Bull Eng Geol Environ
DOI 10.1007/s10064-014-0570-3
ORIGINAL PAPER
An expert judgement approach to determining the physical
vulnerability of roads to debris flow
M. G. Winter • J. T. Smith • S. Fotopoulou •
K. Pitilakis • O. Mavrouli • J. Corominas •
S. Argyroudis
Received: 30 May 2013 / Accepted: 3 January 2014
 Springer-Verlag Berlin Heidelberg 2014
Abstract The physical vulnerability of roads to debris flow
may be expressed through fragility functions that relate flow
volume to damage probabilities. Fragility relationships are
essential components of quantitative risk assessments as they
allow for the estimation of risk within a consequence-based
framework. To the best of the authors’ knowledge, this is the
first time that fragility curves have been produced in order to
provide the conditional probability for a road to be in, or to
exceed, a certain damage state for a given debris flow volume.
Preliminary assessments were undertaken by means of a
detailed questionnaire. A total of 47 returns were received
from experts in 17 countries: 32 % academia, 51 % the
commercial sector and 17 % governments. Fragility curves
have been defined for three damage states (limited damage,
serious damage and destroyed) for each of low-speed and
high-speed roads in order to cover the typical characteristics
of roads vulnerable to debris flow. The probability of any
given damage state being reached or exceeded by a debris
flow of a given volume (10–100,000 m3) was derived from
M. G. Winter (&)
Transport Research Laboratory (TRL), 13 Swanston Steading,
109 Swanston Road, Edinburgh EH10 7DS, UK
e-mail: mwinter@trl.co.uk
J. T. Smith
Golder Associates (formerly TRL), Cavendish House, Bourne
End Business Park, Cores End Road, Bourne End SL8 8AS, UK
S. Fotopoulou  K. Pitilakis  S. Argyroudis
Research Unit of Geotechnical Earthquake Engineering and Soil
Dynamics, Department of Civil Engineering, Aristotle
University of Thessaloniki, 54124 Thessalonı́ki, Greece
O. Mavrouli  J. Corominas
Department of Geotechnical Engineering and Geosciences,
Technical University of Catalonia, 08034 Barcelona, Spain
the mean of the responses received. Inevitably there was a
degree of scatter in the results, and the treatment of such
variation, or ‘experimental errors’, was crucial to under-
standing the data and developing the fragility curves. Fra-
gility curves are quantitative expressions of vulnerability.
The method adopted is based upon qualitative, expert judg-
ment of quantitative probabilities. In addition to an assess-
ment of the probabilities of given damage states being
exceeded, respondents to the questionnaire were polled as to
their level of experience and confidence in their ability to
provide a valid and coherent set of answers to the questions
posed. The development of the fragility curves and their
validation are described in the paper.
Keywords Landslide  Debris flow  Road  Hazard 
Vulnerability  Risk
Introduction
Fragility curves are a graphical means of describing the
physical vulnerability of elements at risk to a given hazard.
They give the conditional probability of a particular ele-
ment at risk to be in, or to exceed, a certain damage state as
a result of a hazard of a particular type or intensity
(Mavrouli and Corominas 2010a, b). Fragility relationships
are essential components of quantitative risk assessments
(QRA), as they allow for the estimation of risk within a
consequence-based framework.
The losses, or economic impacts, resulting from debris
flow activity take a variety of forms and include (Winter
and Bromhead 2012) those that are:
• Direct such as damage to the road and associated
structures and the consequential need for repair.
123

