Professional Documents
Culture Documents
Comparison and Modelling of Building Losses in South Iceland Caused by Different Size Earthquakes
Comparison and Modelling of Building Losses in South Iceland Caused by Different Size Earthquakes
Comparison and Modelling of Building Losses in South Iceland Caused by Different Size Earthquakes
A R T I C L E I N F O A B S T R A C T
Keywords: In June 2000, two earthquakes o Mw6.52 and Mw6.44 occurred with a 4-day interval in the
Seismic vulnerability South Iceland seismic zone, and in May 2008, a Mw6.30 earthquake occurred urther west in the
Fragility unctions zone. Their epicentres and ault ruptures were close to small towns and arms, and almost 5000
Zero-inated beta regression residential buildings were aected in each o these two seismic events. Despite substantial
Seismic codes damage, no residential building collapsed and there were no atalities. Catastrophe insurance o
GEM taxonomy
buildings is mandatory in Iceland. Driven by insurance claims, repair cost was estimated or every
ESRM20
aected building, and registered in two independent datasets, called hereater the 2000 and the
2008 dataset. Two complete loss datasets rom dierent size earthquakes, aecting the same
building typologies in the same region is rare to fnd in the literature. Valuable inormation can be
drawn by comparing the losses caused by these events on the housing stock. In this study an
advanced empirical vulnerability model based on zero-inated beta regression was ftted to fve
building typologies, based on GEM taxonomy, independently or the 2000 dataset and the 2008
dataset. The peak ground acceleration (PGA) was used as intensity measure. The vulnerability
model and the ragility curves are substantially dierent rom these two datasets. For a given
PGA, the mean losses based on the 2000 dataset was typically roughly twice that based on the
2008 dataset. Such a large dierence in losses or similar building types located in the same area
indicates that PGA is not alone an adequate intensity measure to predict losses. The study also
showed that high code buildings perormed better than low code buildings.
1. Introduction
Seismic hazard is high in Iceland [1,2]. In this century alone, three destructive earthquakes have struck in the South Iceland Seismic
Zone (SISZ). On 17 and June 21, 2000, two, similar size, Mw6.52, and Mw6.44, earthquakes struck in the eastern part o the zone [2].
These events occurred on two shallow strike-slip aults oriented north-south, and approximately 16 km apart (Fig. 1). They aected
nearly 5000 low-rise residential buildings. On May 29, 2008, a Mw6.30 earthquake struck urther west in the zone and again aected
nearly 5000 buildings [2,3]. The area inuenced by these events is an agricultural region with small towns, arms, bridges and all the
inrastructure assets o a modern society. Mandated by law, all properties in Iceland are insured against natural hazards, including
earthquakes [4]. Thereore, ater the two 2000 earthquakes and again ater the 2008 event, monitory losses were estimated or each
damaged building to address insurance claims. Registers Iceland maintains a detailed property database containing comprehensive
inormation about all building units in Iceland [5]. Combination o the property database and loss paid out or damaged buildings
* Corresponding author.
E-mail addresses: bb@hi.is (B. Bessason), rajesh@hi.is (R. Rupakhety), jon@nti.is (J.Ö. Bjarnason).
https://doi.org/10.1016/j.jobe.2021.103806
Received 11 September 2021; Received in revised orm 15 November 2021; Accepted 30 November 2021
Available online 3 December 2021
2352-7102/© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Fig. 1. Epicentre and ault ruptures o the three earthquakes considered in this study. Residential buildings and settlements are marked with red dots, with their size
scaled to the population. The 2008 event occurred closer to the densely populated areas than the 2000 events. The area within 30 km rom the ault rupture was
aected by the 2008 event and within 35 km or the 2000 events. The capital area o Reykjavik is outside these ranges. (For interpretation o the reerences to colour in
this fgure legend, the reader is reerred to the Web version o this article.)
provides a complete loss dataset. Loss datasets corresponding to the 2000 and 2008 events are hereater reerred to as the 2000 dataset
and the 2008 dataset, respectively. The two datasets are complete in the sense that they include every building exposed to estimated
PGA o 005g or more, and thereore give unique possibilities to study dierent aspects o seismic vulnerability [6]. The datasets have to
some extent, been analysed and modelled in previous studies concentrating on losses in residential buildings. The frst vulnerability
model using only the 2008 dataset was presented by Bessason et al., in 2012 [7]. A detailed overview o how the losses in the 2008
dataset were classifed into dierent subcategories o structural and non-structural damage was reported in Re. [8]. Classical ragility
unctions using the lognormal distribution were calibrated using both the 2000 and 2008 datasets combined into one dataset in
Re. [9]. Then Ioannou et al. in Re. [10] used an advanced beta-regression [11] to model the losses caused by the two June 2000
earthquakes. Buildings located between the two aults o the 2000 earthquakes (Fig. 1) were to some degree aected by both events,
and thereore damage accumulation rom the two events was also considered in this beta regression model. Finally, Bessason et al. [12]
improved the loss modelling by applying a zero-inated beta regression model (ZIBRM) [13] to better account or the large proportion
o buildings that suered no damage in the June 2000 events. This last model also has more exibility than the previous beta model in
Re. [10] in two aspects. First, the regression parameters are calibrated or each building typology separately, and second, two extra
parameters are used to model the zero-losses, in total fve model parameters or each building typology [12].
