Digital Time Series Analysis

concepts in time series

Last week, I showed you some of my holiday pictures…
…and we saw some controls on how faithfully we can digitise
an analog signal.

Today - a bit more of a mathematical description of how

frequencies are represented: Fourier Analysis

- and some Fourier Domain operations that are

commonplace in seismic processing.
What is the frequency of
a seismic wavelet?
 Any one wavelet has a range of frequencies.

bandwidth

frequency
Fourier Analysis –
recipes for building wavelets
amplitude spectrum

amplitude
- “how much of each
frequency component”

frequency

phase

frequency

phase spectrum
- “how those components align”
Superposition of
frequency components
Fourier analysis is a means of quantifying the
component frequencies of a wavelet and how they
are aligned.

After Yilmaz (2001)

Phase and amplitude relationships
amplitu The magnitude and
de phase of a sine wave
0 p 2p 3p 4p 5p
phase can be represented on
angle, q
a circle.
sine
wave
t u
p li
am
q
sin θ
θ de
0 p 2p 3p 4p 5p
cos
cosine θ
wave

The cosine wave is a shifted version

of the sine wave. A phase shift = a
time shift.
 The Fourier Transform says…
…any function (wavelet?) w(𝑡) can be expressed as contributions from
shifted and scaled sine and cosine frequency components:
angular frequency in radians
∞ ∞
𝑤 ( 𝑡 ) = 𝐴0 + ∑ 𝐴𝑛 cos𝑛 𝜔1 𝑡+ ∑ 𝐵𝑛 sin 𝑛𝜔1 𝑡
𝑛=1 𝑛=1
A background constant
amplitude scalars
EULER’S THEOREM
+∞ +∞
1
𝑤 (𝑡)=
2𝜋
∫ 𝑊 (𝜔) 𝑒 𝑖𝜔𝑡
𝑑 𝜔 𝑊 ( 𝜔 )= ∫ 𝑤(𝑡 ) 𝑒 −𝑖 𝜔 𝑡 𝑑𝑡
−∞ −∞

Converts from
angular frequency real part: imaginary part:
scalars phase shifts
Fourier Analysis –
recipes for building wavelets
amplitude spectrum

amplitude
- the real part of

frequency

phase

frequency

phase spectrum
- Imaginary part of
Phase and waveforms

Phase shift
constant with
frequency =
change in
waveform

Phase shift
linear with Here, it’s a 90°
frequency = phase shift, so the
bulk time- output wavelet is
shift 90° shifted.

After Sheriff & Geldart (1995)

 A greater linear shift to a phase spectrum
(top) results in a greater linear shift to a
constant waveform (bottom)
9 After Sheriff & Geldart (1995)
A cycle of constant phase shifts (top) gives a
cycle of waveforms (bottom)

After Sheriff & Geldart (1995)

After Yilmaz (2001)
 However…
…an analog vs. digital problem
+∞
𝑊 ( 𝜔 )= ∫ 𝑤(𝑡 ) 𝑒
−𝑖 𝜔 𝑡

amplitude
𝑑𝑡
−∞

frequency

phase

frequency

Because the time domain is discretely

sampled, Dt, the frequency domain is also
discretely sampled, Df
Frequency sampling and its effects

amplitude
1
Nt also known as…?
Nyquist frequency

Frequency (Hz)
0 1/2Dt
If the original time series has:
sample interval Dt and N sample points
the harmonic frequencies of the spectrum are at:

 1  2  3   N 2  1
0,  ,  ,   ,.........  Hz
 Nt   Nt   Nt   Nt  2t
Frequency Resolution

If you want to resolve two different frequency peaks, f1 and f2,

2
they must be more than Nt apart (i.e., two frequency steps).
2
Nt
amplitude

The difference between

f1 and f2 is now less that 2/NDt,
therefore the peaks are unresolved.

frequency
f1 f2
 Important Fourier Transform Identities
A time domain spike has a flat amplitude
spectrum, to f=∞ … i.e., ALL frequencies at
zero phase.
However, if the phase is instead random,
the time series
A constant a(t) is a(t)
amplitude also “DC”
random
has(“white
A(f)
noise”).
with just a spike at f=0Hz.
A ‘boxcar’ in the frequency domain i.e. a
band-limited version of (a) is a “sinc
function” (“{sin x}/x”) in the time domain:
like a spike, but broader and with side
lobes (important for future processing)
A wider boxcar (less bandwidth limited)
amplitude spectrum is nearer to a time-
domain spike
After Kearey & Brooks (2001)
NB All of these are “zero-phase” time series i.e.
Φ(f)=0°
The FFT, and “powers of 2”.
You’ll see Fourier Transforms often described as FFTs, or “Fast” Fourier
Transforms… It’s complicated – but is an analytic approach for speeding
up the computation of spectra.

