14 Seismic Data Processing

By using a single-frequency sinusoid, we see that fre- ed, because 8-ms sampling still is sutficient to sample
this
quencies above the Nyquist really are not lost after low-frequency component. On the other hand, the 75-Hz
sampling, but reappear at frequencies below the Nyquist. component is seen as a lower-frequency component
(50
Now consider the superposition of two sinusoids with Hz). Once again, note that those frequencies in the
12.5- and 75-Hz frequencies as shown in Figure 1-10. original signal above the Nyquist frequency correspond-
Digitization of this signal at 2- and 4-ms sampling inter- ing to the chosen sampling interval are folded back in
the
vals does not alter the original signal, since its frequency amplitude spectrum of the digitized version of the
signal.
components are below the Nyquist frequencies associat- This analysis can be extended to many sinusoids of
ed with 2- and 4-ms sampling intervals (250 and 125 Hz, different frequencies. In particular, the discrete time
respectively). However, when the signal is digitized at a series derived from too coarse sampling
(undersampling)
coarser sampling interval, such as 8 ms, the amplitude of a continuous signal actually contains contributions
spectrum changes. The 12.5-Hz component is not afiect- from high-frequency components of that continuous
signal. Those high frequencies fold back onto the spectrum 1.2.2 Phase Considerations
of the discrete time series and appear as lower frequen-
cies. The phenomenon that is caused by undersampling
the continuous signal is termed aliasing. In Section 1.2, a time-dependent signal was synthesized
To compute the alias frequencyfi,, use the relationf}, = from its frequency components. Consider a signal with a
I2mfN - 1;, where fN is the folding frequency. f, is the zero-phase spectrum. Figure 1-11 shows sinusoids with
signal frequency, and m is an integer such that j;, < fN. frequencies ranging from approximately 1 to 32 1-12.
All of
For example, suppose that f, = 65 Hz, fN = 62.5 Hz these sinusoids have zero-phase lag; thus, the peak
which corresponds to 8 ms sampling rate. The alias amplitudes align at t = 0. The time-domain signal on the
frequency then is fa = I2 x 62.5 - 65] =-= 60 Hz. trace identified by an asterisk in Figure 1-11 is derived by
In conclusion, undersampling has two effects: (a) band summing all these sinusoids. This summation is an in-
limiting the spectrum of the continuous signal, with the verse Fourier transform. Such a time-domain signal is
maximum frequency being the Nyquist, and (b) contami- called a wavelet. A wavelet usually is considered a
nation of the digital signal spectrum by high frequencies transient signal, that is, a signal with a finite duration.
It
beyond the Nyquist, which may have been present in the has a start time and an end time, and its energy is
continuous signal. Nothing can be done about the first confined between these two time positions. The wavelet
problem. The second problem is of practical importance. that was just constructed is symmetric around I = 0
and
To keep the recoverable frequency band between zero has a(positive) peak amplitude at t = 0. Such a wavelet
is
and the Nyquist frequency free from aliased frequencies, called zero phase. In fact, the wavelet was
synthesized
a high-cut antialiasing filter is applied in the field before using the zero-phase sinusoids of equal peak
amplitude.
analog-to-digital (A/D) conversion of seismic signals. A zero-phase wavelet is symmetric with respect to zero
This filter eliminates those frequency components that time and peaks at zero time. Figure 1-12 shows the
results
would have been aliased during sampling. Typically, the of
applying
a linear phase shift to the sinusoids in Figure
high-cut antialiasing filter has a cutoff frequency that is 1-11. Linear phase shift is defined as phase = can-
either three-quarters or half of the Nyquist frequency. stant - frequency. The wavelet (identified by an asterisk)
This filter rolls oif steeply so that frequencies above the has shifted in time by -0.2 s, but its shape has not
Nyquist are highly attenuated. changed. Thus, a linear phase shift is equivalent to a constant time shift. The
slope of the line describing
phase spectrum is proportional to the time shift.
The wavelet can be shifted by any amount oftime simply
by changing the slope ofthe line that describes the phase
spectrum (Figure 1-13). Starting with the zero-phase
wavelet, Figure l-13 shows the effect of increasing
amounts of linear phase shift on a zero-phase wavelet.
Although not shown, by changing the sign of the slope in
the phase spectrum, the wavelet can be shifted in the
opposite time direction.

If a 90-degree phase shift is applied to each of the
sinusoids in Figure 1-11, as shown in Figure 1-14, then
the zero crossings are aligned at t = 0. The result of this
summation yields the wavelet shown on the trace identi-
fied by an asterisk. A_n antisymmetric wavelet is pro-
duced. Note that the two wavelets in Figures 1-1] and I-
14 have the same amplitude spectrum because they have
the same frequency content. The difference lies in their
phase spectra. The wavelet in Figure 1-11 has zero-phase
spectrum, while that in Figure 1-14 has a constant-phase
spectrum (+90 degrees). Therefore, the difference in
wavelet shape is due to the diflerence in their phase
spectra.

