EMM 2513 Geostatistics Brief Notes

Prof Rop. 2023 JKUAT, MMPE

Introduction to Geostatistics — Course Notes
Overview:
What is geostatistics?
 Data analysis and spatial continuity modeling (Journel, 1989).
 Establish quantitative measure of spatial correlation to be used
for sub- sequent estimation and simulation (Deutsch, 2002).
The following introduction and overview materials are based on
compilation of several source materials (see full references in Sec.
1.5.1).

1.1 Why Geostatistics?

Classic statistics is generally devoted to the analysis and
interpretation of un- certainties caused by limited sampling of a
property under study.
Geostatistics however deviates from classic statistics in that
Geostatistics is not tied to a population distribution model that
assumes, for example, all samples of a population are normally
distributed and independent from one another.
Most of the earth science data (e.g., rock properties, contaminant
concentrations) often do not satisfy these assumptions as they can
be highly skewed and/or possess spatial correlation (i.e., data values
from locations that are closer together tend to be more similar than
data values from locations that are further apart).
To most geologists, the fact that closely spaced samples tend to be
similar is not surprising since such samples have been influenced by
similar physical and chemical depositional/transport processes.

Compared to the classic statistics which examine the statistical

distribution of a set of sampled data, geostatistics incorporates both
the statistical distri- bution of the sample data and the spatial
correlation among the sample data. Because of this difference, many
earth science problems are more effectively ad- dressed using

1
geostatistical methods. As stated by Marc Cromer (in Geostatis- tics
for environmental and geotechnical applications, 1996, ASTM
International, edited by Rouhani et al.):

“Geostatistical methods provide the tools to capture, through

rigorous exami- nation, the descriptive information on a
phenomenon from sparse, often biased, and often expensive sample
data. The continued examination and quantitative rigor of the
procedure provide a vehicle for integrating qualitative and quantita-
tive understanding by allowing the data to “speak for themselves”. In
effect, the process produces the most plausible interpretation by
continued examination of the data in response to conflicting
interpretations. ... The application of geo- statistics to environmental
problems (e.g., groundwater contaminant cleanup) has also proven a
powerful integration tool, allowing coordination of activities from
field data acquisition to design analysis. For example, data collection
is often incomplete, resulting in uncertainties in understanding the
problem and increasing the risk of regulatory failure. While this
uncertainties can often be reduced with additional sampling, the
benefits must be balanced with increasing cost. ... Thus, geostatistics
offers a means to quantify uncertainty, while leveraging existing data
to support sampling optimization.”

1.2 Geostatistical Prediction

The goal of geostatistics is to predict the possible spatial distribution
of a prop- erty. Such prediction often takes the form of a map or a
series of maps. Two basic forms of prediction exist: estimation
(Figure 1.1) and simulation (Fig- ure 1.2). In estimation, a single,
statistically “best” estimate (map) of the spatial occurrence is
produced. The estimation is based on both the sample data and on a
model (variogram) determined as most accurately representing the
spatial correlation of the sample data. This single estimate or map is
usu- ally produced by the kriging technique. On the other hand, in
simulation, many equal-likely maps (sometimes called “images”) of
the property distribution are produced, using the same model of

2
spatial correlation as required for kriging. Differences between the
alternative maps provide a measure of quantifying the uncertainty,
an option not available with kriging estimation.

Geostatistics has played an increasing role in both groundwater

hydrology and petroleum reservoir characterization and modeling,
driven mainly by the recognition that heterogeneity in petrophysical
properties (i.e., permeability and porosity) dominates groundwater
flow, solute transport, and multiphase migra- tion in the subsurface.

3
Geostatistics, by transforming a sparse data set from the field into a
spatial map (kriging estimation), offers a means to recreate het-
erogeneity to be incorporated into numerical flow and transport
modeling. On the other hand, by transforming a sparse data set into
multiple spatial maps (unconditional/conditional simulations), it
offers a means of evaluating the un- certainties on modeling due to
the uncertain nature of each map (Figure 1.3). In both reservoir
simulation and groundwater modeling, for example, Monte Carlo
simulation is a popular technique. Note that this uncertainty reflects
our lack of knowledge about the subsurface, though the geological
“groundtruth”, albeit unknown, is deterministic and certain.

