A Bayesian Network methodology for coastal hazard assessments on a regional scale: The BN-CRAF

Journal Pre-proof A Bayesian Network methodology for coastal hazard assessments on a regional scale: The BN-CRAF M. Sanuy, J.A. Jiménez, N. Plant PII:

S0378-3839(19)30170-X

DOI:

https://doi.org/10.1016/j.coastaleng.2019.103627

Reference:

CENG 103627

To appear in:

Coastal Engineering

Received Date: 10 April 2019 Revised Date:

18 November 2019

Accepted Date: 29 December 2019

Please cite this article as: Sanuy, M., Jiménez, J.A., Plant, N., A Bayesian Network methodology for coastal hazard assessments on a regional scale: The BN-CRAF, Coastal Engineering (2020), doi: https://doi.org/10.1016/j.coastaleng.2019.103627. This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 Published by Elsevier B.V.

1

A Bayesian Network methodology for coastal hazard assessments on a regional scale:

2

the BN-CRAF M. Sanuy1*, J. A. Jiménez1, N. Plant2

3

1

4

c/Jordi Girona 1-3, Campus Nord ed D1, 08034 Barcelona, Spain.

5

2

6

*Corresponding author: Marc Sanuy, [email protected]

Laboratori d’Enginyeria Marítima, Universitat Politècnica de Catalunya, BarcelonaTech,

United States Geological Survey, Saint Petersburg, FL, USA.

7

8

Abstract

9

Hazard assessment is one of the key elements to be included in any coastal risk assessment

10

framework. Characterizing storm-induced erosion and inundation involves the assessment of

11

the coastal response under the forcing of a stochastic source (the storm), acting on a variable

12

morphology (the beach) and inducing some damages. Hazard assessment under any present or

13

future scenario will be affected by uncertainties either associated to the models used, the

14

definition of climate conditions, and the characterization of the coastal morphology. In this

15

context, Bayesian Networks (BN) can effectively address the problem as they allow

16

accounting for these uncertainties while characterizing stochastically the system response and

17

giving insight on the dependencies among involved variables. In this work, a BN-based

18

methodology for storm-induced hazard assessment at regional scale is presented. The

19

methodology is able to account for uncertainties associated with included models and forcing

20

conditions through Monte-Carlo simulations. It produces distributions of erosion and

21

inundation hazards under given scenarios allowing conditioned hazard assessments as a

22

function of storm and morphological variables. Results are compared to hazards evaluated

1

23

using an existing Coastal Risk Assessment Framework (CRAF), at two locations of the

24

Catalan coast already identified as hotspots for storm-induced erosion and/or flooding.

25

Keywords: Bayesian Networks, Coastal hotspots, Monte Carlo, uncertainty, inundation,

26

erosion, storms, SLR.

27

28

1. Introduction

29

As coastal regions continue to grow in population size and density, they are also

30

increasingly exposed to coastal storm-induced hazards, such as inundation and erosion (IPCC,

31

[1, 2]). The projections of rising sea levels, the existence of background coastal erosion and

32

other concerns related to climate change, such as increases in the magnitude and/or frequency

33

of storms (Lionello et al. [3], Conte and Lionello [4]; IPCC, [5]) or changes in the

34

directionality of incoming waves (Cases-Prat and Sierra, [6]), highlight the need for hazard

35

assessments that consider these factors. Applications of these assessments include designing

36

coastal management plans, which will require a specific chapter on coastal risks, as

37

recognized in the protocol of Integrated Coastal Zone Management in the Mediterranean plan

38

(UNEP/MAP/PAP, [7]). This may include the need to identify coastal hotspots at regional

39

scale, to compare assessments for different scenarios, and the development of early warning

40

systems (EWS) among other elements (e.g., Ciavola et al., [8, 9]; Vojinovic et al., [10];

41

Oumeraci et al., [11]; Van Dongeren et al., [12]). The problem to be solved with these

42

approaches is complex due to its multidimensionality (both in terms of multiple and

43

interdependent hazards and vulnerabilities) and their multiscale nature (both in space and

44

time). Moreover, hazard estimation comes with a number of associated uncertainties that can

45

become especially large when dealing with future projections (Vousdoukas et al., [13]).

2

46

A widely used approach in risk assessment is the Source-Pathway-Receptor (SPR)

47

model (Sayers et al., [14]; Narayan et al. [15]). This is a conceptual model which describes

48

how a given risk propagates across a given domain from the source to the receptors. In this

49

particular case, it is applied to storm-induced hazards (erosion and inundation) assessment,

50

and the problem is schematized in terms of a source (storms), the pathway (beach or coastal

51

morphology) and receptors (elements of interest) at the coast. To this end and under the

52

common scarcity of direct observations, storm-induced hazards are usually assessed by using

53

predictive models which are fed with information on both the source and the pathway.

54

However, the implementation of a suite of complex numerical models at a high resolution

55

(e.g., 1~10 m spatially, 1 s ~ 1 h temporally) at the regional scale (e.g., hundreds of km) may

56

not be feasible or time/cost efficient. Furthermore, the probabilistic definition of the hazard

57

component should preferably be performed based on the response (Garrity et al. [16];

58

Callaghan et al. [17]; Sanuy et al., [18]) due to the nonlinear and multidimensional

59

dependencies involved in the driving processes (see e.g., Hawkes et al., [19]; Masina et al.,

60

[20]; Lin-Ye et al., [21]). This implies the simulation of the erosion and inundation induced

61

by a large set of conditions characterizing the existing storm climate. Additionally, model

62

error is an important source of uncertainty that should be included in the assessment. Simple

63

parametric models which use only bulk information on the source (e.g. peak Hs, Tp, direction

64

and duration) and the pathway (e.g., slope, grain size, or beach height and width) can be a

65

suitable option to generate sufficient hazard estimations at a reduced computational expense

66

(e.g., Stockdon et al., [22]; Mendoza and Jiménez, [23]). The drawback is the simplification

67

of the storm characteristics to a set of parameters and the simplification of the morphology to

68

the bulk definition of a 1-D cross-shore profile. When a long coastal stretch (~1km) is

69

represented by the bulk characteristics of a single profile, the pathway is treated

3

70

deterministically, as neither spatial (alongshore) nor temporal (seasonal) variability is

71

included in the assessment.

72

All this makes that any robust methodology for hazard assessments needs both to

73

account for the existing variability in source and pathway and to tackle model uncertainties.

74

However, this has to be done in a reasonable cost-effective manner, and it is in this aspect

75

where Bayesian Networks (BNs) have demonstrated their versatility and utility in efficiently

76

combining multiple variables to predict system behaviour. They have been used e.g. in

77

groundwater flow predictions supporting decision making (Fienen et al. [24]), habitat

78

protection against natural hazards (Palmsten et al. [25]; Gieder et al. [26]), coastal

79

vulnerability and shoreline evolution to sea level rise (Gutierrez et al. [27]; Plant et al. [28]),

80

postevent hazard and damage assessment (Van Verseveld et al. [29]; Poelhekke et al. [30])

81

and disaster risk reduction assessments (Plomaritis et al. [31], Sanuy et al. [32]). BNs can

82

handle nonlinear systems by means of conditional probability tables, which represent the

83

interdependencies of the variables involved, and are solved through data training. They are

84

not computationally expensive, can work with data from different sources (e.g., modeled,

85

observed, or even opinions) and have a simple and intuitive graphical structure that is easily

86

understood by nontechnical users (Uuscitalo, [33]). Notably, they can be used to represent the

87

SPR scheme through the dependency relations that physically, or even psychologically, exist

88

between the different steps (e.g., Straub [34], Jäger et al. [35]) and thus, they can be adapted

89

to assess many kinds of natural hazards and impacts on many kinds of receptors.

90

The general aim of this work is to develop a methodology for the assessment of storm-

91

induced erosion and inundation at regional scale with the purpose of hotspot identification

92

and characterization under present conditions and future scenarios. Within this context, the

93

development and implementation of a BN-based hazard assessment methodology is

94

presented. It is based on the Coastal Risk Assessment Framework (CRAF) phase 1 4

95

methodology (Viavattene et al., [36]), which has been validated and applied across different

96

study sites within the RISC-KIT project (Van Dongeren et al. [12]; Ferreira et al. [37]),

97

thereby indicating its robustness. The CRAF phase 1 combines a probabilistic treatment of the

98

source (storms) with a deterministic treatment of the pathway (coastal morphology), without

99

including model error uncertainties. The here presented new methodology is able to deal with

100

the small-scale variability of coastal morphology (20-30 profiles per kilometre) and with

101

model uncertainties by using Monte-Carlo simulations based on known model errors. The

102

method is applied under current conditions and given scenarios defined in terms of SLR and

103

observed rates of background erosion, where Monte-Carlo simulations are used to incorporate

104

their associated variability. The proposed methodology, while maintaining a relatively simple

105

structure, represents an advantage with respect to other existing approaches where the

106

treatment of the source is deterministic or do not perform the statistics focusing on the hazard

107

variables (e.g., use of hurricane categories as levels, Stockdon et al., [28]; use of event

108

approach, e.g., Villatoro et al., [39]; Armaroli and Duo, [40]; use of a homogenous

109

representation of the source Poelhekke et al. [30]). Moreover, by accounting for the spatial

110

variability of the coastal morphology, it also represents an improvement with respect to

111

existing methodologies adopting a deterministic treatment of the pathway (Callaghan, [17];

112

Bosom and Jimenez, [41]; Ballesteros et al. [42], Viavattene et al., [36]). Notably, none of the

113

referenced works included model errors in the analysis.

