration kernel one uses to migrate prestack active data when one assumes that impulsive virtual sources are located at all receiver locations. In summary, it is well established that time differences calculated by correlation are informative about the medium between and below the receivers. Data collected using subsurface sources are called transmission waveelds or T. The conventional reection-seismic data volume is referred to as R. The rst section of this paper explains the basic kinematics of extracting R from T. Next, I introduce the use of shotprole wave-equation depth migration to create an image of subsurface structure without rst correlating the traces from the transmission waveelds. One important attribute of truly passive data is that the bulk of the raw data is likely to be worthless. Useful seismic energy captured in

Manuscript received by the Editor March 19, 2005; revised manuscript received January 24, 2006; published online August 17, 2006. 1 Stanford University, Stanford Exploration Project, Mitchell Building, Department of Geophysics, Stanford, California 94305.E-mail: brad@sep. stanford.edu. 2006 Society of Exploration Geophysicists. All rights reserved.

SI177

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

SI178

Artman produced from a single-plane wave will produce another plane wave. However, each planar reection is moved to the lag time associated with a two-way trip from the surface to the reector. Correlation removes the wait time for the initial arrival and maintains the time differences between the direct arrival and the reections. Summing the correlations from a full suite of plane waves builds hyperbolic events through constructive and destructive interference. Analysis of seismic data in terms of plane-wave constituents is a commonly invoked tool in seismic processing. Summing the correlations from incident plane waves is a plane-wave superposition process that builds hyperbolas. Passive seismic imaging is predicated on raypaths bouncing every which way from every direction. Cartoons depicting the transmission experiment always leave something out, which causes an inconsistency that needs more raypaths and receivers to explain it. Unfortunately, the trend continues nearly forever. Such a complication arises with the inclusion of a second reector. The two reection rays correlate positively with each other. The two travel paths share the time through the shallow layer, so they correlate at a lag or time difference equal to the two-way traveltime through the deep layer. Such a correlation is not a problem, however. An upcoming intrabed multiple event arrives after the direct arrival that was delayed by the two-way traveltime through the deeper layer. This multiple event has a polarity opposite to that of the direct arrival after it changes its propagation direction from upward to downward. The multiples arrival at receiver r2 is delayed, compared with the timing of the direct arrival at receiver r1 by the same lag as that of the correlation between the reectors, and it has the opposite sign. Therefore, the internal multiple cancels potential artifacts of the correlation. This shows the importance of modeling transmission data with a two-way extrapolator. Without all possible multiples, correlation artifacts will quickly overwhelm the earths structure in the correlated output. Crosscorrelation of each trace with every other trace resolves the three main difculties associated with transmission data: timing, waveform, and interference. First, the output of the correlation is in lag units, which when multiplied by the time-sampling interval, provide the time delays between like events on different traces. The zero lag of the correlation is the zero time for the synthesized shot gathers. Second, each trace records the character and duration of the incident energy as it is reected at the surface. This becomes the source wavelet analogous to a recorded vibrator sweep. Third, overlapping wavelets are separated by correlation. To calculate the Fourier transform of the reection response of the subsurface, R xr,xs, , Wapenaar et al. 2004 proved

the transmission waveeld may include random distributions of subsurface noise, downhole source noise, or earthquake arrivals. Assuming that these noise sources are not happening continuously, and not knowing when they will occur, the passive seismologist must record continuously. Through Fourier analysis, one can explore the ramications of processing an entire passively recorded data volume rather than individual waveelds from separate sources. Finally, results from a small eld experiment are presented. Passive data were recorded over several days on clean beach sand with a 2- 2-m array of 72 geophones. A hollow pipe was buried beneath the array to provide an imaging target. Several data volumes from different times of the day and various preprocessing strategies were imaged for comparison with an active survey collected at the same location.

FROM TRANSMISSION WAVEFIELDS TO REFLECTION WAVEFIELDS

Figure 1, simplied for clarity, shows the basic kinematics exploited in processing transmission data. The gure includes two recording stations capturing an approximately planar wavefront emerging from a two-layer subsurface. Figure 1a shows a raypath associated with the direct arrival and a raypath reected both at the free surface and at a subsurface interface. The second travel path labeled reection ray has the familiar kinematics of the reection seismic experiment with a source that is excited at receiver one. The transmission waveeld is depicted in Figure 1b. Wavelet polarity is appropriate for direct arrivals and for reection. The three main features of transmission waveelds can be appreciated here. First, the exact timing of the arrival is unknown. Second, the phase and duration of transmitted energy are unknown and likely are complicated. Third, if the incident wavetrain coda is long, arrivals in the transmission record can interfere. If we choose trace r1 as the comparison trace, Figure 1c depicts the correlation spikes associated with the arrivals in the data panel of Figure 1b, where  is correlation. A solid line with linear moveout is superimposed across the correlated traces corresponding to the direct arrival recorded at each receiver location. The dashed line on Figure 1c has hyperbolic moveout. However, no correlation peak exists on the r1  r1 trace under the hyperbola. Not drawn, the second arrival on r2 will have a counterpart on r1 from a ray reaching the free surface farther to the left of the model. In fact, the correlations

