Transcript ppt - International Seismological Centre

Improved Location Procedures at the
International Seismological Centre
István Bondár
ESC 32nd General Assembly
Montpellier, September 6-10, 2010
Outline
• Motivation
• Current
ISC locator
• Proposed ISC locator
• Validation tests
•
•
Relocation of ~7,000 GT0-5 events
Comparison with EHB
• Conclusions
2
Motivation
Correlated travel-time prediction errors along similar ray
paths introduce location bias and result in underestimated
error ellipses
• Example:
•
•
•
The effect of USArray on event locations in South America
Without accounting for correlated errors, locations get worse!
3
Current ISC Locator
ak135, P[g|b|n], S[g|b|n] up to 100º
• Jeffrey’s uniform reduction algorithm
•
•
•
•
Iterative reweighting, all phases are equal
Reidentifies phases in each iteration (S could become pP!)
Duplicates are explicitly down-weighted
Uncertainties are scaled to 95% confidence level
• Attempts free-depth solution and cycles through all reported
hypocentres until a convergent solution is found
•
•
•
If fails, fixes depth to that of the trial hypocentre
If fails, fixes depth to region-dependent default depth (10/35 km)
Depth-phase depth solution is obtained by inverting pP-P
times
• Network magnitude can be obtained from a single station
magnitude; no magnitude uncertainties are calculated
•
4
New ISC Locator
•
Uses all ak135 predicted phases (including depth phases)
in location
•
•
Ellipticity, elevation corrections
Bounce point and water depth corrections for depth phases
Attempts free-depth solution only if there is depth resolution
• Accounts for correlated model error structure
• Obtains initial guess via neighbourhood algorithm search
• Performs iterative linearized inversion using a priori
estimate of data covariance matrix
• Scales uncertainties to 90% confidence level
• Obtains depth-phase depth via depth-phase stacking
• Provides robust network magnitude estimates with
5
uncertainties
•
Data Covariance Matrix
When correlated errors are present, the data covariance
matrix is no longer diagonal
 The network covariance matrix accounts for correlated
travel-time prediction errors due to similar ray paths

 Estimated
by an isotropic stationary variogram (Bondár and
McLaughlin, 2009) derived from GT residuals
The a priori picking error variances add to the diagonal
Generic variogram model
C
D = CN + CR

2
CN (i, j)   sill
  ((stai , sta j ))
6
Depth Resolution
• Attempt
free-depth solution if and only if there is
depth resolution
• We have depth resolution if
•
There is a local network
•
•
Or there are sufficient number of depth phases
•
•
5 or more defining depth phases reported by at least two
agencies
Or there are sufficient number of core reflections
•
•
one or more defining stations within 0.2º
5 or more defining core reflections (PcP et al)
Or there are sufficient number of local/near-regional S
•
5 or more defining S and P pairs within 5º
• Otherwise
fix the depth to regional default depth
• Explosions are fixed to zero depth
7
Neighbourhood Algorithm Grid Search
The prime location may or may not be close to the global
optimum in the search space
• Reported hypocentres may exhibit a large scatter
• Neighbourhood Algorithm (Sambridge and Kennett, 2001) to
get an initial hypocentre
•
•
Grid search around the median of reported hypocentre parameters
NA explores the search
space and rapidly closes in
on the global optimum
• Once close to the global
minimum, we switch to the
linearized inversion algorithm
to obtain the final solution
and formal uncertainties
•
8
Location Algorithm
• Minimize d  GmT CD1 d  Gm by solving WGm  Wd
–G is the design matrix containing the travel-time derivatives for an
event-station path, m is the model adjustment vector [Δx, Δy, Δz,
ΔT]T and d is the vector of time residuals.
–W is the projection matrix that orthogonalizes CD and projects
redundant observations into the null space (Bondár and
McLaughlin, 2009).
• The SVD of CD is CD  U D ΛDVD
• We keep only the first p largest eigenvalues from the eigenvalue spectrum
p
N

