Deconvolution
DP START Oslo
2004
At the end of this Session, you will be able to:
• Define the term ‘reflectivity series’
• Define the term ‘convolution’
• Define the term ‘deconvolution’
• Define the term ‘seismic wavelet’
• State a mathematical equation that describes the convolutional model in terms of the
following components: recorded seismogram, basic seismic wavelet, earth impulse
response, random ambient noise.
• Define the term predictive deconvolution
• Define the term gap length
• Define the term total operator length
• Describe how gap length selection affects the success of deconvolution processing
• Describe how total operator length selection affects the success of deconvolution
processing
DP START Oslo
2004
At the end of this Session, you will be able to:
• Define the term 'pre-whitening' as used in deconvolution processing
• Describe how the use of pre-whitening can improve results of deconvolution
processing
• Define the relationship between pre-whitening level and seismic data bandwidth
• Define the term 'white noise' as used in deconvolution processing
• Define the term 'spiking deconvolution'
• Describe at least 3 ways in which application of pre-stack deconvolution can
improve the quality of seismic data.
• Describe at least 2 ways in which application of post-stack deconvolution can
improve the quality of seismic data.
• Define the term 'deterministic deconvolution'
• Define the term 'statistical deconvolution'
DP START Oslo
2004
At the end of this Session, you will be able to:
• Define the term 'multi-window deconvolution‘
• Explain how multi window deconvolution can be used to improve the quality of
data in comparison to single window deconvolution
• Define the term 'single channel deconvolution'
• Define the term 'autocorrelation'
• Define the term 'first zero crossing'
• Define the term 'second zero crossing'
• Explain how an autocorrelation display can be used to help determine the optimum
gap length for deconvolution processing
• Explain how an autocorrelation display can be used to help evaluate deconvolution
performance for a given dataset
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Introduction
• Deconvolution is a processing tool which has been used for
– Wavelet Shaping
– Multiple Removal
– Noise Attenuation
DP START Oslo
2004
Introduction
• Deconvolution is a processing tool which has been used for
– Wavelet Shaping
– Multiple Removal
– Noise Attenuation
• However
– Occasionally applied without thought as to whether it is actually
necessary
– More specialised wavelet processing or multiple attenuation
programs may be more appropriate
DP START Oslo
2004
Introduction
• Deconvolution is a processing tool which has been used for
– Wavelet Shaping
– Multiple Removal
– Noise Attenuation
• However
– Occasionally applied without thought as to whether it is actually
necessary
– More specialised wavelet processing or multiple attenuation
programs may be more appropriate
• Even so, decon is sometimes the most robust and straightforward way
of achieving your goals
DP START Oslo
2004
Types of Deconvolution
• Designature
• Predictive Deconvolution
• Surface Consistent Deconvolution
• Designature is explained in the DPSTART designature
presentation, this presentation we will discuss Predictive
Deconvolution.
DP START Oslo
2004
Ideal vs Reality
The recorded seismic trace may be modelled as a series of interactions
between the source signature (a finite, band-limited wavelet) and the earth.
Actual
seismogram
A plot of the reflection
coefficients of the layers
as a function of two way
travel time.
Sedimentary Reflectivity
Column Function
DP START Oslo
2004
Convolution
Convolution is the change of a wave shape as a result of passing it
through a linear filter
When a signal passes through any filter (such as the earth), it is replicated
many times with different amplitudes and time delays, by the filter.
Assuming that the signal, itself, does not alter with the passing of time (ie.
it is time shift invariant), then the filter produces a linear superposition of
these copies of the signal. The mathematical description of this process
is known as convolution.
To revise convolution read the material in the DPSTART lecture on Signal
Processing.
DP START Oslo
2004
Deconvolution
The objective of deconvolution:
in theory……….
• to extract the reflectivity component of the seismic trace by
removing the effects of the seismic wavelet, including ghosts
and short-period multiples.
in practice………
• to arrive at a better estimate of the reflectivity.
• output trace to represent reflectivity functions in terms of amplitude,
polarity and depth within limits of seismic signal properties.
Dcon can compress the basic wavelet in the recorded seismic trace, attenuate
reverberations, & thus increase the resolution.
