Applied Mathematical Modelling 89 (2021) 710–730

Contents lists available at ScienceDirect

Applied Mathematical Modelling

journal homepage: www.elsevier.com/locate/apm

Assessing failure probability of coastal structures based on

probabilistic representation of sea conditions at the
structures’ location
Dimitra I. Malliouri a,∗, Constantine D. Memos a, Takvor H. Soukissian b,
Vasiliki K. Tsoukala a
a
Laboratory of Harbour Works, School of Civil Engineering, National Technical University of Athens, 5, Heroon Polytechniou Street, 15780,
Zografos, Greece
b
Hellenic Centre for Marine Research, Institute of Oceanography, 46.7 km Athens-Sounion Ave, 19013, Anavyssos, Greece

a r t i c l e i n f o a b s t r a c t

Article history: In the present paper a thorough probabilistic methodology is presented, aiming at esti-
Received 18 October 2019 mating the reliability of coastal structures, such as rubble mound breakwaters during their
Revised 15 June 2020
lifetime, based on the probabilistic representation of load environmental and resistance
Accepted 4 August 2020
parameters. One of the innovative points and main objectives of this study is the estima-
Available online 12 August 2020
tion of the failure probability of a coastal structure based on the long-term wave climate
Keywords: at the structure’s location, usually met in intermediate waters, using wave observations
Probabilistic design methodology or measurements in deeper waters. This task is accomplished by applying a wave propa-
Wave parameters gation statistical model in order that the joint probability density function of all random
Sea level data load parameters be estimated at the structure’s location. Moreover, a relation between an
Intermediate waters event-based extreme value analysis and an analysis on sea-state conditions within storm
Direct integration method events is derived in order that both of these two approaches could be compared estimat-
Monte carlo method ing the same kind of failure probability; an unconditional one. The latter is properly de-
fined here as the percentage of the structure’s Lifetime that the structure will be in a
failure situation. This unconditional failure probability provides direct information on the
time period that the structure will be in a situation of failure (since it considers the to-
tal range of loadings and total lifetime), and thus can be incorporated more efficiently in
an integrated risk analysis with consideration of social and economic costs. Besides, an-
other specific issue of scientific originality could be considered the investigation on the
proper time step denoting the sea state in the sea state analysis applied. In this manner,
the actual history and shape of each storm event is taken into consideration. Furthermore,
it is shown that these two approaches could be incorporated in the design of a coastal
structure. Moreover, two different fully probabilistic methods, Direct Integration Method
and Monte Carlo Method, were applied (and compared) by using a combination of vari-
ables with zero and non-zero hazard rate, referred here as a combined time-invariant and
time-variant analysis. Finally, the effect of considering additional and different types of
parameters as random variables on the assessed failure probability of the structure has
been investigated. The aforementioned methodology has been applied to a sample of wave
and sea level data at the structure’s location generated for this purpose. The original wave
data were derived from measurements in deeper waters than the structure’s location, cov-

∗
Corresponding author.
E-mail address: [email protected] (D.I. Malliouri).

https://doi.org/10.1016/j.apm.2020.08.001
0307-904X/© 2020 Elsevier Inc. All rights reserved.
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 711

ering a period of 8 years, obtained from an oceanographic buoy, located in the western
Mediterranean off Malaga, Spain. Sea level data were also obtained from a tide gauge,
located in Malaga’s harbor. Finally, it is shown that the methodology derived from this
study could be incorporated into a coastal structure’s design process to meet specific safety
requirements.
© 2020 Elsevier Inc. All rights reserved.

1. Introduction

Structural reliability assessment, aiming at estimating the probability of a structure to meet predefined safety and perfor-
mance levels, with due consideration of all uncertainties related to actions, resistance, and design tools, has been a topic of
extensive research in many different engineering application areas. The most versatile solution technique available for this
purpose is the Direct Integration Method (DIM) or the alternative Monte Carlo Method (MCM). Although DIM approach is
the most theoretically direct solution for structural reliability problems, in practice reliabilities of many offshore structures
are more likely to be evaluated using simulation-based methods (e.g. direct MCM). Therefore, several advanced approximate
methods have been also developed so far for estimating the structural failure, such as the first and second order reliability
methods, environmental contour methods, response surface methods and emulator-based analysis.
Referring to these approximate methods, a brief literature description is provided below. In the first-order reliability
method (FORM) the failure surface is approximated by a tangent hyperplane at the design point [1], while a more accu-
rate and similar method is the second-order reliability method (SORM), which uses a quadratic approximation to the failure
surface (see e.g. [2]). Moreover, the concept of environmental contours is a method developed for estimating extreme con-
ditions as a basis for design, e.g. [3,4], and is widely used in marine structural design (see e.g. [5]). This approach is close
to the first-order reliability method approximation. Moreover, whereas the response surface method motivated by Box and
Wilson in 1951 [6] was traditionally utilized for chemical or industrial engineering, its application area has been expanded in
structural engineering (see e.g. [7]). One way to address the computational cost incurred in the simulation of the structural
reliability assessment could be to apply a less demanding approximation, e.g. a metamodel or emulator (see e.g. [8]).
Two main criteria for selecting one of the aforementioned methods for structural reliability assessment could be the high
degree of accuracy required in a very narrow region, i.e. the small failure probability, and the computational cost incurred.
The latter could become significantly demanding, depending on the application case.
Focusing on coastal structures, the reliability based design methods are divided into four categories, related to the accu-
racy in determining the reliability of their elements and the consideration of all involved uncertainties, cited below [9]:

• Deterministic method (Level 0)

• Semi-Probabilistic Method (Level I)
• Probabilistic Methods with Approximations (Level II)
• Fully Probabilistic Methods (Level III)

Conventional design practice for coastal structures is often deterministic in nature (Level 0), and its reliability is based on
the exceedance probability of the design wave load. Specifically, the notion of design wave parameters and especially that of
wave height associated with a certain return period is adopted [10]. In addition to the deterministic method, partial factors
can be implemented for resistance and load, based on standards [11]. This calculation is classified as a semi-probabilistic
(Level I) method. However, this approach does not allow accurate determination of the reliability of the design. In order
to overcome this problem, more advanced probabilistic methods should be applied that consider the uncertainties of all
random variables of load and strength of the structure.
Fully probabilistic methods (Level III) belong in the category of advanced probabilistic methods, which consider the joint
probability density function of all random variables involved. There are two main methods classified as Level III methods,
i.e. the Direct Integration Method (DIM) and Monte Carlo Method (MCM). Another possibility, though less accurate regarding
the estimation of the reliability of the design, are methods with approximations (Level II).
Referring, at this point, to the general gap in current practice of reliability assessment of coastal structures, despite that
the probabilistic representation of environmental parameters at the location of coastal structures is vital for the reliability
based design of these structures, there is a lack of long-term environmental data at the location of coastal structures under
design, making thus difficult to estimate their joint probability distribution at this location. Therefore, one of the main
objectives of this paper is to deal with the issue of how to properly describe the long-term wave climate at the coastal
structures’ location, and then to use this core information for the reliability assessment and design of coastal structures.
As for the structure of the present paper, a brief description of the relevant theoretical background is given in Section 2.
In Section 3, some significant issues in the reliability assessment of coastal structures are clarified. The proposed methodol-
ogy developed is presented in Section 4, while in Section 5 the case study of Malaga is described and results are presented
and discussed. Finally, conclusions are provided in Section 6.
712 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

2. Theoretical background

2.1. Reliability of an element

The reliability of an element depends on the safety margin between the strength (i.e. resistance R) and the load or
action A. The limit state function g describes the relation between resistance and action upon an element and is formulated
as follows:
g=R−A (1)
Thus, the probability of failure of a structural element Pf and the reliability level of that element Re are defined by
Eq. (2) and 3, respectively.

Pf = P rob(g < 0 ) = P rob(A > R ) (2)

Re = 1 − Pf (3)
A deterministic calculation method uses nominal values of the basic variables. Often a global safety factor γ is applied
(γ > 1), to deal with the unknown uncertainties in the basic variables and to provide a safety margin between strength and
load. Thus, the reliability inequality is formulated as:
Rnom ≥ γ Anom (4)
where: Rnom and Anom are the nominal values for strength and action, respectively. In the deterministic calculation, the load
variables correspond to certain return periods, which can be extracted from measurements. These measurements usually
cover a time period of data, usually shorter than the return period of interest. In general, the longer period of measurements
makes the estimated return values more reliable.
In the semi-probabilistic approach, partial factors developed initially by PIANC [11] are implemented for strength and
load via the following relation, creating by this way a safety margin:
Rnom
≥ γA Anom (5)
γr
where, γ r (γ r ≥ 1) and γ A (γ A ≥ 1) are the partial factors for strength Rnom and load Anom , respectively.
In short, according to [11], partial factors are calibrated with the following input:
• Design lifetime L (=20, 50 or 100 years)
• Return period of actions equal to Design Lifetime
• Acceptable probability of failure Pf L|e (=0.01, 0.05, 0.10, 0.20, or 0.40) depending on the consequence class of the structure.
This target failure probability used here is equal to the conditional failure probability during the structure’s Lifetime
given that only extreme events are considered, reflecting the failure proportion of the structure in the sample of extreme
events. In this approach of partial safety factors, all random variables’ uncertainties (e.g. of actions, resistance parameters,
model uncertainties) have been considered.

