See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327691570

A Copula-Based Bayesian Network for Modeling Compound Flood Hazard

from Riverine and Coastal Interactions at the Catchment Scale: An
Application to the Houston Ship Channel, Texas

Article in Water · September 2018

DOI: 10.3390/w10091190

CITATIONS READS

74 253

3 authors:

Anaïs Couasnon Antonia Sebastian

Vrije Universiteit Amsterdam Delft University of Technology
29 PUBLICATIONS 854 CITATIONS 41 PUBLICATIONS 1,481 CITATIONS

SEE PROFILE SEE PROFILE

Oswaldo Morales Napoles

Delft University of Technology
112 PUBLICATIONS 1,785 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

RISK ANALYSIS OF INFRASTRUCTURE NETWORKS IN RESPONSE TO EXTREME WEATHER View project

Compound flood events View project

All content following this page was uploaded by Oswaldo Morales Napoles on 12 July 2019.

The user has requested enhancement of the downloaded file.

 water
Article
A Copula-Based Bayesian Network for Modeling
Compound Flood Hazard from Riverine and Coastal
Interactions at the Catchment Scale: An Application
to the Houston Ship Channel, Texas
Anaïs Couasnon 1,2, * ID
, Antonia Sebastian 2 ID
and Oswaldo Morales-Nápoles 2 ID

1 Institute for Environmental Studies, Faculty of Science, Vrije Universiteit Amsterdam, De Boelelaan 1087,
1081 HV Amsterdam, The Netherlands
2 Faculty of Civil Engineering and Geosciences, Delft University of Technology, Stevinweg 1,
2628 CN Delft, The Netherlands; [email protected] (A.S.);
[email protected] (O.M.-N.)
* Correspondence: [email protected]




Received: 3 August 2018; Accepted: 30 August 2018; Published: 4 September 2018


Abstract: Traditional flood hazard analyses often rely on univariate probability distributions;
however, in many coastal catchments, flooding is the result of complex hydrodynamic interactions
between multiple drivers. For example, synoptic meteorological conditions can produce considerable
rainfall-runoff, while also generating wind-driven elevated sea-levels. When these drivers interact in
space and time, they can exacerbate flood impacts, a phenomenon known as compound flooding.
In this paper, we build a Bayesian Network based on Gaussian copulas to generate the equivalent
of 500 years of daily stochastic boundary conditions for a coastal watershed in Southeast Texas.
In doing so, we overcome many of the limitations of conventional univariate approaches and are able
to probabilistically represent compound floods caused by riverine and coastal interactions. We model
the resulting water levels using a one-dimensional (1D) steady-state hydraulic model and find that
flood stages in the catchment are strongly affected by backwater effects from tributary inflows and
downstream water levels. By comparing our results against a bathtub modeling approach, we show
that simplifying the multivariate dependence between flood drivers can lead to an underestimation
of flood impacts, highlighting that accounting for multivariate dependence is critical for the accurate
representation of flood risk in coastal catchments prone to compound events.

Keywords: flood risk; copula; compound events; multivariate; storm surge; spatial dependence;
Bayesian Network

1. Introduction
Coastal cities suffer from extreme flood risks, being both exposed to multiple hazard types and
gathering high asset values and critical infrastructure [1]. They experience different sources of flooding
driven by the interaction of oceanographic, hydrological, geological, and meteorological processes [2].
It was estimated that in 2000, about 10% of the global population (625 million people) lived in the area
below 10 m of elevation with a direct connection to the sea [3]. More specifically, port cities fostering
rapid socio-economic development are flood risk hotspots. Hallegatte et al. [4] estimated more than
an eight-fold increase in global economic average annual flood losses for major coastal port cities
(from US$6 billion in 2005 to US$52 billion by 2050) due to socio-economic changes alone suggesting
that the risk of compound flooding, especially in those areas, will grow over time. Thus, detailed flood
probability estimates are critically needed for future coastal resilience.

Water 2018, 10, 1190; doi:10.3390/w10091190 www.mdpi.com/journal/water

Water 2018, 10, 1190 2 of 19

Modeling flood hazards in coastal catchments is a complex undertaking. Synoptic meteorological

conditions can generate both considerable rainfall and anomalously high coastal water levels.
For example, Atlantic cyclones have been well documented as causing high surge levels and heavy
precipitation [5–10]. The importance of this interaction in determining the flood level and magnitude
in coastal catchments is dependent on the storm and the catchment properties [11]. Large catchments
typically exhibit a lag of several days between high sea levels and high river levels such that the
flood events may not physically interact depending on their magnitude and duration. For example,
the Rhine delta experiences on average a lag of about six days between high sea levels at the coast
and high-water level at Lobith in the Netherlands [12]. In contrast, smaller catchments often have
riverine flood wave travel times of less than a day [13]. If co-occurring with a coastal storm surge event,
the high river discharges are obstructed by the anomalously high sea-level which may exacerbate the
flood stage in the coastal catchment. The joint occurrence of both riverine and coastal floods leading to
an extreme impact, here the flood level, is referred to as compound flooding [14–16].
Despite increasing scientific recognition of compound flooding [14], regulatory flood hazard
maps seldom systematically include the impacts of riverine and coastal interactions [17]. One typical
approach used in the US is to consider these two processes separately and merge the independently
obtained coastal and riverine flood hazard maps by combining the respective rate of occurrence of
a given modeled water surface elevation [17,18]. In doing so, the dependence between these two
flood processes is neglected, providing a simplified representation of the flood extent and thereby
the flood hazard characterization. Continental to global scale coastal or riverine studies often model
flooding due to one flood driver only [19–24]. River discharge and storm surge interactions are
highly dependent on the geometry of the cross-sections and the slope of the river bed, generating
nonlinear responses in the flood extent and amplitude along the river [25]. Several local studies have
highlighted the importance of modeling compound flooding to better characterize flood hazard in
coastal catchments [26–29]. For example, Kumbier et al. [27] found that neglecting river discharge
when modeling the June 2016 storm event in the Shoalhaven Estuary (Australia) underestimated the
flood extent by 30%. As they only focus on one event, these modeling scenarios do not allow further
analysis on the impact of compound flood for flood hazard analysis.
To better capture the behavior of compound events, previous researchers have focused on the
statistical dependence between river and coastal floods since it is inherently linked with the joint
probability analysis [30–32]. The strength of the statistical dependence between flood hazards is
usually quantified by means of proxy variables [16,33]. Zheng et al. [34] pointed out that even a weak
dependence between flood drivers can have significant implications for estimating the flood hazard.
Ward et al. [35] used non-parametric correlation coefficients and copula structures to quantify the
statistical dependence between coastal and river floods. Salvadori et al. [36] developed a theoretical
framework to quantify multivariate return periods from copulas according to predefined hazard
scenarios (for a developed and extensive summary, see also Reference [37]). This framework was
also applied by Moftakhari et al. [38] to characterize compound flooding caused by the interactions
between river flows and sea-level rise. However, while these probability and dependence assessments
of compound extremes are valuable, they do not provide direct guidance on the expected impacts
within the catchment. Moreover, flood drivers of different levels which are not themselves extreme
may combine to create an extreme impact [14,39].
To model compound flood events at the catchment scale, multiple boundary conditions are needed
such that a univariate or bivariate analysis alone may not be sufficient to characterize the flood hazard.
A traditional univariate method consists of compartmenting the river network into single river reaches
and modeling flood extent and/or depth at the latter spatial extent [40]. This becomes problematic
when integrating results at the catchment scale since river reaches are physically connected [41,42].
The discharge downstream of a confluence results from the contribution of its upstream tributaries.
A given downstream river discharge may be obtained by a large combination of upstream river flows
which, when modeled, can result in different flood extents. The spatial dependence of discharge
Water 2018, 10, 1190 3 of 19

between tributaries clearly becomes important to identify possible combinations. Consequently,

in order to derive the flood extent and its associated annual exceedance probability, a realistic spatial
dependence structure between tributaries is needed [40,43].
Few studies have quantified the flood hazard while also including multiple flood drivers. Those
that have, analyzed results at point locations within the catchment rather than showing results for
the whole river segment. For example, Lamb et al. [44] represented the joint multivariate distribution
with a multivariate statistical model based on conditional exceedance. Once the statistical model was
built, they were able to simulate large series of synthetic events and compute the damage at selected
locations, using discharge–damage curves. Bevacqua et al. [28] used pair–copula constructions and a
multiple regression model to predict the water level at one selected location along a coastal river reach.
Their results showed that neglecting the dependence between discharge and sea-level underestimated
the flood hazard at that location. However, to the best of our knowledge, this has not been studied for
the whole river section in a probabilistic manner.
This paper introduces an alternative method to model and assesses the impact of compound
flooding from riverine and coastal interactions at the catchment scale while accounting for spatial
dependence between river tributaries. A Bayesian Network (BN) is constructed and used to simulate
a set of synthetic joint boundary conditions needed to model water levels in a coastal catchment
vulnerable to compound flooding. Unlike conventional univariate models, this model integrates the
spatial dependence between river tributaries as well as the dependence between river and coastal
interactions to generate a set of possible boundary conditions. The simulated events are used to force a
one-dimensional (1D) hydraulic model to generate a large number of modeled water levels along the
whole river reach and to estimate the return frequency of extreme water levels at every location within
the hydraulic model domain, allowing for an impact-based analysis of compound flooding.
As a case study, the framework is applied to the Buffalo Bayou catchment in Southeast Texas.
The catchment drains much of the City of Houston—the fourth largest city in the United States
(U.S.)—and encompasses the Port of Houston and the navigational head of the Houston Ship Channel.
In 2016, the Port of Houston ranked first in the U.S. in terms of foreign waterborne tonnage and was
estimated to have generated an economic impact of more than $265 billion USD in Texas alone [45].
It is also home to the second largest petrochemical complex in the world (first in the U.S.). The region
is prone to both heavy rainfall events and storm surge, as demonstrated during recent flood events
including Hurricane Ike (2008) and Hurricane Harvey (2017), and previous researchers have warned
that flooding of the Port of Houston during extreme events could lead to billions of dollars in economic
impact to the national and global economies [46–48]. In the next section, we provide a theoretical
background of BNs and present the framework applied to the selected catchment. This is followed by
the results, a discussion of the benefits and limitations of the method, suggestions for future research,
and conclusion.