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Wavelet Compression in Theory
In deconvolution, the aim is to resolve the reflectivity R(t)
X(t) = S(t) * E(t) * R(t) + N(t)
seismic trace (seismic wavelet) reflectivity random noise
source ‘system’
signature wavelet
known unknown unknown
Known? unknown
DP START Oslo
2004
Wavelet Compression in Theory
In the frequency domain, the equation can be written as:
X(f) = S(f) x E(f) x R(f) + N(f)
Where X(f), S(f), E(f) and R(f) represent the amplitude spectra of the
corresponding time functions (ignoring the phase for now)
Q. How can we remove the effect of the (S(f) x E(f)) term in this equation?
A. Make it equal to ‘1’ (or any constant value will do)!
Q. What function (in the time domain) will have a ‘constant’ amplitude spectrum
over all frequencies?
A. A SPIKE!
DP START Oslo
2004
Deconvolution Wavelet Compression in Theory
b0
1.0
(b0 ,b1 ,b2 - - - bN ) Deconvolution operator
(1 ,0 ,0 ,0 - -)
(f(f0 , ,ff1, ,ff2 , ,…..)
0 1 2 …..) 0 0 0 0
b1
etc.
b2 IN
t=0
Seismic wavelet Desired (spike) output
We design an inverse filter to deconvolve the data
In the time domain, deconvolution involves finding an inverse of the wavelet which,
when convolved with the seismic trace, outputs the reflectivity series.
The expected result is that the seismic wavelet is converted to a spike.
DP START Oslo
2004
Deconvolutional Model
RECORDED
SEISMIC TRACE
DECON FILTER
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Deconvolution Methods
Generally fall into one of two categories -
Deterministic deconvolution
Deconvolution; where part of the seismic system is known.
For example, where the source wavelet is accurately known
we can do source signature deconvolution.
Statistical deconvolution
Deconvolution where no information is available about any of the
components of the convolutional model. A statistical approach is needed
to derive information about a wavelet (either ‘source’, ‘system’ or
combined wavelets).
DP START Oslo
2004
Statistical Deconvolution
After ‘de-signature’ we are left with :
X(t) = E(t) * R(t)
seismic trace system wavelet reflectivity
known unknown unknown
We can now use Statistical deconvolution techniques to remove
the system wavelet, by compressing it to a spike, and hence derive the
reflectivity.
In practice though, the wavelet is not compressed to a spike but simply
Reduced in length. Remember, predictive deconvolution is also used
to reduce multiples & perform no compression.
Statistical deconvolution may be applied without the prior application
of deterministic deconvolution, as is usually the case with land data.
DP START Oslo
2004
Statistical Deconvolution
Statistical deconvolution is a process where we :
• Have no pre-knowledge of the wavelet
• Derive information about the wavelet from the data itself, specifically
from the autocorrelation of the data.
• Make certain assumptions about the data which justify the statistical
approach.
We assume the autocorrelation of the seismic trace is equivalent to
the autocorrelation of the system wavelet and use Wiener-Levinson algorithm to
derive the deconvolution filter.
DP START Oslo
2004
Auto-correlation – another look
Amplitude spectrum
reflectivity system wavelet seismic trace
X =
frequency
Note the autocorrelation of a seismic wavelet and trace are similar over early part of the trace
Auto-correlation
reflectivity system wavelet seismic trace
* =
time Similar
DP START Oslo
2004
General Assumptions
• Deconvolution is reliant on three major assumptions
– Random Reflectivity Series
• Autocorrelation is a spike
– Stationary (i.e.. time invariant) wavelet
• short design windows
• long design windows
– Minimum phase wavelet
• importance of this assumption depends on decon gap
DP START Oslo
2004
Random Time Series
If we have a perfectly random time series, like white noise as shown in the
figure, its auto-correlation will be a spike at zero lag, since random series
will correlate itself perfectly at zero lag and will be perfectly uncorrelated at
all other time lags.
