Introduction to

Seismology
Diego Quiros, Ph.D.
Lecturer in Geophysics
September 2022
Geophysics Module
Basics of Signal Processing
Topics Spectrogram
1. Signal Processing

frequency
1. Introduction

1.2 Fourier series time

3. Fourier transform

4. Time domain vs frequency domain

5. Spectrograms

6. Sampling theorem

7. Frequency filtering

8. Cross-correlation and convolution

9. 2-D Fourier transform and spatial aliasing

Introduction

In seismology we often use a class of techniques known as t i m e series analysis or

signal processing. Theses techniques consider functions of time or space, also called a series in
general terms. Subjects such as seismology, radar, sonar, and optics make heavy use of these tech-
niques.

For example in seismology we can treat either a continuous (analog) record of ground motion or the
digital data that results from representing the ground motion as being sampled at discrete inter-
vals, providing numbers that can be manipulated using a computer.

In general terms, we can think of f i l t e r i ng a signal, or applying some opperation that modifies the
signal. For example, a seismometer is a filter, in that it yields a record of ground motion that differs
from the actual ground motion.

In this lecture we will extend these ideas by considering mathematical approaches that are common
to such applications and see how these approaches give extraordinary insights into physical
processes.
Fourier Series
In seismology we use an approach based on the idea that any t i m e series can be decomposed into
a sum of sines and cosines (sinusoids). Each sinusoid having a particular amplitude, frequency, and
phase, this is known as Fourier analysis. Fourier Terms
0
Let's first consider the decomposition o f a signal with
finite duration into a Fourier series, or explicitly a sum of
harmonic (sinusoidal) components with different
frequencies. The Fourier series for an a r b i t r a r y function 2

Term number
(e.g., seismogram of an earthquake) of time f(t) defined
over the interval -T/2 < t < T/2 is
∞ 4
(
∞
( Σb
f(t) = a0 + Σan cos 2n π t
(
+ n
sin 2n π t
(
n T n T

This series decomposes f ( t ) into a sum of terms (1, ... , n) 6

that are sine and cosine functions with different periods
(or frequencies) because the arguments (2n π t /T) are pe-
riodic with period T/n, or frequency n/T. 8

Larger values of n correspond to shorter periods. For n = 0

the cosine term equals 1 for all values of t, and there is no
sine term because it would be zero. 10

We can decompose any time series in this way, and study -T/2 T/2
each sinusoid component. Time
Figure. Successive terms of a Fourier
series. Solid lines are sine function and
dashed lines are cosines.
Fourier Series Example
Let look at an example using the concepts from the previous page. Imagine different sinusoids each
with a specific frequency, peak amplitude, and phase, these are shown in the figure. The motions
represented by these different Fourier terms can be sum in the time domain.

The result is a time-dependent signal that

is represent by the bottom trace.
80
Note that this looks very much like a simple
seismogram of an earthquake.
64
The sum o r synthesis allows us to

Frequency (Hz)
transform the motion from frequency to
time. This t r a n s f o r m a t i o n is reversible,
since the time-domain signal (trace at the 48
bottom) can be broken d o w n (analyzed)
into its sinusoidal components in the
frequency domain. 32

Mathematically this two-way process is

achieved by the Fourier transform. 16
In practice the standard algorithm
used on digital computers is the -0.2 00 .2
the sum of all waveforms 0.4
f a s t Fourier t r a n s f o r m (FFT). Time (s)
Figure. Several sinusoids (Fourier terms) with different
frequency, peak amplitude, and phase can be sum to syn-
thesize a time-dependent waveform (shown by arrow).
The Fourier Transform

Remember the complex exponential is related to trigonometric functions by Euler's formula

eix = cos x + i sin x , meaning we can rewrite f(t) using eix instead of sinusoids for simplicity
∞
f(t) = Σ Fn eiωnt
n = -∞

Where ω n = 2 n π / T, which are angular frequencies.