2. Materials and Methods

Figure 1 provides an overview of the framework applied in this paper. The BN was constructed
based on daily observations of discharges in the coastal catchment and storm surge at the coast
(box A in Figure 1). The BN model (box B) was extensively sampled to extract the equivalent of
500 years of daily joint realizations of discharges and storm surge (box C). After translating the storm
surge at the coast to the mouth of the catchment and the tributary discharges to the main river
reach (boxes D and E), we used these boundary conditions to force a 1D steady-state hydraulic model
(box F). The modeled water surface profiles were analyzed and estimates of large return periods—up
to a 500-year—were calculated and compared against empirical observations. We then tested the
influence of spurious boundary conditions relationships on the delineation of the 100-year flood stage
to better understand the importance of the multivariate statistical dependence on the water levels in
the river reach.
Water 2018, 10, 1190 4 of 19
Water 2018, 10, x FOR PEER REVIEW 4 of 20

Schematicworkflow
Figure1.1.Schematic
Figure workflowfor
forgenerating
generatingstochastic
stochasticwater
watersurface
surfaceprofiles
profilesin
inaacoastal
coastal catchment.
catchment.

The
The following
followingsection
sectionprovides
providesan an overview
overviewof of the
the theoretical
theoretical background
background ofof the
the probabilistic
probabilistic
model:
model:the
theBayesian
BayesianNetwork
Network(BN).
(BN).AAreader
readerwho
whoisisalready
already familiar
familiarwith
with Bayesian
BayesianNetworks
Networksmaymay
choose
choose to give less attention to the next section and skip ahead to Section 2.2 which introduces the
to give less attention to the next section and skip ahead to Section 2.2 which introduces the
variables
variables used
used to
to construct
construct the
the BN,
BN, followed
followed by
by an
an explanation
explanation ofof the
the methods
methods (Section
(Section 2.3)
2.3) and
and
assumptions
assumptionsused
used inin the
the hydraulic
hydraulic modeling
modeling (Section
(Section 2.4).
2.4).

2.1.Bayesian
2.1. BayesianNetworks
Networks
Bayesian Networks
Bayesian Networks(BNs) (BNs)are are
probabilistic graphical
probabilistic models models
graphical which explicitly
which include
explicitlydependence
include
between multiple
dependence between random
multiplevariables.
randomRandom
variables.variables
Randomare represented
variables by nodes and
are represented connected
by nodes and
by arcs forming
connected by arcsaforming
directeda acyclic
directedgraph.
acyclicThe direction
graph. given togiven
The direction the arcs establish,
to the in some
arcs establish, cases,
in some
the causality
cases, structurestructure
the causality and, in general,
and, inconditional
general, independence statements between
conditional independence variables
statements [49–53].
between
This structure
variables determines
[49–53]. which determines
This structure nodes are identified as parents
which nodes of a givenasnode,
are identified i.e., of
parents thea predecessors,
given node,
or as
i.e., children,
the i.e., theorsuccessors.
predecessors, as children,Variables can be discrete
i.e., the successors. or continuous,
Variables and the
can be discrete or expression
continuous,ofand
the
(conditional) dependency is quantified by conditional probability functions. In this study,
the expression of the (conditional) dependency is quantified by conditional probability functions. In we will use
the class
this study,ofwe
BNs described
will use the in Hanea
class et al.described
of BNs [52,53]. The joint density
in Hanea for BNsThe
et al. [52,53]. mayjoint
be written
densityas:
for BNs
may be written as:  n 
f X1 , X2 ,··· ,Xn ( x1 , x2 , · · · , xn ) = ∏ f Xi Pa(Xi ) xi XPa(Xi ) = x , (1)
, ,⋯, , ,⋯, i =1
│Pa │ Pa , (1)
where XPa(Xi ) = x is short hand notation for XPa1 (Xi ) = xPa1 (Xi ) , . . . , XPam (Xi ) = xPam (Xi ) and Pa( Xi ) is
where misparents
short hand notation for , … , Pa( X ) = ∅ so that and f Xi PaPa is
the set containing
Pa of node Xi . For nodes without parents, i ( Xi ) = f Xi .
the set containing
When m parents variables
only continuous of node are . For nodes
present in without parents,
the model, ∅ so that
Pa representing
the nodes the continuous
│Pa
.
random variables are usually constructed using their empirical distribution function, and the
When only
dependence continuous
structure between variables
each pair are of present
nodes is in the model,
modeled using the nodesThe
copulas. representing
use of copulas the
continuous
provides anrandom
efficientvariables are usually
way to represent the constructed usingand
joint probability their
thusempirical distribution
the dependence function,
structure [54].
and
The the dependence
advantage structure
of copula between
functions is each pair of the
to separate nodes is modeled
selection of anusing copulas.multivariate
underlying The use of
copulas provides
distribution functionan from
efficient
their way to representmarginal
one-dimensional the joint probability This
distributions. andenables
thus thea highdependence
flexibility
structure [54]. The advantage
in the representation of copuladependence
of the multivariate functions isstructure
to separate the selection
[28,54–57]. For the of an underlying
bivariate case, the


multivariate distributionfunction,
cumulative distribution function F
from
Xi X j their
xi , x jone‐dimensional
is defined as: marginal distributions. This enables a
high flexibility in the representation of the multivariate dependence structure [28,54–57]. For the
F ,


bivariate case, the cumulative distribution function,
FX X xi , x j = C FX ( xi ), FX x j = C(u, v), as:
is defined


i j i j
(2)

,
where FXi ( xi ) and FX j x j areFthe marginal C F , distribution
F C u, v , or uniform ranks, of (2)


cumulative functions, the
random variables Xi and X j . C is the copula function and u and v are uniform variates on the unit
where F and F are the marginal cumulative distribution functions, or uniform ranks, of
square I = [0, 1]2 . We refer to References [58–60] for more theoretical background.
the random variables and . C is the copula function and u and v are uniform variates on the
In the class of BNs used in this research, the multivariate dependence structure is parametrized
unit square 0,1 . We refer to References [58–60] for more theoretical background.
with the empirical Spearman’s (conditional) rank correlation coefficient r between parent–child pairs
In the class of BNs used in this research, the multivariate dependence structure is parametrized
and modeled using Gaussian copulas. The rank correlation coefficient is equivalent to the Pearson’s
with the empirical Spearman’s (conditional) rank correlation coefficient between parent–child
product moment correlation, ρ, applied to the ranks of the individual variables, such that:
pairs and modeled using Gaussian copulas. The rank correlation coefficient is equivalent to the
Pearson’s product moment correlation, , applied  to the ranks
 of the individual variables, such that:
r Xi ,X j = ρ FXi ( xi ), FX j x j , (3)
, , , (3)

Equation (3) is used to determine the parameter of the bivariate Gaussian copula, C :
C u, v Φ Φ u ,Φ v , (4)
Water 2018, 10, 1190 5 of 19

Equation (3) is used to determine the parameter of the bivariate Gaussian copula, Cρ :
 
Cρ (u, v) = Φρ Φ−1 (u), Φ−1 (v) , (4)

where Φ−1 is the inverse of the univariate standard normal distribution, and Φρ is the bivariate
Gaussian cumulative distribution function. While the use of the Gaussian copula provides a
computationally efficient formulation for sampling, in particular for large BN structures [52], it also
restricts the representation of the dependence. Diagnostic tools (so-called informal tests) based on
the Cramér-von-Mises statistic between alternative copulas and semi-correlation, as mentioned in
Wang and Wells [61] and Joe [56], can provide some valuable guidance in the assessment of the
fit of the bivariate data with respect to copula models. Formal hypothesis testing does not always
lead to any valid model as a representation for a particular data set. Moreover, these tests may
be very computationally expensive. For these reasons, we use both approaches as complementary.
A description of these tests is given in Supplementary Material Section S5.2 and applied to the BN
developed for this study.
It is beyond the scope of this paper to provide a complete overview of BNs. For more details on
the class of BNs used in this paper, the reader is referred to Hanea et al. [52] and references therein.
For our purpose, it is sufficient to say that, loosely, the BN to be used attaches: (1) either an empirical
or parametric cumulative distribution function to the nodes representing random variables; (2) a
non-unique assignment of (conditional) rank correlations to the arcs of the BN, and (3) a multivariate
Gaussian copula that represents the dependence structure of the data and conditional independence
structure dictated by the graph of the BN.
This class of Bayesian Networks has been successfully applied to various flood risk applications
for different flood generating processes, namely to characterize extreme river discharge [62], to model
expected riverine flood damages [50], and to estimate tropical cyclone parameters [5]. We extend their
application further to generate stochastic boundary conditions which include compound flooding from
riverine and coastal interactions and model their impact on a coastal catchment in Southeast Texas.

2.2. Data Collection

The location and characteristics of the Buffalo Bayou catchment make it prone to the co-occurrence
of riverine and coastal floods. It is a small catchment (~2000 km2 ) exposed to intense rainfall
events from local convective storms, large-scale frontal systems, and torrential rainfall brought by
tropical cyclones. Historical records of flood events reported as major by the Federal Emergency
Management Agency (FEMA) indicate a relatively equal importance of these different flood sources [18].
The response of the catchment to extreme rainfall has been further exacerbated by rapid development
which has occurred over the century [63], leading to faster runoff rates and larger flood volumes [64,65].
This study focuses on characterizing water surface elevations along the downstream reach of
Buffalo Bayou which flows through downtown Houston and the Port of Houston before joining the
San Jacinto River, see Figure 2a. The catchment drains into Galveston Bay (Figure 2b), a shallow
semi-enclosed estuary connected to the Gulf of Mexico via Bolivar Roads (Figure 2c). Buffalo Bayou
upstream of the Port of Houston has a constant bed slope, approximately 0.55 m/km. Within the Port
of Houston, the channel has the near-uniform bed slope of 0.12 m/km due to regular dredging to
ensure sufficient ship clearance up to 14 m. Seven tributaries feed into the main stem of Buffalo Bayou
before it reaches the Lynchburg Landing (LL) site.
Water 2018, 10, 1190 6 of 19
Water 2018, 10, x FOR PEER REVIEW 6 of 20

(b)

(a) (c)

Figure 2. Location
Figure 2. Location of
of the
the case
case study
study area,
area, the
the Buffalo
Buffalo Bayou
Bayou catchment
catchment (a)
(a) which drains into the
Galveston
Galveston Bay (b) and connected to the Gulf of Mexico (c). The discharge
discharge stations
stations (filled circles in
black) and the tide station (empty circles) Galveston
Galveston Pier 21 are used to construct the BN and model
water surface elevations along the main stem of the Buffalo Bayou catchment
catchment (red line). Water level
stations (green) were used for the comparison with the modeled extreme
extreme water
water levels.
levels.