To our best knowledge, the frst use o ZIBRM to model post-earthquake loss data is in Re. [12] authored by same authors as this
study. The most important novelty o the ZIBRM is that it treats the no-loss buildings specially and help to bend the vulnerability curves
down towards zero loss at low intensity. By this the models better reect the dominance o no-loss data in that intensity range.
Presenting the capability o using this type o advanced statistical tool in empirical vulnerability assessment is thereore believed to be
important contribution in the feld.
The objective o this study is threeold. First, to study the eect o earthquake size by comparing the losses rom the two datasets.
Availability o complete loss data rom two dierent sized earthquakes aecting the same building stock in the same region makes this
comparison interesting and valuable. The comparison is done frst by exploration o loss data and then by comparison o calibrated
vulnerability models. The second objective is to study the inuence o seismic code status on the perormance o building stock. The
buildings in the earthquake zone were designed and constructed over a long time and thereore covering period when no seismic
2
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Table 1
Status o seismic design codes or the Icelandic building stock or dierent construction periods [14].
1) The hazard map or Iceland presented in 1958 in the Icelandic Journal o Engineering [24] initiated seismic design in the country. Most designers began applying
static lateral orce in their structural analysis although not ofcial requested by regulations or codes.
design codes existed to the modern-day seismic code enorcement. Status o code enorcement can be conveniently described in terms
o no-code to high-code classes proposed by Crowley et al. [14]. Existing vulnerability and ragility models commonly take into ac-
count ductility classes (low, medium, high) but in this study we instead connect the models to status o seismic codes (no-code,
low-code, moderate-code and high-code). Seismic loads can or example increase when going rom say moderate-code to high-code,
although the ductility can be classifed as DCL or a given building or both code periods. The third objective is to use the GEM building
taxonomy [15] to classiy the Icelandic loss data or frst time and then to calibrate local vulnerability model based on this classif-
cation. By this the local model can more easily be compared to other global vulnerability models and added to existing seismic risk
platorms [16–18]. In this study the ragility and vulnerability unctions rom ESRM20 [19] will be used or comparison.
Most studies reported in the literature construct ragility curves and convert them to vulnerability models using damage-to-loss
modes [19]. In this study, the statistical vulnerability model is calibrated directly rom the loss data, and ragility curves are iner-
red rom it.
3
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Table 2
Classifcation o residential buildings aected by the June 2000 and May 2008 earthquakes based on the GEM building taxonomy [15] and based on seismic code status
(Table 1).
GEM building Design code June 2000 earthquakes May 2008 earthquake
system or the great majority o the building typologies are structural walls, hence LWAL is used to identiy the structural system or the
three building typologies. Frame structures with inflls or stone/brick masonry units common in South Europe are rare to fnd in
Iceland. Clay is also scarce in Iceland and there is no production or supply o low-cost fred clay bricks. The ground oor slab and the
oundations in both timber and masonry buildings are usually made o RC. The structural walls o the masonry buildings are built o
unreinorced, hollow concrete blocks using light-weight pumice (high porosity volcanic rock) as the main aggregate. The weight density
o the pumice blocks is low, typically around 14 kN/m3, and consequently the inertia orces are considerably lower than in stone or
clay brick masonry buildings with higher weight density. These hollow pumice blocks have not been manuactured or many decades
(not ater ~1975) and masonry buildings have hardly not been built since then (Table 2).
In the 2000 dataset, 54% o the buildings were built o reinorced concrete (RC). Almost all cast-in-place (CIP). Furthermore, 37%
were timber buildings, 9.3% masonry buildings, and the rest (only 0.3%) used other main construction material (not shown in Table 1).
Moreover, 68% o the buildings were one storey, 23% two storeys, 7.9% three storeys, 0.3% our storeys, and no building was higher.