114

This methodology is applied at two sectors of the Catalan coast (NW Mediterranean),

115

and obtained results are compared to those of the semideterministic CRAF phase 1 (Jiménez

116

et al., [43]). One sector is a 2 km long fully rigidized coastline composed by a rip-rap

117

revetment protecting a coastal railway. The other one is a 3 km long sandy coastal fringe

118

protecting a low lying deltaic area with campsites in the hinterland. Both locations have been

119

identified as coastal hotspots to storm impacts in Jiménez et al. [43]. 5

120

The structure of the paper is as follows: (i) the second section describes the data used,

121

(ii) the third section presents the BN- based proposed methodology, (iii) the fourth section

122

shows and example of the application of the method, followed by (iv) the discussion and

123

comparison with CRAF phase 1 and (v) the main conclusions.

124

125

2. Data

126

To perform this analysis, data on waves and water levels to characterize the forcing,

127

and on coastal morphology to characterize the pathway, are required. The assessment of

128

coastal inundation and erosion hazards requires a long-term series of wave conditions and

129

water levels that have the appropriate spatial and temporal coverage. This work uses hindcast

130

waves from the Downscaled Ocean Waves dataset (Camus et al., [44]) derived from the

131

Global Ocean Waves (Reguero et al., [45]) and hindcast surges from the Global Ocean Surge

132

dataset (Cid et al. [46]), which were obtained at a node located in front of the Tordera Delta at

133

~20 m depth and cover the period 1950-2014. Simultaneous mean water levels were available

134

at the same resolution of the GOS dataset.

135

The coastal morphology has been represented by using LIDAR-derived topography

136

from the Institut Cartogràfic i Geològic de Catalunya. The data were provided as high-

137

resolution digital elevation models (DEMs) with 1m x 1m grid cells and a vertical precision

138

of 2-3 cm (Ruiz et al. [47]). This was complemented with data obtained from a topo-

139

bathymetric survey provided by the Ministry of Agriculture, Fish, Food and Environment.

140

Data up to 1.5 m water depth were obtained by total station topographic surveying and, data

141

down to 50 m water depth were recorded with a dual-frequency echosounder with a nominal

142

accuracy of 7 cm.

6

143

The long-term shoreline evolution is obtained from a 25-year shoreline dataset

144

(Jiménez and Valdemoro, [48]) and the global SLR projections correspond to the RCP8.5

145

IPCC [5].

146

147

3. The BN-CRAF methodology

148

In the following, the hazard assessment general scheme is introduced while presenting

149

the different steps involved in the BN-based methodology and comparing with the

150

corresponding steps of the previously developed CRAF phase 1 methodology. Then, methods

151

followed in each step are described in detail.

152

153 154 155

3.1. The methodological scheme The general methodological scheme follows CRAF phase 1 (Viavattene et al., [36]) which was applied to the Catalan coast in Jimenez et al., [43].

156

In CRAF phase 1 (Figure 1A), the source is characterized by identifying all the storms

157

in a long (e.g. 60 years) time series of waves and water levels. The pathway (morphology) is

158

characterized by a single cross-shore profile for each ~1km coastal sector. This sector-

159

characteristic profile can be calculated by simple beach profile averaging, or by choosing the

160

worst-case profile (e.g. lowest dune and narrowest beach). The second option was applied in

161

Jiménez et al. [43]. Then, parametric models are used to estimate hazard magnitudes (e.g.

162

total water level at the coast for inundation, and shoreline retreat for erosion) for each storm.

163

Obtained hazard values are fitted by an extreme probability distribution, and finally, erosion

164

and inundation magnitudes associated with selected probabilities/return periods are

165

transformed to hazard indicators (see e.g. Ferreira et al., [49]) within a scale ranging from 0 to 7

166

5, being 0 no hazard and 5 the most hazardous situation. These indicators combine

167

information of the hazard magnitude with basic information on the relative exposure of the

168

hinterland to each hazard (e.g. beach width for erosion and, dune/berm height for inundation).

169

The final output is, thus, a single intensity value per sector for each hazard associated with a

170

given return period.

171 172 173

Figure 1. Conceptual scheme of the CRAF-phase 1 approach (A), compared to the BN-based methodology (B).

174

8

175

In the here presented BN-based methodology (Figure 1B), the source is characterized

176

as in CRAF phase 1, but the pathway is characterized by using a set of profiles (~10-30,

177

Figure 2) to properly capture the existing morphological variability within each sector.

178

Hazards magnitude is computed by means of parametric models for each storm and profile.

179

To account for the uncertainty associated with model errors, Monte-Carlo simulations

180

(Hastings, [50]) are used. Thus, for each profile and storm, multiple erosion and inundation

181

hazard values are calculated, which are later transformed to hazard indicators using the same

182

scale as in CRAF phase 1. These hazard estimations together their corresponding source

183

(storm characteristics) and pathway (profile characteristics) data are used to train the BN. The

184

final output are probabilistic distributions of the different erosion and inundation hazard

185

indexes at each sector, which consider model uncertainties and preserve the information on

186

conditional dependencies with source and pathway.

187

In addition to detecting hotspots along the coast, the BN methodology is used to assess

188

hazard values under given scenarios associated with future global SLR and local background

189

erosion projections, where the Monte-Carlo approach is applied to simulate multiple

190

conditions per scenario.

191

192

3.2. Source and pathway characterization.

193

For comparison purposes, the areas to be studied are divided into ~1 km sectors, with the

194

same spatial division as in the CRAF-phase1 application (Jimenez et al., [43]). Each sector is

195

represented by a number of profiles (Figure 2). Profile selection is aimed to capture the

196

existing morphological variability in each sector, in terms of slope, beach height (i.e. dune or

197

upper berm or embankment height), and beach width (Figure 3b). The shape of the first part

198

of the hinterland is also taken into account for profile selection. Notably, rigidized coastal

9

199

sectors such as Cabrera South (CS) and Cabrera North (CN) require a lower number of

200

profiles (16 and 9 respectively) than natural beach sectors with larger variability such as

201

Malgrat North (MN) and S’Abanell (SA) with 20 and 32 profiles respectively. Malgrat South

202

(MS) is more homogeneous and was characterized with 15 profiles.

203

204 205 206 207

Figure 2. Locations of the wave and surge data nodes used in this work. Sector limits and used profiles are indicated. The red dot highlights the location of the numerical node for wave and surge data.

208 209

Coastal storms have been identified in the dataset (see location of the node in Figure 2) by

210

means of a double threshold P.O.T analysis as similarly conducted in Sanuy et al. [18]. The

211

0.98 quantile (Hs = 2 m) was used as the first threshold to characterize storms by controlling

212

the event duration. The 0.995 quantile (Hs = 2.6 m) was then applied to the subset to identify

213

the extreme events. The resulting dataset is composed of 179 storms (~3 storms per year), in

214

which the complete hourly evolution of the different wave conditions is stored, i.e.,

215

significant wave height (Hs), peak period (Tp), storm surge and wave direction. The storm 10

216

duration is calculated as the time during which Hs exceeds the selected threshold. Figure 3a

217

shows scatterplots of the main characteristics of the dataset.

218

219 220 221 222

Figure 3. Dataset characteristics. Storm bulk parameters (left) Hs vs Tp, colored by storm duration, and markers by storm duration. Main morphological features (right), beach height vs beach width, colored by beach slope and markers by sector.

223 224

3.3. Hazard indicators

225

In this work, the hazards have been estimated similarly to that of Bosom and Jiménez

226

[41] and applied similarly to Jimenez et al. [43]. The inundation potential is calculated using,

227

as proxy, the total water level at the coast (TWL), characterized by the wave-induced run-up

228

and the surge level. The run-up models used are the Stockton et al. [22] model [eq 1] for

229

sectors composed of sandy beaches and the EurOtop [51] model [eq 2] for rigidized sectors

230

with protected revetments. These models use significant wave height in deepwaters (Hs) or

231

breaking conditions (Hb), wave length in deepwaters (Lo) and beach slope (tanβ, calculated

232

from the -1 to 1 m elevations) to estimate the vertical level potentially reached by waves at

233

the coast. The EurOtop [51] model also includes factors accounting for wave directionality

234

(γβ) and surface friction (γf). 11

235

236

[

]

1/ 2  Hs Lo (0.563 tan β 2 + 0.004) 1/ 2  Ru = 1.1 0.35 tan β ( Hs Lo) +  2 

= 1.65 ∗

∗

∗

∗

  *γ ,  β 

,

(1)

(2)

237

The erosion hazard has been estimated by using the Mendoza and Jimenez [36] model,

238

which has been widely applied to estimate storm-induced erosion potential (see e.g., Armaroli

239

and Duo, [39]; Silva, [52]). The model is given by the following:

240

∆ = 7,9 ∗

241

where ∆ is the eroded volume at the beach face (Figure 4), Tp is the wave period at the peak

242

and wf is the sediment fall velocity.

