E. Varini and R. Rotondi - Copernicus.org · 2020. 5. 1. · E. Varini and R. Rotondi Istituto di...
Transcript of E. Varini and R. Rotondi - Copernicus.org · 2020. 5. 1. · E. Varini and R. Rotondi Istituto di...
-
Probabilistic damage scenarios from uncertainmacroseismic data
E. Varini and R. Rotondi
Istituto di Matematica Applicata e Tecnologie InformaticheConsiglio Nazionale delle Ricerche, Milano – Italy
EGU General Assembly: Sharing Geoscience Online4-8 May 2020
-
Motivations
We consider the beta-binomial model for macroseismic attenuation (Rotondi et al.,BSSA, 2009; Rotondi et al., Bull. Earthquake Eng, 2016), which:
• respects as far as possible the ordinal nature of the intensity scale,• allows for the assumption of spatial isotropy or anisotropy,• allows for the Bayesian treatment of uncertainties,• is a probabilistic tool to produce macroseismic scenarios.
Critical point:
The application of the beta-binomial model typically requires rounding-up or -down theobserved intensities to the nearest integer values (e.g. intensity VIII-IX must be setequal to VIII or IX).
Solution:
We propose an extension of the beta-binomial model in order to include in thestochastic modelling the uncertainty in the assignment of the intensities.
E.Varini, R.Rotondi 2 / 12
-
Original beta-binomial model of the intensity Is at site
J circular bins are drawn around the epicenter (isotropy)E
r_1
r_2
r_3
r_4
•
•
••
••
•
•
•
•
•
•
•
•
•
At a given bin j :
• Is = I0 −∆I follows the binomial distribution Binom(Is | I0, pj):
Pr {Is = i | I0 = i0, pj} = Binom {i | i0, pj} =
=
(i0i
)pij (1− pj)i0−i i ∈ {0, 1, . . . , I0}
• random variable pj ∼ Beta distribution
Beta(pj ; αj , βj) =Γ(αj + βj)
Γ(αj)Γ(βj)
∫ pj0
xαj−1(1− x)βj−1dx
to account for the variability in ground shaking
The prior hyperparameters αj , βj are assigned on the basis of the macroseismic fieldsbelonging to the same class, but of different I0.
E.Varini, R.Rotondi 3 / 12
-
New beta-binomial model of the intensity Is at site
J circular bins are drawn around the epicenter (isotropy)E
r_1
r_2
r_3
r_4
•
•
••
••
•
•
•
•
•
•
•
•
•
At a given bin j :
• Is = I0 −∆I follows the binomial distribution Binom(2Is | 2I0, pj):
Pr {Is = i | I0 = i0, pj} = Binom {2i | 2i0, pj} =
=
(2i02i
)p2ij (1− pj)2i0−2i i ∈ {0, 0.5, 1, 1.5 . . . , I0}
• random variable pj ∼ Beta distribution
Beta(pj ; αj , βj) =Γ(αj + βj)
Γ(αj)Γ(βj)
∫ pj0
xαj−1(1− x)βj−1dx
to account for the variability in ground shaking
The prior hyperparameters αj , βj are assigned on the basis of the macroseismic fieldsbelonging to the same class, but of different I0.
E.Varini, R.Rotondi 4 / 12
-
Updating parameters in the Bayesian framework(Fortran software)Given all the earthquakes with epicentral intensity I0 in a class,
let Nj be the number of felt intensities Dj in the j-th bin
Original beta-binomial model: αj = αj,0 +
Nj∑n=1
i (n)s βj = βj,0 +
Nj∑n=1
(I0 − i (n)s )
New beta-binomial model: αj = αj,0 +
Nj∑n=1
2i (n)s βj = βj,0 +
Nj∑n=1
(2I0 − 2i (n)s )
I0 VIII, class B
0 100 200
Distance [km]
0
0.2
0.4
0.6
0.8
1
Pa
ram
ete
r p
B We estimate the parameter pj through its posterior mean
p̂j = E (pj | Dj) =αj
αj + βjj = 1, . . . , J
E.Varini, R.Rotondi 5 / 12
-
Beta-binomial model with smoothed p = p(d)The estimates p̂j are smoothed by the inverse power function of the distance throughthe least squares method:
p(d) =
(c1
c1 + d
)c2Smoothed binomial distribution at any distance d
Pr(Is = i |I0 = i0, p(d)) =(
i0i
)p(d)i [1− p(d)](i0−i)
Pr(Is = i |I0 = i0, p(d)) =(
2i02i
)p(d)2i [1− p(d)](2i0−2i)
I0 VIII, class B
0 100 200
Distance [km]
0
0.2
0.4
0.6
0.8
1
Pa
ram
ete
r p
B
0 1 2 3 4 5 6 7 8
Is
pro
ba
bili
ty
0.0
0.1
0.2
0.3
0.4
Io VIII
p(d) = 0.9 at d = 10Km
0 1 2 3 4 5 6 7 8
Is
pro
ba
bili
ty
0.0
0.1
0.2
0.3
0.4
Io VIII
p(d) = 0.7 at d = 50Km
The mode ismooth of the smoothed binomial distribution is the predicted value of Is
E.Varini, R.Rotondi 6 / 12
-
Learning set
Analysis of 441 macroseismic fields (MFs) of “good quality” from the ItalianDBMI15 database:
− occurred since 1500,
− at least epicentral intensity V ,
− at least 40 felt reports.