Another possibility of estimating the reliability of an element is to use a Level II method, which is a probabilistic method
with approximations. Its calculations are based on a rather iterative and complex process, making thus Level II methods
difficult to apply, if the random variables are non-normally distributed and statistically correlated, and the failure function
is not linear.
When a fully probabilistic method is applied, e.g. Direct Integration Method (DIM), or Monte Carlo Method (MCM), the
failure probability of an element can be calculated accurately based on the probabilistic framework of all random variables
involved. The core problem of DIM is the exact estimation of the joint probability density function (pdf) of these variables.
Given that the marginal pdfs of the variables considered are known, the calculation of their joint pdf is necessary only if
these variables are correlated; otherwise their joint pdf is equal to the product of the marginal pdfs of the variables.
−
→
→ denotes the joint pdf of all random variables involved, i.e. of the vector X = (X1 , X2 , . . . , Xn ), the probability of
If f−
X
failure can be calculated via the following integral:
 −
→ −
→
Pf = f−
→ x d x (6)
−
→ X
( )
g x <0

The MCM is based on a large number of simulations N, a part of which (Nf ) leads to element’s failure. Thus, it is assumed
that provided N is high enough, Pf attains acceptable convergence and is computed as:
Nf
Pf = (7)
N
One way (e.g. inverse CDF method) to generate the random sample for each variable is described here. Specifically, each
simulation starts by drawing a random number from a uniform probability density function U(0, 1). Through those numbers
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 713

(xui ), a sample of the variable (X) can be produced by the latter’s inverse cumulative distribution function (cdf)FX−1 (x ) as
follows:
Xi = FX−1 (xui ), i = 1, 2, 3, ..., N (8)
Hence, the problem in this approach is to estimate the joint distribution function of all random input variables and
compute its inverse. Let us also note that when the marginal cdf(s) are not continuous then, in relation (8), the generalized
inverse distribution function should be used.

2.2. Capturing the statistical correlation of random variables

In case of correlated random variables, an efficient way to account for their correlation is to apply the conditional prob-
ability model in order to calculate their joint pdf. The said model can be illustrated by the total probability law applied in
the following Eq. (9) indicatively to two variables, but could be generalized to more than two variables, as well:
fX1 ,X2 (x1 , x2 ) = fX1 |X2 (x1 |x2 ) fX2 (x2 ) (9)
where fX1 ,X2 (x1 , x2 ) is the joint pdf of the random variables X1 ,X2 , fX1 |X2 (x1 |x2 ) is the conditional pdf of X2 given X1 , and
fX2 (x2 ) is the marginal pdf of X2 .

2.3. Estimation of an element’s failure probability during the structure’s lifetime

2.3.1. Time-invariant analysis

−
→
Given that f− → i.e. the joint pdf of all random variables X = (X1 , X2 , . . . , Xn ) involved, is independent of time, and the
X
−
→
variables X do not exhibit an extremal behavior, i.e. they are characterized by a zero-hazard rate, then the one-year prob-
ability of failure Pf,1y of a system or an element is equal to the corresponding failure probability referring to the structure’s
Lifetime Pf,L :
Pf,L = Pf,1y (10)
Variables of this type could be deterministic ones, such as tide, whose value can be predicted in time but is combined
with wave parameters in a random way, due to the random structure of the latter parameters. Resistance parameters may
also belong in this category, as they could be regarded either as deterministic or random variables with a zero hazard rate.
This analysis could be referred to as a time-invariant analysis.

2.3.2. Time-variant analysis

The environmental load parameters of a coastal structure, due to wave kinematics, sea level change induced by storm
surge, and sea currents, are commonly time-variant, referring to the inherent variability of those variables. It is thus neces-
sary to implement an extreme value analysis or an extrapolation of their probability distributions to their right tails (i.e., to
lower probabilities of extreme values). In such an analysis, which could be referred to as a time-variant analysis, if Y years
of observations were available, the encounter probability of an extreme value extracted from the dataset (i.e. the probability
that this value will be exceeded once in the mean within Y years), is equal to the average encounter probability of this
extreme value within 1 year. In deterministic design, the latter is usually considered as the average failure probability of 1
year denoted by Pf∗,1y referring to the total sample of data, and by Pf∗,1y|e referred to the extreme sample. Both of them may
be extrapolated to larger time periods such as the design lifetime of the structure L by assuming independent failure events
from one year to another [9,10] and stationary conditions among years. The latter means that long-term trends due e.g.
to climatic variability have not been considered. The aforementioned extrapolation can be accomplished via the following
expression, referring to the extreme sample:
 λe L
Pf∗,L|e = 1 − 1 − Pf∗,1y|e (11)
where: λe is the number of extreme observations considered per year of observations, i.e. the mean hazard rate.
Eq. (11) is based on the following relation [10], written in a more illustrative form here:
 Ne,L
FXL|e (xo ) = PL (x < xo|x ∈ extreme) = FX1|ye (xo ) (12)

1y
where: FXL|e (xo ) and FX |e (xo ) are the average conditional cumulative probabilities of an extreme event xo within L, and 1 year,
respectively, given that only extreme events are considered, extracted from Y years of observations. Ne,L denotes the number
of independent extreme events during L.
Assuming now that the number of extreme events considered per year is only one, i.e. the maximum observation of each
year, then Eq. (11) can be written as follows:
 L
Pf∗,L|e = 1 − 1 − Pf∗,1y|e (13)
since in this case Ne,L = L.
714 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

The time extrapolation that refers to the total sample is accomplished by the following expression:
 L
Pf∗,L = 1 − 1 − Pf∗,1y (14)

where Pf∗,1y is the average one-year probability of failure of a system referring to the total sample of data of time-dependent
random variables and Pf∗,L is the corresponding average failure probability referring to the structure’s Lifetime.

3. Significant issues in the reliability analysis

Currently, there are some significant issues that need to be solved during the reliability assessment, and consequently
the design process of coastal structures. Some of them, which should be duly considered, are listed below focusing on the
case of a rubble mound breakwater. However, the discussion and the methodology that will be presented below can easily
be adjusted to account for every coastal structure.

(1) Input data of the coastal structure’s load parameters that affect the limit state function of its elements should be
collected, before performing the reliability analysis. Input data on the long-term wave climate and sea level should be
collected at the project site. Such data include variables like significant wave height Hs , mean wave period Tm , mean
wave direction θ m , sea level, etc. The data can be acquired through:
(i) Measurements at the structure’s location for a long period of time, ideally of the order of the structure’s design
operating life. Obviously, this requirement can be rarely met.
(ii) Measurements of the same type but in deep water during a time span ideally of the order of the structure’s design
operating life. Then, either linear or non-linear models (e.g. [12,13,14]) may be applied to transfer each sea state
from deep towards shallower waters provided that the location of the structure is in shallower waters.
(iii) Hindcasting methods for a period of several years ([15,9]). Particularly, regarding the long-term wave climate at
the structure’s location, wind data in the wave generation area (e.g. [16]) can be used to provide this information
under a probabilistic framework.

Furthermore, the target of the aforementioned input data collection is to gain a thorough probabilistic representation of
the long-term wave climate, sea level conditions, and other environmental parameters, at the coastal structure’s location,
usually met in intermediate waters. In the above case (ii), it is worth mentioning that non-linear models might be more
accurate than linear models, as they account for wave-wave interactions that may play a considerable role in the final
outcome depending on water depth; however, non-linear models are quite demanding in terms of computing resources,
especially when they have to account for reliability analysis based on fully probabilistic methods. In this case, a fast and
adequately accurate model could be applied that enables the integration of short- with long-term wave statistics [17] and
transfers the probabilistic information to the structure’s location.

(2) A functional dependence may exist between certain environmental parameters involved in the reliability analysis.
Specifically, the total mean sea level (including tide, storm surge, etc.) often plays a role in modifying the wave climate
at the structure’s location, since it forms the water depth conditions, affecting as a result the wave conditions at this
location.
(3) Referring to wave parameters that affect the stability of a rubble mound breakwater, oblique wave attack does play
a significant role in the stability of a breakwater’s armor layer (see e.g. [18]). In particular, it is accepted that the
stability of armor layers can be significantly higher for oblique waves compared against perpendicular wave attack
([19]), thus neglecting of the effect of oblique waves on the structure’s stability leads to a conservative assumption.
According to [18], the consideration of this effect can be accomplished by using the equivalent normal wave height.
Therefore, the principal or mean wave direction that characterizes each sea sate at the breakwater’s location has been
included in the analysis. This type of functional dependence, i.e. between wave direction and resistance of an element
is considered via the limit state equation.
(4) Another issue is related to the hazard rate of some of the parameters considered in the reliability analysis, referred
here as time dependence. Then the random variables involved in a design formula should be examined with rele-
vance to their dependence on time, and a combination of time-dependent and time-invariant reliability analysis may
be applied, when extrapolating the failure probability from the reference period of one year to the lifetime of the
structure.
(5) The exceedance probability of a design value of a time-dependent core load variable is commonly regarded as the
failure probability of the structure in the conventional design, given that only extreme values of the variable are
considered (e.g. values above a predefined high threshold (Peak over Threshold- POT)). In a conventional event-based
POT-analysis (e.g. [20]), the exceedance probability of a storm event is considered as the ratio of the expected number
of events exceeding this specific storm event to the number of extreme events expected to occur in a reference time
period, i.e. annually. It is noted that all these (event-based) probabilities are conditional ones (i.e. estimated in the
sample of extreme events) and do not provide direct information on the time period that the structure will fail during
a reference time period. The latter issue is further discussed in Section 4.3.
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 715