243

∆3 = 5678∗

244

where ∆X is the beach retreat calculated as a function of the eroded volume, the beach height

245

(ZB) and the pivot point depth (d*), (Figure 4).

∆4

∗

!"# + 3.6 , '! ℎ

= )4 −

,

,-

./∗01

)∗

#2 ,

(3)

(4)

246 247 248

Figure 4. Profile erosion schematization used to derive beach retreat from eroded volume (adapted from Mendoza and Jiménez [36]). 12

249 250

Finally, hazard estimations are converted into a 0-5 hazard intensity scale, as in

251

Jiménez et al. [43]. For the inundation hazard, the TWL is compared against the profile to

252

obtain the potential reach of the flooding by means of the bathtub approach. This magnitude is

253

then compared with the beach width (BW) to derive the hazard index (Table 1). The erosion

254

index, in a different manner, is derived by comparing the resulting post-event BW (i.e., the

255

pre-event BW subtracting the storm-induced retreat) with a reference width. Following

256

Jiménez et al. [43], the reference width of the different levels is based on the average beach

257

retreat for the 10-year return period at the study site ∆X10, with a value of 7.5 meters (Table

258

1).

259 260 261 262 263

Table 1. The hazard index levels as a function of the beach width (BW) against inundation reach or beach retreat. All units are in meters, and ∆X10 stands for average profile retreat associated with TR=10 years for the site, which has a value of 7.5 m. Hazard index level 0 1 2 3 4 5

Inundation Reach < 0.5*BW 0.5* BW BW + 60

Erosion BW – Retreat >4 ∆X10 4 ∆X10< BW – Retreat <3 ∆X10 3 ∆X10< BW – Retreat <2 ∆X10 2 ∆X10< BW – Retreat < ∆X10 ∆X10< BW – Retreat <0 BW – Retreat ≤ 0

264 265 266

3.4. Future scenarios

267

The assessment of the possible evolution of hazards requires having information on

268

future morphology (pathway) and forcing (source) conditions. In the present application, this

269

is done by estimating how such elements will vary under a given scenario by a given time

270

horizon. As an example, to assess how hazard will evolve at mid-term scale (from years to

271

few decades), we have considered the current decadal-scale coastline evolution to build the

13

272

future morphology (beach width) at selected time horizons while using Mendoza and Jimenez

273

(2006) to assess the corresponding-storm induced retreats. When the analysis is extended at

274

the long-term scale (several decades to centuries), we have to consider potential changes in

275

forcing conditions. If storminess is not expected to change significantly (see e.g., Somot et al.

276

[53] for the Mediterranean basin or Bender et al. [54] for Atlantic hurricanes), the main

277

variable affecting analysed hazards is the long-term change in sea level. In such a case, this is

278

done by considering the future mean sea level under a given SLR scenario by the selected

279

time horizons. In the case that long-term changes in storminess are expected, storm properties

280

must be modified accordingly.

281

In this study, mid-term background erosion along the sandy beach sectors has been

282

estimated by analysing shoreline position changes from aerial photographs available at

283

different time frequencies that cover the last 25-30 years. The estimated average retreat for the

284

3 sandy beach sectors are 1.1 m/y at SA, 4 m/y at MN and 1.9 m/y at MS (Jiménez and

285

Valdemoro, [48]; see Figure 2 for locations).

286

To introduce SLR-induced effects, the IPCC [5] RCP8.5 scenario was considered in

287

this study, which accounts to 0.30 m and 0.63 m of global mean SLR by the years 2050 and

288

2100, respectively, with respect to the period 1986-2005.

289 290

3.5. Monte-Carlo approach

291

All hazard models and future projections are affected by errors and uncertainties. The

292

available knowledge on these errors, or reasonable assumptions on the uncertainties, may be

293

included in this hazard assessment. Known model root mean squared error (RMSE), and

294

confidence ranges (Table 2) have been included in hazard estimations by means of Monte-

14

295

Carlo simulations (Hastings, [50]). Other existing uncertainties may also be included

296

following a similar procedure.

297

Uncertainty in the run-up has been estimated as follows. For the Stockdon [22] model,

298

we use the RSME of the formula as an estimator of the standard deviation of the variable

299

(Plant et al. [55]). It is assumed that the variable follows a normal distribution with the mean

300

calculated by [eq 1] and a standard deviation equal to the model RSME (0.32 m). For the

301

EurOtop [51] model, the knowledge of the standard deviation of the coefficient in [eq 2]

302

(Table 2) is included in the assessment along with an estimation of the uncertainty of the

303

actual value of the revetment friction coefficient.

304

Uncertainty of erosion potential (∆V) has been similarly estimated. First, the RMSE of

305

the eroded volume given by the Mendoza and Jiménez [23] model is used. Then, the

306

uncertainty on the real height of the eroded profile (i.e., distance from submerged pivot point

307

to beach top) is included by assessing the variability of the pivot point position assessed from

308

simulations with XBeach 1D under different storm intensities (d*, Table 2).

309

Confidence ranges on projected conditions under future scenarios are included in the

310

assessment. In this case, the 95% confidence interval of the SLR estimates (IPCC, [5]) is used

311

to derive the standard deviation of the variable, assuming a normal distribution. The standard

312

deviation of shoreline evolution rates within each sector (Jiménez and Valdemoro, [48]) was

313

used to simulate multiple values of beach width per future scenario (Table 2).

314 315 316 317 318

Table 2. Ranges of the uncertain variables included in the assessment. Note that total number of cases is always 100 for hazard variables and 20 for future scenario variables The total number of cases can be the result of the simulation of a single variable or of the combination of simulations from 2 variables.

319

Tordera Delta- S’Abanell (SA), Malgrat North (MN) and South (MS) Variable Mean and std. dev. Monte-Carlo Total number 15

simulations 100 10 10

Ru on [eq 1] ∆V on [eq 3] d* on [eq 4]

[eq 1] ± 0.32 (m) [eq 3] ± 10 m3 0.8 ± 0.15 m SB: 1.1 ± 0.8 m/yr Retreat trends MN: 4.0 ± 0.5 m/yr 20 MS: 1.9 ± 2.1 m/yr Cabrera de Mar North (CN) and South (CS) Monte-Carlo Variable Mean and std. dev. simulations Coef. (1.65) on [eq 2] 1.65 ± 0.10 10 γf on [eq 2] 0.65 ± 0.10 10 SLR

2050: 0.30 ± 0.04 (m) 2100: 0.63 ± 0.10 (m)

20

of Cases 100 (Ru) 100 (∆X) 20 (BW)

Total number of Cases 100 (Ru) 20 (MSL)

320 321

Therefore, following the Monte-Carlo example, normally distributed values of the

322

uncertain variables for each combination of storm and profile characteristics are simulated.

323

This extends uniformly to the dataset maintaining the natural variability of the storm climate

324

(source) and the morphology (pathway). The total number of cases introduced in the BN

325

training is 100 values of erosion and inundation hazards per profile and storm combined with

326

20 values of MSL or beach width per future scenario.

327 328

3.6. Bayesian Network

329

Bayesian Networks (BN) are probabilistic models based on acyclic graphs (see e.g.,

330

Pearl, [56]; Jensen, [57]). They use Bayes’ probability theory [eq 5] to describe the

331

relationships between different variables.

332

p[Oi|Pi] = p[Pi|Oi]*p[Oi]/p[Pi]

(5)

333

In this application, Oi represents the hazard-estimators, such as inundation reach or

334

beach retreat, and Pi represents the parent conditions for such hazard’s results. Basically, Pi is

335

a set of variables containing the storm and scenario morphological characteristics, e.g., Hs, 16

336

duration, direction, surge, period, slope, beach height and beach width. By this method, the

337

hazard outputs at each profile are mapped through the conditional probability tables to their

338

parent inputs in the BN, and the sector results are obtained by weighting each profile

339

according to their contribution length over the sector. Source variables, which are known to

340

present interdependencies can also be interconnected in the BN, e.g., establishing Hs, Tp,

341

direction and duration as parents of storm surge. Figure 5 shows the BN structure used in this

342

work. Arrows denote dependency relations between the variables. The forcing and

343

morphologic variables are parents of the preliminary hazard output (i.e., Ru, and eroded

344

volume). Ru, eroded volume, and morphology are parents of the main hazard outputs (i.e.,

345

inundation reach and beach retreat). Finally, these hazards are combined with beach width to

346

obtain the final hazard index.

347 348 349 350

Figure 5. The BN structure used in the present work. Source variables are highlighted in green, pathway variables in orange, primary hazard variables in grey and final hazard indexes are in red.