1.3 Geostatistics versus Simple Interpolation

In geostatistical estimation, we wish to estimate a property at an
unsampled location, based on the spatial correlation characteristics
of this property and its values at existing sampled locations. But, why
not just use simple interpo- lation? How is spatial correlation
incorporated in the geostatistical approach? A simple example may
illustrate this point more clearly (Figure 1.4): we know permeability
at n sampled locations, we wish to estimate the permeability at an
4
unsampled location, z0. Using inverse distance, the unknown value
can be evaluated as:

(estimate) (weight)

We can see that the above relation is a linear estimator, i.e., z0 is a

weighted sum of the n known values. Each weight (wi) (assigned to a
known zi) is determined by the distance of the known data point to
5
the unknown data point. For n = 7, for example, the weights can be
calculated easily as shown in Figure 1.5.

Using this scheme, the weights assigned to points 1, 2, 4, 6 are all

equal to 0.2. However, from the understanding of geology, we realize
that permeability within the elongated sand body should be more
similar in the lateral direction. Thus, points 4 and 6 should be given
higher weights than points 1 and 2. This is obviously not the case
when using inverse distance. Thus, in conventional inter- polation
methods (e.g., inverse distance, inverse distance squared),
information on spatial correlation is not incorporated. On the other
hand, geostatistical estimation considers both distance and spatial
correlation. In general, geosta- tistical estimation consists of 3 steps:
(1) examining the similarity between a set of sample (known) data
points via an experimental variogram analysis;
(2) fitting a permissible mathematical function to the experimental
variogram;
(3) conducting kriging interpolation based on this function.
In the above example, the spatial correlation will be revealed by the
more similar values of z4 and z6 (step (1)). It will be modeled via step
(2) (variogram modeling). Then, using kriging, we’ll find that the
weights assigned to points 4 and 6 will increase (those of 1 and 2 will
decrease accordingly since the total weight must sum to 1.0) (step
(3)). In kriging, based on the new weights, a best linear unbiased
estimate of z0 is obtained. Further (though sometimes optional
depending on the goal of the study), uncertainty in the estimated
field is additionally evaluated. In this class, we’ll use many exercises
to illustrate how to conduct a geostatistical study.

Given the same set of sampled data, interpolation results using IDS
(di is replaced by d2i ) and kriging can look drastically different
(Figure1.6). However, does this mean that kriging is the preferred
interpolation method regardless of the data? It turns out, there are
situations when the sampled data are simply not good for kriging
(we’ll explore this aspect when we look at the “pitfalls” of conducting

6
a variogram analysis). Given such data—either too unreliable or too
sparse and widely spaced to capture the spatial correlation of the
true property field, the conventional IDS may give just as good result.
The decision of which method to use is in a way data-driven. Usually,

an increase in sample quality or density will affect which method

may be the most appropriate for the study.

1.4 Limitations
What is not geostatistics?
Interestingly, geostatistics models mathematical objects, not
geological objects. For example, given a set of spatial measurements
of isopach values, a geologist can create various contour maps based
on his/her understanding of the underlying geology (Figure 1.7). This
process is best described as pattern recognition—the geologist has
an existing idea of the underlying geology when doing the
interpretation. Geostatistics, however, does not recognize pattern,
rather, it is based on a set of mathematical principles.

7
As stated by Andre ́Journel (1989), “geostatistics is an art, and as
such, is neither completely automatable nor purely objective”. In an
experiment conducted by the US EPA, 12 independent
geostatisticians were given the same dataset and asked to perform
the same kriging. The 12 results were very different due to widely
different data analysis conclusions, variogram models, choices of

kriging type, and search strategy. As stressed by Journel, “that there

are no accepted universal algorithm for determining a
variogram/covariance model, that cross-validation is no guarantee
that an estimation procedure will produce good estimates at
unsampled locations, that kriging needs not be the most appropriate
estimation method, and that the most consequential decisions of any
geostatistical study are made early in the exploratory data analysis”.
In this class, I will repeatedly emphasize the importance of
understanding our data, via exploratory analysis, trend analysis, error
identification, and dealing with sampling issues and non-stationarity.