4. Proposed methodology

4.1. Transferring the probabilistic information of sea conditions from deeper waters to the structure’s location

The model selected in the present paper to transfer storms and sea states from deeper waters to the structure’s location
is the one developed by Malliouri et al. [17]. This is a statistical wave propagation model that uses the long-term wave
statistics in deep waters as input data and by using the short-term wave statistics for each sea state or storm event in
deep waters, it estimates the long-term wave statistics in shallower waters. Specifically, the short-term joint distribution of
individual wave height H, period T, and direction θ for every sea state or storm event could be produced in deep waters
by using: i) data/measurements of significant wave height Hs , mean wave period Tm , and mean wave direction θ m in deep
waters, ii) the dimensionless short-term images by Memos and Tzanis ([21,22]) in deep waters, iii) a theoretical expression
for wave directionality adjusted in an individual wave statistical analysis by Malliouri et al. [17], and iv) a modification of
Battjes approach [23]. Then, the short- and the long-term joint distribution of H, T, θ could be estimated in intermediate
waters, as well as the long-term joint probability density function of Hs , Tm , θ m , by considering linear wave transformation of
each individual wave, as waves propagate from the open sea towards shallower waters. This model is simpler than nonlinear
wave propagation models, but can easily handle a large number of sea states or storm events with adequate accuracy and
considerable computational cost reduction.
Since, for the response of a structure (such as a rubble mound breakwater) within a sea state or storm event, the design
formulas consider and use the integrated parameters only, e.g. Hs , Tm , θ m , therefore the joint probability density function of
Hs , Tm , θ m should be utilized here for the reliability assessment. Let it be noted that the model by Malliouri et al. [17] also
produces the long-term joint distribution of H, T, θ at the structure’s location, if this is in intermediate water depths.
Additionally, comparisons have been made between the statistical model of wave propagation from deep to shallower
waters, referring to short-term sea conditions, with wave measurements in deep waters and results in intermediate waters
extracted from a Boussinesq wave propagation model. These comparisons [24] showed that the statistical wave propagation
model [17] can well be applied in deep and intermediate waters, up to the surf zone.

4.2. Incorporation of the reliability analysis into a design process

The steps of the methodology adopted in this study are described below and can formulate a reliability analysis of a
coastal structure under wave action, e.g. a rubble mound breakwater considered as a system, with a known simple fault
tree. This reliability analysis is indicatively incorporated in a design process example below:

Step (a): Define design requirements for Limit State [9] adopted (e.g. Serviceability Limit State (SLS), Ultimate Limit State
(ULS)), such as Design lifetime, and allowable failure probability during the lifetime.
Step (b): For every element of the system, apply a preliminary design, e.g. deterministic design, which will provide an
initial result (i.e. some structural characteristics, e.g. mass of armor units), for the fully probabilistic design.
Step (c): For every element of the system apply a fully probabilistic method to estimate the failure probability of the
element.
Step (d): Evaluate the system’s reliability by estimating the system’s probability of failure. The latter can be calculated
via estimation of the probability of the union or of the intersection of the elements’ failure events, referring to the
case of series or parallel system, respectively.
Step (e): Optimize the structure’s characteristics according to the design requirements and economic costs.

A flowchart of the aforementioned methodology is presented in Fig. 1.

The approach for the preliminary design adopted in this study is an event-based extreme value analysis applied to wave
data, aiming to distinguish medium and extreme storm events from the total sample of sequential data at a given step. The
result extracted from this procedure is the time period that the structure will be in a situation of failure during its Lifetime.
Moreover, the fully probabilistic approach is applied directly to the total sample or a subset of sequential wave data, i.e. sea
states, estimating the same type of probability. Firstly, the preliminary design approach will be described, followed by the
fully probabilistic approach.

4.3. Preliminary design based on an event-based extreme value analysis applied to wave data

The peaks over threshold modeling will be used for the predominant variable Hs,p , i.e. the maximum Hs within a storm
event, following the relevant event-based framework for over-threshold modeling introduced by Bernardara et al. [20], which
will then be associated with the concurrent values Tm and θ m of that particular storm [26].
Firstly, the sequential wave data are homogenized at a given time step enabling the separation of data/sea states into
independent sea-state systems (called hereafter “wave events”). This is attempted by firstly selecting an initial threshold u1
for Hs distinguishing “weak” and “intense” storms from the dataset. Then, a minimum calm period, e.g. of 6 hr [26] (i.e.,
a time separation interval), between two consecutive storms was selected to ensure that the two events are independent.
Therefore, the dataset of the maximum significant wave heights Hs,p is derived from this procedure. Referring to the selection
716 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

Fig. 1. flowchart of a reliability analysis incorporated into a design procedure (based on Everts [25]).

of u1 , it should be high enough to distinguish and separate two consecutive storms, but also low enough to be below the
extreme area [27].
The second step is the selection of another threshold u2 higher than u1 . The selection of u2 needs to be more accurate
than that of u1 , since the storm Hs -peaks above u2 should follow the same extreme probability distribution. Referring to
modeling from the perspective of peaks over threshold, Pickands [28] showed that this extreme probability distribution can
be approximated by a Generalized Pareto distribution (see also [29]). That is, for Hs,p > xo given that Hs,p > u2 ,
 xo − u2
−1/ξ
P r ( Hs,p > xo|Hs,p > u2 ) = 1 + ξ (15)
σ
where: xo is a value higher than u2 , ξ ∈ ( − ∞, +∞) and σ > 0 are the shape and scale parameter, respectively, of GP
distribution.
Using the properties of the GP distribution to determine u2 , two different methodologies are commonly used. Specifically,
an appropriate threshold u2 value could be selected by examining the domain of stability of the shape parameter ξ , and the
domain of stability of the modified scale parameter σ or the linearity of the scale parameter σ of the GP distribution with
respect to u2 (see e.g. [30]). The second procedure for threshold selection is to identify the domain of threshold u2 where
the mean residual life plot is approximately linear with respect to u2 [31].
The return (design) value of Hs,p from GP model is estimated as follows:
 
ξ
Hs,p− Tr = u2 + σξ (Tr λe ) − 1 , for ξ = 0
(16)
Hs,p− Tr = u2 + σ log (Tr λe ), for ξ = 0
Therefore, the design storm event corresponds to a return value of Hs,p associated with the most probable values of storm
mean wave period Tm , and storm mean wave direction θ m , as estimated from their scatter diagram. To estimate the average
proportion of the design storm exceedance time period within 1 year, the expected value of the design storm duration is
used in the following equation, developed in the present paper:
mDdesign storm λe mD e Y mDdesign storm
Pex,storm,1y = Pex,storm,1y|e Pu2 = · = (17)
λe Tr mD e Y 365nY dt Tr 365nY dt
where: Pex,storm,1y|e is the exceedance probability of the design storm Tr -Hs,p given exceedances of u2 , Pu2 is the exceedance
probability of u2 by Hs,p regarding the total data of Hs with a time step of dt, λe is the number of extreme events
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 717

(i.e. number of Hs,p exceeding u2 ) per year of observations, mDdesign storm is the expected duration of the design storm, Y
is the number of years of observations, mD e is the mean duration of storms with intensity Hs,p exceeding GP threshold u2 ,
and nY is the number of sequential observations per day at a given time step dt. The expected duration of the design storm
mDdesign storm is considered here equal to mD e , which is in accordance with Leadbetter [32], who used an extremal index for
the transformation from the event-describing peak to the sequential observation. The extremal index may be interpreted as
the reciprocal of the mean number of sequential values, e.g. of Hs , Tm data, per event.
Since Hs varies with time throughout a storm, it would be very conservative to associate Hs,p with the total storm dura-
tion mDdesign storm , thus by assuming a triangle storm history [33], Hs,p is used here, associated with the equivalent duration
mDdesign storm /2. Then Eq. (17) changes to the following equation that estimates the exceedance of the design Hs,p during the
reference period of 1 year:
mDdesign storm /2
Pex,Hs,p ,1y = (18)
Tr 365nY dt
To assess the exceedance of the design Hs,p during the reference period of L years, considered here as an equivalent
∗
failure probability, Eq. (14) should be applied replacing Pf, by Pex,Hs,p ,1y .
1y
However, there are some subjective steps and uncertainties involved in the design approach based on the above extreme
value analysis: specifically, as for the domain of stability of the generalized Pareto parameters with respect to its threshold,
although objective methods ([30]; [31]) have been developed for this purpose, the accurate selection of the generalized
Pareto threshold may still be difficult in some cases [27]. However, this threshold selection is critical, since it determines
the expected number of extreme events per year λe , and consequently the exceedance probability of the design storm,
leading thus to uncertainties of the analysis results. Another subjective issue in this approach relates to the future prediction
of the duration of the design storm. This could be estimated either as the mean duration of the storms observed or as
the most probable one depending on the scatter plot of storm’s duration and Hs,p . Last but not least, the consideration of
the equivalent duration associated to Hs,p of a storm with duration mDdesign storm /2 is rather approximate and needs further
research, because the limit state of a coastal structure, e.g. a rubble mound breakwater, does depend on the number of
waves attacking the structure during a storm, i.e. on its duration.
Also, as far as the multivariate extreme value analysis is concerned, according to the classification proposed by Mazas
and Hamm [34], whereby wave spectral parameters are different physical quantities that belong to the same phenomenon,
i.e. the wave storm, firstly the choice of the event-defining variable should be made, i.e. of Hs,p , which should then be
associated to the concurrent values of the other covariate (spectral) parameters. Referring now to different components of
a single phenomenon, like waves and sea levels conditions, the predominant component should be chosen firstly [35] that
leads to the choice of the event-describing variable, e.g. the event-describing variable is sea level in case of a macro-tidal
environment, with the concomitant Hs value. Furthermore, in case of equivalent components, two event-based describing
variables should be considered [34] in combination with a bivariate threshold by using multivariate generalized Pareto (GP)
distributions (e.g. [36]).