17

351

As previously mentioned, the uncertainties on a number of variables are included in

352

the BN by means of Monte-Carlo simulations. Figure 6 shows a graphical example of how

353

this information goes into the BN and propagates to the output. Simulated variables are

354

assumed to be Gaussian. For each profile, storm and scenario, the BN hazard output is a

355

discretized probability distribution accounting for the considered uncertainties (see the cases

356

in Table 2). The total number of cases is a function of the number of profiles, number of

357

storms, number of uncertain variables considered, and number of simulations per uncertain

358

variable. As an example, a sector with 20 profiles and 150 storms would have 100 Monte-

359

Carlo outputs of TWL for each specific profile, storm and MSL condition. Since each future

360

SLR scenario is defined with 20 Monte-Carlo values of MSL, this leads to a total number of

361

6*106 cases per time projection to feed the BN.

362 363 364

Figure 6. Graphical description of inclusion and propagation of Monte-Carlo simulations in the BN. Example on inundation hazard estimation 18

365 366

4. Application of the methodology

367

4.1. Current state hazard profile

368

The first direct result produced by the BN is the unconstrained distribution of all the

369

variables. Data training fills the conditional probability tables with all simulated cases, being

370

each case a set of the source (waves and water level parameters), pathway (beach

371

characteristics) and hazard variables (Figure 5). When no conditioning is applied to the BN, it

372

shows the prior probabilities, i.e. the marginal distributions of each variable without

373

information on the other variables ([eq 5]).

374

The distributions of the calculated hazard indexes (Figure 7-B) represent the

375

probabilities of each hazard level at each sector, both considering the variability of the source

376

(storms) and the variability of the sector morphology. This consideration allows a more

377

representative and detailed sector intercomparison than from the semideterministic

378

approaches (Figure 7-A), which aims for a single hazard value for the sector associated with a

379

given return period (e.g., Jimenez et al. [43]). Note that results are presented here with the

380

same sector definition (~1 km) as in Jiménez et al. [43] for comparison purposes, but the

381

framework allows integration at any scale.

382

The percentages in the hazard profile results can be interpreted in two ways. For

383

example, looking at the erosion distribution at SA as follows: (i) from the perspective of an

384

incoming event of unknown characteristics, there is a 6% probability of hazard index 5

385

occurring at some location within the sector, or (ii) from the perspective of an average

386

incoming event, 6% of the total length of the sector is estimated to be affected by an intensity

387

5 erosion hazard. This dual interpretation is a consequence of the combination in the BN of

388

the storm-climate and the morphological variabilities which allows comparing differences

19

389

between sectors under the same storm-climate with similar representative profiles but with

390

different morphological variability, which in the case of a deterministic treatment of the

391

pathway would produce the same result.

392 393 394 395

Figure 7. Results for a current state hazard-profile. A) semi deterministic results according to Jiménez et al. [43]. B) stochastic results obtained with the BN-based approach.

396

Focusing on the inundation hazard in Tordera Delta, both MN and MS sectors present

397

percentages > 5% (a typical threshold for statistical significance) of index 5, although they

398

show different distributions. In SA, where the topography is higher, the maximum

399

inundations are less frequent but intermediate hazard indexes present higher probabilities,

20

400

which is a consequence of its narrow beach. In MN the response is the opposite due to its

401

morphology, characterized by wider beaches and a lower hinterland. Since the inundation

402

reach is approximated by using the bathtub approach, when water exceeds the berm height a

403

large extension is potentially inundated, and therefore, the hazard index jumps from low (0-1)

404

to extreme values (5). This dual state behaviour is reinforced in MS as a consequence of its

405

even wider and more homogeneous beaches. Identifying the different hazard responses is of

406

key importance if a hazard profile is to be combined with vulnerability data to obtain a risk

407

profile. In Cabrera de Mar, both sectors have similar inundation hazard distributions, while

408

the semideterministic approach (figure 7-A) showed a different index between the sectors

409

because of its representation by single-transect and no inclusion of uncertainties. Notably,

410

using a single transect fixes a single beach height and hinterland shape, and the inundation

411

reach upon the profile chosen for CS for the TR=100 yr was higher than 40 m and lower than

412

60 m. However, an even higher profile but with lower hinterland topography may exist, which

413

combined with the inclusion of model errors in the analysis (higher run-ups) may result in the

414

possibility of inundation reaches higher than 60 m leading to an intensity 5 inundation hazard.

415

A similar effect can be observed comparing CRAF phase 1 results with the BN-based

416

distributions for inundation in MS.

417

Moving to the erosion result at Tordera Delta (Figure 7-B), it can be observed how the

418

hazard intensity decreases from north (SA) to south (MS). In addition, the use of multiple

419

profiles and the inclusion of model errors shows that erosion intensity 5 is actually probable at

420

SA (6%), whereas it was not shown in the semideterministic approach (figure 7-A). In

421

addition, SA and MN show a quite different profile, whereas in Figure 7-A they appear as

422

equivalent sectors. SA shows a generalized vulnerability to erosion, having significant

423

probabilities throughout all intensity levels. Meanwhile, MN has a large probability of no-

424

hazard at 75%, and the remaining probabilities are mainly at medium intensity levels, with 21

425

some residual probability of occurrence of a top-level erosion episode. This outcome is a

426

consequence of SA having narrower and more uniform beaches along the costal stretch while

427

MN has various areas (wide and narrow) that behave differently, which are also detectable

428

with the BN (see results on backward propagation and hazard source below).

429 430

4.2. Identification of hazard sources

431

One of the main advantages of using BN is its capacity to assess correlations between

432

variables by means of the information stored in the conditional probability tables. Thus, the

433

BN can also be used to answer questions about the sector behaviour because conditional

434

probability tables are built preserving the relationships between variables connected in Figure

435

5. By means of backpropagation, i.e. imposing values in output variables (hazards) and

436

assessing the distributions obtained in their parent variables (source and pathway), the BN can

437

assess storm or morphological characteristics associated with induced hazard indexes of

438

interest.

439

To illustrate the potential of BNs in this aspect, the case of the erosion hazard at MN is

440

presented in Figure 8, where the number at the left side of the boxes represent the ranges in

441

which each variable is discretized, and numbers next to black bars represent the discretised

442

probability density at each bin. This case shows a probability of hazard at ≥ 4 of 6%, with a

443

residual of 3% associated to an index 5. The prior distributions represent unconstrained

444

probabilities of some variables of the sector. As an example, the beach width distribution

445

shows a sector mainly with a 40-100 m wide beach, with some extension (~22%) of a

446

relatively narrower coast.

22

447 448 449 450 451 452

Figure 8. Results of hazard source identification at Malgrat North (MN). The prior shows unconstrained dataset probabilities. The posterior shows probabilities conditioned to Level 4 and 5 erosion indexes. Number at the left side of the boxes represent the ranges in which each variable is discretized while numbers next to black bars represent the discretised probability density at each bin

453 454

The BN can be constrained to show how the source and the morphological variables

455

change, e.g. when constraining the maximum erosion hazard level (i.e., 4 and 5), obtaining

456

the conditional probabilities related to that specific case (posterior distributions) and

457

identifying the combination of variables that result in such hazard values. As observed,

458

sectors that have a beach width between 10 and 40 meters (representing only ~22% of the

459

sector’s locations) concentrate the ~86% of the probability of getting a 4 to 5 erosion

460

intensity.

461

Focusing on the hydrodynamic variables, it can be observed how the duration of the

462

event is the main driver of high intensity erosion episodes at the sector (Figure 8, posterior).

463

Higher Hs and Tp also favor a large erosion but the changes between prior and posterior are

464

less intense.

465

Since the erosion hazard is quite sensitive to the beach width in the analysed sector,

466

the BN can be used to statistically assess the expected erosion hazard for existing width

23

467

values along the area. Thus, Figure 9 shows the mean erosion hazard index and its 95%

468

confidence range, together the probability of occurrence of an intensity 5 by constraining

469

beach widths to a range between 15 and 70 m. As it can be observed, a minimum width of 40

470

m is needed in MN to fully avoid an erosion index 4 to 5 (probability of index 5 < 0.05 and

471

confidence interval out of intensities 4 to 5). With a value of BW of 70 m or more, all index

472

values become zero, which is also important information for other uses of the coast, such as

473

the recreational use of the beach. Note that the results presented in Figure 9 are for current

474

conditions and do not include information on the temporal evolution of the study site.

475 476 477 478 479 480

Figure 9. Beach Width (BW) against mean erosion hazard level (blue) and probability of level 5 erosion (green) at Malgrat North (MN). The shaded area represents the 95% confidence interval of the mean erosion index. Obtained probabilities are assessed at the erosion hazard variable after constraining the beach width variable to its different levels.

481 482 483 484

4.3. Hazard assessment at future scenarios 24

485

When the scenario-datasets are introduced in the BN, the hazard profile at different

486

time horizons is obtained. Semideterministic indexes (Figure 7-A) in future scenarios will

487

only show the changes in the sector hazard’s single-level, and once it reaches its maximum

488

(index 5), it will stop showing changes unless new levels are created. Alternatively, the BN-

489

approach allows the evaluation of changes in the probabilities of the different hazard

490

intensities at different time horizons. To illustrate this, two main results are presented: erosion

491

hazard mid-term horizons due to background erosion in the Tordera Delta assuming that past

492

shoreline trends remain in the future (Figure 10), and long-term horizons of the inundation

493

hazard due to SLR at Cabrera de Mar (Figure 11).