Discharge: Mean daily discharges were collected from the U.S. Geological Survey at the
Discharge: Mean daily discharges were collected from the U.S. Geological Survey at the locations
locations shown in Figure 2a. The available period of data varied significantly per station, with
shown in Figure 2a. The available period of data varied significantly per station, with records starting
records starting between 1936 and 1971. Abrupt changes in the discharge data series indicate a
between 1936 and 1971. Abrupt changes in the discharge data series indicate a possible sign of
possible sign of anthropogenic influences on the hydrologic response of the catchment [66,67]. We
anthropogenic influences on the hydrologic response of the catchment [66,67]. We identified the most
identified the most important change in mean in the data series (function findchangepts() in
important change in mean in the data series (function findchangepts() in MATLAB), which, at most
MATLAB), which, at most stations, was found to be located between 1970 and 1980 (for details see
stations, was found to be located between 1970 and 1980 (for details see Section S1.1 in Supplementary
Section S1.1 in Supplementary Material). Therefore, we selected data from 1 January 1980 onwards
Material). Therefore, we selected data from 1 January 1980 onwards to represent the current developed
to represent the current developed state of the catchment and assumed stationary conditions
state of the catchment and assumed stationary conditions between 1980 and 2016. Of the seven stations,
between 1980 and 2016. Of the seven stations, four stations had a very high temporal coverage for
four stations had a very high temporal coverage for this period (>97%), and three had a limited to poor
this period (>97%), and three had a limited to poor coverage (9–43%).
coverage (9–43%).
Storm surge: Hourly water levels and astronomical tide projections were obtained from the
Storm surge: Hourly water levels and astronomical tide projections were obtained from the
National Oceanic and Atmospheric Administration (NOAA) for the LL and Galveston Pier 21 (GP)
National Oceanic and Atmospheric Administration (NOAA) for the LL and Galveston Pier 21 (GP)
tide stations, Figure 2a,b. The LL tide station has a limited record length, about 11 years worth of
tide stations, Figure 2a,b. The LL tide station has a limited record length, about 11 years worth of data
data scattered between 1995 and 2014, compared to 113 years from 1904 to 2016 at GP. At both
scattered between 1995 and 2014, compared to 113 years from 1904 to 2016 at GP. At both stations,
stations, hourly non‐tidal residuals were calculated by subtracting the measured water level from
hourly non-tidal residuals were calculated by subtracting the measured water level from the predicted
the predicted astronomical tide. The maximum hourly non‐tidal residuals in a day is set to be the
astronomical tide. The maximum hourly non-tidal residuals in a day is set to be the daily non-tidal
daily non‐tidal residual, what is referred to the storm surge in this study. Data for GP were further
residual, what is referred to the storm surge in this study. Data for GP were further corrected for
corrected for mean sea‐level rise using a linear regression of the hourly non‐tidal residuals of the
mean sea-level rise using a linear regression of the hourly non-tidal residuals of the whole record
whole record length, 1904–2016. This was not possible for the LL tide station, so we assumed
length, 1904–2016. This was not possible for the LL tide station, so we assumed stationarity of the
stationarity of the available data.
available data.
For the sake of clarity, station IDs have been simplified. The original station numbers are
For the sake of clarity, station IDs have been simplified. The original station numbers are provided
provided in Section S1 in the Supplementary Material.
in Section S1 in the Supplementary Material.

2.3.
2.3. Bayesian
Bayesian Network
Network Construction
Construction
The BN for
The BN for the
thecase
casestudy
studyareaareawas
wasquantified
quantified based
based onon
thethe empirical
empirical marginal
marginal distributions
distributions of
of the
the daily discharges in the catchment (seven nodes) and the daily storm surge at GP
daily discharges in the catchment (seven nodes) and the daily storm surge at GP (one node). A clear(one node). A
clear advantage
advantage of thisofchoice
this choice
is that is that
the the analysis
analysis of the statistical
of the statistical dependence
dependence givesinto
gives insight insight into the
the physical
physical mechanisms leading to flooding. However, in the case of strong serial dependence within
the time series, this may lead to an incorrect quantification of the joint exceedance probability and
Water 2018, 10, 1190 7 of 19

mechanisms leading to flooding. However, in the case of strong serial dependence within the time
series, this2018,
Water may 10, xlead to anREVIEW
FOR PEER incorrect quantification of the joint exceedance probability and 7therefore of 20
an underestimation of the flood hazard. Visual observations of the time series of the input variables
showedtherefore an underestimation
that both coastal and river of the flood
flood hazard.
events areVisual observations
short-lived, from of the2 time
1 to days.series of the input our
To complement
visualvariables
findings,showed that both coastal
we investigated and river
the presence floodcorrelation
of serial events are based
short‐lived,
on thefrom 1 to 2 days.
methodology To
presented
complement our visual findings, we investigated the presence of serial correlation based on the
in [68,69]. The results are presented in Section S3 of the Supplementary Material. In almost all cases
methodology presented in [68,69]. The results are presented in Section S3 of the Supplementary
except for two variables, the empirical autocorrelation functions dropped very rapidly below 0.1,
Material. In almost all cases except for two variables, the empirical autocorrelation functions
indicating only a weak serial dependence. Therefore, we expected this effect to be of limited influence
dropped very rapidly below 0.1, indicating only a weak serial dependence. Therefore, we expected
here and acknowledge that this
this effect to be of limited choicehere
influence mayand lead to conservative
acknowledge estimates
that this choice mayof flood
lead tolevels.
conservative
We used the Uninet
estimates of flood levels. software package (http://www.lighttwist.net/wp/uninet) to test various
BN structures
We used andthe extract
Uninetthe rank correlation
software matrix of the final model which to
package (http://www.lighttwist.net/wp/uninet) is test
shown in Figure
various BN 3.
structures and
Two discharge extract the
variables withrank correlation
a limited matrix ofcoverage
temporal the final model
were which is shown
inserted in Figure 3. Two
as user-defined random
discharge
variables variables the
to maximize with a limited
number temporal
of joint coverage from
observations were all
inserted
nodes. as This
user‐defined
is further random
described
variables to maximize the number of joint observations from all nodes.
in Section S5.1 in Supplementary Material. The BN was first constructed as a saturated graph This is further described in and
Section S5.1 in Supplementary Material. The BN was first constructed as a saturated graph and
iteratively simplified by removing arcs with absolute conditional rank correlation lower than 0.3
iteratively simplified by removing arcs with absolute conditional rank correlation lower than 0.3
(not shown here). All the arcs with the parent node ‘Surge HGP’ were kept to explicitly include the
(not shown here). All the arcs with the parent node ‘Surge HGP’ were kept to explicitly include the
coastal and riverine interactions.
coastal and riverine interactions.

Figure 3. The
Figure Bayesian
3. The Network
Bayesian Networkmodelmodelfor
forthe
the Buffalo Bayoucatchment.
Buffalo Bayou catchment. TheThe boxes
boxes represent
represent the the
nodesnodes
and the
and arrows
the arrowsthethe
arcs. The
arcs. Thevalues
valuesofofthe
the (conditional) rank
(conditional) rank correlation
correlation coefficient
coefficient are shown
are shown
on each
on each arc. The
arc. The histogram
histogram in in
thethe nodeshows
node shows thethe empirical
empiricalprobability
probabilitydistribution function
distribution for each
function for each
variable
variable and and the values
the values below
below representthe
represent themean
mean and
and standard
standarddeviation, respectively.
deviation, Units
respectively. are inare in
Units
m3 s-1mfor
3s‐1 for discharge and in m for the storm surge.
discharge and in m for the storm surge.

The Gaussian copula model was investigated against two alternative models, the Gumbel and
The Gaussian
Clayton copula, copula modelthewas
by comparing investigated against
Cramér‐von‐Mises two alternative
statistic values models,
as recommended the Gumbel
in Reference
and Clayton copula, by comparing
[61] and semi‐correlation thethese
[56]. Together, Cramér-von-Mises statistic
copulas include a wide values
range as recommended
of dependence observed in
Reference [61] and semi-correlation
in environmental data and express[56]. Together,
different these
types of copulas include
tail dependence a wide range
of importance when of dependence
estimating
extreme
observed joint probabilities
in environmental (Equation
data (S.3) in
and express Supplementary
different types ofMaterial and following
tail dependence discussion)when
of importance
[28,36,70,71].
estimating We complemented
extreme these informal
joint probabilities (Equation tests by ain
(S.3) formal Goodness‐of‐Fit
Supplementary test [55]and
Material usingfollowing
two
representative pair variables, discharge–discharge and surge–discharge. The detailed description of
discussion) [28,36,70,71]. We complemented these informal tests by a formal Goodness-of-Fit test [55]
Water 2018, 10, 1190 8 of 19

using two representative pair variables, discharge–discharge and surge–discharge. The detailed
description of the tests and additional results are included in Sections S5.2 and S5.3 of Supplementary
Material. The results highlight the difficulty of pinpointing a single copula model to capture the
complexity of the joint behavior. Since the simplification of the dependence structure is inherent to
the probabilistic model selected, we consider this choice as an additional source of uncertainty in the
modeled water levels and perform a simple sensitivity analysis presented in the results section.
For the BN inference, we adopted a parametric representation of the marginal distribution
since extreme values were of importance for this study. The BN inference was implemented with
MATLAB [72] following a similar procedure described and developed by Hanea et al. [52] except for
the inverse transformation of the uniform variables which relied on parametric marginal distributions
presented next.
For the mean daily discharges, the marginal distributions were fitted with a parametrized
distribution, the generalized extreme value (GEV) distribution which resulted in the lowest Akaike
Information Criteria measure [73] among the 17 distribution candidates tested at each discharge station,
except one where it ranked second. The GEV cumulative distribution, also referred as the Fisher-Tippet
distribution, is: (   −1 )
x−µ

k
FQ = exp − 1 + k , (5)
σ

where x is a quantile of the discharge variable X, µ is the location parameter, σ the scale parameter,
and k is the shape parameter with k 6= 0. We chose to specifically fit the upper tail of the data to
put the emphasis on moderate to extreme discharge daily events and, therefore, used a truncated
maximum-likelihood method to derive the distribution parameters (presented in Section S2.1 in
Supplementary Material). A comparison against the regulatory riverine flood model provided little
insight due to the difference in methodology applied which relies on a regression at the regional
scale [18].
For the storm surge, the marginal distributions were fitted with a Gaussian mixture model with
c = 2 components [74,75]:
c
f S = ∑ w j Φ · µ j , σj ,


(6)
1

where f S is the density probability function of the maximum hourly residual in a day, Φ · µ j , σj

is the normal density with mean µ j and standard deviation σj , and w j is the mixing coefficient of
each component such that ∑ w j = 1 and w j > 0. The distribution parameters were estimated
j
from the expectation maximization [75] and are presented in Table S2 of Supplementary Material.
The 100-year water level was found to be 2.84 m-NAVD88, 95% bootstrap confidence interval
[2.47, 3.13]. The extreme value analysis performed by NOAA [76] estimates this value to be
3.09 m-NAVD88 with a 95% confidence interval of [2.19, 4.89]. This comparison was not possible
for the LL station since no return period estimates are mentioned for this station by NOAA.