4.4. Fully probabilistic reliability analysis

The fully probabilistic analysis, to be presented in the following, is applied to sequential wave data within “weak” and
“intense” storms by adopting a time step, enabling equally-discrete representations of the temporal variation of sea-state
conditions within storm events, denoted from now on simply as sea states. The result of this approach is the unconditional
failure probability of the structure, extracted more directly from initial data than the approach based on the extreme value
analysis of Section 4.3. In the fully probabilistic analysis, the individual data used describing sea-state conditions, are not
required to be mutually independent. Besides, depending on the intended use of the analysis, the time step applied on the
initial sequence of sea states could be equal or higher from that of the time interval between successive measurements,
in which sea conditions are considered stationary. It is noted that a short time interval between successive measurements,
e.g. of 1 hr, is preferable since it enables avoidance of record loss of a fast change or peak in sea conditions. Furthermore,
the time step used for representation of sea-state conditions for the design and operation of coastal structures, is desirable
to range between 1 hr and 3 hr depending on the type of the structure element to be assessed. Specifically, regarding
the reliability analysis of the main armor of coastal structures, the minimum three hours interval is recommended to be
used in order that the data correspond to more than 10 0 0 waves per loading event, as suggested by limit state function
[37] associated to the said design element. Obviously other limit state functions in the realm of coastal engineering may
require other time lengths for optimum data assimilation. It is underlined that the issue of loading duration is directly
linked with the notion of structural resilience, i.e. the ability of a structure to withstand a storm containing several sea
states, as defined above. This ability is not warranted by the successful result over sea states contained in the storm, when
such result is based on a limit state assuming a small amount of damage as in the armoring of mound breakwaters. The
reason for this is that in reality small damages in such structures cannot be repaired during storms, a fact commonly not
accounted for in relevant design formulae.
By adopting a proper time step denoting representations of sea-state conditions of equal duration to be considered as
loadings events upon the structure, e.g. of 3 hr, the maximum or mean Hs within this duration associated with its concurrent
718 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

covariate parameters, i.e. Tm and θ m , should be selected, in order to represent the corresponding loading of the structure. In
this manner, the actual history and shape of each storm event is taken into consideration. Therefore, in the fully probabilistic
analysis examined, this vital information is not missed in contrast to the event-based extreme value analysis mentioned in
Section 4.3. Besides, it is recommended that first this fully probabilistic reliability analysis should be applied to the sample
of sea states of medium and extreme events and then the estimated conditional failure probability should be multiplied
by the probability of the threshold exceedance in order that the unconditional one be derived. Thus, working on the basis
of sea states of equal duration, defined as above, enables the direct derivation of the percentage of time during which the
structure will be in a failure state, a useful metric in risk analysis.

4.4.1. Multivariate analysis of sea conditions within storm events

In the analysis of sea conditions within storm events applied to initial time series of wave spectral parameters (wave
data), the conditional model is used to fit bivariate distributions of Hs and Tm to the available data set for each directional
sector, written as below:

fTm ,Hs ,θm (Tm , Hs , θmi ) = fθm (θmi ) fHs |θm (Hs |θmi ) fTm |(Hs ,θm ) (Tm |(Hs , θmi )) (19)
i i

where: fθm (θmi ) is the probability density of the i th directional sector with central value of θmi and bin size of dθ m , i.e.
θm ∈ (θmi − dθ2m , θmi + dθ2m ) measured from North, fHs |θm (Hs |θmi ) is the conditional pdf of Hs given that θ m belongs to the
i th directional sector, and fTm |(Hs ,θm ) (Tm |(Hs , θmi )) is the conditional pdf of Tm given that Hs has a certain value and θ m
i
belongs to the i th directional sector.
The conditional bivariate pdf of Tm and Hs , given θ m belongs to the i th directional sector, fHs ,Tm |θm (Tm , Hs |θmi ) fθm (θmi )
i
was modeled according to Mathisen and Bitner-Gregersen [38]. The pdf fHs |θm i (Hs |θmi ) can well be represented by a Weibull
pdf in many cases [39], while the pdf fTm |Hs |θm (Tm |(Hs , θm i )) can be modeled by a lognormal probability distribution, and
i
the location and scale parameters of the conditional lognormal pdf can be written as functions of Hs |θm i as below [38],
adjusted here to consider also θ m :

μ(Hs |θmi ) = a1 + a2 Hsa3 d θm d θm

, θ ∈ θm i − , θm i + (20)
σ (Hs |θmi ) = b1 + b2 exp (b3 Hs ) m 2 2

where a1 , a2 , a3 , b1 , b2 , and b3 are constants determined from fitting of the curves described by Eq. (20) to the data.

4.4.2. Time extrapolation of an element’s failure probability using time-invariant and time-variant random variables via direct
integration method (DIM)
In a reliability analysis, there may be time-variant but also time-invariant parameters, i.e. variables with nonzero and
zero hazard rate, that all play a role in the reliability or the average total failure probability Pf,L of the structure during
its lifetime. Time-invariant parameters, however, are independent of the reference period L. In practice, such time-invariant
parameters can usually be considered those related to the resistance of the structure’s element. Thus, Pf,L is estimated via
the following equation developed here to account for a combined time-variant and –invariant analysis:
   

Pf,L = → (
f− r ) f−
→ ( x )d xd r = → (
f− r) → (
f− x )d x d r (21)
R X ,L R X ,L
(r,x ) (r,x )
−
→ −
→ −
→
where hereinafter f− → ( r ) is the joint pdf of the time-invariant resistance parameters r , f −
→ ( x ) is the joint pdf of time-
R X ,L
−
→ −
→ − →
variant random load variables x during design lifetime L, and ( r , x ) is the failure domain, which is equal to the union
of the definition domains of all random variables involved under the condition that the limit state function is less than zero,
−
→ − →
i.e. g( r , x ) < 0. Therefore, it is evident that:
→ −→  − → 
 − → −
r , x = g r , x <0 (22)
Using Eq. (14), Eq. (21) results in:

L 
 − 
→ −
→ −
→ −
→
Pf,L = f−
→ r 1− 1− f−
→
X ,1y
x dx dr (23)
 (−
→−→
R
r,x)

where f−
→
X ,1y
(−
→ −
→
x ) is the average joint pdf of x referring to the reference period of one year, extracted from Y years of
−
→
observations. As it was aforementioned, the time-variant load parameters x consist of sea-state parameters, such as Hs , Tm ,
and θ m or similar, sea level variation due to storm surge, denoted by SL, current velocity, etc.

4.4.3. Inclusion of sea level variation due to astronomical tide (TL) by dim
The most common environmental parameter present around coasts and not related to the wave field is the sea-level
variation due to the astronomical tide (TL). This variable is deterministic, and depends on the tidal cycle of the site. Nev-
ertheless, it can be combined randomly with waves, and thus can be considered as a random variable. However, it is not
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 719

statistically correlated with wave parameters, but the depth-limited wave parameters depend on the sea level value which
can be strongly affected by tide, storm surge etc.
The timing of tide is mainly associated with the relative position of the Moon to the Earth and advances at about 50 min
a day; tides also vary seasonally, while only minor variations occur outside a period of about 19 years [40]. Therefore,
given that the lifetime of most coastal structures ranges between 20 and 50 years, the astronomical tide can be considered
as a time invariant parameter, and its pdf stable over the years. Thus, the inclusion of the tide component in Eq. (23) is
accomplished as follows:

L 
  − 
→ −
→ −
→ −
→
Pf,L = f−
→ r fT L (tl ) 1 − 1− f−
→
X ,1y
x dx d r dtl (24)
 (−
→− →
R
r , x ,tl )

−
→ − →
where: fTL (tl) is the pdf of sea level due to tide and ( r , x , tl ) is the expanded failure domain.
Eq. (24) could be estimated by defining each definite integral as the limit of a sum:
⎧ ⎡  L ⎤ ⎫
 − ⎨ ⎬
→   −
→ −→ ⎦
Pf,L = −
lim f → r
− lim f T L (tl )⎣1 − 1 − lim  ( −
→−
,
→
,tl ) f −
→ x δ x δ tl δ−
→
r (25)
δ→r →0 −
→
r
R
⎩δ tl→ 0
tl
δ −
→
x → 0 −
→
x
r x X , 1 y
⎭

Eq. (25) implies that the increments of all random variables must be small enough in order that Pf,L be estimated
properly.

