..

..
..
..
.. Gregg Babish
GIS Specialist

Ecological Research Division

Environmental Conservation Branch
2365 Albert Street, Room 300
Regina, Saskatchewan S4P 4K1

Edition 2006.03

July 5, 2006

Environmen
t Canada

Geostatistics Without
Tears
. . . . . . . . . . . . . . . . . .

A practical Guide to Surface

Interpolation, Geostatistics,
Variogramsand Kriging

Gregg Babish
GIS Specialist
Environment Canada
Regina, Saskatchewan

Geostatistics Without Tears

Geostatistics Without Tears 2
PREFACE .........................................................................................................................7

INTRODUCTION ............................................................................................................9

INTERPOLATION TECHNIQUES...............................................................................9
SPATIAL INTERPOLATION ..............................................................................................10
Deterministic/Stochastic ...........................................................................................10
Global/Local.............................................................................................................10
Exact/Approximate ...................................................................................................1 1
Gradual/Abrupt ........................................................................................................11
CLIMATE DATA .............................................................................................................11
ANUSPLIN................................................................................................................12
Centro Internacional de Agricultura Tropical .........................................................13
Climatologically Aided Interpolation .......................................................................13
Daymet......................................................................................................................14
Gradient plus Inverse Distance Square ....................................................................14
Optimal Interpolation ...............................................................................................15
Parameter-evaluation Regressions on Independent Slopes Model ..........................15
Thin-Plate Spline ......................................................................................................16
DIGITAL ELEVATION MODELING ...................................................................................16
FOURIER SERIES ............................................................................................................1 8
GEOSTATISTICS .............................................................................................................18
INVERSE DISTANCE WEIGHTED .....................................................................................1 9
NEAREST NEIGHBOUR ...................................................................................................20
SPLINES .........................................................................................................................20
TREND SURFACE ANALYSIS...........................................................................................2 1

TRIANGULATION............................................................................................................21
HISTORY OF KRIGING ..............................................................................................23

DATA ANALYSIS..........................................................................................................27

SUMMARY DATA ANALYSIS ...................................................................................27

SUMMARY STATISTICS ..................................................................................................27

MEASURES OF SHAPE ....................................................................................................28
Frequency Distribution.............................................................................................28
Cumulative Distribution ...........................................................................................28
Skewness ...................................................................................................................28
Kurtosis.....................................................................................................................29
Normal Probability Plot ...........................................................................................29
MEASURES OF LOCATION ..............................................................................................31
Mean .........................................................................................................................31
Median ......................................................................................................................31
Mode .........................................................................................................................31
Quantiles...................................................................................................................32
MEASURES OF SPREAD ..................................................................................................32
Range ........................................................................................................................32
Variance....................................................................................................................32
Standard Deviation...................................................................................................32
Interquartile Range...................................................................................................33
Coefficient of Variation ............................................................................................33
CORRELATION ANALYSIS ..............................................................................................34
Coefficient of Correlation.........................................................................................34

Geostatistics Without Tears 3

 Coefficient of Determination ....................................................................................34
Standard Error of Estimate ......................................................................................35
EXPLORATORY DATA ANALYSIS..........................................................................37

CENSORED DATA...........................................................................................................37
CO-LOCATED SAMPLES..................................................................................................37
DECLUSTERING..............................................................................................................37
EXTREME VALUES.........................................................................................................38
SKEWED DATA WITH ZEROS ..........................................................................................38
SPATIAL OUTLIERS ........................................................................................................39
STRATIFICATION (DATA SET SUBDIVISION) ...................................................................39
TIME ..............................................................................................................................40
TRANSFORMATION ........................................................................................................40
Arcsine Transform ....................................................................................................40
Box-Cox Transform ..................................................................................................40
Lognormal Transform...............................................................................................41
Normal Score Transform..........................................................................................41
Scale Transform........................................................................................................42
Square Root Transform ............................................................................................42
OFFSETS AND BACK-TRANSFORMS ................................................................................42
Offset.........................................................................................................................42
Back-transformation.................................................................................................42
TRENDS .........................................................................................................................43
APPLIED GEOSTATISTICS .......................................................................................45

VARIOGRAMS..............................................................................................................45
CALCULATING THE VARIOGRAM....................................................................................46

First Lag Distance ....................................................................................................4 6

Second Lag Distance ................................................................................................46
Third Lag Distance...................................................................................................4 7
Fourth Lag Distance.................................................................................................47
Fifth Lag Distance ....................................................................................................4 8
Sixth Lag Distance....................................................................................................48
GRAPHING THE SEMIVARIOGRAM..................................................................................50
THE SEMIVARIOGRAM...................................................................................................5 1
Coordinate System ....................................................................................................5 1
Lag............................................................................................................................51
Range ........................................................................................................................53
Sill.............................................................................................................................5 3
Nugget.......................................................................................................................5 4
Variance Cloud.........................................................................................................56
Scattergram ..............................................................................................................56
SEMIVARIOGRAM MODELS ............................................................................................57
Linear........................................................................................................................57
Spherical...................................................................................................................5 8
Exponential...............................................................................................................5 8
Gaussian ...................................................................................................................5 8
Hole Effect ................................................................................................................5 9
ANISOTROPIC VARIOGRAMS..........................................................................................60
Nugget Variance .......................................................................................................60
Anisotropic Variogram Surface ................................................................................60
Geometric Anisotropy...............................................................................................6 1
Zonal Anisotropy ......................................................................................................6 2
ALTERNATE VARIOGRAMS ANDAUTOCORRELATION MEASURES..................................63

Geostatistics Without Tears 4

 Correlogram .............................................................................................................63
Covariance................................................................................................................6 3
Cross-K.....................................................................................................................63
Drift ..........................................................................................................................6 4
Fractal ......................................................................................................................64
Geary’s C..................................................................................................................65
Inverted Covariance .................................................................................................6 5
Madogram ................................................................................................................6 5
Moran’s I ..................................................................................................................6 6
Relative Variograms.................................................................................................6 7
Ripley’s K .................................................................................................................68
KRIGING........................................................................................................................69

THE KRIGING ALGORITHM ............................................................................................6 9

KRIGING VARIATIONS ................................................................................................... 70
Ordinary Kriging......................................................................................................70
Simple Kriging..........................................................................................................70
Universal Kriging .....................................................................................................70
Kriging with External Drift.......................................................................................71
PUNCTUAL AND BLOCK KRIGING ..................................................................................71
Punctual Kriging ......................................................................................................7 1
Block Kriging............................................................................................................72
KRIGING SEARCH NEIGHBORHOOD................................................................................72
CROSS-VALIDATION ......................................................................................................7 3
KRIGING PREDICTORS....................................................................................................7 4
Cokriging ..................................................................................................................7 4
Disjunctive Kriging...................................................................................................74
Indicator Kriging......................................................................................................74
5
Lognormal Kriging...................................................................................................7
Median Indicator Kriging.........................................................................................7 6
Multiple Indicator Kriging .......................................................................................76
Isofactorial Disjunctive Kriging ...............................................................................76
Probability Kriging...................................................................................................7 7
Regression Kriging...................................................................................................7 7
Residual Indicator Kriging.......................................................................................77
KRIGING AND REMOTELY SENSED IMAGES....................................................................77
GEOSTATISTICAL CONDITIONAL SIMULATION .............................................7 9
CONDITIONAL COSIMULATION ......................................................................................80
GAUSSIAN SIMULATION.................................................................................................80
PLURIGAUSSIAN SIMULATION........................................................................................80
PROBABILITY FIELD SIMULATION..................................................................................80
SEQUENTIAL GAUSSIAN SIMULATION............................................................................8 1
SEQUENTIAL INDICATOR SIMULATION...........................................................................8 1
TRUNCATED GAUSSIAN SIMULATION ............................................................................8 2
TRUNCATED PLURIGAUSSIAN SIMULATION ...................................................................82
TURNING BANDS SIMULATION ......................................................................................82
APPLYING GEOSTATISTICS ....................................................................................8 3
DATA COLLECTION........................................................................................................8 3
DATA, SUMMARY STATISTICS AND HISTOGRAM VISUALIZATION..................................84
SEMIVARIOGRAM VISUALIZATION.................................................................................84
EVALUATION OF MODEL ...............................................................................................8 4

CONCLUSION...............................................................................................................8 7

Geostatistics Without Tears 5

REFERENCES ...............................................................................................................8 9

Geostatistics Without Tears 6

 GEOSTATISTICS WITHOUT
TEARS
A Practical Guide to Surface Interpolation, Geostatistics,
Variograms and Kriging
by Gregg Babish
GIS Supervisor / Data Management
Environment Canada
Illustrations by Mark Gilchrist
Environment Canada

PREFACE
Geostatistics (also known as kriging) was developed for the mining industry
during the 1960s and 1970s to estimate changes in ore grade. The principles of
geostatistics are now applied to many applications that require statistically based
interpolation techniques. Geostatistics provides a data value estimate for
locations that cannot be sampled directly by examining data taken at locations
that can be sampled. If a model of spatial correlation can be established for the
sampled data then this model is used to interpolate data values for those locations
that cannot be sampled directly. The effectiveness of geostatistics depends upon
how well the observation points represent the phenomenon under study as well as
the appropriate selection of the model parameters.

Matheron introduced the term geostatistics in 1962 with this definition:

“Geostatistics is the application of the formalism of random functions to the
reconnaissance and estimation of natural phenomena” (Journel and Huijbregts,
1978). Oliver and Webster (1991) said, “. . . there is to be no geostatistics
without tears”.

The intent of this paper is to provide the reader with an introduction and
understanding of surface interpolation techniques and geostatistics (kriging).
Hopefully no tears will be shed.

Geostatistics Without Tears 7

Geostatistics Without Tears 8
INTRODUCTION
Surface models are used for a variety of purposes including interpolating between
actual data measurements, identifying data anomalies and establishing confidence
intervals around predictions. The principles of surface analysis are applied in a
number of applications such as climatic measurements, contamination plumes, and
population densities. The interpolated surfaces are then used to examine weather
change, contaminant distribution and population change. The accuracy of
subsequent analyses depends directly on the accuracy of the surfaces created.

Data to create these surfaces are usually collected through field work. As a rule,
data collection can only be conducted at a restricted number of point locations
due to limited resources and high costs. In order to generate a continuous surface
some type of interpolation method must be used to estimate data values for those
locations where no samples or measurements were taken.

INTERPOLATION TECHNIQUES
Surface interpolation is the process of estimating a data value for unsampled
sites within an area of existing data observations. The best surface interpolation
method to use will depend on the data, the study objectives, and the ease of
creating a surface versus the accuracy of the surface created.

In order to make spatial predictions there is a set of essential assumptions:

• the measurements taken are precise and reproducible

• the sample measurements are accurate and represent the true value at that
location

• the samples are collected from a physically continuous, homogeneous

population

• the values at unsampled locations are related to values at sampled

locations

All interpolation methods depend on the similarity of nearby sample locations to

derive the surface. Spatial independence means that the location of any data point
is unrelated to the location of any other data point and therefore there is no
spatial relationship between sampling locations. Spatial autocorrelation is the
arrangement of data where the point locations are related to each other. The
concept of spatial autocorrelation is one of the most important aspects of spatial
statistics. Spatial autocorrelation is a normal result of physical and chemical
processes in the environment and environmental parameters will exhibit spatial
autocorrelation at some scale (Englund; Rigsby et al. 2001). Many social
phenomena are spatially autocorrelated (Ned Levine & Associates, 1999).

Geostatistics Without Tears 9

Spatial Interpolation

Deterministic/Stochastic
Deterministic techniques are based on surrounding measurements (mathematical
functions) to calculate the surface. These techniques are based on the measured
values of a parameter at samples near the unmeasured location for which a
prediction is made. For deterministic interpolation there is only one possible
predicted value for a location. Thiessen polygons, IDW and spline interpolation
are deterministic techniques

Stochastic techniques use both mathematical and statistical functions for

prediction. The first step for such methods is to determine the statistical
relationship between samples points. This information is then used to predict
parameter values for unmeasured points. Stochastic interpolation incorporates the
concept of randomness (through autocorrelation). Stochastic techniques are
advantageous because their statistical basis allows for quantification of error.
Polynomial regression and kriging are stochastic interpolation methods. Kriging
attempts to statistically obtain the optimal prediction to provide the Best Linear
Unbiased Estimation (BLUE), specifically when data are sparse.

Global/Local
A global interpolator derives a surface model using all of the available data to
provide predictions for the whole area of interest by applying a single function
across the entire region. The resulting surface gives a best fit for the entire
sample data but may provide a very poor fit at some locations. There are a
number of disadvantages to a global fit procedure. The most obvious is that
global interpolation methods don’t capture short-range variation. A change in one
input value can affect the entire map. Global algorithms tend to produce
smoother surfaces with less abrupt changes.

A local interpolator calculates predictions from the measured points within

neighborhoods or smaller spatial areas within a larger study area to ensure that
interpolated values are determined only by nearby points. Local interpolation applies
an algorithm to a small portion at a time. If the interpolated plane can be bent in one
place, it may be possible to obtain a better overall fit. To allow one bend is the basis
for second-order global interpolation, two bends in the plane would be a third order
and so forth (polynomial trend surface). If a number of planes are used for predicting
many locations in a study area the final surface may fit more accurately. A change in
one input value only affects results within a small area.

IDW and kriging are examples of local interpolation. Kriging interpolation

depends on whether or not there is a trend in the data and probabilistic/stochastic
variation or deviation from a trend (spatially autocorrelated random error) in
addition to uncorrelated random noise.

Geostatistics Without Tears 10

 Exact/Approximate
Exact interpolation honours the data points upon which the interpolation is based
so that the interpolated surface passes through all points whose values are known.
Kriging methods honor the sample data points, but kriging may incorporate a
nugget effect (random noise) and in this case the concept of an exact interpolator
ceases to be appropriate.

Approximate interpolation is used when there is uncertainty in the data values. In

many data sets there are global trends and local variation that produces
uncertainty (error) in the sample values. Approximate interpolation introduces
smoothing so as to reduce the effects of error on the resulting surface.

Gradual/Abrupt
Gradual interpolation produces a surface with gradual changes by applying the
same rules over the entire data source. Gradual interpolation is appropriate for
interpolating data with low local variability.

Abrupt interpolation can involve the application of different rules at different

points by including ‘barriers’ in the interpolation process. Semipermeable
barriers, such as weather fronts, will produce quickly changing but continuous
values. Impermeable barriers, such as a geologic fault, will produce abrupt
changes. Abrupt interpolation produces surfaces with a stepped appearance and is
appropriate for interpolating data of high local variability or data with
discontinuities.

Climate Data
Collection of meteorological data requires specialized instrumentation collecting
data over many years and is subject to siting, data collection errors and biases.
Climate data is available only at a relatively small number of observed points.
Given a set of meteorological data, researchers are confronted with a variety of
interpolation methods to estimate meteorological variables at unsampled
locations. Depending on the spatial attributes of the data the accuracy of the
interpolated surface can vary widely among different methods (MacEachren and
Davidson, 1987; and Rhind, 1975).

Climate interpolation models provide few estimates of error because these estimates
rely on the same assumptions used in the interpolation process itself, and are
therefore, not independent or reliable (Daly, 2006). Experts are the most useful
source of verification because they can integrate information from disparate sources
to give justifiable feedback on how well the model results reproduce their
knowledge from a variety of perspectives. This feedback can be in the form of
evaluation of the spatial patterns and magnitudes of the mapped climate values, as
well as insight into station data quality issues (Daly and Johnson, 1999).

The interpolation of irregular meteorological-climatological point data onto a

uniform grid has been the focus of research and a number of methods have been
proposed ranging from simple Thiessen and distance weighting methods
(Shepard, 1968; Willmott et al., 1985), to geostatistical methods such as kriging

Geostatistics Without Tears 11

 (Phillips et al., 1992), splines (Hutchinson, 1995) and locally-varying regression
techniques such as PRISM (Daly et al., 1994).

Kriging has been used for interpolation of precipitation data (Chua and Bras,
1982; Dingman et al., 1988; Phillips et al., 1992). Meteorologists sometimes
consider that a potential drawback of kriging is that kriging implicitly relies on
the data to directly represent the spatial variability of the actual precipitation
field. If the data are not representative (as is often the case in complex terrain,
such as mountainous regions), the accuracy of the resulting interpolated field will
be in question (Daly et al. 1993).

The choice of spatial interpolator is especially important in mountainous regions

where data collections are sparse and variables may change over short spatial
scales (Collins and Bolstad, 2006). For many climate applications it is important
that elevation is included as a covariate or independent variable because the
climate variable is dependent on elevation in some manner (Willmott and
Mastsuura, 1995; Briggs and Cogley, 1996; New and Hulme, 1997). Goovaerts
(2000) confirmed the finding of Creutin and Obled (1982) that for low-density
networks of rain gauges geostatistical interpolation outperforms techniques that
ignore the pattern of spatial dependence which is usually observed for rainfall
data. Prediction can be further improved if correlated secondary information,
such as a DEM is taken into account.

Ordinary kriging has been applied for developing contour maps of design storm
depth using intensity-duration-frequency (IDF) data (Cheng, et al., 2003). The
design storm, a crucial element in urban design and hydrological modeling, is a
hypothetical storm of specific storm duration, and recurrence interval. Variogram
parameters, the sill and range are functions of the recurrence interval and the
storm duration. The sill accounts for the time non-stationarity of the rainfall field,
and design storms with higher total rainfall depths have higher sill values.

Some of the variations of meteorological interpolations are discussed below.

ANUSPLIN
ANUSPLIN software developed in the late 1980’s by Michael Hutchinson
(Hutchinson, 1991 and 1994), Australian National University (ANU), Centre for
Resource and Environmental Studies (CRES), Canberra is based on the srcinal
thin-plate surface fitting technique (Wahba, 1979 and 1990) for mapping climate
variables, especially for the interpolation of rainfall amounts (Hutchinson, 1995;
1998a; 1998b). ANUSPLIN fits thin-plate splines (usually second- or third-order
polynomials) through station data in three dimensions: latitude, longitude, and
elevation (Hutchinson, 1995). Booth and Jones (1996) conclude that ANUSPLIN

provides a powerful
used to develop set ofconsistent
globally programs30”
for climate
climaticsurfaces
analysis.covering
ANUSPLINmost has been
areas on
the Earth’s surface (Hijmans et al. 2004 a,b).

ANUSPLIN was specifically developed for interpolating climate data and is

made up of nine programs that incorporate additional dependencies (the concept of
“surface independent variables”) such as elevation (Kesterven and Hutchinson,

Geostatistics Without Tears 12

 1996). Despite having the capability of including many more dependencies
ANUSPLIN routinely uses only threevariables in practical applications. Being able
to accommodate several covariates to aid interpolation is a significant advantage. A
disadvantage of the method is ANUSPLIN’s reliance on a dense network of
stations covering all aspects of the topography being mapped. This is particularly
problematic for the interpolation of rainfall, which is often underestimated at high
elevations (Tait and Zheng, 2005). Since ANUSPLIN places a greater reliance a
DEM than other methods ANUSPLIN is more likely to provide more robust
surfaces, but that robustness is dependent on the accuracy of the DEM, and may
results in problems where relief is subtle (Chapman et al., 2005). Because a spline
is by definition smoothly varying, this approach has difficulty simulating sharply
varying climate transitions, which are characteristic of temperature inversions, rain
shadows and coastal effect (Daly, 2006).

ANUSPLIN produces results similar to kriging (Delfiner and Delhomme, 1975;

Hutchinson, 1991b and 1993; Hutchinson and Gessler, 1994; Wahba and
Wendelberger, 1980; Wahba, 1990; and Cressie, 2003). Unlike kriging the thin-
plate spline method does not require development of a covariance function
(variogram). Instead the degree of smoothing is optimized objectively by
minimizing the predictive error of the fitted function as measured by generalized
cross validation (Milewska et al., 2005; Tait and Zheng, 2005).

Centro Internacional de Agricultura Tropical

The Centro Internacional de Agricultura Tropical (the International Centre for
Tropical Agriculture) (CIAT, http://www.ciat.cgiar.org/
), Cali, Columbia, method

uses a simple
between interpolation
the station and thealgorithm based
interpolated on of
point thethe
inverse square
nearest five of the distance
stations (Jones
et al. 1990, Jones, 1995).

The major difference between CIAT and ANUSPLIN is that the CIAT method
uses a standard lapse rate applied over the whole dataset whereas ANUSPLIN
uses a 3-dimensional spline algorithm to determine a local lapse rate from the
data. The CIAT method introduces error when the local lapse rate deviates from
the standard lapse rate function. ANUSPLIN suffers when there are erroneous
data or insufficient data range in the local area, resulting in spurious correction
for elevation (Booth and Jones, 1996). CIAT has the advantage of speed and ease
of use for large data sets where computational capacity is limited. The influence
of a bad data point can be significant and can cause significant circling in the
resultant surface. Because CIAT uses only five data points, it relies less on the
underlying DEM than other methods (Booth and Jones, 1996).

Climatologically Aided Interpolation

Climatologically Aided Interpolation (CAI) is a hybrid approach (incorporating
elements of temporal substitution with distance weighted interpolation) that uses
existing spatial climate data to improve the interpolation of another set of data
(Willmott and Robeson, 1995; Robeson and Janis, 1998). CAI produces low
validation errors, and its accuracy is attributed in part to the incorporation of
terrain effects provided by the high resolution climatology (Thornton, et. al.,

Geostatistics Without Tears 13

 1997). CAI relies on the assumption that local spatial patterns of the element
being interpolated closely resemble those of the existing climate grid (called the
background or predictor grid). This method is useful for interpolating climate
variables and time periods for which station data may be relatively sparse or
intermittent.

Use of CAI fall into two broad categories (Daly, 2006):

1. using a long-term mean grid of a climate element to aid the interpolation of

the same element over a different (usually shorter) averaging period; and

2. using a grid of a climate element to aid the interpolation of a different,

but related, climate element (e.g. interpolating growing degree days
using mean temperature as the predictor grid).

Daymet
Daymet (http://www.daymet.org/) is similar to PRISM in that both methods use
local regression techniques, but Daymet and PRISM are different in their
formulation. Daymet was developed at the University of Montana, Numerical
Terradynamic Simulation Group (NTSG) to fulfill the need for fine resolution, daily
meteorological and climatological data necessary for plant growth model inputs.
Using a digital elevation model and daily observations of minimum/maximum
temperatures and precipitation from ground-based meteorological stations, an 18
year daily data set (1980-1997) of temperature, precipitation, humidity and radiation
have been produced as a continuous surface at a 1 km resolution for the United
States (http://www.daymet.org/dataSelection.jsp
).

Daymet develops local linear regressions between climate and elevation for each
grid cell on a digital elevation model using data from surrounding stations. Each
station is weighted in the regression function by its distance from the target grid
cell. This method takes into account the elevation variation of climate and a
simple station distance weighting algorithm. Daymet does not have the ability to
simulate nonmonotonic relationships between climate and elevation, such as
temperature inversions, and does not explicitly account for terrain-induced
climatic transitions or coastal effects (Daly, 2006).

Gradient plus Inverse Distance Square

Gradient plus inverse distance square (GIDS) method has been used in various
parts of the world to produce useable surfaces for meteorological parameters.
GIDS combines multiple linear regression with distance weighting.

A comparison of GIDS and ANUSPLIN concluded that ANUSPLIN provides

generally superior results. Both subjective assessment and statistical analysis
showed that ANUSPLINE is generally more accurate in predicting climate
variables. In addition ANUSPLIN produces better smoothing and better gradients
at high elevations and in areas where climate station coverage was poor (Price et
al., 2000; Chapman, 2003). GIDS is easy to implement and understand which can
provide a useful baseline to compare with more sophisticated methods (Price et
al., 2000).

Geostatistics Without Tears 14

Optimal Interpolation
Geostatistical techniques were srcinally developed by Soviet scientists for
meteorological data predictions. The first book with complete explanations about
simple and ordinary kriging and cokriging techniques was published in Leningrad
(Gandin, 1963). According to this book, the srcinal name of the technique is
objective analysis. In atmospheric and oceanographic science this technique is
known as optimal interpolation (OI). In spatial statistics this technique is known as
simple kriging (Matheron, 1963). The multivariate analysis of meteorological data
using an optimal interpolation method is presented in Gandin, 1963; Gandin and

Kagan, 1974;
al., 1982; Bretherton
Hiller and Käse,et1983;
al., 1976; Bretherton
Swanson, et. al., and Williams,
2001; 1980; Sarmiento et
AWI, 2006.

The advantage of optimal interpolation is the simplicity of implementation and

its relatively small cost if the right assumptions can be made on the selection of
observation data. A drawback of optimal interpolation is that spurious noise is
produced in the analysis fields because different sets of observations (and
possibly different background error models) are used on different parts of the
model state (Bouttier and Courtier, 1999). It is also impossible to guarantee the
coherence between small and large scales of the analysis (Lorenc, 1981).

Parameter-evaluation Regressions on Independent Slopes Model

Parameter-evaluation Regressions on Independent Slopes Model (PRISM),
(http://www.ocs.orst.edu/prism/) is an expert system developed at Oregon State
University in the early 1990’s for orographic precipitation estimates (Daly,
1996). PRISM uses point data and DEM data to generate gridded estimates of
climate parameters (Daly et al., 1994; Taylor et al., 1997). PRISM is suited for
use in mountainous regions because data relationships can be extrapolated
beyond the lowest and highest station elevation. The use of PRISM has been
extended to map temperature, snowfall, weather generator statistics, and more.
Climate layers for the United States, western Canada and other a few other
countries, have been developed using PRISM and are available on-line
http://www.climatesource.com/ (Daly et al., 1994 and Gibson et al. 2004).

PRISM develops local climate-elevation regression functions for each DEM grid
cell, but calculates station weights on the basis of an extensive spatial climate
knowledge base that assesses each station’s physiographic similarity to the target
grid cell (Daly et al., 2002, 2003). The knowledge base and resulting station
weighting functions currently account for spatial variations in climate caused by
elevation, terrain orientation, effectiveness of terrain as a barrier to flow, coastal
proximity, moisture availability, a two-layer atmosphere (to handle inversions),
and topographic position (valley, midslope, ridge).

While PRISM accounts for more spatial climate factors than other methods, it also
requires more effort, expertise and supporting data to take advantage of its full
capability (Daly 2006). The effects of other variables, such as slope, aspect, coastal
proximity, the influence of the boundary layer, etc. are controlled by the use of
special algorithms, parameters and weights. Some of these values are set as defaults
based on the general knowledge of physiographic and atmospheric

Geostatistics Without Tears 15

 processes; some are inferred by the model from station data; other are assigned
manually by an expert climatologist through the user interface (Daly et al., 2002).

Although PRISM incorporates a number of spatial interpolation quality control

measures in similar ways to ANUSPLIN, the PRISM and ANUSPLIN methods
are fundamentally different (Simpson, et al., 2005). PRISM uses a two-layer
atmosphere to model the effects of atmospheric inversions on surface temperature
whereas ANUSPLIN cannot. PRISM analysis typically extends beyond the
coastline to include near-shore areas, whereas ANUSPLINE analysis is largely
restricted to land areas.

Thin-Plate Spline
The thin-plate spline (TPS) method refers to a physical analogy involving the
bending of a thin sheet of metal constrained not to move at the sample points and
is free from any external force. The resulting shape of the plate represents the
natural bend of the metal to accommodate all of the fixed sample points. The use
of thin-plate splines has been found useful for a number of applications including
DEM construction, fingerprint modeling, image manipulation, geophysical
applications, medical research, and rubber sheeting. Thin-plate splines have
found application in meteorology (Hutchinson and Bischof, 1983; Hutchinson et
al., 1984; Kalma et al., 1988; Hutchinson, 1990; Mitchell, 1991; Zheng and
Basher, 1995; Chapman et al., 2005).

The local thin-plate spline method is an extension of the thin-plate Spline

interpolation technique, and is recommended for use with a large number of grid

points (>200). the

interpolation, Thelocal
only version
difference is that
takes instead of of
a maximum using
10 all the grid
closest points
points for
to the
sample point and fits a spline surface through them. The local spline surface is
then used to determine the sample value (RocScience, 2006)

The key features of the thin-plate smoothing spline analysis are robustness,
simplicity and the advantage of being able to map sparse and noisy data. One of
the main shortcomings of the thin-plate spline method has been the limited range
of practical options for user control of the fit. The thin-plate spline method can be
problematic if sample points are sparse. Small changes in sample position could
produce global changes in the warping transformation (Glasbey and Mardia,
1998). The thin-plate spline is equivalent to kriging with a specific spatial
covariance structure (Kent and Mardia, 1994).

Digital Elevation Modeling

There areNetwork
Irregular two forms of Digital
(TIN) Elevation
with their Model (DEM):
own advantages Grid and Triangular
and disadvantages (Cadell,
2002). A number of algorithms and computer programs have been developed to
create a DEM.

ESRI’s ARCTIN method makes a Triangular Irregular Network (TIN) and then
converts the TIN to a DEM. ANUDEM and ESRIs ArcInfo TOPOGRID are two
common spline-based methods of obtaining a reasonable interpolation between

Geostatistics Without Tears 16

spot elevations, and elevation contours while maintaining stream connectivity.
ANUDEM was developed at the Centre for Resource and Environmental Studies
(CRES) in Canberra (Hutchinson, 1988, 1989). A version of ANUDEM is
included in ArcInfo as TOPOGRID.

ANUDEM preserves ridge and valley lines as well as maintaining drainage

(Hutchinson, 2006). Input data to ANUDEM may include point elevations,
elevation contours, streamlines, sink data points, cliff lines, boundary polygons,
lake boundaries and data mask polygons. TOPOGRID is based on an older
version of ANUDEM. ANUDEM/TOPOGRID use iterative interpolation starting
from a very coarse model to end up at a user-specified grid spacing (Jaakkola and
Oksanen, 2000). The strength of the ANUDEM method over most other methods
is that ANUDEM imposes a global drainage condition through an approach
known as drainage enforcement, to produce elevation models that represent more
closely the actual terrain surface and which contain fewer artifacts than those
produced with more general-purpose surface interpolation routines (USGS 2003).