2R R xr,xs,

xs xr T x r, ,
Dm

T * x s, ,

d2 , 1

Figure 1. a An approximately planar arrival with rays showing important propagation paths for passive imaging. b Idealized traces from a transmission waveeld. c A shot gather reection waveeld synthesized using trace r1 as the source. Many details explained in the text are omitted here for simplicity.

where * represents conjugation. The vector x will correspond herein to horizontal coordinates, where subscripts r and s indicate different station locations from a transmission waveeld. After correlation, r and s acquire the meaning of receiver and source locations, respectively, that are associated with an active survey. The right-hand side RHS of equation 1 represents summing correlations of windows of passive data around the occurrence of individual sources from loca-

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Imaging passive seismic data tions . The symbol Dm represents the domain boundary that surrounds the subsurface region of interest on which the sources are located. The transmission waveelds T xr, , contain the arrival and reverberations from only one subsurface source. For exact synthesis of the reection experiment, impulsive sources should completely surround the volume of the subsurface that one is trying to image.Alternatively, many impulses can be substituted by a full suite of plane waves emerging from all angles and azimuths, as in the kinematic explanation above.

SI179

DIRECT MIGRATION
Artman and Shragge 2003 introduced the applicability of direct migration for transmission waveelds. Artman et al. 2004 provided the mathematical justication for zero-phase source functions. Shragge et al. 2006 showed results for the special case of imaging with teleseisms. Direct migration of transmission waveelds requires an imaging algorithm composed of waveeld extrapolation and a correlation-based imaging condition. Shot-prole wave-equation depth migration Claerbout, 1971 fullls these requirements. Shot-prole migration uses one-way extrapolators to independently extrapolate upgoing energy, U, and downgoing energy, D, through the subsurface velocity model. The term U, extrapolated acausally, is a single shot gather. The term D, extrapolated causally, is a waveeld modeled to mimic the source used in the experiment. I will dene the causal extrapolation operator, E+, and the acausal operator, E. These are adjoint phase-shift operators calculated Claerbout, 1985 by

E = eikz z ,
where

kz =

2 kr ,

kr is the wavenumber dual variable for receiver location xr, and s is slowness. These one-way operators cannot extrapolate evanescent waves, but that limitation is not crippling. To avoid adding needless complexity, I do not include the details of the more accurate phaseshift-plus-interpolation PSPI extrapolation operators Gazdag and Sguazzero, 1984 used for migration examples herein. Waveelds are extrapolated to progressively deeper levels, z, by recursive application of E:

Uz+1 kr ;xs, Dz+1 kr ;xs,

= EUz kr ;xs, = E+Dz kr ;xs,

and .

4 5

nal of a square matrix multiplied by the second signal vector. As such, the two diagonal-square operations are commutable. This means that the correlation required to calculate the earths reection response from transmission waveelds can be performed after extrapolation as well as at the acquisition surface, if the velocity model is accurate. Table 1 demonstrates pictorially how direct migration of transmission waveelds ts into the framework of shot-prole migration to produce the 0th and 1st depth levels of the zero-offset image. The correlation in the imaging condition takes the place of preprocessing the transmission waveeld. The summations over shot locations and frequency in equation 6 are omitted to reduce complexity. Note, however, that the sum over shot locations in equation 6 takes the place of the integral over source locations in equation 1. Also, after the rst extrapolation step, using the two different phase-shift operators, the two transmission waveelds are no longer identical and can be redened as U and D. This is noted with superscripts on the T waveelds at depth. Transmission waveelds from 225 impulsive sources across the bottom of the velocity model in Figure 2c were modeled with a twoway time-domain extrapolation program. The time sampling rate was 0.004 s, and the source functions were Ricker wavelets with a dominant frequency of 25 Hz. Figure 2a is the image created by correlating each transmission waveeld implementing equation 1 and migrating the shot gathers by the algorithm described on the left side of Table 1. Figure 2b was produced by migrating each transmission waveeld directly and summing the images as described by the right column of Table 1. They are identical to machine precision. The term T is the superposition of U and D. Extrapolating the transmission waveeld with a causal phase-shift operator propagates the energy reected downward from the free surface that is the source function for later reections. This allows the use of T in place of D in shot-prole migration. Extrapolating the transmission waveeld with an acausal phase-shift operator, reverse-propagates the upcoming events reected from subsurface structure. This allows the use of T in place of U in shot-prole migration. In effect, the extrapolations redatum the experiment to successively deeper levels in the subsurface. At those deeper levels, the waveelds are correlated in accordance with the general theory of equation 1. The physics captured in the formulation of shot-prole migration remains the same, regardless of the temporal or areal characteristics of initial conditions in the waveelds. Extraction of only the zero lag of the correlation for the image discards energy in the two waveelds that is not collocated. This inTable 1. The equivalence of shot-prole migration of reection data and direct migration of transmission waveelds.2 Shot-prole migration Uz=0 xr ;xs, E Uz=1 xr ;xs,
2