2.4. 1D Hydraulic Model

Longitudinal water profiles along the selected reach of Buffalo Bayou were obtained by solving
the steady-state one-dimensional (1D) shallow water equations [77] with the standard step method [78]
(as described in Section S4.1 in Supplementary Material). While this is a clear simplification of flood
processes, water levels in the catchment are currently derived under similar assumptions to create
the FEMA flood insurance rate maps (FIRMs) [18,26]. The spatial discretization was obtained from
surveyed cross-sections [79] at approximately every kilometer, and the water surface elevations were
calculated every 100 m, here forth referred to as river calculation points. The contributions from the
tributaries were added as point sources along the river reach. Because the discharge stations do not
cover the whole drainage area, discharge values were corrected based on an area-weighted average
Water 2018, 10, 1190 9 of 19

and lateral inflow was neglected. The downstream boundary condition at the LL site, i.e., the total
water level, was reconstructed by adding the storm surge to the mean high tide level. The storm surge
at LL Water
was 2018,
predicted
10, x FORbased on the storm surge obtained for GP using a linear regression (Equation
PEER REVIEW 9 of 20 (S.1)
in Supplementary Material). The hydraulic model was forced with the derived boundary conditions to
obtaintide
thelevel.
water The storm elevation
surface surge at LL was predicted
along based
the selected on the storm surge obtained for GP using a
reach.
linear regression (Equation (S.1) in Supplementary
The model showed a very good agreement with the 10-, Material). The hydraulic
50-, 100- and model was forced
500-year returnwith
period
the derived boundary conditions to obtain the water surface elevation along the
riverine flood levels obtained from the validated Hydrologic Engineering Center’s River Analysis selected reach.
The model showed a very good agreement with the 10‐, 50‐, 100‐ and 500‐year return period
System (HEC-RAS) software model for Buffalo Bayou used to produce the aforementioned flood
riverine flood levels obtained from the validated Hydrologic Engineering Center’s River Analysis
insurance maps (R2 > 0.98) [79]. Moreover, the model was also validated for Tropical Storm Frances
System (HEC‐RAS) software model for Buffalo Bayou used to produce the aforementioned flood
(11 September 1998)(R
insurance maps and Tropical
2 > 0.98) Storm Allison
[79]. Moreover, (6 June
the model wasand
also9validated
June 2001)for and showed
Tropical StormaFrances
reasonable
performance, considering
(11 September 1998) andthat water
Tropical level
Storm observations
Allison from
(6 June and tropical
9 June cyclone
2001) and showed events often have
a reasonable
high performance,
uncertainty considering
[80]. Results thatand
water figures from the runs
level observations fromare shown
tropical in Sections
cyclone S4.2have
events often andhigh
S4.3 in
Supplementary
uncertainty Material.
[80]. Results and figures from the runs are shown in Sections S4.2 and S4.3 in
Supplementary Material.
3. Results
3. Results
The BN model was sampled 182,500 times, equivalent to 500 years of daily observations,
to generate Thepossible
BN model was sampled
realizations of 182,500 times, equivalent
daily discharge to 500surge
and storm yearsfrom
of daily
theobservations,
multivariatetojoint
generate possible realizations of daily discharge and storm surge from the multivariate joint
distribution and model the resulting water surface elevations along the selected river reach. At each
distribution and model the resulting water surface elevations along the selected river reach. At each
of the river calculation points, we extracted the 90th, 95th, 99th, and 99.99th percentile of the daily
of the river calculation points, we extracted the 90th, 95th, 99th, and 99.99th percentile of the daily
waterwater
level level
distribution, as shown
distribution, as shown in Figure 4. These
in Figure 4. Thesepercentiles
percentilesare
areexceeded on average
exceeded on averageabout
about36,
36, 18,
4 days18,per
4 days per year, and once every 28 years, respectively. The maximum water surface elevation line
year, and once every 28 years, respectively. The maximum water surface elevation (red
in Figure 4) represents
(red line in Figure 4)the highestthe
represents water level
highest obtained
water at eachatriver
level obtained each calculation point
river calculation from
point fromall the
modeled daily
all the joint daily
modeled occurrences of discharges
joint occurrences and storm
of discharges andsurge. Similarly,
storm surge. the minimum
Similarly, the minimumwater surface
water
surface
elevation (blueelevation (blue line)
line) indicates the indicates
lowest dailythe water
lowest surface
daily water surface
elevation elevation
at each river at each riverpoint.
calculation
calculation
The model point. The
reproduces model
a wide reproduces
range a wide range
of conditions from of conditions
extremely from
low extremely
to normal lowextremely
and to normalhigh
and extremely high
water surface elevations. water surface elevations.

Figure
Figure 4. Maximum
4. Maximum (red),
(red), minimum
minimum (blue)water
(blue) water surface
surface elevations
elevationsand
andpercentiles (90th,
percentiles 95th,95th,
(90th, 99th,99th,
and 99.99th)
and 99.99th) of daily
of daily water
water levels
levels at at eachriver
each rivercalculation
calculation point
pointobtained byby
obtained inferring the the
inferring BN 182,500
BN 182,500
timestimes and modeling
and modeling the the resulting
resulting water
water surfaceprofile.
surface profile.

We note that there is a nonlinear response in the propagation of the backwater effects on the
We note
water that elevations
surface there is a modeled
nonlinear response
along inreach.
the river the propagation
The imposed of the level
water backwater effects on the
at the Lynchburg
waterLanding
surfacesite
elevations modeled
controls the along the
water surface river reach.
elevations The imposed
in the downstream waterfrom
section, level at the Lynchburg
approximately 0
to −25 km. As the bed slope steepens in the upstream section and the geometry of the river
Landing site controls the water surface elevations in the downstream section, from approximately 0 to becomes
moreAs
−25 km. complex,
the bedthe watersteepens
slope surface elevations are determined
in the upstream sectionbyand
the backwater effects
the geometry ofofthe
theriver
incoming
becomes
tributary discharges along the river and the imposed downstream water level.
more complex, the water surface elevations are determined by the backwater effects of the As a result, there is a
incoming
significant variation in the differences across all modeled scenarios between these two river sections.
Water 2018, 10, 1190 10 of 19
Water 2018, 10, x FOR PEER REVIEW 10 of 20

tributary
The range discharges
in modeled along the waterriver and the
levels imposed downstream
is indicated water level.
by the difference between As atheresult,
minimumthere is and a
significant variation in the differences across all modeled scenarios
maximum water surface elevation profiles. Upstream, this difference is about 20 m whereas between these two river sections.
The range in modeled
downstream water levels
the difference is indicated
is about 5 m. Thisby the
showsdifference between
that the the minimum
magnitude and maximum
of the backwater effects
water
from the riverine and coastal interactions is influenced by the river characteristics which inthe
surface elevation profiles. Upstream, this difference is about 20 m whereas downstream turn
difference
determines is about 5 m. Thisflood
the dominant showsdrivers
that theinmagnitude
the modeled of water
the backwater effects from the riverine and
surface elevations.
coastal Weinteractions
further focus is influenced
on extreme by the river
water characteristics
levels which flood
and characterize in turnhazard
determinesat anythe dominant
location in the
flood drivers in the modeled water surface elevations.
model domain by constructing annual exceedance probability curve from the modeled daily water
We further
surface focus We
elevations. on extreme
define the water levels and
modeled ‘annualcharacterize
maxima’floodof the hazard
wateratsurface
any location in theby
elevations
model
randomly separating the modeled water levels at each river calculation point into 500 binswater
domain by constructing annual exceedance probability curve from the modeled daily of 365
surface
days elevations.
and extracting We define the modeled
the maximum ‘annual
of each bin.maxima’
By doing ofso,
thewe
water surface
assume theelevations
extreme by randomly
modeled daily
separating
water surfacethe modeled
elevations waterin levels at each
the series to beriver calculation of
independent point
eachinto 500 which
other, bins of may365 dayslead andto an
extracting the maximum
overestimation of each
of the flood bin.for
level Byadoing
givenso, we assume
annual the extreme
exceedance modeled
probability. Wedaily
comparewaterour surface
results
elevations in the series to be independent of each other, which may
at three locations along the river where observed annual maxima of water levels are present. lead to an overestimation of theTwo
flood level for a given annual exceedance probability. We compare our results
stations, HM and HTB, are located close to each other in the Port of Houston, while the third station at three locations along
theHBB,
riveriswhere
located observed
upstream. annual
Themaxima of water
exact location of levels are present.
the stations Two stations,
is indicated in Figures HM1and andHTB,
4. A areGEV
located close toiseach
distribution fittedother in thethe
to both Port of Houston,
observed while theannual
and modeled third station
maxima HBB, datais located
series, aupstream.
commonly
The exact location
applied statistical of the stations isfor
distribution indicated in Figures 1of
the quantification and
the4. probabilities
A GEV distribution of extremeis fitted
waterto both
levels
themaxima
observed and
[81–83]. modeled annual maxima data series, a commonly applied statistical distribution for
the quantification
Figure 5 shows of thethe probabilities of extreme
annual exceedance water levels
probability maxima
curves [81–83].
obtained from the model at the three
locations where observations of annual maxima water levels are present. Inthe
Figure 5 shows the annual exceedance probability curves obtained from modelthe
general, at model
the three is in
locations
agreement where with observations
the empirical of annual maxima except
return periods water levels are present.
at station HM (Figure In general,
5a). When the model
comparing is
in Figure
agreement5a,b,with the empirical
the model indicates return periods
a similar exceptvalue
extreme at station HM while
behavior (Figure 5a).
this is When
not thecomparing
case for the
Figure 5a,b, annual
empirical the model indicatescurves.
exceedance a similarThisextreme
differencevaluein behavior
extreme valuewhilebehavior
this is not the case
might be duefor tothethe
empirical annualthat
uncertainties exceedance
stem from curves. This differenceobserved
the discontinuous in extremeannual
value behavior
maxima series might at be station
due to the HM.
uncertainties that stem from the discontinuous observed annual maxima
Because the model framework outputs similar data series length, this contrast is absent, and a series at station HM. Because
thepositive
model framework
shape parameter outputs is similar
calculateddataatseries
both length,
locations, thiswhich
contrast is absent,
is also and awith
consistent positive shape
the modeled
parameter
hydraulic is calculated
behavior. The at both locations, which
performance of theishydraulic
also consistent
model with
alsotheaffects
modeled the hydraulic behavior.in
results presented
The performance of the hydraulic model also affects the results
Figure 5. Compared with the observed extreme water levels, the model underestimates extremepresented in Figure 5. Compared
with the surface
water observed extremein
elevations water levels, the model
the downstream reach underestimates extreme waterinsurface
(Figure 5b) and overestimates elevations
the upstream reach
in (Figure
the downstream
5c). In allreach cases,(Figure
the annual 5b) and overestimates
exceedance in the upstream
probabilities of water reachlevels (Figure
are within 5c). the
In all95%
cases, the annual exceedance probabilities of water levels are within the 95%
confidence interval of the empirical distribution parameters, except in the high‐frequency region for confidence interval of the
empirical distribution
station HBB (Figure 5c).parameters, except in the high-frequency region for station HBB (Figure 5c).