Care must be taken with spline-based techniques, such as ANUDEM or

TOPOGRID because certain topography (such as steep slope close to flat terrain)
can cause undershoot and overshoot errors. Biasing towards contour elevations
can occur and is sometimes quite noticeable.

When there is interest in maintaining an assumed network of hydrology then

methods such as ANUDEM/TOPOGRID probably give a more acceptable result
than the TIN method. This advantage tends to be more obvious in low-relief
topography than in topography where the hydrology is constrained by strong
slopes. If maintaining the hydrological networks isn’t important then
ANUDEM/TOPOGRID might not be the best answer. ANUDEM can give an
inferior results compared to a DEM generated when the hydrological constraints
are withdrawn and a linear interpolator is used. Then again, a simple linear terrain
interpolation might not be good for some visualization work since it introduces
visual artifacts. Analysis of DEMs suggests that the quality of the input data from
which the DEM is generated has a more significant effect on DEM quality than do
the algorithms employed by different methods (Barringer and Lilburne, 1997).
Techniques have been developed that use the results derived from Radarsat
imagery in conjunction with ANUDEM to provide a DEM of much greater
accuracy. This approach has been valuable in areas with little terrain variability,
and has been used to create a DEM of the Antarctic (Liu et al. 2001).

Triangular models have been used in terrain modeling since the 1970s, and are
sometimes used to create Digital Terrain Models (DTMs). Commercial systems
using TIN began to appear in the 1980's as contouring packages. Due to
limitations of computers and the complexity of TIN data structures, gridded

models have
effectively overshadowed
divided triangular
into triangles models.
with plane Certain
facets. Thistypes of terrain true
is particularly are very
with
fluvially eroded landscapes. However, other landscapes, such as glaciated ones,
are not well represented by flat triangles. Triangles work best in areas with sharp
breaks in slope where TIN edges can be aligned with breaks such as mountainous
terrain, but triangular models aren’t as suitable for representing prairie
landscapes (Goodchild, 1999).

Geostatistics Without Tears 17

Fourier Series
The Fourier series approximates a surface by overlaying a series of sine and cosine
waves. The Fourier series is best suited for data sets that exhibit marked periodicity,
such as ocean waves. Fourier series is a high-order surface representation with
triangular mesh application (Bruno and Pohlman, 2002, 2003).

Geostatistics
Interpolation methods such as spline, IDW, triangulation, trend surfaces and
Fourier series are based on mathematical models (deterministic). These methods
assume that the sampled data has no errors, which is often an incorrect
assumption since errors can be minimized but not eliminated. The best
representative map created with these techniques may look like they model the
spatial process but this model may not provide a correct representation of the
spatial process (Krivoruchko, 1998).

Geostatistics is based on random processes with dependence. A variable that

takes on values according to its spatial location (or time) is known as a
regionalized variable. Regionalized variables describe phenomena with
geographical distribution, such as disease incidence, geophysical measurements,
insect counts, soil/water contamination, and surface elevation. Regionalized
variables recognize the fact that properties measured in space (or time) follow a
pattern that cannot be described by a mathematical function. In the spatial or
temporal context such dependence is called autocorrelation. Statistical
dependence provides valuable information for prediction. Geostatistical
estimation is a two stage process:

i. study the data to establish the predictability of values from place to

place (uncover the dependency rules)

ii. interpolate values at those locations which have not been sampled
based on the degree of spatial pattern (spatial autocorrelation).
Kriging is the process of prediction of the unknown values

The results from kriging are generally of higher quality and have a more realistic
look compared to techniques such as IDW and triangulation. The “bull’s eye”
effect is avoided; a measure of error or uncertainty is provided; confidence
placed in the estimates can be calculated; extreme weighting of values caused by
irregular distribution of sample points is reduced; redundant data (clustering) and
anisotropies is accounted for; and kriging can be used to determine where more

data is needed if future sampling is planned.

The major disadvantage of the kriging process is that the construction of the
model and evaluation of the surface interpolation results can be time consuming.
Kriging is not a suitable method for data sets that have spikes, abrupt changes or
break lines.

Geostatistics Without Tears 18

 Inverse Distance Weighted
Inverse Distance Weighted (IDW) (also known as
Inverse Distance to a Power) assumes that each
coordinate (X,Y) and surface value (Z) has a local
influence that diminishes with distance. Points closer
1 to the processing cell are given more weight than
t
g D those farther away. The power parameter in the IDW
e 1 interpolation controls the significance of the
2
D surrounding points on the interpolated value. A
14 higher power results in less influence from distant
D
1 points.
8
D
Simplicity, speed of calculation, programming ease,
Distance and reasonable results for certain types of data are
some of the advantages associated with IDW
interpolation. The IDW method is based on the assumptions that the data values
form a continuous surface across the whole area, and that the data is strongly
correlated with distance. One potential advantage of IDW is the ability to control
the influence of distance. IDW has been used to interpolate surfaces for consumer
purchasing (distant locations have less influence because people are more likely to
shop closer to home) and IDW is used for noise analysis since noise falls off very
predictably with distance.

Although distance-weighted methods are one of the more commonly used

interpolation approaches, they are far from ideal (Goodchild, 1999; Clark and
Harper, 2000; Lembo, 2005). IDW doesn’t work as well with phenomena that are
dependent on complex variables because IDW can only account for the effects of
distance. IDW works best with sample points that are dense and evenly
distributed. IDW is easily affected by the uneven distribution of data points since
an equal weight is assigned to each of the data points even if the data points are
clustered. IDW is an averaging (smoothing) technique and cannot interpolate
above or below the surrounding data which tends to generate flat areas. IDW has
a tendency to generate patterns of concentric contours around the actual data
points (“bull’s eyes”) and will average out trends and emphasize anomalies
(outliers). For an elevation surface, this has the effect of flattening peaks and
valleys (unless their high and low points are part of the sample). Smoother IDW
surfaces can be created by decreasing the power, increasing the number of
sample points used, or increasing the search radius – the opposite is done to
create a more locally influenced surface (Zamkotowicz, 2005).

Shepard’s Method (Shepard, 1968) is a variation on the inverse power with two
different weighting functions (an inverse distance weighted least squares method,
and the inverse distance to a power interpolator). Shepard’s Method, while
similar to IDW, eliminates or reduces “bull’s eye” patterns, especially when a
smoothing factor is used. The Shepard’s Method can extrapolate values beyond
the data’s range.

Geostatistics Without Tears 19

Nearest Neighbour
The nearest neighbour technique is a very simple method in which grid point
value are assigned the value at the station nearest them. The nearest neighbour
method is similar to IDW, except that the grid values are not a weighted average
of the values from nearby sampling locations. There are obvious problems with
this method when there are no nearby stations. This method is the least accurate
method for estimating values away from their sampling location.

For climate studies, both the nearest neighbor and IDW can be significantly
improved by including topographical effects. For example, an environmental
lapse rate (–6.5 °C/km) can be applied to temperature values at the stations
before the interpolation is performed. Then, after the interpolation, the grid
values can be transformed back by re-applying the lapse rate. This normalizing
technique standardizes the temperature data with respect to elevation, resulting in
a more accurate interpolation of ‘like’ data (Tait and Zheng, 2005).

Splines
Splines were srcinally flexible pieces of wood or metal used by draftsmen to fit
curvilinearly smooth shapes when the mathematics and/or tools were not
available to create the shapes directly (airplane wings/fuselages, boat hulls,
railway tracks). Spline interpolation was later developed (early 1960’s) to help in
the computer aided manufacturing of car bodies, for computer graphics and to fit
isopleths (contours) to continuous data.

Splines are polynomials (usually cubics) which are fitted to coordinate (X,Y)
and surface values (Z) and forced to be smooth and continuous at the data joining
points called 'knots' (Smith, et al., 2004). To visualize the spline in action,
imagine a rubber sheet being used to interpolate the surface. The rubber sheet
must be bent and stretched so that it passes through all of the measured values.
The rubber sheet can either be forced to form nice curves (regularized spline
interpolation) or the edges of the rubber sheet can be tightly controlled (tension
spline). The spline curve goes through data points exactly and assumes a curved
shape elsewhere. A spline interpolation demo can be seen at
http://www.math.ucla.edu/~baker/java/hoefer/Spline.htm.

Splines are easy to compute mathematically and are useful in fitting a smooth
line or surface to irregularly spaced data points while still retaining localized
features. Spline functions produce good results for gently varying surfaces, such
as elevation, water table heights or pollution concentrations.

The spline function is inappropriate if there are large changes in the surface
within short horizontal distances and/or when the sample data are prone to error
or uncertainty. With this type of data the spline method can wildly overshoot
estimated values and closely spaced data points can develop embolisms.
Extrapolating beyond the edges of the data domain often yields erratic results
because there is no information beyond the data to constrain the extrapolation
and splines will grow to large values (positive or negative). Splines tend to
emphasize trend rather than anomalies (outliers).

Geostatistics Without Tears 20

Trend Surface Analysis
Statistically based surface interpolation techniques assume that the sample data
comes from a simple distributed in space distribution which means that the irregular
spatial pattern cannot be describe by a mathematical function. Statistical techniques
cannot handle the idea of a trend in the data. If there is a trend then the trend must
be modeled so that the situation can be stabilized. With trend surface analysis drift
can be analyzed for and subtracted out of the data much the same way an offset can
be subtracted out of a data set (Glover, 1998, 2000).

A trend surface can be used to filter out large scale spatial trends in order to focus
on the smaller scale variation (residuals) (Goodman, 1973, 1999, and
Klinkenberg, 2002). Variability in spatial data is generally the product of two
effects: the first the result of broad-scale regional changes in the value of the
phenomena, the second the result of smaller-scale ‘local’ variations (Goodman,
1973, 1999; Davis, 1973). Trend surfaces typically fit one of three
mathematically defined ideal surface models: linear (a constant dip in some
single direction); quadratic (a bowl or dome shape, anticline or syncline); or
cubic (saddle point, perhaps large scale folding).

Trend surface interpolation is highly sensitive to outliers (extreme values) and

uneven distribution of the sample data points. The problem is further complicated
by the fact that some data points are more informative than others. Polynomial
trend surfaces have a tendency to estimate higher or lower values in areas where
there are no data points, such as along the edges of maps. A polynomial model
produces a rounded surface, which is rarely the case in many human and physical
applications.

Triangulation
When a set of coordinate (X,Y) and surface value (Z) data points are connected
in a triangular network of varying size the resulting lattice work is irregular and
is called a Triangular Irregular Network (TIN) or irregular triangular mesh. The
TIN partitions a surface into a set of contiguous, non-overlapping, triangles. The
TIN is a vector data structure and is made from a set of points called mass points.
Triangular structures were invented in the early 1970’s because computers were
limited and grids weren’t as inefficient (Thurston, 2003).

Triangulation is a fast, flexible and popular interpolation method that offers

several advantages. When a surface has significant relief features, triangulation
generates more accurate surface representation by maintaining break line features
such as stream channels, shorelines, and slope features such as ridges.
Triangulation can represent a surface with fewer data points than grid based
methods. Triangulation works best when data points are evenly distributed over
an area. The triangulation method honors the srcinal data points exactly.

The main disadvantage of triangulation is that the generated surfaces are not
smooth and may have a jagged appearance. The size and shape of the areas
depend on the sample layout, which can lead to irregularly shaped polygons,
especially on the edges. Sparse data sets result in distinct triangular facets.

Geostatistics Without Tears 21

Triangulation is not suitable for extrapolation beyond the observed data points.
Analysis involving comparison with other layers of data is difficult. Estimation
error can’t be determined as the value assigned to each cell is based on only one
value. Computation of a value at an unsampled point is a function of the polygon
it lies within, rather than the values of the points closest to it.

The Voronoi diagram, sometimes named Dirichlet tessellation breaks an area into
pieces or tiles. Lines that bisect the lines between a center point and its
surrounding points define a Voronoi polygon. The bisecting lines and the
connection lines are perpendicular to each other. Voronoi polygons are created so
that every location within a polygon is closer to the sample point in that polygon
than any other sample point. Interest in the Voronoi diagram srcinates from the
climatologist A.H. Thiessen’s use of Thiessen polygons to define regions that
surround unevenly distributed weather stations.

Using the Voronoi diagram as a basis, Delaunay (named after B. Delaunay)

triangulation is constructed by drawing lines between the points in adjacent
polygons. Delaunay triangulation is a common and preferred technique since it
provides nearly unique and optimal triangulation (Watson and Philip 1984; Tsai
1993).

Geostatistics Without Tears 22

HISTORY OF KRIGING

Given that concentrations of high-grade ore are easier and more profitable to
mine, while regions of low-grade ore should be ignored, the estimation of
recoverable ore is very important to the mining industry because local variability
can make or break a mining venture. The first steps to resolve this problem were
taken in the early 1950’s in South Africa with the work of the mining engineer
Danie Krige (Krige, 1951; Krige et al. 1989) and the statistician Herbert Sichel
(1915-1995) working on the Witwatersrand goldfields.

The book “Historical Overview of the Witwatersrand Goldfields” (Handley,

2004) indicates that the discovery of the Witwatersrand goldfields in 1886
probably exerted a greater influence on the course of South African history than
any other event. The Witwatersrand goldfield of South Africa has dominated
world gold production for the last century and has been the source of almost one
third of all the gold mined.

Krige and Sichel studied the best ways of predicting the value of a block to be
mined using only samples from the faces already worked. Krige assumed that there
was a relationship between neighboring blocks of ground in that blocks close to the
stope (area of a mine from which ore is or has been extracted) would have a
different and stronger relationship than those being estimated that are further away.
Which is the basis for the assumption: “two blocks of ground will have a
covariance which depends on the distance between them and on their relative

locations.” Krige
necessary, To simplify
devised athe calculations
template andbe cut
that could laid down the
over the amount
mined areas.of work

Georges Matheron (1930-2000) an engineer with Ecoles des Mines, Centre de

Morphologie Mathématique, Fontainebleaue, France, became aware of Krige’s
approach to ore reserves calculation while working with the French Geological
Survey in Algeria and France from 1954 to 1963 (Rivoirard, 2000). Matheron
adopted the pioneering work being done in South African and formalized the
major concepts of the theory that he named geostatistics. In 1955 Matheron
collaborated in a paper (Duval, et. al., 1955) that presented the work of Krige.
Matheron’s early work culminated in two books (Matheron, 1962 and Matheron,
1965). Out of Matheron’s research came a spatial interpolation method that he
called kriging in honor of Dr. Krige.

The word “krigeage” appears in Matheron's srcinal work (in French) and the
English translation was kriging. Matheron’s first English paper appeared in 1963
(Matheron, 1963a). The two most common ways to pronounce kriging in English
are “kree-ging” or “kree-jing”. Danie Krige pronounced kriging with a hard 'g'
(as in 'grand' or 'organize'), and not with the guttural Afrikaans 'g' (which is how
he pronounced his name).

While Matheron was developing his theory of prediction in France, the

meteorologist Lev S. Gandin (1921-1997) in the Soviet Union was doing
remarkably similar work in meteorology and atmospheric sciences. Gandin first
started publishing in 1959 (Krivoruchko, 2000, Armstrong and Galli, 2001), and

Geostatistics Without Tears 23

Gandin’s first book complete with explanations about simple and ordinary
kriging and cokriging techniques was published in Leningrad (Gandin, 1963)
where the geostatistical technique was known as objective analysis or optimal
interpolation. This work did not appear in English until much later when Gandin
emigrated to Isreal (Myers, 1999).

B. Matern working in Sweden developed essentially a parallel theory to

Matheron as a forestry application. Matern’s work appeared in Swedish in 1960
and was not translated into English until 1986 (Myers, 1999). Matheron who is
considered to have laid the foundation of geostatistics acknowledges that his
work is similar to or duplicates the work of Matern and Gandin in some of his
writings. Matheron knows Russian, was a communist supporter, studied statistics
in the Soviet Union and was aware of Gandin’s work.

From 1964 to 1968, Georges Matheron turned his attention to the mathematical
characterization of geometric shapes and in collaboration with Jean Serra, created
the discipline of "Mathematical Morphology" which has since become an
essential branch of the field of image analysis (Rivoirard, 2000). In 1986 the
Centre de Morphologie Mathématique became two programs, one on
mathematical morphology and one on geostatistics (Centre de Geostatistique,
2000). Two of Matheron’s first students (Journel and David) would start new
centers of teaching and research in the USA and Canada.

In 1978 Shell Oil and the Bureau de Recherche Geologie Mathematique

cooperated to develop a commercial software package called BLUEPACK.
BLUE stands for Best Linear Unbiased Estimator. In 1980 the software
MAGMA brought together BLUEPACK and the complete geostatistical library
of the Centre de Géostatistique. Among the first users of MAGMA were Total,
Exxon, Shell, British Petroleum, Agip, Amoco, and Gaz de France.

In 1993 ISATIS was released which offered in one package all the techniques
previously available in MAGMA. ISATIS (http://www.geovariances.com/) is the
result of 40 years of experience in industrial applications and applied research in
geostatistics. Today, ISATIS is widely used the world over by more than 250
private oil & gas companies, consultant teams, mining corporations and
environmental agencies.

In the mid 1980’s the Environmental Protection Agency (EPA) commissioned a

geostatistical software package, GEO-EAS, which was subsequently released in
the public domain. GEO-EAS was a DOS program but included a menu system
that made it fairly easy to use and was fairly inexpensive. The EPA did not
continue to support the software and it has not been updated for a number of
years. Go to http://www.epa.gov/ada/csmos/models/geoeas.html to find version
1.2.1 – April 1989.

In 1992 Andre Journel (Stanford University) and Clayton Deutsch published

GSLIB (http://www.gslib.com/). GSLIB is an acronym for Geostatistical
Software LIBrary. GSLIB is an extensive set of geostatistical programs and a
user manual (second edition completed in 1997). The code is available on the
website. GSLIB is compiled for a variety of platforms.

Geostatistics Without Tears 24

In 1996, Yvan Pannatier published VARIOWIN . Version 2.1 of the VARIOWIN
(http://www-sst.unil.ch/research/variowin/) software was developed as part of a
Ph.D. thesis that was presented on October 9, 1995 at the Institute of Mineralogy,
University of Lausanne, Switzerland. VARIOWIN is a Microsoft Windows
version of two of the components of GEO-EAS. VARIOWIN allows for much
larger data sets than GEO-EAS and also allows for interactive variogram
modeling. VARIOWIN 2.21 is still available.

Gstat is an open source computer code for multivariable geosatistical modeling,

prediction and simulation. Gstat has been around from 1996 under the GNU General
Public License (GPL) and is available from http://www.gstat.org/index.html. In the
srcinal form, gstat is a stand-alone executable, interfaced to various GIS. Gstat was
not initially written for teaching purposes, but for research purposes, emphasizing
flexibility, scalability and portability. As of 2003, the gstat functionality is also
available as an S extension, either as R package or S-Plus library. Current
development mainly focuses on the S extension. The gstat package provides
multivariable geostatistical modeling, prediction and simulation, as well as several
visualization functions.

GEMS (Geostatistical Earth Modeling Software) was designed at Stanford

University (http://sgems.sourceforge.net/) with two aims in mind: to provide user
friendly software that offers a large range of geostatistics tools complete with the
ability to visualize data and results in an interactive 3D environment; and to
design software whose functionalities could conveniently be augmented through
a system of plug-ins.

Many applications of geostatistics continue to appear in: agroforestry, agronomy,

air and water pollution, aerial distribution of acid rain and aerial contaminants,
atmospheric sciences, disease outbreaks, entomology, environmental sciences,
monitoring and assessment, epidemiology, fishery, forestry, geography, global
change, meteorology, migratory bird population estimates, mining,
oceanography, petroleum, plant pathology, surface hydrology, radioecology and
more.

Geostatistics Without Tears 25

Geostatistics Without Tears 26
DATAANALYSIS
Geographic Information Systems (GIS) includes tools to explore and visualize
data to identify unusual data values or errors, detect patterns in data and to
formulate hypothesis from data. Unfortunately this also inspires users to draw
conclusions visually from maps and this often leads to faulty decision making
(Krivoruchko, 2002b). Before starting any surface interpolation project it is
important to do summary and exploratory data analysis to understand the
classical statistics and spatial correlation of the data. Luc Anselin (Jacquez, 2005)
defined exploratory spatial data analysis as "techniques to describe and visualize
spatial distributions, identify atypical locations (spatial outliers), discover
patterns of spatial association (spatial clusters) and suggest different spatial
regimes and other forms of spatial non-stationarity. . . "

The first step is to verify three data features: dependency, stationarity and
distribution (Krivoruchko, 2002a). If data are independent it makes little sense to
analyze them geostatistically. Autocorrelation assumes stationarity, meaning that the
spatial structure of the variable is consistent over the entire domain of the data set. If
the data are not stationary, they need to be made so, usually by data detrending and
data transformation. Geostatistics works best when input data are Gaussian (normal).
If not, the data need to be made to be close to Gaussian distribution.

SUMMARY DATAANALYSIS

Summary Statistics
Summary statistics includes four groups:

• measures of shape –histogram (frequency and cumulative distribution),

coefficient of skewness, kurtosis, normal probability plot, and quantile-
quantile plot

• measures of location – number of samples, minimum and maximum

values, mean, median, mode and quantiles

• measures of spread – variance and

Untransformed Transformed standard deviation and the coefficient
mean 22.383 2.9600 of variation

std deviation 10.822 0.6650 • measures of correlation– coefficient of

sample variance 117.112 0.4423 correlation, coefficient of determination,
minimum value 0.00 0.000 and standard error of estimate
maximum value 55.10 4.009
n ( n missing ) 36 (0) 36 (0)
frequency distibution

skewness ( se ) 0.73 (0.39) -2.39 (0.39)

kurtosis ( se ) 0.89 (0.77) 9.03 (0.77)
Geostatistics Without Tears 27
 Measures of Shape

Frequency Distribution
The histogram is the graphical version of a table
Histogram which shows what proportion of cases fall into
each of several or many specified categories. The
y
c
100 100.00% frequency distribution summarizes discrete data
n Frequency
e
u
by counting the number of observations falling
q 50 50.00% Cumulative %
e
r
into each category. The number associated with
F 0 .00% each category is called the frequency and the
20 25 30 35 40 45 50
collection of frequencies over all categories gives
Value the frequency distribution of that variable.

Cumulative Distribution
The cumulative distribution determines the number of observations that lie above
(or below) a particular value in a data set. The normal distribution produces an s-
shaped graph when the frequency value is plotted against cumulative proportion.

Skewness
Skewness measures the degree to which a distribution is asymmetric. Normal
distributions (symmetrical distribution) will have a skewness value near zero. If a
distribution's mean and median do not coincide, the distribution is skewed.

Positive Skew
If the distribution's mean is greater than (to the right of) the median, the
distribution is said to be skewed to the right. The tail of the distribution points to
the right, and the mound of data are on the left. Data found in pollution studies or
geological applications tend to have a lot of values clustered below the mean
value and a long tail into the high values.

Negative Skew
The opposite is referred to as a negative skew. Data found in geological
applications in limestone and sedimentary iron ores tend to have a lot of high
values with a long tail into the lower values.

Right-skewed distributions will have a positive skewness value; left-skewed

distributions will have a negative skewness value. Typically, the skewness value
will range from negative 3 to positive 3.

Geostatistics Without Tears 28

Kurtosis
Kurtosis is a statistical measure used to
describe the distribution of observed
data around the mean. Kurtosis
measures how flat or peaked the
distribution is, and how thick or thin the tails of the distribution are. The graph
on the right has a higher kurtosis than the graph on the left. It is more peaked at
the centre and it has fatter tails.

Many classical statistical tests depend on normality assumptions. Significant

skewness and kurtosis clearly indicate that data are not normal. If a data set
exhibits significant skewness or kurtosis (as indicated by a histogram or the
numerical measures) someway to deal with the problem must be attempted. One
approach is to apply some type of transformation to try to make the data normal,
or more nearly normal. The Box-Cox transformation is a useful technique for
trying to normalize a data set. In particular, taking the log or square root of a data
set is often useful for data that exhibit moderate right skewness.

Mesokurtic
A normal random variable has a kurtosis of 3 irrespective of its mean or standard
deviation. Kurtosis should be near 3 for a normal distribution (normal tails).

Leptokurtic
Kurtosis is greater than 3 for distributions that have thicker than normal tails.
Leptokurtosis is associated with distributions that are simultaneously peaked and
have fat tails. The graph on the right is leptokurtic.

Platykurtic
Kurtosis is less than 3 for distributions with thinner than normal tails.
Platykurtosis is associated with distributions that are simultaneously less peaked
and have thinner tails. Platykurtic distributions are said to have shoulders. The
graph on the left is platykurtic.

Normal Probability Plot

The following explanation of the normal probability plot and the explanation of the
quantile-quantile plot were adapted from the “Engineering Statistics Handbook” A
complete discussion on this topic can be found at the National Institute of Standards
and Technology (NIST) – Information Technology Laboratory – Statistical
Engineering Division website:
http://www.itl.nist.gov/div898/handbook/eda/section3/eda33.htm. The normal
probability plot is graphic technique for assessing whether or not a data set is
approximately normally distributed. The data are plotted against a theoretical normal
distribution in such a way that the points should form an approximate straight line.
Departures from the straight line indicate departure from normality. It is not easy to
judge how far the pattern in the plot can deviate from linearity before the assumption of
normality should be judge implausible. Be careful about deciding against the
plausibility of a normal distribution based on a normal probability plot when the
sample size is small.

Geostatistics Without Tears 29

Normally Distributed Data
The normal probability plot based on normally distributed data shows
a strong linear pattern. In this example, the normal distribution
appears to be a good model for these data which is verified by the
correlation coefficient of 0.999 of the line fit to the probability plot.
The fact that the points in the lower and upper extremes of the plot do
no deviate significantly from the straight line pattern indicates that
there are no significant outliers.

Short Tails
The normal probability plot based on data that has short tails shows a
non-linear pattern. In this example, the normal distribution is not a
good model for this data which shows up in two ways. First, the
middle of the data shows an s-like pattern. This is common for both
short and long tails. Second, the first few and the last few points
show a marked departure from the reference fitted line. In comparing
this plot to the long tail example the important difference is the
direction of the departure from the fitted line for the first few and last
few points. For short tails, the first few points show increasing
departure from the fitted line above the line and the last few points show
increasing departure from the fitted line below the line.

Long Tails
The normal probability plot based on data that has long tails shows a
reasonably linear pattern in the centre of the data. However the tails,
particularly the lower tail, show departures from the fitted line. For
data with long tails relative to the normal distribution, the non-
linearity of the normal probability plot can show up in two ways.
First, the middle of the data may show an s-like pattern. This is
common for both short and long tails. In this particular case, the s-
pattern in the middle is fairly mild. Second, the first few and the last
few points show marked departure from the reference fitted line. This
is most noticeable for the first few data points. In comparing this plot to the
short-tail example, the important difference is the direction of the departure from
the fitted line for the first few and the last few points. For long tails, the first few
points show increasing departure from the fitted line below the line and the last
few points show increasing departure from the fitted line above the line.

Geostatistics Without Tears 30

 Skewed
The normal probability plot based on data that are skewed right
(example) shows a strongly non-linear quadratic pattern in which all
the points are below a reference line drawn between the first and last
points. The quadratic pattern in the normal probability plot is the
signature of a significantly right-skewed data set. Similarly, if all the
point on the normal probability plot fell above the reference line
connecting the first and last points, that would be the signature
pattern for a significantly left-skewed data set.

Quantile-Quantile Plot
A Quantile-Quantile (QQ) plot is a plot of the quantiles of one data set
against the quantiles of a second data set. The QQ plot can determine if
two data sets come from populations with a common distribution. The
QQ plot has a number of advantages: sample sizes do not need to be
equal; many distributional aspects can be simultaneously tested (shifts in
location, shifts in scale, and changes in symmetry); the presence of
outliers can be detected. For example, if the two data sets came from
populations whose distributions differ only by a shift in location, the
points should lie along a straight line that is displaced either up or down
from the 45 degree reference line.

Measures of Location

Mean
The arithmetic mean (commonly called the average) is the sum of all data values
divided by the number of data values. The mean is a good measure of central
tendency for roughly symmetric distributions but can be misleading in skewed
distributions since the mean can be greatly influence by extreme data values.

Median
The median is the middle of a distribution. To calculate the median the data must
be ranked (sorted in ascending order). The median is the number in the middle.
The median value corresponds to a cumulative percentage of 50% (i.e., 50% of
the values are below the median and 50% of the values are above the median).

If the mean and the median are approximately the same value this is evidence
that the data may be normally distributed. A frequency distribution is said to be
skewed when its mean and median are different.

Mode
The mode is the most common (frequent value). The easiest way to look at
modes is on the histogram. There may be no mode if no value appears more than
any other. There may also be two modes (bimodal), three modes (trimodal), or

Geostatistics Without Tears 31

 four or more modes (multimodal). The mode will be close to the mean and
median if the data have a normal or near-normal distribution.

Quantiles
A quantile is the variable value that corresponds to a fixed cumulative frequency.
Any quantile can be read from the cumulative frequency plot:

 First quartile = 0.25 quantile

 Second quartile = median = 0.5 quantile

 Third quartile = 0.75 quantile

A quantile is the percent of points below a given value. That is, the 30% (0.3)
quantile is the point at which 30 percent of the data fall below and 70 percent fall
above that value. Probability intervals can also be read from the cumulative
frequency plot, i.e. the 90% probability interval.

Measures of Spread
Measures of spread describe the variability of a data set. This group includes the
range, variance, standard deviation, interquartile range and the coefficient of
variation.

Range
The range is the largest data value minus the smallest data value. As a general
rule of thumb, the range should be somewhere between 4 and 6 standard
deviations depending on how many samples are present.

Variance
One of the most useful statistics to measure the variation of the sample values is the
variance, which is found by considering the deviation of each sample from the
average value. The variance is calculated by squaring the deviations, then adding
them up and taking the average. Variance is a mean squared deviation and its units
are squared rather than in the srcinal sample units.