The imaging condition combines the two waveelds and outputs a subsurface image. The zero-offset imaging condition for shot-prole migration is dened as the zero lag of the crosscorrelation of the two waveelds

iz xr =
xs

Uz xr ;xs,

* Dz xr ;xs,

Dz=0 xr ;xs, E+

Transmission imaging  Tz=0 xr ; , = Tz=0 xr ; , E

= Tz=1 xr ; ,

The sum over frequency extracts the zero lag of the correlation. The sum over shots, xs, stacks the overlapping images from all the individual shot gathers. Extrapolation in the Fourier domain, across a depth interval, is a diagonal matrix whose values are phase shifts for each frequencywavenumber component in the waveeld. Correlation of two equallength signals in the Fourier domain has one signal along the diago-

E+

+ Tz=1 xr ; ,

Dz=1 xr ;xs,

T xr, ,t are the waveelds of equation 1. The term xs has a meaning similar to . The sums xs, and produce the image iz xr for both methods. Only the rst and second levels of the recursive process are depicted.

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

SI180

Artman identical image, not shown, was produced by correlating and then migrating the synthesized gathers. Combining the weak redundant signal within each shot gather with all of the others through migration produces a very good result, despite the low quality of the shot gather produced with the same data.

cludes energy that has been extrapolated in the wrong direction because the same data are used initially for both U and D waveelds . Conveniently, the only modication needed to make a conventional shot-prole migration program into a transmission-imaging program is to use T as D instead of using a modeled waveeld. Very important among the motivations for migrating transmission waveelds, is the need to increase the signal-to-noise ratio of the output. If the experiment records only a small amount of energy, the synthesized gathers from correlation may be completely uninterpretable. The synthesized shot gather in Figure 3a was produced using the same 225 transmission waveelds used in Figure 2, although here each was convolved with a random source function. The waveelds were individually correlated and summed. A few events can be seen centered around 4000 m, but the gather is dominated by noise. Figure 3b shows an image produced by direct migration of the transmission waveelds, followed by stacking the 225 images. An

TRULY PASSIVE DATA

I differentiate passive seismic data from transmission waveelds when the timings of individual sources are unknowable and may overlap. For passively collected transmission waveelds, the time axis and the shot axis are combined naturally. The parameterization of eld data, T f , is in terms of xr and time rather than xr, ,t, which is the parameterization of equation 1. Within , many individual sources, and the reections that occur t seconds afterward, are distributed randomly across the total recording time. Both t and represent the real time axis and therefore share the sampling rate t = . However, I will parameterize waveelds with the understanding that max t is the two-way time to the deepest reector of interest, and is the time axis from the beginning to the end of the total recording time. The difculty in separating the data into constituent waveelds from a single source, parameterizing as function of , comes from not knowing the time separation between events and from the possibility that the sources may overlap in time and space. If it is impossible to separate eld data into individual waveelds, transforming the long data to the frequency domain has important

Figure 2. Images created by a shot-prole migration of correlated shot gathers and b summing of images from direct migration of individual transmission waveelds. c The velocity model used for modeling and migration.

Figure 3. a A synthesized shot gather implementing equation 1, with a source trace from x = 4000 m from a passive data set modeled with the velocity model in Figure 2c and random source functions. b A zero-offset image produced by direct migration of the modeled data used to create the shot gather in a .

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Imaging passive seismic data ramications on further processing. The denition of the discrete Fourier transform DFT for an arbitrary signal f can be evaluated for a particular frequency, , as

SI181

Tf

1 8

DFT
n=1

gn t

11

= DFT f

1 n

n 1

f n
n=0

i n

where n is the number of samples in f. The frequency-domain dual variable of is , and I will use for the frequency domain dual variable of t. If the long function f is broken into N sections, gn t , of length n /N, the amplitude of any frequency = also can be calculated as

N n=1

DFT gn t

1 N

DFT
n=1

gn t 8

where the number of samples in each gn t was 512. An identical result can be produced by subsampling T f by 8. To facilitate plotting, the trace was padded with zeros, although the function actually is undened at times greater than the last visible energy. Both versions of the output the middle and bottom traces have late correlation lags that do not correspond to physical raypaths found in R. Most efciently, they are windowed away in the time domain before further processing. From the bottom trace, notice that increasing the decimation factor to 16 would alias the symmetric, acausal lags into the reection response. To avoid this problem, the short time windows, gn t , must be more than twice as long as the time to the deepest reection of interest. In this example, the rst two correlations of are R. Producing the bottom trace involves O 10 R fewer computations than producing the middle trace.

by simply changing the order of summation for convenience and scaling the result. Therefore, F calculated with equation 8 is an N-times-subsampled version of F calculated with equation 7. The center equality of equation 8 the sum of short transforms shows that frequency components common to the two transforms are identical after a simple amplitude scaling. The right equality the sum of time windows shows that the long transform stacks the constituent windows at common frequencies. The two alternatives interpreted together show that subsampling the long transform aliases the time domain. This relationship between sampling and aliasing in a signals representation in two domains is the Fourier dual to subsampling a time signal to reduce its Nyquist frequency at the risk of aliasing high frequencies. Passive seismic eld data, T f , take the role of the general signal f , and the gn t are T xr, ,t =n as well as background noise beis a subset from tween sources . Equation 8 states that T f xr, T f xr, and that

Waveeld summation
Unfortunately, the beginning of is not always R. The stacking of R waveelds implicit in processing eld data can be explored by considering eld data composed of transmission waveelds a xr,t and b xr,t from individual sources. As is the case for the input trace of Figure 4, the transmission waveelds are placed on the eld record at unknown times a and b so that

Tf

=a*

+b*
a b

.
b a

12

Correlation in the Fourier domain to implement equation 10 yields

T f T* = AA* + BB* + AB*ei f

+ BA*ei

. 13

T f xr,t =

T xr, ,t .