In the 2008 dataset, 45% o the dwellings were built o RC, 48% timber and 7.6% masonry dwellings. Furthermore, 74% were one
storey buildings, 19% two storeys buildings, 5.9% three storeys, and 0.5% our storeys. Since only a low raction o the aected
buildings are 3 storeys or higher all the buildings in both the datasets are placed in HBET:1,2 class. Typical oor area o a dwelling is in
the range 100–150 m2. In buildings built beore 1976 (CDN and CDL), the concrete strength was low, typically corresponding to C16
concrete (ck = 16 MPa). Ater 1976 (CDM and CDH) it was increased, usually to C20 or C25 concrete. Today, only ribbed high grade
steel bars with yk = 500 MPa are used or reinorcement, but beore 1965, non-ribbed low-grade bars with yk = 235 MPa were the only
alternative. It was common in this period, i.e. beore 1965, to use only one or two horizontal 12 mm steel bars above window and door
openings. Between 1965 and 1976, this reinorcement was generally increased to one or two 12 mm bars around all openings [7]. Ater
1976 structural walls usually have reinorcement grid. It should be noted that no buildings in the 2000 dataset belong to high code as
they were all constructed beore the Eurocodes were implemented in Iceland in 2002 (Table 1).
4
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Table 3
Subcategories o the loss data used or both datasets.
Table 4
Subcategories o structural damage and non-structural damage in the 2008 dataset.
Fig. 2. Number o buildings within dierent PGA bins, or dierent construction material, and or both the datasets.
fxtures, as well as technical systems (plumbing, electrical installations etc.) but does not include damage o loose household items like
urniture, TVs, computers, etc. In the 2000 dataset the damage (estimated repair cost) in each building was classifed into fve sub-
categories, two covering structural damage and three or non-structural damage (Table 3). In the 2008 dataset ten subcategories were
used (Table 4) [8]. The 2008 dataset can eortlessly be combined to obtain identical categories as in the 2000 dataset. The damage
actor, DF, is defned as:
5
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Fig. 3. Proportions o ‘No Loss’, ‘Loss’, and ‘Total loss’ buildings, within dierent PGA bins, or dierent construction materials, and or both the datasets.
Fig. 4. Mean damage actor within dierent PGA bins, or dierent construction materials, and or both the datasets. The bars also show how the loss was split into
structural and non-structural loss in each case.
6
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Fig. 5. Number o dwellings (N), observed mean damage actor (DF), and observed damage stages in Seloss (PGA ~ 0.55g) and Hveragerdi (PGA ~0.60g), or
dierent construction material and dierent status o seismic codes in the May 2008 Ölus earthquake.
5
∑
ECi
Estimated loss
DF = = i=1 (1)
Replacement value RV
where ECi is the estimated repair cost in each subcategory i (Table 3) and RV is the replacement value taken as the fre insurance
value rom the ofcial property database. The RV is estimated as the depreciated replacement value plus the cost o dismantling and
transporting the debris. Depreciation is based on age, main construction material and general condition. On the other hand, the repair
cost was not depreciated.
The DF can reach 1.0 (100%) or buildings with total damage and correspond to payment o ull replacement values to the building
owners. Damage equivalent to “total loss” was assigned to buildings that suered an estimated repair cost o more than 70% o their
replacement value in the 2000 dataset. In 2008, the level was lower and estimated on a case-by-case basis or each building with
estimated loss in the range 50–70%.
3. Data exploration
3.1. Classifcation with respect to intensity measure
To explore the loss data, computed PGA at each building site as well as main construction material (RC, timber and masonry) was
used to classiy the loss data into our PGA bins, i.e., 0.05–0.10g, 0.10–0.20g, 0.20–0.40g and >0.40g (Fig. 2). In the 2000 dataset most
o the buildings were exposed to low PGA, i.e., they belong to the two frst PGA bins (Fig. 2abc). For the 2008 dataset this was opposite
(Fig. 2de), and most o the buildings were exposed to PGA in the two higher bins. The main reason or this is that the two largest towns
in the region (Seloss and Hveragerdi) were located within 5 km rom the ault rupture o the 2008 event (Fig. 1).
For both the June 2000 and the May 2008 earthquakes, most o the buildings in the two lowest PGA bins did not suer any loss. The
higher PGA bins, on the other hand, contained buildings with no loss, partial loss, and total loss (Fig. 3). This is also reected in the
mean DF, which is very low in the two lower PGA bins in all cases but substantially higher in the two higher PGA bins (Fig. 4). The
7
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
mean DF in the 2000 dataset is much higher than that in the 2008 dataset (Fig. 4). In more details, or a given main construction
material the two datasets diered in structural losses (category E1 and E2 in Table 3) but showed comparable levels o non-structural
losses (categories E3, E4 and E5). The similarity in the extent o non-structural damage can partly be explained rom the analysis o the
loss data rom the 2008 earthquake where most o the losses were due to cosmetic damage o ooring and paint work [8]. Such minor
damage oten results in replacement o all the ooring and repainting o the entire wall in question. Thereore, even minor cosmetic
damage can result in notable repair costs or non-structural elements.
Since both the 2000 and 2008 earthquakes occurred in the same region and aected alike buildings (same codes, same work-
manship, same detailing, same construction material etc.), the observed dierence in mean DF in each PGA bin (Fig. 4), indicates that
the PGA alone is not an adequate ground motion IM to predict the structural losses. Other parameters, or example the size o the
earthquake, which controls strong shaking duration among other things, are also expected to aect the losses suered by the structural
elements (see also [30]).