(a)

Figure 5. Cont.
 Water 2018, 10, x FOR PEER REVIEW 11 of 20
Water 2018, 10, 1190 11 of 19

(b)

(c)
Figure
Figure 5. Comparison
5. Comparison of theofannual
the annual exceedance
exceedance probability
probability distributions
distributions from and
from observed observed
modeled and
annual maxima at stations HM (a), HTB (b) and HBB (c). The dashed lines indicate the fit of the 95% of
modeled annual maxima at stations HM (a), HTB (b) and HBB (c). The dashed lines indicate the fit
the 95% interval
confidence confidence
frominterval from GEV
the empirical the empirical GEV
distribution distribution
parameters. Theparameters. The stations
stations are presented fromare
presented from downstream to upstream along the Buffalo Bayou river, with stations
downstream to upstream along the Buffalo Bayou river, with stations HM and HTB located in the Port HM and HTB
of located
Houston,in see
the also
PortFigures
of Houston,
1 andsee also
4 for Figures
their exact1location.
and 4 for their exact location.

To investigate the importance of the statistical dependence on the water levels in the river
To investigate the importance of the statistical dependence on the water levels in the river reach,
reach, we forced the hydraulic model with other joint boundary conditions derived independent of
we forced the hydraulic model with other joint boundary conditions derived independent of the BN
the BN model and compared them with the 100‐year return period for each river calculation point
model and compared them with the 100-year return period for each river calculation point obtained
obtained from the model framework. We selected two contrasting options often applied in large
from the model framework. We selected two contrasting options often applied in large global flood
global flood hazard models, which we refer to as Case A and Case B:
hazard models, which we refer to as Case A and Case B:
 Case A: The 100‐year marginal return period for each discharge variable and the storm surge
• Case A: The
variable 100-year marginal
is calculated returnThis
and modeled. period for each
represents thedischarge
(untrue) variable andofthe
assumption fullstorm surge
dependence.
 variable is calculated and modeled. This represents the (untrue) assumption of full dependence.
Case B: The boundary conditions of the model are set to the marginal 100‐year return period for
• Case
the B: The boundary
storm conditionsand
surge downstream, of the
the model are setmean
distribution to thefor
marginal 100-year
the upstream return period
boundary for
conditions.
theThis
storm surge downstream, and the distribution mean for the upstream boundary
represents the (untrue) assumption of physical ‘independence’ between the downstream conditions.
This represents the (untrue) assumption of physical ‘independence’ between the downstream
 Water 2018, 10, x FOR PEER REVIEW 12 of 20
Water2018,
Water 2018,10,
10,1190
x FOR PEER REVIEW 12 of
12 of 19
20

waterlevel
water leveland andthethedischarge.
discharge.Such Suchan anapproach
approachisiscomparable
comparabletotoaabathtubbathtubapproach
approach[84],[84],even
even
though
water
though levelin
in thethe latter
andlatter method
the discharge. discharges are
Such an approach
method discharges usually completely
is comparable
are usually neglected
to a bathtub
completely neglected and
and not
approach modeled.
[84], even
not modeled.
though
Asseen in the latter
seenininFigures method
Figures66and discharges
and7,7,water
watersurfaceare usually
surfaceelevations completely
elevationsfor forCase neglected
CaseAAare and
arehigher not
higherthan modeled.
thanthethe100‐year
100‐year
As
return
return period
Asperiod
seen in obtained
obtained
Figuresfromfrom
6 andthe the modeled
7, modeled outputs,
outputs,
water surface with no
with no for
elevations difference
difference
Case Aatare at the Lynchburg
thehigher
Lynchburg
than the Landing
Landing
100-year site
site
andaa1period
and
return 1mmdifference
difference
obtained within
within
from the therest
the restofofthe
modeled theoutputs,
riverreach.
river reach. This
withThis corresponds
corresponds
no difference totoaaLynchburg
at the medianrelative
median relative error
error
Landing ofof
site
6.5%,
6.5%,
and with
a 1with a maximum
a maximum
m difference of 9.4%.
of 9.4%.
within the rest The diagnostic
Theofdiagnostic tools presented
tools presented
the river reach. in Sections
in Sections
This corresponds S5.2
to aS5.2 in Supplementary
in Supplementary
median relative error
ofMaterial
Material
6.5%, with suggest
suggest
a maximum that
that thethe joint dependence
joint dependence
of 9.4%. The diagnostic might
might deviate
tools deviate
presented from
from the Gaussian
the Gaussian
in Sections copula model,
copula model,
S5.2 in Supplementary
especially
especially
Material between
between
suggest discharge
discharge
that the variables.
variables. might
joint dependence In
In thethe hypothetical
hypothetical
deviate case of
case of upper
from the Gaussian upper tail dependence,
tail dependence,
copula model, especially
accounting
accounting for
for this
this dependency
dependency relationship
relationship would
would result
result in
in the
the 100‐year
100‐year
between discharge variables. In the hypothetical case of upper tail dependence, accounting for water
water level
level return
return period
period
this
totobe
beininbetween
dependency between thecurves
the
relationship curves obtained
obtained
would from
resultfrom
in the the
the currentwater
current
100-year model
model framework
framework
level andCase
and
return period Case A.in between the
A.
to be
curves obtained from the current model framework and Case A.

Figure6.6.Sensitivity
Figure Sensitivityofofthe
themultivariate
multivariatedependence
dependenceassumptions.
assumptions.The The100–year
100–yearwater
waterlevel
levelobtained
obtained
Figure 6. Sensitivity of the multivariate dependence assumptions. The 100–year water level obtained
from the
from the current
the current model
current model framework
model framework
framework is is represented
is represented
represented byby the
by the black
the black full
black full line.
full line. Case
line. Case
Case A A and
A and B represent
and BB represent full
represent full
full
from
dependence
dependence and and ‘independence’
and ‘independence’
‘independence’of of
ofthethe boundary
theboundary conditions,
boundaryconditions, respectively
conditions,respectively
respectively(see (see text).
(seetext).
text).
dependence

(a)
(a)
Figure 7. Cont.
 Water2018,
Water 2018,10,
10,1190
x FOR PEER REVIEW 1313ofof1920

(b)

(c)
Figure7.7.Similar
Figure SimilartotoFigure
Figure66but
butininthe
thetransverse
transversedirection
directionatatstations
stationsHM
HM(a),
(a),HTB
HTB(b)
(b)and
andHBB
HBB(c).
(c).
The location of the bank stations is extracted from the regulatory HEC‐RAS riverine flood model
The location of the bank stations is extracted from the regulatory HEC-RAS riverine flood model and and
roughlyindicates
roughly indicatesthe
thechange
changeininchannel
channelconveyance.
conveyance.

In contrast to Case A, Case B’s assumption leads to much higher differences in the model
In contrast to Case A, Case B’s assumption leads to much higher differences in the model outputs.
outputs. While we again observe no difference in water surface elevations at the Lynchburg Landing
While we again observe no difference in water surface elevations at the Lynchburg Landing site, the
site, the difference increases moving upstream along the river and reaches a maximum difference of
difference increases moving upstream along the river and reaches a maximum difference of 12 m at
12 m at the upstream end of the river reach. The median relative error is −12%, with a maximum of
the upstream end of the river reach. The median relative error is −12%, with a maximum of −75%.
−75%. This result clearly highlights that neglecting discharge interactions can result in large
This result clearly highlights that neglecting discharge interactions can result in large underestimates
underestimates of flood stage and, therefore, potential flood risks, especially upstream in the river
of flood stage and, therefore, potential flood risks, especially upstream in the river reach.
reach.
4. Discussion
4. Discussion
In this study, we used observations of mean daily discharge, daily storm surge, and their statistical
In this study, we used observations of mean daily discharge, daily storm surge, and their
dependence to generate stochastic joint occurrences of boundary conditions and model water surface
statistical dependence to generate stochastic joint occurrences of boundary conditions and model
elevations resulting from their interactions. Unlike previous studies, our model framework provides
water surface elevations resulting from their interactions. Unlike previous studies, our model
a method for characterizing the compound flood hazard along the entire river reach. Yet several
framework provides a method for characterizing the compound flood hazard along the entire river
limitations inherent to the construction of our model are present and discussed in this section, namely
reach. Yet several limitations inherent to the construction of our model are present and discussed in
the simplification of the extreme flood events, the uncertainties in the BN generated from the lack
this section, namely the simplification of the extreme flood events, the uncertainties in the BN
of observations and the selection of the Gaussian copula and the assumption of stationarity in the
generated from the lack of observations and the selection of the Gaussian copula and the assumption
flood hazard.
of stationarity in the flood hazard.
First, while the current model framework can capture a wide range of hydraulic conditions of
First, while the current model framework can capture a wide range of hydraulic conditions of
relevance when studying compound events, it does not include the effect of flood duration on water
relevance when studying compound events, it does not include the effect of flood duration on water
surface elevations. The characteristics of both the riverine and coastal flood waves, such as volume
surface elevations. The characteristics of both the riverine and coastal flood waves, such as volume
Water 2018, 10, 1190 14 of 19