4.4.4. Resistance parameters

Resistance parameters are those related to the strength/ resistance of the structure and usually have some variability
from their characteristic value. The density of stone or artificial armor units, the sea water density, geometric parameters
or even empirical formulas’ parameters derived from calibration based on experimental studies belong in this category. In
the preliminary design of coastal structures, resistance parameters are replaced by their characteristic values, e.g. median,
mean etc., in the design formula for every element. The type of the resistance parameter value is determined by the design
formula, i.e. limit state function. However, in probabilistic design (see Fig. 1), resistance parameters can follow probability
distributions. A straightforward way to model their variability is to consider them as normal random variables.

4.4.5. Monte carlo method (MCM)

In MCM, the frequency density of TL and the joint pdfs of SL, θ m , Tm , and Hs are also used, aiming at generating a random
sample for each variable that should follow its pdf, but also at maintaining the statistical correlation between correlated
variables. Therefore, Eqs (24) and (25) still apply in this method. The generation of the artificial sample of θ m , Tm , and
Hs that corresponds to a total sea level value at the structure’s location can be produced in the simple way, by randomly
generating a number of groups of Tm , Hs , and θ m between the limits of each joint class, equal to the new frequency of each
joint class.
Alternatively, Eq. (8) could be used to produce a MCM sample, firstly by generating θ m values that follow their pdf. Then,
for every θ m value depending on its class, i.e. directional sector, a conditional inverse cumulative distribution function (icdf)
of Hs |θ mi can be used to produce Hs values whose sum is an integer linearly related to the frequency of θ m class. Therefore,
for every value of θ m , there is a corresponding value for Hs that subsequently determines the conditional icdf’s parameters
of Tm |(Hs , θ m ), which are functions of Hs . Thus, by using again Eq. (8), N groups of θ m , Hs , Tm are produced, where N is large
enough to attain convergence of the solution of MCM. It is noted that the two approaches give similar results, but the first
one used here is simpler.
Denoting by Ns , Ntl , Nr , the total number of the random samples of sea level data components SL, TL and resistance
parameters R produced by MCM respectively, then by combining Eq. (8) and Eq. (25), the following relation is obtained:
⎛ ⎡  L ⎤ ⎞
   Nf
Pf,L = ⎝ ⎣1 − 1− /Ns ⎦/Ntl ⎠/Nr (26)
N
Nr Ntl Ns

where: Nf is the number of occurrence of the system’s failure (see Table 1) and N is the total number of simulations/ sea
state events.

4.4.6. Modeling the probability of union of the elements’ failure events

A simple way to model the union probability of the elements’ failure events is to check whether all elements of the
structure fail under the same load conditions. In this way, the dependence of the elements and the system with the en-
vironmental loadings has been considered, while the physical dependence between the elements (e.g. if the damage of
an element can make another element more vulnerable to damage) has not been taken into consideration. An illustrative
example of this procedure concerning a series system of three elements of a rubble mound breakwater is presented in
Table 1 referring to the case of MCM.
In case of DIM, Table 1 can be rewritten in the following form (see Table 2):
720 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

Table 1
Estimation of a series system’s failure frequency via consideration of the frequency of union of the elements’ failure
events ([41,25]) for MCM.

Sea state Simulation number Sea side armor failure Toe failure Rear side armor failure System failure

1 1 0 1 1
2 0 0 0 0
3 0 1 1 1
. . . . .
. . . . .
. . . . .
N 1 0 0 1

Total Ns Nt Nr Nf

Table 2
Estimation of a series system’s failure probability via consideration of the probability of union of the elements’ failure
events for DIM.

Load parameters’ joint probability Sea side armor failure Toe failure Rear side armor failure System failure

p1 p1 0 p1 p1
p2 0 0 0 0
p3 0 p3 p3 p3
. . . . .
. . . . .
. . . . .
pn pn 0 0 pn

Total=1 pfs pft pfr pf

5. Case study

5.1. Input data and structural characteristics

The methodology presented in this study is applied to measured long-term historical wave data of Hs , Tm , and θ m that
cover the time period from 2010 to 2017 obtained from an oceanographic buoy. The latter is located in intermediate water
depth of about 15 m (36.69° N, 4.42° W), near the port of Malaga (Spain) in the Mediterranean Sea. Wave data were taken
hourly, while missing values have been excluded from the analysis. Furthermore, long-term sea level measurements have
also been used covering the period 1992–2019 obtained from a tide gage located in the port of Malaga (36.71° N, 4.42° W).
The recording interval is 1 h and the sea level data is divided by the gage in two components, i.e. the astronomical (tide)
TL and the meteorological component SL (e.g. storm surge). The locations of these two stations (buoy and tide gage) are
depicted in Fig. 2.
Regarding the statistical correlation between wave parameters obtained from the buoy, the Pearson correlation coefficient
between Tm and Hs is estimated at 0.68, given as an initial presentation of their correlation, while directional effects on the
joint pdf of Tm and Hs were also noticed. The scatter diagram of Tm and Hs measurements is depicted in Fig. 3.
The Pearson correlation coefficient between TL and SL in the case of Malaga is estimated at 0.001; thus they can be
considered uncorrelated. Furthermore, the Pearson correlation coefficient between TL and Hs , and between SL and Hs , are
about 0.01 and 0.1, respectively. Despite that SL and Hs seem to be more correlated compared against the other two variables
with Hs , this correlation is not considered significant, and thus will not be considered in this study. The histograms of TL
and SL as measured by the tide gage are presented in Fig. 4. As seen in Fig. 4, the case of Malaga refers to a micro-tidal and
storm surge sea, thus wave parameters are more critical than the sea level variations, for the design of a coastal structure
off Malaga.
The structure under design is a rubble mound breakwater supposed to be located at the depth of 6 m. The breakwater’s
structural characteristics are presented in Table 3 and a typical cross section of the structure is depicted in Fig. 5. The
permeability parameter used in limit state functions was 0.40 according to Van der Meer [37], while the damage parameters
S and NOD correspond to initial damage for both seaside and rear-side armor, and toe stability, respectively [9].

5.2. Define design requirements and the fault tree

In this step, the design requirements should be defined, as well as the fault tree that should include all the fundamental
individual failure modes. Specifically, the design lifetime, the design return period of the load variables of the structure
(when referring to the deterministic design), the allowable failure probability during the structure’s lifetime and the design
limit state of the structure should be determined. Since a rubble mound breakwater with armor slopes commonly does
not fail immediately after a storm, but the damage progresses gradually during its lifetime, the design Limit State here is
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 721

Fig. 2. Locations of the buoy (red point) and tide gage (yellow point) near the port of Malaga (Spain) http://www.puertos.es/en-us/oceanografia/Pages/
portus.aspx. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3. Scatter diagram of Tm and Hs measurements obtained from the buoy.

Fig. 4. Histograms of TL and SL as measured by the tide gage. (For interpretation of the references to colour in this figure legend, the reader is referred to
the web version of this article.)
722 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

Table 3
Structural characteristics of a specific rubble mound breakwater.

Parameter Value

Seaward slope 1:2

Freeboard at the seaward side (m) 2.0
Water depth in front of the structure’s toe (m) 6.0
Water depth above the toe (m) 4.8
Crest width (m) 4.0
Rear side slope 1:2
Freeboard at the rear side (m) 2.0
Permeability parameter P 0.4
Damage level S 2
Damage level NOD 0.5

Fig. 5. A typical cross section of the rubble mound breakwater.

selected to be the Serviceability Limit State (SLS). Thus, the design rubble mound breakwater in the present paper will
be checked for initial damage to its elements and to the total system. Besides, the design lifetime of the rubble mound
breakwater is selected to be 50 years, and two design return periods of 10 and 100 years concerning time dependent wave
parameters are selected for the deterministic design.
As far as the fault tree is concerned, a series fault tree is used, as is usually the case for coastal structures [9], that
consists of three individual main failure modes, i.e. the sea side armor failure, the toe instability, and the rear side armor
failure due to overtopping. The expressions used for the aforementioned structural elements have been developed by [37,42],
and [43], respectively. Regarding the sea side rock armor stability, after conducting a comparative analysis on three stability
formulas, i.e. [37]; modified Van der Meer by Van Gent et al. [44]; the new formula by Van Gent et al. [44], the former was
selected based on the design constraints recommended in the Rock Manual [40].
It is also worth mentioning that the effect of oblique wave attack on these three elements’ stability can be taken into
consideration via the use of the equivalent normal Hs proposed by Galland [18].

5.3. Identification of storm events at the buoy location

An initial threshold u1 was selected for Hs distinguishing weak and strong storms from the dataset. Besides, a minimum
calm period of 6 hr [26] between two consecutive storms was selected to ensure that the two events are independent result-
ing in reducing storm events from 83 dependent to 66 independent events in this dataset. Then, the dataset of the maximum
significant wave heights Hs,p throughout the storms is derived from this procedure. In this study, the initial threshold u1 is
selected equal to 2 m, which corresponds to the 95-quantile of Hs data. The histograms and the scatter diagram of duration
and Hs,p of independent events exceeding this initial threshold 2 m are shown in Fig. 6. Furthermore, the histograms and
the scatter diagrams of storm mean wave direction and Hs,p and that of storms mean wave period and Hs,p are presented
in Figs. 7 and 8, respectively. As shown in Fig. 8, all storms mean wave direction fall within the sector (100° to 125°). The
most severe storm has a mean wave direction of 113°.

