Integrating Prior Knowledge and Locally

Varying Parameters with Moving-GeoStatistics:

Methodology and Application to Bathymetric
Mapping
Cedric Magneron, Nicolas Jeannee, Olivier Le Moine,
and Jean-Francois Bourillet
Abstract The paper aims at presenting an innovative methodology, called M-GS
(M-GeoStatistics), which is fully dedicated to the local optimization of parameters
involved in variogram-based models. M-GS considers the structural and computational parameters as a set of dependant parameters to be spatially optimized. The
optimization process, which may be guided by objective or subjective criteria, is
carried out during a M-structural analysis phase that leads to a set of spatially variable structural and computational parameters.
The methodology is applied for bathymetry mapping. The availability of accurate seafloor estimates is essential for numerous oceanographic projects, including
hydrographic, oceanographic and biological models, sedimentary processes, etc.
Seafloor usually presents strong non stationarity and complex structures, such as
small channels with varying orientations, spatially varying measurements errors, local heterogeneities for coastal areas, or deep canyons within general gentle slope
for continental margins. The adequacy of the M-GS methodology in this framework
is illustrated and compared with classical estimates for the Marenne-Oleron coast
(West of France). Moreover such methodology could be used to input different local
structures into a general model in the aim of a regional synthesis.

C. Magneron ()
ESTIMAGES, 10 Avenue de Quebec, 91140 Villebon-sur-Yvette, France
e-mail: cedric.magneron@estimages.com
N. Jeannee
GEOVARIANCES, 49bis Avenue Franklin Roosevelt, BP91, 77212 Avon, France
e-mail: jeannee@geovariances.com
O.L. Moine
IFREMER, Laboratoire Environnement-Ressource des Pertuis Charentais, Avenue de Mus de
Loup, 17390 La Tremblade, France
e-mail: olemoine@ifremer.fr
J.-F. Bourillet
IFREMER, Dep. Geosciences Marines, Laboratoire Environnements Sedimentaires, BP70,
29280 Plouzane, France
e-mail: Jean.Francois.Bourillet@ifremer.fr

P.M. Atkinson and C.D. Lloyd (eds.), geoENV VII Geostatistics for Environmental
Applications, Quantitative Geology and Geostatistics 16,
c Springer Science+Business Media B.V. 2010
DOI 10.1007/978-90-481-2322-3 35, 

405

406

C. Magneron et al.

1 Introduction
Today, most geostatistical methods rely on a global variogram model. The variogram
allows to build effective estimation (kriging) and simulation operators by catching
the mean spatial correlation inherent to a data set. These methods commonly assume
stationarity for the underlying random function. This assumption is too constraining
in numerous applications, as soon as the target area becomes large or involves complex structural patterns. Applying stationary approaches in such cases, even locally
with a moving neighbourhood, can lead to unsuitable estimates and non stationary approaches are preferable to some extent, provided that one is ready to accept
to loose some control on the underlying structural model. Furthermore, even non
stationary algorithms hardly handle prior knowledge nor reproduce precisely complex structures, such as local anisotropies, spatially varying small-scale structures
or heterogeneity, etc.
The M-GS methodology is suitable for processing data in a wide range of such
non stationary contexts.

2 Conventional Variogram-Based Models

2.1 Global Approach
The majority of geostatistical models that are daily implemented in the industry
are variogram-based models see Dubrule (2003), for example. They are used for
processing spatially distributed data, especially in natural resources domains such as
mines, petroleum and environment. They are mainly devoted to mapping, filtering
and uncertainty management applications.
Variogram-based models rely generally on the modelling of a statistical function,
the experimental variogram, which depicts the mean spatial correlation between
data samples. When data can be considered as the result of a stationary random
process, the variogram model is fitted directly to the experimental variogram, which
is supposed to be representative of the whole data field or of a well-separated area
of the data field. Based on the variogram model, effective estimation (kriging) and
simulation operators are built and applied to the data set.
In the second-order stationary case, the variogram-based approach is rather
intuitive as some parameters of the model may be related directly to the observation of the data themselves. Non-stationary models, such as IRF-k models
(Matheron, 1971; Chil`es and Delfiner 1999), are more intricate and lead to less
control on the underlying structural model. It justifies the common strategy of transformation for working in a stationary framework as in the universal kriging case,
despite the observed bias of the variogram of the residuals (Pardo-Iguzquiza and
Dowd, 1998).

Integrating Prior Knowledge and Locally Varying Parameters

407

2.2 Variogram-Based Model Parameters

2.2.1 Structural Parameters
In the stationary case, variogram modelling is driven through a two-steps phase
called structural analysis. The first step consists of interpreting the experimental
variogram computed from the data. This step is rather likely to involve the users
knowledge about their data set. Based on the first step conclusions, the second step
aims at fitting a single or a set of parameterized functions to the experimental variogram, thus defining the variogram model. Broadly speaking, structural parameters
are the parameters that are related to the variogram model such as range(s), sill(s),
anisotropy coefficient(s), etc.

2.2.2 Computational Parameters

In order to run variogram-based estimation and simulation algorithms, some computational parameters must be tuned. They are mainly tied to the moving neighbourhood used for selecting data points surrounding the target point (the point to be
estimated or simulated). In practice, the computational parameters are often utilized
for managing processing times, specifically when dealing with large data sets, or for
adjusting the neighbourhood according to the samples pattern.

2.3 Limits
Variogram-based estimation and simulation results are sensitive to structural and
computational parameters. Although sensitivity may be highly variable depending
on some data characteristics, such as sampling density or variable continuity for
example, it is usually a factor. This point is often not appreciated while running
variogram-based models.
More specifically, sensitivity to the parameters can be very problematic when
faced with complex structural environment or specific acquisition patterns. In such
cases, global stationary models may correspond to local data characteristics and can
lead to unexpected poor results.

408

C. Magneron et al.

3 M-GS (Moving-GeoStatistics) Models

3.1 Principle
M-GS methodology is fully dedicated to the local optimization of parameters
involved in variogram-based models. M-GS considers the structural and computational parameters as a set of dependant parameters to be spatially optimized. The
optimization process, which may be guided by objective or subjective criteria, is carried out during a M-structural analysis phase that leads to a set of spatially variable
structural and computational parameters.

3.2 M-Parameters
M-parameters are locally optimized versions of structural and computational parameters of variogram-based models. They vary spatially over the data field. In the
past, non-stationarity has been explored for several parameters, such as anisotropy,
especially in the environment domain (see Caetano et al., 2004 for example). When
dealing with these models the major challenge is to get stable variations of the parameters and as far as possible to automate the parameter determination process.
Several approaches are possible to compute M-parameters. A simple one merely
consists in computing local variogram parameters in adjacent areas of the data field
and then to smooth the obtained parameters in order to make them available at every
target grid node. More sophisticated algorithms currently under development are
based on automatic validation techniques. They simplify the determination of the
M-parameters and lead to promising results on various real cases that have been
tested.
One example of results obtained with an automatic validation approach is presented in Fig. 1, which displays a 2D seismic data set (Fig. 1a) and one associated
M-parameter map corresponding to the range variations of an isotropic spherical
model (Fig. 1b). An interpolation error criterion has been used for determining the
optimal parameters. The north-eastern part of the data field appears to be less structured (range smaller) than the rest of the data field. The M-parameters are used to
map the seismic data by ordinary kriging (Fig. 1c).
It should be noted that the M-structural analysis process involves some dependency relationship between several parameters. For example, in the second-order
stationary case, the size of the moving neighbourhood in one direction is related
linearly to the range of the largest scale structure in that direction. More complex
relationships can be introduced into the optimization process.

Integrating Prior Knowledge and Locally Varying Parameters

409

Fig. 1 Seismic data mapping by kriging: (a) data set, (b) spatially varying range, (c) M-GS mapping (d) conventional mapping by global kriging

3.3 Advantages
M-GS ensures a better correspondence between the geostatistical model and the
data. As a consequence, spatial estimation and simulation results are more precise
than those obtained with conventional variogram-based models. Regarding the previous seismic data mapping example, the improvement has been quantified through
a cross-validation process. The M-GS map is on average 20% more precise than
the conventional kriging map (Fig. 1d) in the north-eastern part of the field. In other
words the estimation errors have been reduced by 20%.
Moreover, M-GS opens the way to advanced geostatistical mapping (even simulating) practices by allowing the user to introduce his structural a priori knowledge
about the data field directly into the spatial estimation model. In that way geostatistical mapping is no longer a variogram guided process aiming at generating the

410

C. Magneron et al.

Fig. 2 M-GS guided mapping: (a) M-GS mapping guided by channel interpretation, (b) conventional mapping by global kriging

most probable map, but a human process aiming at generating the most probable
desired map. This last case is illustrated in Fig. 2. Channel information, that could
result from subjective interpretation, is translated in terms of M-parameters and then
introduced into the kriging model for mapping 25 depth data samples leading to a
channel-driven map (Fig. 2a) to be compared with a conventional global approach
map (Fig. 2b). The former presents a greater continuity for the channel (red arrow)
than the conventionally-derived map which displays several individual depressions.

4 M-GS Application to Bathymetric Mapping

4.1 Context
The availability of accurate seafloor estimates is essential for numerous oceanographic projects, including hydrographic, oceanographic and biological models,
sedimentary processes, seismic interpretation of buried channels or canyons, etc.
The seafloor usually presents strong non stationarity and complex structures, such
as small channels with varying orientations, spatially varying measurement errors,
local heterogeneities for coastal areas, or deep canyons within general gentle slope
for continental margins.
Conventional variogram-based models often fail to produce consistent maps
within such complex structural environment. More advanced models, such as M-GS
models, can be applied advantageously.

4.2 Data Set Description

Marenne-Oleron (West of France) is a semi-enclosed Bay, and the first oyster farming zone in Europe. Shellfish culture activity induces silting on large intertidal mud
and sandy-mud flats. Several channels incise the inlet between the coast line and
Oleron Island. They are mainly controlled by strong tidal currents (up to 1.4 knots
during the spring tides) with a residual ebb delta offshore the SW channel. The data
set used in this work consists in more than 2,000 sample points, organized along

Integrating Prior Knowledge and Locally Varying Parameters

411

Fig. 3 Marenne-Oleron data set: (a) data set, (b) target area

lines from West to East (Fig. 3a). Samples are separated by few meters within lines.
The (North-South) gap between two lines is about 100 m. Data were acquired with a
single beam echoes sounder for the monitoring of the evolution of the muddy layer.
A target area (Figs. 3b and 4) is selected for illustrating conventional and M-GS
mapping result differences.

4.3 Conventional Variogram-Based Mapping

For kriging purposes, an experimental variogram is computed within the target area.
An anisotropic spherical model (range 800 m along X direction, 1,200 m along
Y direction) is fitted to the experimental variogram (Fig. 5) and used to map the
depth data.
The resulting bathymetric map is shown in Fig. 6. Major structures have been
well imaged. However when looking into detail, the map contains some artefacts on

412

C. Magneron et al.

Fig. 4 Target area

Fig. 5 Global variogram modeling

the walls of the channels which are mainly due to the line-oriented organization of
the data within strongly anisotropic areas. Moreover, one micro-channel (red arrow),
which is interpretable on the original data set, has not been reproduced at all.
Therefore, a more refined model is needed to reduce the artefacts and to image
correctly the interpreted micro-channel.

Integrating Prior Knowledge and Locally Varying Parameters

413

Fig. 6 Conventional mapping results

4.4 M-GS Mapping

The M-GS methodology enables the determination of locally optimized structural
and computational parameters. For the current application, a specific emphasis is put
on the range, the anisotropy and the related orientation of a generic spherical model.
Firstly parameters are optimized during a M-structural analysis step, leading to several M-parameters maps. One resulting M-parameter map is shown in Fig. 7a. This
map illustrates the spatial variations of the shortest axes of the anisotropy ellipsoid.
Afterwards prior knowledge is integrated into the model: additional information regarding the interpreted micro-channel is introduced into the M-parameter maps. The
previous M-range map is transformed as shown in Fig. 7b.
Finally the M-parameters are used to estimate the bathymetry. Mapping results
are displayed in Fig. 8. The artefacts identified on the conventional map are no
longer visible and the interpreted micro-channel is imaged. In this case it is evident that the M-GS map is of better quality than the conventional map.

414

C. Magneron et al.

Fig. 7 Short range map: (a) short range map without micro-channel interpretation, (b) short range
map with micro-channel interpretation

Fig. 8 M-GS mapping results

Integrating Prior Knowledge and Locally Varying Parameters

415

5 Conclusion
The popularity of stationary variogram-based models is mainly explained by the
easy interpretation which is made of the involved parameters. In particular, some
structural parameters can be directly linked to the observation of the structural
properties of the data. Advanced methodologies, which allow management of spatial variations in these parameters, increase the accuracy of variogram-based model
results, especially when processing large data sets and/or areas with complex structural patterns.
In this direction, the M-GS methodology, which is dedicated to the optimization
of variogram-based models parameters, has proved to be promising when applied
to bathymetric or seismic interpretation data in a complex structural environment.
It could be useful too for mapping aquifers bottom architecture, for example. The
adequacy of the M-GS methodology in the framework of bathymetric mapping for
Marenne-Oleron coast (West of France) is obvious. Moreover such a methodology
could be used to input different local structures into a general model with the aim of
a regional synthesis.
Acknowledgments The authors would like to thank Mr. Fazilleau (Port Autonome de La
Rochelle) and The Conseil general de Charente Maritime for having provided the MarenneOleron dataset.

References
Caetano H, Pereira MJ, Guimar`aesC (2004) Use of factorial kriging to incorporate meteorological
information in estimation of air pollutants. In: Sanchez-Vila X, Carrera J, Gomez-Hernandez
J (eds) GeoENV IV geostatistics for environmental applications, Part 2. Kluwer, Dordrecht,
pp 5565
Chil`es JP, Delfiner P (1999) Geostatistics: modeling spatial uncertainty. Wiley series in probability
and statistics, New York
Dubrule O (2003) Geostatistics for seismic data integration in earth models. Distinguished Instructor Short Course Distinguished Instructor series, N 6, SEG & EAGE
Matheron G (1971) La theorie des functions aleatoires intrins`eques generalisees. Note
Geostatistique N 117. Technical report N-252. Centre de Geostatistique, Fontainebleau, France
Pardo-Iguzquiza E, Dowd PA (1998) The second-order stationary universal kriging model revisited. Math Geol 30(4):347378 (32)