and duration, determines when and where they will interact within the catchment. The analysis in
Supplementary S3 confirmed that most flood events in the case study area tend to be short-lived. Yet,
long flood events spanning multiple days have been observed in the catchment, such as Hurricane
Harvey in August 2017. The tropical cyclone stalled over Texas generating extreme coastal water
surface elevations and precipitation lasting for several days which impeded runoff from the system [85],
causing critical road cutoffs, the release of toxic materials, and more than 70 fatalities [86,87]. Therefore,
it is important to consider the temporal aspect of flood not only to correctly determine the flood hazard
but also to properly model their impact on flood risk assessments.
Second, the lack of observations for three discharge variables (station QBB, QG, and QS) and
for the storm surge at the Lynchburg Landing site introduces uncertainties in the quantification
of their marginal distribution and of their statistical dependence with other variables in the BN,
which propagate through the model framework. For example, we expect uncertainties in the water
surface elevation imposed at the downstream boundary of the hydraulic model domain to strongly
influence extreme water levels in the downstream reach. Wave contributions, such as wave setup
and wave propagation, are also not directly represented by the BN but can strongly influence water
levels [88–90]. While this study can be improved by using complementary data at these stations, this
typically requires dedicated and extensive studies to properly capture complex coastal and hydrological
processes [64,91,92] and is left for future studies.
Third, we restricted the multivariate dependence representation by relying on the Gaussian copula
in the BN. However, our diagnostic tools indicated that especially for the pairs regarding discharges,
asymmetries are present in the data which may not be captured adequately by the Gaussian copula
(see Section S5.2 in Supplementary Material). Further research is needed to investigate the influence of
the copula selection on the modeled water levels. Furthermore, the discharge data resolution leads to
numerous repeated values for low discharges (as shown in Supplementary Figure S6b). The presence
of ties can affect copula inference and parameter estimates [93]. Randomization procedures might be an
option for further research [94]. A simple sensitivity test was conducted by assuming full dependence
between the considered variables for the 100-year water level return period. We estimated the median
relative error in water level along the river reach to be 6.5%, with a maximum of 9.4%. In future work,
a comparison of the current BN model with a probabilistic model allowing for a higher flexibility in
the selection of the dependence structure, such as Vine copula constructions [28], will help refine this
result and better quantify the importance of the statistical dependence structure on the exceedance
probability of high water levels due to compound flooding.
Finally, future flood hazards will be exacerbated by changes in environmental conditions and
anthropogenic factors. Several recent publications have highlighted that climate change has increased
precipitation in the study area [85,95]. Projected trends in urbanization in the United States also indicate
a steady growth in urban land cover [96], which is expected to increase runoff peaks and volumes [97].
As a result, current estimates of flood hazard may become rapidly outdated [38]. This study provides
a static characterization of the flood hazard, but future studies should use a modified framework and
include such dynamics. This could be done, for example, by imposing a non-stationary parametric
correlation coefficient [16,98] or by artificially shifting the marginal distribution [82,83].

5. Conclusions
In this paper, we presented a first attempt to characterize flood hazard in a coastal catchment
prone to compound flooding from riverine and coastal interactions using a BN based on Gaussian
copulas. We constructed and inferred the BN based on daily values of discharges and storm surge,
and propagated the joint occurrences of discharges and storm surge to a hydraulic model to obtain the
water level along Buffalo Bayou, in Southeast Texas. While uncertainties are introduced due to the
selection of the Gaussian copula, the simplification in the hydraulic modeling and the limited data
available for some variables, complex coastal and riverine interactions could be captured from the
multivariate joint probability to characterize compound flood hazard along the whole model domain.
Water 2018, 10, 1190 15 of 19

Because the BN is based on the statistical dependence, it provides a holistic procedure to

stochastically derive joint boundary conditions while accounting for multivariate and spatial
dependencies. The model framework generates daily water surface elevations resulting from various
combinations of riverine and coastal conditions, including both moderate and extreme realizations,
which is necessary for comprehensively analyzing the potential impacts of compound flood events.
Moreover, the analysis of the modeled water levels underlined the importance of considering backwater
effects due to high downstream water levels and tributary discharges. We also highlighted the
effect of different spurious dependence assumptions between flood drivers on the modeled water
level. We conclude that such differences can lead to an over—or underestimation of the annual
exceedance probabilities when compared against measured dependence. Future work will focus on
the characterization of flood hazard in diverse coastal catchments to better understand the propagation
of flood drivers and their impact in the estuarine region in combination with other non-stationary
drivers, such as relative sea-level rise and land cover changes.

Supplementary Materials: The following are available online at http://www.mdpi.com/2073-4441/10/9/1190/s1,

Section S1: Data Collection, Section S2: Marginal Distribution Fit, Section S3: Autocorrelation function, Section S4:
1-D Hydraulic Model Performance, Section S5: Bayesian Network Construction and Validation. Figure S1: Available
records of mean daily discharge for the stations of interests. The most important change in the mean is shown in red.
Table S1/S2: GEV distribution parameters for the discharge/storm surge distributions. Figure S2: Autocorrelation
function (ACF) at the stations of interests. Figures S3 and S4: Performance of the simplified hydraulic model developed
for this study. Table S3: Semi-correlation and Cramér-von-Mises statistic for all variables used in the BN except for
station QG, QS and QBB. Figure S5: Comparison of the maximum water levels observed for Tropical Storm Allison
and Frances with the results from the 1-D hydraulic model. Figure S6: Selected examples of semi-correlation.
Author Contributions: Conceptualization, A.C., A.S. and O.M.-N.; Methodology, A.C., A.S. and O.M.-N.;
Validation, A.C.; Writing-review and editing, A.C., A.S. and O.M.-N.; Visualization, A.C.
Funding: A.C. was supported by the Netherlands Organisation for Scientific Research (NWO) in the form of
a VIDI grant (grant No. 016.161.324). A.S. was supported by the Netherlands America Foundation/Fulbright
Fellowship for Water Management and the NSF PIRE Grant No. OISE-1545837.
Acknowledgments: The authors would like to thank H. Winsemius for his comments and suggestions during the
writing of this manuscript. We would also like to thank three anonymous reviewers for their valuable comments
which helped to improve this manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Kron, W. Coasts: The high-risk areas of the world. Nat. Hazards 2013, 66, 1363–1382. [CrossRef]
2. Rueda, A.; Camus, P.; Tomás, A.; Vitousek, S.; Méndez, F.J. A multivariate extreme wave and storm surge
climate emulator based on weather patterns. Ocean Model. 2016, 104, 242–251. [CrossRef]
3. Neumann, B.; Vafeidis, A.T.; Zimmermann, J.; Nicholls, R.J. Future coastal population growth and exposure
to sea-level rise and coastal flooding—A global assessment. PLoS ONE 2015, 10. [CrossRef]
4. Hallegatte, S.; Green, C.; Nicholls, R.J.; Corfee-Morlot, J. Future flood losses in major coastal cities. Nat. Clim.
Chang. 2013, 3, 802–806. [CrossRef]
5. Sebastian, A.; Dupuits, E.J.C.; Morales-Nápoles, O. Applying a Bayesian network based on Gaussian
copulas to model the hydraulic boundary conditions for hurricane flood risk analysis in a coastal watershed.
Coast. Eng. 2017, 125, 42–50. [CrossRef]
6. Svensson, C.; Jones, D.A. Dependence between sea surge, river flow and precipitation in south and west
Britain. Hydrol. Earth Syst. Sci. 2004, 8, 973–992. [CrossRef]
7. Saleh, F.; Ramaswamy, V.; Wang, Y.; Georgas, N.; Blumberg, A.; Pullen, J. A Multi-Scale Ensemble-based
Framework for Forecasting Compound Coastal-Riverine Flooding: The Hackensack-Passaic Watershed and
Newark Bay. Adv. Water Resour. 2017, 110, 371–386. [CrossRef]
8. Needham, H.F.; Keim, B.D.; Sathiaraj, D. A review of tropical cyclone-generated storm surges: Global data
sources, observations, and impacts. Rev. Geophys. 2015, 53, 545–591. [CrossRef]
9. Karl, T.R.; Melillo, J.M.; Peterson, T.C. Global Climate Change Impacts in the United States; Cambridge University
Press: Cambridge, UK; New York, NY, USA, 2009; ISBN 9780521144070.
Water 2018, 10, 1190 16 of 19