5.4. Transferring storm events from the buoy to the structure’s location

The sea bottom is considered plane with a mean slope of 5% in deep and intermediate waters. Thus, sea level and wave
data need to be transferred to the structure’s water depth. In contrast to deep waters, where wave parameters are not
affected by water depth and sea level variations, in intermediate water depths the latter two do affect wave parameters. In
Fig. 9, a comparison is made between the scatter diagrams of storm duration and Hs,p at the buoy (depth of 15 m) and the
structure (depth of 6 m) location, and in Fig. 10 the histograms and the scatter diagrams of storm mean wave direction and
Hs,p are presented. As shown in Fig. 10, the most severe storm has a mean wave direction of 113°, which is normal to the
shoreline, as in the buoy location.
The next step is the selection of another threshold u2 higher than u1 . Judging from the two methods results for the
selection of the most appropriate GP parameters and threshold value, the value of 2.409 m is selected for threshold u2 . The
parameter λe that corresponds to this value of u2 is 3.125, i.e. approximately 3 extreme events per year.
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 723

Fig. 6. Histograms and scatter diagram of Hs,p and storm duration at the buoy location.

Fig. 7. Histograms and scatter diagram of Hs,p and storm Tm at the buoy location.

Fig. 8. Histograms and scatter diagram of Hs,p and storm θ m at the buoy location.

The return periods and values for Hs,p associated with the most frequent values of their covariate parameters at the
structure’s location selected for the preliminary design are presented in Table 4.
As noted above, the expected duration of the design storm mDdesign storm is considered here equal to the mean duration of
storms, whose Hs,p exceeds u2 . In the case of Malaga, mDdesign storm is equal to 23.21 hr.

5.5. Preliminary design based on the design storm events

Since the input data for the preliminary design have been estimated, the median mass of the rock units of the structure’s
elements (i.e. seaside armor, rear side armor and toe) will be computed at this step. Median mass M50 is defined [40] as
the mass of the theoretical block for which half of the mass of the sample is lighter. Some other deterministic parameters
724 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

Fig. 9. Scatter diagram of storm duration and Hs,p at the buoy (depth of 15 m) and structure (depth of 6 m) location.

Fig. 10. Histograms and scatter diagram of Hs,p and storm θ m at the structure location.

Table 4
Return periods and values for Hs,p associated with the most probable values of the covariate pa-
rameters at the structure’s location (depth of 6 m) selected for the preliminary design.

Return Period Tr for Hs,p (years) Return value - Hs,p (m) storm Tm (s) storm θ m (deg)

10 4.01 6.25 113

100 4.81 6.58 113

Table 5
Median mass M50 of the elements’ units estimated by the preliminary design.

Tr for Hs,p (years) Seaside armor units (kg) Rear-side armor units (kg) Toe units (kg)

10 7446 1516 373

100 12,151 2679 660

that have not been specified yet are the rock density ρ s , and the sea water density ρ w , which are considered equal to
2.650 t/m3 and 1.025 t/m3 , respectively.
As for the structure’s orientation, judging from Fig. 10, the structure’s perpendicular axis has been selected to have a
direction of 113° so as to protect the shoreline from the most extreme storm.
The results of the preliminary design after using the design formulas, i.e. by Van der Meer [37]; Van der Meer [42]; Van
Gent and Pozueta [43], for seaside armor, toe, and rear side armor, for initial damage are shown in the following Table 5.
The conditional and unconditional equivalent failure probabilities per annum (estimated by Eq. (18)) and during structure’s
Lifetime (estimated by Eq. (14)), associated to the return periods of Hs,p at the depth of 6 m are displayed in Table 6.
However, in the above analysis, the variability of resistance parameters and of sea level variation, and the system’s prob-
ability of failure have not been taken into account, in contrast to fully probabilistic design methodologies that will follow
next.
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 725

Table 6
Conditional and unconditional equivalent failure probabilities per
annum and over structure’s Lifetime (50 years), associated to return
period of Hs,p at the depth of 6 m (Malaga).

Tr for Hs,p (years) Pex,storm, 1y Pf,L

10 1.14×10–4 6.60×10–3
100 1.14×10–6 6.62×10–4

Fig. 11. Histograms and scatter diagram of Hs and θ m of sea states at the buoy location.

5.6. Statistical analysis of sea states conditions at the buoy location

In this section, the statistical analysis of the sea state conditions within storm events was carried out to be considered as
wave loadings upon the structure. In this application, a simplification has been made concerning the time step concerning
the representation of sea conditions in the reliability analysis of the rubble mound breakwater. Particularly, this time step is
initially considered equal to 1 hr, and then a constraint is applied in order that the number of waves per sea state should
not be lower than 10 0 0 waves, as suggested by a commonly used limit state function [37]. In case when the number of
waves of a sea state is lower than 10 0 0, a new time step is applied corresponding to 10 0 0 waves. This limitation usually
refers to sea states with large wave periods which tend to be associated with large wave heights that are not probable to
occur but only as part of a storm event of longer duration.
Furthermore, as noticed in Fig. 11, a threshold was applied to Hs data at the buoy location to filter the most significant
sea states that will be transferred to the structure location. In this way, not only the most critical sea states are distin-
guished from the total sample, but also the amount of data to be transferred to the structure location has been significantly
reduced. It is noteworthy that the relative frequency of the threshold exceedance (Hs >1.65 m in this case) was equal to
6.10%, i.e. slightly higher than that applied for identification of storm events in order to avoid significant reduction of data
(see Section 5.3). It is noted that the same threshold to that applied for the identification of storm events could be also
applied in the sea state approach when the available data exceeding threshold is sufficient enough for the derivation of the
appropriate probability distribution.
In Fig. 11 the histograms and scatter diagram of Hs and mean wave direction of sea states at the buoy locations are
depicted. Furthermore, by comparing Figs. 8 and 11, the directions of the most extreme storms and severe sea state both
range between 100 and 125 o from North at the buoy location.

5.7. Transferring the reduced sample of sea states from the buoy to the structure location

By applying the same wave propagation model, the subset of sea states is transferred to the structure location (at the
depth of 6 m) and long-term wave statistics are extracted concerning this location. The histograms and scatter diagram of
Hs and mean wave direction of the propagated subset of sea states at the depth of 6 m are depicted in Fig. 12.
The results depicted in Fig. 12, suggest that the mean wave direction of subset of sea states selected for the design is
concentrated in a rather narrow range of directions, and actually the same with that of storms, as it was expected. Therefore,
in this case, a simplification could be made by considering that all sea states of this subset have a mean wave direction that
is perpendicular to the structure, i.e. wave attack is considered normal for all them. It is noted that this simplification should
be avoided in cases of wide range of mean wave direction of sea states and storms.
By applying, thus, the conditional model for only two variables, i.e. Hs and Tm , the conditional joint pdf of Hs and Tm
given that sea states belong to the subset selected previously are presented in Fig. 13 concerning the buoy and the structure
location. As seen in Fig. 13, the conditional joint pdf of Hs and Tm at the structure location differs from that at the buoy
location and is more depth-limited, despite that both locations belong in intermediate water depths.
726 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

Fig. 12. Histograms and scatter diagram of Hs and θ m of the subset of sea states at the structure location.

Fig. 13. Conditional joint pdf of Hs and Tm given that sea states belong to the subset selected concerning the buoy (1) and the structure location (2)
location (isoprobability density contours step equal to 0.05 1/s/m).

Fig. 14. Histograms of sea level variation from SWL due to storm surge and tide estimated at the structure location.

5.8. Sea level parameters at the structure location

As for the sea level data, they have been transferred to the buoy location by considering sea level variations due to tide
and storm surge as long waves, undergone only to shoaling effects. The histograms of sea level variation from still water
level (SWL due to storm surge and tide estimated at the structure location are displayed in Fig. 14.
It is noted that storm surge total sample can well be represented by a normal pdf, with a mean value of −0.0015 m and
standard deviation of 0.0712 m.

5.9. Assessing failure probability via DIM and MCM

Since one of the objectives of the present paper was to investigate the effect of considering more and different types of
parameters as random variables on the assessed failure probability of the structure, three cases were examined here; firstly
only wave parameters, secondly wave and sea level parameters, and finally wave, sea level parameters and one resistance
parameter were considered as random variables. It is noted that, in the first case, SL and TL were considered equal to zero.
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 727

Fig. 15. Comparison of the random sample generated by Monte Carlo simulation (red) with the transferred reduced data sample (green) to the structure’s
location (isoprobability density contours step equal to 0.05 1/m/s). (For interpretation of the references to colour in this figure legend, the reader is referred
to the web version of this article.)