4. Statistical modelling
In total, 84% and 50% o buildings/dwellings suered no losses (DF = 0) in the 2000 and 2008 datasets, respectively. Total losses
(DF = 1) were on the other hand very rare. When proportion data, like damage actor (DF), contains both “zero” values and “one”
values, i.e. the data is bounded in the range [0,1], it is preerable to use a mixed continuous-discrete regression to model the data [13].
That is, discrete models to cover the “zeros” and “ones”, and then continuous regression or data in the range (0,1). In our case where
the data includes high raction o zero loss incidents but negligible number o total losses, a zero-inated beta regression model is
well-suited [13]. This model was used to ft the 2000 dataset in a two-step regression by Bessason et al. [12]. First, a logistical
regression model was used to evaluate the probability o obtaining loss (DF > 0) or a given PGA. Then a beta regression was applied to
model the continuous loss distribution conditioned on loss. This approach is explained briey in the ollowing sections, but more
details are available elsewhere [10–13].
where β0 and β1 are the regression parameters, which are calibrated separately or each building typology and each dataset.
8
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Table 5
Building typologies based on GEM taxonomy [15] and seismic code categories [14].
Based on their defnition the probability density unction, expected value and variance o the beta distribution, are respectively given
as:
Γ(φ)
f (x; μ, θ) = xμφ1 (1 x)(1μ)φ1 0<x<1
Γ(μφ)Γ(1 μ)φ
E[x] = μ 0<μ<1 (3)
μ(1 μ)
Var[x] = φ>0
1+φ
where μ is the mean value; φ is the precision, and Γ(⋅) is the gamma unction [11]. To consider that the shape o the distribution change
with the intensity o the ground motion a beta regression model uses link unctions, g1(⋅) and g2(⋅) to connect μ and φ, with linear
predictors, η1 and η2 that are commonly unctions o a one or more explanatory variables:
g1 (μ) = η1
(4)
g2 (φ) = η2
The link unctions must be strictly monotonic and twice dierentiable. In this study, the logit link unction was adopted or μ and
the log link unction or φ:
( )
μ
g1 (μ) = logit(μ) = log = η1
1μ (5)
g2 (φ) = log(φ) = η2
Having determined the main properties o the statistical model, the linear predictors η1 and η2 need to be expressed as unctions o
the explanatory variables. As mentioned beore the building typologies were treated independently to have maximum exibility or the
regression parameters but at same time to keep the model simple and exclude categorial explanatory variables. Only one explanatory
variable was used or the linear predictors η1, i.e., PGA like in Res. [10,12]. I the residuals appear to be randomly distributed, the use
o a constant precision is adequate or the linear predictor η2 or the precision. Based on observations and arguments put orward in
Re. [10], a constant precision model was adopted or φ in this study:
η1 = θ0 + θ1 × log(PGA)
(6)
η2 = θ 0
′
here θ0, θ1, and θ’0, are the parameters o the beta regression model which are calibrated separately or each building typology as
mentioned beore. Log-transormed PGA has been reported (see, or example [10]) to provide better ft or the linear predictor η1.
E[DF] = p · μ (7)
μ · (1 μ)
Var[DF] = p · + (1 p) · p · (μ)2 (8)
φ+1
The total probability theorem can then be used to compute desired prediction interval [12]:
P[X < x] = 1 + p ·(FX (x, μ, φ) 1) (9)
where FX(x,μ,φ) is the beta cumulative distribution unction (CDF) or a given building typology or case. Eq. (9) can also be used to
construct ragility curves by defning threshold values, x, or exceeding a damage state or given IM (see section 5.2).
9
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Fig. 6. Probability o getting DF > 0 as a unction o PGA or three dierent building typologies in South Iceland based on logistic regression models o the 2000
dataset (solid lines) and 2008 dataset (dotted lines).
Table 6
Number o buildings in each typology class (Ntot) and number o damaged buildings (Ndam), and estimated model parameters and standard error (SE), or the 2000 and
2008 loss dataset.