Random series
Auto-correlation
Statistical deconvolution techniques are based on the assumption that
the earth’s reflectivity is also random and that its auto-correlation is
therefore a spike
DP START Oslo
2004
Random Time Series
If the random time series assumption is not met, the deconvolution will
try to remove reflections that it sees as periodic, and will also introduce noise
at other times.
No decon filter applied
Spike at centre of
autocorrelation.
Normally we only
display half the auto-
correlation because it
is symmetrical.
Therefore autocorrelation energy is assumed to be multiple energy
DP START Oslo
2004
Stationary Wavelet Assumption
•White reflectivity assumption
– Use long windows (10% rule)
– Data within the design
window must be well
balanced
• use geospread
• use expgain if necessary (be
careful with multiples)
• try to exclude areas of very
strong reflectivity
DP START Oslo
2004
Stationary Wavelet Assumption
•Stationary wavelet P
– Wavelet (and multiple period
& amplitude) should not
change within the design M1
window
• choose windows carefully M2
• use multiple windows
• consider using more M3
appropriate sfms
DP START Oslo
2004
Minimum Phase Assumption
• Minimum phase assumption
– Long-gap decons don’t modify wavelet
• wavelet phase is less important
– Multiples must be minimum phase
• don’t distort them with badly chosen amplitude balancing
INPUT GAP
DP START Oslo
2004
Statistical Deconvolution
Two most common examples -
Predictive Deconvolution- uses information from the earlier part of
the seismic trace to predict and deconvolve the latter part of the
trace. This method is also known as ‘Gapped Deconvolution’.
Predictive Deconvolution can be used to attenuate multiples, and less
commonly these days shape the wavelet.
Spiking Deconvolution- is a special case of predictive deconvolution
where the ‘gap’ is one sample interval
DP START Oslo
2004
Statistical Deconvolution - ‘Spiking’
X(t) = W(t) * R(t)
seismic trace seismic wavelet reflectivity
Could use E(t) in the above equation if deterministic deconvolution had been
applied
We know that the ideal wavelet would be a spike (1,0,0,0...) - this
would result in the seismic trace being a good representation of the
reflectivity function.
We try to achieve this by applying ‘Spiking Deconvolution’.
DP START Oslo
2004
Statistical Deconvolution - ‘Predictive’
A Seismic trace
Decomposes
to
B C
Reflection Series Common multiple
Auto-correlation of A train
Unpredictable is used to derive Predictable
White noise Information about Repetitive
series B and C systems
Predictive deconvolution ‘predicts’ the repetitive element within the
seismic trace (multiples, ringing etc) and generates an operator (of a specific
length) which will remove them, leaving only the random element i.e. the
reflection series.
Predictive deconvolution uses information from the earlier part of the seismic
trace to predict and deconvolve the latter part of the trace
VG016
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Key Components of Decon Operator
We can delay the point on our wavelet where our deconvolution
operator begins to operate - via the prediction ‘lag’ or ‘gap’
Zero-lag point
Amplitude
Time
2nd zero-crossing
Predictive gap is measured from zero-lag point on the autocorrelation plot
Prediction Lag or Active length
Gap(passive)
Total decon operator length
DP START Oslo
2004
Deconvolution in practice
Gap length
Will have effect on -
pulse stabilisation - to ‘equalise’ the basic wavelet throughout
the data
wavelet compression - degree of ‘spiking’
Occasionally, which multiple system is targeted - long gap length
with short active operator to straddle long period multiples
(Too long a gap may result in short period reverberations remaining)
DP START
VG026
Oslo
2004
Deconvolution in practice
Active operator length
•If the assumption that the auto-correlation of the seismic trace
is equal to the auto-correlation of the wavelet is correct, then
the length of the operator is not so critical
• However, if we make the operator too long, we may end up
distorting the very information we are trying to extract.
• Deconvolution operator length parameters are arrived at after
extensive testing.
• Whenever possible, a 10% rule should be obeyed, where the
total operator length is not more than 10% of the deconvolution
design window
DP START
VG026
Oslo
2004
Deconvolution in practice
Operator Design Window
This is the length of input seismic data (shot or CMP gathered) that
will be auto-correlated In order to determine the length and the
shape of the deconvolution Operator.