Therefore, implementing a single correlation of passive eld data does not produce R, but instead some unknown quantity : R

The sum of the rst two terms is the result dictated by equation 1. The last two terms are crosstalk. If b a max t , one term is acausal, and the other is at late lags that can be windowed away in the time domain. That was the case in the middle trace of Figure 4. If b a max t , the cross terms are included in the correlated gathers. In light of the previous section, the last scenario will be the predominant situation if the source delays a and b are truly random. For this reason, Figure 4 is not a fair representation of eld data because values for a,b were selected very carefully.

x ,x , R r s

1 T f x r, N

T * x s, f

10

where N is the number of time windows of length max t into which the total recording time can be divided and where T f can be calculated with equation 9. The relationship is undened at frequencies . Figure 4a shows a processing ow of a simple 1D time-domain signal, and a zoomed-in view of each trace appears in Figure 4b. The with 4096 samples. It represents a top trace is the input signal T f long eld recording, including three transmission waveelds exploring a subsurface with a single layer. The middle trace is its autocorre. = DFT1 T f T* lation calculated with all frequencies, R f The bottom autocorrelation was calculated by rst summing eight constituent windows of T f before discrete Fourier transformation DFT , so that

Figure 4. The top row is a trace representing three identical subsurface sources in a model with a single reector. The middle row is an autocorrelation of top trace. The bottom row is an autocorrelation of the top trace, performed with every eighth frequency and padded with zeros to facilitate plotting . Direct arrivals and reections from each source do not interfere. Column b is the rst 1.0 s of column a for all rows.

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

SI182

Artman aliased into the early lags of the bottom result at this level of decimation. Both methods produce the wrong result at almost all times. However, they are correct and identical at zero lag. This analysis shows that without invoking long, purely random source functions, a single correlation of passive data will not model shot gathers from a reection survey. In the next section, I use 2D synthetic data to show the extent of the problem and the ability of direct migration to produce a correct subsurface image, despite the summation of waveelds.

Next, we redene A and B as the impulse response of the earth, Ie, convolved with source functions, F, which now contain their phase delays. As such, the cross terms of equation 13 are

AB* = FaIe FbIe

= F aF *I 2 = F cI 2 . b e e

14

As do the rst two terms in equation 13, the cross terms here contain the desired information about the earth. However, the source function f c included is not zero phase. These terms may be the otherterms or virtual multiples mentioned in Schuster et al. 2004 . If the 2 source functions are random series, terms with residual phase FcIe within the gathers will decorrelate and diminish in strength as the length of f and the number of cross terms increases. Although we hope to collect many sources, it probably is unreasonable to expect very many of them to be random series of great length. That is not probable if the noise sources have similar source mechanisms. Inclusion of these cross terms in the correlation explains why R produced with equation 10 is not equivalent to R from equation 1. Virtual multiple events resulting from ns sources likely will be more problematic than conventional multiples because every reector can be repeated ns ! / ns 2 ! times. The ratio of desirable zero-phase 2 terms, Ie, to cross terms, F jIe , decreases as 1/ ns 1 if the source terms contain the same wavelet. Figure 5 shows the effect of the cross terms when one is calculating correlations. The gure is directly analogous to Figure 4, although with a more realistic input trace. First, there are overlapping source-reection pairs: The second source arrives at the receiver before the reection from the rst source arrives. Second, the direct arrivals are spaced randomly along the time axis, in contrast to Figure 4, in which the direct arrivals were placed at samples 1, 513, and 2049. That contrivance allowed the summing of constituent time windows 256 or 512 samples long without introducing residual phase functions as described in equation 14. A seismologist using passive seismic data to produce a zero-offset trace from R, the bottom trace in Figure 4, cannot do so by autocorrelating the top trace in Figure 5. The middle trace is the autocorrelation with input T . The bottom autocorrelation was computed after rst stacking eight constituent windows. In contrast to the two autocorrelations in Figure 4, here not even the early lags are the same. The presence of more aphysical correlations in the middle trace for this example has caused the symmetric peaks at late lags to be