To make things even faster, you’ll often see (e.g., in Matlab) FFT lengths,
N, equal to a power of 2 (e.g., 256, 512, 1024, etc). This is another
computational efficiency trick that I won’t go in to here…

(…because I don’t claim to understand it)

The Processing Power of Fourier Transforms
Beyond simply showing spectra, Fourier Transforms have very powerful
processing applications.

* =

Recall that convolution represented the action of a filter on a given

signal? We can ‘undo’ the effects of the filter using deconvolution.

Time domain deconvolution is seldom straightforward. However, there is

a neat convolution trick in the frequency domain…
 A Schematic Processing Sequence…
amplitude and
e series, a(t)  Fourier Transform, A(w) 
phase spectra
max +2π
A(f) Φ(f)

0 Some sort of
0 processing
time
0
frequenc
-2π
frequenc operation…
y y • frequency
filtering
A(f) Φ(f)
+2π
• deconvolution
0

-2π
time frequenc frequenc
y y
processed time Inverse Fourier processed
series, a(t) Transform spectra
 Applying multipliers in the frequency domain…
time series, a(t)  Fourier Transform, A(w)
max

0
time frequenc
y
DEconvolution
Convolution ininthe
thetime
timedomain
domain==
DIVISION in in
multiplication the frequency
the frequencydomain
domain

frequenc
y
Deconvolution operations are generally CONVOLUTION
frequency domain operations
 1-D Frequency Filtering
Having and understanding the amplitude spectrum of our data allows us
to design some useful ‘filtering’ operations. Signal and noise may
overlap in the time domain, but they are resolvable in the frequency
domain.
suppress suppress
“low- ed ed
pass” “high-pass” filter
filter

suppress suppress
ed ed
“bandpa “notch”
ss” filter filter
Variously Filtered Spikes…

Spikier wavelet form a broader amplitude spectr

- there’s a shorter-duration main peak
- the peak-to-sidelobe amplitude ratio is greater

after Sheriff & Geldart (1995)

However –
remember this?
Care is required when designing filters to avoid processing artefacts,
known as “Gibbs’ Phenomena”, or “Gibbs Effects”.

steep gentle
filter – filter
like a
boxcar
 Careful Filtering Required
Same bandwidth but
Filter slopes are defined in dB/octave: 12dB/octave (left) or
24dB/octave (right)
the higher the number, the steeper the slope

However: there are pro’s and con’s…

Baker (1999)

Simm & Bacon (2014)

Windowing Traces
Sometimes, we want to extract part of a trace for Fourier analysis

modified after Yilmaz (2001)

1-

…but this is a steep-sided filter!

0-

We multiply the trace by 0, except for the section we wish to extract,

which is multiplied by 1.
Spectral Leakage

time freq

frequency domain side lobes,

termed ‘spectral leakage’

1-

Better than a steep-sided

window is the use of ‘tapers’
0-

modified after Jenkins & Watts (1968

Tapers, tapers everywhere…
From… erm…
https://en.wikipedia.org/wiki/Window_function
 Hepburn, D (2016)

Leakage + Clipping
Clipped data are bad news for
spectra, because they introduce
leakage effects.
2-D Filtering
- time and space domains
You’ll recall from last week that there is an equivalence between
sampling in the time and spatial domains…

sample interval (s)  sample interval (m)

period (s)  wavelength (m)

frequency (Hz)  wavenumber (m-1)

Nyquist frequency (Hz)  Nyquist wavenumber, spatial Nyquist (m-1)

We can filter in the time and also spatial domains.

Fourier transforms in the offset direction
offset
(m)

…here we have a ‘spatial series’, which

can still be transformed into the Fourier
Domain

cycles per unit distance, or wavenumber,

k (m-1)

Familiar with the Fourier

two- transform in time…
way i.e. cycles per unit time
time (Hertz)
(ms)
Velocity recognition in a 2-D transform
fNYQ -
Different slopes in an
offset-time plot faster

frequency
…

(Hz)
express different
velocities.
speed (m/s) = distance (m)

time (s)

speed (m/s) = . frequency

(s-1) .

faster wavenumberslower
(m-1)
… Negative k
…
means?
slower |
wavenumber
|

… –kNYQ (m-1) +kNYQ

The ‘frequency-wavenumber’ (f-k) spectrum
offset (m) wavenumber (m-1)

Each of these trends is a

response to a velocity
trend in the data.

Polygons can be defined

that act like spatial-
temporal (and therefore
velocity) filters.

