Geostatistical Analysis of Three Dimensional Current Patterns in Coastal Oceanography: Application To The Gulf of Lions (NW Mediterranean Sea)
Geostatistical Analysis of Three Dimensional Current Patterns in Coastal Oceanography: Application To The Gulf of Lions (NW Mediterranean Sea)
Geostatistical Analysis of Three Dimensional Current Patterns in Coastal Oceanography: Application To The Gulf of Lions (NW Mediterranean Sea)
Abstract: Two geostatistical methods are used to map hydrodynamic patterns in the
Gulf of Lions (Mediterranean Sea). The aims are both methodological –
mapping vectorial data raises some difficulties – and applied – sampling
schemes from boat cruise are not convenient to get maps or to compare with
model output. From a large data set that was obtained from a shipboard ADCP
(Acoustic Doppler Current Profiler), stationary isotropic geostatistical models
were fitted for several horizontal layers. Vectors of current are characterized
by two components or by intensity and direction. A linear model of
coregionalization was used on vector components and compared to a second
approach that considers vectors as elements of the complex plane ¢. Then
interpolated maps were computed by ordinary cokriging and by ordinary
kriging in the complex plane for two different depths. Although some
difficulties remain unsolved due to the effect of time in the sampling scheme
or to some constrains (physical equations and limit conditions) that currents
must satisfy, the first results are already satisfactory and allow a better
understanding of spatial patterns than the simple plots of original data. The
same data set were also used in parallel for hydrological modelling using a
physical circulation model. Then the complex kriging approach was used to
address the spatial analysis of the residuals, i.e. difference between predicted
and observed current vectors. Residuals were highly structured in space.
367
X. Sanchez-Vila et al. (eds.), geoENV IV – Geostatistics for Environmental Applications, 367-378.
© 2004 Kluwer Academic Publishers. Printed in the Netherlands.
368 P. Monestiez, A. Petrenko, Y. Leredde and B. Ongari
1. INTRODUCTION
The hydrodynamics of coastal areas is a central issue to assess and to
understand spatio-temporal distribution of biological and chemical
parameters related to resources, global fluxes or pollution assessment. The
Gulf of Lions system is mainly forced by strong physical and meteorological
influences as the Rhone river plume, very strong winds and mesoscale
current patterns (Millot, 1990). Preliminary studies at one permanent
monitoring site (SOFI monitoring station) and on the whole gulf during
cruises (MOOGLI cruises) were not able to deal with spatial pattern
descriptions but showed their importance and the necessity to get accurate
measurement of currents for the whole area in a short time interval (Petrenko
et al., 2002). This was done during the ten SARHYGOL cruises in 2000 and
2001 to identify main patterns and seasonal effects. In this study we focused
on one cruise (June 14-15, 2000) which was used for a physical modeling
study, that allowed us to compare data interpolation and circulation model
output.
It is not frequent in geostatistics to deal with directional or vectorial data.
Lajaunie and Béjaoui (1991) proposed a kriging in the complex plane and
developed some theory on the spatial covariance models. It is probably the
first example of covariance modelling in the complex plane. They applied it
to a case study on tidal data which was very similar to our problem. In fact,
their data derived from a physical model solved by a finite element method
and then sampled to test the geostatistical approach. However, they did not
restraint the model to a variogram structure in the complex plane which
would have been a less rich model. Some other cases can be found in Chilès
and Delfiner (1999) with the interpolation of directional derivatives of a
variable. They used a model of coregionalisation and then the cokriging.
Grzebyk (1993) in a different way developed some theory in the field, but
did not work on vectorial structures. His main results were an extention of
the classical linear model of coregionalization for two or more variables to
complex framework in order to model asymmetrical crosscovariances. A
synthesis of all theses first approaches is given in Wackernagel (1995)
completed by some considerations and analysis on the available models and
methods.
Geostatistical analysis of three dimensional current patterns 369
Figure 1. Trajectory of the cruise on June 14-15, 2000 from and to the port of Marseilles.
2773 localized points with 58 different depths were measured but only the 367 regularized
points are plotted. The coast line is plotted as solid line. Dashed line is the 100m-deep isoline
and dotted line the 500m-deep isoline. Current data at the depth of 8m are symbolized by
vectors.
370 P. Monestiez, A. Petrenko, Y. Leredde and B. Ongari
1
J ij (h) E Z i ( x) Z i ( x h) Z j ( x) Z j ( x h) where i, j ^ E , N ` u ^ E , N `
2
They are modelled using a linear model of coregionalization and fitted using
the least squares procedure described in Goulard and Voltz (1992).
S
ī( h) ¦A
u 0
u g u ( h)
n n
Z E* ( xo ) ¦ ZD Z
D 1
E
E ( xD ) ¦ ZDN Z N ( xD )
D 1
n n
*
Z ( xo )
N ¦ Z 'D Z
D 1
N
N ( xD ) ¦ Z 'DE Z E ( xD )
D 1
n n n n
¦ ZD ¦
D D1
E
Z 'D
1
N
1 and ¦
D
ZD ¦ Z 'D
D 1
N
1
E
0
where the left terms remain constant for a given neighborhood xi,…,xn and the
right terms depend on k {E,N} xo.
Geostatistical analysis of three dimensional current patterns 371
3. RESULTS
Figure 3. Experimental variograms and covariogram for North and East components of
current at the depth of 8 meters. Distances are in km on the horizontal plane. The fitted model
of linear coregionalization is plotted with solid lines.
Figure 4. Experimental variograms and covariogram for North and East components of
current at the depth of 80 meters. Distances are in km on the horizontal plane. The fitted
model of linear coregionalization is plotted with solid lines.
374 P. Monestiez, A. Petrenko, Y. Leredde and B. Ongari
The two variables are the East and the North components of the horizontal
current. Experimental variograms and crossvariogram are shown in figure 3. We
assumed isotropy. A model based on three elementary structures was fitted. The
structures were a nugget effect and two nested spherical models of range 20 km
and 50 km respectively. The nugget effect remains very small and differs
slightly from zero for the variogram on North component. The cross variogram
shows that spatial variations of North and East components are quite
uncorrelated at short distance, but become correlated when distance increases,
showing dependence patterns at the scale of 40 km or more.
Figure 5. Map of currents at the depth of 8 meters that was obtained by cokriging.
Figure 6. Map of currents at the depth of 80 meters that was obtained by cokriging. Image
legend is the same than for figure 5.
At the depth of 80 meters, the cross structure vanish and the two variograms
show more regularity close to the origin (Figure 4). The coregionalization was
Geostatistical analysis of three dimensional current patterns 375
Complex kriging was then applied to the map of currents at the depth of 8m and
results are shown in Figure 8. This method seems to work as well as the cokriging of
the current components. In fact, in this study, differences between maps obtained by
the two different krigings are smaller than differences we can observe when we
choose other variogram models as nested or simple exponential, suppressing or not
the nugget effect, in place of the two nested spherical models.
Figure 8. Map of currents at the depth of 8 meters that was obtained by complex kriging.
Image legend is the same than for figure 5.
Using these residual vectors as data, we reapply the complex kriging method -
after residual variogram fitting - in order to get a kriged map of residuals. Figure 10
shows the result for the depth of 8 meters. It looks like an over estimation of main
stream current (or a bad positioning too close to the coast) by the model and a wrong
rotation pattern in the western part of the Gulf of Lions.
Figure 9. Measurement of currents (in black) along the trajectory compared with the output of
the circulation model (in gray, same scale) at same time, same location and at the depth of 8
meters.
Figure 10. Map of the residuals (data versus circulation model) at the depth of 8 meters that
was obtained by complex kriging.
A last point concerns the time. The boat cruise takes some time and the
currents may change in between. This can be checked when the boat crosses
it own trajectory. We tried to take into account the time but there was some
time-space confounding effects in the measurement procedure itself. For
most data, because of the constant boat speed, a pair of points at a given
distance corresponds to a given time lag. So it is impossible to get points at
different distances for a given time lag excepted if two boats are
simultaneously measuring currents, and it was not possible to get regularly
spaced time intervals for points at a given small distance for this kind of boat
trajectories. Improvement could be easily done with a boat trajectory that
378 P. Monestiez, A. Petrenko, Y. Leredde and B. Ongari
often comes back on itself or stops to get a better and “orthogonal” sampling
of time and space pairs. Modelling fully time and space components in the
variogram model should improve comparisons between circulation models
and data.
ACKNOWLEDGMENTS
This work was supported by the PNEC (National Program for Coastal
Environment) of the CNRS (French National Institute for Scientific
Research). During most part of this research, Pascal Monestiez was visiting
the Centre d'Océanologie de Marseille (Université de la Méditerranée -
Marseille II) which supported his stay.
REFERENCES
1. Chelton D. B. (1994). Physical oceanography: a brief overview for statisticians. Statistical
Science, 9:150-166
2. Chilès, J-P. (2001). Hydrogeology and geostatistics. In geoENV III: Geostatistics for
Environmental Applications, Monestiez P., Allard D. and Froidevaux R. Eds, Kluwer
Academic Publishers, Dordrecht, 1-16.
3. Chilès, J-P. and Delfiner, P. (1999).Geostatistics: Modeling Spatial Uncertainty. Wiley Series in
Probability and Statistics, Wiley. 695 p.
4. Courault, D. and Monestiez, P. (1999). Spatial interpolation of air temperature according to
atmospheric circulation patterns in Southeast France. International Journal of Climatology,
19:365-378.
5. Cressie, N. (1993). Statistics for Spatial Data, rev. edn. Wiley: NY, 900 pp.
6. Estournel C., Durrieu de Madron, X., Marsaleix, P. Auclair, F., Julliand, C. and Vehil, R. (2002).
Oberservation and modelisation of the winter coastal oceanic circulation in the Gulf of Lions
under wind conditions influenced by the continental orography (FETCH experiment). J.
Geophys. Res. (in press).
7. Goulard, M. and Voltz, M. (1992). Linear coregionalization model: tools for estimation and
choice of multivariate variograms. Mathematical Geology, 24:269-286.
8. Grzebyk, M. (1993). Ajustement d'une Coré gionalisation Stationnaire. Thesis, Ecoles des Mines
de Paris, 154 pp.
9. Lajaunie Ch. and Béjaoui R. (1991). Sur le krigeage des fonctions complexes. Note interne N-
23/91/G. Centre de Géostatistique, Ecole des Mines de Paris, Fontainebleau, 24pp.
10. Millot, C. (1990) The Gulf of Lion's hydrodynamics. Continental shelf Research, 10:885-894.
11. Panel on statistics and oceanography (1994). Report on statistics and physical oceanography
(with discussion) Statistical Science, 9:167-221
12. Petrenko, A. (2002). Circulation features in the Gulf of Lions, NW Mediterranean sea; summer
versus winter conditions. Oceanol. Acta, N° special PNEC-Golfe du Lion (in press).
13. Wackernagel, H. (1995). Multivariate Geostatistics, Springer-Verlag, Berlin, 256 pp.