Standard Deviation
To get a descriptive statistic the square root of the variance is required to obtain the
standard deviation (root mean squared deviation). The standard deviation can be
interpreted as the difference between any data set member and the average of the
data set. The mean value fixes the centre of the distribution, and the standard
deviation scales the horizontal axis. In a normal distribution, 68% of the
population values lie within one standard deviation of the mean value. Just over
95% of the values lie within two standard deviations. Approximately 99.7% of
the observations are within 3 standard deviation of the mean.

Geostatistics Without Tears 32

 Contour plots of the standard deviations of predicted values at non-sampled
locations can be very useful. These contours show areas of higher uncertainty
(higher standard deviations). Sampling from these locations can substantially
improve the accuracy of predictions.

Geostatistics played a role in making the Channel Tunnel Project successful by

assessing the geological risks and optimizing the alignment of the tunnel
(RocNews, 2003). One of the most import criteria in optimizing the alignment
was to ensure that the tunnel was bored within the Chalk Marl avoiding the Gault
Clay material. Kriging was used to determine the boundary between the Chalk
Marl and the Gault Clay, based on data available before construction. Contours
of the standard deviations of predicted depths of this boundary were also
generated.

The standard deviation contours helped engineers to realize that improved

precision was required at certain tunnel sections. As more data became available
from surveys and ongoing construction, geostatisticians enabled the tunnel
engineers to readily improve the spatial model of the Chalk Marl/Gault Clay
interface. Geostatistical analysis reduced the risk of penetrating the Gault Clay to
acceptable levels. Penetration of the Gault Clay occurred only twice and in areas
that already been predicted from the geostatistical model.

Interquartile Range
The interquartile range is used to describe the spread of a data set. The
interquartile range is the difference between the first quartile (a number for which

25% of
75% of the
the data
data is
is less
less than
that that
that number) anda third
number) of set ofquartile (a interquartile
data. The number for which
range
is the range of the central 50% of the sample values.

Coefficient of Variation
The coefficient of variation is calculated by dividing the mean by the standard
deviation. The general rule of the coefficient of variation says, “if the standard
deviation is relatively large compared to the arithmetic average, the average can’t
be used to make decisions.” As a rule of thumb, when the standard deviation is
smaller than the mean, the data are relatively closely clustered and mean is
considered a reasonably good representation of the full data set. By contrast, if the
standard deviation is greater than the mean, then the data are relatively widely
dispersed and the mean is a rather poor representation of the full data set.

Geostatistics Without Tears 33

Correlation Analysis
Correlation analysis is used to determine the manner in which one (an
independent) variable affects another (dependent variable). In correlation
analysis, the coefficient of correlation, coefficient of determination, and standard
error of estimate provide assessments of the reliability.

Coefficient of Correlation
Pearson’s coefficient of correlation (r) is the most common measure of
correlation or predictability between two variables. Pearson’s r ranges in value
from -1 to 1. The lager the r (ignoring the sign) the stronger the association and
the more accurate it is to predict one variable from knowledge of the other
variable. If r is 0 there is no correlation (no relationship) between the two
variables.

The sign of the correlation implies the “direction” of the association. A positive
correlation means that relatively high scores on one variable are paired with
relatively high scores on the other variable, and low scores are paired with
relatively low scores. A negative correlation means that relatively high scores on
one variable are paired with relatively low score on the other variable.

A zero correlation does not necessarily mean there is no relationship between two
variables – it means there is no linear relationship. For this reason, scatter plots
are important supplements to statistical measures of association by providing a
graphical picture of where variables may be associated with each other.
Correlation measures the extent of association, but association does not imply
causation. It can happen that two variables are highly correlated, not because one
is related to the other, but because they are both strongly related to a third
variable.

Deciding if r is "good enough" is an interpretive skill developed through

experience. As a rule of thumb, if r > 0.95, there is good evidence for positive
linear correlation, if r < -0.95, there is good evidence for negative linear
correlation, and if -0.95 < r < 0.95, the data is either not linear, or "noisy". As an
informal rule of thumb, call the relationship strong is r is greater than +0.8 or less
than (more negative than) -0.8, weak if r is between -0.5 and 0.5, and moderate
otherwise.

Coefficient of Determination
The coefficient of determination (r2, r2 or R2 – the square of the coefficient of
correlation) is the ratio of the explained variation to the total variation and
provides the proportion of the variation (fluctuation) of one variable that is
predictable from the other variable. The coefficient of determination is a measure
that provides the mechanism to determine how certain one can be in making
predictions from a certain model or graph.

The coefficient of determination ranges from 0 to 1 and denotes the strength of

the linear association between two variables. The coefficient of determination
represents the percent of the data that is closest to the line of best fit. If the

Geostatistics Without Tears 34

 coefficient of correlation (r) is 0.922, then the coefficient of determination r2 =
0.850, which means that 85% of the total variation in the variable Y can be
explained by the linear relationship between X and Y (described by the
regression equation). The other 15% of the total variation in Y remains
unexplained.

The better the regression line fits the data the larger the r2. If the regression line
passes exactly through every point on the scatter plot, it would be able to explain
all of the variation. The further the line is away from the points the less it is able
to explain. A correlation coefficient of one (perfect correlation) is rarely, if ever,
achieved. The coefficient of determination should be at least 0.6 and preferably
between 0.8 and 1.0 for a reliable surface model.

Standard Error of Estimate

The standard error of estimate (se, SE, Se or SEE) is a measure of the accuracy of
predictions. The top regression example show the points are closer to the
regression line then they are in the bottom example. The predictions in the top
graph are more accurate than the bottom graph. A regression line is the line that
minimizes the sum of square deviations of prediction (also called the sum of
squares error) . For example, if the standard error of the estimate is 775 then the
prediction will be within +/- 775 units. The standard error of estimate is a
measure of the variation (scatter) of the points about the regression line and is
analogous to the standard deviation. However, the standard deviation measures
the scatter about the arithmetical mean of a sample whereas the standard error of
estimate measures the scatter of the dependent variable about the regression line.
If the standard error of estimate is equal to the standard deviation of the
dependent variable, then the correlation coefficient is equal to zero. The standard
error of estimate provides a measure of the closeness of the point to the curve
relation; it does not show the degree of correlation.

Geostatistics Without Tears 35

Geostatistics Without Tears 36
EXPLORATORYDATAANALYSIS
Exploratory data analysis (EDA) is a set of graphical tools that help bridge the gap
between data and modeling. The purpose of summary analysis is to arrive at a few
key statistics such as mean and standard deviation, whereas the goal of exploratory
data analysis is to gain insight into the process behind the data. Summary statistics
are passive and historical, where as EDA is active and futuristic.

Censored Data
Censored data is data having a <,> or estimated qualifier, as is typical with water
quality data. An accepted practice is to convert not detected (ND) values to one-
half of the detection limit; to convert less than detection (<DL) values to the
detection limit; and to convert estimated (E) values to the value estimated.

Co-located Samples
If the samples represent replicates of the same sample collected at the same time, an
accepted practice is to average the samples and only include the average in the data
set. If the sample represents a different element or a different time of collection an
accepted practice is to change the location coordinates slightly. A problem arises
with exact kriging when multiple values are measured at locations. A method of
dealing with multiple samples was documented in Matheron's srcinal works. With
multiple samples at certain locations, the diagonal entry in the kriging system can
be modified. Instead of γ(0) = 0 use γ(0) = (n-1)/n times nugget effect. This tells the
kriging system you have replicates and it will adjust weights and optimal estimator
accordingly (Isobel Clark, 2001b)

Declustering
In many situations the spatial location of data collection sites are not randomly or
regularly spaced. Data may have been sampled with a higher density in some
places that in others. The mean value will be over-estimated if the samples are
spatially clustered in the high value areas. A declustering method must be
implemented to adjust for preferential sampling. One solution to preferential
sampling is to weight the data, with data in densely sampled areas receiving less
weight and data in sparsely sampled areas receiving greater weight. This can be
accomplished in a couple of ways.

One method is cell declustering. A grid of cells is overlaid the data locations, and
the weight attached to each data point is inversely proportional to the number of
data points in its cell. Some researchers suggest that the cell size can be chosen
corresponding to the minimum weighted mean if the data have been
preferentially sampled in areas of high values. Conversely pick the cell size
corresponding to the maximum weighted mean if the data have been

Geostatistics Without Tears 37

 preferentially sampled in areas of low value. This can lead to bias in other cases,
and in general this method should not be used.

Another scheme uses a polygon method that defines a polygon around each
spatial data location, such that all locations within that polygon are closer to the
data location than any other data location. The idea is to weight each data point
in proportion to the area that it represents. The problem with this method is that it
is difficult to define weights toward the edge. The edge points can often receive
large weights unless a border encloses the data.

Extreme Values
The first step in dealing with extreme values is to establish why the data contain
extreme highs. Is there a high degree of imprecision in measuring values (the
sample observations are actually inaccurate)? Is the distribution of data values
skewed? Are there two (or more) populations, only one of which gives the high
values? Are there other reasons?

Once the reason for extreme values has been determined it will be possible to
deal with this problem more objectively. You may be able to:

• use the error statistics in cross validation to assist in identifying

erroneous sample measurements.

• use transformations or distribution-free approaches to geostatistics.

• use a mixed model together with indicator geostatistical approaches.

• identify extreme values as errors so they can

be eliminated from the data set.

• stratify the data set by grouping extreme values together and treating
them as a separate statistical population. Stratified kriging can be done
by interpolating for different sub-regions separately and then the results
of the interpolation can be combined as a single map.

• extreme values can be accepted and their effect on the distribution can be
mathematically moderated. There are two sets of mathematical methods
to do this. The first is to transform the data set before applying statistical
measures. The other is to apply robust statistical measures that are better
designed mathematically to reduce the effect of outliers or a small
number of samples.

Skewed Data with Zeros

Isabel Clark (1987, 2002) suggests several ways of tackling skewed data with
zeroes. Try a lognormal probability plot and see whether it is a straight line or if
it drops off the line at low values. This is indicative of a three parameter
lognormal distribution which needs an additive constant. Find the additive
constant that makes the line straightest (Isabel’s criterion) or the skewness closest
to zero (Sichel's recommendation). Treat the zeroes as a different population. Are
the sampled locations zero because there are no fish there or

Geostatistics Without Tears 38

 because no fish were caught? If the later, use an indicator approach to separate
the “no fish” population from the “some fish” population. Then do lognormal
transformations on the “some fish” and recombine for a final result. The ‘not as
nice’ approach is to use the lognormal probability plot to choose a threshold
value to replace the zeroes. This assumes that all areas sampled are “some fish”
areas and but no fish were caught.

Spatial Outliers

Special attention
since these valuesshould be given to
will seriously abnormal
distort values (spatial
the interpolation outliers)
process. A
good way to detect the spatial outliers is to map or visually display
the data set. Map display provides insights about the data set, how the
samples are spatially distributed (uniform or clustered), their general
pattern, trend and the extreme high or low values associated with the
data set. The spatial location of extreme values is helpful in detecting
erroneous data.

Care should be taken before deciding that a given outlier is actually a

bad data point. An isolated extreme value may be suspicious, but this
may not be sufficient to justify its removal. In some cases outlier values are
obvious. There are a variety of methods that are used to treat outliers including
inclusion, accommodation, replacement and rejection. The method to use
depends on the objectives of the interpolation (depiction versus prediction; trend
versus anomalies) and the nature of the data point. The decision to discard
extreme values must be made with care, and data should be dismissed only if
they are clearly wrong. It is possible to set a missing value indicator prior to
surface generation in some software packages so that the missing values are
ignored during analyses.

Stratification (Data set Subdivision)

If a phenomenon behaves differently in different locations the sample values
collected will also differ significantly based on where they were taken. In this
case, if a surface is created using all the measured points the basic principle of
geography may be violated (things closer to one another are more alike). If this is
the case it may be beneficial to divide or stratify the data into homogenous
regions (or strata). Once the data is divided into parts the surface can be
interpolated for each part separately and then the results can be combined.

The decision to split the data into more homogeneous subsets should be based on
physical considerations and the density of the available data. Each subset should
have enough data to allow the reliable inference of statistics for each population.
Stratification should be based on information other than the data values
themselves. Don’t stratify by putting high values into one stratum and low value
into another. Secondary information (soil type, vegetation class) can be used
when there is justification for stratification. Alternatively the data could be
divided according to geographic location or elevation. Care should be taken
because there may be no justification for stratification.

Geostatistics Without Tears 39

Time
According to Clark (2001c) there are two ways of approaching data that has a
time element:

1. Treat time as a covariable and use cokriging.

2. Treat time as a dimension (an additional coordinate).

Transformation
Data transformations can be used to make data normally distributed. Kriging
relies on the assumption of stationarity, and this assumption requires in part, that
all data values come from distributions that have the same variability. Data
transformations can improve predictions in many but not all cases.
Transformations and trend removal are often applied to data to justify
assumptions of normality and stationarity. The goal of transformations is to
remove the relationship between the data variance and the trend. When data are
composed of counts of events, such as crimes, the data variance is often related to
the data mean. That is, if you have small counts in part of your study area, the
variability in that region will be larger than the variability in a region where the
counts are larger. The histogram and normal QQ plots can be used to see what
transformations, if any, are needed to make a data set more normally distributed.
The same transformation will likely equalize variances as well, helping to satisfy
the stationarity assumption.

Arcsine Transform
For percentages or proportions (data between 0 and 1) the arcsine transform can
be used. When data consists of proportions, the variance is smallest near 0 and 1
and largest near 0.5. The arcsine transformation often yields data that has
constant variance throughout the study area and often makes the data appear
normally distributed (Krivoruchko, 2005). Kriging prediction with a Box-Cox or
arcsine transformation is known as transGaussian kriging (Krivoruchko, 2002a).
To create an output map, the results of transGaussian kriging require back
transformation to the srcinal data, however, it should be noted that this can only
be done approximately (Cressie, 1993).

Box-Cox Transform
Box-Cox (also knownas power transformation) is a useful method to alleviate
heteroscedasticity when the distribution of the dependent variable is not known.

These
causedtransformations
by nonnormality areof
defined
one ofonly
thefor positive an
variables, dataindirect
values.relationship
Heteroscedasticity
betweenis
variables, or the effect of a data transformation.

Geostatistics Without Tears 40

Lognormal Transform
Highly positively skewed data are encountered in many fields such as
environmental, pollution and mining data. Kriging is a linear estimator and is
sensitive to a few large samples, which may bias the results. To cope with highly
positively skewed data geostatisticians developed lognormal kriging, which is
kriging applied to lognormal transforms of data followed by a back-transform of
the final estimates (Journel and Huijbregts, 1978). A back-transformation is not
required if the srcinal variable is not important, for example if the logarithm
itself is a useful index. The log transformation requires that all data are positive.

Lognormal kriging must be used with caution because it is non-robust against

departures from the lognormal model and the back-transform is very sensitive to
semivariogram fluctuations. The back-transform tends to exaggerate any error
associated with the interpolation and is most dramatic for extreme values.
Although lognormal transforms attenuate the impact of high values it is not the
most appropriate for censored data, which are better viewed as a different
statistical population (Sito and Goovaert, 2000).

Sichel (1971, 1973) found that lower frequency distribution drop off can be
brought back into line by adding a constant number to every one of the sample
values before taking logarithms. If the probability plot shows a straight line, the
population is said to be “three parameter lognormal.” According to Clark and
Harper (2000) the additive constant has no physical meaning and is basically, just
an artifact to normalize the logarithms. The size of the additive constant is a
debatable issue with the most common rule of thumb being that it would be a
problem if the additive constant starts to get to the same order of magnitude as
the average value.
Sampling can reflect a mixture of normal or lognormal populations. For example
if measurements are taken from both male and female populations the results may
be comprise of two different normal populations. This structure will be reflected
in the histogram and probability plots.

Normal Score Transform

The normal score transform ranks a data set, from lowest to highest values, and
matches these ranks to equivalent ranks from a normal distribution. The
transform is then defined by taking values from the normal distribution at that
rank. Unlike the lognormal transform the normal score transform allows the
symmetrization of the distribution of data regardless of the shape of the sample
histogram (Goovaerts, 1997 and Deutsch and Journel, 1998).

A normal score transform allows the data distribution to be made symmetrical

regardless of the shape of the sample histogram (Goovaerts, 1997 and Deutsch and
Journel, 1998). The kriging of normal scores is referred to as multi-Gaussian
kriging. Normal score transforms requires apriori specifications of the mean. An
unfortunate feature of the normal score transform is the loss of the interval
properties of the original data because the quantile definition depends on whether it
occurs in the middle or at the tails of a standard normal distribution.

Geostatistics Without Tears 41

 The normal score transformation cannot deal with any systematic trend in the
data and should only be used in situations where the mean of the data is constant.
Functional transformations such as lognormal or power work globally to
transform data and fixed effects to a model with normal errors, where as the
normal score transformation is working to make errors normally distributed. The
normal score transform must occur after detrending since covariance and
variograms are calculated on residuals after trend correction. Back transformation
of a normal score transform is approximate and can be seriously biased when a
trend exists in the data (Krivoruchko, 2002a). Although normal-score transforms
attenuate the impact of high values it is not the most appropriate for censored
data, which are better viewed as a different statistical population (Sito and
Goovaert, 2000).

The goal of the normal score transform is to make all random errors for the whole
population normally distributed. Thus it is important that the cumulative
distribution from the sample reflect the true cumulative distribution of the whole
population (Krivoruchko, 2005). An unfortunate property of the normal score
transformation is that it destroys the interval property of the srcinal data,
because the quantile has a different meaning depending whether it occurs in the
middle or at the tails of a standard normal distribution.

Scale Transform
In some cases it can be useful to scale data to a range of 0 to 1 if the values are
extremely large.

Square Root Transform

The square-root transformation is a special case of the Box-Cox transformation.
For count data the square root transforms can help make the variance more
constant throughout the study area, and often makes the data appear normally
distributed (Krivoruchko, 2005). Square-root transforms retain the measurement
units. For square root, the back-transform through squaring tends to exaggerate
any error associated with the interpolation. Such exaggeration of errors is most
dramatic for extreme values (Sito and Goovaert, 2000).

Offsets and Back-transforms

Offset
In the case of a lognormal or square root transform if the Z value spans the range
of less than one to greater than one (<1 to >1) all values should be made greater
than one prior to transformation by adding an offset value.

Back-transformation
When a transformation is chosen, after analysis of the transformed data, the
output data are customarily (but not necessarily) back-transformed to the srcinal

Geostatistics Without Tears 42

 data domain. Three potential back-transformation choices are: none, standard or
weighted.

Standard
The standard back-transform is simply the converse of the transformation –
scaled values are rescaled to the srcinal range, log transformed values are raised
to the natural exponent (e) and squared values are raised to the square root (0.5).
Offset values are subtracted from the back-transformed values.

Weighted
The weighted back-transformation is a complex back-transformation that more
closely approximates true population statistics than simple back transformations
(Hann, 1977 and Krige, 1981).

Trends
When there is a trend in the data the first task is to identify the pattern and degree
of the trend and then to decide whether the trend needs to accommodated. Trends
can occur because of the measurement process, sampling design, or can be due to
the physical qualities of the sampled object. Climatic data may follow a strong
directional pattern. Increasing rainfall from west to east leads inevitably to
constantly increasing local means in the same direction.

The semivariogram shows an example of data with

definite trend. The fit appear to be fairly good for
20 distances up to 45 kilometres, but beyond this the
trend must be taken into account. The trend may or
may not cause interpolation problems until after
15
h the range of influence is passed, however this is
not always the case and the closer the parabolic
n
10 behavior is to the srcin the more attention will
need to be paid to the trend.
m

5 There are several quantitative management tools

for trends. One possibility is to remove the trend
component. The trend component can be
0
0K m 1 0K m 2 0K m 3 0K m 4 0K m 5 0K m 60K m subtracted from the samples before calculation of
variogram and identification of the spatial
lag dis tance h
( )
continuity. The trend surface can then be added
back after kriging (Agterberg, 1974) or can be used as data input in universal
kriging methods (Pebesma and Wesseling, 1998). Despite criticisms that the
detrended data often have a different covariance function or variogram from the
srcinal data, they do have the same generalized covariance function (Kitanidis,
1993).

Secondary populations in the histogram, not removed by trend analysis, are

strong suggestions to look for the existence of multiple populations, or perhaps
an indicator method should be used. Geostatistical analysis is deeply dependent
on the assumptions of single population with a consistent behavior.

Geostatistics Without Tears 43

Geostatistics Without Tears 44
APPLIED GEOSTATISTICS
The underlying principle behind geostatistics (and spatial interpolation in general)
is that observation points closer together are more likely to have similar values than
points further away. Tobler's Law of Geography (Tobler, 1979) is often cited as the
First Law of Geography, where “everything is related to everything else, but near
things are more related than distant things” (also known as autocorrelation). This
means that samples collected close to one another are often more similar than
samples collected further away, whether in space or in time.

Geostatistical interpolation is a two stage process:

1. Study the data to establish the predictability of values from place to place
(determine dependency rules) and define the degree of autocorrelation.

The sample semivariance is used to estimate the shape of the variogram (the
curve that represents the semivariance as a function of distance). The variogram
describes the spatial relationship between the data points by showing the spatial
structure and the data associations.

2. Once spatial or temporal dependency is established, the value at locations that

have not been sampled are interpolated based on the degree of spatial
autocorrelation. The process of prediction of the unknown values is called
kriging.

The estimated semivariance function is used to determine the weights needed to

define the contribution of each sampled point to the interpolation. Sample points
close to the point for which an estimated value is to be generated contributes the
most to the interpolation.

VARIOGRAMS
Geostatistics defines the correlation between any two values separated by a
distance (known as the lag distance) then uses this information to make
predictions at unsampled locations. Semivariance is a measure of the degree of
spatial correlation among sample data points as a function of the distance and
direction between the sample data points. The independent variable is the
distance between pairs of points. The dependent variable is the semivariance of
the differences in the data values for all samples a given distance apart. The
semivariogram (or simply the variogram) controls how kriging weights are
assigned to points during interpolation, and consequently controls the quality of
the results. The semivariance increases with distance until the semivariance
equals the variance around the average value and no longer increases, causing a
flat region to occur on the semivariogram called the sill.

Geostatistics Without Tears 45

 Calculating the variogram

Scale
It is not difficult to calculate and plot the semivariogram.
0 K1m The first step is to define a lag increment h(), which is the
spacing between any two points in the data. For the
20.1 14.6 13.0 15.4 12.9 16.1 16.5
7 Km example data a 1 kilometer lag increment is arbitrarily
defined. For each pair separated by 1 kilometer, calculate
15.7 11.8 20.9 17.1 24.3
6 Km
) the difference, and then square the difference. Sum up all
g the differences and divide by twice the number of pairs.
n i
th 5 Km
28.5 13.6 35.4 27.2
r This gives the measure of similarityγ(h) for the variogram
o
(N 30.2 0.0 33.3 19.6 23.8
for a particular lag increment or distance. The same is done
te 4 Km for other lag distances of 2, 3, 4, 5, and 6 kilometers.
a
in Grouping the pairs by location is referred to as binning.
d
r 29.9 31.3 7.5 33.3 23.3 18.2
That is, all points that are within 0 to 1 kilometres apart are
o 3 Km
o
c
- grouped into the first bin, those that are within 1 to 2
Y
2 Km
43.7 14.4 12.1 16.3 15.5 55.1
kilometres apart are grouped into the second bin, and so on.
N
33.6 29.7 31.9
1 Km

1 Km 2K m 3K m 4K m 5K m 6K m 7K m
γ(h) = (Σ [Z(xi) - Z(xi + h)]
2
) / 2N
i=1
X-coordinate ( Easting )

1
First Lag Distance
Twenty (N = 20) pairs are found for lag distance h = 1. The pairs in the east-west
direction and their calculated differences:
20.1 – 14.6 = 5.5 14.6 - 13 = 1.6 13 - 15.4 = -2.4 15.4 - 12.9 = 2.5
12.9 - 16.1 = -3.2 16.1 - 16.5 = -0.4 15.7 - 11.8 = 3.9 20.9 - 17.1 = 3.8
17.1 - 24.3 = -7.2 28.5 - 13.6 = 14.9 33.6 - 19.6 = 13.7 19.6 - 23.8 = -4.2
29.9 - 31.3 = -1.4 31.3 - 7.5 = 23.8 7.5 - 33.3 = -25.8 33.3 - 23.3 = 10
14.4 - 12.1 = 2.3 12.1 - 16.3 = -4.2 16.3 - 15.5 = 0.8 15.5 -55.1 = -39.6

The sum of the squared differences:

30.25 + 2.56 + 5.76 + 6.25 + 10.24 + 0.16 + 15.21 + 14.44 + 51.84 + 222.01 + 187.69 +
17.64 + 1.96 + 566.44 + 665.64 + 100.00 + 5.29 + 17.64 + .64 + 1568.2 = 3489.82

The gamma for the first lag distance is:

γ(1) = 3489.82/(2*20) = 87.25

Second Lag Distance

The lag distance h is incremented such that h = h + h = 2h. In this example the
incremental value would be two kilometers. A search is conducted for pairs at a
lag distance of two kilometers in the east-west direction. All pairs used for the h
= 1 search and calculation are not included in the new lag interval.

1
The example variogram calculations have been adapted from Carr (1995).

Geostatistics Without Tears 46

 Twenty (N = 20) pairs are found for lag distance h = 2.
20.1 - 13.0 = 7.1 14.6 - 15.4 = -0.8 13.0 - 12.9 = 0.1 15.4 - 16.1 = -0.7
12.9 - 16.5 = -3.6 11.8 - 20.9 = -9.1 20.9 - 17.1 = -3.4 35.4 - 27.2 = 8.2
30.2 - 0 = 30.2 0 - 33.3 = -33.3 33.3 - 23.8 = 9.5 29.9 - 7.5 = 22.4
31.3 - 33.3 = -2.0 7.5 - 23.3 = -15.8 23.3 - 18.2 = 5.1 43.7 - 14.4 = 29.3
14.4 - 16.3 = -1.9 12.1 - 15.5 = -3.4 16.3 - 55.1 =-38.8 29.7 - 31.9 = -2.2

The sum of the squared differences:

50.41 + 0.64 + 0.1 + 0.49 + 12.96 + 82.81 + 11.56 + 67.24 + 912.04 + 1108.89 + 90.25 +
501.76 + 4.00 + 249.64 + 26.01 + 858.49 + 3.61 + 11.56 + 1505.40 + 4.84 = 5502.65

The gamma for the second lag distance is:

γ(2) = 5502.65/(2*20) = 137.57

Third Lag Distance

Pairs of observations in the east-west direction and with lag distances of three
kilometers are now listed.

Fifteen (N = 15) pairs are found for lag distance h = 3.

20.1 - 15.4 = 4.7 14.6 - 12.9 = 1.7 13.0 - 16.1 = -3.1 15.4 - 16.5 = -1.1
15.7 - 20.9 = -5.2 11.8 - 17.1 = -5.3 13.6 - 35.4 = -21.8 0 - 19.6 = -19.6
29.9 - 33.3 = -3.4 31.3 - 23.3 = 8.0 33.3 - 18.2 = 15.1 43.7 - 12.1 = 31.6
14.4 - 15.5 = -1.1 12.1 - 55.1 = -43.0 33.6 - 29.7= 3.9

The sum of the squared differences:

22.09 + 2.89 + 9.61 + 1.21 + 27.04 + 28.09 + 475.24 + 384.16 + 11.56 + 64.00 + 228.01
+ 998.56 + 1.21 + 1849.00 + 15.21 = 4117.88

The gamma for the third lag distance is:

γ(3) = 4117.88/(2*15) = 137.26

Fourth Lag Distance

Pairs of observations in the east-west direction and with lag distances of four
kilometers are now listed.

Twelve (N = 12) pairs are found for lag distance h = 4.

20.1 - 12.9 = 7.2 14.6 - 16.1 = -1.5 13.0 - 16.5 = -3.5 15.7 - 17.1 = -1.4
11.8 - 24.3 = -12.5 28.5 - 35.4 = -6.9 30.2 - 33.3 = -3.1 0 - 23.8 = -23.8
29.9 - 23.3 = 6.6 7.5 - 18.2 = -10.7 43.7 - 16.3 = 27.4 14.4 - 55.1 = -40.7

The sum of the squared differences:

51.84 + 2.25 + 12.25 + 1.96 + 156.25 + 47.61 + 9.61 + 566.44 + 43.56 + 114.49 +
750.76 + 1656.49 = 3413.51

Geostatistics Without Tears 47

 The gamma for the fourth lag distance is:

γ(4) = 3413.51/(2*12) = 142.23

Fifth Lag Distance

Pairs of observations in the east-west direction and with lag distances of five
kilometers are now listed.

Eight (N = 8) pairs are found for lag distance h = 5.

20.1 - 16.1 = 4.0 14.6 - 16.5 = -1.9 15.7 - 24.3 = -8.6 13.6 - 27.2 = -13.6
30.2 - 19.6 = 10.6 31.3 - 18.2 = 13.1 43.7 - 15.5 = 28.2 33.6 - 31.9 = 1.7

The sum of the squared differences:

16.00 + 3.61 + 73.96 + 184.96 + 112.36 + 171.61 + 795.24 + 2.89 = 1360.63

The gamma for the fifth lag distance is:

γ(5) = 1360.63/(2*8) = 85.04

Sixth Lag Distance

Pairs of observations in the east-west direction and with lag distances of six

kilometers are now listed.