In the third case, the rock mass density was treated as random variable. Because its distribution is not known, a normal
distribution was used in order to predict its variability. The mean value of this normal distribution is equal to the one used
when this variable was regarded as a deterministic one, i.e. 2650 kg/m3 , while the standard deviation is considered here
equal to 10 kg/m3 , which could be a reasonable variability of quarry stones density.
It is noted that in order to simplify the computations, and because the case of Malaga refers to a micro tidal and storm
surge sea, the functional dependence between wave parameters and the two components of sea level TL, SL at the struc-
ture’s location has not been taken into consideration in this study. To clarify this statement, this could be described by the
following equation:
fTL,SL,Hs ,Tm (tl, sl, Hs , Tm ) = fTL (tl ) fSL (sl ) fHs ,Tm |(TL,SL) (Hs , Tm |(tl, sl ) ) (27)
where:
f Tm ,Hs |(TL, SL) (Tm , Hs |(tl, sl ) ) = fTm ,Hs (Tm , Hs ) (28)

The right-hand side of Eq. (28) has been estimated for t l = t l , sl = sl . The accurate estimation of f Tm ,Hs (Tm , Hs |(tl, sl )), for
every tl and sl value, is preferable than the estimation of the right hand side of Eq. (28), especially in seas with significant
tide and storm surge variability.
As for the evaluation of Eq. 25 in this study via DIM, δ tl and δ sl were kept equal to 0.05 m, and δ Hs and δ Tm equal to
0.10 m and 0.10 s, respectively, which were considered satisfactorily small. Further reduction of the value of the steps did
not seem to alter significantly the results by DIM, but contrariwise did increase the computational cost considerably.
In MCM, a random sample was generated for each variable that followed its pdf, but also at maintaining the statistical
correlation between correlated variables, e.g. Tm and Hs . In Fig. 15, the scatter diagram of Tm and Hs transferred at the
structure location, their joint pdf, estimated via the conditional model, and the MC sample are depicted, showing good
agreement.
The selection of the total number of simulations for each variable depends on its variability and its contribution to the
structure’s failure. For example, in this study the total number of simulations for waves was 10,0 0 0, while for SL and TL
50, and 20 for resistance parameters, which results in 5 × 108 iterations. From the above, it is seen that the computational
demands of MCM simulation can become very high.

5.10. Results and discussion

The estimated failure probabilities derived from the preliminary design based on the design storm events at the struc-
ture’s location is compared with those estimated by DIM for the first case, since both approaches consider only wave pa-
rameters as random variables (see Table 7). As seen in Table 7, the results by the preliminary design based on the design
storm events are close to those by DIM for case 1. This could be attributed to the fact that both approaches have been trans-
formed to estimate the same kind of failure probability, i.e. an unconditional one. However, some differences between the
two methods’ results are noticed. The latter is due to the fact that the fully probabilistic method DIM considers the effect of
a significantly larger number of sea conditions on the structure’s failure, and the statistical correlation of individual failure
modes, compared against the extreme event-based preliminary design, which only calculates the exceedance probability of
the design storm. Also, the expected duration of the design storm is also a random variable and its consideration as the
mean duration of storms exceeding GP threshold might not be adequately precise.
728 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

Table 7
Unconditional probability of failure during L = 50 yr derived from the extreme
event-based preliminary design and the fully probabilistic reliability method
(Direct Integration) for the first case.

Return Period (years) for Preliminary design based DIM (Case 1)

Hs,p (years) on the design storm events

10 6.60×10–3 6.40×10–3
100 6.62×10–4 6.88×10–4

Table 8
Unconditional probability of failure during L = 50 yr by DIM for the first case based on the
time step of 1 hr and wave number limitation and on the time step of 3 hr.

Return Period (years) Preliminary design DIM (Case 1) (1hr DIM (Case 1) (3 hr)
for Hs,p (years) based on the design & min 10 0 0 waves)
storm events

10 6.60×10–3 6.40×10–3 6.70 ∗ 10–3

100 6.62×10–4 6.88×10–4 7.07×10–4

Table 9
System’s unconditional probability of failure during L = 50 yr derived from fully
probabilistic reliability method (Direct Integration) for the three cases considered.

Return Period (years) for Hs,p (years) (Case 1) (Case 2) (Case 3)

10 6.40×10–3 7.20×10–3 7.20×10–3

100 6.88×10–4 7.51×10–4 7.50×10–4

It is noted that the event-based extreme value analysis was applied only to wave data. As noted above, to account for
different components of a single phenomenon, e.g. wave and sea level conditions, a multivariate event-based extreme value
analysis should be applied (see e.g. [35,34]) for the design of a coastal structure. This is highly recommended in cases of
macro-tidal environments.
Besides, an investigation, concerning the time step that denotes the sea state, was also carried out. Numerical tests
showed that within a time-step range of 1hr to 3hr the results are well balanced between statistical accuracy (low values)
and resilience-type requirements (high values). Specifically, the previous results of the 1 hr time step and the limitation of
at least 10 0 0 waves per sea state are compared with those derived from the analysis based on a 3 hr time step, which are
all presented in Table 8. As seen in this table, the results of the two approaches are close to each other but the results
of the 3 hr time step are a little higher than those of the 1 hr time step. This could be attributed to the fact that in the
application of the time step of 3hr, the maximum Hs in a sequence of three hourly measurements of Hs within a storm
is selected to represent the sea state, which results in an overestimation of the sea state’s contribution to the final result.
This is due to the fact that the maximum wave height may not be present during the entire time step; this discrepancy
obviously increases with increasing time step. The time step of 3 hr, however, enables working with sea states of equal
duration without the need to apply the condition of 10 0 0 waves in most cases. Nevertheless, since the results estimated by
the two approaches are quite similar, the following analysis continues by adopting the 1 hr time step and the limitation of
at least 10 0 0 waves per sea state, which avoids the abovementioned overestimation.
Moreover, DIM results for the three cases examined are presented in Table 9. According to Table 9, the probability of fail-
ure of the structure is higher in the second case, when the sea level parameters have been considered as random variables,
than in the first case. This could be avoided, if the structure had been designed for a higher tide and storm surge value.
This optimization is recommended for the design of the structure.
Also, the variability of environmental loadings and especially those with extremal behavior, such as Hs and Tm , has a
greater impact on the structure’s failure probability than that of resistance parameters, since the latter have been considered
to have a zero hazard rate. This has been verified by case 3, whereby the failure probabilities estimated in this case are
almost the same with those in case 2 (when resistance parameter is treated deterministically). This could be explained by
the fact that in the case 2 computations have been made considering the mean value of the resistance parameter, and in
case 3 considering a symmetrical distribution, i.e. a normal pdf, with the same mean value.
In Table 10, a comparison between DIM and MCM results for case 3 is presented. Particularly, both methods’ results
are close, with MCM probabilities slightly higher than those of DIM. This could be attributed to the difficulty in MCM
solution convergence by altering the total number of simulations. The latter might lead to the conclusion that DIM is more
appropriate for structural reliability assessment than MCM. However, the problem examined here might be quite simple and
simulation-based methods could give more efficient solutions in more complex cases.
It is noted that the unconditional annual probabilities by DIM and MCM were estimated from conditional annual proba-
bilities via total probability law.
 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730 729

Table 10
System unconditional failure probability within a Lifetime of 50 years esti-
mated by DIM and MCM (for case 3).

Return Period (years) for Hs,p (years) DIM (Case 3) MCM (Case 3)

10 7.20×10 –3
7.88×10–3
100 7.50×10–4 8.01×10–4

Table 11
Median mass M50 of the elements units and system failure probability within a Lifetime of 50 years estimated by
DIM for the design of the structure.

Solution Seaside armor units (kg) Rear-side armor units (kg) Toe units (kg) Pf,L

Initial solution 12,151 3652 1303 1.14×10–4

Optimized solution 12,758 3834 1310 9.98×10–5

Finally, the step of optimization (see Fig. 1) based on economic criteria can then be made by applying the fully prob-
abilistic approach for different solutions (i.e. by altering the nominal mass of the elements’ units, geometric parameters
etc.). This step is not included in the present paper but is necessary for the design of a coastal structure in order to meet
predefined safety and cost requirements.

5.11. Design of the structure

Regarding the design of the rubble mound breakwater (see Fig. 5), the first step is to define a target failure probability,
considered in the present paper equal to 10–4 . As can be seen, this is close to the second solution that correspond to the
return period of 100 years for Hs,p . The values of SL and TL examined for the design are the return values corresponding to
100 years for SL (SL-100 yr) and the maximum TL (max TL) at the structure’s location. Therefore SL-100 yr is estimated at
0.45 m, similarly to Hs,p via POT method, and max TL is equal to 0.48 m, both above the Still Water Level (SWL) of 6 m.
Therefore, by using SL-100 yr, max TL and Hs,p -100 yr associated to the most probable values of Tm and θ m , presented
in Table 4, the median mass of rock armor units of the elements of the structure and the corresponding failure probability
are presented in the first row of Table 10, as extracted from DIM. In this case, wave, sea level parameters and the resistance
parameter rs were considered as random variables.
In the second row of Table 11 the optimized solution selected for the design of the structure’s elements is presented
resulting in a failure probability Pf,L slightly lower than the target Pf,L . The optimized solution has been derived from a
simple optimization technique regarding the target reliability Pf,L , i.e. excluding any economic and social costs.