such that 95% of the total variance is explained.
 j j i i  0.95
• This reduces the effective number of degrees of freedom of the data from N
to p, with N-p dimensions of null space.
• Let B  U p Λ1p/ 2 , CD  BBT then W  B1  Λp1/ 2U Tp ; Gw  WG , dw  Wd
T
• Solution via singular value decomposition
Gwm  dw ; Gw1  Vw Λw1U wT ; mest  Gw1dw
• A posteriori model covariance matrix
1
CM  G CDG
1T
 Vw Λw2VwT
m j 1  m j  mest
9
Depth-phase Depth
2001/01/28 17:15:27.363, 23.28, 70.04, 28.5
Depth-phase stack (Murphy
and Barker, 2006)
• Generate predicted moveout
curves
•
•
•
Generate depth traces for
each station
•
•
depdp=26.0 ± 8.9
•
TTdepth-phase – TTfirstP
Parametric on delta
Boxcar at the corresponding
depth for the observed moveout
The width of boxcar is the prior
measurement error estimate
Stack depth traces
•
Calculate median and SMAD
10
Validation Tests
•
Relocation of some 7,200 GT0-5 events
•
•
•
Mimic automatic ISC location
•
•
•
•
Earthquakes and explosions
Some explosions are poorly recorded
ISC associations
Use only reported hypocentres
Ignore GT, EHB and ISC hypocentres
Cases
•
•
•
•
•
•
•
Baseline: current ISC locator (isc)
Independent errors, current ISC phases (ii)
Independent errors, all ak135 phases (ia)
Independent errors, all phases, grid search (iag)
Correlated errors, current ISC phases (ci)
Correlated errors, all ak135 phases (ca)
Correlated errors, all phases, grid search (cag)
11
Explosions
• Abysmal
coverage when correlated error
structure is ignored
12
Test Sites
French Polynesia
Novaya Zemlya
Semipalatinsk
•
NTS
Correlated errors
•
•
Significant improvements in French Polynesia
Deterioration in Novaya Zemlya
•
•
Lop Nor
When the effect of conspiring stations is taken out, the
poor ak135 regional TT predictions become apparent
Grid search
•
Does a good job in rectifying poor initial
hypocentres
13
GT5 Earthquakes (free-depth)
Improved locations, depth and OT
14
GT5 Earthquakes (depth-phase depth)
Significant improvements in location, depth and OT15
Regional Default Depth
•
Get a reasonable default depth where there is seismicity
•
First attempt:
•
•
Second attempt:
•
•
•
•
1x1 degree grid of EHB free-depth solutions
Improvements were shown for free-depth hypocentres of GT events with
the new locator
Relocated the entire ISC bulletin with the new locator
0.5x0.5 degree grid from relocated free-depth solutions + EHB free-depth
and EHB reliable depth solutions (~800K events)
Otherwise use CRUST2.0
16
GT5 Earthquakes (fixed depth)
•
Default depth grid: version2 vs version1
•
•
•
Significant improvements in depth
Improvements in location w.r.t. EHB grid
17
Most of the events now have a default depth based on seismicity
Effect of Correlated Errors
All events
•
•
•
90% percentile
Area of error ellipse does not
vanish with increasing number
of stations
Actual coverage is maintained
Location and depth bias is
reduced for large number of
stations
50% percentile
18
Comparison with EHB
EHB earthquakes in IASPEI
Reference Event List
• Overall improvements in
location
• Depth estimation is
consistent with EHB
• Results are comparable or
better than EHB
•
19
Conclusions
• Accounting
•
•
for correlated errors
Provides honest ~85-90% actual coverage
Reduces location bias due to correlated travel-time
prediction errors
• Improvements
in location, depth and origin time for
free-depth solutions are due to
•
•
Using all phases (including depth phases) in location
Testing for depth resolution
• Default
depth grid provides reasonable depth for
fixed-depth earthquakes
• Initial guess by neighbourhood algorithm grid search
•
Crucial for poorly recorded events with unreliable
reported hypocentres
• Location
20
and depth accuracy is comparable to EHB