Examples with synthetic data

To demonstrate the effects of waveeld summation, I use two synthetic acoustic passive-data sets. Transmission waveelds from 225 impulsive sources across the bottom of the velocity models were propagated with a two-way extrapolation program, as before. To then simulate a passive recording campaign, I convolved a unique source function with each waveeld before summing all of them together. The total length of the source-function trace mimics the duration of the recording campaign. The shape, time location, and length of the wavelet used within the source-function trace, otherwise zero, embodies my assumptions on the nature of the ambient subsurface noiseeld. The source functions incorporate many of the features of the toy example in Figures 4 and 5. Figure 6 shows synthetic data from a model containing two diffractors in a constant-velocity medium. Figure 6a is a transmission waveeld from a subsurface source below the far left of the model, whereas the source in Figure 6b was below x = 5000 m. Source wavelets were 50 ms long, with a dominant frequency of 25 Hz, and were placed on the time axis to align their direct arrivals. This experiment is the 2D analog to Figure 4. Figure 6c is the sum of all 225 similar waveelds from shots below the entire width of the surface array. The coherent summation of the direct arrivals makes a strong plane wave at t 0.6 s, and the diffractors are captured well by means of constructive interference. Crosstalk between the waveelds has been canceled using destructive interference. The summed waveeld below the strong plane wave is the same as would be collected from a zero-offset reection experiment. Correlation of the traces in Figure 6c does not produce shot gathers. Instead, each xs plane from contains a series of plane waves that completely mask R the diffractors. However, the slice from the correlated volume xs = xr is interpretable. The autocorrelation produces a result similar to Figure 6c, in which the plane wave is moved from t 0.6 s to zero lag. In contrast, Figure 7 shows synthetic data from the same model, with the addition here of random phase delays for the source functions throughout the experiment. Figure 7ac correspond to its respective panels in Figure 6. Note that the minimum traveltime of the direct arrivals in Figure 6a and b is the same whereas those in Figure 7a and b are not the same. The strong plane-wave and coherent diffractors from Figure 6c have been replaced by an uninterpretable superposition. Correlating traces from this waveeld returns that is R completely unusable, rather than the reection waveeld R. This is the 2D equivalent to Figure 5. Data also were synthesized using a model that contained two synclines. Source functions were identical to those used for the diffractor exercise above. Figure 8a shows summed waveelds with correction for the onset time of the 225 subsurface sources, and Figure

Figure 5. The top row is a trace representing three identical subsurface sources as in Figure 4. The middle row is an autocorrelation of the top trace. The bottom row is an autocorrelation of the top trace, performed with every eighth frequency and padded with zeros to facilitate plotting . The direct arrival of the second source is before the reection resulting from the rst source. Column b is the rst 1.0 s of column a for all rows.

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Imaging passive seismic data 8b shows the waveelds without correction. Again, correlating the superposition of waveelds, from either panel, returns an unusable data volume rather than shot gathers of R.

SI183

DIRECT MIGRATION OF PASSIVE DATA

If we use equation 10 to correlate eld data because we are unable to collect T as a function of individual source functions , we cannot process the result of the correlation by using conventional reection data tools. Without separating the waveelds from each source for independent processing, we cannot eliminate all time delays completely, as shown in equation 13. However, the superposi-

tion of waveelds inherent in the Fourier transform of passive eld data, equation 8, has a direct counterpart to migration of active source data. Shot-prole depth migration produces the correct image if the source waveeld, D, is correct for the data waveeld, U. Shot-prole migration becomes plane-wave migration if all shot gathers are summed for the upcoming waveeld xsU, and if a horizontal planar source is used for the downgoing energy, xsD. The waveeld in Figure 6c could be imaged successfully with a horizontal plane wave at t = 0.6 s modeled for D. This shot-prole migration method can introduce crosstalk between the individual experiments while we are attempting to process them all together. Such crosstalk will be avoided in two cases. First,

Figure 6. Representative transmission waveelds with sources a below x = 1200 m and b below x = 5000 m. c The sum of 225 transmission waveelds from sources evenly distributed across the bottom of the model. Minimum traveltime of the direct arrival in each waveeld is the same.

Figure 7. Representative transmission waveelds with sources a below x = 1200 m and b below x = 5000 m. c The sum of 225 transmission waveelds from sources evenly distributed across the bottom of the model. In contrast to Figure 6c, the minimum traveltime of the direct arrival in each waveeld here is random.

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

SI184

Artman Table 2 demonstrates pictorially how direct migration of passive eld data ts into the framework of plane-wave migration analogous to Table 1. Summation of shot gathers from an active data set before migration is represented on the left. The top row of the right column recognizes the sum over the source axis inherent in the Fourier transform of passive eld data shown in equation 8. Regardless of the initial conditions of U, D, or T, they are extrapolated correctly to the next subsurface level, if the velocity model is accurate. Figure 9a and b shows zero-offset images produced by direct migration of the data shown in Figure 8a and b, respectively. The quality of Figure 9a is equivalent to a zero-offset reection migration. Dipping reectors and sharp changes in slope produce migration tails. Figure 9b is a remarkably clean image, given the appearance of the input data, although it is not as high quality as its counterpart. A faint virtual multiple that mimicks the rst event can be seen at z = 350 m. The physical multiple from the rst reector at z = 485 m is very dim which actually highlights the second reector that gets masked in Figure 9a . The summed waveelds shown in Figure 8a and used as the data for the image in Figure 9a also could be imaged with a source function modeled as a horizontal plane wave at t = 0.6 s. However, the summed data shown in Figure 8b and used for the image in Figure 9b require a complicated source function with temporal topography that would be impossible to extract from the data. The lengths of the source functions used to synthesize the data were equal to the modeling time of the transmission waveeld. Thus, summing the waveelds as dictated by equation 9 does not accommodate the requirement that the waveelds be at least twice as long as max t to avoid aliasing the symmetric acausal lags. The short time axis also allows late time events to wrap around the time axis when one is applying the random phase delays. For these reasons, the image in Figure 9b represents a worst-case scenario of careless data preparation.