494

In the Tordera Delta, SA is the sector showing the current highest index of the erosion

495

hazard, but MN shows a faster progression towards intense erosion events in future scenarios.

496

Notably, in 5 years, SA still shows a general situation with larger frequencies of medium-high

497

erosion episodes, but MN reaches a similar frequency for hazard index 5. At the 10-year

498

horizon, MN is clearly in a worse situation than SA for index 5 frequencies but still better for

499

probabilities of 0 to 4 erosion intensities. At the 20-year horizon, MN is the most vulnerable

500

sector to erosion with almost all events causing at least an index 3 situation, with 86%

501

frequencies of index 5. In contrast, MS remains unaltered across scenarios showing a no-

502

impact profile. (figure 10).

25

503 504 505 506 507

Figure 10. Changes in erosion hazard-profiles at Tordera Delta at different time horizons due to background shoreline retreat. S’Abanell (SA), Malgrat North (MN) and Malgrat South (MS) at current scenario, and +5, +10 and +20-year horizons, assuming past measured shoreline trends remain unchanged in the future.

508 509 510

Figure 11. Changes in inundation hazard-profiles as at different time horizons under RCP 8.5 SLR. Cabrera de Mar North (CN) and South (CS) at current state, 2050 and 2100.

511

26

512

In Cabrera de Mar, both CN and CS show similar responses to SLR at the 20150 and

513

the 2100 RCP 8.5 scenarios. CN shows a higher increase in frequency of medium-intensity

514

events whereas CS keeps on showing slightly larger probabilities of index 5 situations. In this

515

case, this distinction is crucial, since index 3 intensities already reach the infrastructure

516

behind the revetment also causing significant disruptions which makes the northern sector

517

more vulnerable (in general terms) than the southern one (Figure 11).

518

519 520 521

5. Discussion In this work, a BN-based methodology for regional storm induced hazard assessment has been presented and applied at two sectors on the Catalan coast to illustrate its potential.

522

Obtained results highlight the benefits of estimating a hazard probability distribution

523

for each sector. The explicit inclusion of uncertainties (i.e. including model and parameter

524

errors when producing results) and the extensive coverage of the morphological

525

characteristics showed the potential highest-intensity hazards not detected in CRAF phase 1

526

(e.g., erosion at SA or inundation at CS, Figure 7). In other words, the methodology has been

527

successful in hotspot identification, as in Jimenez et al. [43], but allowing detailed sector

528

intercomparison and reducing hazard underprediction due to the omission of uncertainties. In

529

order to illustrate this, Figure 12 shows an additional comparison between results obtained

530

here and those using CRAF phase 1. The BN-method outputs the complete curve of Ru vs the

531

hazard index. This is obtained by constraining the BN at different Ru levels and assessing the

532

distribution of the hazard indicator. On the other hand, the application of CRAF phase 1

533

results in a single hazard index value associated to a given return period (illustrated by a point

534

in the graph). At MN, both results are equivalent (see also Figure 7) and the BN output

535

allocates the Ru associated to TR = 100 yr to a hazard index 5 with a ~95% of probability. At 27

536

CS, the CRAF phase 1 output is a hazard index 4 for TR=100 yr (which has a 0% probability

537

of reaching index 5 for the associated Ru, which corresponds to 5.6m), while the BN-method

538

outputs some probability of having index 5. This difference between both approaches is due

539

to the configuration of the area, which is controlled by the embankment height and the

540

topography of the hinterland. The consequence is the existence of an abrupt increase in the

541

expected mean hazard index and the probability of occurrence of hazard intensity 5 (and

542

associated 95% confidence ranges) for Ru values higher than 5.5 m. The inclusion of model

543

uncertainty in the BN assessment produces values for Ru up to 6.5 meters which will fall into

544

the index 5 category. This explains the difference in hazard intensities obtained using both

545

methods showed in Figure 7. In both cases, it can be observed how the BN-method gives

546

more detailed information on the relation between Ru values and expected hazard indexes

547

(Figure 12) than the CRAF phase 1 single value, showing explicitly the expected variability

548

due to the morphology and model uncertainties.

28

549 550 551 552

Figure 12. Run-up vs mean inundation hazard indicator (left) and probability of Level 5 inundation hazard (right) at Malgrat North (top) and Cabrera South (bottom). The BN results are in blue and Jimenez et al. [43] hazard value for TR = 100 yr is in red.

553 554

Accounting for this variabilities and uncertainties is especially important when

555

performing mid-term and long-term hazard assessments. When assessing hazards at future

556

horizons, changes in the frequencies of low-intermediate hazard intensities can be as

557

important as those of the extreme (large TR) (Figures 10 and 11). Thus, obtained changes in

558

the hazard distributions give complete information about the sector response. This allows a

559

preliminary identification of tipping points for coastal adaptation, as information of

560

intermediate hazard indexes at different time horizons is also needed for such purpose. This 29

561

also allows identification of sectors with different responses over time but with nearly the

562

same current hazard profile (Figure 11). Mid-term and long-term scenarios are based on

563

modelled projections with associated uncertainties that are often estimated with a given error

564

or expected range (e.g., expected SLR for a given horizon and the associated 95% confidence

565

range). This is included in the hazard assessment, and it should be considered for sector

566

intercomparison since it prevents hazard underprediction that may arise from single-value

567

future estimations.

568

The methodology can also produce results of sector morphology and forcing

569

characteristics conditioned to a given hazard intensity (Figures 8 and 9). Thus, the BN,

570

through back-propagation, indicates the potential causes of given hazard index values, giving

571

insights into which processes are important or even to what remediation measures could be

572

effective. Moreover, the BN can produce hazard distributions conditioned to specific storm

573

conditions (i.e., known values of Hs, Tp, duration, direction and surge) in real time. Thus,

574

since these variables are produced by operational hydrodynamic forecasting systems, the tool

575

can perform as a preliminary regional EWS without the need of time-consuming, detailed,

576

morphodynamic, process-based modeling.

577

In this work, we have applied the developed methodology by integrating hazards at ~1

578

km spatial units. This was done mainly for comparison purposes with the results obtained in

579

Jimenez et al. [43]. However, the BN can integrate results at any spatial and scale depending

580

on the physical-process or management questions to be addressed.

581

One limitation of the presented methodology is that a long record of storms is needed

582

to produce robust results. As opposed to CRAF phase 1, no extreme analysis is done and

583

therefore, no extrapolation of storm conditions is performed to consider events of higher

584

magnitude than recorded ones. However, as highlighted in Figure 12 and from the direct

30

585

comparison of obtained results, for the analysed record length (60 years), this seems to have a

586

smaller impact than not accounting for model errors and morphological variability in the

587

sense that CRAF phase 1 results associated to a TR =100 yr underestimate hazard indexes

588

when compared to the BN-CRAF using 60 years of storm events.

589

In the current application, some assumptions have been imposed for the sake of

590

simplicity and test the output. However, they are independent of the method, and they could

591

easily be substituted by other choices. Thus, future scenarios due to background shoreline

592

retreat have been built assuming that measured past trends will keep unchanged in the future.

593

However, they could be substituted by time-varying rates or results from other models.

594

Moreover, process-based models (e.g., 1D-Xbeach, Roelvink et al. [58]; SBeach, Larson and

595

Kraus [59]; LISFLOOD, Bates and de Roo, [60]; Bates et al., [61]) could be used to quantify

596

erosion and inundation hazards instead of simple parametric models. The uncertainty

597

associated with numerical modeling can be included through an ensemble of simulation

598

outputs from different parametrizations or different models, similar to model ensembles used

599

to transfer the uncertainties in future projections of SLR to the inundation maps (Purvis et al.

600

[62]). Following this concept, many different sources of uncertainty (e.g., seasonality of the

601

coast morphology), which were omitted in the present application could be included, with the

602

only consequence being an increase in the size of the case-dataset feeding the BN. The current

603

application was executed on a regular laptop, meaning that there is still room,

604

computationally speaking, to explore new elements to include in the analysis, or increase the

605

number of simulations (e.g. >20 for future projections), increasing the confidence of the

606

results. In this sense, the approach could also be extended with socioeconomic data in order to

607

translate hazard profiles into impact profiles.

608

31

609

7. Conclusions

610

A Bayesian Network-based methodology for storm-induced hazards assessment at

611

regional scale has been developed. It has successfully identified coastal erosion and

612

inundation hotspots in two sectors of the Catalan coast (NW Mediterranean) similarly to what

613

was previously done with the CRAF-phase 1 method while giving further information.