Figure 1-15 shows the effect of various amounts of
constant phase shift on a zero-phase wavelet. The 90-
degree phase shift converts the zero-phase wavelet to an
antisymmetric wavelet. The 180-degree phase shift
changes the polarity of the zero-phase wavelet. The 270-
degree phase shift changes the polarity of the zero-phase
wavelet, while converting it to an antisymmetric wavelet.
Finally, the 360-degree phase shift retains the shape of
the original wavelet. A constant phase shift changes the
shape ofa wavelet. In particular, a 90-degree phase shift
converts a symmetric wavelet to an antisymmetric wave-
let, while a I80-degrcc phase Slllfl changes its polarity.

So far, two basic phase spectra have been examined;
namely, linear and constant phase shifts. The combinedeffect of the two now is considered. The phase
spectrum
is a function defined as a + I) - frequency, where a is the
constant phase shift and b is the slope of the linear phase
shift. Figure 1-16 shows the results of
applying
a 90-
degree constant phase plus a linear phase component to
the sinusoids in Figure 1-1 1. The zero-phase wavelet with
the same amplitude spectrum as that in Figure 1-11 was
shifted in time by -0.2 5 because of the linear phase shift,
and converted to an antisymmetric form because of the
constant 90-degree phase shift.
Other variations in phase spectrum are shown in Figure
0 1-17. The zero-phase wavelet (Figure I-17a) can be
modified to difierent shapes simply by changing the phase
spectrum. It can be modified so much that it may not
resemble the original wavelet shape as illustrated by the
last example (Figure l-17d). By keeping the amplitude
spectrum unchanged, the wavelet shape can be changed
by modifying the phase spectrum.
1.2.3 Time-Domain Operations
Consider a reflectivity sequence represented by the time
series (1, 0, ). Also consider an impulsive source that
causes an explosion at t = 0 with an amplitude of 1. The
response of the reflectivity sequence to an impulse is
called the impulse response. This physical process can be
described as follows:
Time of Reflectivity
Onset Sequence Source Response
0 1 0 1/2 1 I 0 1/2
One unit time later, suppose that the impulsive source
generates an implosion with an amplitude of 1/2. This process is described as: Wavelet l: (2, I, - 1, 0, 0)
Time of Reflectivity Wavelet 2: (0. 0, 2, 1, -1)

Onset Sequence Source Response

Note thatthe response is the reflectivity sequence scaled mined. To do this, perform the operation on wavelet I
as
by impulse strength. Since a general source function is described in Table 1-3 without reversing wavelet 2
(omit
considered to be a sequence of explosive and implosive Step 0). This is crosscorrelation and the result is
shown in
impulses, the individual impulse responses are added to Table 1-5. Crosscorrelation measures how much two
time
obtain the combined response. This process is called series resemble each other. Crosscorrelation of a time
linear superposition and is described in Table 1-1. series with itself is known as autocorrelation.

In Table 1-1, the asterisk denotes convolution. The FromTable1-5,notethatmaximumcorrelationoccurs at
response of the reflectivity sequence (1, 0, ) to the lag -~2. This suggests that if wavelet 2 were shifted two
source wavelet (1, -) was obtained by convolving the samples back in time, then these two wavelets would
two series. This is done computationally as shown in have maximum similarity.

Table 1-2.
Afixed arrayis set up from the reflectivity sequence.The when the arrays are interchanged. This time the maxi-
SOUFCC W3Vlt iS FVFSd (fo1dd) and moVd (lagged) mum correlation occurs at lag 2. Thus, if
wavelet l were
one sample atatime. At each lag, the elements that align shifted two samples forward in time, then these two
are multiplied and the resulting products are summed. wavelets would have maximum similarity. Also note that,
The mechanics of convolution are described in Table l-3. unlike convolution, crosscorrelation is not
commutative:
The number ofelements ofoutput array c-(k) is given by m i.e., the output depends on which array is fixed and
2 - l, where m and n are the lengths of the operand which is moving (compare the results listed in Tables 1-5
array an) and the operator array /9(j), respectively. and I-6).

When the roles of the arrays in Table 1-2 are inter- Table 1-7 shows the autocorrelation lags of wavelet I.
changed, the output array in Table 1-4 results. Note that Note that maximum correlation occurs at zero lag, an
the output response is identical to that in Table l-2. important property of autocorrelation. Moreover, the
Hence, it does not matter which array is fixed and which autocorrelation function is symmetric. This is a
property
is moved as the convolution is performed. of real time series. Therefore. only one side of the

Seismic processing often requires measurement of the autocorrelation needs to be computed.
similarity or time alignment of two traces. Correlation is Itis heuristically shown in Section l.2.4that convolution
another time-domain operation that is used to make such in the time domain is equivalent to multiplication in
the
measurements. Consider the following two wavelets: frequency domain (Bracewell, 1965). Since correlation is

-f- . Although these wavelets are identical in shape, wavelet 2
4 IS shifted by two samples with respect to wavelet I. The
time lag at which they are most similar can be deter-

Table 1-6 shows the crosscorrelation values that result