sources that are separated areally will not interfere, such as when we are migrating the rst and last shots of a sail line, and those shots do not investigate the same subsurface volume. Second, an approximately contiguous distribution of sources will cancel the crosstalk by means of destructive interference and will synthesize a zero-offset data acquisition. The modeled data used herein have sources beneath the entire surface array and thus fulll the latter requirement. The information lost in this sum is the redundancy across the offset axis. For simple structural imaging, the zero-offset image is satisfactory. For more complicated mediums or further processing, a full plane-wave migration can be implemented Sun et al., 2001; Liu et al., 2002 .

FIELD EXPERIMENT
Crosscorrelating seismic traces of passively collected waveelds has a rich history in the study of the sun Duvall et al., 1993 . However, only two dedicated eld campaigns that test the practicality of passive seismic imaging on earth can be found in the current literature: Baskir and Weller 1975 and Cole 1995 . Neither experiment Figure 8. Summed transmission waveelds from 225 sources evenly produced convincing results. Hoping that hardware limitations or lodistributed across the bottom of the syncline model in Figure 2c. a cality could explain their lack of success, I conducted a shallow, Minimum traveltime of the direct arrival in each waveeld is the meter s -scale, passive-seismic experiment in the summer of 2002. same i.e., is analogous to Figure 6c . b Minimum traveltime of the direct arrival in each waveeld is random i.e., is analogous to Figure Seventy-two 40-Hz geophones were deployed on a 25-cm grid on 7c . the beach of Monterey Bay, California, and were linked to a Geometrics seismograph. The experiTable 2. The equivalence between direct migration of passive eld data and ment was combined with an active investigation 3 simultaneous migration of all shots in a reection survey. of the same site using the same recording equipment and a small hammer Bachrach and Mukerji, 2002 . A short length of 15-cm-diameter plasPlane-wave migration Wavefront imaging tic pipe was buried a bit less than 1 m below the   = Tz=0 xr, Tz=0 xr, surface. The array was approximately 100 m xsU z=0 x r ;x s, xsD z=0 x r ;x s, from the waters edge. The water table was approximately 3 m deep. The velocity of the sand, + + E E E E derived from the active survey, was a simple gra dient of 180 to 290 m/s from the surface to the +   = Dz=1 xr, Tz=1 xr, Tz=1 xr, Uz=1 xr, water table, below which it was 1500 m/s. 3Only the rst and second levels of the iterative process are depicted. Figure 10 shows the time-migrated activeproduces the source image, which has a clear anomaly associimage iz for both methods.

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Imaging passive seismic data

SI185

sidering the low velocity of the beach sand and the small areal extent ated with the hollow pipe and the water table. A simple rms gradient of the array. Figure 12 is the image produced at midday. Figure 12a velocity to the water table was used for imaging. The water-table reand b are the x and y sections corresponding to the center of the burection at 26 ms in the section carries a wavelet approximately ied pipe. 2.5 ms in duration, which gives it a dominant frequency of 400 Hz. Figure 13 was produced with data from around midnight. The imThe high quality of the beach sand resulted in recording useful enerage planes are the same as for the previous gure. One-dimensional gy as high as 1000 Hz Bachrach and Mukerji, 2002 . spectral whitening also was tried, although the simple application rePassive data were collected at several sampling rates over the mained stable only during night acquisition. course of 2 days, 2 weeks after the active survey. Because of limitaFigure 14 was produced with the whitened version of the data tions of the recording equipment, only 1 hour of data exists from the used for the previous gure. Notice the instability at shallow depths campaign. The seismograph was only able to buffer 1 s of data in before the wavefront healing has interpolated across the empty tracmemory before writing to a le. The time required to write, reset, and es. Data collected in the morning did not yield images sufciently retrigger was at least ve times greater than the length of data capdifferent from the night data to warrant inclusion. tured, depending on sampling rates. By continuously triggering the seismograph, the system wrote subsequent records as soon as the buffer/write cycle was complete. Then the individual records were spliced together along the time axis to produce long traces. The gaps in the traces do not invalidate the assumptions of the experiment, as long as the individual recordings were at least as long as the longest two-way traveltime to the deepest reector. Because the array was only eight-by-nine stations, shot gathers produced by correlation, even when resampled as a function of radial distance from the center trace, had too few traces to nd consistent events. Migrating the data provides both signal-to-noise enhancement, as described above, as well as interpolation. In this case, ve empty traces were inserted between the geophone locations for migration, as shown in Figure 11. During extrapolation, the energy on the live traces moves laterally across to ll the empty traces. After the distance propagated is approximately equal to the separation between live traces, wavefronts have coalesced into close approximations to their expression if the data were interpolated rst. Data were collected to correspond to distinct environmental conditions throughout the course of the experiment. Afternoon data were collected during high levels of cultural activity and wind action. Night data had neither of these features, whereas morning data had no appreciable wind noise. In all cases, the pounding of the surf remained mostly consistent. By processing data within various time windows, I hoped that images of the water table could be produced at different depths. However, given the 2-m maximum offset of the array, ray parameters less than 17 from the vertical would be required to image a 3-m table reector at the very center of the array. Very little energy was captured at such steep incidence angles. Had we not been careful to avoid walking around the array during recording, this might not have been the case. Figure 9. a A zero-offset image produced by direct migration of Figures 1214 show the images produced during the different Figure 8a. b A zero-offset image produced by direct migration of times of the day. Approximately 5 minutes of Figure 8b. 0.001 s/sample data were used to produce each image by direct migration. Usable energy past 450 Hz is contained in all the data collected. Abiding by the one-quarter-wavelength rule, and using 200 m/s with 400 Hz, the data should resolve targets to 12.5 cm. Other data volumes corresponding to various faster and slower sampling rates were processed, although these results are the most pleasing. All output images contain an appreciable anomaly at the location of the buried pipe. In most cases, preprocessing consisted of a low-cut lter that removed energy below 150 Hz Figure 10. An inline, x, and crossline, y, time-migrated active-seismic image. The hollow to eliminate electrical grid harmonics. Also, the pipe causes an overmigrated anomaly at x = 12 m, t = 12 ms. A strong water-table rehigher octaves carry the only useful signal conection is imaged at 28 ms. After Bachrach and Mukerji 2002 .

