Lecture 13:
Deconvolution, part 2
• Wiener filtering
• Deconvolution design
• Prewhitening
• Prediction distances
• Types of deconvolution
• Spiking deconvolution
• Predictive deconvolution
• Waveshaping deconvolution
The convolutional model:
x(t) is the recorded seismogram
w(t) is the source wavelet
r(t) is the earth’s impulse response (e.g., the reflectivity
series)
n(t) is random ambient noise
The goal of deconvolution:
To remove the affect of the source wavelet and of
reverberations and short period multiples in order to
isolate the earth’s reflectivity
Yilmaz, 2001
Deterministic deconvolution
• If the wavelet is known, we can design inverse filters
to remove the effect of the source and isolate the
reflectivity series of the earth
• Filters with more terms provide results that are
closer to the desired output
• Better results are achieved if the desired output
resembles the energy distribution of the input
• For example, if the desired output is a spike with
zero time lag, minimum phase input is required
to achieve good results
Do we know the
source wavelet?
• The farfield source signature of an
airgun array can be recorded with
a hydrophone (or modeled) and
used for deterministic
deconvolution
• However, we usually do not really
know w(t) (or what we do know
does not account for all of the
affects on our seismogram besides
the earths’ reflectivity series )
• Need to find a way of determining
a deconvolution filter that does
not require knowledge of the
source wavelet
Revisit example of least squares filtering
for minimum phase wavelet
Find the filter that has the minimum difference between
the squared difference of the desired output and the
actual output
Input wavelet: (1, -1/2)
Filter (a, b)
Desired output: (1,0, 0)
Sum of squared differences between desired and
actual output:
Revisit example of least squares filtering
for minimum phase wavelet
We seek to minimize L:
Find the minima:
L
slope=0
optimal
a
a
Least squares filtering for minimum
phase case expressed in matrix form
Re-arranging…
Cross-correlation of the
Auto-correlation of the
desired output with the
input wavelet
input wavelet
Least squares filtering for maximum
phase case expressed in matrix form
Re-arranging…
Cross-correlation of the
Auto-correlation of the
desired output with the
input wavelet
input wavelet
Earth’s reflectivity series: a white spectrum
The earth’s impulse response is assumed to be a white reflectivity
series and thus have a flat spectrum. This means that the
amplitude spectra of the seismogram is a scaled version of the
amplitude series of the source wavelet.
Autocorrelations and the convolutional model
Where rx, rw, and rr are the autocorrelations of the seismogram,
source wavelet and reflectivity series, respectively
• Where r0 is the autocorrelation of a random series, which is
zero everywhere but the zero lag. Here it is the cumulative
energy contained in the time series.
• Key point: The autocorrelation of the seismogram is an
approximation for the autocorrelation of the input wavelet
Cross-correlation of the
Auto-correlation of the
desired output with the
input wavelet
input wavelet
Approximate as the
Approximate as the
auto-correlation of the
cross-correlation of the
seismogram
desired output with the
seismogram
The Main Message:
• We can approximate the source wavelet with
the seismogram because the reflectivity series
of the earth is random
• As a result, we can design an inverse filter if
we know the seismogram and the desired
output!!
Yilmaz, 2001
Can also demonstrate by generalizing
least squares filter
Sum of squared differences between desired output (dt) and
actual output (yt)
where is the lag time
Autocorrelation of xt: rt
Cross-correlation of xt and dt: gt
The normal equations for Wiener filter
ri: autocorrelation of the input wavelet
ai: the desired filter
gi: crosscorrelation of the desired output with the input wavelet
Robinson & Treitel, 1980
This example demonstrates:
ri = r-i
r0 = x02+x12+x22+x32+x42
r1 = x0x1+x1x2+x2x3+x3x4
Wiener filter
Yilmaz, 2001
Assumptions of deconvolution
• The primary reflection series is random
• The source wavelet is minimum phase and is
doesn't vary though the earth (stationary).
• The noise is random and is of minimal level.
• The multiple period is fixed (stationary).
• The data are zero offset and dip is ignored.
Consideration in deconvolution design
• Pre-whitening
• Filter length (also called operator length)
• Noise
• Design windows
Pre-whitening
• The spectra of the spiking deconvolution
operator is approximately the inverse of the
amplitude spectra of the input data
• If there are zeros in the original data, these are
blown up by deconvolution, causing artifacts
• To avoid this, add ‘white noise’ to the spectra
of the input spectra to stabilize deconvolution
Pre-whitening
Amplitude Amplitude Result of
spectrum of spectrum of multiplying
input wavelet
inverse of the two
input wavelet
Yilmaz, 2001
Adding a constant to the zero lag of the autocorrelation is the
same as adding white noise to the spectrum
Other Effects of Prewhitening
• Pre-whitening narrows the spectrum, but does not decrease its
flatness
• Use a relatively small number: 0.1-1% prewhitening
Yilmaz, 2001
Filter length
Yilmaz, 2001
Filter length
Yilmaz, 2001
Effects of random noise
• The autocorrelation of random noise should
be zero except for zero lag, where it will be a
constant (e.g., akin to pre-whitening)
• In practice, it effects other lags as well
• The unavoidable presence of random noise in
seismic data means that only a very small
amount of pre-whitening is need
Without noise
Yilmaz, 2001
With random noise
Yilmaz, 2001
Design windows
To account for changes in the source wavelet with depth/time
due to attenuation, etc, it is common to use windows for
deconvolution, which allow you to determine different filters and
apply them to different parts of the data. Considerations for
design window:
• It needs to be much longer than the length of the filter (rule of
thumb: at least 10x the filter length)
• It should avoid particularly noisy areas, multiples, etc
• Ideally, merges between different windows should not occur in
particular areas of interest
Types of deconvolution
• Spiking deconvolution: turn source into
ideal frequency content – spike
• Predictive deconvolution: remove
multiples and reverberations by specifying
prediction distance
• Waveshaping: normalize wavelets from
different surveys, apply deconvolution to non-
minimum phase data
• Remove instrument effects
Spiking deconvolution
Purpose: sharpen the source
|G(f)|
Actual
Source
wavelet
|H(f)|
Filter
Ideal
output
Before
After
Yilmaz, 2001
After
Before
Bubble pulse
The normal equations for spiking
deconvolution
In the case where the desired output is a spike, g is a spike scaled
by the input wavelet
Designing spiking deconvolution
operators in practice
Minimum
phase or zero
phase
Length
Prewhitening
Gates for the
determination
Filter after of an inverse
deconvolution to filter.
remove artifacts
When spiking deconvolution does not work…
Yilmaz, 2001
Predictive deconvolution
• Used to remove ‘ringy’ parts of source or
multiples
• Seeking a time-advanced form of the input
series
– For input series x(t), we seek x(t+α) where α is
the prediction lag
A common application of
predictive deconvolution:
Multiple suppression
Main steps of predictive deconvolution
Yilmaz, 2001
The normal equations for predictive
deconvolution
In the case where the desired output is a time-advanced version
of the input. is the prediction lag.
Choosing a prediction distance or lag
• Measure off of seismic record
– Sometimes it is possible to simply determine the
prediction distance by examining the data
• Use autocorrelation
– Peaks in the autocorrelation function indicate time
delays where the two traces are most similar
[Link]
Before deconvolution
After deconvolution
Designing prediction deconvolution
operators in practice
Length
Prediction lag
Prewhitening
Gates for the
determination
Filter after
of an inverse
deconvolution to
filter.
remove artifacts
Waveshaping deconvolution: can be applied to
mixed phase or maximum phase wavelets
Input wavelet
Desired output
Shaping filter
Shaping filter
Yilmaz, 2001