Ntotal Ndam β0 β1 θ0 θ1 ′
θ0
2000 dataset:
RCNo-Low Code 1665 346 3.471 0.152 12.85 0.81 1.597 0.044 0.304 0.027 1.450 0.027
RCMod-High Code 775 85 3.638 0.216 10.46 1.02 2.490 0.145 0.316 0.088 2.673 0.102
WNo-Low Code 692 89 3.804 0.256 11.61 1.27 1.590 0.067 0.304 0.041 1.498 0.043
WMod-High Code 1047 129 3.569 0.189 6.73 0.59 2.752 0.111 0.071 0.081 2.405 0.084
MRNo-Low Code 443 110 3.025 0.259 11.37 1.35 0.360 0.031 0.724 0.020 1.012 0.017
2008 dataset:
RCNo-Low Code 1112 634 2.405 0.183 7.245 0.461 2.260 0.032 0.303 0.037 2.682 0.024
RCMod-High Code 1003 546 2.346 0.191 5.952 0.421 2.900 0.029 0.321 0.038 3.742 0.026
WNo-Low Code 649 299 1.424 0.194 2.876 0.392 2.297 0.056 0.148 0.060 2.365 0.046
WMod-High Code 1623 694 1.968 0.136 4.324 0.315 3.001 0.011 0.232 0.012 3.635 0.010
MRNo-Low Code 359 209 1.940 0.285 5.937 0.692 1.446 0.077 0.268 0.084 1.255 0.050
seismic codes. We note here that the eect o earthquake size is potentially reected in other ground shaking parameters such as
duration, Arias Intensity, etc., but in lack o appropriate and local ground motion prediction models or such parameters and or
comparative purpose, a simple classifcation based on earthquake size is used in this study. Based on data exploration described in
section 3.2, two code classes are defned or RC buildings, one consisting o CDN and CDL and the other consisting o CDM and CDH.
For masonry buildings, a single class consisting o CDN and CDL is defned. This results in fve building typologies. The ull names and
short names o the building typologies based on the GEM nomenclature and code status as defned in Crowley et al. (2021) is given in
Table 5. Calibrating the model parameters independently or the 2000 and the 2008 dataset thereore results in 10 parameter sets (5
typologies × 2 datasets = 10 cases).
10
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Fig. 7. Comparison o the vulnerability model rom the two datasets. The dots show the mean DF based on the loss data within each acceleration bin. The PGAs o the
binned data are average values in each bin. (see Fig. 4).
replaced with a max value o DFmax = 0.85. In the 2000 dataset this was done or 15 RC, 5 timber and 13 masonry buildings, and in the
2008 dataset or 4 RC, 7 timber, and 12 masonry buildings.
Another important issue is weighting o loss data beore ftting the conditional beta model. Newer buildings have higher
replacement value than older buildings. Replacement value is also aected by building size. These issues are not reected in the
normalised DF but aect the prediction o total accumulated loss. It is thereore sensible to put the suered losses in the context o the
replacement value o the property by introducing a suitable weighting scheme, see also [12]. The ollowing actor, wi, was used to
weight the DF o each property, i, beore the conditional beta model was calibrated.
RVi
wi = (10)
RVmin
where RVi is the replacement value or property i and RVmin is the lowest replacement value among all properties in the same building
typology. The more expensive buildings thereore get a higher weight in the regression than the low-cost properties. Having modifed
the max DF value (DFmax = 0.85) and included the weighting (Eq. (10)), the betareg( ) unction in the R package was used to estimate
the regression parameters o the beta model [31].
The model parameters or the 10 cases are given in Table 6.
The curves or the mean DF based on the vulnerability model or each dataset (2000 and 2008) and or each building typology are
shown on Fig. 7. The dots show the observed average DF in the same our PGA bins as given in Fig. 4. The PGA value (x-coordinate) or
each dot was ound by averaging PGA or all data points within the corresponding bin. The mean curves ft the aggregated data points
quite well but at same time it must be underlined that the distribution around the mean is quite wide.
The curves show a considerable dierence in the mean DF curves between the 2000 and 2008 dataset. Furthermore, the timber
buildings show better perormance than the RC-buildings. The outstanding perormance o the timber buildings can be explained by
very high prescribed wind loads in Iceland, (Vb0 = 36 m/s or the entire country [29]). Light-weight timber houses are thereore
11
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Table 7
Ratio o predicted mean DF to mean DF rom loss data (RDF), or the fve building typologies and the two datasets. Weighting based on Eq. (10) was used in the regression
and DFmax = 0.85.
Table 8
Ratio o predicted accumulated loss to actual accumulated loss (RLoss) or the fve building typologies and the two datasets. Weighting based on Eq. (10) was used in the
regression and DFmax = 0.85.
Table 9
Defnitions o threshold values or exceeding a given damage states in ESRM20 [19] and in this study.
Fig. 8. Fragility curves based on the statistical model or RC buildings (let) based on the 2000 dataset and (right) based on the 2008 dataset. Both curves or no-code &
low code (upper panel) and or moderate-code & high code (lower panel) are shown. high code buildings are shown. For comparison ragility unctions rom ESRM20 are
shown (CR_LWAL-DUL_H1 (upper panel) and CR_LWAL-DUM_H1 (lower panel)).
12
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Fig. 9. Fragility curves based on the statistical model or timber buildings (let) based on the 2000 dataset and (right) based on the 2008 dataset. Both curves or no-
code & low code (upper panel) and or moderate-code & high code (lower panel) are shown. For comparison ragility unctions rom ESRM20 are shown (W_LFM-
DUL_H1 (upper panel) and W_LFM-DUM_H1 (lower panel).
strongly built and well suited to withstand both wind and earthquake orces.