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

SI186

Artman surface sources, the length of assumed source functions, and migration. Rickett and Claerbout 1996 identied increasing the signalto-noise ratio by the familiar 1/ n factor, where n can be time samples in the source function, or the number of subsurface sources captured in the records. If the natural rate of seismicity within a eld area is constant, accumulation of sufcient signal dictates how long to record. When one is interpreting the increase in signal by the factor 1/ t, with application to short subsurface sources, t represents the mean length of the source functions rather than the total recording time. Assuming some rate of seismicity associated with each eld site, the total recording time will control the quality of the output by 1/ ns, where ns is the number of sources captured. Another method for increasing the quality of the experiment is to eld more instruments and thereby to migrate more traces. Migration facilitates the constructive summation of information reected from a subsurface reection point and captured by each receiver in the survey. Therefore, the use of more receivers sampling the ambient noise eld results in more-constructive summation to each image location in the migrated image. In this manner, migration increases the signal-to-noise ratio of a subsurface reection by the ratio 1/ r, where r is the number of receivers that contain the reection. This allows the production of very interpretable images, despite raw data or correlated gathers that show little promise. Denitive parameters for the numbers of geophones required, and sufcient length of time to assure high-quality results for a passive seismic experiment, are ongoing research topics because few eld experiments have yet been analyzed. However, it is clear that a very complete sampling of the surcial waveeld will be required and that the length of time required will be dictated on the activity of the local ambient noise eld. Considering the layout of equipment, very complete sampling means that more receivers are better and that areal arrays are much better than linear ones. The limitations of linear arrays can be seen by considering a plane wave propagating along an azimuth other than that of a linear array. After the direct arrival is captured, the subsequent reection path pierces the earths surface again in the crossline direction away from the array. This acquisition limitation requires a 2.5D approximation. The apparent-ray parameter of the arrival will sufce, given an areally consistent and planar source wave and a limited velocity contrast perpendicular to the receiver line. Because the true direct arrival associated with a reec-

In the whitened-night-data image and the bandpassed-day-data image, there is a hint of a reector at a depth that could be the water table. High tide on that day was at 4:30 in the afternoon, and low tide was at about 9:30 P. M., so the relative change of this hint of a reector is consistent. However, because of the limitations of the array, discussed above, and the lack of strength and continuity along the crossline direction, I do not consider this a very reliable interpretation.

DISCUSSION
Draganov et al. 2004 systematically explored the quality of a passive seismic processing effort as a function of the number of sub-

Figure 11. A small time window of inline and crossline sections of a raw passive-transmission waveeld inserted on a ve-times-ner grid for migration.

Figure 12. A migrated image from passive data collected during the windy afternoon. This inline and crossline depth section was extracted at the coordinates of the buried pipe.

Figure 13. A migrated image from passive data collected during the night. This inline and crossline depth section was extracted at the coordinates of the buried pipe.

Figure 14. A migrated image from passive data collected during the night. One-dimensional spectral whitening was applied before migration to the same raw data used in Figure 13. This inline and crossline depth section was extracted at the coordinates of the buried pipe.

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/

Imaging passive seismic data tion travel path is not recorded, erroneous phase delays and wavelets may occur, and they may distort the result.