Five (N = 5) pairs are found for lag distance h = 6.
20.1 - 16.5 = 3.6 28.5 - 2.7.2 = 1.3 30.2 - 23.8 = 6.4 29.9 - 18.2 = 11.7
43.7 - 55.1 = -11.4

The sum of the squared differences:

12.96 + 1.69 + 40.96 + 136.89 + 129.96 = 322.46
The gamma for the sixth lag distance is:

γ(6)= 322.46/(2*5) = 32.25

Geostatistics Without Tears 48

 a) b)
20.1 14.6 13.0 15.4 12.9 16.1 16.5 20.1 14.6 13.0 15.4 12.9 16.1 16.5
7 Km 7 Km

15.7 11.8 20.9 17.1 24.3 15.7 11.8 20.9 17.1 24.3
6 Km 6 Km
) )
g g
in in
h
t 5 Km
28.5 13.6 35.4 27.2 ht 5 Km
28.5 13.6 35.4 27.2
r r
o o
N N
( 30.2 0.0 33.3 19.6 23.8 ( 30.2 0.0 33.3 19.6 23.8
te 4 Km te 4 Km
a a
n i ni
d d
r 29.9 31.3 7.5 33.3 23.3 18.2 r 3 Km 29.9 31.3 7.5 33.3 23.3 18.2
o 3 Km o
o o
c c
-Y -Y
43.7 14.4 12.1 16.3 15.5 55.1 43.7 14.4 12.1 16.3 15.5 55.1
2 Km 2 Km

33.6 29.7 31.9 33.6 29.7 31.9

1 Km 1 Km

1 Km 2 Km 3 Km 4 Km 5 Km 6 Km 7 Km 1 Km 2 Km 3 Km 4 Km 5 Km 6 Km 7 Km

X-coordinate ( Easting ) X-coordinate ( Easting )

(c) (d)
20.1 14.6 13.0 15.4 12.9 16.1 16.5 20.1 14.6 13.0 15.4 12.9 16.1 16.5
7 Km 7 Km

15.7 11.8 20.9 17.1 24.3 15.7 11.8 20.9 17.1 24.3
6 Km 6 Km
) )
g g
in in
th 5 Km th 5 Km
28.5 13.6 35.4 27.2 28.5 13.6 35.4 27.2
r r
o o
N N
( 30.2 0.0 33.3 19.6 23.8 ( 30.2 0.0 33.3 19.6 23.8
e t 4 Km e t 4 Km
a a
n
i n
i
d d
r 29.9 31.3 7.5 33.3 23.3 18.2 r 29.9 31.3 7.5 33.3 23.3 18.2
o 3 Km o 3 Km
o o
c - c -
Y 43.7 14.4 12.1 16.3 15.5 55.1 Y 43.7 14.4 12.1 16.3 15.5 55.1
2 Km 2 Km

33.6 29.7 31.9 33.6 29.7 31.9

1 Km 1 Km

1 Km 2 Km 3 Km 4 Km 5 Km 6 Km 7 Km 1 Km 2 Km 3 Km 4 Km 5 Km 6 Km 7 Km

X-coordinate ( Easting ) X-coordinate ( Easting )

(e) (f)
20.1 14.6 13.0 15.4 12.9 16.1 16.5 20.1 14.6 13.0 15.4 12.9 16.1 16.5
7 Km 7 Km

15.7 11.8 20.9 17.1 24.3 15.7 11.8 20.9 17.1 24.3
6 Km 6 Km
) )
g g
n
i n
i
ht 5 Km
28.5 13.6 35.4 27.2 ht 5 Km
28.5 13.6 35.4 27.2
r r
o o
N N
( 30.2 0.0 33.3 19.6 23.8
( 30.2 0.0 33.3 19.6 23.8
te 4 Km te 4 Km
a a
n
i n
i
d d
r 3 Km 29.9 31.3 7.5 33.3 23.3 18.2 r 3 Km 29.9 31.3 7.5 33.3 23.3 18.2
o o
o o
c
- c
-
Y 2 Km 43.7 14.4 12.1 16.3 15.5 55.1 Y 2 Km 43.7 14.4 12.1 16.3 15.5 55.1

33.6 29.7 31.9 33.6 29.7 31.9

1 Km 1 Km

1 Km 2 Km 3 Km 4 Km 5 Km 6 Km 7 Km 1 Km 2 Km 3 Km 4 Km 5 Km 6 Km 7 Km

X-coordinate ( Easting ) X-coordinate ( Easting )

Transects used to estimate the semi-variances at varying lag distances ( h)

Geostatistics Without Tears 49

 Graphing the Semivariogram
A summary table of the east-west directional semivariogram calculations for the
data represented is:
Interval of h Number of Pairs γ(h)
1 20 87.25
2 20 137.57
3 15 137.26
4 12 142.23
5 8 85.04
6 5 32.25

The semivariogram provides a graphical and numerical measure of the spatial

continuity within a study area by plotting semivariance γ(h) versus lag distance
(h). The semivariogram values cannot be used directly instead a model must be
fitted to the semivariogram. A line is fit through the plotted points to assess the
spatial correlation. Once the model is fit, then this model is used to determine
semivariogram values for various distances. The initial value of h is called the
class size. Variogram modeling requires experience and practice.

y 200 ar
ti il
r m
si
la ss
u le
im
s
f 150 Spherical model
o sill 140
e
r )
u (C
s e
a
e n
c no further corralation between
ia sample points ( covariance becomes 0)
m 100 r
a
- v
l
) a
r
(h tcu ( C0 + C)
u
e
c rts
n 50
a
ir n u g g et ( C0) 2 0
a r
v a A0
ie
m e ilm
r
o ra ng e of in flu en ce ( ) 3
m si
s 0
0Km 1Km 2Km 3Km 4Km 5Km 6Km

lag distance (h ) - distance between samples

Geostatistics Without Tears 50

The Semivariogram
Kriging is a robust technique and minor errors in estimation of the semivariogram
parameters make little difference to the reliability of interpolation (Trangmar, Yost
and Uehara, 1985). Semivariograms are statistical measures that assume the input
sample data are normally distributed and that there are no trends in the local
neighborhood means and standard deviations. The semivariogram has a number of
components to which the semivariogram model is fit.

Coordinate System
The geographic latitude/longitude coordinate reference system uses angular
measurements to describe a position on the surface of the earth. Only at the
equator is the distance represented by one degree of longitude equal to the
distance represented by one degree of latitude. Moving towards the poles the
distance between lines of longitude becomes progressively smaller until at the
exact location of the pole all 360° of longitude are represented by a single point.
Because lines of longitude are not uniform, a latitude/longitude is not a suitable
coordinate system for lag distances.

Choosing a suitable projection is necessary to ensure that the value of x and y

units are equivalent and constant across the study and that distance and area
calculations are in real distance units. Coordinates for lag separation distances are
presumed to be in Cartesian space. The coordinates start with a 0,0 srcin that
increases for X in an easterly direction and for Y in a northerly direction. Values

can be less than or greater than 0.

Lag

Lag Distance
The lag distance is the range over which autocorrelation is calculated. The
number of lags and the size of the lags must be specified based on knowledge of
the phenomena being analyzed and the reason for modeling the variogram. The
selection of a lag size has important effects on the semivariogram.

• If the lag distance is too large, short-range autocorrelation may be

masked by forcing data pairs of widely varying separation distances into
a single lag increment and therefore short-range autocorrelation may not
be detected.

• If the lag distance is too small, there may be too many empty sample and

class sizes to get representative averages for a particular lag distance

because there are relatively few pairs contributing to any one lag
increment resulting in a ‘noisy’ semivariogram plot. The number of
samples in each bin may be too small to be representative.

• The general rule is that lag size times the number of lags is
approximately equal to one-half of the largest distance among all points.

Geostatistics Without Tears 51

Lag Class
Lag class distance is selected by experimenting with increasing or decreasing the
value and noting its influence on the semivariogram plot. The goal is to find lag
distances that generate a relatively smooth continuous structure. If the range of
the fitted variogram model is very small relative to the extent of the
semivariogram then the lag class distance can be decreased. If the range of the
fitted semivariogram model is large relative to the extent of the semivariogram
then the lag size can be increased. Typically variograms should generally be
limited to a maximum distance equal to half the sampled distance.

Optimally, a class size should be found where approximately the same number of
pairs result for each lag increment. For some data sets such a class size cannot be
found due to sampling geometry. For these data sets a class size should be sought
yielding an interpretable semivariogram showing a clear sill and range. It is
common that the sill will be reached in less than 20 lags (Myers, 1997).

Use the smallest lag class distance spacing that produce acceptable number of
pairs. Ideally 30 to 50 data pairs are consider a practical minimum to obtain a
reliable estimate of the variogram at any lag (Armstrong, 1984 and Russo, 1984).
In statistics, this approximates the number of pairs needed to produce a normal
distribution if drawing samples from a random distribution. Webster and Oliver
(1992) reported that variograms computed on fewer than 50 sample points are of
little value and that at least 100 data are needed. Their experiments suggest that for
a normally distributed isotropic variable a variogram computed from a sample of
150 data might often be satisfactory, while one derived from 225 data will usually
be reliable. The best lag spacing may differ for an isotopic versus anisotropic
variograms of the same data (Welhan, 2005).
Variable Lag Class Distance
Specifying a variable non-uniform lag class is appropriate when locations are
irregularly spaced or clustered across the interpolation area, or if the phenomenon
being studied is known or suspected to be composed of several components. In
this situation, a suggestion is to use lags with small separation distances at short
distances and use lags with increasingly larger separation distances at longer
distances. Changing the width of each lag may help improve detection of
continuity.

Distance Tolerance
Lag distance tolerance is defined to deal with irregularly spaced data. Typically
lag tolerance is defined to be half the lag increment used. For a lag increment of
one kilometer the lag tolerance would be ± 500 meters. When the variogram is
plotted the lag distance will be the average of all the distances between pairs
falling within the tolerance. The larger the lag tolerance the more pairs that can
be defined and the variogram will look smoother. Larger lag tolerances may also
lead to unstable variograms that mask the continuity structure. The smaller the
tolerance the fewer number of pairs will be defined and the variogram will not be
as smooth.

Geostatistics Without Tears 52

 Directional Tolerance
Choosing a lag distance and angular tolerance is similar to focusing a
o
0

camera. There will be a certain lag (in each direction) at which the
22.5
angular
o

calculation will balance between detail on distance, detail on angles

5
L
ag tolerance
and number of pairs. Use the smallest directional tolerance angle that
4 retains an acceptable number of data pairs per lag. Using too large a
s
i ag
x
a
-
y
L
o
tolerance angle can over generalize by incorporating more pairs and
22.5

L
ag
3
le
angular
tolerance
will mask anisotropy that may be present. A tolerance of 90 degrees
g
an
2 ec
t io
n means an omni-directional variogram, regardless of azimuth.
ag d ir
L

1 nce
Directional
examining thetolerance must be chosen carefully and the process of
h -scatter plots at different lag widths or angular
ag ar tolera
L angul

o
tolerances is essential.
90
x-axis

lag distance interval

Range
The range (A0) is the distance where the semivariance reaches the sill. The range
of influence defines a neighborhood within which all locations are related to one
another. Samples separated by distances in excess of this range are said to be
spatially independent (not correlated). The range signifies the distance beyond
which sample data should not be considered in the interpolation. A larger range
means more continuous behavior, the variable is very well correlated and
predictions result in fairly smooth maps. The range typically tends to increase as
more and better data become available.

Sill
When the distance between sampling points is zero the value at each point is
compared to itself. Therefore the difference is zero and the calculated
semivariance for γ(0) should be zero. If h is a small distance, the points being
compared tend to be very similar, and the semivariance will be a small value. As
the distance h increases, the points being compared are less and less closely
related to each other and their differences become larger, resulting in larger
values of γ(h) . Spatial variability in most environmental data between sample
pairs increases as the separation distance increases.

At some distance the points being compared are so far apart that they are no
longer related to each other, and their squared differences become equal in
magnitude to the variance around the average value. The semivariance no longer
increases and the semivariogram develops a flat region called the sill. The sill
indicates that there is no further correlation between pairs of samples. The higher
the sill value, the higher the prediction variances.
The semivariogram sill is theoretically equivalent to the variance of the spatial
samples – however the sample variance is often an inappropriate measure of the
population variance (Clark, 1979b, Rossi, et al. 1992; Isaaks and Srivastava
1989). The standard formula for estimating the variance assumes independent
data, which is invalid in most environmental situations (Goovaerts, 2002, Barnes,
1991). In general if a sill is clearly present its value should be used as an estimate
of the population variance, and the sample variance should not be used as an

Geostatistics Without Tears 53

 estimate of the variogram sill (Barnes, 1991). If the following conditions are met,
then the sample variance is a reasonable approximation for the variogram sill
(Barnes, 2005):

• the data are evenly distributed across the area of interest; there is no
significant trend in the data across the area of interest; and

• the dimension of the area of interest is more than three times the effective
variogram range.

Nugget
When the semivariogram is plotted back to the zero lag distance the nugget effect
(nugget variance) is found. The nugget effect would be expected to be zero since
two samples from the same point should have the same value. If the intercept is
greater than zero then a random or unstructured component is present. No matter
how close sampling is done the difference between neighboring samples will
never be zero.

The term nugget effect was coined from gold mining. In most gold deposits the
nugget effect tends to be quite large due to the nuggety nature of the
mineralization, so that samples taken close together can potentially have very
different grades. No matter how close the samples get, there will be large
differences in value between the samples because of the nuggety occurrence of
the gold. The value of the nugget effect will be close to zero in those deposits that
have a very uniform grade distribution, such as porphyry coppers (Surpac, 2004).

The nugget effect represents the inherent variability of the data which can be
interpreted as sampling errors, measurement errors, reproducibility issues, short
scale variations or random variation at distances smaller than the sampling
interval (white noise). A nugget structure increases the variability uniformly
across the entire variogram because the variability is not related to distance or
direction of separation. Increasing the nugget effect inflates the prediction error
(kriging variance) and hence increases uncertainty of estimated values (Siska &
Kuai Hung, 2001). Specifying a nugget effect causes kriging to become more of
a smoothing interpolator, implying less confidence in individual data points
versus the overall trend of the data. The higher the nugget affect the smoother the
resulting grid.

Ideally the nugget effect should be as small as possible. A large nugget effect
denotes unreliable data and more effort should be put into getting the data correct
rather than using the unreliable data with large nugget effect (ACE, 2005). A
high nugget variance indicates large point-to-point variation at short distances
and increased sampling often reveals more detail in structure. The nugget effect
tends to increase with lag tolerance and with data scarcity. Typically, the nugget
effect decreases, as more and better data become available.

The way in which the sample variogram behaves at near zero separation
distances is critical in describing the spatial continuity. A variogram that consists
of pure nugget effect indicates complete lack of spatial dependence. It is not

Geostatistics Without Tears 54

possible to get a nugget effect that is different in different directions. If there is
an apparently higher nugget effect in one direction than another, there is most
likely a short range component for which the sampling effort is not dense enough
to pick up (Clark and Harper, 2000).

There are two possible ways in which geostatistical software will deal with the
semivariogram model at zero distance:

1. force the model to go through 0 and a zero distance ( γ(0) =0)

2. allow the model to hit the vertical axis (γ(0) = nugget effect.
The first option makes kriging an exact interpolator and the sample location will
be honored and the kriging variance will be zero. This is what Matheron
srcinally specified and is described in early test books. The second option means
that kriging will not exactly honor the data but will put most of the weight on the
sample and some weight on the other samples.

If the software only allows the second option the only way to honor the sample
values is to have a zero nugget effect. The nugget effect does not have to be
removed from the model, but another component (perhaps spherical) is added to
the model where the sill equals the real nugget effect and whose range of
influence is just below the closest sampling spacing. If it is not known which
option is implemented in the software, run the interpolation with the nugget
effect and with this alternative. If there is no difference in the results the software
does the first option.

If the sampling
combination canerrors can where
be used be quantified and the
short range second(for
spherical method is implemented
example) replaces thea
smaller scale variability and the nugget effect reflects the ‘true’ replication error.
The choice to filter out the replication error by removing the nugget effect from
the model must be made. It is important to keep in mind that when using the
second method and/or removing the nugget effect when kriging, the calculated
variances will be too low by a factor of 2*nugget effect. If the nugget effect is
divided as suggested, the kriging variance will be too low by a factor of
2*replication error (Clark, 2006b).

Species of North American birds censused on the breeding bird survey vary
considerably in their nugget variances (Villard and Maurer, 1996). Temperate
migrants tend to have higher nugget variances than Neotropical migrants, and
species that use a number of different habitat types have higher nugget variances
than species using a single type of habitat. For estimates of abundances obtained
from censuses like the breeding bird survey, part of the nugget variance could be
due to the inherent variability among observers in counting birds at different
breeding bird survey routes. The second component is due to the discontinuous
nature of the process in space. For birds, this would be related to the continuity of
the habitat in which they were being counted. Since the habitat of a species is
generally discontinuous in space, this is a likely component of the nugget
variance for data like the breeding bird survey.

Geostatistics Without Tears 55

Variance Cloud
The variance cloud (variogram cloud) is a graph of the semivariances plotted
against separation distance which is used to indicate how well the interpolated
values fit the defined semivariogram model. The variance cloud can be used to
refine the model. Variance clouds are specific to both direction and to a particular
lag class. Spatial autocorrelation quantifies the basic principle of geography that
things that are closer are more alike than those farther apart. Pairs of locations
that are closer (far left on the x-axis of the variogram cloud) should have more
similar values (low on the

y-axis
cloud). of As
the pairs
variogramof
locations become farther
apart (moving to the right
on the x-axis of the
variogram cloud), pairs of
locations should become
more dissimilar and have a
higher squared difference
(high on the y-axis of the
semivariogram cloud). The
variance cloud is useful for
discovering outliers.

Scattergram

The scattergram (or scatter plot) is an XY representation of two variables used to

analyze the spatial continuity of the data by displaying all the pairs of samples
which are separated by a certain distance along a given direction. The
coordinates correspond to the value of the first variable at the first sample
location versus the value of the second variable (which can be identical to the
first one) at the second
sample location. The shape
of the cloud of points
spreads out as the spatial
correlation between the two
samples decreases or the
relationship between the two
variables weakens. The
stattergram is specific to
both direction and to its lag
class. The stattergram can be
used to look for outliers.

Geostatistics Without Tears 56

 Semivariogram Models
The semivariogram provides a spatial picture of the data. Choosing which
variogram model to be applied to the semivariogram is an art. Selecting the
appropriate variogram model requires many correct decisions based on a solid
understanding of the data and the underlying processes from which the data are
drawn. The shape of the semivariogram function is typically one of four forms:
linear, spherical, exponential, or Gaussian.

The major
for the purpose of between
relationship fitting a model
values isatto specified
give to provide a mathematical
distances. formula
Semivariances are
estimated at discrete values of h, whereas the true variogram is continuous. The
estimates are subject to error, and unless a large sample is taken the experimental
variogram will appear erratic. The better correlated the variable the smoother the
interpolated surface. Semivariograms can be a mix of two or more models.

The model selected influences the prediction of the unknown values, particularly
when the shape of the curve near the srcin changes significantly. The steeper the
curve as it nears the srcin, the more influence the closest neighbors will have on
the prediction. As a result, the output surface will be less smooth. To consider the
effect of the model on predictions, think in terms of the shape of the variogram
for early lags

Linear

e
The linear model is the simplest model for a semivariogram and applies when the
c
n spatial variability increases linearly with distance and never levels off. The linear
a
r
a
v
model indicates that the greater the separation of two samples, the greater the
m
e
difference in the two samples. The linear model does not have a plateau and may be
s
considered to be the beginning of the spherical or exponential model (Matheron,
distance 1963a, 1965). The range is defined arbitrarily to be the distance interval for the last
lag class in the variogram. Because the range is an arbitrary value it should not be
compared directly with the ranges of other models. There is no sill, and the sill is
the calculated semivariance for the arbitrarily defined range.
e
c
n The linear/sill model is similar to the linear model except that at some distance
a
r
a (the range), pairs of points will no longer be autocorrelated and the variogram
v
m
e will reach a sill. This model should not be used unless the variation is limited to
s
one dimension (Webster, 1985).
distance
The semivariogram immediately takes its maximum value (represented by a flat
semivariogram) if there is no correlation. In this case the phenomenon is

e
c completely
n
a the scale of random
sampling.andMeasurement
there is no spatial
errorsdependence
or samplingamong thefar
sites too observations at
apart (and as
r
a a result are spatially independent) or both may lead to noisy semivariograms that
v
m appear as pure nugget effect. Choosing a pure nugget effect model is an extreme
se
modeling decision that precludes use of kriging.
distance

Geostatistics Without Tears 57

 Spherical
The spherical model first proposed by Matheron (and sometimes referred to as
e the Matheron model) is based on the concept that a sample has a sphere of
c
n influence around it. Samples within this sphere have values that are related to the
a
r
a value at the central point. Imagine a second point with its own sphere of
v
m influence. If the spheres do not touch there is no relationship between the values
e
s
at the two central points. If the spheres overlap, there will be a relationship, and
distance the more the spheres overlap, the stronger the relationship. The spherical
semivariogram is the simple geometric calculation for the volume of non-overlap

of the two spheres, given the distance between their centres (Clark, 2006a)
The spherical model shows a progressive decrease of spatial autocorrelation
(increase of semivariance) of the variable with increasing distance. Dependence
fades away altogether when the model reaches a sill. The spherical model is a
modified quadratic function for which at some distance (the range) pairs of points
will no longer be autocorrelated and the semivariogram reaches a sill (becomes
asymptote). The spherical model is one of the most commonly used models.

Exponential
The exponential model was developed to represent the notion of exponential
decay of “influence” between two samples. The exponential model applies
e
when the spatial dependence decreases exponentially with increasing distance.
c
n
a
This dependence disappears completely at infinite distance. The exponential
r model is similar to the spherical model in that it approaches the sill gradually,
a
v
m but different from the spherical model in the rate at which the sill is
e
s approached and in the fact that a definite sill is never actually achieved. The
distance
exponential model reaches 98% of its sill value at around three times the
range of influence. The exponential model is also a commonly used model.

Gaussian
The Gaussian (hyperbolic) model is similar tothe exponential model but assumes a
e
c
n gradual rise for the y-intercept. The Gaussian semivariogram model is named
a
r because the formula is basically identical to that for the normal (Gaussian)
a
v
m probability distribution. Spatial dependence vanishes only at an infinite distance.
se This model represents phenomena that are extremely continuous or similar at short
distance
distances. The main feature of this model is its parabolic shape at the srcin. It
expresses a rather smooth spatial variation of the variables. The Gaussian model
does not have a true range but rather a parameter used in the model to provide
range (called the effective range). For a Gaussian model the range of influence is
square root of three times the distance scaling parameter. The Gaussian model is
generally unstable in the presence of a nugget effect or of very dense data. This sort
of model occurs in topographic applications or where samples are very large
compared to the spatial continuity of the values being measured.

Geostatistics Without Tears 58

 Hole Effect
The hole effect (holesin) model suggests repetition in the variation that is neither
e
c wholly random nor periodic. This model was developed to represent a cyclic or
n
a
r
a
periodic relationship between two samples. In practical circumstances, the
v
m
relationship usually tends to trail off, but sometimes a dampening or decay
e
s parameter is included otherwise the cyclic effect would continue on to infinity.
The hole effect model is not mathematically stable and can lead to some weird
distance
results if not used with care. Cyclicity may be linked to underlying periodicity or

could be due to limited data.

Semivariograms from breeding bird survey data typically do not indicate the
existence of an asymptotic maximum semivariance (Maurer 1994). Rather, they
indicate that the maximum semivariance occurs at intermediate distances, and
that points farther apart than this distance are actually more highly correlated.
This increasing correlation with large distances separating sampling sites can be
explained by the characteristic distribution of abundances across a geographic
range. Generally a species is most abundant in one or a few places near the center
of its geographic range, and its abundance decreases away from these points
(Maurer and Villard, 1994). Abundances at points relatively large distances away
from one another should be positively correlated because they represent
peripheral sites in the geographic range.

Abundance may be determined by a number of factors, each operating at a

different scale. For example, large-scale patterns in climate variation across a
continent may determine large-scale patterns of vegetation productivity, and
consequently variation in habitat for birds. Small-scale variation exists among
sites located in the same biotic province, such as the physiographic regions used
in the breeding bird survey. Sites that occur within the same physiographic
regions, for example, might have abundances more similar than those that fall in
different regions.

Geostatistics Without Tears 59

Anisotropic Variograms
Spatial autocorrelation can depend on more than the distance between two
locations. Directional influence occurs when there is an effect that causes
autocorrelation in one direction to be different from autocorrelation in another
direction. Both isotropic (omnidirectional) and anisotropic (directional)
variograms are examined to determine if autocorrelation is dependent on
direction. Anisotropy alters the basic principle of geography to . . . things that are
close together are more alike for distances in a certain direction than those in
other directions. The isotropic variogram curve should be used to detect or
confirm lag distance parameters to establish a clear continuity before examining
anisotropic variograms.

Typically four anisotropic variograms with

different orientations are constructed. These
orientations are usually E/W (0°), NE/SW (45°),
N/S (90°), and NW/SE (135° or -45°). The
directional component is set at one of these
directions, plus or minus a directional tolerance
angle. Sometimes eight spatial directions
analysis is used to yield more directional
resolution. Larger data sets are usually required
to yield enough information for the eight
directional calculations.

If the directional variograms are reasonably

similar then the spatial correlation can be
assumed to be isotropic in which case the
variability between samples would be a function of distance only. When the
experimental variograms show different behaviors in different directions an
anisotropic variogram model should be used since spatial correlation varies with
direction and distance. For example, anisotropy associated with plant diseases
can be interpreted as indicative of directional spread (Matheron, 1963a,
Schotzko, et al., 1989, and Tangmar, et al., 1985.

Nugget Variance
Nugget variance should be viewed as an isotropic parameter because it is not
physically meaningful to say that uncorrelated (random) variability at a point is
different depending on the direction away from that point. If the apparent nugget
in two directional variograms is substantially different the data should be
modeled as an additional anisotropic nested (non-nugget) structure with a
maximum range so large as to add negligibly to the variogram at small lags
(Welhan, 2005).

Anisotropic Variogram Surface

The anisotropic variogram (or variogram map) provides a visual picture of the
semivariance in every direction by plotting the semivariances versus distance and
direction. Each pair of samples corresponds to a distance and direction

Geostatistics Without Tears 60

 (converted to a grid cell), and to a variability (the cell value). Each location on
the graph represents an approximate average of the pairs semivariance for the set
of pair separation distances and directions, see Isaaks and Srivastava (1989; page
150) and Goovaerts (1997; page 98).

The anisotropic variogram is a good visual tool to highlight possible anisotropy

in the data. The anisotropic variogram surface is used to find the axis of
maximum spatial dependence (the lowest semivariances at a given distance) by
viewing the anisotropic variogram surface to define and set the principal
anisotropic axis to the direction aligned with the lowest semivariance values (the
direction of maximum spatial continuity, or major axis of the anisotropic variogram
model). The principal axis is the direction of maximum spatial continuity, or base
axis from which the offset angles for anisotropic analyses are calculated. The axis
orientation should correspond to the axis of maximum spatial continuity – the
major anisotropic axis.

Geometric Anisotropy
Geometric (or affine) anisotropy is a situation where the variogram exhibits
different ranges in different directions and the sill is constant in all directions. A
longer range in one direction means that the values are more continuous in that
direction than for a direction with a shorter range (Liebhold et al., 1993). For
example, in an Aeolian deposit, permeability might have a larger range in the
wind direction compared to the range perpendicular to the wind direction (Syed,
1997).
The anisotropy ratio is the ratio between the smallest range and biggest range
(these directions are assumed to be approximately perpendicular to each other). A
ratio of one denotes an isotropic variogram (Syed, 1997). The ratio between the
lowest range and highest range, and the difference between their two angles can
be used to transform the model to an isotropic model suitable for kriging (Ricci,
A.K., 1998).

Geostatistics Without Tears 61

Zonal Anisotropy
Zonal anisotropy is a situation where the range remains constant in all directions
but the sill varies with direction (see Isaaks and Srivastava, 1989, Figure 16.7 for
a graphical representation). A variable sill means that the magnitude of spatial
variation changes with direction (Liebhold et al., 1993). For example, a
variogram in a vertical well bore typically shows a bigger sill than a variogram in
the horizontal direction (Syed, 1997).

Since the sill of a semivariogram is equal to the population variance it is

theoretically impossible to have different sills in different directions. A sill that
varies with direction is likely a symptom of deeper problems such as non-
stationarity, trend, discontinuities, etc. (Isobel Clark, 2001a). If different sills are
seen in different direction look for one of the following:

• the total sill has not been reached one direction;

• the data is skewed – a transformation should be tried;

• the assumption of homogeneity may have been violated – the sampling

field may be crossing from one population group into another – the
semivariogram in the direction of maximum variation will have the
steepest slope, whereas the semivariogram in the direction of minimum
variation will have the lowest slope (Issaks, et. al., 1989, Larkin, et. al.,
1995, Trangmar, et. al., 1985, and Xiao, et. al., 1997);

• there may be a trend in the data

If an explanation cannot be found that fits one of these points or something

similar there may be true zonal anisotropy, although true zonal anisotropy never
occurs (Clark and Harper, 2000). Some variograms are a combination of both
geometric and zonal anisotropies. For example geometric and zonal anisotropy is
present in highly stratified conditions with strong horizontal layering (Onsoy, et
al., 2005).

Geostatistics Without Tears 62

Alternate Variograms and Autocorrelation Measures
Alternative variograms and autocorrelation measures can provide additional
estimates of spatial variability because these procedures are sometimes less
sensitive to outliers, skewed distributions, or clustered data than ordinary
variograms and these methods may aid in the recognition of data structure when
the ordinary variogram is too noisy.

Correlogram
The correlogram shows the correlation among sample points by
plotting autocorrelation against distance. Correlation usually
decreases with distance until it reaches zero. A correlogram mirrors
the variogram with the expectation that the correlogram assumes a
stationary mean (Srivastava, 1996). As a variogram gradually rises to
a plateau the correlogram gradually drops and levels out.
Correlograms can be used to detect and describe regular spatial
patterns between patches (Radeloff, et al. 2000). For a repeating
pattern of patches separated by some distance, the patch size appears
as a peak in autocorrelation for distances up to average patch size; the
distance between patches appears as a second peak.

Covariance
The covariance (like the semivariogram) measures the strength of
statistical correlation as a function of distance – when two locations
are close to each other they are expected to be similar and so their
covariance will be large. As the two locations move further apart,
they become less similar and their covariance eventually becomes
zero. Since covariance decreases with distance, covariance can be
thought of as a similarity function.