The ratio o predicted mean DF (using the statistical model) to the observed mean DF rom the loss data, RDF, are shown in Table 7
or all the 10 cases. Furthermore, the ratio o predicted loss to real estimated loss, RLoss, are shown in Table 8. The overall predicted loss
is 13.8% higher than the observed loss or the 2000 dataset and 4,8% higher or the 2008 dataset.
In summary, Fig. 7, and Tables 7 and 8 indicate that the models ft the data quite well and give air predictions.
13
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
Fig. 10. Comparison o vulnerability unctions in this study to with or (let) the 2000 dataset and (right) 2008 dataset or both RC buildings (upper panel) and timber
buildings (lower panel) to selected ESRM20 vulnerability unctions indicated in the legend.
and the comparison or RC buildings is shown in Fig. 8. As an example, the probability o exceeding DS1 and DS2 is much higher in this
study compared to ragility unctions rom ESRM20 at the lower PGAs. The dierence is more in the 2000 dataset than in the 2008
dataset. The ragility unctions and the comparison or timber buildings is shown in Fig. 9. Here, the ESRM20 ragility unctions are in
general higher than the local ragility curves. Again, the dierence is more or the 2008 dataset.
Similarly, i the vulnerability unctions rom this study (Fig. 7) are compared to the ESRM20 unctions the dierence is considerable
(Fig. 10). At low PGA the ESRM20 models are predicting much lower losses or RC buildings than the local models. In addition, the
shape o the curves is quite dierent. The dierence or the timber buildings is also large. The dierence can be partly explained by the
act that the global ESRM20 unctions are entirely based on analytical models where the losses are related to the deormations o the
structural elements whilst orm the loss data it can be seen that large proportion o the losses are related to non-structural losses,
especially in the 2008 dataset (Fig. 4) [8].
6. Conclusions
This paper compares structural losses in residential buildings in two strong seismic events in the South Iceland Seismic Zone. The
frst event is a sequence that consisted o a Mw 6.52 and a Mw 6.44 event occurring in 17th and June 21, 2000, respectively. The
second event is a Mw 6.30 earthquake occurring on the May 29, 2008. These events aected and damaged similar building typologies
in towns and arms in South Iceland. Complete loss datasets consisting o property registers data and loss data or insurance payments
exist or each o these events. Each dataset includes nearly 5000 residential buildings.
This study explores and compares the losses suered by residential buildings during these events rom dierent angles. To acilitate
vulnerability modelling and comparison with similar structures, the GEM taxonomy is used to classiy the aected buildings and in
addition the status o seismic codes (no-, low-, moderate- and high-code) is used to subclassiy them in the two loss datasets. A novel
statistical vulnerability models based on zero-inated beta regression [12] were calibrated rom the 2000 and the 2008 datasets. These
models are denoted as ICE2000VM and ICE2008VM.
One o the main fndings o the study is the dierences in loss patterns, and resulting vulnerability models, o buildings aected by
the June 2000 (Mw~6.5) and the May 2008 events (Mw6.3). The losses caused by the 2000 events (larger magnitude) is substantially
higher. The results show that two earthquakes that produce similar PGA can have very dierent impact on structures. An ideal in-
tensity measure should reect the eect o earthquake size, source-to-site distance, and other local eatures that might aect ground
shaking intensity and duration and control the structural damage. This study demonstrates the limitation o PGA as a ground motion
IM or the Icelandic building-by-building loss data which was obtained rom a housing stock o similar nature exposed to earthquakes
14
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
o dierent sizes. It is important to report the observed discrepancy in the loss pattern because it highlights the pitalls o combining
loss data rom dierent-sized earthquakes in vulnerability modelling with simple intensity measures such as PGA. This all points
towards a need or better ground motion intensity measures that can capture the event size eect better. At same time it could be useul
and inormative to use the available subclassifcation o the loss data (see section 2.4) to model separately the structural and non-
structural losses (and subclasses herein) rather than only working with one loss parameter (DF) that includes all kinds o losses.
This is a challenge or urther research.
The loss data and the calibrated vulnerability models showed consistently that moderate-code and high-code buildings showed
better perormance than no-code and low-code buildings, which is expected, but important to report. Timber/wood constructions
showed overall the best perormance and masonry buildings the poorest perormance. Nevertheless, no residential buildings collapsed
during these events.
Finally, there was a considerable dierence in the local vulnerability model developed in this study to global models (ESRM20
[19]) or similar building classes which points out the importance o applying local loss models when available.
