Download as pdf or txt
Download as pdf or txt
You are on page 1of 41

Lecture 3b: Automatic Isola in Python

(Bayesian Isola by Vackar et al.)


Short introduction for ISOLA users
J. Zahradník and E. Sokos

Charles Univ., Prague, Czech Republic


Univ. of Patras, Greece
Repetition
Uncertainty measure for a fixed
centroid position and time - new
d = Gm d = data, m= model parameters
−1 T
Cd a Cm = their covariance matrices
T −1
m =(G C d G ) G C d−1 d
~ PDF = probability density function
−1 in 6-dimensional model space
−1
C m = (G T
Cd G ) for full MT is given by m and Cm.

PDF ( mand
1 
) =mathematically
1
exp − correct
( d −way ~ T −1 ~ 
d − Gmm) and Cm
A simple d (best-fitting
Gismto) useCthe
 2many Gaussian random samples of m.
c and draw
(fully defining PDF),
MATLAB code MVNRND.
PDF of parameters PDF(m) is known analytically. Gaussian m2
in each dimension. Around 2012 (with S. Custódio) A we
constructed 6-dim. error ellipsoids and analyzed
variability of MT ‘inside ellipsoid’ of a given threshold,
in a regular grid of m1, m2….

m1
Multivariate normal distribution

https://peterroelants.github.io/posts/multivari
ate-normal-primer/
Multivariate normal distribution

Analogy with uncertainty of location (NonLinLoc)


Here 2D, but for MT of fixed position
the parameter space is 6D… https://peterroelants.github.io/posts/multivari
ate-normal-primer/
Histograms of parameters derived from
sampled distribution of Full MT
of the Korea nuclear test
Simple idea: Knowing samples of m, for each m we can evaluate e.g. its
Mo, Mw, DC%, CLVD%, etc. (whose distribution is non-Gaussian)
and plot their histograms!

Detonation

Collapse

liu_etal_GRL2018_suppl.pdf
Parameter trade-off –example Korea
test (can be easily plotted out of GUI)
Having samples of
MT, and also derived
DC%, CLVD%, ISO%, Mw,
we can study their
trade-off.

Note strong inter-relation


(correlation)! E.g. we have
big uncertainty of Mw
since Mw trades-off with
depth and DC.

Good example that any SINGLE


estimate (i.e. one DC, one CLVD…)
is not reliable.

Liu_etal_SRL2018.pdf
Part 1 – Effects of noise upon MT
uncertainty
This code aims at automatically
suppress frequencies with a low
signal-to-noise ratio
Simple uncertainty measure for a fixed
source position
d = Gm
T −1 −1 T −1
~
m = (G C d G ) G C d d
−1 −1
C m = (G T
Cd G )
CN … Condition number.
d = data, m= model1parameters
m) =  1 Here wTi are−square
− ( d − Gm ) C ( d − Gm ) 
exp matrices ~ 1 ~ 
roots of
Cd aPDF
Cm = (their covariance eigenvaluesd
c
PDF = probability density function
2 of the matrix (GT Cd-1 G) 
in definition
From 6-dimensional model
we see that space
CN depends on Green’s function, i.e.
on the assumed source position, station positions {=network geometry},
velocity model and considered frequency range.
However, CN depends also on data error, e.g. natural noise, represented by
the data-error matrix Cd.
Uncertainty assessment when also source
position (e.g. depth) is the parameter to seek

Vackar_etal_GJI2017.pdf
Data-error covariance matrix Cd
from pre-event noise
Data-error covariance matrix Cd
from pre-event noise
Synth. test: automatic suppression of
frequencies with a low signal-to-noise ratio

Almost
0 noise

High noise

Almost
0 noise
Still unclear ? Another explanation:
We do not simply minimize L2-norm
misfit between real data and synthetics,
but the L2-norm misfit between
so-called standardized real and synthetic
data (terminology by Dettmer) obtained
by means of the covariance matrix Cd.

We can imagine standardized data as


filtered in such a way that just low-noise
ranges are amplified (and high-noise
ranges are suppressed).

We avoid good fitting of a noisy signal.


The method cannot provide perfect MT if
noise is large at stations are few!
If noise is strong, uncertainty is large
Part 2 – Example: Application to 36
formerly studied earthquakes in
Switzerland
Comparison with independent MT
solutions
Validation by comparison with
SeisComP3
Algorithm
• Data and metadata were downloaded automatically. All broad-
band stations within a radius controlled by the ML magnitude were
used.

• Stations closer than 2 km were removed, because they can make


inversion unstable.

• Also eliminated were records with data gaps.

• Code setup required at least 2 stations which have at least 5


usable components, otherwise the event was removed.