• Direct consequential such as the effects of delays and
detours. The costs to society of fatal and non-fatal
injuries would also be included in this category.
• Indirect consequential such as reduced levels of
business activity as access to an area is restricted.
When considering dispersed consequential economic
impacts, the concept of vulnerability shadow (Winter and
Bromhead 2012) becomes important for understanding the
consequences of an event. In such cases, an event that closes a
road at a critical location on the network has a relatively small
footprint, but the vulnerabilities that result may be experi-
enced across a much wider geographical area. This wider
geographical area is referred to as the vulnerability shadow.
Extensive vulnerability shadows are particularly likely to
occur on rural networks, including routes in Scotland (some of
which are used as examples later in this paper), and the range
of losses illustrates the need for a robust risk assessment, of
which the proper assessment of vulnerability forms an
essential part. Fragility relationships are widely adopted in
seismic ‘expected loss’ and risk assessments, being a valuable
tool to explicitly assess the vulnerability of structures to
earthquake hazard (Pitilakis et al. 2006). For the purposes of
this work, debris flow is defined as classified by Cruden and
Varnes (1996), and broadly falls into the initial slurry flow and
viscous slurry flow, viscous granular flow categories defined
by Pierson and Costa (1987). The sediment–water flows are
defined as plastic, with movement occurring over a wide range
of potential velocities. These features are broadly character-
istic of the debris flow types experienced in, for example,
Scotland in recent years (Winter et al. 2005).
The estimation of the vulnerability of road infrastructure to
landslide, let alone specifically rainfall-induced debris flow,
seems to be relatively poorly served by the available litera-
ture. A generic approach is often applied to different vul-
nerable elements’ typologies, and/or simple damage criteria
such as the percentage of affected road are considered
(Wieczorek et al. 2004; Galli and Guzzetti 2007; Geertsema
et al. 2009; Mansour et al. 2011). Unsurprisingly, there is a
greater focus upon human as opposed to infrastructural los-
ses, and typically traffic volume—with or without other
factors—is taken as a proxy for such vulnerabilities
(McMillan and Matheson 1997; Winter et al. 2009).
For the purposes of this work, the element at risk is a
road and the hazard is debris flow. Damage probabilities
have been assigned for specific debris flow volumes. Fra-
gility curves have been produced that indicate the proba-
bility of a debris flow of a given volume exceeding each of
three damage states. Typically, damage to roads resulting
from debris flow may include one or more of the following:
• Debris covering the carriageway, preventing vehicle
movements.
• Damage to the carriageway surfacing materials.
123
M. G. Winter et al.
• Blockages and other types of damage to the drainage
system.
• Damage to vehicle restraint systems.
• Damage to support structures including slopes and
retaining walls downhill from the road.
The vulnerability to debris flow for impacted buildings has
been expressed using fragility curves and/or probabilities
of exceedance of damage states (Haugen and Kaynia 2008;
Jakob et al. 2012; Quan Luna et al. 2011; Papathoma-
Khöle et al. 2012). However, to the best of the authors’
knowledge, this is the first time that fragility curves have
been developed for the effects of debris flow on roads.
While several possible approaches were available for the
development of fragility curves, including analytical
approaches, it was decided that expert engineering judgement
should be used due to the lack of a comprehensive empirical
dataset as well as the complex nature of the problem.
This paper describes the development of the question-
naire that was sent to experts globally in order to collect
data for the development of fragility curves. It then goes on
to describe the analysis of the data collected, the inter-
pretation of those data, and finally its validation using real-
world examples. The paper forms part of a larger piece of
work undertaken as part of the SafeLand project (http://
www.safeland-fp7.eu/).
Methodology
Road characterisation
Many different classifications of roads could potentially be
considered, covering numerous key factors, such as con-
struction type, stiffness, and traffic speed. However, in
order to reduce the questionnaire to a reasonable size, some
simplification was needed.
Primarily, it was decided that, for the purposes of this
exercise, all roads could be considered to be relatively stiff
and brittle (the low strain stiffness of even an unbound
pavement, for example, may be typically up to around one
gigapascal) in comparison to most debris materials. In
order to further simplify the analysis, roads were divided
into low- and high-speed roads, characterized as follows:
• High-speed roads speed limit between 80 and 110 km/h
and at least one running lane in each direction, very
often in conjunction with a hard strip or hard shoulder.
• Local (or low-speed) roads speed limit typically
\50 km/h on a single-carriageway (one lane for each
traffic direction) or single track. This category is
intended to encompass both paved (bituminous, unre-
inforced or reinforced concrete) and unpaved
constructions.
Determining physical vulnerability of roads to debris flow

Clearly there is a gap between the speed limits of the two
classes of road, reflecting the transition between local roads
and high-speed roads, which is by no means geographically
consistent. This reflects reality—in some countries and
regions certain road geometries are more closely aligned
with the definition of local roads, and in others they are
more closely aligned with the definition of high-speed
roads. Speed limit is not, and should not be, the only
determinant of the category of road. In most instances, the
category that a particular geometry and speed limit
combination will belong to is relatively self-evident, but
extending the speed limits between 50 and 80 km/h in the
category descriptions could lead to uncertainty and
potential incorrect categorisation.
Damage states
Representative damage states associated with the conse-
quences of a debris flow of a given volume intersecting a
road were defined. The damage states considered in the
questionnaire are defined in Table 1. The damage states
range from ‘limited damage’, which is unlikely to signifi-
cantly affect the passage of vehicles, through ‘serious
damage’ that completely blocks the carriageway, to
‘destroyed’, involving complete blockage and damage to
the road itself that for, high-speed roads at least, will
almost certainly need to be repaired prior to reopening to
traffic without restrictions on speed.
Probability of exceedance
combination of damage states (limited damage, serious
damage and destroyed) and landslide volumes (\10, 10, 100,
1,000, 10,000, 100,000 m3). (Note that the codes A–G in
Table 2 were used to avoid confusion between the visually
rather similar probabilities with which they correspond and
which were used to construct the fragility curves.) It was
particularly noted in the questionnaire that the qualitative
descriptors ‘highly improbable’ and ‘extremely likely’
should be used with caution, and only where an extensive,
high-quality dataset supports the classification.
Respondents were asked to consider a segment of road
500 m long. The questionnaire specifically noted that the
volume ranges refer to volumes deposited at the level of the
road. However, this could perhaps have been clearer as to
whether this implied the total volume remaining on the
road immediately post-event, or the volume that passes
through the level of the road. In the former instance, for
volumes of 10,000 and 100,000 m3, it would be impossible
not to have reached the destroyed damage state, since the
whole road would be covered. The fact that not all of the
questionnaire responses indicate a probability of unity for
these volumes does imply that this may not have been fully
understood by all respondents, and goes some way to
explaining the lower probabilities for high volumes as
discussed under ‘‘Interpretation’’.
Experience of the respondents
Hazard perception varies from person to person, and
clearly the experience of the respondents is a critical metric