10. Pielke, J.R.; Roger, A.; Gratz, J.; Landsea, C.W.; Collins, D.; Saunders, M.A.; Musulin, R. Normalized
Hurricane Damage in the United States: 1900–2005. Nat. Hazards Rev. 2008, 9, 29–42. [CrossRef]
11. Zheng, F.; Leonard, M.; Westra, S. Application of the design variable method to estimate coastal flood risk.
J. Flood Risk Manag. 2015. [CrossRef]
12. Klerk, W.J.; Winsemius, H.C.; van Verseveld, W.J.; Bakker, A.M.R.; Diermanse, F.L.M. The co-incidence of
storm surges and extreme discharges within the Rhine–Meuse Delta. Environ. Res. Lett. 2015, 10, 035005.
[CrossRef]
13. Allen, G.H.; David, C.H.; Andreadis, K.M.; Hossain, F.; Famiglietti, J.S. Global Estimates of River Flow Wave
Travel Times and Implications for Low-Latency Satellite Data. Geophys. Res. Lett. 2018. [CrossRef]
14. Zscheischler, J.; Westra, S.; van den Hurk, B.J.J.M.; Seneviratne, S.I.; Ward, P.J.; Pitman, A.; AghaKouchak, A.;
Bresch, D.N.; Leonard, M.; Wahl, T.; et al. Future climate risk from compound events. Nat. Clim. Chang. 2018,
8, 469–477. [CrossRef]
15. Field, C.B.; Barros, V.; Stocker, T.F.; Dahe, Q.; Dokken, D.J.; Ebi, K.L.; Mastrandrea, M.D.; Mach, K.J.;
Plattner, G.-K.; Allen, S.K.; et al. Managing the Risks of Extreme Events and Disasters to Advance Climate
Change Adaptation. A Special Report of Working Groups I and II of the Intergovernmental Panel on Climate Change;
Cambridge University Press: Cambridge, UK, 2012; ISBN 978-1-107-60780-4.
16. Wahl, T.; Jain, S.; Bender, J.; Meyers, S.D.; Luther, M.E. Increasing risk of compound flooding from storm
surge and rainfall for major US cities. Nat. Clim. Chang. 2015, 5, 1093–1097. [CrossRef]
17. Federal Emergency Management Agency. Guidance for Flood Risk Analysis and Mapping; Combined Coastal
and Riverine; Federal Emergency Management Agency: Washington, DC, USA, 2015.
18. Federal Emergency Management Agency. Flood Insurance Study. Harris County, Texas and Incorporated Areas;
Federal Emergency Management Agency: Washington, DC, USA, 2017; Volumes 1–2.
19. Ward, P.; Jongman, B.; Salamon, P.; Simpson, A.; Bates, P.; De Groeve, T.; Muis, S.; de Perez, E.C.; Rudari, R.;
Trigg, M.A.; Winsemius, H. Usefulness and limitations of global flood risk models. Nat. Clim. Chang. 2015, 5,
712–715. [CrossRef]
20. Muis, S.; Verlaan, M.; Winsemius, H.C.; Aerts, J.C.J.H.; Ward, P.J. A global reanalysis of storm surges and
extreme sea levels. Nat. Commun. 2016, 7, 11969. [CrossRef] [PubMed]
21. Cid, A.; Camus, P.; Castanedo, S.; Méndez, F.J.; Medina, R. Global reconstructed daily surge levels from the
20th Century Reanalysis (1871–2010). Glob. Planet. Chang. 2017, 148, 9–21. [CrossRef]
22. Vousdoukas, M.I.; Voukouvalas, E.; Mentaschi, L.; Dottori, F.; Giardino, A.; Bouziotas, D.; Bianchi, A.;
Salamon, P.; Feyen, L. Developments in large-scale coastal flood hazard mapping. Nat. Hazards Earth
Syst. Sci. 2016, 16, 1841–1853. [CrossRef]
23. Winsemius, H.C.; Van Beek, L.P.H.; Jongman, B.; Ward, P.J.; Bouwman, A. A framework for global river flood
risk assessments. Hydrol. Earth Syst. Sci. 2013, 17, 1871–1892. [CrossRef]
24. Ward, P.; Jongman, B.; Weiland, F.; Bouwman, A.; van Beek, R.; Bierkens, M.; Ligtvoet, W.; Winsemius, H.
Assessing flood risk at the global scale: Model setup, results, and sensitivity. Environ. Res. Lett. 2013, 8,
44019. [CrossRef]
25. Maskell, J.; Horsburgh, K.; Lewis, M.; Bates, P. Investigating River-Surge Interaction in Idealised Estuaries.
J. Coast. Res. 2014, 294, 248–259. [CrossRef]
26. Ray, T.; Stepinski, E.; Sebastian, A.; Bedient, P.B. Dynamic Modeling of Storm Surge and Inland Flooding in a
Texas Coastal Floodplain. J. Hydraul. Eng. 2011, 137, 1103–1111. [CrossRef]
27. Kumbier, K.; Carvalho, R.C.; Vafeidis, A.T.; Woodroffe, C.D. Investigating compound flooding in an estuary
using hydrodynamic modelling: A case study from the Shoalhaven River, Australia. Nat. Hazards Earth
Syst. Sci. 2018, 2, 463–477. [CrossRef]
28. Bevacqua, E.; Maraun, D.; Hobæk Haff, I.; Widmann, M.; Vrac, M. Multivariate statistical modelling of
compound events via pair-copula constructions: analysis of floods in Ravenna (Italy). Hydrol. Earth Syst. Sci.
2017, 21, 2701–2723. [CrossRef]
29. Zheng, F.; Leonard, M.; Westra, S. Efficient joint probability analysis of flood risk. J. Hydroinform. 2015, 17,
584–597. [CrossRef]
30. Hawkes, P.J.; Gouldby, B.P.; Tawn, J.A.; Owen, M.W. The joint probability of waves and water levels in
coastal engineering design. J. Hydraul. Res. 2002, 40, 241–251. [CrossRef]
Water 2018, 10, 1190 17 of 19

31. Leonard, M.; Westra, S.; Phatak, A.; Lambert, M.; van den Hurk, B.; Mcinnes, K.; Risbey, J.; Schuster, S.;
Jakob, D.; Stafford-Smith, M.A. Compound event framework for understanding extreme impacts.
Wiley Interdiscip. Rev. Clim. Chang. 2014, 5, 113–128. [CrossRef]
32. Zheng, F.; Westra, S.; Leonard, M.; Sisson, S.A. Modeling dependence between extreme rainfall and storm
surge to estimate coastal flooding risk. Water Resour. Res. 2014, 50, 2050–2071. [CrossRef]
33. Kew, S.F.; Selten, F.M.; Lenderink, G.; Hazeleger, W. The simultaneous occurrence of surge and discharge
extremes for the Rhine delta. Nat. Hazards Earth Syst. Sci. 2013, 13, 2017–2029. [CrossRef]
34. Zheng, F.; Westra, S.; Sisson, S.A. Quantifying the dependence between extreme rainfall and storm surge in
the coastal zone. J. Hydrol. 2013, 505, 172–187. [CrossRef]
35. Ward, P.J.; Couasnon, A.; Eilander, D.; Haigh, I.; Hendry, A.; Muis, S.; Veldkamp, T.I.E.; Winsemius, H.;
Wahl, T. Dependence between high sea-level and high river discharge increases flood hazard in global deltas
and estuaries. Environ. Res. Lett. 2018. [CrossRef]
36. Salvadori, G.; De Michele, C.; Kottegoda, N.T.; Rosso, R. Extremes in Nature: An Approach Using Copulas;
Springer Science & Business Media: Berlin, Germany, 2007; pp. 1–294, ISBN 978-1-4020-4415-1.
37. Salvadori, G.; Durante, F.; De Michele, C.; Bernardi, M.; Petrella, L. A multivariate copula-based framework
for dealing with hazard scenarios and failure probabilities. Water Resour. Res. 2016, 52, 3701–3721. [CrossRef]
38. Moftakhari, H.R.; Salvadori, G.; AghaKouchak, A.; Sanders, B.F.; Matthew, R.A. Compounding effects of sea
level rise and fluvial flooding. Proc. Natl. Acad. Sci. USA 2017, 114, 9785–9790. [CrossRef] [PubMed]
39. Van den Hurk, B.; van Meijgaard, E.; de Valk, P.; van Heeringen, K.-J.; Gooijer, J. Analysis of a compounding
surge and precipitation event in the Netherlands. Environ. Res. Lett. 2015, 10, 035001. [CrossRef]
40. Neal, J.; Keef, C.; Bates, P.; Beven, K.; Leedal, D. Probabilistic flood risk mapping including spatial
dependence. Hydrol. Process. 2013, 27, 1349–1363. [CrossRef]
41. Keef, C.; Tawn, J.A.; Lamb, R. Estimating the probability of widespread flood events. Environmetrics 2013, 24,
13–21. [CrossRef]
42. Bender, J.; Wahl, T.; Müller, A.; Jensen, J. A multivariate design framework for river confluences. Hydrol. Sci. J.
2016, 61, 471–482. [CrossRef]
43. Apel, H.; Thieken, A.H.; Merz, B.; Blöschl, G. Flood risk assessment and associated uncertainty. Nat. Hazards
Earth Syst. Sci. 2004, 4, 295–308. [CrossRef]
44. Lamb, R.; Keef, C.; Tawn, J.; Laeger, S.; Meadowcroft, I.; Surendran, S.; Dunning, P.; Batstone, C. A new
method to assess the risk of local and widespread flooding on rivers and coasts. J. Flood Risk Manag. 2010, 3,
323–336. [CrossRef]
45. Port of Houston Authority. Available online: http://www.porthouston.com/ (accessed on 27 March 2018).
46. Davlasheridze, M.; Atoba, K.O.; Brody, S.; Highfield, W.; Merrell, W.; Ebersole, B.; Purdue, A.; Gilmer, R.W.
Economic impacts of storm surge and the cost-benefit analysis of a coastal spine as the surge mitigation
strategy in Houston-Galveston area in the USA. Mitig. Adapt. Strateg. Glob. Chang. 2018, 1–26. [CrossRef]
47. Christian, J.; Fang, Z.; Torres, J.; Deitz, R.; Bedient, P. Modeling the Hydraulic Effectiveness of a Proposed
Storm Surge Barrier System for the Houston Ship Channel during Hurricane Events. Nat. Hazards Rev. 2015,
16, 04014015. [CrossRef]
48. Torres, J.M.; Bass, B.; Irza, J.N.; Proft, J.; Sebastian, A.; Dawson, C.; Bedient, P. Modeling the Hydrodynamic
Performance of a Conceptual Storm Surge Barrier System for the Galveston Bay Region. J. Waterw. Port Coast.
Ocean Eng. 2017, 143, 05017002. [CrossRef]
49. Niedermayer, D. An Introduction to Bayesian Networks and Their Contemporary Applications; Innovations in
Bayesian Networks; Springer: Berlin/Heidelberg, Germany, 2008; pp. 117–130. ISBN 9783540850656.
50. Vogel, K.; Riggelsen, C.; Korup, O.; Scherbaum, F. Bayesian network learning for natural hazard analyses.
Nat. Hazards Earth Syst. Sci. 2014, 14, 2605–2626. [CrossRef]
51. Weber, P.; Medina-Oliva, G.; Simon, C.; Iung, B. Overview on Bayesian networks applications for
dependability, risk analysis and maintenance areas. Eng. Appl. Artif. Intell. 2012, 25, 671–682. [CrossRef]
52. Hanea, A.; Morales Napoles, O.; Ababei, D. Non-parametric Bayesian networks: Improving theory and
reviewing applications. Reliab. Eng. Syst. Saf. 2015, 144, 265–284. [CrossRef]
53. Hanea, A.M.; Kurowicka, D.; Cooke, R.M. Hybrid Method for Quantifying and Analyzing Bayesian Belief
Nets. Qual. Reliab. Eng. Int. 2006, 22, 709–729. [CrossRef]
54. Salvadori, G.; De Michele, C. Frequency analysis via copulas: Theoretical aspects and applications to
hydrological events. Water Resour. Res. 2004, 40, 1–17. [CrossRef]
Water 2018, 10, 1190 18 of 19