614

The method accounts for the contribution of different sources of variability (i.e.,

615

variability in storms and coastal morphology), as well as uncertainties associated with the

616

hazard models (i.e., model errors, parameter uncertainties and confidence ranges on future

617

scenarios). As a result, instead of resulting in a single value, the Bayesian Network-based

618

method provides hazard distributions which will indicate the probability of occurrence of

619

hazards at any intensity.

620

The method can be easily used to forecast changes in storm-induced hazard levels by

621

applying it under given future scenarios, based on changes in morphology and/or storm

622

properties. In such a case, the method will predict the expected change in the probability of

623

occurrence for different ranges of hazard indexes.

624

The inherent capability of Bayesian Network’s to assess the interdependence between

625

variables is used to identify source (storms) and pathway (beach morphology) characteristics

626

responsible for given hazard distributions. When this is applied under different scenarios

627

permit to identify harmful combinations taking place at specific locations and at given time

628

horizons. From the management standpoint this will provide essential information for making

629

decisions on selecting coastal adaptation alternatives.

630

The application of the method at two known hotspots of the Catalan coast highlighted

631

concerning scenarios of erosion (Tordera Delta) and inundation (Cabrera de Mar) in 10 to 20

32

632

years from baseline, while allowing a better comparison and characterization of the ~1km

633

sectors within hotspots.

634

635

Acknowledgements

636 637

This study was conducted in the framework of the RISC-KIT (Grant No 603458) and

638

the M-CostAdapt (CTM2017-83655-C2-1-R) research projects, funded by the EU and the

639

Spanish Ministry of Economy and Competitiveness (MINECO/AEI/FEDER, UE)

640

respectively. The first author was supported by a PhD grant from the Spanish Ministry of

641

Education, Culture and Sport. The authors express their gratitude to IH-Cantabria for

642

supplying wave and water level data, to the Spanish Ministry for Ecological Transition for the

643

bathymetric data, and to the Institut Cartogràfic i Geològic de Catalunya by for the LIDAR

644

data used in this study.

645 646

References

647

[1] IPCC, 2012: Climate Change 2012. Managing the risks of extreme events and disasters to

648

advance climate change adaptation. A Special Report of Working Groups I and II of the

649

Intergovernmental Panel on Climate Change, edited by C. B. Field, V. Barros, T. F.

650

Stocker, D. Qin, D. J. Dokken, K. L. Ebi, M. D. Mastrandrea, K. J. Mach, G.-K.

651

Plattner, S. K. Allen, M. Tignor, and P. M. Midgley, Cambridge University Press,

652

Cambridge, UK and New York, NY, USA., 2012.

653

[2] IPCC, 2013: Climate Change 2013. The Physical Science Basis. Contribution of Working

654

Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate

655

Change, edited by T. F. Stocker, D. Qin, G.-K. Plattner, M. Tignor, S. K. Allen, J.

33

656

Boschung, A. Nauels, Y. Xia, V. Bex, and P. M. Midgley, Cambridge University Press,

657

Cambridge, United Kingdom and New York, NY, USA., 2013.

658

[3] Lionello, P., Boldrin, U. and Giorgi, F., 2008. Future changes in cyclone climatology over

659

Europe as inferred from a regional climate simulation, Clim. Dyn., 30(6), 657–671,

660

doi:10.1007/s00382-007-0315-0.

661

[4] Conte, D. and Lionello, P., 2013. Characteristics of large positive and negative surges in

662

the Mediterranean Sea and their attenuation in future climate scenarios, Glob. Planet.

663

Change, 111, 159–173, doi:10.1016/j.gloplacha.2013.09.006.

664

[5] IPCC, 2014: Climate Change 2014. Synthesis Report. Contribution of Working Groups I,

665

II and III to the Fifth Assessment Report of the Intergovernmental Panel on Climate

666

Change [Core Writing Team, R.K. Pachauri and L.A. Meyer (eds.)]. IPCC, Geneva,

667

Switzerland, 151 pp.

668

[6] Casas-Prat, M., Sierra, J.P., 2013. Projected future wave climate in the NW Mediterranean

669

Sea. J. Geophys. Res. Ocean. 118, 3548–3568. https://doi.org/10.1002/jgrc.20233

670

[7] UNEP/MAP/PAP. (2008) Protocol on Integrated Coastal Zone Management in the

671

Mediterranean. Priority Actions Programme, Split.

672

[8] Ciavola, P., Ferreira, O., Haerens, P., Van Koningsveld, M., Armaroli, C. and Lequeux,

673

Q., 2011. Storm impacts along European coastlines. Part 1: The joint effort of the

674

MICORE

675

doi:10.1016/j.envsci.2011.05.011.

and

ConHaz

Projects,

Environ.

Sci.

Policy,

14(7),

912–923,

676

[9] Ciavola, P., Ferreira, O., Haerens, P., Van Koningsveld, M. and Armaroli, C., 2011. Storm

677

impacts along European coastlines. Part 2: Lessons learned from the MICORE project,

678

Environ. Sci. Policy, 14(7), 924–933, doi:10.1016/j.envsci.2011.05.009.

679

[10] Vojinovic, Z., Abebe, Y., Sanchez, A., Medina Pena, N., Nikolic, I., Monojlovic, N.,

680

Makropoulos, C., Pelling, M. and Abbott, M., 2014. Holistic flood risk assessment in

34

681

coastal areas - the PEARL approach, in 11th International Conference on

682

Hydroinformatics, pp. 1–8, CUNY Academic Works, New York City, USA..

683

[11] Oumeraci, H., Kortenhaus, A., Burzel, A., Naulin, M., Dassanayake, D. R., Jensen, J.,

684

Wahl, T., Mudersbach, C., Gönnert, G., Gerkensmeier, B., Fröhle, P. and Ujeyl, G.,

685

2015. XtremRisK — Integrated Flood Risk Analysis for Extreme Storm Surges at Open

686

Coasts and in Estuaries: Methodology, Key Results and Lessons Learned, Coast. Eng.

687

J., 57(1), 1540001, doi:10.1142/S057856341540001X.

688

[12] Van Dongeren, A., Ciavola, P., Martinez, G. and Viavattene, C., 2018. Introduction to

689

RISC-KIT.

Resilience-increasing

strategies

690

doi.org/10.1016/j.coastaleng.2017.10.007.

for

coasts,

Coast.

Eng.,

691

[13] Vousdoukas, M.I., Bouziotas, D., Giardino, A., Bouwer, L.M., Mentaschi, L.,

692

Voukouvalas, E., Feyen, L., 2018. Understanding epistemic uncertainty in large-scale

693

coastal flood risk assessment for present and future climates. Nat. Hazards Earth Syst.

694

Sci. 18, 2127–2142. https://doi.org/10.5194/nhess-18-2127-2018

695

[14] Sayers, P.B., Hall, J.W., Meadowcroft, I.C., 2002. Towards risk-based flood hazard

696

management in the UK. Proc. Inst. Civ. Eng. - Civ. Eng. 150, 36–42.

697

https://doi.org/10.1680/cien.2002.150.5.36

698

[15] Narayan, S., Nicholls, R. J., Clarke, D., Hanson, S., Reeve, D., Horrillo-Caraballo, J., le

699

Cozannet, G., Hissel, F., Kowalska, B., Parda, R., Willems, P., Ohle, N., Zanuttigh, B.,

700

Losada, I., Ge, J., Trifonova, E., Penning-Rowsell, E. and Vanderlinden, J. P., 2014.

701

The SPR systems model as a conceptual foundation for rapid integrated risk appraisals:

702

Lessons from Europe, Coast. Eng., 87, 15–31, doi:10.1016/j.coastaleng.2013.10.021.

703

[16] Garrity, N.J., Battalio, R., Hawkes, P.J., Roupe, D., 2006. Evaluation of the event and

704

response approaches to estimate the 100-year coastal flood for Pacific coast sheltered

705

waters. Proc. 30th Int. Conf. on Coastal Engineering, ASCE, 1651-1663.

35

706

[17] Callaghan, D.P., Nielsen, P., Short, A., Ranasinghe, R., 2008. Statistical simulation of

707

wave

climate

and

extreme

beach

erosion.

708

https://doi.org/10.1016/j.coastaleng.2007.12.003

Coast.

Eng.

55,

375–390.

709

[18] Sanuy, M., Jiménez, J.A., Ortego, M.I. and Toimil, A. (in review). Differences assigning

710

probabilities to coastal inundation hazard estimators: event vs. response approaches.

711

Submitted to Journal of Flood Risk Management.

712

[19] Hawkes, P.J., Gouldby, B.P., Tawn, J.A., Owen, M.W., 2002. The joint probability of

713

waves and water levels in coastal engineering design. J. Hydraul. Res. 40, 241–251.

714

https://doi.org/10.1080/00221680209499940

715

[20] Masina, M., Lamberti, A., Archetti, R., 2015. Coastal flooding: A copula based approach

716

for estimating the joint probability of water levels and waves. Coast. Eng. 97, 37–52.

717

https://doi.org/10.1016/j.coastaleng.2014.12.010

718

[21] Lin-Ye, J., Garcia-Leon, M., Gracia, V., Sanchez-Arcilla, A., 2016. A multivariate

719

statistical model of extreme events: An application to the Catalan coast. Coast. Eng.