Respondents to the survey were asked to use their expert Table 2 Description of probabilities
judgement to assess the probability of each damage state
Qualitative Description Code Values
(Table 1) being exceeded. This was achieved using the
descriptor for
qualitative descriptors and descriptions of the quantitative analysis
probabilities of damage state exceedance (Table 2) for a
range of volumes deposited at the level of the road for each Highly Damage state almost certainly not A 0.000001
improbable exceeded, but cannot be ruled
out
Table 1 Damage state definition
Improbable Damage state only exceeded in B 0.00001
Damage High-speed roads Local (low-speed) roads (remote) exceptional circumstances
state Very Damage state will only be C 0.0001
unlikely exceeded in very unusual
P1 (limited Encroachment limited to Partial blockage of
circumstances
damage) verge/hard strip carriageway
Unlikely Damage state may be exceeded, D 0.001
P2 (serious Blockage of hard strip Complete blockage of
but would not be expected to
damage) and one running lane carriageway and/or
occur under normal
damage to ancillaries
circumstances
P3 Complete blockage of Complete blockage of
Likely Damage state expected to be E 0.01
(destroyed) carriageway and/or carriageway and/or
exceeded under normal
repairable damage to damage to surfacing. For
circumstances
surfacing unpaved roads the
surfacing may remain Very likely Damage state expected to be F 0.1
damaged but passable at exceeded
reduced speeds post Extremely Damage state is almost certainly G 1.0
clean-up likely exceeded

123

in terms of understanding, evaluating, analysing and
interpreting the data resulting from the questionnaire.
Having been pre-selected by the authors as individuals with
known expertise in this area, the respondents were asked to
assess the levels of both their experience of debris flow and
their confidence in their responses for each debris flow
volume assessed on a scale of 0 (no experience) to 10
(extensive experience).
The scores for the experience-based self-assessment
were weighted towards the higher end of the range, as
might be expected from a sample of respondents who were
selected for their known expertise in this area. It thus seems
potentially appropriate to place a greater confidence in the
responses received from those who reported that their level
of experience was higher than the average, and a number of
approaches is possible. The scores for the level of confi-
dence in the responses for each volume of debris flow
assessed are discussed in the ‘‘Interpretation’’ section.
Data collection
In total, the questionnaire was sent to 176 experts by the
three organisations involved in this work. Of those, 47
responded by completing the questionnaire, giving a
response rate of 27 %. Responses were received from 17
countries: UK (34 %), Greece (23 %), other European
countries (26 %), Asia (4 %), Australasia (4 %), North
America (4 %), and the Middle East (2 %). The back-
ground of the respondents spanned academia (32 %), the
commercial sector (51 %) and government bodies (17 %).
As with all data, a reality (or ‘sense’) check is advis-
able. Figure 1 illustrates the way in which one might
anticipate that the basic probability data should plot. For
smaller landslide volumes, the majority of the responses
should plot to the left hand side, while for larger volumes,
the majority of the data should plot to the right hand side.
For the intermediate volumes, the majority of the data
should plot somewhere between these two extremes. The
precise shape of the curve will, of course, vary between
these conceptual ‘norms’. Although not illustrated in
Fig. 1, it would also be reasonable to anticipate that for
limited damage, the data would be skewed to the right, as
more events would be expected to induce such a damage
state. Similarly, for the destroyed category, it might be
expected that the data would be skewed to the left, as
fewer events would be likely to induce this extreme
damage state.
The data from the questionnaire responses (Fig. 2) show
that, in broad terms, the results are reasonable. While there
is a greater or lesser degree of scatter and/or variation from
the norm set out in Fig. 1, most of the data broadly follow
the pattern illustrated.
123
M. G. Winter et al.
Fig. 1 An idealized plot of the probability of exceeding a given
damage state against the frequency of response, illustrating how the
data might be expected to appear and to thus provide an immediate
‘sense’ check
Analysis
Preliminary fragility curves
It is a relatively straightforward matter to construct pre-
liminary fragility curves from the average of the proba-
bility responses at each volume, at each damage state and
for both high speed roads and local roads, as illustrated in
Fig. 3. These curves have the basic attributes of typical
fragility curves. The curves generally show that landslides
of a given volume are associated with higher probabilities
of exceeding a certain damage state when they affect local
roads than when they affect high-speed roads, as would be
expected. In addition, the curves for high-speed roads
generally show little effect at small landslide volumes,
below a few 100 cubic metres.
It is noticeable that the mean probabilities do not reach
unity for any of the curves. This implies that the damage
states as defined can never be reached or exceeded with
complete certainty. However, this is an inevitable function
of using the average of the responses, as the maximum
possible response coincides with the desired termination
point of each curve (a volume at which exceeding the given
damage state is inevitable and the probability is unity).
Further stages of analysis were undertaken to interrogate
and better understand the data.
Curve fitting
Standard Microsoft ExcelTM curve-fitting was applied to the
basic curves (third-order polynomial for limited damage
states, and fourth-order polynomial for serious damage and
destroyed states). This gave better separation of the individ-
ual curves than is shown in Fig. 3, and there was greater
contrast between the results for high-speed and local roads
(Pitilakis and Fotopoulou 2011). This technique allows a
Determining physical vulnerability of roads to debris flow

Fig. 2 Probability of exceeding each damage state on each type of road plotted against the number of responses for each landslide volume: left
local roads, right high-speed roads, top limited damage state, middle serious damage state, bottom destroyed damage state

degree of extrapolation of the data to higher volumes/prob-
abilities. However, the over-application of such extrapolation
tends to distort the curves at lower volumes by increasing the
divergence of the curve from the data. Such a technique
proved unsuitable for extending the curves by other than
small amounts of the volume considered (certainly consid-
erably less than an order of magnitude). Other curve-fitting
methods could, of course, have been used, including two-
parameter cumulative log-normal distributions.
This has been achieved by visually judging the appropriate
value of probability at 1,000,000 m3 in order to maintain
the broad trend of the curves (Fig. 4)—generally, in this
high-volume part of the curve, the probability is increasing
with increasing volume but at a decreasing rate. It is
noticeable that, even when the volume is increased to
1,000,000 m3 in this way, none of the fragility curves reach
unity. Only for limited damage of local roads is a value of
around 0.95 achieved at the highest volume.