After some manipulation the Fourier series becomes an integral

Inverse Fourier Transform

Fourier Transform
∞

∫
(going from the time ∞

∫ f(t) = 1
domain to the freq. domain)
F(ω) e iωt dω
F(ω) = f(t) e-iωt dt 2π
-∞ -∞
An important feature of the transform and inverse transform is that their dimensions are different.
For example if f(t) is a seismogram that has dimensions of displacement, its transform F(ω) has di-
mensions of displacement multiplied by time (from the dt term). Thus if f(t) gives the ground motion
in centimeters, F(ω) gives the transform of ground motion in centimeter-seconds.

The Fourier transform, a complex-valued function, can be written in terms of two real-valued func-
tions
F(ω) = ǀF(ω)ǀ eiφ(ω)

where ǀF(ω)ǀ is called the a m p l i t u d e spectrum and φ(ω) is called the phase spectrum.
The Amplitude Spectrum
In many applications only the amplitude spectrum is shown because it indicates how the energy
(amplitude squared) in the time series (e.g., seismogram) depends on frequency.

The figure below shows a seismogram for an earthquake and below two panels that include only
body waves, and surface waves.
Vanuatu earthquake (Ms 6.5)

(p-wave coda)
Amplitude

0 200 400 600 800 1000 1200 fast fourier

Time (s) transform is
basically a filter,
And after calculating the Fourier transform of the 2 panels separately we get can filter out
Amplitude Spectrum specific
1.0 waveforms.
Normalized Amplitude

0.5

Surface waves Body waves

0.0
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14

Frequency (Hz)
surface waves have lower frequencies.
Time domain vs Frequency domain

Looking at the seismogram in the previous page we see that the surface waves contain
longer-period energy than the body waves (i.e., body waves have higher frequencies). The spectra
actually shows this, body waves are dominated by energy with frequencies between 0.08 - 0.1 Hz
whereas the surface wave is dominated by energy with frequencies between 0.05 - 0.07 Hz.

The Fourier transform F(ω) is another way of representing the time series f(t). We speak of f(t) as
being in the t i m e domain and F(ω) as being in the f r e que nc y domain. The two representations are
equivalent because we can easily convert data from one domain to the other without losing any
information.

The transform and its inverse relate a function of time to a function of angular frequency. Similar
relations apply between other pairs of variables, in seismology for example the other commonly used
pair is distance and spatial f r e que nc y (wavenumber).

(lamda) equals 1/k(spatial freq./wavenumber) --> could also write velocity in terms of freg. and spatial freq. -> V equals f/k

previously used T(period) equals 1/f

 Seismogram
Spectrograms

Another way of visualizing data via a Fourier transform is the spectrogram.

A spectrogram is a way of plotting the frequency content of a seismo-

gram as it changes with time. For example with a spectrogram we can see
if there is more energy or less energy at 20 Hz vs 40 Hz, and also see how
spectogram of a train passing by:
energy levels vary over time.

Spectrograms are 2-D graphs, with a t hi r d dimension represented by

colors. Time is usually plotted from left to right on the horizontal axis.
The vertical axis represents frequency increasing from bottom to top. The
amplitude (or energy) of a particular frequency at a particular time is
represented by the third dimension, color.

Creating a spectrogram is done via the FFT. The data in the time domain is
broken into chunks, which usually overlap, the chunks are Fourier
transformed to calculate the Amplitude Spectra for each. Each chunk then
corresponds to a vertical line in the image, these spectra are then laid side
by side to create the spectrogram graph.

The figure shows an example. The seismogram on top was divided into
very small time chunks and Fourier transformed to get the spectrogram.
At the bottom the amplitude spectra of the entire seismogram (full 500 s)
is shown for comparison. Figure. Seismogram,
spectrogram, and
to make a spectogram -> first break the ampl vs time seismogram into distinct time intervals and create a amplitude amplitude spectra.
sepctra diagram of all time intervals (each time interval will have a amplitude and freq signal to be used in each time interval). Then you put all of these together
and flip the axis to get freq vs. time to get the spectrogram (you represent the amplitudes using colours -> high amplitudes as yellow, low ampl.s as blue)
Analog vs. Digital
A seismic signal is a continuous time function. In digital recording, the continuous (analog) seismic
signal is sampled at a fixed rate in time, called the sampling i nt e r v a l (or sampling rate). The figure
below shows a continuous signal in time. The discrete samples that might actually be recorded are
shown by dots. The bottom curve in the figure is an a t t e m p t e d reconstruction of the original ana-
log signal, which is shown as the curve on top.