6. Conclusions

Some significant issues concerning the reliability assessment of coastal structures have been dealt with, in the present
paper. In particular, it was shown that the failure probability of a coastal structure, based on the long-term wave climate at
a coastal structure’s location using wave observations or measurements in deeper waters, could be estimated by applying
a wave propagation statistical model. Moreover, a relation between an event-based extreme value analysis and an analysis
of sea-state conditions within storms was derived and both of these two approaches provided the same order and kind of
failure probability; the latter was properly defined here as a proportion of the structure’s failure period over its Lifetime.
These two approaches were included in the design of a coastal structure.
Furthermore, two different fully probabilistic methods, DIM and MCM, were applied giving satisfactorily similar results by
using a combination of variables with zero and non-zero hazard rate, referred here as a combined time-invariant and time-
variant analysis. Regarding the case of the ruble mound breakwater examined here, DIM was a more straightforward method,
since it was applied directly to the joint pdf of the random variables, while MCM simulations were more time consuming,
since the total number of iterations required was really huge. Besides, some difficulty in MCM solution convergence was
met. Nevertheless, DIM seemed to be more appropriate in the problem examined here, this case might be considered quite
simple, and simulation-based methods could give more efficient solutions in more complex cases.
In short, the problem of assessing failure probability of coastal structures was presented and attempted to be solved by
giving application examples, a flowchart, and mathematical equations that describe the procedure, focusing on the case of
rubble mound breakwaters.

Acknowledgements

The wave and sea level measurements near the port of Malaga have been kindly provided by Puertos del Estado (Spain).

References

[1] R.E. Melchers, Structural Reliability Analysis and Prediction, Ellis Horwood Limited, UK., 1987.
730 D.I. Malliouri, C.D. Memos and T.H. Soukissian et al. / Applied Mathematical Modelling 89 (2021) 710–730

[2] S.R.K. Koyluoglu, H.U. Nielsen, New Approximations for SORM Integrals, Struct. Saf. 13 (4) (1994) 235–246.
[3] S.R. Winterstein, T.C. Ude, C.A. Cornell, P. Bjerager, S. Haver, Environmental parameters for extreme response: inverse FORM with omission factors, in:
Proc. 6th international conference on structural safety and reliability, 1993.
[4] S. Haver, S. Winterstein, Environmental contour lines: a method for estimating long term extremes by a short term analysis, Trans. Soc. Nav. Archit.
Mar. Eng. 116 (2009) 116–127.
[5] E. Vanem, Environmental contours for describing extreme ocean wave conditions based on combined datasets, Stoch. Environ. Res. Risk Assess. 33
(4–6) (Jun. 2019) 957–971.
[6] K. Box, G.E.P. Wilson, On the experimental attainment of optimum conditions, J. R. Stat. Soc. Ser. B 13 (1951) 1–45.
[7] U. Bucher, C. Bourgund, A fast and efficient response surface approach for structural reliability problems, Struct. Saf. 7 (1990) 57–66.
[8] P.O. Hristov, F.A. DiazDelaO, K.J. Kubiak, U. Farooq, Adaptive emulation-based reliability analysis, in: Proceedings of the 2nd International Conference
on Uncertainty Quantification in Computational Sciences and Engineering (UNCECOMP 2017), 2017, pp. 198–211.
[9] U.S. Army Corps of Engineers, Coastal Engineering Manual. Washington, D.C., 2006.
[10] Z. Liu, H.F. Burcharth, Design Wave Height Related to Structure Lifetime, Coastal Engineering 1996 (1997) 2560–2572.
[11] PIANC, “Analysis of rubble mound breakwaters. Report of working group 12 (PTC II). Supplement to Bulletin no. 78/79.,” Brussels, 1992.
[12] P.A. Madsen, O.R. Sørensen, A new form of the Boussinesq equations with improved linear dispersion characteristics. Part 2. A slowly-varying
bathymetry, Coast. Eng. 18 (3–4) (Dec. 1992) 183–204.
[13] O.R. Sørensen, H.A. Schäffer, L.S. Sørensen, Boussinesq-type modelling using an unstructured finite element technique, Coast. Eng. 50 (4) (Feb. 2004)
181–198.
[14] M.K. Chondros, C.D. Memos, A 2DH nonlinear Boussinesq-type wave model of improved dispersion, shoaling, and wave generation characteristics,
Coast. Eng. 91 (Sep. 2014) 99–122.
[15] C.L. Bretschneider, Forecasting relations for wave generation, Look Lab 1 (3) (1970).
[16] D. Malliouri, V. Afentoulis, V. Bonaros, C. Memos, V. Tsoukala, Integrated Design of Coastal Structures and Adaptation to Climate Change Hazards, CEST
(2017).
[17] D.I. Malliouri, C.D. Memos, N.D. Tampalis, T.H. Soukissian, V.K. Tsoukala, Integrating short- and long-term statistics for short-crested waves in deep
and intermediate waters, Appl. Ocean Res. 82 (3) (2019) 346–361.
[18] J.C. Galland, Rubble mound breakwater stability under oblique waves: and experimental study, in: Coastal Engineering Proceedings, 1994.
[19] M.R.A. Van Gent, Oblique wave attack on rubble mound breakwaters, Coast. Eng. 88 (Jun. 2014) 43–54.
[20] P. Bernardara, F. Mazas, X. Kergadallan, L. Hamm, A two-step framework for over-threshold modelling of environmental extremes, Nat. Hazards Earth
Syst. Sci. 14 (3) (. 2014) 635–647.
[21] C.D. Memos, K. Tzanis, Numerical results of the joint probability of heights and periods of sea waves, Coast. Eng. 22 (3–4) (Feb. 1994) 217–235.
[22] C.D. Memos, K. Tzanis, Joint Distribution of Wave Heights and Periods in Waters of Any Depth, J. Waterw. Port, Coastal, Ocean Eng. 126 (3) (May 20 0 0)
162–172.
[23] J.A. Battjes, Long-term wave height distribution at seven stations around the British Isles, Hydrogr. Zeitschrift 24 (4) (1972) 180–189.
[24] D.I. Malliouri, Probabilistic design of coastal structures, National Technical University of Athens.
[25] P.S. Everts, Probabilistic Design of a rubble mound breakwater, Delft University of Technology (2016).
[26] F. Li, P.H.A.J. van Gelder, D.P. Callaghan, R.B. Jongejan, C. de. Heijer, R. Ranasinghe, Probabilistic modeling of wave climate and predicting dune erosion,
J. Coast. Res. 65 (Jan. 2013) 760–765.
[27] F. Mazas, L. Hamm, A multi-distribution approach to POT methods for determining extreme wave heights, Coast. Eng. 58 (5) (May 2011) 385–394.
[28] J. Pickands, Statistical Inference Using Extreme Order Statistics, Ann. Stat. 3 (1) (1975) 119–131.
[29] P. Embrechts, C. Klüppelberg, T. Mikosch, Modelling Extremal Events, Springer Berlin Heidelberg, Berlin, Heidelberg, 1997.
[30] P. Embrechts, C. Klüppelberg, T. Mikosh, Modelling Extremal Events For Insurance and Finance, Springer, Berlin, 2003.
[31] S. Coles, An Introduction to Statistical Modeling of Extreme Values, Springer London, London, 2001.
[32] M.R. Leadbetter, Extremes and local dependence in stationary sequences, Zeitschrift for Wahrscheinlichkeitstheorie und Verwandte Gebiete 65 (2)
(1983) 291–306.
[33] H.F. Burcharth, P. Frigaard, “Ciervana Breakwater, Puerto Autonorne de Bilbao: model test results.,” 1992.
[34] F. Mazas, L. Hamm, Extreme Meteo-Oceanic Events, Coast. Eng. (2018).
[35] T. Wahl, N.G. Plant, J.W. Long, Probabilistic assessment of erosion and flooding risk in the northern Gulf of Mexico, J. Geophys. Res. Ocean. 121 (5)
(May 2016) 3029–3043.
[36] A. Kiriliouk, H. Rootzén, J. Segers, J.L. Wadsworth, Peaks Over Thresholds Modeling With Multivariate Generalized Pareto Distributions, Technometrics
61 (1) (Jan. 2019) 123–135.
[37] J.W. Van der Meer, Rock Slopes and Gravel Beaches Under Wave Attack, Delft University, the Netherlands, 1988.
[38] J. Mathisen, E. Bitner-Gregersen, Joint distributions for significant wave height and wave zero-up-crossing period, Appl. Ocean Res. 12 (2) (Apr. 1990)
93–103.
[39] S. Haver, Wave climate off northern Norway, Appl. Ocean Res. 7 (2) (Apr. 1985) 85–92.
[40] CIRIA; CUR; CETMEF, The Rock Manual, The Use of Rock in Hydraulic Engineering, 2nd edition, CIRIA, London, 2007 C683.
[41] D.I. Malliouri, C.D. Memos, T.H. Soukissian, V.K. Tsoukala, Reliability analysis of rubble mound breakwaters - An easy-to-use methodology, DMPCO
2019 (2019).
[42] J.W. Van der Meer, K. d’ Angremond, E. Gerding, Toe Structure Stability of Rubble Mound Breakwaters, Proc. Coastlines, Marine Structures and Break-
water (1995) 308–321.
[43] M.R.A. Van Gent, B. Pozueta, Rear-side Stability of rubble mound structures, Coastal Engineering 2004 (2004) 3481–3493.
[44] M.R.A. Van Gent, A.J. Smale, C. Kuiper, Stability of Rock Slopes with Shallow Foreshores, in: Proceedings of 4th International Coastal Structures Con-
ference, 2004.