Manual extrapolation Weighting the data

It is thus interesting to manually extrapolate the data pre- In order to incorporate the experience of the respondents as
sented in Fig. 3 by a further logarithmic cycle of volume. a measure of the reliability of the data, a weighting

123
 M. G. Winter et al.

approach may be taken. However, care is needed to ensure

that the sample is weighted rather than the individual
responses; otherwise bias will be introduced into the
results. (Weighting the individual responses will, depend-
ing upon the precise approach taken, either increase or
decrease the individual probabilities contained within the
questionnaires for those with higher expertise and the
converse for those with lower levels of expertise. There is
no logical justification for such a change and this approach
has therefore been avoided.) Weighting the sample may be
done as follows:
Pn
p i Ei
p ¼ Pi¼1 n ; ð1Þ
i¼1 Ei

where p is the weighted mean probability of a particular

damage state being exceeded, pi is the individual responses
of the probability of a particular damage state being
exceeded, Ei is the individual responses in terms of self-
assessed experience, and n is the number of responses.
However, the question of what a weighted average
means remains, and statistical advice (Sexton, personal
communication) indicates that the results should be treated
with a degree of caution. This approach yields fragility
curves (Fig. 5) with lower probabilities of given damage
states being exceeded for a given volume of event com-
Fig. 3 Preliminary fragility curves: top local roads, bottom high-
speed roads
pared to those derived from the unweighted data set
(Fig. 3). This potentially indicates that either those who
perceive their expertise as greater take a less conservative
approach to assessing the likely damage incurred for a
given event size, or that those who are less experienced
overestimate the potential damage.
The second approach involves rejecting the data from
those respondents reporting less experience, leaving only
the data from those who assessed themselves as more
experienced in this area. Statistical advice (Sexton, per-
sonal communication) indicates that (approximately) only
scores from the upper 25 % of the available range should
be examined. This implies that the analysis should be
undertaken for those judging their experience level as eight
or above (33 % of respondents). However, plotting the data
led to curves that were rather difficult to interpret and to the
conclusion that those 16 respondents (or 33 %) were
insufficient in number to present a coherent picture (Piti-
lakis and Fotopoulou 2011).

Interpretation

The data points that make up the fragility curves described

above represent the mean responses of the entire sample at
Fig. 4 Extrapolated fragility curves: top local roads, bottom high- each event volume. As such, it is both necessary and
speed roads desirable to examine the statistical variation in the data.

123
Determining physical vulnerability of roads to debris flow
Fig. 5 Weighted fragility curves: top local roads, bottom high-speed
roads
In order to do this, the standard deviation may be
assessed and the curves representing the mean plus and
minus one standard deviation plotted alongside the mean
curve; these represent the bounds within which 68 % of the
data fall (Fig. 6). Clearly the results cover a wide range and
reach into the zone corresponding to negative probability
and probabilities greater than unity in some instances.
While both states are ‘impossible’, this reflects the genuine
statistical variation in the data.
The lines at zero and unity are important as probability
states outside these bounds are purely of a statistical nature
and have no meaning in terms of the relative likelihood of a
damage state being reached. Where the curves extend
beyond these conceptual limits, it is clear that there is a
wide range within the reported data.
In broad terms, the standard deviation of the probability of
a given damage state being reached or exceeded increases
with the volume of the event considered. This is not altogether
surprising, as larger events occur less frequently, and thus the
collective experience of the respondents may be expected to
be both less and more variable; both of these factors have the
potential to increase the range of responses. This increase in
the variation in responses seems to be more marked as the
damage level increases, seemingly indicating a high level of
confidence on the part of the respondents to questions relating
to the probability of low-volume events creating high damage
states. Again, this result conforms to what might reasonably
be expected.
The curves illustrated in Fig. 3 do not stretch between
zero and unity. Even when manually extrapolating to a
landslide volume of 1,000,000 m3 (Fig. 4), the curves do
not reach to unity, as would be expected if they had been
derived from modelling in which a constraint would have
been placed upon the process in order to ensure such a
result.
As previously mentioned, using the current approach, it
is inevitable that the mean probability of each damage state
being reached or exceeded is less than unity unless all of
the respondents return such a value. This seems more than
a little unlikely (see also the ‘‘Analysis’’ section). This
facet is also likely compounded by the fact that, in the
questionnaire, the respondents were encouraged to exercise
caution and to use their direct experience to support their
use of the lowest and highest qualitative descriptors of the
likelihood of given damage states (i.e. highly improbable
and extremely likely), and thus discouraging their use. In
addition, as articulated in the description of the method-
ology, the questionnaire could perhaps have been clearer
on what was meant by the volume of a debris flow, and this
may have led to some confusion amongst respondents and
inconsistency of response. The questionnaire also asked
respondents to consider a length of road 500 m long, and
this may have led to some debate as to when the different
damage states are reached (i.e. when affecting the full
500 m length or only a fraction thereof). It is clear that an
event of a few thousand cubic metres can seriously damage
or destroy the road while only affecting a fraction of the
500 m length. How this was perceived and rated in terms of
road vulnerability by the respondents remains uncertain.
This then begs the question of how to account for such
an inevitable, and seemingly contradictory, facet of the
results. If a constant ratio is assumed between the different
probabilities of exceeding a damage state at any given
volume, it is a simple matter to ‘force’ (or stretch) the
fragility curves to reach unity at the greatest value of
landslide volume. This can be achieved by adopting a
straightforward ratio approach (the forced probability at
any value of landslide volume pif = pi[1/pn], where pi is
the mean probability and pn is the mean probability at the
maximum value of the landslide volume).
In order to determine whether such an approach can be
justified, one must examine the more detailed responses of
the respondents to the questionnaire, and in particular the
responses of those where a probability of unity was
assigned to the combinations of landslide volume and
damage state. These data illustrate, as might be anticipated,
123
 M. G. Winter et al.

Fig. 6 Fragility curves from Fig. 3 showing the mean and the mean plus and minus one standard deviation: left local roads, right high-speed
roads, top limited damage state, middle serious damage state, bottom destroyed damage state

that the number of respondents assigning a probability of
unity increases markedly with landslide volume while
decreasing with increased damage state severity. Most
importantly, by the time high landslide volumes are con-
sidered, the majority of respondents return a value of unity
for the likelihood of a given damage state being reached or
exceeded. This lends considerable justification to the
approach of ‘forcing’ the curves to reach unity.
As discussed, the preliminary fragility curves of Fig. 3
can be forced to unity at the highest value of landslide
volume. As these have already been manually extrapolated
to the next order of magnitude in terms of landslide volume
(i.e. 1,000,000 m3) (Fig. 4), the next logical step is to
combine these two actions, as illustrated in Fig. 7. The
curves illustrated therein conform to the ‘s’ shape generally
perceived as being the correct form for fragility curves.
Notwithstanding this, one would normally expect that the
curves for different damage states would reach unity at
different landslide volumes; that they do not is a function
of the type of analysis undertaken, and it seems reason-
able—as none of the curves reach unity—to force all to
such a level at the highest volume considered.
Clearly it is possible to make alternative assumptions in
relation to the presentation of the fragility curves; for exam-
ple, forcing the curve for limited damage to unity (at the
greatest landslide volume), allowing that for destroyed to

123
Determining physical vulnerability of roads to debris flow
Fig. 7 Fragility curves ‘forced’ to unity and manually extrapolated to
the next order of magnitude for volume: top local roads, bottom high-
speed roads
remain as derived from the data (Fig. 4), and forcing the
curve for serious damage to a level midway between the two.
It does, however, seem that there is no greater or lesser jus-
tification for either approach, except inasmuch as it might be
anticipated that the destroyed state should always be reached
in response to a landslide volume of 1,000,000 m3, and thus
the approach taken in Fig. 7 appears reasonable.
Geographical variations and variations potentially
caused by respondents’ backgrounds (academic, the com-
mercial sector, and governments) were investigated (Piti-
lakis and Fotopoulou 2011). It was concluded that the
datasets were generally too small to draw definitive con-
clusions, albeit that the data appeared to suggest that:
• Responses from the UK exhibited slightly higher
probabilities for larger landslide volumes compared to
those for the rest of the world.
• Similarly, the response for academia exhibited slightly
higher probabilities for larger landslide volumes com-
pared to those for the commercial sector.
Not withstanding these very tentative observations, no
major difference in the pattern of responses was observed
for the different groups.
As noted in the ‘‘Analysis’’ section, the experience of
the respondents is a critical metric in terms of this data set,
and respondents were asked to assess their experience on a
scale of 0–10. The scores of this self-assessment were
weighted towards the higher end of the range, as might be
expected from a sample of respondents who were selected
for their known expertise in this area, and the data are
plotted in Fig. 8. As might be expected, the confidence of
respondents for low-volume events is markedly higher than
that for high-volume events, reflecting the higher frequency
of low-volume events. Indeed, examination of the indi-
vidual charts for each landslide volume considered (Fig. 8)
indicates that:
• For low-volume events, the distribution of the
responses is skewed to the right (high confidence).
• For medium-volume events (1,000–10,000 m3), the
distribution becomes broadly normal.
• For high-volume events (100,000 m3), the distribution
shows signs of beginning to skew to the left (lower
confidence).
Use of the fragility curves
In order to assess the validity of the curves developed, a
view must be taken on the most appropriate set of curves to
consider. In light of the comments received from ques-
tionnaire respondents, it was decided to use curves which
have been extrapolated to a debris flow volume of
1,000,000 m3 and which have been forced to unity as
shown in Fig. 7. Events from Scotland in the UK and the
Republic of Korea are considered here.
Figure 9 illustrates hypothetically shaped curves in
which the numbers given relate to a 5,000 m3 event as
applied to a high-speed road in Fig. 7. The probabilities of
the damage being equal to or greater than a given level are
as follows:
• Damage greater than or equal to limited, p = 0.6.
• Damage greater than or equal to serious, p = 0.4.
• Damage greater than or equal to destroyed, p = 0.3.
The discrete, or conditional, damage state probabilities (i.e.
the probabilities of the occurrence of a given damage state)
are then estimated from those probabilities of a given
damage state being exceeded as follows:
• Probability of no damage = 1.0–0.6 = 0.4.
• Probability of limited damage = 0.6–0.4 = 0.2.
• Probability of serious damage = 0.4–0.3 = 0.1.
• Probability of destroyed damage state = 0.3.
These should, of course, always total unity.
123

Fig. 8 Self-assessed confidence ratings relative to given event volumes: top left \10 m3, middle left 10 m3, bottom left 100 m3, top right
1,000 m3, middle right 10,000 m3, bottom right 100,000 m3
Note that the conditional probability of the destroyed
damage state is always equal to the probability of that state
being exceeded. Vulnerability assessment using fragility
curves is, of course, probabilistic in nature and the models
used in their construction—in this case a model based
purely upon expert judgment—have inherent uncertainties.
In this context, the validation examples are not expected to
precisely predict the observed damages.
A85 Glen Ogle, Scotland
In August 2004, two debris flow events occurred at Glen
Ogle (Fig. 10). These blocked the A85 strategic road,
123
M. G. Winter et al.
culverts and other drainage features, and necessitated a full
repair to the road pavement, safety barriers and parapets
(Winter et al. 2005, 2006). Some 20 vehicles were trapped
by the events and 57 people were airlifted to safety; one
vehicle was swept away in the latter stages of the event
(Winter et al. 2005). The smaller southerly and larger
northerly events were estimated to have deposited around
3,200 and 8,500 m3 in their respective debris lobes, having
been triggered by smaller translational slides of around 285
and 280 m3 (Milne et al. 2009). The depositional figures
are believed not to include material deposited on the road,
and it seems reasonable therefore to round these figures up
to around 5,000 and 10,000 m3. This perhaps also
Determining physical vulnerability of roads to debris flow
Fig. 9 Hypothetical fragility curve in which the numbers given relate
to a 5,000 m3 event as applied to a high-speed road (see Fig. 7),
illustrating the probabilities of a given event exceeding a given
damage state and also the conditional probabilities for a given level of
damage (not to scale)
Fig. 10 The larger northerly event of the two August 2004 debris
flows at A85 Glen Ogle
illustrates the uncertainty when dealing with debris flow
volumes between the amount that is mobilised and the
amount that is deposited at around road level.
Figure 11 shows how these event volumes plot on the
fragility curves originally derived in Fig. 7. For the smaller
(5,000 m3) event, the conditional probabilities for no
damage, limited, serious and destroyed damage states are
0.4, 0.2 (0.6), 0.1 (0.4) and 0.3 (0.3) (the probabilities of
the damage states being met or exceeded are given in
Fig. 11 The fragility curves of Fig. 7 with vertical lines added at
200, 500, 1,000, 5,000 and 10,000 m3. These lines are drawn
horizontally across to the probability axis. In order to retain a degree
of clarity, only horizontal lines where the limited damage curve is
intersected are shown: top local roads, bottom high-speed roads
brackets); for the larger (10,000 m3) event, the conditional
probabilities are around 0.3, 0.15 (0.7), 0.15 (0.55) and 0.4
(0.4), in both cases for a high-speed road. Certainly the
damage caused by the larger event would have been
described as destroyed using the scheme considered here,
and the probability of this state, 0.4, seems to be broadly in
line with observations made in its immediate aftermath,
albeit affecting a road length of around 200 m. Similarly,
the damage caused by the smaller event, although signifi-
cantly less in terms of physical damage to the infrastruc-
ture, would also be classified as destroyed, and this seems
to be broadly reflective of the probability of 0.3 returned
from Fig. 11 (Table 3).
A83 Rest and be Thankful, Scotland
Events at the Rest and be Thankful on the A83 strategic
road (Winter et al. 2009, 2010) are frequent and have
certainly occurred on at least an annual basis over the
preceding 20–25 years. The event magnitude is, however,
rather small, generally ranging between 200 and 1,000 m3.
Following an event in October 2007, the volume of
material deposited at road level was estimated by the road
123
 M. G. Winter et al.

Table 3 Probabilities of damage states being exceeded and condi-

tional probabilities for high-speed roads for given event volumes
Event volume (m3)
200 m3 1,000 m3 5,000 m3 10,000 m3