Figure. (top) Analog continuous

signal. (middle) Digitized for of the
signal sampled at Δt. (bottom) Recon-
structed signal

Note that the reconstructed signal lacks the details present in the original analog signal. These
details correspond to high-frequency components that were lost by sampling. If a smaller sampling
interval were chosen, then the reconstructed signal would more accurately represent the original
signal.
Nyquist Frequency
Is there a measure of the highest restorable frequency of the digitized data?

In general, given the sampling interval Δt,

the highest f r e que ncy that can be restored
accurately is called the Nyquist frequency
and is given by

f Nyq = 1
2Δt

We can calculate the Nyquist frequency for

each series in the figure. For Δt = 2 ms the
Nyquist is 250 Hz, for Δt = 4 ms and Δt = 8 you lose all the
freq.s above
ms, it results in 125, and 62.5 Hz respective- 62.5Hz due to the
ly. Nyquist Freq.

After the analog signal is sampled can fre-

Figure. (left) Time series sampled at different sampling
quencies above the Nyquist be recovered?
rates. (right) Correspinding amplitude spectrum.
No. Once a continuous signal is digitized, the highest frequency that can be restored accurately is
the Nyquist frequency.

Note Interpolation does not recover the frequencies lost by sampling; it only generates extra
samples.
Frequency Aliasing
What happens when an analog signal that is digitized with sampling rate Δt, originally had frequen-
cies above the Nyquist for that particular sampling rate?

The figure shows the reconstructed signal of a 7 5 H z

sinusoid. Previously, we calculated the Nyquist
frequency for the Δt's shown (250, 125, 62.5 Hz).
Sampling the original 75 Hz signal at Δt = 2 ms and
Δt = 4 ms, reconstructs the signal properly as shown
in the amplitude spectra plots (both show a 75 Hz
signal).

However, sampling the original signal at

Δt = 8 ms changed the signal and made it appear to
be a lower frequency sinusoid (see Amp. Spectra).
The recovered signal has a frequency of 50 Hz in-
stead of the true signal frequency of 75 Hz. Figure. (left) Time series sampled at different
sampling rates. (right) Correspinding amplitude
As a result of undersampling the original 75 Hz spectrum.
signal folded back into the spectrum and appears as
an alias frequency of 50 Hz.
The phenomenon that is caused by undersampling the continuous signal is termed frequency
aliasing.

frequency aliasing will ruin your life if you're not careful.

 noise or white noise has a constant frequency
Frequency Filtering
What happens to a seismogram when we change its amplitude spectra ǀF(ω)ǀ but leave its phase
spectra φ(ω) the same?

Let's look at an example. Start with a signal in the time-domain made of a 1 Hz and a 22 Hz
sinusoids on top of a normal-distributed noise floor.

Figure. (left) Time series made

of noise and 1, and 22 Hz
sinusoids. (right) Zoom in.

The zoom in shows hints of the

1 Hz sinusoid. We can take the
Fourier transform (FFT) of the
signal to reveal its frequency
components.
If we take the Inverse Fourier Transform of
the amplitude spectra on the right we get
the original signal back. But if for example
we select only frequencies b e l o w 2 H z and
then apply the IFFT we effectively f i l t e r the
signal.

Figure. (top) Amplitude spectra with sharp

peaks at 1 and 22 Hz as expected. (bottom)
Filtered and unfiltered signal after IFFT.
once you remove the high feq.s it will look like the red line, vs. the unfilter blue:
Time Domain operations: Cross-correlation
In seismology we often want to measure how similar two signals are, or we are often interested in
finding arrivals within a seismograms that look like the direct P and direct Swave, as they might
represent reflections off discontinuities in the earth (e.g., PmP).
For example, the figure shows a noisy signal, the
same signal has been shifted in time 0.6 s and that
is the signal in the middle panel. If we compare
the similarity of the two signals we will see that
they are most similar when we shift signal one 0.6
s forward in time. This is what the lower panel is
telling us, it is giving us a measure of when the
two signals are most similar.