720

117, 138–156. https://doi.org/10.1016/j.coastaleng.2016.08.002

721

[22] Stockdon, H.F., Holman, R.A., Howd, P.A., Sallenger, A.H., 2006. Empirical

722

parameterization

of

setup,

swash,

and

runup.

723

https://doi.org/10.1016/j.coastaleng.2005.12.005

Coast.

Eng.

53,

573–588.

724

[23] Mendoza, E.T. and Jiménez, J.A., 2006. A Storm Classification Based on the Beach

725

Erosion Potential in the Catalonian Coast. Coast. Dyn. 2005 40855, 1–11.

726

https://doi.org/10.1061/40855(214)98

727

[24] Fienen, M.N., Masterson, J.P., Plant, N.G., Gutierrez, B.T., Thieler, E.R., 2013. Bridging

728

groundwater models and decision support with a Bayesian network. Water Resour. Res.

729

49, 6459–6473. https://doi.org/10.1002/wrcr.20496

36

730

[25] Palmsten, M.L., Todd Holland, K., Plant, N.G., 2013. Velocity estimation using a

731

Bayesian network in a critical-habitat reach of the Kootenai River, Idaho. Water Resour.

732

Res. 49, 5865–5879. https://doi.org/10.1002/wrcr.20361

733

[26] Gieder, K.D., Karpanty, S.M., Fraser, J.D., Catlin, D.H., Gutierrez, B.T., Plant, N.G.,

734

Turecek, A.M., Robert Thieler, E., 2014. A Bayesian network approach to predicting

735

nest presence of the federally-threatened piping plover (Charadrius melodus) using

736

barrier

737

https://doi.org/10.1016/j.ecolmodel.2014.01.005

island

features.

Ecol.

Modell.

276,

38–50.

738

[27] Gutierrez, B. T., Plant, N. G. and Thieler, E. R., 2011. A Bayesian network to predict

739

coastal vulnerability to sea level rise, J. Geophys. Res. Earth Surf., 116(2), 1–15,

740

doi:10.1029/2010JF001891.

741

[28] Plant, N. G., Robert Thieler, E., & Passeri, D. L.., 2016. Coupling centennial‐scale

742

shoreline change to sea‐level rise and coastal morphology in the Gulf of Mexico using

743

a Bayesian network. Earth's Future, 4(5), 143-158.

744

[29] Van Verseveld, H. C. W., Van Dongeren, A. R., Plant, N. G., Jäger, W. S. and den

745

Heijer, C., 2015. Modelling multi-hazard hurricane damages on an urbanized coast with

746

a

747

doi:10.1016/j.coastaleng.2015.05.006.

Bayesian

Network

approach,

Coast.

Eng.,

103,

1–14,

748

[30] Poelhekke, L., Jäger, W. S., van Dongeren, A., Plomaritis, T. A., McCall, R. and

749

Ferreira, Ó., 2016. Predicting coastal hazards for sandy coasts with a Bayesian Network,

750

Coast. Eng., 118, 21–34, doi:10.1016/j.coastaleng.2016.08.011.

751

[31] Plomaritis, T.A., Costas, S., Ferreira, Ó., 2018. Use of a Bayesian Network for coastal

752

hazards, impact and disaster risk reduction assessment at a coastal barrier (Ria Formosa,

753

Portugal). Coast. Eng. 134, 134–147. https://doi.org/10.1016/j.coastaleng.2017.07.003

37

754

[32] Sanuy, M., Duo, E., Jäger, W. S., Ciavola, P., and Jiménez, J. A., 2018. Linking source

755

with consequences of coastal storm impacts for climate change and risk reduction

756

scenarios for Mediterranean sandy beaches, Nat. Hazards Earth Syst. Sci., 18, 1825-

757

1847, https://doi.org/10.5194/nhess-18-1825-2018.

758

[33] Uusitalo, L., 2007. Advantages and challenges of Bayesian networks in environmental

759

modelling.

760

https://doi.org/10.1016/j.ecolmodel.2006.11.033

761 762

Ecol.

Modell.

203,

312–318.

[34] Straub, D., 2005. Natural hazards risk assessment using Bayesian networks. 9th Int. Conf. Struct. Saf. Reliab. ICOSSAR 05 Rome Italy 2509–2516.

763

[35] Jäger, W. S., Christie, E. K., Hanea, A. M., den Heijer, C. and Spencer, T., 2018. A

764

Bayesian network approach for coastal risk analysis and decision making, Coast. Eng.,

765

134, 48-61, doi:10.1016/j.coastaleng.2017.05.004.

766

[36] Viavattene, C., Jiménez, J.A., Ferreira, O., Priest, S., Owen, D., McCall, R., 2018.

767

Selecting coastal hotspots to storm impacts at the regional scale: a Coastal Risk

768

Assessment

769

https://doi.org/10.1016/j.coastaleng.2017.09.002

Framework.

Coast.

Eng.

134,

33–47.

770

[37] Ferreira, O., Viavattene, C., Jiménez, J. A., Bolle, A., Das Neves, L., Plomaritis, T. A.,

771

McCall, R., Van Dongeren, A. R. 2018. Storm-induced risk assessment: Evaluation of

772

two tools at the regional and hotspot scale. Coastal Engineering, 134, 241-253, doi:

773

10.1016/j.coastaleng.2017.10.005.

774

[38] Stockdon, H. F., Sallenger, A. H., Holman, R. A. and Howd, P. A., 2007. A simple

775

model for the spatially-variable coastal response to hurricanes, Mar. Geol., 238(1–4), 1–

776

20, doi:10.1016/J.MARGEO.2006.11.004.

777

[39] Villatoro, M., Silva, R., Méndez, F.J., Zanuttigh, B., Pan, S., Trifonova, E., Losada, I.J.,

778

Izaguirre, C., Simmonds, D., Reeve, D.E., Mendoza, E., Martinelli, L., Formentin, S.M.,

38

779

Galiatsatou, P., Eftimova, P., 2014. An approach to assess flooding and erosion risk for

780

open

781

https://doi.org/10.1016/j.coastaleng.2013.11.009

782

beaches

in

a

changing

climate.

Coast.

Eng.

87,

50–76.

[40] Armaroli, C. and Duo, E., 2018. Validation of the Coastal Storm Risk Assessment

783

Framework

along

the

Emilia-Romagna

784

doi:10.1016/j.coastaleng.2017.08.014.

coast,

Coast.

Eng.,

134,

159-167,

785

[41] Bosom, E., Jiménez, J.A., 2011. Probabilistic coastal vulnerability assessment to storms

786

at regional scale – application to Catalan beaches (NW Mediterranean). Nat. Hazards

787

Earth Syst. Sci 11, 475–484. https://doi.org/10.5194/nhess-11-475-2011.

788

[42] Ballesteros, C., Jiménez, J.A., Viavattene, C., 2018. A multi-component flood risk

789

assessment in the Maresme coast (NW Mediterranean). Nat. Hazards 90, 265–292.

790

https://doi.org/10.1007/s11069-017-3042-9.

791

[43] Jiménez, J.A., Sanuy, M., Ballesteros, C., Valdemoro, H.I., 2018. The Tordera Delta, a

792

hotspot to storm impacts in the coast northwards of Barcelona (NW Mediterranean).

793

Coast. Eng. 134, 148–158. https://doi.org/10.1016/j.coastaleng.2017.08.012.

794

[44] Camus, P., Mendez, F.J., Medina, R., Tomas, A., Izaguirre, C., 2013. High resolution

795

downscaled ocean waves (DOW) reanalysis in coastal areas. Coast. Eng. 72, 56–68.

796

https://doi.org/10.1016/j.coastaleng.2012.09.002

797

[45] Reguero, B. G., Menéndez, M., Méndez, F. J., Mínguez, R., Losada, I. J., 2012. A Global

798

Ocean wave (GOW) calibrated reanalysis from 1948 onwards, Coast Eng 65:38-55.

799

https://doi.org/10.1016/j.coastaleng.2012.03.003

800

[46] Cid, A., Castanedo, S., Abascal, A. J., Menéndez, M., Medina, R., 2014. A high

801

resolution hindcast of the meteorological sea level component for Southern Europe: the

802

GOS dataset, Clim Dyn, 43:2167-2184.

39

803

[47] Ruiz, A., Kornus, W., Talaya, J. 2009. Coastal applications of Lidar in Catalonia. 6th

804

European Congress on Regional Geoscientific Cartography and Information Systems,

805

Munich.

806

[48] Jiménez, J.A., Valdemoro, H.I., 2019. Shoreline evolution and its management

807

implications in beaches along the Catalan coast. In: Morales, J.A. (ed), The Spanish

808

Coastal Systems Dynamic Processes, Sediments and Management, Springer, 745-764.

809

[49] Ferreira, Ó., Plomaritis, T.A., Costas, S., 2017. Process-based indicators to assess storm

810

induced

811

https://doi.org/10.1016/J.EARSCIREV.2017.07.010

812

coastal

hazards.

Earth-Science

Rev.