Cross-K
The Cross-K (Cressie, 1993) is similar to Ripley’s K. A point from the first data
set is chosen. A radius ( h) is delineated and the points from the second data set
that are within the circle are counted. A line above that designated by the
interaction between two random point sets would imply that the data points are
attracted to each other or tend to exist together in the same place. A line below
shows that the points repel each other or tend to exist in opposing space.

If Monte-Carlo envelopes were constructed around a K representation and a

random point process was simulated and the K function for each was determined

Geostatistics Without Tears 63

 a plot of the maximum and minimum values at each distance (h) can be
constructed. The envelope around the expected random K function represents a
99.9% confidence interval that the data are randomly distributed. If the K
function from the observed data lies outside of this envelope there is very strong
evidence that the data exhibit non-randomness.

Drift
Drift (also called trend) is calculated as the difference between
values at different sample locations within a particular lag
distance. Drift function is used to assess non-stationarity – does
the mean value change over distance).

Fractal
Mandelbrot introduced the term fractal for temporal or spatial
phenomena that are continuous but not differentiable, and that

exhibit partial correlation

fractal properties overaccepted
have been many scales. Since
widely by the mid-1980’s
geologists and
geophysicists as a quantitative characterization of complex
geological phenomena and a fractal law has been accepted for
many geological parameters. Many environmental variables
suggest that not only are they fractals, but that they may have a
wide range of fractal dimensions, including values that imply that
interpolation mapping may not be appropriate in certain cases
(Burrough, 1981).

In ecology, fractal theory has been used mostly as an analytical

tool either to identify characteristic scales (Frontier, 1987) structural hierarchy
(Burlando, 1990; Collins and Glenn, 1990) or to assess chaotic behavior
(Sugihara and May, 1990) of phenomenon showing a spatial distribution or a
temporal dynamics. One of the main appeals of fractals to ecologists lies in their
ability to summarize the complexity and heterogeneity of a spatial or temporal
distribution in a single value, the fractal dimension that is purportedly
independent of scale (Leduc, et al., 1994).

The fractal dimension is directly related to the slope of the best-fitting line
produced when the log of the distance between samples is regressed against the
log of the mean-squared difference in the elevations for that distance. A graphical
representation of variations of fractal dimension values as a function of scale has
been called a fractogram (Palmer, 1988). According to Burroughs (1983), the
slope of a variogram model plotted on log-log axes can be translated directly into

Geostatistics Without Tears 64

 a fractal (Hausdoff-Besicovitch) dimension. A linear variogram model on log-log
axes would therefore indicate constant fractal dimension, and lead to simulations
with similar textures at all scales (Englund, 1993). When a variogram is used as a
means of determining the fractal dimension of the land surface, the normal
practice of removing non-stationarity in the data should not be performed
(Armstrong, 1986; Cressie and Hawkins, 1980). Fractal methods are not widely
used in the mining industry (Vann, et. al., 2002).

Geary’s C

Geary’s C (Geary, 1954) statistic is similar to Moran’s I. Geary’s C statistic is a

squared difference statistic for assessing spatial autocorrelation. The values of C
typically vary between 0 (similar) and 2 (dissimilar) although 2 is not a strict
upper limit (Griffith, 1987). If values of any one location are spatially unrelated
to any other location, then the expected value of C would be 1 (the theoretical
value). Values between 0 and 1 indicate positive spatial autocorrelation while
values between 1 and 2 indicate negative spatial autocorrelation. Gear’s C is
inversely related to Moran’s I. Geary’s C compares the squared differences in
values between the center cell and its adjacent neighbors to the overall difference
based on the mean of all the values. If the adjacent differences are less, then there
is a positive correlation. If the differences are more, then there is a negative
correlation. And if the adjacent differences are about the same, the variables are
unrelated. See Comparison of Moran’s I and Geary’s C below.

Inverted Covariance

The inverted covariance (InvCov) is based on estimates of covariance rather than

variance and is calculated by subtracting lag covariance from the sample variance
(Srivastiva, 1988). This approach compensates for cases where the mean of the
sample pairs in one direction is not the same as the mean of the sample pairs in
another direction. The inverted covariance variograms have the same units
(measurement units squared) as ordinary variograms and may be modeled and
used for kriging in the same way.

Madogram
The madogram is a graph of mean absolute difference of sample
measurements as a function of distance and direction.
Madograms are less sensitive to extreme data values and can be
useful for inferring range and anisotropy of data sets with outlier
values that may make variogram features difficult to discern
using traditional methods (Sobieraj, Elsenbeer and Cameron,
2003). The madogram should not be used for modeling the
nugget of the semivariograms (Deutsch and Journel, 1992;
Goovaerts, 1997).

Geostatistics Without Tears 65

Moran’s I
Moran's I (Moran, 1948 and 1950) is a weighted correlation
coefficient used to describe autocorrelation by comparing the
value of a variable at one location with values at all other
locations. Moran’s I provides an indication of spatial patterns
such as clusters and geographic trend. The semivariogram is a
measure of dissimilarity while the Moran’s I is a measure of
similarity. Moran’s I is one of the oldest indicators of spatial
autocorrelation. Moran’s I varies between -1 and +1. Moran’s I
shows a large positive value (close to 1) if a strong spatial
correlation exists. Moran’s I becomes negative (close to -1) if
the spatial autocorrelation is negative. Although it can be hard to
detect spatial autocorrelation for shorter lag distances due to the
lack of samples, a strong Moran’s I for smaller distances suggest
a lower nugget effect. Moran’s analysis on a data set at different spacing
(resolutions) provides insight on whether there is more than one controlling
factor (Petrone et al., 2004).

No variography or kriging should be carried out for data with a zero Moran’s I
for all lag distances since this does not respect Tobler’s Law. Deeper slopes in
the Moran’s I correlogram indicate a shift from a spherical to an exponential
model while flat profiles with a high spatial autocorrelation reveal a shift from a
spherical to a Gaussian one (Garrott, 2003).

Comparison of Moran’s I and Geary’s C

Moran’s I is slightly more robust than the Geary’s C but Geary’s is often used as
well. Moran’s I and Geary’s C statistics are essentially equivalent and
interchangeable. Moran’s I gives a more global indicator whereas Geary’s C is
more sensitive to differences in small neighborhoods.

The general interpretation of the Geary’s C and Moran’s I statistics can be

summarized as:

autocorrelation Moran’s I Geary’s C

strong positive autocorrelation I>0 0<C<1

random distribution of values I=0 C=1

strong negative autocorrelation I<0 1<C<2

Geostatistics Without Tears 66

Relative Variograms
Often there is dependence between the semivariance and the mean of the data
value for each lag distance. This leads to a proportional effect that complicates
the interpretation of analysis. To detect the proportional effect a moving window
statistic can be used. Relative variograms are analogous to the relative standard
deviation often used to measure analytical variability. Relative variograms tend
to be smoother than normal variograms and may aid in resolving the structure of
the variogram. Relative variograms can be calculated only for data sets for which
all Z value are positive. Relative variograms are unitless (decimal fraction

squares).
deviationsWhen modeled
must be and by
multiplied used
thefor kriging values
estimated the relative kriging standard
to be comparable with
kriging standard deviations produced with ordinary variogram models.

Local
The local relative variogram considers the data within each region as a separate
population and the values are proportional to their means. To describe the overall
spatial continuity the local variogram values are scaled by their means. This
method is rarely used because of the difficulty of determining the regions.

General
The general relative variogram weights the semivariance by the
mean value of the data by dividing each semivariance by the
squared mean of all samples used to estimate the semivariance.
General relative variograms are less sensitive to clustering and
data outliers and therefore may provide cleaner estimates of
spatial correlation (Deutsch and Journel, 1997, p.45).

Pairwise
The pairwise relative variogram calculations are adjusted by a
square mean performed for each pair of sample values by dividing
each sample pair’s difference by its mean. Experience has shown
that the general relative and pairwise relative semivariograms are
effective in revealing spatial structure and anisotropy when the

scatter points
relative are sparse
variograms should (Deutsch & on
only be used Journel, 1992).
positively Pairwise
skewed data
sets. Pairwise relative variograms are less sensitive to clustering
and data outliers and therefore may provide cleaner estimates of
spatial correlation (Deutsch and Journel, 1997, p.45).

Geostatistics Without Tears 67

 Rodogram
The rodogram is similar to the traditional variogram except that
the square root of the absolute difference is used rather than the
square difference. The Rodogram is a robust variogram that is
useful for data with outliers.

Standardized
In standardized variogram analysis the variogram is computed on
the natural logarithms of the variates.

Ripley’s K
Ripley’s K (Ripley, 1977; Cressie, 1993) is a spatial statistics function that is
used to analyze spatial autocorrelation (usually clustering, rather than
dispersion). Ripley’s K is also called the reduced second moment measure – that
is it measures second order trends (local clustering as opposed to a general
pattern over the region. The function chooses a point within the data set and
counts all the other points within a distance of that point. Like any statistic
Ripley’s K is prone to biases which include: edge biases, sample size, scale of
interpretation, and the shape of the boundaries (Ned Levine & Associates, 2004).

Geostatistics Without Tears 68

KRIGING
Kriging can interpolate values for points at unsampled locations by using
knowledge of the underlying spatial autocorrelation provided by the
semivariogram to find an optimal set of weights to estimate the surface at
unsampled locations. The kriging process involves the construction of a weighted
moving average. Since the semivariogram is a function of distance, the weights
change according to the geographic arrangements of the samples. Low weights
are assigned to distant samples and vice versa. Kriging also take into account the
relative position of the samples to each other.

The Kriging Algorithm

The understanding of the kriging algorithm is beyond what is needed for a
pragmatic introduction to geostatistics, however the implementation of the
kriging algorithm can be summarized (Ramirez-Beltrán, et al., 2006). Given a set
of known points with each point having an X,Y coordinate value and an
associated Z variable value, and a set of unknown points with X,Y coordinates
and a Z variable value to be estimated.

The unknown values can be estimated as follow:

n

zij = Σ wijzi j = 1, …, m
i=1

where the weight values Wij can be calculated by solving the following system of
equations:
n

Σ uijwkj + λj = vij i=1, …, m j=1, …, m

k=1

Σ wki = 1 j = 1, …, m
k=1

vij uis the

where ij is the distance from the known point (x i,y i) to the known point (x j,y j)
and distance from the known point (x i,yi) to the unknown point (x j,yj)
and λj is a weight value that can be estimated using the calculated variogram with
the assumption that weights sum for each X should equal 1.

Geostatistics Without Tears 69

Kriging Variations
Four variations of kriging interpolation are ordinary kriging, simple kriging,
universal kriging, and kriging with an external drift. Any one of these kriging
methods can be applied as one of two forms: punctual or block.

Ordinary Kriging
Ordinary kriging is done when there is no underlying spatial trend (drift), the
mean of the variable is unknown and the sum of the kriging weights is equal to
one. This method assumes that the data set has a stationary variance but also a
non-stationary mean within the search radius. Ordinary Kriging is highly reliable
and is recommended for most data sets. One of the main issues concerning
ordinary kriging is whether the assumption of a constant mean is reasonable.
Sometimes there are good scientific reasons to reject this assumption.

Simple Kriging
Simple kriging assumes that the data set has no underlying spatial trend, a
stationary variance and a known mean value. Because the mean is assumed
known, it is slightly more powerful than ordinary kriging, but in many situations
the selection of a mean value is not obvious. Simple kriging also assigns a weight
to the population mean, and is in effect making a strong assumption that the
mean value is constant over the site. Simple kriging also requires that the
available data be adequate to provide a good estimate of the mean. Simple
kriging uses the average of the entire data set (a constant known mean) while
ordinary kriging uses a local average (the average of the scatter points in the
kriging subset for a particular interpolation point). Simple kriging can be less
accurate than ordinary kriging, but it generally produces a result that is
"smoother" and more aesthetically pleasing.

Universal Kriging
Universal kriging combines trend surface analysis (drift) with ordinary kriging
(Bonham-Carter, 1994). Universal kriging is more difficult to apply. Once trends
have been accounted for (or assumed not to exist), all other variation is assumed
to be a function of distance. This method represents a true geostatistical approach
to interpolating a trend surface of an area.

Universal kriging is a two-stage process.

1. Fit a trend (local or global) and calculate the residuals. From these
residuals obtain a de-trended (driftless) semivariogram.

2. Using the semivariogram derived in step one together with the

established trend do the kriging with the trend from the srcinal sample
values -- not from the residuals. The only time the residuals are used is to
develop the semivariogram model.

The recommended setting for the trend is a first degree polynomial which will
avoid unpredictable behavior at the outer margins of the data set (Loránd, 2005).

Geostatistics Without Tears 70

 The problem with this method is that the semivariogram itself is sensitive to the
form of the deterministic surface. It is possible to over fit the trend surface, which
does not leave enough variation in the random errors to properly reflect
uncertainty in the model. When used properly universal kriging is more powerful
than ordinary kriging because it explains much of the variation in the data
through the nonrandom trend surface.

Kriging with External Drift

Kriging with an external drift (KED) is an extension of universal kriging in that
the external drift is intended to be defined by a secondary variable that may
incorporate some information of relevance to the primary variable. Universal
kriging assumes that the shape of the trend is known, but its magnitude (or
coefficients) is unknown. It is also possible to specify a trend mathematically but
the srcinal intent of kriging with external drift was to incorporate a secondary
variable relevant to estimation of the first.

Punctual and Block Kriging

Any one of the four kriging methods can be applied as either a punctual (point)
or block method. In punctual kriging the goal is to predict the value of a variable
over a specified region. In block kriging the goal is to predict the average value
of a variable over a specified region (a block).

Punctual Kriging
Punctual (point) kriging provides an estimate for a given point by estimating the
value of that point from a set of nearby sample values. Punctual kriging can be
used if sampling was done to represent point values in a field or in time.

Punctual kriging is an exact interpolator – estimated values are identical to

measured values where interpolated points coincide with sample locations. In the
presence of a nugget variance there will be local discontinuities at the sampling
points (Dorsel & La Breche, 1998). If the nugget variance is large, undesirably
large estimation variances may be produced. These discontinuities depend on the
particular locations where the samples were collected. Changing the sampling
pattern could result in a different map of kriged values and their estimation
variances. Therefore, results obtained by punctual kriging depend strongly on the
sampling methodology.

Despite being computationally simple, punctual kriging punctual has some

drawbacks: it does not work unless the variable being mapped is stationary; it is
not unbiased in the presence of a trend; the estimates shift up or down from the
true values depending upon the arrangements of points and the trend direction.
Block kriging eliminates some of these short comings (Castrignanò and Lopez,
2003).

Geostatistics Without Tears 71

 Block Kriging
Block kriging (block averages), can be used to estimate an attribute value over an
area instead of at one point location. Due to an averaging out effect, block
kriging variances are generally much smaller than corresponding punctual
kriging variances (Castrignanò and Lopez, 2003). Block kriging provides better
variance estimation and has the effect of smoothing interpolated results. Block
kriging is an approximate interpolator (based on a block of points) so that
estimated values are not necessarily identical to measured values where the
interpolated points coincide with sample locations.

A fine spatial resolution (small blocks) will require denser sampling than a
coarser resolution (large blocks) to predict block means at a given level of
precision. An optimal sampling scheme will attempt to find the combination of
block size and sampling spacing in order to achieve a given level of precision at
the lowest total cost (Castrignanò and Lopez, 2003).

Block kriging can be used if sampling was done to represent the area around the
actual sample point. For example if soil samples from an area around the
sampling location were composited before analysis then block kriging may be
more appropriate. Block kriging produces smoother maps than punctual kriging
by interpolating average values over blocks with the effect of smoothing local
discontinuities. These effects are particularly desirable when the main interest is
to study regional patterns of variation rather than local details (Castrignanò and
Lopez, 2003). Consequently block estimates appear more reliable than those for
points. In environmental work block kriging is usually more appropriate than
punctual kriging.

By computing block estimates, the nugget effect, which may be due either to
measurement error or very short-range variation, can be avoided. Block
interpolation may be more appropriate than punctual interpolation where average
values of properties are more meaningful than exact single-point values,
especially where spatial or temporal dependence is weak (Burgess and Webster
1980).

Kriging Search Neighborhood

The search neighborhood specifies which data points are included in the
weighted moving spatial average. The use of a search neighborhood limits the
estimates to just the data within some predefined radius of estimation called the
search radius. The search radius does not need to be a circle or a sphere. To
account for possible anisotropies in the data the search radius can be elliptical or
ellipsoidal (in 3D). Kriging weights become small for sample points at a distance
away from the location estimated and have a negligible influence especially if
there is a strong spatial correlation. Points beyond the sill contribute in no way to
the kriged estimate and so can be ignored. The use of the search neighborhood is
the assumption of stationarity within the search window (quasi-stationarity). This
concept can be useful if there is a trend in the data by assuming a small enough
search neighborhood that the trend component can be ignored.

Geostatistics Without Tears 72

 Cross-Validation
Cross-validation is used to check kriging model assumptions. After specifying
the variogram model, the search neighborhood and kriging the surface the
resulting kriged values and the true values are compared. The difference between
these two values is called the cross-validated residual. A graph is constructed of
the estimated versus the actual values for each sample location. Each point on the
graph represents a location in the input data set for which an actual and estimated
value are available.

1.250 Regression coefficient = .416 If all estimated values were the same as
(SE = .045, r2 = .399, y intercept = 0.222)
the true value the scatter plot would be
a 45-degree straight line. The distance
0.952
Z
each point is from the 45-degree line is
d the estimation error for that sample
et
a 0.655 value. The closer the cloud of points is
ts to this line, the better the estimation.
E The slope of the cross-validation scatter
0.358
plot of predicted versus true values is
usually less than one. It is a property of
0.060
kriging that it tends to under-predict
0.060 0.358 0.665 0.952 1.25 large values and over-predict small
Actual Z values. Systematic deviations from the
line indicate lack-of-fit with the model.
Individual points deviating from the line may be considered outliers.

Cross-validation is used to assess performance of the model assumptions used in

kriging the surface. Is the variogram model appropriate, is the search
neighborhood too small, are some regions being over or under estimated, should
a trend model be incorporated, are the prediction errors comparable? Cross-
validation does not prove that the variogram model is correct but rather that it is
not grossly incorrect [Cressie, 1990].

Geostatistics Without Tears 73

Kriging Predictors
There are a wide variety of predictors used in geostatistics.

Cokriging
Cokriging is a multivariate extension of kriging to allow the incorporation of two
or more variables that are correlated with each other by using one set of data to
help explain and improve the description of variability in another (Goovaerts,
1997, pp 203-248, Wackernagel, 1995). Cokriging assumes that the second data
set is highly correlated with the primary data set to be interpolated. For example
altitude is considered as an important additional variable when estimating air
temperature and hence cokriging should provide a better estimation than kriging
because cokriging would take altitude as an additional variable. As another
example, the quantity of vegetation (the biomass) may be related to elevation and
soil moisture, and if this relationship can be determined, it would be possible to
use cokriging to predict biomass based on both elevation and soil moisture.

Disjunctive Kriging
Disjunctive kriging (full indicator cokriging) is a specific nonlinear transformation
of the srcinal data that does not have simple (Gaussian) distributions (Rivoirard,
1994). The first step is to perform a normal score transformation of the data. The
next step is to express the normalized function as an expansion of hermitian
polynomials. Disjunctive kriging assumes all data pairs come from a bivariate
standard normal distribution. The bivariate normal distribution is weaker than the
multi-Gaussian condition. If bivariate normality exists, disjunctive kriging can
produce better results than other geostatistical methods. If not, another type of
kriging should be used. Disjunctive kriging tries to do more than ordinary kriging
and indicator kriging by considering functions of the data and requires making
stronger assumptions. The technique is computational daunting even with the use of
variogram modeling software.

Indicator Kriging
Indicator kriging (Journel, 1993, and Journel and Arik, 1988) is a method
developed for use on data populations with any distribution. The method is easy
to understand. A cutoff value is defined, and all data below the cutoff are
assigned a value of 0, all data above the cutoff are assigned 1. Variogram
modeling and kriging is done on binary (0 or 1) values. An advantage of
indicator kriging is to minimize the influence of outliers, to contend with skewed

distributions, to deal with populations of both high value and values below the
detection limit. Indicator kriging is not recommended for data having a trend.

Kriged indicator values can be viewed either as probabilities (the probability of

exceedence), with a higher value denoting a higher probability or vice versa; or
as proportions (the proportion of the attribute above the specified cut-off). The
final result of indicator kriging is a set of probabilities that a grid cell exceeds a
specific set of cut-off values with the values ranging between 0 and 1. This

Geostatistics Without Tears 74

 method is especially useful when a decision will be made based on a specific
cutoff, or if the probability of exceeding some threshold is the final use of the
interpolated data. Although geostatistical analysis is not specifically designed for
binary data, it is tolerant of such data (Chellemi et. al, 1988, Lecaust et. al., 1989,
Maison, et. al. 1976, Materson, 1963a, Tangmar, et. al., 1985).

Although sampling error is not taken into account the indicator transform can
improve estimation of spatial continuity parameters in two ways. By reducing the
scale of the variogram the behavior of the sample variograms frequently improves.
Being an arithmetic average kriging can be drastically affected by even a single
outlier. Deleting outliers is one approach to dealing with unusual values, but this
may not be desirable because data are thrown away. Alternatively an indicator
transform may be applied to the data, because no matter how unusual the data
values are the data are transformed to either 0 or 1. Since the indicators are 0 or 1
the indicator variogram is very well behaved and resistant to outliers. Indicator
variograms have a linear relationship (Lemmar, 1985).

When the threshold level does not correspond well with the continuity structure
the indicator variogram will not produce good result, for example, when a
skewed proportion of the data fall into one level relative to the other. Experience
shows transforming the data into indicators using the median of the distribution
as the threshold can provide a variogram revealing the approximate description
of variability which at times is a reasonable approximation (Myers, 1997).

Among the shortcomings of indicator kriging is the loss of information after

coding data through indicator functions. For example, if the data values are in the
range of 1 to 100 and the indicator value (cutoff) selected is 60, then the data
points with values 1 and 59 will be interpreted as being equivalent. The practice
of using what was developed as a linear statistical model for continuous data to
operate instead on discrete data (ones and zeroes) could be considered an
“engineer’s method” – it works well as long as you don’t think too hard about its
theoretical underpinning (Henley, 2001).

Lognormal Kriging
To cope with the presence of a few high grades South African mining
geostatisticians developed lognormal kriging, which is kriging applied to
lognormal transforms of data followed by a back-transform of final estimates
(Journe, Huijbregts, 1978). The lognormal hypothesis is very strict in that any
departure can result in completely biased estimates. Lognormal kriging is rarely
used in mining applications today (Vann and Guibal, 2000). Some environmental
parameters are lognormally distributed, and in such cases, lognormal kriging may
be useful. The lognormal transform is referred to as a “sledgehammer” approach
to data manipulation and can significantly reduce or even eliminate true
distribution characteristics of the data (Myers, 1997). It is also believed that log
transforming data often goes too far, especially if data are not srcinally
lognormally distributed (Cressie and Hawkins, 1980).

Geostatistics Without Tears 75

Median Indicator Kriging
Median indicator kriging is an approximation of multiple indicator kriging which
assumes that the spatial continuity of indicators at various cut-offs can be
approximated by a single function.

Multiple Indicator Kriging

Multiple indicator kriging can be used if the data has a strange distribution or is a
mixture of different populations. Multiple indicator kriging involves kriging of

indicators
recoverableatresource
several estimation
cut-offs. Multiple indicator
which is robust kriging values
to extreme is an and
approach to
practical
to implement (Vann and Guibal, 2000).

The distribution histogram, probability plot, and other characteristics of the data
should be examined before resorting to the complicated multiple indicator
approach. If a decent semivariogram is obtained from ordinary values, indicator
kriging is not needed. Theoretically multiple indicator kriging gives a worse
approximation of the conditional expectation than disjunctive kriging, which can
be shown to approximate a full cokriging of the indicators at all cut-offs, but does
not have the strict stationarity restriction of disjunctive kriging (Vann and
Guibal, 2000).

Isofactorial Disjunctive Kriging

There are several versions of isofactorial disjunctive kriging.

Gaussian Disjunctive Kriging
By far the most common is Gaussian disjunctive kriging which is based on an
underlying diffusion model where the variable tends to move from lower to
higher values and vice versa in a relatively continuous way. The initial data are
transformed into values with a Gaussian distribution, which can easily be
factorized into independent factors called Hermite polynomials (Rivoirard,
1994). Gaussian disjunctive kriging has proved to be relatively sensitive to
stationarity decisions (Vann and Guibal, 2000).

Uniform Conditioning
Uniform condition is a variation of Gaussian disjunctive kriging more adapted to
situations where stationarity is not very good. In order to ensure that the
estimation is locally well constrained, a preliminary ordinary kriging of relatively

large
value.panels is made,
Uniform and the is
conditioning proportions per panelbut
a robust technique aredoes
conditional
depend to that kriging
heavily on the
quality of the kriging of the panels.

Geostatistics Without Tears 76

 Probability Kriging
Probability kriging (Sullivan, 1984) represents an attempt to alleviate the order
relationship problems associated with multiple indicator kriging by considering
discretized (0 and 1) standardized rank transforms (the variables are sorted by
increasing value and assigned their cumulative frequency which is referred to as
the rank order transform) in addition to the indicator values. Probability kriging
is thus cokriging between the indicator and the rank transform of the data.

Since cokriging must be done for a discrete number of cutoffs, there are a large
number of variograms required to implement probability kriging. If ten cutoff
values are used, then twenty-one variograms are necessary. The hybrid nature of
this estimate as well as the time-consuming complexity of the structural analysis
makes probability kriging rather unpractical (Vann and Guibal, 2000). Since
probability kriging uses additional information it should theoretically produce
better estimates than indicator kriging (Knudsen, 1987). Probability kriging is not
recommended for data having a trend.

Regression Kriging
Regression kriging (RK) is a spatial interpolation technique that combines a
regression of the dependent variable on auxiliary variable (such as terrain
parameters, remote sensing imagery and thematic maps) with kriging of the
regression residuals. Regression kriging is mathematically equivalent to
interpolations methods such as universal kriging and kriging with external drift
(KED) where auxiliary predictors are used directly to solve the kriging weights
(Hengl, et. al., 2003). Hengl (2006) has a web page that presents step-by-step
instruction for running regression kriging in a statistically sound manner – called the
generic framework for regression-kriginghttp://spatial-analyst.net/regkriging.php
( ).

Residual Indicator Kriging

Residual indicators (Rivoirard, 1994, chapter 4 and 13) are one way to cokrige
indicators by separately kriging independent combinations of the indicators and
then recombining these to form the cokriged estimate. The residuals are defined
from indicator functions. In practice the residuals are calculated at each data
point, their variograms are then evaluated and independent kriging is performed.

Kriging and Remotely Sensed Images

Kriging has been found to be useful for assessing spatial patterns in remotely
sensed images (Curran, 1988; McBratney and Webster, 1981; Ten Berge et al.,
1983; Webster et al., 1989; Webster and Oliver, 1992; Woodcock et al., 1988a;
and Woodcock et al., 1988b). Variograms of remotely sensed images should be
interpreted with care because these variograms may differ from variograms
resulting from ordinary samples. In remote sensing, the support size (support is a
geostatistical term used to describe the size, geometry and orientation of the
space on which an observation is defined) is equivalent to the spatial resolution.
For example, reflection values are average over the field of view or pixel size of

Geostatistics Without Tears 77

the measuring device. The shape and distribution of elements in the image
influence variograms of data collected by remotely sensed devices.

According to the authors cited above some of the major points for variogram
interpretation are:

• the range presents information on spatial dependence or reflectance – the

range is related to sizes of objects in the terrain (groupings of shrubs)

• the variogram model or shape of the variogram reveals information on

the spatial of
variability behavior
the sizeofofthe data –inthe
objects theshape
terrainof the variogram is related to

• the sill gives information on the total variability – the height of the
variogram is influenced by the density of coverage of the objects and the
spectral differences between the objects

• the nugget reveals information on variability between adjacent pixels

• as the spatial resolution become coarser the overall variance of the data is
reduced and the fine scale variation becomes blurred – consequently, the
sill height will reduce, the range will increase, and the nugget will
increase

• anisotropy in the image is expressed by the variation of variogram

parameters with the direction of the transect

Although the variogram seems to be a robust tool, a number of disadvantages of

variograms can be identified. A fractal approach to assess spatial patterns from
images provides an easier and more rapid method to assess spatial patterns from
remotely sensed images (Ardini et al., 1991, Burrough, 1993, De Cola, 1989, de
Jong, 1993, de Jong, et at. 1995, Jones et al., 1989, LaGro, 1991, Lovejoy, 1982,
Vasil’yev and Tyuflin, 1992, Walsh et al. 1991).

Geostatistics Without Tears 78

GEOSTATISTICALCONDITIONAL SIMULATION
Spatial uncertainty modeling has been a growing area of research in geostatistics for
the last couple of decades (Goovaerts, 1997, 2001 and 2006; Chiles and Delfiner,
1999). Geostatistical conditional simulation (stochastic simulation) is more
computationally demanding than geostatistical estimation. Increasing computer
processor speed, memory and data storage capacity have provided an opportunity to
apply these techniques. Geostatistical conditional simulation is a spatial extension of
the concept of Monte Carlo simulation and provides a method to quantify
uncertainty and to minimize risk. Conditional simulation is an interpolation
technique that builds many realizations in which known data values are honored at
their locations (hence ‘conditional’). Each realization is different to others because
there is always uncertainty away from known data points, hence individually such
realizations are simulations not estimates. A conditional simulation passes through,
or honors, the known data (hence it is said to be conditioned by the data).

A geostatistical simulation model is different than geostatistical estimation

(kriging) in that a family of model realizations is generated. A series of
realizations are produced that presents a range of plausible possibilities. The
plausibility of these realizations is dependent on the assumptions and
methodology of the simulation process. There is a clear distinction between
methods where conditioning is built in (sequential methods) and where
conditioning takes place as a separate kriging step (turning bands).