Now keep in mind that cross-correlation of signal

1 with signal 2 is not equal to the cross-correlation 0 12 3
of signal 2 with signal 1. In fact the latter would time (s)
result in a spike at -0.6 s as signal 2 would have to
be shifted backwards in time to have a high
similarity with signal 1.

-1.0 -0.5 0 0.5 1.0

time lag (s)
Figure. (top and middle) Two signals, one a
shifted version of the other. (bottom) Cross-cor-
relation of the two signals.
Cross-correlation
Mathematically we can show that correlation can be written as T/2

where f(t) can be the part of the signal we seek to identify re-
Cc(L) = lim
T ∞ T
1

-T/2
∫ x(t) f(t + L) dt

peated in the seismogram (e.g., direct P), and the remaining por-
tion of the time series is x(t). You can also think of f(t) as one
signal and x(t) as the other. Cc(L) is the cross-correlation of x(t) and f(t) and it measures the similari-
t y between the two signals by shifting f(t) by different lag times, L, and evaluating the integral of the
product as a function of L. The lag for which Cc(L) is maximum is the time shift that makes the two
functions most similar.

In the case of discrete time series we can write a discrete form of Cc(L). Imagine two time series xm
and fm. In discrete f o r m Ck we simply write N

Cc(k) = ∑
m=-N
xm f m + k

the 2 seismogram's frequencies at the diff. time series:

An example of the cross-correlations of two time series xm = [2, 1, -1, 0, 0], f m = [0,0, 2, 1, -1].

We start by lining up the last element of f m 2 1 −1 0 0 Output Lag

with the first element of xm. We multiply amplitude: axis of time:
0 0 2 1 −1 −2 −4
the element of f m (-1) with that of xm (2) and 0 0 2 1 −1 1 −3
that's the first output. Then we shift f m one 0 0 2 1 −1 6 −2
position. Again we multiply the elements of 0 0 2 1 −1 1 −1
each time series (1 X 2) + (-1 x 1) = 1, then we 0 0 2 1 −1 −2 0 -> 0 lag
shift again and multiply the elements (2 x 2) + 0 0 2 1 −1 0
when
1 the 2
(1 x 1) + (-1 x -1) = 6. And keep going until the 0 0 2 1 −1 0 2 series
are
first element of f m lines up with the last ele- 0 0 2 1 −1 0 3 aligned
ment of xm. 0 0 2 1 −1 0 4
What can we use cross-correlation for?
Cross-correlation methods are an important part of a field of
seismology called ambient noise tomography (ANT). We will
discuss tomography and ANT more in depth in following lectures.

The figures shown correspond to a ANT study by Young et al.

(2011) of Tasmania. The stations are shown as triangles and
squares. An example of cross-correlations between a station and
all the others is shown below. The waves shown in the cross-cor-
relation are Rayleigh waves propagating through the network.
By deriving dispersion
curves for these surface
waves the authors study
the velocity structure
(tomography) of Tasmania.

The "ambient noise" in

ANT comes from what the
source of the seismic signal
is (what causes the Rayleigh
wave energy). In this study
the majority of the signal
comes the interaction of this is telling us there are two tectonic zones, this was
done by filtering noise and create something like this. This
ocean energy hitting the is a very important current technique.

coastline and coupling onto the island as low amplitude

seismic energy, which turns out to be mostly surface waves.
can measure velocity from this diagram
dont need expensive sources for this (explosions), just record the
noise -> good for academia which is apparently very poor.
 Convolution
The definition of convolution is very similar to that of cross-correlation. ∞
Although both operations look to be very similar from their integral
definitions, they accomplish very different things. Cv(L) = ∫ x(t) f(L - t) dt

-∞
We can use deconvolution to describe a seismogram as the output
resulting from sending a source signal through a set of linear systems such as the effects of earth
structure on a earthquake signal, and the effect of the seismometer on the signal itself. This is gener-
ally written as u(t) = x(t) * g(t) * i(t), where the * signifies convolution.