173,

159–167.

[50] Hastings, W. K., 1970. Monte Carlo sampling methods using Markov chains and their

813

applications,

Biometrika,

Volume

814

https://doi.org/10.1093/biomet/57.1.97

57,

Issue

1,

Pages

97–109,

815

[51] EurOtop., 2007. Wave Overtopping of Sea Defences and Related Structures–Assessment

816

Manual. UK: NWH Allsop, T. Pullen, T. Bruce. NL: JW van der Meer. DE: H.

817

Schüttrumpf, A. Kortenhaus. www. overtopping-manual. com.

818

[52] Silva, 2019. Coastal Hazard Assessments for Sandy Coasts: Appraisal of Five

819

Methodologies. J. Coast. Res. https://doi.org/10.2112/JCOASTRES-D-18-00083.1

820

[53] Somot, S., Jordà, G., Harzallah, A., Darmaraki, S., 2019. Sub-chapter 1.2.3. The

821

Mediterranean Sea in the future climate projections In : The Mediterranean region under

822

climate change : A scientific update [en ligne]. Marseille : IRD Éditions, 2016 (généré

823

le

824

. ISBN : 9782709922203. DOI :

825

10.4000/books.irdeditions.23100.

27

mars

2019).

Disponible

40

sur

Internet

:

826

[54] Bender, M. A., T. R. Knutson, R. E. Tuleya, J. J. Sirutis, G. A. Vecchi, S. T. Garner, and

827

I. M. Held, 2010. Modeled Impact of Anthropogenic Warming on the Frequency of

828

Intense Atlantic hurricanes, Science, 327, 454.

829

[55] Plant, N. G., J. Flocks, H. F. Stockdon, J. W. Long, K. K. Guy, D. M. Thompson, J. M.

830

Cormier, C. G. Smith, J. L. Miselis, and P. S. Dalyander, 2014. Predictions of barrier

831

island berm evolution in a time-varying storm climatology, Journal of Geophysical

832

Research F: Earth Surface, 119(2), 300-316.

833 834

[56] Pearl, J., 1988. Probabilistic reasoning in intelligent systems: networks of plausible inference, Morgan Kaufmann, San Francisco, California, USA.

835

[57] Jensen, F. V., 1996. An introduction to Bayesian networks, UCL Press, London, UK.

836

[58] Roelvink, D., Reniers, A., van Dongeren, A., van Thiel de Vries, J., McCall, R. and

837

Lescinski, J., 2009. Modelling storm impacts on beaches, dunes and barrier islands,

838

Coast. Eng., 56(11–12), 1133–1152, doi:10.1016/j.coastaleng.2009.08.006.

839

[59] Larson, M., Kraus, N.C., 1989. SBEACH. Numerical model for simulating storm-

840

induced beach change; Report 1: Empirical foundation and model development. Tech.

841

Rep. CERC (US Army Eng. Waterw. Exp. Stn. Coast. Eng. Res. Center).

842

https://doi.org/Technical Report CERC-89-9

843 844

[60] Bates, P.D., De Roo, A.P.J., 2000. A simple raster-based model for floodplain inundation. Journal of Hydrology 236, 54-77.

845

[61] Bates, P.D., Dawson, R.J., Hall, J.W., Horritt, M.S., Nicholls, R.J., Wicks, J., Hassan,

846

M.A.A.M., 2005. Simplified two-dimensional numerical modeling of coastal flooding

847

and example applications. Coast. Eng., 52, 793-810.

848

[62] Purvis, M.J., Bates, P.D., Hayes, C.M., 2008. A probabilistic methodology to estimate

849

future coastal flood risk due to sea level rise. Coast. Eng. 55, 1062–1073.

850

https://doi.org/10.1016/j.coastaleng.2008.04.008

41

851

Tables

852

Table 1. The hazard indicator levels as a function of the beach width (BW) against inundation

853

reach or beach retreat. All units are in meters, and ∆X10 stands for average profile retreat

854

associated with TR=10 years for the site, which has a value of 7.5 m. Hazard indicator level 0 1 2 3 4 5

Inundation Reach < 0.5*BW 0.5* BW BW + 60

Erosion BW – Retreat >4 ∆X10 4 ∆X10< BW – Retreat <3 ∆X10 3 ∆X10< BW – Retreat <2 ∆X10 2 ∆X10< BW – Retreat < ∆X10 ∆X10< BW – Retreat <0 BW – Retreat ≤ 0

855 856

Table 2. Ranges of the uncertain variables included in the assessment. Note that total number

857

of cases is always 100 for hazard variables and 20 for future scenario variables The total

858

number of cases can be the result of the simulation of a single variable or of the combination

859

of simulations from 2 variables. Tordera Delta- S’Abanell (SA), Malgrat North (MN) and South (MS) Monte-Carlo Total number Variable Mean and std. dev. simulations of Cases Ru on [eq 1] [eq 1] ± 0.32 (m) 100 100 (Ru) ∆V on [eq 3] [eq 3] ± 10 m3 10 100 (∆X) d* on [eq 4] 0.8 ± 0.15 m 10 SB: 1.1 ± 0.8 m/yr Retreat trends MN: 4.0 ± 0.5 m/yr 20 20 (BW) MS: 1.9 ± 2.1 m/yr Cabrera de Mar North (CN) and South (CS) Monte-Carlo Total number Variable Mean and std. dev. simulations of Cases Coef. (1.65) on [eq 2] 1.65 ± 0.10 10 100 (Ru) γf on [eq 2] 0.65 ± 0.10 10 SLR

2050: 0.30 ± 0.04 (m) 2100: 0.63 ± 0.10 (m)

860

861 862

42

20

20 (MSL)

863

Figure Captions

864

Figure 1. Conceptual scheme of the CRAF-phase 1 approach (A), compared to the BN-based

865

methodology (B).

866

Figure 2. Locations of the wave and surge data nodes used in this work. Sector limits and

867

used profiles are indicated. The red dot highlights the location of the numerical node for wave

868

and surge data.

869

Figure 3. Dataset characteristics. Storm bulk parameters (left) Hs vs Tp, colored by storm

870

duration, and markers by storm duration. Main morphological features (right), beach height vs

871

beach width, colored by beach slope and markers by sector.

872

Figure 4. Profile erosion schematization used to derive beach retreat from eroded volume

873

(adapted from Mendoza and Jiménez [36]).

874

Figure 5. The BN structure used in the present work. Source variables are highlighted in

875

green, pathway variables in orange, primary hazard variables in grey and final hazard indexes

876

are in red.

877

Figure 6. Graphical description of inclusion and propagation of Monte-Carlo simulations in

878

the BN. Example on inundation hazard estimation

879

Figure 7. Results for a current state hazard-profile. A) semi deterministic results according to

880

Jiménez et al. [43]. B) stochastic results obtained with the BN-based approach.

881

Figure 8. Results of hazard source identification at Malgrat North (MN). The prior shows

882

unconstrained dataset probabilities. The posterior shows probabilities conditioned to Level 4

883

and 5 erosion indexes. Number at the left side of the boxes represent the ranges in which each

884

variable is discretized while numbers next to black bars represent the discretised probability

885

density at each bin. 43

886

Figure 9. Beach Width (BW) against mean erosion hazard level (blue) and probability of

887

Level 5 erosion (green) at Malgrat North (MN). The shaded area represents the 95%

888

confidence interval of the mean erosion level. Obtained probabilities are assessed at the

889

erosion hazard variable after constraining the beach width variable to its different levels.

890

Figure 10. Changes in erosion hazard-profiles at Tordera Delta at different time horizons due

891

to background shoreline retreat. S’Abanell (SA), Malgrat North (MN) and Malgrat South

892

(MS) at current scenario, and +5, +10 and +20-year horizons, assuming past measured

893

shoreline trends remain unchanged in the future.

894

Figure 11. Changes in inundation hazard-profiles as at different time horizons under RCP 8.5

895

SLR. Cabrera de Mar North (CN) and South (CS) at current state, 2050 and 2100.

896

Figure 12. Run-up vs mean inundation hazard indicator (left) and probability of Level 5

897

inundation hazard (right) at Malgrat North (top) and Cabrera South (bottom). The BN results

898

are in blue and Jimenez et al. [43] hazard value for TR = 100 yr is in red.

44

Highlights Hazard assessments involve a stochastic source (storms), a variable pathway (coast morphology) and model uncertainties (hazard models and future projections). Bayesian Network framework for regional coastal hazard assessment (BN-CRAF) is developed to cope with these variabilities and uncertainties. The approach has been successful in hotspot identification and the results have been compared to the CRAF-phase 1 robust assessment module. The tool outputs current and future hazard distributions, allows a deep understanding of sector characteristics and provides useful information for the selection of coastal adaptation alternatives.

Sanuy, M.: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Resources; Validation; Visualization; Writing – original draft; Writing – review & editing Jimenez, J.A.: Conceptualization , Funding acquisition, Supervision, Writing – review & editing Plant, N.: Conceptualization, Software, Supervision, Writing – review & editing

Declaration of interests ☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. ☐The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: