Extreme storm surge distributions and morphological

storm impacts, application at Marseilles

Pierre Gaufrès1 and François Sabatier2
1
Centre d'Etudes Techniques Maritimes Et Fluviales, CETMEF, 2 Bd du Président Kennedy, B.P. 543,
13092 Aix-en-Provence cedex 02, France
2
Centre Européen de Recherche et d’Enseignement en Géosciences de l’Environnement, CEREGE, Europôle de
L’Arbois, BP 80, 13545 Aix-en-Provence cedex 04, France

Abstract
Extreme sea levels threaten sand-beaches and low-lying densely populated coastal areas due to ocean atmosphere
exchanges, at all space and time scales. In micro-tidal conditions, the longest hourly tide-gauge record available on
the French Mediterranean coast (Marseilles Endoume 1885-2003) has enabled statistical analysis usually computed
on continental hydrological parameters. Frequential analysis tools, applied to time series near to the million data, are
implemented in order to provide a calculation method for stochastic meteorological storm surge component to
characterize coastal inundation and erosion risk. Surge probability distributions tested with associated confidence
interval (GEV, Jenkinson, Gamma, …) and extreme values of the data fitting (Maximum Annual and Peaks-Over-
Threshold) are discussed with relative tests (stationarity, independence, homogeneity,...). Results are compared with
sea-levels of morphological significance correlated to coastal impact observations during storm events. Statistic
methods and tools put forward contribute to a better knowledge of intensity, frequency and duration of extreme sea
levels associated with morphogenic storms. They also provide guidelines for the coastal risk mitigation.

1 Introduction Rhône delta (Bruzzi, 1998), the control of the

Shoreline moves permanently under the combined
action of sea, wind and inland waters. The result is
alongshore alternation of projection or retreat. The
retreat trend on the Camargue coast observed during
the twentieth century forced the state administrators,
coastal managers and scientific community to study
the morphological evolution of the delta and meteo-
marine agents forcing these processes (Sabatier et
al.,2003) carefully.
At the national level, the frequential analysis of
the extreme surge was carried out by the SHOM
(Service Hydrographique et Océanographique de la
Marine) on the Atlantic and English Channel coasts
(Simon, 1994), mainly for navigation needs. More
recently, following the storms of December 1999 and
the flooding of the nuclear power station of Blayais
(Gironde), EDF (Electricité de France) was led to
recalculate the altimeter setting of the platforms of the
nuclear power stations. Recent variations of the
measurement of sea level in Marseilles (Wöppelmann,
1997) and the statistical inference of extreme surges
together with the behaviour of upper tail of surge
distributions (Walton, 2000).
2 Physical background
The area of study, Figure 1, extends on the coastline
influenced by the Rhône (Camargue) to the Cap
Croisette at the end of the Gulf of Marseilles (45 km
from the Grand Rhône mouth). The tide range is 0.3 m
and sea-level rise is estimated to 2 mm/yr. On the west
side, the coast of the Rhône delta is 90 km long; 85 %
of which is coastal protection equipped.
43°30
Rhône delta
Marseilles
43°20

weather factors on the Atlantic coast of France have

been studied (Pirazzoli, 2000). Up to now, no
exhaustive study of the coastal maregraphic data has 43°10

been carried out for the French Mediterranean sea. Mediterranean Sea
This study is part of the IMPLIT program
0 10 20 km

43°

(IMPact des évènements extrêmes - tempters et 5° 40 5° 5° 20

surcôtes - liés au Changement Climatique sur les

hydro-systèmes du LITtoral méditerranéen français).
IMPLIT is supported by the GICC group (Gestion et
Impacts du Changement Climatique) appointed by the
MEDD (Ministère français de l'Ecologie et du
Développement Durable). This work fits with the
studies concerning the impact of the storms on the
Figure 1. Area of the study and tide gauges location
around Marseilles
Shoreline moves can reach 10 m/yr locally. The
Camargue island is subject to both flood hazards and
marine inundations. Alluvial valley can be influenced
by sea levels as far as 50 km inland. The Eastern part

1
is composed successively of the Gulf of Fos occupied
by an industrial harbor complex and the Gulf of
Marseilles - Prado with strong stake in term of
protection of people and goods.
3 Tide-gauge records and data
pretreatment
A significant number of tide-gauges were listed after
investigation into the area coastal managers.
Eventually, the data gauge of Marseilles in the Calvo
Cove (Endoume), in operation since 1885 (Figure 2)
is chosen for a statistic exploration taking into account
the length of the series, unique on the French
Mediterranean coast. The sea level observatory
produced two series of measurements: recordings on
roller paper which provide the temporal variation of
the sea level and monthly averages of the sea level
obtained with an exceptional mechanical adding
machine. The historical equipment worked until 1988
and was replaced in 1998 by an acoustic tide gauge in
the same stilling well. Marseilles hourly series (1885-
2003) and predicted tide were provided by SHOM.
Figure 2. Marseilles historical tide-gauge (Endoume)
The control of the series of values highlighted a
variable temporal shift. One notes that the
meteorological residue (predicted tide deduced from
sea-level data) preserves an undeniable periodic
character which could be explained by signal filtering
function of the geometry of the observatory (feeder
canal to float chamber and water gate) or obstruction
of this during years of poor maintenance. As an
explanation has not yet been found, the repercussion of
this temporal shift in term of uncertainty over extreme
surge was quantified before the statistical analysis. On
the one hand, it proves that the maximum error is
lower than the centennial result variations of studied
laws adjustment, and on the other hand, that the
variations - positive or negative - tend to be balanced
considering the data set on the whole (characteristics
summarized in table 1).
Beginning 12 February 1885
End 31 December 2003
Absolute length (years) 119
Availability (%) 75,6 %
Missing data (%) 24,4 %
Equivalent full years 89,96
Hourly records 911 650
Table 1. Marseilles Endoume tide gauge data
characteristics
4 Statistical models and distribution fitting
control
Modeling theory of the hydrological extremes is
particularly developed in continental field, taking into
account the stakes in term of regional planning and
density of measurements. Tools of statistical inference,
elaborated during decades, adjust extreme statistical
laws to the data considered as random variables
(Gumbel, 1958), (Jenkinson, 1969),… . It is classically
based on two methods: Maximum annual (MA)
analysis constitutes a traditional approach, easy to
implement. On the other hand, MA analysis selects
only one value per year: rich information of intense
years is not taken into account. POT (Peaks-Over-
Threshold) method usually leads to better adjustments
because it takes into account the whole information
available and corresponds to a measurable physical
reality on ground. However the choice of the threshold
influences the quality of results Mousavi (1997).
The bibliographical analysis provides few
examples of thorough statistical assessments of storm
surge distribution. A more complete description is
given in meso-tidal conditions (Sandy Hook, New
Jersey, US) by Walton (2000). Having a long hourly
series (1935-1996), he notices that if many probability
distributions can be used for the current frequencies,
none can be dissociated as particularly fitting over the
entire experimental data. One notes significant
discrepancy in the surge magnitudes for large return
periods used for hazard prevention (T > 50 yr).
The parameter estimation of probability laws is
carried out according to two methods: maximum
likelihood (ML) and method of moments. The
goodness-of-fit, measured initially according to
Pearson's traditional test of Khi-square appears not
very relevant. Besides, the Bootstrap method is used
for robustness analysis.
2
4.1 Maximum annual analysis
MA adjustments are computed with SAFARHY
software developed by IRD (Institut de Recherche et
Développement). From the available data set (1885-
2003), 15 years of completely missing data were
excluded. No additional criterion of exclusion of
uncomplete year was selected. As a matter of fact,
application of 80 % availability for a year and/or main
surge season (November to February) lead to draw
aside four of ten greater observations.
laws: Jenkinson and Gamma guarantee, a priori, a
better fitting for the interval T = [50 yr, 100 yr].
Beyond, the confidence interval widens significantly,
they are not more relevant and the GEV 2 type laws
(Frechet, Pareto, Cauchy, …) are recommended
(T>100 yr).

Figure 4. Jenkinson's distribution of Maximum annual

(3 parameters)

Bootstrapping the data, Figure 5, illustrate the

variability of 2p and 3p distributions with MA method,
Figure 3. Gumbel EV1 distribution of Maximum
and enables one to draw the following conclusions:
annual - Moments estimation method
Two parameters distributions keep all their entire
relative position. Gumbel's method (ML – 2p), which
For ML estimation parameters method, Gumbel
the use is strongly widespread for the analysis of the
adjustment is relevant for T<10 yr. Beyond, one notes
extreme values revealed not very efficient
a clear divergence: the twenty strongest surges are
(systematically overestimating for T > 50 yr), Weibull
located below the adjustment. The two strongest
(ML - 2p) undervaluating as well. Jenkinson (3p)
events leave the confidence interval at 95%. For
which appeared to fit the best MA data for T < 100 yr,
moments estimation parameters method, Figure 3, only
also reveals the strongest sensitivity to the resampling
two strongest surges move away from the adjustment;
(multiple intersections). All laws provide the
maximum plot leaves the confidence interval at 95 %.
centennial quantile with a variation close to 5 %
Jenkinson's distribution shown in Figure 4 proves to be
(either 5 cm) except Jenkinson (3p) with 7,3 % of
equivalent to Gumbel's in comparison with the Khi-
average deviation to the original data on 20 series
square test. However, graphically the goodness-of-fit
resampled.
is much better for the tail of the sample, which is
explained by an additional parameter. Jenkinson's third Bootstrap method - Marseille (1885-2003)
Surge ajustment evolution for T = 100 ans
parameter (K = 0.153) is significantly different from 135

zero which confirms that the Gumbel's law (K = 0) is 130

Centennial surge quantile (cm)

125

not adapted to this series. Lastly K is positive, the MA 120

series has an hyperexponential behavior of GEV2 type 115

(K > 0), also typical of the rains and flows in 110

105

continental hydrology. 100

The synthesis of adjustments confirms Walton 95

90

(2000)’s conclusions: all the usual laws tested: Square 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Bootstrap sample data: i = 2 to 20 (original data set i = 1 )
16 17 18 19 20 21

Normal (ML), Gumbel (ML and moments), Log Square Normal (ML) - 2p
Log Normal (ML) - 2p
Gumbel EV 1(Moment) - 2p
Incomplete Gamma (ML) - 3p
Gumbel EV 1(ML) - 2p
Weibull (ML) - 2p

Normal (ML), Gamma (ML), Weibull (ML), Poisson- Poisson-Exponential (Moment) - 2p Jenkinson (spe) - 3p

Exponential (moments), Jenkinson (spe.) with MA

method are applicable to the series of Marseilles for Figure 5. Bootstrap method - Robustness of maximum
the current frequencies (T<50 yr). Three parameters annual distributions (T = 100 yr)

3
 Peaks-Over-Threshold analysis evolution of sedimentary stock and shoreline; the
Constitution of the POT series is much more complex. highest surge episodes corresponding to the
The choice of the parameters of sampling (threshold morphogenic storms too (figure 7). Thus, thresholding
and independence duration of successive events) isolates exceptional events and confirms the interest of
affects the quality of the adjustments. Therefore the surges as a magnitude indicator of the storms,
constitution of the sample must be carefully attended correlated with coastal impacts. Finally, the POT
to. Walton (2000) had fixed independence interval at method with exponential distribution provides
72 hours for US Atlantic coast without justification. adjustment results similar to those obtained with MA
Here, temporal independence criterion is obtained with method, the theoretical curve tending to over-estimate
24 hours, using autocorrelation test of step 1 advised slightly extreme events.
by Miquel (1984). Various thresholds were then tested
with methodology developed by Lang (1999): the test
of mean number of over-threshold events and the test 4.2 Intensity-Duration-Frequency curve
of the dispersion index. The distribution of mean In the preceding sections, storm surges are studied in
number of over-threshold events per year, Fig. 6, is intensity term. In this last part the following
identical to the description made by Lang ( 1999); adjustment results suggests a synthetic statistical
domain n°2 corresponding to the decreasing mean indicator used in continental hydrology: Intensity-
number of over-threshold events where the statistical Duration-Frequency curves, Figure 8.
optimum is reached (2 to 4-6) according to Mousavi
(1997). Intensity-Duration-Frequency curve - Marseille (1885-2003)
Storm events with Gumbel (ML) ajustement
140

Mean number of over-threshold values per year versus threshold

(Marseilles 1885-2003) - Independance duration : 24 h 120
32
30 100
1 2 3
28
Storm surge (cm)
Mean number of over-threshold events

26
80
24
22
60
20
18
per year

16 40
14
12 20
10
8
0
6 0 10 20 30 40 50 60 70 80
4 Duration (hour)
2
T = 2 ans T = 5 ans T = 10 ans T = 20 ans
0 T = 50 ans T = 100 ans T = 300 ans January 1978
-20 -10 0 10 20 30 40 50 60 70 80 90 100 110 120
September 1947 December 2003 September 2000
Surge threshold (cm)

Figure 7. Storm events plotted on Intensity Duration

Figure 6. Mean number of over-threshold events per Frequency curves
year
As a matter of fact, intensity is not the whole
parameter of physical signification; duration must be
This value (45-50 cm) corresponds to the one
combined with magnitude to appreciate an impact
recognized as geomorphological the reference
event. The permanence of surge over a threshold for a
threshold by Bruzzi (1998).
long time can generate more damage than a "flash"
surge even if more intense, taking into account joint
probability of wave occurrence.

5 Conclusions
The comparative analysis of several frequential models
of Marseilles (1885-2003) storm surge data set
confirms Walton (2000)’s results: all the usual laws
tested (MA method) are applicable for T<50 yr and 3 p
laws adjustment is more accurate for 50 yr <T<100 yr.
Beyond, the confidence interval of 3p distributions
widens clearly and GEV 2 type laws are here
recommended (Frechet's domain of attraction ). POT
Figure 7: Faraman Coast (Camargue) after 1997 storm
method adjustments fit best but require more care in
the data processing. The major interest of this study
Ground observations during six storms in Camargue
lies in the coincidence of the statistical optimum
(1995-1998) highlight this threshold concerning the

4
threshold and the morphological threshold observed on
the littoral of Camargue. Eventually, a synthetic
representation of storm surge including the temporal
dimension of the phenomenon is put forward.
Intensity-Duration-Frequency curves are an
operational tool for the analysis of the surge climate
and the comparison of the associated storms, which the
hydrosedimentary answer on the coast is both intensity
and duration dependent.
6 Acknowledgements
First author would like to thank G. Caude, chairman
and O. Piet vice-chairman, J. L'Her and G. Grimaldi of
CETMEF who authorized to publish this information,
B. Simon (SHOM) and G. Wöppelmann (LAREG) for
access to observation data.
7 References
Bruzzi C. (1998) “Les tempêtes et l’évolution
morphosédimentaire des plages orientales du
delta du Rhône”. Thèse sci.: Univ. de Provence
(Aix - Marseille I).
Gumbel E.J. (1958) “Statistics of extremes”,
Columbia University Press, New York
Jenkinson A.F. (1969) “Statistics of extremes.
Estimation of maximum floods”, Tech. Note
No .98, World Meteorological Organization,
Geneva, pp. 183-227.
Lang M., Ouarda T., Bobée B. (1999)
“Towards operational guidelines for over-
threshold modeling”, Journal of Hydrology,
Vol. 225, pp. 103-117
Miquel J. (1984) “Guide pratique
d'estimation des probabilités de crues”, Ed
Eyrolles, Paris.
Mousavi N. S. (1997) “Composition des lois
élémentaires en hydrologie régionale :
application à l’étude des régimes de crue”,
Thèse Univ. J. Fourier, Grenoble.
Pirazzoli P.A. (2000) “Surges, atmospheric
pressure and wind change and flooding
probability on the atlantic coast of France”,
Oceanologica Acta, Vol. 23, pp. 643-661.
Sabatier F, Suanez S. (2003) “Evolution of
the Rhône delta coast since the end of the 19th
century”, Géomorphologie: relief, processus,
environnement, Vol. 4, , pp. 283-300.
Simon B. (1994) “Statistique des niveaux
marins extrêmes le long des cotes de France”,
rapport d’étude SHOM n°0001/04, DPNM,
Ministère de l’Equipement, .
Walton T. L. (2000) “Distributions for surge
extremes”, Ocean Ing., Vol. 27, pp. 1279-
1293.
Wöppelmann G. (1997) “Rattachement
géodésique des marégraphes dans un système
de référence mondial par techniques de
géodésie spatiale”. Th. Sci. Obs. Paris.