• The code did not remove any station with high noise because we
wanted to test practical ability of the covariance matrix to manage
data sets with noisy stations.

• Also removed were the stations components, where instrumental


disturbances were detected automatically by the MouseTrap
code (Vackar et al., 2015) - discussed later.
Automatic solutions for 139 events of Swiss
Digital Seismic Network provided 45 trustful
MTs
The 45 trusted solutions:

• good waveform fit


• well resolved
(low condition number)
• dominant DC part
(assuming tectonic origin)
• small uncertainty
Code to download

http://geo.mff.cuni.cz/∼vackar/isola-obspy/
Sample output

Vackar_suppl_GJI2017.pdf
Limitation of the code – natural noise is the
only considered type of “data error”
Solution = to extend the code in future for including uncertainty
of velocity models and Green’s functions:

The authors analyzed Cd using Monte Carlo


sampling of velocity variations and associated waveforms shifts,
and based on the tests suggested a closed-form expression for Cd
Need to generalize the code for effect of
inaccurate velocity model Hallo_Gallovic_GJI2016.pdf
Examples of a single-component
covariance matrix
The preferred Cd type
to describe inaccurate
Green’s functions

Hallo_Gallovic_GJI2016.pdf
Part 3 – Automatic elimination of
records with disturbances
Disturbance – a “mouse”
The Mouse-Trap code

vackar_etal_SRL2015.pdf
A common “mouse” type is the response of BB
sensor to an acceleration step (possibly due to tilt)

Output of
BB sensor
A common “mouse” type is the response of BB
sensor to an acceleration step (possibly due to tilt)

Output of Output of
BB sensor BB sensor
after
INTEGRATION
A common “mouse” type is the response of BB
sensor to an acceleration step (possibly due to tilt)
Code detects records disturbed by an
input-acceleration step through modeling

Records with detected disturbances


are removed from MT inversions.
Of course, the detection can sometimes fail
if a disturbance is of a more complex type.
Disturbances are universal, not limited to
certain stations or sensor types
Limitation of the code: Disturbances of other
types than step in acceleration are not detected,
although they exist
Note a disturbance on
vertical (Z) component,
different from N and E
Disturbances are
not always due to tilt.

zahradnik_plesinger_bssa2010.pdf
“Mouse” Type 1 and Type 2

Unit steps in acceleration Full line: integrated record of STS-2


and velocity Dashed: integrated record of CMG-40T

Disturbances of Type 2 are not due to tilt.


Their origin remains enigmatic! We have no
physical explanation for a step in velocity…
zahradnik_plesinger_bssa2010.pdf
Thanks for
your attention.

Regards
from Prague.

http://www.kamvpraze.info/g
More about the tilt
Can the Earth produce an acceleration
step? Yes, on EW and NS, during tilt.
‘Step’ = everything shorter that corner period of instrument

Due to ‘sudden’ tilt,


acc_t inclination by angle φ,
the sensor
φ is affected by the horizontal
component of the acceleration
of gravity g, acc:
Example:
φ = 1e-7 rad
g . tg φ ~ g .φ = acc_t
acc_t = 1e-6 m/s2

which is large, often acc_t


acc_t > seismic acceleration
time
Relation between integrated broad
band output and input acceleration

𝑢̇ ξ ∫ξ
h int.

ξ = 𝑢̇ ∗ ℎ ∫ ξ = 𝑢̇ ∗ ∫ ℎ = 𝑢̈ ∗ ∫ ∫ ℎ
schematic log-log response
𝑢̈ ∫∫ℎ
∫ξ
f2

Response is flat to acceleration f-2


down to statics (f=0)
Integrated velocity sensor is an low-freq corner of
“expensive (low-noise) acceleration sensor” BB instrument
Why do we call integrated BB record
‘raw displacement’?

𝑢̇ ξ

ξ = 𝑢̇ ∗ ℎ ∫ξ = 𝑢∗ℎ

𝑢 ∫ξ

Response is flat to displacement
in limit of f → ∞, or f > instrument corner.
But ‘mouse’ is a low-frequency phenomenon
(check its spectrum!)
The same instrument shows different
ground-motion parameters
at different frequency ranges
True displacement value
(high frequency)

A
True acceleration value (step)
(low frequency)
‘Dot’ and ‘integral’ rules
The rules follow from spectral representation.
Dot in time domain … multiplication by (2π i f) in spectral domain.
Integral (primitive function) … division by (2π i f)

𝑢̇ ∗ ℎ … 2π𝑖𝑖 . 𝑈. H

“Moving dots”: 𝑢̇ ∗ ℎ = 𝑢 ∗ ℎ̇
“Adding dot = adding int.:” 𝑢̇ ∗ ℎ = 𝑢̈ ∗ ∫h

You might also like