An important property of convolution is that is commutative x(t) * g(t) = g(t) * x(t), unlike cross-cor-
relation. In discrete form is written as N

Cv(k) = ∑
m=-N
xm f k - m
very similar to cross-correlation, except you flip one
of the time series

An example of the convolution of two time series xm = [1, -1/2], f m = [1, 0, 1/2].
We start by reversing the time series that is going to be Reflectivity Output
shifted xm , line up the last element of x-m with the first Sequence Response
element of f m . Multiply the elements of both series
1
and shift x -m . We follow the same procedure as in the 1 0 2
cross-correlation example. 1 − 2 1 1

− 1 − 1
2 1 2

1 1
− 2 1 2

− 1 − 1
2 1 4
Convolution is a way of representing a mixture -> could use this to study the earth, or even a soup.
Properties of Convolution

An important property of convolution is the following

Convolution in the time domain is equivalent to multiplication in the frequency domain

similarly,

Convolution in the frequency domain is equivalent to multiplication in the time domain

A consequence of this property is that frequency filtering can be applied in the frequency domain as
multiplication or in the time domain as convolution.

Cv(t)

Thus Cv(t) equals F(w) * S(w)

2D Fourier Transform

When multiple seismic stations are available, seismic processing operations can be defined to operate
on several seismic traces simultaneously. This type of processing is called multichannel processing.

The 2D Fourier transform is an operation that can work on multiple seismic traces at a time. Imagine
having a line of seismometers spaced apart with constant interstation distance. This type of seismic
networ is sampling the earthquake signal in both time and space.

Before we defined frequency, particularly temporal frequency or the number of cycles per unit time.
Temporal frequency is the Fourier dual of the time variable. However, a seismic wavefield is not only a
function of time, but also a function of a space variable (distance). The fourier dual for the space vari-
able is defined as spatial frequency, which is the number of cycles per unit distance or wavenumber
(k).

Just as the temporal Nyquist frequency, the Nyquist wavenumber (k) is defined as
kNyq = 1
2Δx
where Δx is the spatial sampling interval (spatial sampling rate).
How to compute the Wavenumber k?
12 15
The top panels of the figure on the right are seismic sections.
Each vertical trace records a 12 Hz sinusoid. You can check this by
counting the number of peaks on the last trace of each section
numbered 12 and 15. The numbers on the top refer to the dip (slope

time (s)
or 1/v) in the horizontal direction (e.g. 12 ms/trace, or 15 ms/trace).

First we need the total time difference along a peak (or trough)
across the section. On the 15 ms/trace section, let's follow the peak
that starts at t = 0 s, for that peak the time difference is ~ 0.35 s.

Convert this number to cycles by dividing by the temporal period

12 Hz Sinusoids
0.35 s / T = 0.35 s / (1/12 cycles/s) = 4.2 cycles

frequency (Hz)
The spacing between traces is 25 m, there are 24 traces, thus the
spatial extend of the section is 23 * 25 = 575 m. Therefore the
wavenumber associated with the 15 ms/trace dip and the 12 Hz
frequency is
-20 20
k = 4.2/0.575 = 7.2 cycles / km (wavenumber for the particular signal)
Wavenumber
(cycles/km)
The bottom panels on the figure are known as f-k plots for
frequency vs wavenumber plots. In the figure both f-k plots
show the same frequency since both come from a 12 Hz signal, but the wavenumber is different
because the dip (velocity) for the two seismic sections is not the same. Events that have negative
slopes (dip to the left) will map into the left quadrant.
How does Spatial Aliasing looks?

Before we considered a single frequency sinusoid. For a con-

tinuum of frequency components associated with a single
dip (slope), this events map along a straight line in the f-k
domain, just like the one shown in the figure.

As with temporal aliasing, where frequencies above the

Nyquist fNyq folded back into the signal, when events are
spatially aliased, the radial line wraps around (folds) at the
Nyquist wavenumber.

In summary 2D Fourier transform is a way to decompose a

wavefield into its plane-wave components

can use this as a way to filter out

either P or S-waves.

--> freq. and wavenumber filtered

Figure. A real f-k plot showing spatial aliasing of

some events, folding back at around 8 Hz.