Owing to the dierences in the two models, or seismic risk assessment in Iceland, it is recommended that ICE2000VM is used or
scenarios with Mw in the range 6.4–6.6 and the ICE2008VM or those with Mw range 6.2–6.4. These dierences also highlight the
uncertainties associated with extrapolating the models to earthquake o other size. Future studies on the loss data need to search or
ground motion intensity parameter (s) than can either by themselves or in a set explain the losses observed during these events in-
dependent o the size o the earthquakes.
Author statement
Bjarni Bessason: Methodology, Formal analysis, Data curation, Writing - Original drat preparation. Rajesh Rupakhety: Writing -
Reviewing and Editing, Jón Örvar Bjarnason: Writing - Reviewing and Editing.
Acknowledgements
The authors thank the Natural Catastrophe Insurance o Iceland or placing the earthquake damage database and other relevant
inormation at their disposal. This research was supported by the Icelandic Research Fund (grant no. 218149–051), and the University
o Iceland Research Fund.
References
[1] P. Einarsson, Plate boundaries, rits and transorms in Iceland, Jokull 58 (2008) 35–58.
[2] K. Jónasson, B. Bessason, Á. Helgadóttir, P. Einarsson, G.B. Gudmundsson, B. Brandsdóttir, K.S. Vogjörd, K. Jónsdóttir, A harmonised instrumental earthquake
catalogue or Iceland and the northern Mid-Atlantic Ridge, 2021, Nat. Hazards Earth Syst. Sci. 21 (2021) 2197–2214, https://doi.org/10.5194/nhess-21-2197-
2021.
[3] R. Sigbjörnsson, J.Th Snæbjörnsson, S.M. Higgins, B. Halldórsson, S. Ólasson, A note on the Mw6.3 earthquake in Iceland on 29 May 2008 at 15:45 UTC, Bull.
Earthq. Eng. 7 (1) (2009) 113–126, https://doi.org/10.1007/s10518-008-9087-0.
[4] Natural Catastrophe Insurance o Iceland, 2021. http://www.nti.is/en/. (Accessed 10 September 2021).
[5] Registers Iceland, 2021. http://www.skra.is/english. (Accessed 10 September 2021).
[6] T. Rossetto, D. D’Ayala, I. Ioannou, A. Meslem, Evaluation o existing ragility curves, in: K. Pitilakis, H. Crowley, A.M. Kaynia (Eds.), SYNER-G: Typology
Defnition and Fragility Curves or Physical Elements at Seismic Risk, Geotechnical, vol. 27, Geological and Earthquake Engineering, 2014, pp. 1–28, https://
doi.org/10.1007/978-94-007-7872-6.
[7] B. Bessason, J.Ö. Bjarnason, A. Gudmundsson, J. Sólnes, S. Steedman, Probabilistic earthquake damage curves or low-rise buildings based on feld data, Earthq.
Spectra 28 (4) (2012) 1353–1378, https://doi.org/10.1193/1.4000082.
[8] B. Bessason, J.Ö. Bjarnason, A. Gudmundsson, J. Sólnes, S. Steedman, Analysis o damage data o low-rise buildings subjected to a shallow Mw6.3 earthquake,
Soil Dynam. Earthq. Eng. 66 (2014) 89–101, https://doi.org/10.1016/j.soildyn.2014.06.025.
[9] B. Bessason, J.Ö. Bjarnason, Seismic vulnerability o low-rise residential buildings based on damage data rom three earthquakes (Mw6.5, 6.5 and 6.3), Eng.
Struct. 111 (2016) 64–79, https://doi.org/10.1016/j.engstruct.2015.12.008.
[10] I. Ioannou, B. Bessason, I. Kosmidis, J.Ö. Bjarnason, T. Rossetto, Empirical seismic vulnerability assessment o Icelandic buildings aected by the 2000 sequence
o earthquakes, Bull. Earthq. Eng. 16 (12) (2018) 5875–5903, https://doi.org/10.1007/s10518-018-0413-x.
[11] S.L.P. Ferrari, F. Cribari-Neto, Beta regression or modelling rates and proportions, 2004, J. Appl. Stat. 31 (7) (2004) 799–815, https://doi.org/10.1080/
0266476042000214501.
[12] B. Bessason, J.Ö. Bjarnason, R. Rupakhety, Statistical modelling o seismic vulnerability o RC, timber and masonry buildings rom complete empirical loss data,
Eng. Struct. 209 (2020), 109969, https://doi.org/10.1016/j.engstruct.2019.109969.
[13] R. Ospina R, S.L.P. Ferrari, A general class o zero-or-one inated beta regression models, Comput. Stat. Data Anal. 56 (2012) 1609–1623, https://doi.org/
10.1016/j.csda.2011.10.005.