55. Genest, C.; Favre, A.-C. Everything You Always Wanted to Know about Copula Modeling but Were Afraid
to Ask. J. Hydrol. Eng. 2007, 12, 347–368. [CrossRef]
56. Joe, H. Dependence Modeling with Copulas; Chapman & Hall/CRC: London, UK, 2015; ISBN 9781466583221.
57. Graler, B.; Van Den Berg, M.J.; Vandenberghe, S.; Petroselli, A.; Grimaldi, S.; De Baets, B.; Verhoest, N.E.C.
Multivariate return periods in hydrology: A critical and practical review focusing on synthetic design
hydrograph estimation. Hydrol. Earth Syst. Sci. 2013, 17, 1281–1296. [CrossRef]
58. Sklar, A. Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Stat. Univ. Paris 1959, 8, 229–231.
59. Nelsen, R.B. An Introduction to Copulas, 2nd ed.; Springer Science & Business Media: New York, NY,
USA, 2006.
60. Joe, H. Multivariate Models and Multivariate Dependence Concepts; CRC Press: London, UK, 1997.
61. Wang, W.; Wells, M.T. Model Selection and Semiparametric Inference for Bivariate Failure-Time Data. J. Am.
Stat. Assoc. 2000, 95, 62–72. [CrossRef]
62. Paprotny, D.; Morales-Nápoles, O. Estimating extreme river discharges in Europe through a Bayesian
network. Hydrol. Earth Syst. Sci. 2017, 21, 2615–2636. [CrossRef]
63. Qian, Z. Without zoning: Urban development and land use controls in Houston. Cities 2010, 27, 31–41.
[CrossRef]
64. Gori, A.; Blessing, R.; Juan, A.; Brody, S.; Bedient, P. Characterizing urbanization impacts on floodplain
through integrated land use, hydrologic, and hydraulic modeling. J. Hydrol. 2018, in press.
65. Gori, A. Quantifying Impacts of Development on Floodplain Evolution and Projection of Future Flood
Hazard: Applications to Harris County. Master’s Thesis, Rice University, Houston, TX, USA, 2018.
66. Mallakpour, I.; Villarini, G. The changing nature of flooding across the central United States. Nat. Clim. Chang.
2015, 5, 250–254. [CrossRef]
67. Villarini, G.; Smith, J.A.; Baeck, M.L.; Krajewski, W.F. Examining Flood Frequency Distributions in the
Midwest US. JAWRA J. Am. Water Resour. Assoc. 2011, 47, 447–463. [CrossRef]
68. Serinaldi, F. Can we tell more than we can know? The limits of bivariate drought analyses in the United
States. Stoch. Environ. Res. Risk Assess. 2015, 30, 1691–1704. [CrossRef]
69. Claps, P.; Laio, F. Can continuous streamflow data support flood frequency analysis? An alternative to the
partial duration series approach. Water Resour. Res. 2003, 39, 1–11. [CrossRef]
70. Hobaek Haff, I.; Frigessi, A.; Maraun, D. How well do regional climate models simulate the spatial
dependence of precipitation? An application of pair-copula constructions. J. Geophys. Res. D Atmos.
2015, 120, 2624–2646. [CrossRef]
71. Serinaldi, F.; Bárdossy, A.; Kilsby, C.G. Upper tail dependence in rainfall extremes: would we know it if we
saw it? Stoch. Environ. Res. Risk Assess. 2015, 29, 1211–1233. [CrossRef]
72. Morales-Nápoles, O.; Worm, D.; van den Haak, P.; Hanea, A.; Courage, W.; Miraglia, S. Reader for Course:
Introduction to Bayesian Networks; TNO-060-DTM-2013-01115; TNO: Delft, The Netherlands, 2013.
73. Mutua, F.M. The use of the Akaike Information Criterion in the identification of an optimum flood frequency
model. Hydrol. Sci. J. 1994, 39, 235–244. [CrossRef]
74. Morales-Nápoles, O.; Steenbergen, R.D.J.M. Large-Scale Hybrid Bayesian Network for Traffic Load Modeling
from Weigh-in-Motion System Data. J. Bridg. Eng. 2014, 20, 1–10. [CrossRef]
75. McLachlan, G.; Peel, D. Finite Mixture Models; Wiley Interscience: Hoboken, NJ, USA, 2000.
76. NOAA National Oceanic and Atmospheric Administration. Tides and Currents. Annual Exceedance
Probability Curves 8771450 Galveston Pier 21, TX. Available online: https://tidesandcurrents.noaa.gov/est/
curves.shtml?stnid=8771450 (accessed on 15 December 2017).
77. Henderson, F.M. Open Channel Flow; Prentice-Hall: Upper Saddle River, NJ, USA, 1966; ISBN 0023535105.
78. Brunner, G.W. HEC-RAS River Analysis System. Hydraulic Reference Manual; US Army Corps of Engineers,
Institute for Water Resources, Hydrologic Engineering Center: Davis, CA, USA, 2016.
79. Harris County Flood Control District Model and Map Management System. Available online: http://www.
m3models.org/ (accessed on 1 July 2016).
80. Landsea, C.W.; Franklin, J.L. Atlantic Hurricane Database Uncertainty and Presentation of a New Database
Format. Mon. Weather Rev. 2013, 141, 3576–3592. [CrossRef]
81. Arns, A.; Wahl, T.; Haigh, I.D.; Jensen, J.; Pattiaratchi, C. Estimating extreme water level probabilities:
A comparison of the direct methods and recommendations for best practise. Coast. Eng. 2013, 81, 51–66.
[CrossRef]
 Water 2018, 10, 1190 19 of 19

82. Arns, A.; Dangendorf , S.; Jensen, J.; Talke, S.; Bender, J.; Pattiaratchi, C. Sea-level rise induced amplification
of coastal protection design heights. Sci. Rep. 2017, 7, 40171. [CrossRef] [PubMed]
83. Vitousek, S.; Barnard, P.L.; Fletcher, C.H.; Frazer, N.; Erikson, L.; Storlazzi, C.D. Doubling of coastal flooding
frequency within decades due to sea-level rise. Sci. Rep. 2017, 7, 1399. [CrossRef] [PubMed]
84. Teng, J.; Jakeman, A.J.; Vaze, J.; Croke, B.F.W.; Dutta, D.; Kim, S. Flood inundation modelling: A review of
methods, recent advances and uncertainty analysis. Environ. Model. Softw. 2017, 90, 201–216. [CrossRef]
85. Van Oldenborgh, G.J.; van der Wiel, K.; Sebastian, A.; Singh, R.; Arrighi, J.; Otto, F.; Haustein, K.; Li, S.;
Vecchi, G.; Cullen, H. Attribution of extreme rainfall from Hurricane Harvey, August 2017. Environ. Res. Lett.
2017, 12, 124009. [CrossRef]
86. Sebastian, A.G.; Lendering, K.T.; Kothuis, B.L.M.; Brand, A.D.; Jonkman, S.N. Hurricane Harvey Report.
A Fact-Finding Effort in the Direct Aftermath of Hurricane Harvey in the Greater Houston Region; Delft University
Publishers: Delft, The Netherlands, 2017.
87. Jonkman, S.N.; Godfroy, M.; Sebastian, A.; Kolen, B. Brief communication: Post-event analysis of loss of life
due to hurricane Harvey. Nat. Hazards Earth Syst. Sci. 2018, 18, 1073–1078. [CrossRef]
88. Serafin, K.A.; Ruggiero, P.; Stockdon, H.F. The relative contribution of waves, tides, and nontidal residuals
to extreme total water levels on U.S. West Coast sandy beaches. Geophys. Res. Lett. 2017, 44, 1839–1847.
[CrossRef]
89. Vousdoukas, M.I.; Bouziotas, D.; Giardino, A.; Bouwer, L.M.; Mentaschi, L.; Feyen, L. Understanding
epistemic uncertainty in large-scale coastal flood risk assessment for present and future climates. Nat. Hazards
Earth Syst. Sci. 2018, 18, 2127–2142. [CrossRef]
90. Melito, L.; Postacchini, M.; Sheremet, A.; Calantoni, J.; Zitti, G.; Darvini, G.; Brocchini, M. Wave-Current
Interactions and Infragravity Wave Propagation at a Microtidal Inlet. Proceedings 2018, 2, 628. [CrossRef]
91. Sebastian, A.; Proft, J.; Dietrich, J.C.; Du, W.; Bedient, P.B.; Dawson, C.N. Characterizing hurricane storm
surge behavior in Galveston Bay using the SWAN+ADCIRC model. Coast. Eng. 2014, 88, 171–181. [CrossRef]
92. Dietrich, J.C.; Zijlema, M.; Westerink, J.J.; Holthuijsen, L.H.; Dawson, C.; Luettich, R.A.; Jensen, R.E.;
Smith, J.M.; Stelling, G.S.; Stone, G.W. Modeling hurricane waves and storm surge using integrally-coupled,
scalable computations. Coast. Eng. 2011, 58, 45–65. [CrossRef]
93. Salvadori, G.; Tomasicchio, G.R.; D’Alessandro, F. Practical guidelines for multivariate analysis and design
in coastal and off-shore engineering. Coast. Eng. 2014, 88, 1–14. [CrossRef]
94. Pappadà, R.; Durante, F.; Salvadori, G. Quantification of the environmental structural risk with spoiling ties:
Is randomization worthwhile? Stoch. Environ. Res. Risk Assess. 2017, 31, 2483–2497. [CrossRef]
95. Emanuel, K. Assessing the present and future probability of Hurricane Harvey’s rainfall. Proc. Natl. Acad.
Sci. USA 2017, 201716222. [CrossRef] [PubMed]
96. Hamilton, C.M.; Martinuzzi, S.; Plantinga, A.J.; Radeloff, V.C.; Lewis, D.J.; Thogmartin, W.E.; Heglund, P.J.;
Pidgeon, A.M. Current and Future Land Use around a Nationwide Protected Area Network. PLoS ONE
2013, 8. [CrossRef] [PubMed]
97. Hollis, G.E. The effect of urbanization on floods of different recurrence interval. Water Resour. Res. 1975, 11,
431–435. [CrossRef]
98. Sarhadi, A. Time-varying nonstationarymultivariate risk analysis using a dynamic Bayesian copula.
Water Resour. Res. 2016, 52, 2327–2349. [CrossRef]

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

View publication stats