Each realization
from known datais points
different
andto so
others because there
individually is always
realizations areuncertainty
simulationsaway
not
estimates. Simulation has different objectives to estimation. The point of
simulation is to reproduce the variance of the input data, both in a univariate
sense via the histogram and spatially via the variogram. Thus simulations provide
an opportunity to study any problem relating to variability, for example risk
analysis, in a way that estimates cannot (Vann, et. al., 2002)

Simulation provides a method to arrive at a theoretically infinite number of

realizations (renditions) of the surface, each of which has approximately the same
variogram and variance of the srcinal data. These are referred to as equiprobable or
equally likely images. Any individual simulation is a poorer estimate than kriging;
however averaging a very large set of simulations can yield a good estimate.
Simulation strives for realism whereas estimation strives for accuracy.

Simulations can be used to deal with a range of problems that involve variability,
and for such problems they are far superior than any estimate (including kriging).
One example of an application is to generate confidence intervals for an
estimated block. Other applications include modeling local spatial variability for
blending, stockpiling or mining selectivity studies (Bertoli, 2005).

Gaussian Simulation and Sequential Gaussian Simulation are the most common
conditional simulation methods. Non-parametric simulation techniques such as
Sequential Indicator Simulations and Probability Field Simulations are becoming
more and more popular.

Geostatistics Without Tears 79

Conditional Cosimulation
Conditional Cosimulation is a multivariate conditional simulation of more than
one variable.

Gaussian Simulation
Variogram modeling is performed on normally distributed data. The variogram
model is used to krig the data at all locations providing a base map. In
simulation, multiple realization maps are created that honor the actual data value
(a conditional map), and approximately the same variogram and distribution. To
generate each realization, an unconditional map is simulated which honors the
variogram model but not the data at the data locations. Then the same
unconditional map is used to create another map by kriging the values at the
same locations as the actual data. At each grid node a simulated error is obtained
– the difference between the kriged and simulated value. This error is added to
the base map at each grid location. This gives the first realization which can then
be back transformed to its srcinal distribution. This is repeated for other
realization using a different random number sequence to generate multiple
realizations of the map (Syed, 1997).

Plurigaussian simulation
Plurigaussian simulation tries to simulate categoricalvariables by the intermediate
simulation of continuous Gaussian variables (Armstrong et al., 2003). A
Plurigaussian simulation is way of simulating the geometry of reservoirs and
deposits. Fluid flow in oil reservoirs depends on their internal architecture - on the
spatial layout of the lithofacies. A similar problem arises in mining when different
facies with different grades or metallurgical recoveries are intermingled in a
complicated way. Plurigaussian simulations make it possible to simulate complex
types of geometry such as double (or multiple) anisotropies (e.g. gold bearing veins
in one dominant direction with conjugate veins in another direction) or ore types
that cannot touch each other. Plurigaussian simulation is an extension of Truncated
Gaussian simulation. Plurigaussian simulation has proved to be very efficient in
providing images reproducing the main features of the geology encountered in
kimberlite crater deposits (Deraisme et al., 2004).

Probability Field Simulation

Probability Field (p-field) simulationis performedonce to estimate thehistogram at
each grid location. Then using a uniform random number generator, a simulated
value is drawn at each node using the estimated probability density function (or
histogram). The only other constraint to honor is that the distribution of uniform
random numbers should honor the variogram of the data, or simply that they should
be correlated. This makes for fast conditional simulations compared to Sequential
Indicator Simulation (Syed, 1997). Probability Field Simulation is very efficient and
conceptually simple. Probability Field Simulation has two main drawbacks which
can be detrimental. Nodes close to the conditioning data commonly appear as local

Geostatistics Without Tears 80

 minima or maxima of the simulated realizations. The simulated values usually show
greater continuity than the srcinal data. Pyrcz and Deutsch (2001) detail these
problems and recommend that Probability Field Simulation not be used.

Sequential Gaussian Simulation

Sequential Gaussian simulation is an efficient method widely used in the mining
industry (Vann et al., 2002). Variogram modeling is performed on normally
distributed data. One grid node is selected at random then the value at that location

is kriged(Gaussian)
normal which also gives the kriged
distribution that hasvariance. A random
a variance number
equivalent to theiskriged
drawnvariance
from a
and a mean equivalent to the kriged value. This number is then the simulated
number for that grid node. Another grid node is selected at random and the process
is repeated. This procedure defines a random path through all the gird nodes. For
the kriging all previously simulated nodes are included to preserve the spatial
variability as modeled in the variogram. When all the nodes have been simulated a
back transform to the srcinal distribution is performed. This gives the first
realization. This process is repeated for all the other realizations using a different
random number sequence to generate multiple realizations of the map (Syed, 1997).
For Sequential Gaussian Simulation the biggest problem is the search neighborhood
selection. The selection of small neighborhoods can lead to poor condition and poor
replication of the variogram.

Sequential Indicator Simulation

Sequential Indicator simulation was developed for application in petroleum

reservoir modeling where extreme values (high permeabilities) are well
connected in space. The high connectivity of extreme values is difficult to model
in the multigaussian paradigm. Another major advantage of the Sequential
Indicator Simulation is that hard data and soft data can be easily mixed (Vann, et.
al., 2002)

Variables are divided into several thresholds (or cut-offs). The data indicators are
then coded (1 or 0) depending on exceedence or not of the thresholds. Sequential
indicator simulation proceeds as follows: a random location is chosen and
indicator kriging is performed for the threshold at that location. After correcting
for possible order relations violations the indicator kriging will define a full
cumulative probability density Function at that location. A simulated value from
this Cumulative Probability Density Function is drawn based on a uniform
random number generator. The sequence simulations then moves on to another
location with this simulated value assumed as “hard” data until all nodes are

simulated. This gives one realization (Syed, 1997).

A main difficulty with Sequential Indicator Simulation is the same as for multiple
indicator kriging which is order relation problems (Vann and Guibal, 2000; Vann et
al., 2000). Because indicator variogram models may be inconsistent from one cut-off
to another a higher prediction about a certain cut-off have be predicted than for a
lower cut-off. A further drawback is that the quality of the simulation is sensitive to
the kriging neighborhood employed which is often too small (Vann et al., 2002).

Geostatistics Without Tears 81

Truncated Gaussian Simulation
Truncated Gaussian simulation was first designed to provide stochastic images of
sedimentary geology, mostly in fluvio-deltaic environments. The basic principle
is to replace the handling of the geological description by the handling of a
Random Function with multigaussian distribution for which geostatistical
simulations are used routinely (Carrasco et al., 2006).

Truncated Plurigaussian Simulation

The truncated plurigaussian model was introduced by Le Loc’h (1994), Gallie
(1994), Le Loc’h and Gallie (1997) and others as a more flexible alternative to
the standard truncated Gaussian model for describing facies distributions with
complicated facies arrangements (Liu and Oliver, 2003)

Turning Bands Simulation

Turning Bands was the first large-scale 3D Gaussian simulation algorithm
implemented (Journel, 1974, Mantoglou and Wilson, 1982). The Turning Bands
method works by simulating a one-dimensional process on lines regularly spaced
in 3D. The one-dimensional simulations are then projected onto the spatial
coordinates and average to give the required 3D simulated value.

The Turning Bands method is very efficient for generating non-conditional

simulations and reproduces the variogram better than other methods (Vann, et al.,
2002). The Turning Bands method used to suffer mechanical limitations. Only
certain specific variogram models (including spherical and exponential models)
could be simulated and Turning Bands is considerably slower to process than
sequential methods.

Geostatistics Without Tears 82

APPLYING GEOSTATISTICS
Rather than simply seeking to produce a visually-pleasing map geostatistics
provides tools to explore and understand the nature of a data set. Do geostatistical
tools produce useful results? “Useful” has to be decided by the user. Across a
wide spectrum of applications the answer is yes but in some cases the answer
may be not positive due to: a lack of data, difficulty in estimating or modeling
the variogram, the sensitivity of the estimator to unusual data values, etc. It has
been stated that “the interpretation of variograms has been found in practice to be
almost as difficult as Freud’s interpretation of dreams”(Henley, 2001).
Geostatisticians have developed a number of “rules of thumb” to help. There are
no “correct answers in geostatistics, only the opportunity to gain more
knowledge about the data and the interpolated surface, and to improve on the
surface model.

Geostatistics is concerned with spatial data and each data value is associated with
a location in space and there is at least an implied connection between the
location and the data value. The semivariogram provides a spatial picture of the
data and choosing the correct variogram model to use is an art. The development
of an appropriate variogram model requires numerous correct decisions based on
a solid understanding of the data and the underlying processes from which the
data are drawn. The following is provided as a summary guide to understanding
the kriging process.

Data Collection
• State the objective of the modeling exercise in clear quantitative term.

• Select a sample size and a sample collection strategy that are consistent
with the objectives.

• Try to obtain a sample size of 100 data points or more – any calculations
using less than 100 samples may give poor results.

• Errors made during sampling add a random component of variance to the

variogram and increases the nugget effect. If the data are really bad, the
variogram may show a total nugget effect.

• If necessary, re-project data from geographic to a projected coordinate

system.

•
Mixed data sets – if a number of sampling methods were used, each data
set should be statistically analyzed separately to see if there are any
significant differences.

Geostatistics Without Tears 83

Data, Summary Statistics and Histogram Visualization

• Examine a graduated symbols map looking for data trends, anomalies

and errors.

• Is the standard deviation larger than the mean?

• Are the data skewed?

• If the data are multi-modal, the data may need to be split into different sets.

• Are there data anomalies?

• Possibly transform the data of skewed distributions (i.e., logarithmic

transforms; indicator).

Semivariogram Visualization
• Check for enough number of data pairs at each lag distance (a
minimum of 30 to 50 pairs).

• Truncate at half the maximum lag distance to ensure enough pairs.

• Use a large lag tolerance to get more pairs and a smoother variogram.

• Sometimes the variogram is sensitive to the choice of class size, lag

distance spacing
average and tolerance.
betweenAthe
good starting point for the class size is the
samples.

• Is the variogram nice, parabolic, discontinuous, erratic or completely

random.

• If the spatial structure in the semivariogram doesn’t look good, don’t

revert to IDW.

• Start with an omnidirectional variogram before proceeding with

directional variograms.

• Remove trends.

Evaluation of Model
• Correct data errors.

• Manage extreme values or remove outliers – the estimator of the

variance is very sensitive to outliers. A chaotic looking variogram
can often be improved by limiting the maximum value used in the
calculation.

• Select an appropriate model for the data and the objectives.

Geostatistics Without Tears 84

 • Use other variogram measures to take into account lag means and
variances (inverted covariance, correlogram or relative variograms).

• Interpret the model to obtain the best reliable fit possible.

• Understand the uncertainty and the precision of the predictions.

• Make decisions that are sensible.

• Reevaluate the model with new data when environmental conditions

change.

Geostatistics Without Tears 85

Geostatistics Without Tears 86
CONCLUSION
Geostatistics (kriging) is an advanced technique for interpolation and mapping of
data that relies on statistics (theory of regionalized variables) and computer
processing. While there are individuals who would identify themselves as
"geostatisticians" it is more likely that individuals using geostatistics would call
themselves: ecologists, environmental scientists, geographers, geologists,
hydrologists, mathematicians, meteorologists, mining engineers, petroleum
engineers, plant pathologists, soil scientists, statisticians, etc.

Geostatistics is not a cure-all nor is it useful for all problems. It is not a black box
scheme and must be used with some care. In some cases, and for certain data sets
given some objectives, a user need only have access to a geostatistical software
package. In other cases it may be necessary to seek the advice and assistance of
someone more experienced and with a stronger understanding of the
mathematics/statistics in order to adequately apply geostatistics to a data set.

The information presented in this paper was developed to provide a good base
understanding to the field of geostatistics to either design or to analyze the data
from environmental monitoring surveys. Hopefully no tears were shed during
this learning process. An extensive bibliography is given which goes beyond the
references cited in this paper and is provided to give readers an entry to the
primary literature on geostatistics and surface interpolation.

Geostatistics Without Tears 87

Geostatistics Without Tears 88
REFERENCES

ACE, 2005. Options for Variogram. Applied Computer Engineering (ACE), Inc. Integrated
Reservoir Studies. Houston Texax 77036.http://www.gridstat.com/ACEweb/example8.htm

Acevedo. M.F., 1997. Geog 5190 Advanced Stats and Quantitative Methods, Lecture
Outlines, Lab Guides and Assignments.
http://www.geog.unt.edu/~acevedo/courses/590/week10/lectur10.htm

Agterberg, F.P., 1974. Geomathematics, Elsevier Science Inc.

Annales des Mines (Accueil), 2000. Georges François Paul Marie Matheron (1930-
2000). http://www.annales.org/archives/x/matheron.html

Australian National University. 2004. Centre for Resource and Environmental Studies
(CRES)-ANUSPLIN Version 4.3. http://cres.anu.edu.au/outputs/anusplin.php

ANUDEM Version 5.2. The Australian National University.

http://cres.anu.edu.au/outputs/anudem.php

Arad, N., Dyn, N., Reisfeld, D. and Yeshurun, Y. 1994. Image warping by radial basis
functions: applications to facial expressions. CVGIP: Graphical Models and Image
Processing, 56, 161-172.

Ardini, F., Fioavanti, S. and Giusto, D.D., 1991. Multifractals towards Remote Sensing
Surface Texture Characterization. Proc. Int. Geoscience and Remote Sensing Symp.
(IGARSS’91), 3-6 June, Espoo, Finland, pages 317-320.

Armstrong A., Galli A., Le Loc’h G., Geffroy G. and Eschard R., 2003. Plurigaussian
Simulations in Geosciences, Springer.

Armstrong, M., 1984. Common problems with universal Kriging, Mathematical Geology
16(1): 101-108.

Armstrong, M., 1986. On the Fractal Dimensions of some Transient Soil Properties,
Journal of Soil Science, 37, 641-652.
Armstrong, M., and Alain G., 2001. Matheron's Contributribution to Geostatistics, Centre
de Géostatistique, Fontainebleau, France, 2001 Annual Conference of the International
Association for Mathematical Geology, Cancún, Mexico, September 6-12, 2001.
Ashraf, M., Loftis, J.C., and Hubbard, K.G., 1997.Application of Geostatistics to Evaluate
Partial Weather Station Networks. Agricultural & Forest Meteorology. 84(3-4): 225-271.

Atkinson, P.M. and Lloyd, C.D., 1998. Mapping Precipitation in Switzerland with
Ordinary and Indicator Kriging. Journal of Geographic Information and Decision
Analysis, Vol. 2, No. 2.: 72-86. ftp://ftp.geog.uwo.ca/SIC97/Atkinson/Atkinson.html

AWI. 2006. Objectivein

Interpolation Method. Stiftung Alfred-Wegener-Istitut für Polar-
und Meeresforschung der Helmholtz-Gemeinschaft. http://www.awi-
bremerhaven.de/Atlas/SO/A4.html

Barnes, R.J. 1991. The variogram sill and the sample variance. Mathematical Geology,
Vol. 23, No. 4: 673-678.

Barnes, R. 2005. Variogram Tutorial. Golden Software, Inc. Golden, Colorado 80401-
1866, U.S.A. http://www.goldensoftware.com/variogramTutorial.pdf

Geostatistics Without Tears 89

Barringer, J.R.F. and Lilburne, L., 1997. An Evaluation of Digital Elevation Models for
Upgrading New Zealand Land Resource Inventory Slope Data. Presented at the Second
Annual Conference of GeoComputation ’97 & SIRC ’97, University of Otago, New Zealand,
26-29 August 1997.
http://divcom.otago.ac.nz/sirc/webpages/Conferences/SIRC97/97Proceedings/97Barringer.pd
f

Bastin, G., Lorent, B., Doque, C., and Gevers, M., 1984 Optimal estimation of the average areal
rainfall and optimal selection of rain gauge locations, Water Resources Res. 20(4):
463-470.

Bates, D., Lindstrom, M., Wahba, G. and Yandell, B. 1987 GCVPACK - Routines for

Generalized
ComputationCross Validation. Communications in Statistics B - Simulation and
16: 263-297.
,

Berry, J.K, 1999. Extending Spatial Dependency to Maps. Beyond Mapping, Geoworld,
Vol 12 (1): 26-27.

Bertoli, O., 2005. Conditional Simulation. Quantitative Geoscience.

http://www.quantitativegeoscience.com/references_simulation.asp

Billings, S.D., Newsman, G.N., and Beatson, R.K. 2006. Fourier Transformation of
Geophysical Data Using Continuous Global Surfaces. University of British Columbia.
http://www.geop.ubc.ca/~sbilling/papers/fourier_transformation_using_cgs.PDF

Binkley, M.R., Carlson, R. and Lee, J., 1997. Spatial Interpolation of Lake Water Quality
Data. Kent State University. http://humboldt.kent.edu/~binkley/poster.html

Bogaert, P., Mahau, P. and Beckers, F., 1995. The Spatial Interpolation of Agro-Climatic
Data – Cokriging Software and Source Code,Use’ Manual, Version 1.0b. Unité de
Biométrie, Faculté des Sciences Agronomiaques, Université Catholique de Louvain,
Louvain-la-Neuve, Belgium. Agrometeorology Series Working Paper, Number 12, FAO
Rome, Italy

Bonham-Carter, G.F., 1994. Geographic Information Systems for Geoscientists:

Modelling with GIS, Pergamon, Tarrytown, New York.

Boomer, K.M. and Brazier, R.A., 1997. Geostatistics, Seismic Risk Analysis of Northern
Arizona from Acceleration Contour Maps, Enhancing the Sampling Procedure through a
Geostatistical Analysis.http://www.geo.arizona.edu/geophysics/students/brazier/geo.html

Booth, T.H. and Jones, P.G. 1996. Climatic Databases for use in Agricultural Management
and Research Proceedings of the Arendal II Workshop on UNEP/GRID and CGIAR
cooperation to meet requirements for the use of digital data in agricultural management and
research. Arundal, Norway 9-11 May 1995.
http://www.grida.no/cgiar/htmls/climpres.htm#climpres

Boutteir F. and Courtier, P. 1999. Data Assimilation Concepts and Methods.

European
Centre for Medium-Range Weather Forecasts(ECMWF).
http://www.ecmwf.int/newsevents/training/rcourse_notes/D ATA_ASSIMILATION/ASSI

M_CONCEPTS/Assim_concepts8.html
Bretherton F.P., R.E. Davis and C.B. Fandry. 1976. A Technique for Objective Analysis and
Design of Oceanic Experiments applied to MODE-73. Deep Sea Research, 23, p. 559-582.

Bretherton F.P. and J.C. Williams. 1980. Estimations from Irregular Arrays. Reviews of
Geophysics and Space Physics, Vol.18, 4, p.789-812.

Briggs, P.R., and Cogley, J.G. 1996. Topographic Bias in Mesoscale Precipitation
Networks. Journal of Climate, 9, 205-218.

Geostatistics Without Tears 90

Brill, D.J., 1997. Skewness and Kurtosis - A Discussion
http://www.uark.edu/plscinfo/pub/methods/help/skkrdis.htm

Bruce, J.P. and Clark, R.H., 1980. Introduction to Hydrometeorology. Pergamon Press,
Ontario. 324 pages.

Bruno, O. and Pohlman, M. 2002. High-order Surface Interpolation. Applied and

Computational Mathematics, California Institute of Technology.
http://www.acm.caltech.edu/~bruno/matt_lockheed_12_2_2002.pdf

Bruno, O.P, and Pohlman, M.M., 2003. High-order Surface Representation. California
Institute of Technology. http://www.acm.caltech.edu/~bruno/afosrposter03.pdf

Burgess, T.M. and Webster, R., 1980. Optimal Interpolation andIsarithmic Mapping of Soil
Properties I. The semivariogram and punctual kriging. Journalof Soil Science 31: 315-331.

Burgess, T.M. and Webster, R., 1980. Optimal Interpolation and Isarithmic Mapping of
Soil Properties II. Block kriging. Journal of Soil Science 31: 333-341.

Burrough, P.A., 1981. Fractal dimensions of landscapes and other environmental data.
Nature 294:240-242.

Burrough, P.A., 1993a. Fractals and Geostatistical Methods in Landscape Studies.

Fractals in Geography (N.S.N. Lam and L. De Cola, editors), Prentice Hall.

Burrough, P.A., 1993b. Soil Variability: A Lake 20 th Century View. Soils and Fertilizers,
pages 529-562.

Cadell, W., 2002. Report on the Generation and Analysis of DEMs for Spatial
Modelling. The Macaulay Institute. http://www.macaulay.ac.uk/ladss/documents/DEMs-
for-spatial-modelling.pdf

Calter, P., 1990. Technical Mathematics with Calculus. Prentice Hall, Engelwoods Cliffs,
New Jersey.

Carr, James R., 1995. Numerical Analysis for the Geological Sciences. Prentice Hall,
Englewoods Cliffs, New Jersey.

Carrasco, P., Ibarra, F., Le Loc’h, G., Rojas, R., Seguret, S., 2006. Application of the
Truncated Gaussian Simulation Method to the MM deposit at Codelco Norte, Chile.
http://www.geovariances.com/IMG/pdf/EAGE_Madrid_Leloch.pdf

Carrat, F. and Valleron, A., 1992. Epidemiological Mapping Using the “Kriging”
Method: Application to an Influenza-like Illness Epidemic in France. American Journal
of Epidemiology, 135(11): 1293-1300.

Castrignanò, A. and Lopez, R., 2003. GIS and Geostatistics: An Essential Coupling For
Spatial Analysis. Istituto Sperimentale Agronomico. Italian Interest Group on Spatial
Statistics. Text: http://194.119.196.14/istituti/ieif/spatialstatistics/res/gg.pdf, PPT:
http://194.119.196.14/istituti/ieif/spatialstatistics/res/gg_ppt.pdf

Centre de Geostatistique de l'Ecole des Mines de Paris, 2005. Professor Georges

Matheron (1930-2000), A life in probabilistic modelling. 35 Rue St Honore, 77300
Fontainebleau, FRANCE. http://cg.ensmp.fr/Presentation/Matheron/Matheron_en.shtml

Geostatistics Without Tears 91

Chandila, P.K., 2004. Strategy for Global Optimization and Post-Optimality Using Local
Kriging Approximations. A Thesis submitted to the Graduate School of the University of
Notre Dame in Partial Fulfillment of the Requirements for the Degree of Master of
Science. Graduate Program in Aerospace and Mechanical Engineering, Notre Dame,
Indiana. http://etd.nd.edu/ETD-db/theses/available/etd-04142004-
144454/unrestricted/ChandilaP042004.pdf

Chapman, A.D., 2003. The Case for a 3-minute Climate Surface for South America.
SpeciesLink, Appendix_e.pdf. http://cres.anu.edu.au/outputs/anudem.php

Chapman, A.D., Muňoz, M.E.S., and Koch, I. 2005. Environmental Information: Placing
Biodiversity Phenomena in an Ecological and Environmental Context. Biodiversity
Informatics, 2, 24-41. http://jbi.nhm.ku.edu/index.php/jbi/article/viewFile/5/3

Chang, T.J. and Teoh, C.B., 1995. Use of the Kriging Method for Studying Characteristics
of Ground Water Droughts. Water Resources Bulletin, American Water Resources
Association, Vol. 31, No. 6: 1001- 1007.

Cheng, K-S, Wei, C., Cheng, Y-B, and Yeh, H-C. 2003. Effect of Spatial Variation
Characteristics on Contouring of Design Storm Depth. Hydrological Processes, 17, 1755-
1769. Published online 28 January 2003 in Wiley InterScience (www.interscience.wiley.com).
DOI: 10.1002/hyp.1209. http://www.rslabntu.net/KSC/Publications/IDF_Contouring.pdf

Chellemi, D.O., Rohrbach, K.G., Yost, R.S. Yost, and Sonoda, R.M., 1988. Analysis of
the Spatial Pattern of Plant Pathogens and Diseased Plants Using Geostatistics. The
American Phytopathological Society. Vol. 78, No. 2: 221-226.

Chilès, J.-P., and Delfiner, P., 1999. Geostatistics: Modeling Spatial Uncertainty, John
Wiley & Sons, New York and Toronto, 695 pages.

Chua, S. H. and Bras, R. L., 1982. Optimal Estimators of Mean Areal Precipitation in
Regions of Orographic Influence. J. Hydrology, 57: 23-48.

Clark, I., 1979a. Practical Geostatistics.http://uk.geocities.com/drisobelclark/practica.html

)
th
Clark, I., 1979b. Does Geostatistics Work? 16 International APCOM Symposium, T.J.
O’Neil (Ed), McGraw-Hill, New York, 213-225.

Clark, I., 1983. Regression Revisited. Mathematical Geology, Vol. 15, No. 4,
1983.
th
Clark, I., 1986. The Art ofCross Validation in Geostatistical Applications. 19 Application
of Computers and Operations Research in theMineral Industry, R.V. Ramani , Editor.
Sponsored by The Pennsylvania State University, April 14-16, 19 86. Published by Society of
Mining Engineers, Inc., Littleton, Colorado 1986.

Clark, I., 1987. Turning the tables an interactive approach to the traditional estimation of
reserves", J. S. Afr. Inst. Min. Metall., Vol 87, No 10, pp.293-306. Full copy at
http://uk.geocities.com/drisobelclark/resume/Publications.html

Clark, I., 1998. Geostatistical Estimation Applied to Highly Skewed Data. JointStatistical
Meetings, Dallas, Texas, August, 1998.

Clark, I., 2000. Erratic highs -- a perennial problem in resource estimation SME 2000, Salt
Lake City, Utah.

Clark I., 2001. Practical Geostatistics. Geostokos Limited, Alloa, Scotland.

Clark, I., 2001a. AI-GEOSTATS listserver email: Anisotropy with varying Range and Sill.

Clark, I., 2001b. AI-GEOSTATS listserver email: Kriging of multiple samples.

Geostatistics Without Tears 92

Clark, I., 2001c. AI-GEOSTATS listserver email: interannual spatial "stability"
of variable.

Clark, I., 2002. AI-GEOSTATS listserver email: Log transformation and zeros.

Clark, I., 2006a. AI-GEOSTATS listserver email: spherical model.

Clark, I., 2006b. AI-GEOSTATS listserver email: kriging without a nugget.

Clark I. and Harper,W.V., 2000. Practical Geostatistics 2000. Ecosse North America Llc,
Columbus, Ohio, U.S.A.

Clark I. and Harper,W.V., 2000. Practical Geostatistics 2000, Answers to the Exercises.

Ecosse North America Llc, Columbus, Ohio, U.S.A.

Cox D.D., Cox L.H., and Ensor, K.B., 1997. Spatial Sampling and the Environment -
Some Issues and Directions. Environmental & Ecological Statistics. 4(3): 219-233.

Cressie, N., 1985. Fitting variogram models by weighted least squares. Mathematical
Geology 17: 563 - 586.

Cressie N., 1990. The Origins of Kriging. Mathematical Geology, Vol. 22, No. 3, 1990: 239 - 252.

Cressie, N.A.C. 1991. Statistics for Spatial Data. John Wiley, New York.
nd
Cressie, N., 1993. Statistics for Spatial Data. 2edition. John Wiley and Sons, Inc, New York.

Cressie, N., 2003. Statistics for Spatial Data. Revised edition. John Wiley and Sons, Inc,
New York.

Cressie, N.A.C. and Hawkins D.M., 1980. Robust Estimation of the Variogram. Journal
International Association Mathematical Geology 12:115-125.

Cressman,
367-374. G., 1959. An Operation Objective Analysis System. Monthly Weather Review, 87,

Creutin, J.D., and Obled. C., 1982. Objective Analysis and Mapping Techniques for Rainfall
Fields: an objective comparison. Water Resource Research 18(2):413-431.

Daly, C. 1996. Overview of the PRSIM Model.

http://www.ocs.orst.edu/prism/docs/overview.html

Daly, C. 2006. Guidelines for Assessing theSuitability of Spatial Climate Data Sets.
International Journal of Climatology. 26: 707-721. Published online 27 March 2006 in Wiley
InterScience (www.interscience.wiley.com) DOI: 10.1002/joc.1322.

Daly, C., Gibson, W.P., Taylor, G.H. Johnson, G.L., and Pasteris P. 200
2. A Knowledge-
based Approach to the Statistical Mapping of Climate. Climate Research 22: 99-113.

Daly, C., Gibson, W.P., Doggett, M.,Smith, J and Taylor, G. 2004. A Probabilistic-spati
al
Approach to the Quality Control ofClimate Observations. In Proceedings of the 14th AMS
Conference on Applied Climatology, American Meteorological Society, Seattle,
Washington, January 13-16, 2004. http://www.ocs.orst.edu/pub/prism/docs/appclim04-
probabilistic_spatial_approach-daly.pdf

Daly, C., Hlemer, E.H., Quinones, M. 2003. Mapping the Climate of Puerto Rico,
Vieques,
and Culebra. International Jounal ofClimatology 23: 1359-1381.

Daly, C, and Johnson, G.L. 1999.PRISM Spatial Climate Layers: Their development and
th
use. Short course on topics in applied climatology. 79 Annual Meeting of the American
Meteorological Society, American Meteorological Society: Dallas, Texax, 10-15 January.
http://www.ocs.orst.edu/pub/prism/docs/prisguid.pdf

Geostatistics Without Tears 93

Daly, C., Nielson, R.P., and Phillips, D.L. 1994. A Statistical-topographic Model for Mapping
Climatological Precipitation over Mountainous Terrain. Journal of Applied Meteorology,
33:140-158.http://www.ocs.orst.edu/prism/docs/overview.html

Dandurand, L.M., Knudsen, G.R., and Schotzko, D.J., 1995. Quantification of

Phythium
Ultimum var. sporangiiferumZoospore Encystement Patterns Using Geostatistics.
Phytopathology, Vol. 85, No. 2: 184 - 190.

David, M., 1977. Geostatistical Ore Reserve Estimation, Elsevier, 364 pages.

Davis, J.C., 1973. Statistics and Data Analysis in Geology. JohnWiley and Sons, New York.
nd
Davis, J.C., 1986. Statistics and Data Analysis in Geology, 2 edition, Wiley, New York.
Davis, B.M., and Borgman L.E., 1979. A Test of Hypothesis Concerning a Proposed
Model for the Underlying Variogram, Proceedings 16th APCOM, 163-181.