8
From both the environmen- tal engineering and petroleum reservoir
modeling literature, I present “rules of thumb” or “best practice”
guide that is recommended by experts in the field. Further, it is my
recommendation that before you embark on a geostatistical study,
you should research the literature for analysis conducted on similar
data in the past. You can often learn a lot from past studies and
hopefully, you can try to avoid pitfalls that others had stumbled upon
before you. In the end of this class, I will present a lecture on
literature search and point to further resources that you can use to
solve your own problems.

Thus, geostatistics is not a black box. Without understanding its

fundamental assumptions and limitations, an untrained person is
more likely use it incorrectly. As summarized by Journel (1989):
Geostatis- tics is a tool: it cannot produce good results from bad
data. It cannot replace common sense, good judgment, or
professional insight. Throughout the course, I’ll pay equal attention
to its limitations as well as its useful applications. In practice, as more
data become available, the geostatistical procedure often need to be
repeated, the data re-analyzed or reinterpreted.

Another point to make is that estimation or simulation based on

variograms cannot very well capture curvilinear features, e.g., curved
channels (Figure 1.8). To overcome such limitations, recent
development includes multiple point statis- tics (where correlation is
characterized among multiple data points and then incorporated into
simulations), pluralGaussian simulation (several correlated
populations can be superimposed), and hybrid or hierarchical
approaches (e.g., kriging is used to create property distribution
within a higher-order geobody created via either deterministic or
stochastic means, often object-based). These are currently areas of
active research.
1.5 This Class
1.5.1 References

9
In this class, a fairly rigorous mathematical treatment is presented.
This course is thus designed at the upper undergrad and graduate
level, appropriate for the level of rigor contained herein. Course
lecture is the key, though most materials are assembled based on
several textbooks, tutorials, and lecture notes, each with its own
emphasis:
 Engineering Geostatistics, Course Notes, Randal Barnes, 2000,
Depart-ment of Civil Engineering, U of Minnesota.
 An Introduction to Applied Geostatistics, Isaaks & Srivastava
(I&S), 1989, Oxford University Press.
 Geostatistical Reservoir Modeling, Clayton Deutsch, 2002,
Oxford Uni- versity Press.
 Fundamentals of Geostatistics in Five Lessons, Andre Journel,
1989, Short Course in Geology, vol 8, Presented at the 28th
International Geological Congress, Washington, D. C.
 Introduction to Geostatistics, Application in Hydrogeology,
Peter Kitani- dis, 1997, Cambridge University Press.
 GSLIB: Geostatistical Software Library and User’s Guide, Clayton
Deutsch & Andre Journel, 2nd Edition, Oxford University Press,
1997.

Bear in mind that the current course can only serve an

introductory purpose: it cannot hope to cover every aspect of the
subject as presented in the refer- ences, nor will we have time to
explore many advanced topics, as they are being continuously
developed and refined in the literature. In particular, the topics of
this course are limited to stationary random space function (RSF)
(station- arity here, roughly speaking, means that the mean,
variance, and variogram do not change with position in the data
field; if we have time, we’ll cover Simple Kriging which is not based
on assuming stationary RSF). Although there are geostatistical
estimation methods developed for non-stationary RSF, the most
widely used ones are based on stationary RSF. However, the major
kriging techniques used in practice will be given in another separate
full class notes to follow this introductory one.

10
Finally, all lectures are rooted in a fairly rigorous statitistical raw data
framework to interpret statistics analysis of the samples, population,
universe, etc.

Hopefully, such an approach will better prepare you for the more
advanced topics or doing independent research. The exercises are
designed to help you understand both the strength of the
geostatistical methods and the various pitfalls you may encounter
when working with raw data and the suggested solutions.

The suggested reading list at the end of each chapter presents either
example applications of geostatistics in different geoscience
specialties or select topics specific to geostatistical reservoir
simulation. I will also post additional notes in brief which will hel you
understand this brief unit course.

11