[14] H. Crowley, V. Despotaki, V. Silva, J. Dabbeek, X. Romão, N. Pereira, J.M. Castro, J. Daniell, E. Veliu, H. Bilgin, C. Adam, M. Deyanova, N. Ademovic, J. Atalic,
A. Karatzetzou, B. Bessason, V. Shendova, A. Tiganescu, D. Toma-Danial, Z. Zugic, S. Akkar, U. Hancilar, Model o seismic design lateral orce levels or the
existing reinorced concrete European building stock, Bull. Earthq. Eng. 19 (2021) 2839–2865, https://doi.org/10.1007/s10518-021-01083-3.
[15] S. Brzev, C. Scawthorn, A.W. Charleson, L. Allen, M. Greene, K. Jaiswal, V. Silva, GEM building taxonomy version 2.0, GEM Technical Report 2013-02 (2013).
[16] H. Crowley, D. Rodrigues, V. Silva, V. Despotaki, L. Marins, X. Romão, J.M. Castro, N. Pereira N, A. Pomonis, A. Lemoine, A. Roullé, B. Tourlière, G. Weatherill,
K. Pitilakis, L. Danciu, A.A. Correira, S. Akkar, U. Hancilar, P. Covi, The European Seismic Risk Model 2020 (ESRM20), 2nd International Conerence on Natural
Hazards and Inrastructure, ICONHIC, 2019.
[17] V. Silva, D. Amo-Oduro, A. Calderon, C. Costa, J. Dabbeek, V. Despotaki, L. Martins, M. Pagani, A. Rao, M. Simionato, D. Viganò, Development o a global
seismic risk model, D, Earthq. Spectra 36 (1 suppl) (2020) 372–394, https://doi.org/10.1177/8755293019899953, 2020.
15
B. Bessason et al. Journal of Building Engineering 46 (2022) 103806
[18] S. Molina, D.H. Lang, C. Lindholm, Selena - an open-source tool or seismic risk and loss assessment using a logic tree computation procedure, Comput. Geosci.
36 (2010) 257–269, https://doi.org/10.1016/j.cageo.2009.07.006.
[19] L. Martins, V. Silva, Development o a ragility and vulnerability model or global seismic risk analyses, Bull. Earthq. Eng. (2020), https://doi.org/10.1007/
s10518-020-00885-1.
[20] Ó. Thórarinsson, B. Bessason, J.Th Snæbjörnsson, S. Ólasson, R. Sigbjörnsson, G. Baldvinsson, The South Iceland earthquakes 2000: strong motion
measurements, in: Proceedings o the 12th European Conerence on Earthquake Engineering. Paper No. 321, Paris, France, 2002.
[21] B. Halldórsson, R. Sigbjörnsson, The Mw6.3 Ölus earthquake at 15:45 UTC on 29 May 2008 in South Iceland: ICEARRAY strong-motion recordings, Soil Dynam.
Earthq. Eng. 29 (2009) 1073–1083, https://doi.org/10.1016/j.soildyn.2008.12.006.
[22] R. Rupakhety, R. Sigbjörnsson, Ground-motion prediction equations (GMPEs) or inelastic response and structural behaviour actors, Bull. Earthq. Eng. 7 (3)
(2009) 637–659, https://doi.org/10.1007/s10518-009-9105-x.
[23] S. Ólasson, Attenuation o earthquake waves, in: J. Sólnes, F. Sigmundsson, B. Bessason (Eds.), Natural Hazard in Iceland, Volcanic Eruptions and Earthquakes,
University o Iceland Press and Natural Catastrophe Insurance o Iceland, 2013 (in Icelandic).
[24] E. Tryggvason, S. Thoroddsen, S. Thórarinsson, Report on earthquake risk in Iceland, Icelandic J. Eng. Tímarit Verkrædingaélags Íslands. 43 (1958) 81–97.
[25] ÍST 13, Earthquakes, Loads and Design Rules, Icelandic Industrial Development Agency/Idnthróunarstoun Íslands. Reykjavík, Iceland, 1976.
[26] ÍST 13, Earthquakes, Loads and Design Rules, Icelandic Industrial Development Agency/Idnthróunarstoun Íslands. Reykjavík, Iceland, 1989.
[27] CEN Eurocode 8, Design o Structures or Earthquake Resistance, European Standard, European Committee or Standardisation, Brussels, 2004.
[28] SI, National, Application Documents (NAD) or Iceland, Standards Council o Iceland/Stadlarád Íslands (SI), 2002.
[29] SI, Icelandic National Annexes to EUROCODES, Standards Council o Iceland/Stadlarád Íslands (SI), 2010.
[30] R. Rupakhety, R. Sigbjörnsson, S. Ólasson, Damage to residential buildings in Hveragerdi during the 2008 Ölus Earthquake: simulated and surveyed results,
Bull. Earthq. Eng. 14 (7) (2016) 1945–1955, https://doi.org/10.1007/s10518-015-9783-5.
[31] R Core Team, R: A Language and Environment or Statistical Computing, R Foundation or Statistical Computing, Vienna, Austria, 2020. https://www.R-project.
org/.
16