De Cola, L. 1989. Fractal Analysis of a Classified Landsat Scene. Photogrammetric

Engineering & Remote Sensing, 55:601-610.

De Jong, S.M. 1993. An Application of Spatial Filtering Techniques for Land Cover
Mapping Using Tm Images, Geocarto International, 8(1):43-49.

Delfiner, P. and Delhomme, J.P. 1975. Optimum Interpolation y Kriging. Display and
Analysis of Spatial Data, Davis, J.C. and McCullagh, M.J. (Eds.) Wiley and Sons, New
York. pp. 96-114.

Delhomme, J.P., 1979. Kriging in the Hydrosciences, Advance in Water Res. 1(5): 251-266.

Deraisme. J., 1997. Interpolation and Reality. Geovariances News Letter #7, page 3.
http://www.geovariances.fr/publications/n1/7/page3.html

Deraisme, J. and Farrow, D. 2004. Geostatistical Simulation Techniques Applied to

Kimberlite Orebodies and Risk Assessment of Sampling Strategies. Banff.
http://www.geovariances.com/IMG/pdf/Banff2004.pdf

Deutsch, C.V., 1996. Correcting for Negative Weights in Ordinary Kriging. Computers
and Geosciences Vol. 22, No. 7: 765-773.

Deutsch, C.V. and A.G. Journel 1992. GSLIB Geostatistical Software Library. Oxford
University Press, New York. www.gslib.com

Deutsch, C.V. and A.G. Journel 1997/1998. GSLIB Geostatistical Software Library and
Users Guide, 2nd Edition. Oxford University Press, New York.

Dingman, S.L., Seely-Reynolds, D.M., and Reynolds III R.C. 1988. Application of
Kriging to Estimate Mean Annual Precipitation in a Region of Orographic Influence.
Water Resource Bulletin, 24,329-339.

Dorsel, D. and La Breche, T., 1998. Environmental Sampling & Monitoring Primer,
Kriging. http://www.ce.vt.edu/enviro2/smprimer/kriging/kriging.html

Duchon, J. 1976. "Interpolation des fonctions de deux variables suivant le principe de la

flexion des plaques minces." RAIRO Analyse Numérique 10, 5-12.

Duval, R., Lévy, R., Matheron, G., 1955. Travaux de D. G. Krige sur l’évaluation des
gisements dans les mines d’or sud-africaines. Annales des Mines, Paris, vol. 12, p. 3-49.

Eastman, J., 1999. IDRISI32, Guide to GIS and Image Processing – Volume 2, p 139 and
p 155-161. Clark Labs, Clark University.

Geostatistics Without Tears 94

Eastman, J., 1999. IDRISI32, Tutorial. 179-201. Clark Labs, Clark University.
Ehlschlaeger, C.R., 2001. Representing Multiple Spatial Statistics in Generalized Elevation
Uncertainty Models: Moving Beyond the Variogram. Department of Geography, Hunter
College, New York, New York.
http://www.geo.hunter.cuny.edu/~chuck/RGSIBG/beyondVariogramSmallMaps.pdf
Englund, E.J. Undated. Spatial Autocorrelation: Implications for Sampling and Estimation.
U.S. Environmental Protection Agency, P.O. Box 93478, Las Vegas, Nevada 89193-3478.
http://www.epa.gov/nerlesd1/cmb/research/papers/ee107.pdf

Englund, E.J., 1990. A Variance ofGeostatisticians. J. Math. Geology. vol. 22, no. 4, pp.
417-455.
Englund, E.J., 1993. Spatial Simulation: Environmental Applications. A version of this
manuscript has been published in Environrnental Modeling with GIS, M.F. Goodchild et
al., eds., Oxford University Press (1993) 432-437.
http://www.epa/gov/nerlesd1/cmb/research/papers/ee104.pdf
Franke, R. 1985. Thin-plate Splines with Tension. Computer Aided Geometric Design 2,
87-95, North-Holland.
Freek, D., 1997. Classification of High Spectral Resolution Imagery Using an Indicator
Kriging Based Technique. International Institute for Aerospace Survey and Earth Sciences
(ITC), Department of Earth Resources Surveys. 350 Boulevard 1945, P.0. Box 6, 7500 AA,
Enschede, The Netherlands.
Galli, A., Beucher H., Le Loc’h, G., Doligez, B. and Group H., 1994. The pros and cons of
the truncated Gaussian method, in Geostatistical Simulations, pp. 217-233, Kluwer
Academic, Dordrecht.
Gamma Design. 2004. GS+ for Windows is a Windows 95/98/NT version of a
geostatistical mapping program for the PC. Gamma Design Software, LLC, Plainwell,
Michigan 49080 USA. http://www.gammadesign.com
Gambolatic, G. and Volpi, G., 1979a. A Conceptual Deterministic Analysis of the Kriging
Technique in Hydrology. Water Resources Res. 15(3): 625-629.
Gandin, L.S., 1963. Objective Analysis of Meteorological Fields: Gidrometeorolgicheskoe
Izdatel’stvo (GIMIZ), Leningrad (translated by Israel Program for Scientific Translations,
Jerusalem, 1965, 238 pages.
Gandin, L.S., and Kagan, R.L. 1974. Construction of a System for Objective Analysis of
Heterogeneous Data Based on the Method of Optimum Interpolation and Optimum
Agreement. Meteorological I Gidrol., No. 5 (Moscow), 1-11. [Translated by Joint
Publications Research Service.]
Garrott, J., 2003. SAKWeb: (Spatial Autocorrelation and Kriging Web) A W3
Geocomputation Perspective. Universidade Nova De Lisboa.
http://www.isegi.unl.pt/labnt/presentations/2003/031215Garrot.pdf
Geary, R., 1954. The Contiguity Ratio and Statistical Mapping. The Incorporated
Statistician, Volume 5, 115-145.
Genoton,, M.G. and Furrer, R., 1998. Analysis of Rainfall Data by Simple Good Sense: is
Spatial Statistics Worth the Trouble? Journal of Geographic Information and Decision
Analysis, Vol. 2, No. 2: 11-17. ftp://ftp.geog.uwo.ca/SIC97/Genton_1/Genton_1.html

Geostatistics Without Tears 95

Genton, M.G. and Furrer, R., 1998. Analysis of Rainfall Data by Robust Spatial Statistics
using S+Spatialstats. Journal of Geographic Information and Decision Analysis, Vol. 2,
No. 2: 12 -136.ftp://ftp.geog.uwo.ca/SIC97/Genton_2/Genton_2.html
Gibson, W.P., Daly, C., Kittel, T., Nychka, D, Johns, C., Rosenbloom, N., McNab, A. and
Taylor, G. 2002. Development of a 103-year High-resolution Climate Data Set for the
th
Conterminous United States in Proceedings 13 AMS Conference on Applied Climatology,
American Meteorological Society, Portland, Oregon, may 13-16, 2002. pp 181-183.
http://www.ocs.orst.edu/pub/prism/docs/appclim02-103yr_hires_dataset-gibson.pdf
Glasbey, C.A., and Mardia, K.V. 1998. A Review of Image Warping Methods. Journal of
Applied Statistics, 25, 155-177. http://www.bioss.sari.ac.uk/staff/chris/warp.pdf
Glover, D.M., 1998,2000. 12.747 Lecture5:Section 4: Objective Mapping andKriging.
Woods Hole Oceanographic Institution.
http://w3eos.whoi.edu/12.747/notes/lect05/l05s04.html
Goodchild, M., 1999. Spatial Analysis and GIS. 1999 ESRI User Conference, Pre-
Conference Seminar.
Goodman, A., 1973,1999. Trend Surface Analysis in the Comparison of Spatial
Distributions of Hillslope Parameters. http://www.deakin.edu.au/~agoodman/masters/
Goovaerts, P., 1997. Geostatistics for Natural Resources Evaluation. Oxford University
Press, New York. 483 pages.
Goovaerts, P., 1999. Performance Comparison of Geostatistical Algorithms for Incorporating
Elevation into the Mapping of Precipitation. Proceedings of, the IV International Conference
on GeoComputation, Mary Washington Colleg e in Fredericksburg, VA, USA,25-28 July
1999. http://www.geovista.psu.edu/sites/geocomp99/Gc99/023/gc_023.htm

Goovaerts, P., 2000. Geostatistical approaches for incorporating elevation into the spatial
interpolation of rainfall. Journal of Hydrology 228 (2000) 113-129.
http://www.terraseer.com/training/geostats/jhydrol00.pdf
Goovaerts, P., 2001. Geostatistical Modelling of Uncertainty in Soil Science. Geoderma
103:3-26.
Goovaerts, P. 2002. AI-GEOSTATS: Tuning a semivariogram.
Goovaerts, P. 2005. Geostatistical Analysis of Disease Data: Estimation of cancer mortality
risk from empirical frequencies using Poisson kriging. International Journal
of Health
Geographics 4:31.http://www.ij-healthgeographics.com/content/pdf/1476-072X-4-31.pdf
Goovaerts, P. 2006. Geostatistical Analysis of Disease Data: Visualization and
propagation of Spatial Uncertainty in Caner Mortality Risk using Poisson kriging and p-
field simulation. International Journal of Health Geographics 2006, 5:7. http://www.ij-
healthgeographics.com/content/pdf/1476-072X-5-7.pdf
Gottwald, T.R., Avinent, L, Llacer, G., de Mendoza, H. and Cambra, M., 1995. Analysis
of the Spatial Spread of Sharka (Plum Pox Virus) in Apricot and Peach Orchards in
Eastern Spain. Plant Disease. Vol. 79 No. 3: 266-278.
Gottwald, T.R., Cambra, M., Moreno, P, Camarasa, E. and Piquer, J., 1996. Spatial and
Temporal Analyses of Citrus Tristeza Virus in Eastern Spain. The American
Phytopatholgical Society. Vol. 86, No. 1: 45 - 55.
Griffith, D.A., 1987. Spatial Autocorrelation. A Primer. Resource Publication in
Geography, Association of American Geographers, Washington D.C.

Geostatistics Without Tears 96

Griffith, D.A., Amrhein, C.G. and Desloges, J.R., 1991. Statistical Analysis for
Geographers. Prentice Hall, Englewoods Cliffs, New Jersey.
Guocheng, P. Garrd, D. Moss, K. and Heiner, T., 1993. A Comparison Between
Cokriging and Ordinary Kriging: Case Study with a Polymetallic Deposit. Mathematical
Geology, Vol. 25, No. 3: 377-398.
Haan, C.T., 1977. Statistical Methods in Hydrology. Iowa
State University Press, Ames, Iowa.
Handley, J.R.F., 2004. Historic Overview of
the Witwatersrand Goldfields. Available from the
Geological Society of South Africa. ISBN 0-620-32127-X. 224 pages.

Hansen,
No. 10: R.O., 1993. Interpretive Gridding by Anisotropic Kriging. Geophysics, Vol. 58,
1491-1497.
Hengl, T. 2006. Regression-Kriging. Spatial-analyst.net. http://spatial-
analyst.net/regkriging.php
Hengl, T., Heuvelink, G.B.M., and Stein, A. 2003. Comparison of Kriging with External
Drift and Regression-kriging. Technical note, ITC, Available on-line at
http://www.itc.nl/library/Academic_output/
http://spatial-analyst.net/GRK/Hengl_et_al_Comparison_RK_KED.pdf
Henley, S. 2001. Geostatistics – Cracks in the Foundations? Stephen Henley, Resources
Computing International Ltd. Derbyshire, U.K.
http://www.infomine.com/softwaremine/articles/GeostatCracks.asp and
http://www.earthsciswinfo.com/swmine/swarticles/GeostatCracks.html
Herzfeld, U.C, Eriksson, M.G., and Homlund P., 1993. On the Influence of Kriging
Parameters on the Cartographic Output – A study in Mapping Subglacial Topography.
Mathematical Geology, Vol. 25, No. 7: 881-900.
Hevesi, J.A., and Istok, J.D. 1992. Precipitation Estimation of Mountainous Terrain Using
Multivariate Geostatistics. Part I: Structure Analysis. Journal of Applied Meteorology, 31,
661-667.
Hijmans, R.J., Cameron, S., and Para J. 2004a. WorldClim, version 1.2. A Square
Kilometre resolution database of global terrestrial surface climate.
Hijmans, R.J., Cameron, S., and Para J. 2004b. WorldClim, a new high-resolution global
climate database. Abstract: Inter-American Workshop on Environmental Data Access.
Campinas, Brazil.
Hiller W. and R. Käse. 1983. Objective Analysis of Hydrographic Data Sets from
Mesoscale Surveys. Berichte aus dem Institut für Meereskunde an der Christian-
Albrechts-Universität Kiel, 116 78 pages.
,

Holawe, F. and Dutter, R., 1999. Geostatistical Study of Precipitation Series in Austria:
Time and Space. Vienna University of Technology.

Hu, J., 1998.

Bechtel Methods ofInc.,
Environmental Generating Surfaces
151 Lafayette in Environmental
Drive, GIS Applications.
Oak Ridge, Tennessee 37922.
http://pasture.ecn.purdue.edu/~edu/~aggrass/esri95/to100/p089.html
Hutchinson, M.F., 1988. Calculation of Hydrologically Sound Digital Elevation Models.
Proceedings of the Third International symposium on Spatial Data Handling, August 17-
19, Sydney. International Geographic Union, Columbus, Ohio.
Hutchinson, M.F., 1989. A New Method for Gridding Elevation and Stream Line Data
with Automatic Removal of Pits. Journal Hydrology 106:211-232.

Geostatistics Without Tears 97

Hutchinson, M.F., 1990. Robust Calibration of Seasonally Varying Stochastic Weather
Models Using Smoothing Splines. Math. Compt. Simulation, 32, 125-130.
Hutchinson, M.F. 1991a. The Application of Thin-plate Smoothing Splines to Continent-
Wide Data Assimilation. In BRMC Research Report No. 27, Data Assimilation Systems,
Bureau of Meterology, Melbourne, Jasper, J.D. (Ed.) pp 104-113. In Data Assimilation
Systems: papers presented at the 2nd BMRC modeling workshop, September 1990. BMRC
Research Report No. 27, 104-113.
Hutchinson, M.F. 1991b. Climatic Analysis in Data Sparse Regions. Climatic Risk in
Crop Production, Muchow, R.C. and Bellamy, J.A. (Eds), CAB International, pp. 55-71.

Hutchinson, M. F. 1993. On Thin-plate Splines and Kriging. Computing Science and

Statistics, Vol. 25. Edited by M.E.Tarter and M.D.Lock, pp. 55-62. Berkeley: Interface
Foundation of North America.
Hutchinson, M.F. 1994. ANUSPINE (Revisions 1 July 1991). Centre for Research and
Environmental Studies. Available from CRES, Australian National University, G.P.O.
Box 4, Canberra, ACT 2601, Australia.
Hutchinson, M. 1995. Interpolating Mean Rainfall Using Thin-plate Smoothing Splines.
International Journal of Geographic Information Systems, 9, 385-403.
Hutchinson, M. 1998a. Interpolation of Rainfall Data with Thin-plate Smoothing Splines
– Part I: Two Dimensional Smoothing of Data with Short Range Correlation. Journal of
Geographic Information and Decision Analysis, vol. 2, no. 2, pp. 139-151, 1998.
http://www.ai-geostats.org/events/sic97/Hutchinson_1.pdf
Hutchinson, M. 1998b. Interpolation of Rainfall Data with Thin-plate Smoothing Splines
– Part II: Analysis of Topographic Dependence. Journal of Geographic Information and
Decision Analysis, vol. 2, no. 2, pp. 152-167, 1998. http://www.ai-
geostats.org/events/sic97/Hutchinson_2.pdf
Hutchinson, M.F. 2006. ANUDEM Version 5.2. Centre for Resource and Environment
al
Studies, The Australian NationalUniversity, Caneberra.
http://cres.anu.edu.au/outputs/anudem.php
Hutchinson, M. F. and Bischof, R. 1983.
J. A New Method for Estimating Mean Seasonal and
Annual Rainfall for the Hunter Valley, New SouthWales. Australian Meteorologi
cal Magazine
,
31; 179-184.
Hutchinson, M. and P. Gessler. 1994. Splines – more than
just a smooth interpolator.
Geoderma,
62, 45-67.
Hutchinson, M.F., Kalma, J.D., and Johnson, M.E. 1984. Monthly Estimates of Wind
Speed and Wind Run for Australia. Journal of Climatology, 4, 311-324
Inggs, M.R. and Lord R.T., 1995. Interpolating Satellite Derived Wind Field Data Using
Ordinary Kriging, with Application to the Nadir Gap, Proc. IEEE Geoscience Remote
Sensing Symp., IGARSS'95, Firenze, Italy, vol. 1, pp. 141-143, July 1995.
Ingram, P., 1998. An Introduction to Geostatistics. School of Earth Sciences, and Kriging.
Macquarie University, Sydney, NSW, 2109, Australia.
http://137.111.98.10/users/pingram/geostat.html and
http://atlas.es.mq.edu.au/users/pringram/krig.html#hrgprcss
Isaaks, E.H., 1989. Applied Geostatistics. Oxford University Press, New York and
Toronto. 561 pages.

Geostatistics Without Tears 98

Isaaks, E.H. and Srivastava, R.M., 1989. An Introduction to Applied Geostatistics. Oxford
University Press, New York.
Jaakkola, O.A. and Oksanen, J., 2000. Creating DEMs from Contour Lines --- Interpolation
Techniques Which Save Terrain Morphology. GIM International, September.
Jacquez, G., 2005. Exploratory Spatial Data Analysis. University of Michigan, D
epartment
of Epidemiology.http://zappa.nku.edu/~longa/geomed/modules/esda/index.html
Jones, J.G., Thomas, R.W., and Earwicker, P.G. 1989. Fractal Properties of Computer-
Generated and Natural Geophysical Data. Computer & Geosciences, 15(2):227-235.

Jones, T.A.,
With the Hamilton,
Computer. VanD.E., and Johnson,
Nostrand C.R.,
Reinhold, 1986. England.
London, Contouring Geological Surfaces

Jones, P.G. 1995. Centro Internacional de Agricultura (CIAT) climate database version
3.41, Digital data tape, Cali, Columbia: CIAT.
Jones, P.G., Robison, D.M. and Carter, S.E. 1990. A geographical information approach
for stratifying tropical Latin America to identify research problems and opportunities in
natural resource management for sustainable agriculture in Centro Internacional de
Agricultura Tropical (CIAT). Cali, Colombia: Agroecological Studies Unit., CIAT.
Journel, A.G., 1974 Geostatistics for Conditional Simulation of Ore Bodies. Economic
.

Geology: Volume 69 pp.673-687.

Journel, A.G., 1983. Mathematical Geology, 445-468.
Journel, A.G. and Arik, A., 1988. In Computer Applications in the Mineral Industry;
Fytas, et. Al., Editors; Balkema: Rotterdam; 161-171.
Journel, A.G. and Hujibregts, C.J., 1978.Mining Geostatistics. Academic Press, New York.
Kalma, J.D., Speight, J.D. and Wasson, R.J. 1988. Potential Wind Erosion in Australia: A
continental perspective. Journal of Climatology, 8, 411-428.
Kent, J.T. and Mardia, K.V. 1994. The link between kriging and thin-plate splines. In
Probability, Statistics and Optimization (ed. F.P. Kelly). Wiley, NewYork. 324-339.
Kesterven J., and Hutchinson, M. 1996. Spatial Modeling of Climate Variables on a
Continental Scale. Proceedings of the Third International Conference Integrating GIS and
Environmental Modeling. University of California, Santa Barbara, National Center for
Geographic Information and Analysis: CD-ROM.
Kitanidis, P., 1985. ‘Minimum-Variance Quadratic Estimation of Covariances of
Regionalized Variables.’ Mathematical Geology, 17(2):195-208.
Kitanidis, P.K., 1997. Introduction to Geostatistics: Applications to Hydrogeology.
Cambridge University Press, New York. 249 pages.
Kitterod NO. Langsholt, E. Wong, W.K. and Gottschalk, L., 1997. Geostatistic
Interpolation of Soil Moisture. Nordic Hydrology. 28(4-5): 307-328.
Klinkenbery, B., 2002. Geography 516, Graduate GIS Seminar: Linking Statistical
Models with Spatial Data. Department of Geography, University of British Columbia,
Vancouver, B.C. http://www.geog.ubc.ca/courses/geog516/notes/statistics_spatial.htm
Klinkenberg, B. and Goodchild, M.F. 1992. The Fractal Properties of Topography: A
Comparison of Methods. Earth Surface Process and Landforms, Vol. 17, 217-234.

Geostatistics Without Tears 99

Knudsen, H.P. and Baafi, E.Y., 1987. Indicator Kriging and Other New Geostatistical
Tools. Pacific Rim Congress 87.
Kohl, M. and Gertner, G., 1997. Geostatistics in Evaluating Forest Damage Surveys -
Considerations on Methods for Describing Spatial Distributions. Forest Ecology &
Management. 95(20): 131-140.
Krige, D.G., 1951. A Statistical Approach to Some Basic Mine Valuation Problems on the
Witwatersrand. Journal of the Chemical and Metallurgical Mining Society of South Africa
52: 119-139.
Krige, D.G., 1959. A Study of the relationship between development values and recovery
grades on the South African goldfields”. Journal of the South African Institution of
Mining and Metallurgy, No. 61, 317-331.
Krige, D.G., 1966. Two Dimensional Weighted Moving Average Trend Surfaces for Ore-
Evaluation. Journal of the South African Institute of Mining and Metallurgy 66:13-38.
Krige, D.G., 1981. Lognormal-de Wijsian geostatistics for ore evaluation. South African
Institute of Mining and Metallurgy Monograph Series. Geostatistics I. South Africa
Institute of Mining and Metallurgy, Johannesburg, South Africa.
Krige DG, Guarascio M, and Camisani-Calzolari FA., 1989. Early South African
Geostatistical Techniques in Today's Perspective. In: Armstrong M, editor. Geostatistics.
Amsterdam: Kluwer Academic Publisher. p 1-19.
Krivoruchko, K., 1998. GIS and Geostatistics: Spatial Analysis of Chernobyl’s
Consequences in Belarus. Workshop on Status and Trends in Spatial Analysis. Santa
Barbara, California., December 10-12, 1998. 20 pages.
http://www.ncgia.ucsb.edu/conf/sa_workshop/papers/ncgia.html

Krivoruchko, K., 2000. Applied Reservoir Characterization Using Geostatistics: The Value
of Spatial Modeling. AAPG Hedberg Conference, Houston, Texas, December 3-6.
Krivoruchko, K., 2002a. Introduction to Modeling Spatial Processes Using Geostatistical
Analyst. ESRI Australia.http://zappa.nku.edu/~longa/geomed/modules/esda/index.html
Krivoruchko, K. 2002b. Bridging the GapBetween GIS and Solid Spatial Statistics. New
Tools for Spatial Data Analsysis: Procceding of the CSISS Specialist Meeting, Santa Barbara,
California, May 10-11, 2002. University of California, Santa Barbara, Center for Spatially
Integrated Social Science.http://www.dpi.inpe.br/gilberto/csiss/papers/krivoruchko.pdf
Krivoruchko, K., 2002. Working on Nonstationarity Problems in Geostatistics Using
Detrending and Transformation Techniques: An Agricultural Case Study. paper was
presented at the Joint Statistical Meetings, New York City, August 2002. Available at
http://www.esri.com/software/arcgis/arcgisxtensions/geostatistical/research_papers.html
and
http://campus.esri.com/campus/library/ConfProc/GeostatisticsTeam/Krivoruchko_2002_
Working.pdf
Krivoruchko, K. 2005. Introduction toModelling Spatial Processes Using Geostatisti
cal
Analyst. ESRI White Papers.http://www.esri.com/library/whitepapers/pdfs/intro-
modeling.pdf
Kufs, C., 1987. Trend-Surface Modeling of Groundwater Data. Management of
Uncontrolled Hazardous Waste Sites. Silver Springs, MD. Hazardous Materials Control
Research Institute. http://members.home.net/terrabyte/t_surf.pdf
Kufs, C., 1998. Interpolation for Fun and Profit.
http://members.home.net/terrabyte/didyou.htm

Geostatistics Without Tears 100

LaGro, Jr., J. 1991.Assessing Patch Shape inLandscape Mosaics. Photogramme
tric Engineering
& Remote Sensing,57(3):285-293.
Larkin, R.P., Gumpertz, M.L., and Ristaino, J.B., 1995. Geostatistical analysis of
Phytophthora
epidemic development in commercial bell pepper fields. Phytopathology 85:191-203.
Le Loc’h, Beucher G., Galli, A., Doligz, B., and
Group, H., 1994. Improvement in the truncated
Gaussian method: Combining several Gaussian Functions, in Proceedings of ECMORE IV,
Fourth European Conference onthe Mathematics of Oil Recovery.
Le Loc’h, G. and Galli, A.1997 Truncated plurigaussian method: Theoretical and practical points
of view, in Geostatistical Simulations, pp. 211-222, Kluwer Academic.
Lecoustre, R. Fargette, D., Fauquet, C. and de Reffye, P., 1989. Analysis and Mapping of
the Spatial Spread of African Cassava Mosaic Virus Using Geostatistics and the kriging
technique. The American Phytopatholgical Society. Vol 79, No. 9: 913-920.
Leduc, A., Prairie, Y.T., Bergeron, Y., 1994. Fractal dimension estimates of a fragmented
landscape: sources of variability. Landscape Ecology Volume 9 Number 4 pp 279-286.
SPB Academic Publishing bv, The Hague.
http://landscape.forest.wisc.edu/landscapeecology/articles/v09i04p279.pdf
Liebhold, A.M., Rossi, R.E., and Kemp, W.P., 1993. Geostatistics and Geographic
Information Systems in Applied Insect Ecology. Gypsy Moth Slow the Spread
Foundation, Inc. http://www.gmsts.org/pubs/liebhold_1993.pdf
Lembo, A.J., Jr., 2005. Spatial Interpolation, Lecture Power Point Presentation, Cornell
University. http://www.css.cornell.edu/courses/420/lecture18.ppt
Lemmar, I.C., 1985. The Mononodal Cutoff-Application to Robust Variography and
Spatial Distribution Estimation, Ph.D. Thesis, Stanford University.
Levine, N. & Associates. 1999. CrimeStat (Version 1.0) – A Spatial Statistics Program for
the Analysis of Crime Incident Locations. The National Institute of Justice, Washington,
D.C. http://www.ojp.usdoj.gov/cmrc/
Little, L.S., Edwards, D. and Porter, D.E., 1997. Kriging in Estuaries - As the Crow Flies,
Or As the Fish Swims. Journal of Experimental Marine Biology & Ecology. 213(1):1-11.
Liu, H.K., Jezek, B. L., and Zhao, Z. 2001. Radarsat Antarctic Mapping Project Digtial
Elevation Model Version 2. NDSIC, Boulder, Colorado.
Liu, N., and Oliver, D., 2003. Conditional Simulation of Truncated Random Fields Using
Gradient Methods. International Association for Mathematical Geology (IAMG). IMAG
2003 Proceedings. http://www.iamg.org/meetings/Proceedings_2003/papers/Liu.pdf
Loránd, 2005. Kriging Interpolation. Eötvös Loránd University, Faculty of Informatics,
Department of Cartography and Geoinformatics.
http://lazarus.elte.hu/hun/digkonyv/havas/mellekl/vm25/vma07.pdf

Lorenc,
Scheme.A., 1981: A
Monthly GlobalReview,
Weather Three-Dimensional Multivariate Statistical Interpolation
109, 701-721.
Lovejoy, S., 1982. Area-Perimeter Relation for Rainand Cloud Areas. Science, 216:185-187.
Lynch, S.D. and Schulze, R.E., 1997. Techniques for Estimating Areal Daily Rainfall.
Department of Agricultural Engineering, University of Natal, Private Bag X01,
Scottsville, 3209, South Africa. http://www.ccwr.ac.za/~lynch2/p241.html

Geostatistics Without Tears 101

Maclean, A.L., Cleland, D.T., The Use of Geostatistics to Determine the Spatial Extent of
Historical Fires as an Aid in Understanding Fire Regimes for Northern Lower Michigan.
USDA Forest Service, Great Lakes Ecological Assessment Reports. In press: Proceedings
of a Conference on Fire, Fuel Treatments and Ecological Restoration. April 16-18, 2002.
Ft.Collins, CO. Rocky Mountain Research Station General Technical Report.
http://www.ncrs.fs.fed.us/gla/reports/geostatisticsinfire.pdf
Mandelbrot, B.B., 1975. Stochastic Models for the Earth’s Relief, the Shape and the Fractal
Dimension of Coastlines, and the Number-Area Rule for Islands. Proceedings of the National
Academy of Sciences, U.S.A. 72, 3825-3828.
Mandelbrot, B.B., 1977. Fractals, Form, Chanceand Dimension. Freeman, San Francisco.
Mandelbrot, B.B., 1982. The Fractal Geometry of Nature. W.H. Freeman, London, New York.
Mandelbrot, B.B. and van Ness, J.W. 1968. Fractional Brownian Motions, Fractional
Noises and Applications, Siam Review, 10(4), 422-437.
Mantoglou, A. and Wilson, J.L., 1982. The turning bands method for simulation of
random fields using line generation by a spectral method: Water Resources Research Vol.
18, No.5, pp.1379-1394.
Mark, D.M., 1997. The History of Geographic Information Systems: Invention and
Reinvention of Triangulated Irregular Networks (TINs). Proceedings, GIS/LIS’97.
http://www.geog.buffalo.edu/ncgia/gishist/GISLIS97.html
Matheron, G., 1962. Traité de geostatisque appliqueée, vol. I: Memories du Bureau de
Géologiques et Miniéres, no 14, Editions Techniq, Paris. 333 pages.
Matheron, G., 1963a. Principles of Geostatistics. Economic Geology 58: 1246-1266.