Database DBMI15
Distribution of maximum macroseismic intensity
E.Varini, R.Rotondi 7 / 12
-
Hierarchical agglomerative clustering (R package cluster)
The learning set is first analyzed by the Ward’s hierarchical agglomerative clusteringmethod.
Four attenuation classes are identified.
class A class B class C class D132 MFs 192 MFs 82 MFs 35 MFs
10.69°E 10.99°E 11.28°E 11.57°E 11.87°E 12.16°E 12.45°E
43.29°N
43.5°N
43.72°N
43.93°N
44.14°N
44.35°N
44.56°N
1919/06/29, Io X: observed
Is
IIIIIIIVVVIVIIVIIIIXX
15.1°E 15.39°E 15.69°E 15.98°E 16.27°E 16.57°E 16.86°E
37.65°N
37.87°N
38.1°N
38.33°N
38.56°N
38.79°N
39.02°N
1907/10/23, Io VIII−IX: observed
Is
IIIIIIIVVVIVIIVIIIIX
10.72°E 11.9°E 13.07°E 14.24°E 15.41°E 16.58°E 17.76°E
43.77°N
44.58°N
45.39°N
46.2°N
47.01°N
47.82°N
48.63°N
1895/04/14, Io VIII−IX: observed
Is
IIIIIIIVVVIVIIVIII
7.73°E 8.91°E 10.08°E 11.25°E 12.42°E 13.59°E 14.77°E
42.58°N
43.41°N
44.23°N
45.06°N
45.89°N
46.71°N
47.54°N
1914/10/27, Io VII: observed
Is
IIIIIIIVVVIVII
0 50 100 150 200 250
0
2
4
6
8
10
12
0 50 100 150 200 250
0
2
4
6
8
10
12
0 50 100 150 200 250
0
2
4
6
8
10
12
0 50 100 150 200 250
0
2
4
6
8
10
12
E.Varini, R.Rotondi 8 / 12
-
Smoothed posterior distribution of intensity Is for class A
Pr {Is = i | I0 = i0, pi0 (d)} = Binom {2i | 2i0, pi0 (d)} where pi0 (d) =(
c1c1 + d
)c2
I0 n.MFs c1 c2V 22 13846.42 219.98
V-VI 17 - -VI 21 113.96 2.16
VI-VII 3 - -VII 17 1989.44 29.01
VII-VIII 10 - -VIII 8 9674.80 126.72
VIII-IX 0 - -IX 14 3573.55 37.52
IX-X 4 - -X 10 27752.59 254.40
X-XI 2 - -XI 4 36102.69 332.45
The number of MFs having uncertain epi-central intensity I0 is sometimes very smallor even null; it follows that the correspond-ing estimated coefficients c1 and c2 mightbe quite unrealiable.
In order to obtain more reliable and stableestimates for uncertain epicentral intensityI0 = i0, its parameter pi0 (d) is chosen asfollows:
pi0 (d) =pdi0e(d) + pbi0c(d)
2
E.Varini, R.Rotondi 9 / 12
-
Probability of the intensity attenuation versus distance
Original model New model New model
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250I
II
III
IV
V
Italy: Io V , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
Italy: Io V , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V Italy: Io V−VI , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250I
II
III
IV
V
VI
Italy: Io VI , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
Italy: Io VI , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI Italy: Io VI−VII , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250I
II
III
IV
V
VI
VII
Italy: Io VII , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
Italy: Io VII , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII Italy: Io VII−VIII , class A
Km
Is
E.Varini, R.Rotondi 10 / 12
-
Probability of the intensity attenuation versus distance
Original model New model New model
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250I
II
III
IV
V
VI
VII
VIII
Italy: Io VIII , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
VIII
Italy: Io VIII , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
VIII Italy: Io VIII−IX , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250I
II
III
IV
V
VI
VII
VIII
IX
Italy: Io IX , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
VIII
IXItaly: Io IX , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
VIII
IX Italy: Io IX−X , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250I
II
III
IV
V
VI
VII
VIII
IX
X
Italy: Io X , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
VIII
IX
XItaly: Io X , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
VIII
IX
X Italy: Io X−XI , class A
Km
Is
E.Varini, R.Rotondi 11 / 12
-
Probability of the intensity attenuation versus distance
Original model New model
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250I
II
III
IV
V
VI
VII
VIII
IX
X
XI
Italy: Io XI , class A
Km
Is
0.0
0.2
0.4
0.6
0.8
1.0
0 50 100 150 200 250
I
II
III
IV
V
VI
VII
VIII
IX
X
XIItaly: Io XI , class A
Km
Is
References
Rotondi R., Zonno G. (2004), Bayesian analysis of a probability distribution for local intensityattenuation, Ann. Geophys., 47 (5), 1521-1540
Zonno G., Rotondi R., and Brambilla C. (2009), Mining macroseismic fields to estimate theprobability distribution of the intensity at site, BSSA, 99 (5), 2876-2892
Azzaro R., D’Amico S., Rotondi R., Tuvè T., ZonnoG. (2013), Forecasting seismic scenarios onEtna volcano (Italy) through probabilistic intensity attenuation models: A Bayesianapproach, J. Volcanol. Geotherm. Res., 125, 149-157
Special issue Bull. Earthq. Eng. (2016) 14, 7
E.Varini, R.Rotondi 12 / 12
MotivationsClustering