Probability of damage state being met or exceeded

Limited 0.2 0.3 0.6 0.7
Serious 0.05 0.2 0.4 0.55
Destroyed 0.01 0.02 0.3 0.4
Conditional probability
No damage 0.8 0.7 0.4 0.3
Limited 0.15 0.1 0.2 0.15
Serious 0.4 0.18 0.1 0.15
Destroyed 0.01 0.02 0.3 0.4 Fig. 12 General view of the main hillside at the A83 Rest and be
Thankful, including the October 2007 event

operating company to be around 400 m3. On-site obser-

vations indicate that it is reasonable to use a figure of
around 1,000 m3 for the total amount of material mobi-
lised. Figure 12 illustrates the main hillside at the Rest and
be Thankful.
For smaller events (200 m3 in Fig. 11), the conditional
probabilities of the damage states no damage, limited,
serious, and destroyed are 0.8, 0.15 (0.2), 0.04 (0.05), and
0.01 (0.01), respectively. For larger events (1,000 m3), the
conditional probabilities of the damage states limited,
serious, and destroyed are 0.7, 0.1 (0.3), 0.18 (0.2), and
0.02 (0.02) (Fig. 11; Table 3).
Following the October 2007 event, the road was closed
for a considerable period while major structural repairs
were undertaken to the area below the road. The observed
effects exceeded the destroyed damage state. This does
then beg the question as to how such a small event caused
so much damage. In simple terms, two events occurred.
The first flow above the road simply blocked the A83 and
in isolation required clean-up and pavement repairs.
However, it also blocked an open drain on the upslope side
of the road which caused water to overtop the road and to
erode a further wash-out-type failure below the road
(Fig. 13); it was this secondary event that did most of the
damage (as described in more detail by Winter et al. 2009).
In addition, the profile of this road and some of the con-
struction details, such as the aforementioned open drain,
make an argument for describing this road as a local road
(within the scheme described herein and notwithstanding the
higher speed limit). Thus, the conditional probability of the
destroyed damage state occurring in the context of the initial
1,000 m3 flow is higher, at around 0.2 (Fig. 11; Table 3).

Chuncheon National Road 46, Republic of Korea

For the Seoul to Chuncheon National Road 46, landslide Fig. 13 Detailed view of the October 2007 event at the A83 Rest and
deposits were up to 5,000 m3 (Lee and Winter 2010) be Thankful

123
Determining physical vulnerability of roads to debris flow
Fig. 14 Source area for a debris flow site from July 2009 at the Seoul
to Chuncheon National Road 46 in the Republic of Korea
(Fig. 14). For an event of this volume, the conditional
probabilities for no damage, limited, serious and destroyed
damage states are 0.4, 0.2 (0.6), 0.1 (0.4) and 0.3 (0.3)
(Fig. 11; Table 3).
The damage state corresponded to the destroyed cate-
gory, and this is reflected in the probability of such an
occurrence of 0.3 (an approximately one in three chance).
It should also be noted that the damage that occurred was
most likely exacerbated by the highly constrained nature of
both the natural topography and by the channelling effect
of concrete drainage channels on the lower slopes close to
the road (Lee and Winter 2010).
Seoul to Chuncheon National Highway tunnel portals,
Republic of Korea
For the Chuncheon National Highway tunnel portals
(Fig. 15), landslide deposits were 500–1,000 m3. For an
event of this volume (1,000 m3), the conditional proba-
bilities of the damage states no damage, limited, serious,
Fig. 15 Debris flow site from July 2009 above tunnel portals on the
Seoul to Chuncheon National Highway in the Republic of Korea: top
source area, and bottom view from the source area looking out over
the tunnel portals
and destroyed are 0.7, 0.1 (0.3), 0.18 (0.2), and 0.02 (0.02)
(Fig. 11; Table 3).
Only very minor damage was incurred at the Seoul to
Chuncheon National Highway tunnel portals, and this
reflects the small volumes; the probability of either limited
damage or no damage (the combined conditional proba-
bilities) is 0.8 (Fig. 15). The road was not open at the time
of the event, and there is every possibility of both further
and larger events that have the potential to meet or exceed
higher damage states.
Seoul to Gangnung Highway at Pyeong-chang Services,
Republic of Korea
For the Seoul to Gangnung Highway at Pyeong-chang
Services, landslide volumes were up to 10,000 m3
(Fig. 16). For an event of this volume, the conditional
probabilities of no damage and the damage states limited,
serious, and destroyed being met are 0.3, 0.15 (0.7), 0.15
(0.55), and 0.4 (0.4), respectively (Fig. 11; Table 3).
Less detail is available for the event at Seoul to Gang-
nung Highway at Pyeong-chang Services, but the damage
123