Must be long enough to ‘predict’ the multiples targeted!
Design window (or gate) usually at least 10 times the operator length.
First break noise is excluded from the design window
One or two derivation windows at the most (multi-window decon)
DP START Oslo
2004
Spiking or Gapped Deconvolution?
• Spiking • Gapped
•Maximum resolution - •Control of resolution via gap size -
•better estimation of reflectivity? •can leave the wavelet unchanged
•Loss of apparent continuity compared to •Can be used solely for multiple
gapped deconvolution? •suppression.
•Possibility of boosting noise? •Less possibility of boosting noise
•To be fully effective, wavelet must be •Minimum phase assumption not so
minimum phase
•critical
VG025
DP START Oslo
2004
Deconvolution in practice
Pre-whitening
Addition of white noise to data (autocorrelogram) during
operator design. It is achieved by adding a constant to the
zero-lag of the Autocorrelation function to prevent:
• Numerical instability (divisions by zero when calculating wavelet inverse) during the
design of an operator.
DP START Oslo
2004
Pre-whitening
amp Input amp Filter
frequency Nq frequency Nq
amp Output
noise
frequency Nq
DP START Oslo
2004
Pre-whitening
The amount of white noise to add will generally be in the range
0.1 % - 1.0 %
• Too little white noise may - • Too much white noise may -
cause the deconvolution decrease the effectiveness of the
operator to become unstable deconvolution process
decrease S/N ratio of the narrow the bandwidth of the data
data
DP START Oslo
2004
Deconvolution in practice
Effects of random Noise
• Random noise in the data itself can have a similar effect to pre-
whitening.
• Increasing random noise decreases the effect of deconvolution
• Can result in spurious ‘spikes’ in the data and amplitude
spectrum
• Presence of random noise should mean minimal pre-whitening
DP START Oslo
2004
Where Pre-Stack ?
Spherical Divergence correction
Exponential Gain
F-K filtering (if applicable)
to remove coherent noise.
Deconvolution
demultiple data prior to velocity analysis
Velocity analysis
DP START Oslo
2004
Where Post-Stack ?
Stack
Deconvolution
Pulse compression and residual
demultiple prior to migration
F-K filtering
This can be done before or after
deconvolution - no set “rules”
Migration
DP START Oslo
2004
Deconvolution in practice
Generally
If “tight” compression or spiking is to be done, it is done after stack
when the signal to noise ratio has been improved.
Otherwise, the risk with the spiking approach is that noise levels
are boosted as well as the signal.
Stacking enhances residual multiples left by pre-stack
deconvolution, hence we normally need to do some form of
demultiple after stack as well as before stack
DP START Oslo
2004
Deconvolution in practice
Multi-window deconvolution
• Sometimes the time-variant character of the seismic
wavelet tends to violate the assumption that the wavelet is
stationary in time and depth.
• In such cases, multi-window deconvolution is advantageous.
• For best results (where necessary) we must design different
deconvoultion operators from different parts of the record
and apply them to corresponding time gates (windows).
DP START Oslo
2004
Deconvolution in practice
Multi-channel deconvolution
Sometimes advantageous over single-channel deconvolution
More ‘randomness’ in the data due to differential moveout
Noise changes from trace to trace - more randomness
Noise cancellation within ‘noisy’ data
However, disadvantages such as smearing of the extracted
wavelet can occur
DP START Oslo
2004
Autocorrelations & Multiples
Autocorrelation is the cross correlation of a trace with itself
Amplitude
Multiple period
Time
Primary energy Multiple energy Multiple energy
The autocorrelation function of a seismic signal is an effective tool for
determining the presence and time lag (duration) of reverberations/multiples.
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Deconvolution Operator Example
• Let us first detail what the key components of a typical
deconvolution operator are.