The same stuff applies

as in 1-D –
spatial/temporal filters
must be applied if
you’re going to avoid
ringing.
 Spatial Aliasing
Output
frequency
(normalised to the
Nyquist frequency)
1
fN )

0fN 1fN 2fN 3fN 4fN 5fN 6fN

Input frequency
…appears at ½ fN (normalised to the
Nyquist frequency
fN )
faithful,
Input frequency
unaliased
of 4½ fN…
 offset (m) wavenumber (m-1)

The signature of
spatial aliasing is a
wrap-around in the
k direction.

Back-Scattered
Noise

speed (m/s) = . frequency

(s-1) .

wavenumber (m-1)

What is the aliased event?

speed (m/s) = 160 Hz ÷ 0.5

-1
Spatial Aliasing
Dx
Spatial aliasing occurs when
our trace spacing is coarser
than the spatial Nyquist
wavenumber, kNYQ.

Steeper dips (i.e., slower

events) and higher frequencies
risk spatial aliasing.
Spatial Aliasing - causes
Just like how frequency aliasing happens because we have an
insufficient time sampling interval, spatial aliasing occurs because our
spatial sampling interval is too sparse.

It’s a major consideration in survey design – particularly in 3D. Aliased

traces are often recorded, but ‘anti-aliasing’ or ‘de-aliasing’ approaches
are used to mitigate the effects.
Claerbout (2009), Basic Earth Imaging
Stanford Exploration Project
http://sepww.stanford.edu/sep/prof/.
Fomel (2002), Application of Plane
Wave Destruction Filters
Geophysics, 67, 1946-1960.
Shows the typical
wrap-around in the
f-k domain

after Yilmaz (2001)

 Aliased
airwaves

Baker (1999)

Jones (2015)
 After all that…
- other transforms are available!

Stockwell Transform: varies the window length as a function of

frequency, said to give better time-localised spectral content.

Gabor Transform: windows always have a Gaussian form (useful

since the Fourier Transform of a Gaussian is another Gaussian),
allowing time and frequency resolution to be analytically linked.

Matching Pursuit: unlike the Fourier Transform, it is not restricted to

sines and cosines, but considers any arbitrary family of wavelets.
 (e)

(a) (d)

(b
)
(c)

New time-frequency transform (SDD) from speech processing literature. Sequence of Hilbert
Transforms, autocorrelations, and Fourier Transforms gives the “Instantaneous Autocorrelation
Function” (a) and “Ambiguity Plane” (b) that shows interferences and allows optimum time-
frequency transform (c) to be found by inverse-transform of the selected region.

Finding true amplitude spectrum requires integration over a time window, not a time-slice (d, a
Ricker wavelet, correctly represented). (e) Sample application to synthetic trace predicted for a
See Reine at al. (2009) Geophysics 74 WA123-WA135; Beckwith et al. (2016) Geophysical Prospecting, doi: 10.1111/1365-2478.12407
well-log, where much interference is present.
 That’s that!

Hopefully you’ll take with you an understanding of the ‘internal

workings’ of the frequency domain. It’s an important and widely-visited
part of seismic processing.

We’ll reinforce some of these lecture concepts in practical classes,

once you have some experience of SOEE5166. Revisit this material
ahead of class in Week 5?
 Some analytic
seismic wavelets

44
 Ricker wavelet
- a convenient proxy for a zero-phase seismic wavelet

• reasonable resemblance to real seismic

wavelets (though amplitude spectrum too
‘peaked’)

• zero-phase i.e. centred on its arrival time

• some useful statistical properties (it’s the 2nd

derivative of a Gaussian distribution)

• manageable analytic expressions for time

series a(t) and amplitude spectrum A(f) as
a function of (only) centre frequency fc

After Sheriff & Geldert (1995)

45
 Küpper Wavelet
• It can have as many “half-cycles” (“lobes”) m as you wish
• It has a manageable analytic expression for its time series a(t) for a given
m and duration T

From Smith, P.J. (2010) unpublished Ph.D. Thesis, University of Leeds.

Available at http://http://etheses.whiterose.ac.uk/1131/
46
More wavelets…
There are many others used in generating synthetic seismograms.

• Klauder wavelet – investigate this in background reading. Its

‘desirable property’ is a more realistic low-frequency content compared
to the widely-used Ricker Wavelet

• Berlage wavelet – A ‘continuous’ sine wave (for positive θ), whose

magnitude is modulated by a pair of exponential ‘envelopes’. Its
‘desirable property’ is that it’s NOT zero-phase, and it can be tuned to
represent many different real-world wavelets (e.g., radar).

• Ormsby wavelet, Butterworth Wavelet – ‘Unique selling points’ are

that they are spikes filtered with a simply-defined pass-band

47
 After Simm & Bacon
(2014)
48
And now that really is it!