Fig. 16 Debris flow site above the Seoul to Gangnung Highway at
Pyeong-chang Services in the Republic of Korea: top view of the flow
channel above the road; bottom view from the flow channel above the
road, looking at the services located on the opposite side of the road
stage exceeded at least serious damage and, most likely,
destroyed. Certainly, significant remedial works have been
constructed since the flow occurred. Thus, the associated
combined conditional probabilities of those damage states
being met is 0.3 (or an approximately 1 in 3 chance of
occurrence), and this seems to be broadly reflective of the
events.
Conclusions
A survey of experts in the field of debris flow has been
conducted to ascertain their views on a wide range of
pertinent issues related to the construction of preliminary
fragility curves for the effects of debris flows on roads,
using an empirical approach.
Included in the questionnaire was the opportunity for the
respondents to make ‘free text’ responses to defined
questions. Their responses have been used, in part, to
determine the form of analysis. In particular, it was clear
that many respondents considered that flows up to
1,000,000 m3 should be considered rather than the
123
M. G. Winter et al.
maximum of 100,000 m3 contained in the questionnaire.
Consequently, the proposed fragility curves have been
extrapolated to include events one order of magnitude
greater than the largest considered in the questionnaire. In
addition, this form of determining fragility curves renders it
almost impossible for the probabilities to range from zero
to unity; accordingly, the proposed fragility curves have
been stretched to ensure such a spread.
The derived fragility curves have been compared to
known events in Scotland (UK) and the Republic of Korea.
In general, the curves tend to give results that might be
deemed reasonable, with probabilities of around 0.3 gen-
erally being suggested for known damage states. Excep-
tions to this occur when detailed site characteristics
introduce complexities that are not, and could not be,
accounted for in the analysis.
The method of data acquisition, the statistical analyses,
and the interpreted results, as well as the perceived inter-
pretations of the questionnaire for this first approach raise
some interesting issues that are explored in the body of the
report. Continued efforts are needed, potentially including
the use of modelling and empirical data to further validate,
and potentially improve, the curves reported herein.
Acknowledgments The work described in this paper was (partially)
supported by the European Commission through the project SafeLand
‘‘Living with landslide risk in Europe: assessment, effects of global
change, and risk management strategies’’ under grant agreement no.
226479 in the 7th Framework Programme of the European Com-
mission. This support is gratefully acknowledged. The TRL authors
gratefully acknowledge additional funding from Transport Scotland.
The first author is grateful to Professor Sugon LEE (of the University
of Seoul) who hosted his visit to the Republic of Korea.
References
Cruden DM, Varnes DJ (1996) Landslide types and processes. In:
Turner AK, Schuster RL (eds) Special report 247: landslides:
investigation and mitigation. National Academy of Science,
Washington, DC, pp 36–75
Galli M, Guzzetti F (2007) Landslide vulnerability criteria: a case
study from Umbria, Central Italy. Environ Manag 40:649–664
Geertsema M, Schwab JW, Blais-Stevens A, Sakals A (2009)
Landslides impacting linear infrastructure in west central British
Columbia. Nat Hazards 48:59–72
Haugen ED, Kaynia AM (2008) Vulnerability of structures impacted
by debris flow. In: Chen Z, Zhang J-M, Ho K, Wu FQ (eds)
Landslides and engineered slopes: from the past to the future.
CRC, London, pp 381–387
Jakob M, Stein D, Ulmi M (2012) Vulnerability of buildings to debris
flow impact. Nat Hazards 60(2):241–261
Lee S-G, Winter MG (2010) The effects of debris flow in the
Republic of Korea and some issues for successful management
and mitigation. In: Williams AL, Pinches GM, Chin CY,
McMorran TM, Massey CI (eds) Proceedings of the 11th IAEG
congress: geologically active. CRC, London, pp 1243–1250
Mansour MF, Morgenstern NR, Martin CD (2011) Expected damage
from displacement of slow-moving slides. Landslides 8:117–131
Determining physical vulnerability of roads to debris flow
Mavrouli O, Corominas J (2010a) Vulnerability of simple reinforced
concrete buildings in front of the rockfall impact. Landslides
7(2):169–180
Mavrouli O, Corominas J (2010b) Rockfall vulnerability assessment
for reinforced concrete buildings. Nat Hazards Earth Syst Sci
10(10):2055–2066
McMillan P, Matheson GD (1997) A two stage system for road rock
slope risk assessment. Int J Rock Mech Miner Sci 34(3–4):196
Milne FD, Werritty A, Davies MCR, Browne MJ (2009) A recent
debris flow event and implications for hazard management. Q J
Eng Geol Hydrogeol 42:51–60
Papathoma-Khöle M, Keiler M, Totschnig R, Glade T (2012)
Improvement of vulnerability curves using data from extreme
events: debris flow event in South Tyrol. Nat Hazards
64(3):2083–2105
Pierson TC, Costa JE (1987) A rheological classification of subaerial
sediment–water flows. Rev Eng Geol VII:1–12
Pitilakis K, Fotopoulou S (eds) (2011) Physical vulnerability of
elements at risk to landslides: methodology for evaluation,
fragility curves and damage states for buildings and lifelines.
SafeLand Deliverable 2.5. http://www.safeland-fp7.eu/
Pitilakis K, Alexoudi M, Argyroudis S, Monge O, Martin C (2006)
Earthquake risk assessment of lifelines. Bull Earthq Eng
4:365–390
Quan Luna B, Blahut J, van Westen CJ, Sterlacchini S, van Ach TWJ,
Akbas SO (2011) The application of numerical debris flow
modelling for the generation of physical vulnerability curves.
Nat Hazards Earth Syst Sci 11:2047–2060
Wieczorek GF, Mossa GS, Morgan BA (2004) Regional debris-flow
distribution and preliminary risk assessment from severe storm
events in the Appalachian Blue Ridge Province, USA. Land-
slides 1:53–59
Winter MG, Bromhead EN (2012) Landslide risk—some issues that
determine societal acceptance. Nat Hazards 62(2):169–187
Winter MG, Macgregor F, Shackman L (eds) (2005) Scottish road
network landslides study. The Scottish Executive, Edinburgh,
p 119
Winter MG, Heald AP, Parsons JA, Macgregor F, Shackman L (2006)
Scottish debris flow events of August 2004. Q J Eng Geol
Hydrogeol 39:73–78
Winter MG, Macgregor F, Shackman L (eds) (2009) Scottish road
network landslides study: implementation. Transport Scotland,
Edinburgh, p 278
Winter MG, Dent J, Macgregor F, Dempsey P, Motion A, Shackman
L (2010) Debris flow, rainfall and climate change in Scotland.
Q J Eng Geol Hydrogeol 43:429–446
123