Prediction distance or ‘gap’
Active operator length
Total operator length
• The following slide shows that the total operator equates to:
Total Operator = Active Operator + Prediction Distance
DP START Oslo
2004
Time Series Containing 140 msec Period Multiples
140 msec
140 msec Period Multiples
Primary Signal
DP START Oslo
2004
1-Sided Autocorrelation
140 msec
DP START Oslo
2004
124/300 msec Deconvolution Operator
Gap = 124 msec
Active Operator = 176 msec
Total Operator = 300 msec
DP START Oslo
2004
Time Series After 124/300 msec Deconvolution
Primary signal
Remnant multiple
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Predictive Decon - “Normal Flow”
DP START Oslo
2004
Predictive Decon - “Normal Flow”
TRACE BY TRACE RAW ATTRIBUTE
DP START Oslo
2004
Predictive Decon - “Normal Flow”
DESIGN OPERATORS FROM
ATTRIBUTES
DP START Oslo
2004
Predictive Decon - “Normal Flow”
MATCH SEISMIC TRACE TO
OPERATOR TRACE THEN APPLY
DP START Oslo
2004
PRD_DCN_SPCTRL_ANL
The autocorrelation half length specified here,
will by default be equivalent to the ‘Total
Operator Length’
DP START Oslo
2004
PRD_DCN_OPR_DESIGN
‘Prediction Distance’
or ‘Gap’
If the user so desires, the ‘Total Operator
Length’ may be changed here from the
default value which is equal to the
autocorrelation half length
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Deconvolution Goals
• Decon can do several things
– Compress the wavelet to a sharper, less reverberatory
pulse
– Remove multiples
– Remove noise (or at least damp down its effects)
• Different objectives may require different, sometimes
contradictory, parameterisation
• When testing decon, you need to know exactly what you
want to achieve
DP START Oslo
2004
Pre-Stack Wavelet Compression
• Pre-stack decon for wavelet compression
– Apply geospread, check amplitude balance
– Mute 1st breaks, or set START_TIME appropriately
– Apply deterministic minimum phase conversion
– Run autocorrelations, check for bubble pulse
– Choose design window(s) - try to use single window
– Choose length and gap(s) to test
• length > bubble period (e.g.. 150 - 200 ms)
• long gap (e.g.. 64 ms) if no whitening required, otherwise test various short
gaps (e.g.. 4, 16, 24, 32, 48)
– Test decon(s), test shot-averaging of “best” decon
DP START Oslo
2004
General deconvolution parameterisation
• Pre-stack decon for wavelet compression
– Source bubble is consistent over all offsets, multiples are
not
– Decon operator needs to be longer than 1st reverberation
Source
bubble
Autocorrelated shot gather
DP START Oslo
2004
Summary
• Omega deconvolution
– Very flexible
• General deconvolution parameterisation
– Decon is a general purpose tool - more specialised programs will
often be more appropriate
– Parameterisation is complicated
– Have a clear idea of what you want to achieve with decon
– Careful testing pays dividends
– Remember that deconvolution ‘Design windows’, ‘Prediction
distances’, ‘Active and total operators’ should all be integer
multiples of the data sampling interval
• Always look at autocorrelations pre and post decon
DP START Oslo
2004
Agenda
• Introduction
• Deconvolution’s wavelet compression theory
• Statistical Deconvolution
• Deconvolution in Practice
• Multiple Removal – Synthetic example.
• Deconvolution SFM in Omega
• Conclusions
• Tau-P Demultiple
DP START Oslo
2004
Tau-P Demultiple
• It is becoming more & more common to perform
Deconvolution in the Tau-P domain.
• Tau-P Dcon is essentially a multiple attenuation process
• Prestack input data is transformed from the T-X domain to
the Tau-P domain. While in the Tau-P domain predictive
trace by trace deconvolution is applied in order to attenuate
predictable multiple energy.
DP START Oslo
2004
Why Dcon in the linear Tau-P domain?
• In T-X domain water bottom multiples period varies in both
time and offset.
• Deconvolution expects that the multiple period is constant
within the design window. So these multiples are not
effectively attenuated.
DP START Oslo
2004
Why Dcon in the linear Tau-P domain?
• Transforming the data into the Tau-P domain effectively
splits the data into traces, which represent given time dips.
These traces contain multiple energy of a given period.
As the multiple energy is now periodic deconvolution will
perform more effectively.
DP START Oslo
2004
Why Dcon in the linear Tau-P domain?
DP START Oslo
2004