SI187

CONCLUSIONS
To avoid crosstalk, previous literature discussing passive seismic imaging assumes that only individual source waveelds are correlated or that the source functions are random series. Direct migration, however, is not constrained by this assumption. In its place, sources need to be either roughly contiguous or widely distributed. Direct migration can be performed with only a minor adjustment to conventional reection-seismic migration algorithms that use extrapolation and a correlation-based imaging condition. Direct migration of all the sources simultaneously also saves substantial computational cost because only O n traces must be migrated rather than the O n2 traces associated with synthesized shot gathers. Migration moves the correlation into the subsurface to the reector position. Casting the processing of transmission waveelds as an imaging problem allows direct comparison of the process to phaseencoding and reverse-time migration strategies and introduces the possibility for more advanced imaging conditions such as deconvolution. Also, converted mode imaging strategies may provide a convenient framework within which to explore multicomponent data. Migrating all sources at the same time removes redundant information from a reector as a function of incidence angle. This makes velocity updating after migration impossible. At this early stage, I contend that passive surveys will be conducted only in actively studied regions where very good velocity models are already available. If this becomes a severe limitation, incorporation of plane-wave migration strategies can ll the offset dimension of the image. The eld experiments on the beach at Monterey Bay indicate that passive seismic experiments can work in practice. A signicant anomaly is present at the location of the hollow pipe target. However, the target was small and localized in all dimensions, which does not remove the possibility of an anomalous focusing of energy within the aperture of the array. The probability of such an occurrence at exactly the location of the pipe is minimal, however. Because of a multifaceted eld program, two weeks passed between burial of the pipe and the passive survey. Because the ends of the pipe were not sealed, sand may have lled the pipe and reduced its linear dimension. Future surveys should be executed to avoid this problem and should include burial of a target with an unnatural shape.

versity for successful collaboration and for his modeling efforts. Ran Bachrach provided the Michigan State shallow-seismic-acquisition equipment for the beach experiment as well as his image from the active seismic experiment. Emily Chetwin and Daniel Rosales helped collect data on the beach. Partial funding for this research was enjoyed from Petroleum Research Fund grant ACS PRF# 37141-AC 2, the Stanford Exploration Project, and National Science Foundation grant number 0106693 to S. L. Klemperer and J. Claerbout. Many thanks go to my patient reviewers, whose constructive comments greatly improved the quality of this manuscript.

REFERENCES
Artman, B., and J. C. Shragge, 2003, Passive seismic imaging: EOS, Transactions, American Geophysical Union, 84, AGU Fall Meeting, Abstract S11E0334. Artman, B., D. Draganov, C. P. A. Wapenaar, and B. Biondi, 2004, Direct migration of passive seismic data: 66th Conference and Exhibition, EAGE, Extended abstracts, P075. Bachrach, R., and T. Mukerji, 2002, The physics of seismic reections within unconsolidated sedi-ments: Technology for near surface 3D imaging using dense receiver array: EOS, Transactions, American Geophysical Union, 83, AGU Fall Meeting, Abstract T22B1141. Baskir, C. J., and C. E. Weller, 1975, Sourceless reection seismic exploration: Geophysics, 40, 158. Claerbout, J. F., 1968, Synthesis of a layered medium from its acoustic transmission response: Geophysics, 33, 264269. , 1971, Toward a unied theory of reector mapping: Geophysics, 36, 467481. , 1985, Imaging the earths interior. Blackwell Scientic Publications, Inc.. Cole, S. P., 1995, Passive seismic and drill-bit experiments using 2-D arrays: Ph.D. thesis, Stanford University. Derode, A., E. Larose, M. Campillo, and M. Fink, 2003, How to estimate the Greens function of a heterogeneous medium between two passive sensors? Application to acoustic waves: Applied Physics Letters, 83, 3054 3056. Draganov, D., C. P. A. Wapenaar, and J. Thorbecke, 2004, Passive seismic imaging in the presence of white noise sources: The Leading Edge, 23, 889892. Duvall, T. L., S. M. Jefferies, J. W. Harvey, and M. A. Pomerantz, 1993, Time-distance helioseismology: Nature, 362, 430432. Gazdag, J., and P. Sguazzero, 1984, Migration of seismic data by phase-shift plus interpolation: Geophysics, 49, 124131. Liu, F., R. Stolt, D. Hanson, and R. Day, 2002, Plane wave source composition: An accurate phase encoding scheme for prestack migration: 72nd Annual International Meeting, SEG, Expanded Abstracts, 11561159. Rickett, J., and J. F. Claerbout, 1996, Passive seismic imaging applied to synthetic data: Stanford Exploration Project, Annual Report, 92, 8390. Schuster, G. T., J. Yu, J. Sheng, and J. Rickett, 2004, Interferometric/daylight seismic imaging: Geophysics Journal International, 157, 838852. Shragge, J. C., B. Artman, and C. Wilson, 2006, Teleseismic shot-prole migration: Geophysics, this issue. Sun, P., S. Zhang, and J. Zhao, 2001, An improved plane wave prestack depth migration method: 71stAnnual International Meeting, SEG ExpandedAbstracts, 10051008. Wapenaar, C. P. A., J. Thorbecke, and D. Draganov, 2004, Relations between reection and transmission responses of three-dimensional inhomogeneous media: Geophysical Journal International, 156, 179194. Zhang, L., 1989, Reectivity estimation from passive seismic data: Stanford Exploration Project, Annual Report, 60, 8596.

ACKNOWLEDGMENTS
Thanks go to my colleagues and advisers at Stanford University for insightful discussions and development of the infrastructure to perform these experiments. I thank Deyan Draganov of Delft Uni-

Downloaded 28 Feb 2012 to 116.0.5.136. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/