Matheron, G., 1963b. Traité

Bureau de Recherches de geostatistique
Géologiques appliqée,
et Miniéres, vol. II, LeBureau
no 24. Éditions krigeage: Memories du
de Recherche
Géologiques et Miniéres. Paris. 171 pages.
Matheron, G., 1965. Les Variables Régionalisées et Leur Estimation. Une Application de
la Thérie des Fonctions Aléatoires aux Sciences de la Nature, Masson, Paris. 305 pages.
Matheron, G., 1971. The Theory of Regionalized Variables and its Applications. Cahiers
du Center de Morphologie Mathematique, Fontainbleau, No. 5.
Maurer, B.A., 1994. Geographical Population Analysis. Blackwell Scientific, Oxford, UK.
Maurer, B.A. and Villard, M.-A., 1994.
Population Density: Geographic Variation in Abundance
of North American birds. National Geographic Research
and Exploration 10: 306-317.
Meinguet, J. 1979. Multivariate Interpolation at Arbitrary Points Made Simple, Journal of
Applied Mathematics and Physics, 30, pp. 292–304
McBratney, A.B. and Webster, R., 1981. Spatial Dependence and Classification of the
Soil along a Transect in Northeast Scotland. Geoderma, 26:63:82.
McBratney, A.B. and Webster, R., 1986. Choosing Functions for Semi-variograms of Soil
Properties and Fitting them to Sampling Estimates, Journal of Soil Science 37: 617-639.
Milewska, E.J., Hopkinson, R.F. and Niitsoo, A. 2005. Evaluation of Geo-Referenced
Grids of 1961-1990 Canadian Temperature and Precipitation Normals. Atmosphere-Ocean
43 (1) 2005, 49-75. Canadian Meteorological and Oceanographic Society.
http://www.cmos.ca/Ao/Abstracts/v430104.pdf

Geostatistics Without Tears 102

Mitchel, N.D. 1991. The Derivation of Climate Surface for New Zealand, and Their
Application to the Bioclimatic analysis of the Distribution of Kauri (Agathis australis).
Journal Royal Society of New Zealand. 21, 13-24.
Moran, P.A.P., 1948. The interpretation of statistical maps. Journal of the Royal Statistical
Society, Series, B, 10, 243-250.
Moran, P.A.P., 1950. Noteson Continuous Stochastic Phenomena. Biometrika, 37, 1950; 17-23.
Mitasova, H. and Mitas, L., 1993. Interpolation by Regularized Spline with Tension: I.
Theory and Implementation. Mathematical Geology, Vol. 25, No. 6: 641-655.

Mitasova,
ApplicationH.toand Hofierka,
Terrain J., 1993.
Modeling and Interpolation by Regularized
Surface Geometry Spline with Tension:
Analysis. Mathematical II.
Geology,
Vol. 25, No. 6: 657-669.
Myers, J., 1997. Geostatistical Error Management: Quantifying Uncertainty for
Environmental Sampling and Mapping, Van Nostrand Reinhold, New York.
Myers, D.E., 1999. What is Geostatistics?http://www.u.arizona.edu/~donaldm/whatis.html
Nalder, I.A. and Wein, R.W. 1998. Spatial interpolation of climatic Normals: test of a new
method in the Canadian boreal forest. Agric. For. Meteorol.92: 211-225.
Ned Levine & Associates, 2004. Chapter 5 – Distance Analysis I and II. CrimeStat III, The
national Institute of Justice, Washington, DC.
http://www.icpsr.umich.edu/NACJD/crimestat.html/and
http://www.icpsr.umich.edu/CRIMESTAT/files/CrimeStatChapter.5.pdf
New, M., and Hulme. M. 1997. Monthly Rainfall and Temperature Surfaces for Africa:
1951-1995 (CD-ROM) information: Production of Climate Surfaces. Climatic Research
Unit, University of East Anglia Norwich NR4 7TJ, UK.
http://www.mara.org.za/climatecd/info.htm
NIST, 2005. Engineering Statistics Handbook. NIST/SEMATECH e-Handbook of Statistical
Methods, Home:http://www.itl.nist.gov/div898/handbook/index.htm , Normal Probability
Plot: http://www.itl.nist.gov/div898/handbook/eda/section3/normprpl.htm
, Quanitile-
Quantile Plot:http://www.itl.nist.gov/div898/handbook/eda/section3/qqplot.htm
Olea, R., Ed., 1991. Geostatistical Glossary and Multilingual Dictionary, New York:
Oxford University Press.
Oliver, M.A. and Webster, R., 1991. How Geostatistics Can Help You. Soil Use and
Management. Volume 7, Number 4, December 1991: 206-217.
Oliver, M., Webster, R. and Gerrard, J., 1989. Geostatistics in Physical Geography. Part I:
theory. Trans. Inst. Br. Geogr. N.S. 14: 259-269 (1989) ISSN:0020-2754.
Oliver, M. Webster, R. and Gerrard, J., 1989. Geostatistics in Physical Geography. Part II:
applications. Trans. Inst. Br. Geogr. N.S. 14: 270 -286 (1989) ISSN:0020-2754.
Omre, H., 1984. The Variogram and its Estimation, in Verly, G., David, M., Journel, A.
G. and Marechal, A. (eds) Geostatistics for Natural Resources Characterization (Reidel,
Dordrecht): 107-125.
Onsoy, Y.S., Harter, T., Ginn, T.R., and Horwath, W.R., 2005. Spatial Variability and
Transport of Nitrate in a Deep Alluvial Vadose Zone. Published in Vadose Zone Journal
4:41-54 (2005), Soil Science Society of America 677 S. Segoe Rd., Madison, WI 53711
USA. http://groundwater.ucdavis.edu/Publications/Onsoy%20et%20al%20-
%20Spatial%20variability%20of%20nitrate%20-%20Vad%20Zone%20J%202005.pdf

Geostatistics Without Tears 103

Oregon State University. 2006.Spatial Climate Analysis Service.
http://www.ocs.orst.edu/prism/
Palmer, M.W., 1988. Fractal geometry: a tool for describing spatial patterns of plant
communities. Vegetatio 75: 91-102.
Parker, H.M. Journal, A.G. and Dixon, W.C., 1979. The Use of the Conditional
Lognormal Probability Distribution for the Estimation of Open-pit Ore Reserves in
Stratabound Uranium Deposits – A Case Study”, Proceedings 16th APCOM, 133-148.
Pebesma, E.J. and Wesseling, C.G., 1998. Gstat: a program for geostatistical modelling,
prediction and simulation. Computers & Geosciences, 24(1): 17-31.

Petrone, R.M., Price,

characterization of theJ.S., Carey,
spatial S.K., and
variability Waddington,
of soil J.M.,
moisture in 2004. Statistical
a cutover peatland.
Hydrological Processes 18, 41-52. Published online 2003 in Wiley InterScience. DOI:
10.1002/hyp.1309. http://www.gret-perg.ulaval.ca/Petrone_et_al_HP18_2004.pdf
Phillips, D.L., Dolph, J. and Marks, D.1992. A Comparison of Geostatistical Procedures for
Spatial Analysis of Precipitation in Mountainous Terrain. Agric. Forest. Meteor., 58, 199-141.
Pincock. 2001. Geostatistics – What Exactly Does All this Mumbo-Jumbo Mean? Pincock
Solutions newsletter, Issue No. 20 – July 2001. Lakewood, Colorado 80228.
http://www.pincock.com/Perspectives/Issue20-Geostatistics.pdf
Poiker, T. and Strobl, J., 2000. UniGIS: Course 7, Spatial Operations, Class Notes. Simon
Fraser University and the University of Salzburg, 2 nd Edition. 99 pages.
Price, D.T., McKenney, D.W., Nalder, I.W., Hutchinson, M.F. and Kesteven, J.L. 2000. A
Comparison of Two Statistical Methods for Spatial Interpolation of Canadian Monthly
Mean Climate Data. Agriculter For. Meteorology. 101: 81-94 (2000).

Pyrcz, M.J., and

Mathematical Deutsch,
Geology: C. V., 2001.
Volume Two
33 No.7: artifacts of probability field simulation.
pp.775-800.
Radeloff, V.C., Miller, T.F., He, H.S. and Mladenoff, D.J., 2000. Periodicity in spatial
data and geostatistical models: autocorrelation between patches. Ecography 23: 81-91.
Copenhagen. http://landscape.forest.wisc.edu/pdf/Radeloff_etal2000_Ecog.pdf
Ramierz-Beltrán, N.D., Winter, A., Veneros, A., Rengifo, F.C., Escalante, N.R., 2006.
Neural Networks Estimate Atmospheric Variables in the Caribbean Basin. American
Meteorological Society.http://ams.confex.com/ams.confex.com/ams/pdfpapers/55135.pdfm
Rendu, J-M.M., 1979. Kriging, Logarithmic Kriging, and Conditional Expectation:
Comparison of Theory with Actual Results, Proceedings 16 th APCOM, 199-212.
Ricci, A.K., 1998. Geostatistics Using SAS. NorthEast SAS Users Group (NESUG) 1998
Conference. October, 4-7, Pittsburgh, PA.
http://www.nesug.org/html/Proceedings/nesug98/stat/p003.pdf
Ringsby, H., B-E Saether, J. Tufto, H. Jensen, and E.J. Solberg. 2001. Asynchronous
Spatiotemporal Demography
Environment. Ecology: of aNo.
Vol. 83, House Sparrow Metapopulation in a Correlated
2, 561-569.
Ripley, B.D., 1977. Modelling spatial point patterns (with discussion). Journal of the
Royal Statistical Society. B39, 1972-212.
Rivoirard J. 1994 Introduction to Disjunctive Kriging and Non-Linear Geostatistics.
Claredond Press, Oxford, New York. 180 pages.

Geostatistics Without Tears 104

Rivoirard J., 2000. Concepts and Methods of Geostatistics. Centre de Géostatistique,
Ecole des Mines de Paris. http://www.fast.u-psud.fr/~cargese/papers/chiles/chiles1.pdf
Robeson, S.M. and Janis, M.J., 1998. Comparison of temporal and unresolved spatial
variability in multiyear time-average of air temperature. Climate Research, Vol. 10: 15-26.
http://www.int-res.com/articles/cr/10/c010p015.pdf
Robertson, G.P., 1987. Geostatistics in Ecology: Interpolating with Known Variance.
Ecology 68:744 -748.
RocNews, 2003. Geostatistics. Article prepared for RocNews Spring 2003.
http://www.rocscience.com/library/rocnews/Spring2003/GeostatisticsArticle.pdf
RocScience. 2006.Interpolation Method.
http://www.rocscience.com/downloads/phase2/webhelp/phase2_mod
el/Interpolation_Method.ht
m
Rouhani, S., Srivastava, R., Desbarats, A, Cromer, M, and Johnson, A. editors. 1996.
Geostatistics for Environmental and Geotechnical Applications, ASTM STP 1283. Special
Technical Publication of the ASTM.
Rosenbaum, M., and Söderström, M., 1997. Geostatistics as an Aid to Mapping.
http://www.esri.com/base/common/userconf/europroc96/PAPERS/PN11/PN11F.HTM
Rossi, R.E., Mulla, D.J., Journel, A.G.
and Franz, E.H., 1992.Geostatistical Tools forModeling
and Interpreting Ecological Spatial Dependence. Ecological Monographs 62: 277-314.
Royle, A.G., 1979. A Practical Introduction To Geostatistics. Department of Mining and
Mineral Sciences, University of Leeds. 103 pages.
Russo, D., 1984. Design of an Optimal Sampling Network for Estimating the Variogram.
Soil Science Society of America J. 48: 708 - 716.
Saito, H., and Goovaerts, P. 2000. Geostatistical Interpolation of Positively Skewed and
Censored Data in a Dioxin-Contaminated Site. Environmental Science & Technology,
volume 34, number 19, 4228-4235.
Sarmiento, J.L., J. Willebrand and S. Hellerman. 1982. Objective Analysis of Tritium
Observations in the Atlantic Ocean During 1971-74. Ocean Tracers Laboratory Technical
Report #1.
Schotzko, D.J., and O’Keeffe, L.E., 1989. Geostatistical Description of the Spatial
Distribution of Lygus Hesperus (Heteroptera: Miridae) in Lentils. Journal Economic
Entomology. 82:1277-1288.
Sharov, A., 1996. Quantitative Population Ecology.
http://www.gypsymoth.ento.vt.edu/~sharov/PopEcol/popecol.html
Shepard, D., 1968. A Two-Dimensional Interpolation Function for Irregularly Spaced Data.
Proceedings, 23rd National Conference, Association for Computing Machinery, Brabdon
Syst. Press, 517-524.
Shibli, S.A.R., 1997. Geostatistics FAQ. http://curie.ei.jrc.it/faq/index.html
Sichel, H.S., 1971. On a Family of Discrete Distributions Particularly Sited to Represent
Long Tailed Frequency Data, in Laubscher, N.F., (editor). Proceedings of the third
symposium on Mathematical Statistics, Pretoria, 51-97.
Sichel, H.S., 1973. Statistical Valuation of Diamondiferous Deposits. Journal of the South
African Institute of Mining and Metallurgy, February, 2335-243.

Geostatistics Without Tears 105

Simpson, J.J., Hufford, G.L., Daly, C., Berg, J.S., and Fleming, M.D. 2005. Comparing
Maps of Mean Monthly Surface Temperature and Precipitation for Alaska and Adjacent
Areas of Canada Produced by Two Different Methods. Arctic. Vol. 58, No. 2, pp 137-161
http://www.ocs.orst.edu/pub/prism/docs/ComparingMapsFinalProof.pdf
Siska, P.P. and I-Kuai Hung, 2001. Propagation of Errors in Spatial Analysis. Presented
in: the 24th Applied Geographic Conferences Vol. 24, Fort Worth, Texas. College of
Forestry, Stephen F. Austin University, Nacogdoches, Texas, 75692-6901.
http://www.faculty.sfasu.edu/ihung/pdf/research/agc24.pdf
Smith, L., R.J. Hyndman, S.N. Wood. 2004. Spline interpolation for demographic
variables: the monotonicity problem. Journal of Population Research, May, 2004.
http://www.findarticles.com/p/articles/mi_m0PCG/is_1_21/ai_n6155266#continue
Sobieraj, J.A., Elsenbeer, H., Cameron, G., 2003. Scale dependency in spatial patterns of
saturated hydraulic conductivity. www.elevier.com/locate/catena and http://www.uni-
potsdam.de/u/Geooekologie/download/elsenbeer/publikationen/RG5.pdf
Sokal, R.R. and Oden, N.L., 1978. Spatial Autocorrelation in Biology 1. Methodology 2.
Some Biological Implications and Four Applications of Evolutionary and Ecological
Interest. Biological Journal of the Linnean Society 10:199-228.
Solow, A.R., 1990. Geostatistical Cross-Validation: A Cautionary Note. Mathematical
Geology, Vol. 22, No. 6: 637-639.
Sotter, A.P.R, Ipia, A.H.S., Pulido,J.W.C., Vaca,W.L.S., 2003. Applied Geostatistics to
Studies of Environmental Contamination, Alternative Analysis Viewpoints for Air
Pollutants in Bogotá, D.C. ESRI 2003 User Conference Proceedings.
http://gis.esri.com/library/userconf/proc03/p0442.pdf
Srivastava, R.M., 1988. A Non-ergodic Framework for Variograms and Covariance
Functions. SIMS Technical Report No. 114, Department of Applied Earth Sciences,
Stanford University.
Stein, Michael, L., 1999. Interpolation of Spatial Data. Springer Series in Statistics. 256 pages.
Sterk, G. and Stein, A., 1997. Mapping Wind-Blown Mass Transport by Modeling
Variability in Space and Time. Soil Science Society of American Journal. 61(1): 232-239.
Sullivan, J., 1984. Conditional recovery estimation through probability kriging – theory
and practice. In: Geostatistics for natural resources characterization, Part 1. Ver, G. et al.
(Eds), Reidel (Dordrecht), pp. 365-384.
Surfer, 1995. Surfer for Windows, Version 6 User’s Guide – Contouring and 3D Surface
Mapping. Golden Software, Inc. Golden, Colorado, 80401-1866, U.S.A.
Surfer, 1999. Surfer for Windows, Version 7 User’s Guide – Contouring and 3D Surface
Mapping for Scientists and Engineers. Golden Software, Inc. Golden, Colorado, 80401-
1866, U.S.A. www.goldensoftware.com

Surpac Minex, 2004. Polygon Kriging. Surpac Minex Group, Perth Western Australia
6850. http://www.surpac.com/refman/default/stats/pkrig.htm
Swanson, R.T., Ritz, R. L. , McAtee, M.D. 2001. Operational Mesoscale Data
Assimilation at Air Force Weather Agency Using a Parallelized Multivariate Optimal
Interpolation Scheme. PSU/NCAR Mesoscale Modeling System Users’ Workshop 2001.
http://www.mmm.ucar.edu/mm5/workshop/ws01/swanson.doc
Syed Abdul Rahman Shibli. 1997. Spatial Correlation: General. AI-GEOSTATS.
http://www.ai-geostats.org/Geostats_Faq/Syed/sc_general.html

Geostatistics Without Tears 106

Syed Abdul Rahman Shibli., 1997. Conditional simulation: implementation. AI-
GEOSTATS. http://www.ai-geostats.org/Geostats_Faq/Syed/cs_implement.html
Syed Abdul Rahman Shibli., 1998. Geostatistics FAQ. AI-GEOSTATS. Desa Pandan
55100 Kuala Lumpur, Malaysia. http://curie.ei.jrc.it/faq/index.html

Taghvakish, S. and Amini, J., 2004. Optimum Weight in Thin-plate Spline for Digital
Surface Model Generation. TS26 Positioning and Measurement Technologies and
Practices II. FIG Working Week, 2004. Athens, Greece, May 22-27, 2004.
http://www.fig.net/pub/athens/papers/ts26/TS26_6_Amini_Taghvakish.pdf
Tait, A., and Zheng, X. 2005. Final Report: Optimal Mapping and Interpretat
ion of Fire
Weather Information, prepared for New Zealand Fire Service. NWA Client Report:
WLG2005-1, June 39, 2005. National Institute of Water andAtmospheric Research (NIWA)
Project: NZF03301 / Year 2.http://www.fire.org.nz/research/reports/reports/Report_49.pdf
Tangmar, B.B., Yost, R.S., and Uehara, G., 1985. Application of Geostatistics to Spatial
Studies of Soil Properties. Advanced Agronomics, 38:45-94.
Taylor, G.H., Daly, C., W.P. Gibson, andSibul-Weisberg, J. 1997. Digital andMap Products
Produces using PRISM. Proceedings 10th AMS Conference on Applied Climatology,
American Meteorological Society 20-24 October 1997,Reno, Nevada pp. 217-218
Ten Berge, H.F.M., Stroosnijder, L., Burrough, P.A., Bregt, A.K., and de Haes, M.J.
1983. Spatial Variability of Soil Properties Influencing the Temperature of the Soil
Surface, Agricultural Water Management, 6:213-226.
Thomas, G.S., 1997. Interactive Analysis and Modeling Semi-Variograms. Snowden
Associates Pty Ltd. Australia.
http://www.metal.ntua.gr/msslab/MineIT97/papers/GT67/GT67.html.
Thornton, P.E., Running, S.W., and White, M.A. 1997. Generating Surfaces of Daily
Meterological Variable over Large Regions of Complex Terrain. Numerical
Terradynamics Simulation Group, School of Forestry, University of Montana, Missoula.
Journal of Hyrdology 190 (1997) 214-251.
http://www.cgd.ucar.edu/tss/aboutus/staff/thornton/pubs/thornton_jh_1997.pdf
Thurston, J. (editor), 2003. Looking Back and Ahead: The Triangulated Irregular Network
(TIN). GeoInformatics, October. http://www.vectorone.info/Publish/tin.pdf
Tobler, W., 1979: A Transformational View of Cartography. The American Cartographer,
6, 101-106.
Todini, E., and Ferrarcsi, M., 1996. Influence of Parameter Estimation Uncertainty in
Kriging. Journal of Hydrology 175: 555-566.
Tolosana-Delgado, R., and Pawlowsky-Glahn. 2006. Kriging Coordinates: What Does that
Mean? Departament d’Informàtica i Matemàtica Aplicada Universitat de Girona, E-17071
,
Girona, Spain.http://ima.udg.es/Activitats/CoDaWork03/paper_Tolosana_Pawlowsky.pdf
Tomczk, M., 1998. Spatial Interpolation and it Uncertainty Using Automated Anisotropic
Inverse Distance Weighting (IDW) - Cross-Validation/Jackknife Approach. Journal of
Geographic Information and Decision Analysis, Vol. 2, No. 2: 18-33.
ftp://ftp.geog.uwo.ca/SIC97/Tomczak/Tomczak.html
Tsai, V.J.D., 1993. Delaunay triangulations in TIN creation: An overview and a linear-
time algorithm. Int. J. Geographic Information. Science, 7~6!, 501-524.

Geostatistics Without Tears 107

Turcotte, D.L. 1992. Fractals and Chaos in Geology and Geophysics, Cambridge
University Press, Cambridge.
University of Queensland, Mathematics. 1999. MN309 -- Advanced Numerical Analysis:
22 pages. http://www.maths.uq.oz.au/~gac/mn309/bspl.html
USGS. 2003. GTOP30 Documentation. USGS-NASA.
Vann, J., and Guibal, D., 2000.Beyond ordinary kriging – An overview of non-linear estimation.
In: Mineral Resource and Ore Reserve Estimation – The AusIMM guide to good practice
(Monograph 23): pp. 249-256.The Australasian Institute of Mining and Metallurgy: Parkvill
e.
[srcinally published as: Vann, J., andGuibal, D., 1998. Beyond ordinary kriging – An
overview
of non-linear estimation. In: Vann, J., (Ed.), Beyond
Ordinary Kriging Seminar, October 30th,
1998 Perth, Western Australia. Geostatistical Association ofAustralasiaMonograph 1:Perth.].
http://www.quantitativegeoscience.com/images/pdf/vann_guibal_beyond_ok. pdf
Vann, J., Guibal, D., and Harley, M., 2000 Multiple indicator kriging – Is it suited to my
.

deposit? Proceedings of the 4 th International Mining geology Conference, Coolum,

Queensland: pp.171-179.The Australasian Institute of Mining and Metallurgy: Parkville.
Vann, J., Bertoli, O. and Jackson, S. 2002. An Overview of Geostatistical Simulation for
Quantifying Risk. Paper srcinally published at a Geostatistical Association of Australasia
symposium “Quantifying Risk and Error” March 2002.
http://www.quantitativegeoscience.com/images/pdf/vann_bertoli_jackson_simulation_for
_risk_distribution.pdf
Vanouplines, P., 1995. Rescaled Range Analysis and the Fractal Dimension of pi.
University Library, Free University Brussels, Brussels Belgium.
http://homepages.vub.ac.be/~pvouplin/pi/fracwhat.htm
Vasil’yev, L.N., and Tyuflin, A.S. 1992. Fractal Characteristics of Geostystem Spatial
Structure from Space Imagery, Mapping Sciences and Remote Sensing, 29:93-102.
Verly, G., and Sullivan, J., 1985. Multigaussian and probability kriging – an application to
the Jerrit Canyon deposit. Mining Engineering, June, 1985, pp. 568-574.
Villard, Marc-André and Maurer, B.A., 1996. Geostatistics as a Tool for Examining
Hypothesized Declines in Migratory Songbirds. Ecology, 77(1), 1996: 59 - 68.
Vivoni, E.R., Ivanov V.Y., Bras R.L., and Entekhabi D., 2004. Generation of Triangulated
Irregular Networks Based on Hydrologic Similarity.
http://hydrology.mit.edu/viva/research/publications/Vivoni2004TINprint.pdf
Wackernagel, H., 1994. Cokriging versus Kriging in Regionalized Multivariate Data
Analysis. Geoderma, 62 (1994): 83-92.
Wackernagel, H., 1995. Multivariate Geostatistics: An Introduction with Applications.
Springer, Berlin, New York. 256 pages.
Wahba, G. 1979. How to Smooth Curves and Surfaces with Splines and Cross-validation.
Proceedings 24th Conference on the Design of Experiments. US Army Research Office
79-2, Research Triangle Park, NC: 167-192.
Wahba, G. 1990. Spline Models for Observational Data. CBMS-NSF Regional
Conference Series in Applied Mathematics 59, SIAM, Philadelphia, Pennsylvania.
Wahba, G. and Wendelberger, J. 1980 Some New Mathematical Methods for Variational
Objective Analysis Using Splines and Cross Validation. Monthly Weather Review , 108,
1122-1143.

Geostatistics Without Tears 108

Walsh, S.J., Bian, L., and Brown, D.G. 1991. Issues of Spatial Dependency for Surface
Representation through Remote Sensing and GIS. The Integration of Remote Sensing and
Geographic Information System (J.L. Star, editor), ASPRS, Bethesda, Maryland.
Warnes, J.J., 1986. A Sensitivity Analysis for Universal Kriging. Math. Geol, Vol. 18, No.
7: 653-676.
Watson, D.F., and Philip, G.M., 1984. Systematic triangulations, Computer Vision
Graphic Image Processing, 26, 217–223.
Webster, R., 1985. Quantitative Spatial Analysis of Soil in the Field. Advances in Soil
Science Volume 3. Springer-Verlag, New York: 1-70.
Webster, R., Curran, P.J., and Munden, J.W. 1989. Spatial Correlation in Reflected
Radiation from the Ground and Its Implication for Sampling and Mapping by Ground-
Based Radiometry, Remote Sensing of Environment, 29:67-78.
Webster, R. and Oliver, M.A., 1992. Sample Adequately to Estimate Variograms of Soil
Properties. Journal of Soil Science, 43, 177-192.
Webster, R. Oliver, M.A., Muir, K.R. and Mann, J.R., 1994. Kriging the Local Risk of a
Rare Disease from a Register of Diagnoses. Geographic Analysis, 26(2) 168 - 185.
Weber, D. And Englund D., 1992. Evaluation and Comparison of Spatial Interpolators.
Mathematical Geology, Vol. 24, No. 4: 381-391.
Wen, R., and Sinding-Larsen, R. 1997. Uncertainty in Fractal Dimension Estimated from
Power Spectra and Variograms. Mathematical Geology, 29(6):727-753.
Welhan, J. 2005. Geology 606: Geostatistics, Course notes: Week 6-10 – Week 6.a
Lecture Material, Experimental Variography. Idaho State University.
http://giscenter.isu.edu/training/
Wiener, N., 1949. Extrapolation, Interpolation and Smoothing of Stationary Time Series:
MIT Press Cambridge, Massachusetts, 158 pages.
Willmott, C.J., Ackleson, S.G., Davis, R.E., Feddema, J.J., Klink, K.M., Legates, D.R.
O’Donnell L., and Rowe, C.M. 1985. Statistics for the Evaluation and comparison of
Models. Journal of Geophysical Research 90: 8995-9005.
Willmott, C.J., Matsuura, K. 1995. Smart Interpolation of Annually Averaged Air
Temperature in the United States. Journal of Applied Meteorology, 34, 2577-2586.
Willmott, C.J., and Robeson S.M. 1995. Climatologically Aided Interpolation (CAI) of
Terrestrial Air Temperature. International Journal of Climatology 15:221-229.
Willmott, C.J., Rowe, C.M. and Philpot, W.D. 1985. Small-scale Climate Maps: A
sensitivity analysis of some common assumptions associated with grid point interpolation
and contouring. American Cartogr. 12, 5-16.
Wingle, W.L., 1992. Examining Common Problems Associated with Various Contouring
Methods, Particularly Inverse-Distance Methods, Using Shaded Relief Surfaces. Geothech
’92 Conference Proceedings, Lakewood, Colorado.
http://magma.mines.edu/fs_home/wwingle/pub/contour/
Wingle, W.L., and Poeter, E.P., 1998. Classes Versus Thresholds: A Modification to
Traditional Indicator Simulation. American Association of Petroleum Geologists (AAPG)
’98 Annual Convention, Advances in Geostatistics, Salt Lake City, Utah, May 17-20,
1998. http://www.mines.edu/fs_home/wwingle/pubs/aapg_98/class.html

Geostatistics Without Tears 109

Woodcock, C.E., Strahler, A.H., and Jupp, D.L.B. 1988a. The Use of Variograms in
Remote Sensing: I. Scene Models and Simulated Images, Remote Sensing of
Environment, 25:323-348.
Woodcock, C.E., Strahler, A.H., and Jupp, D.L.B. 1988b. The Use of Variograms in
Remote Sensing: II. Real Digital Images, Remote Sensing of Environment, 25:349-379.
Wu, B.M., vanBruggen, A.H.C., Subbarao, K.V., and Pennings, G.G.H., 2000. Spatial
Analysis of Lettuce Downy Mildew Using Geostatistics and Geographic Information
Systems. Phytopathology, The American Phytopathological Society. Publication no. P-
2000-1206-01R. http://www.apsnet.org/phyto/pdfs/2000/1206-01R.pdf

Zamkotowicz, M. 2005. Advanced Applications in GIS, Week 8, Interpolating Surfaces Power

Point Presentation. Burlington County College, New Jersey.
http://staff.bcc.edu/mzamkoto/advancedgis/Powerpoi
nt/Spatial%20Analyst%20Lesson%204.ppt
Zeng, X. and Basher, R. 1995. Thin-Plate Smoothing Spline Modeling of Spatial Climate
Data and Its Application to Mapping South Pacific Rainfalls. National Institute of Water
and Atmospheric Research Limited, Wellington, New Zealand. American Meteorological
Society. Monthly Weather Review, 3086, Volume 123.
http://ams.allenpress.com/pdfserv/10.1175%2F1520-
0493(1995)123%3C3086:TPSSMO%3E2.0.CO%3B2
Zimmerman, D., Pavlik, C., Ruggles, A., Armstrong, M.P. An Experimental Comparison
of Ordinary and Universal Kriging and Inverse Distance Weighting. Mathematical
Geology, Vol. 31. No. 4, 1999.
Zirschky, J., 1985. Geostatistics for Environmental Monitoring and Survey Design.
Environmental International, Vol. 11: 515 -524.
Zrinji, Z., and Burn, D.H., 1992. Application of Kriging to Surface Water Level
Estimation. Canadian Journal of Civil Engineering, 19: 181 - 185.
Zuring, Hans, 2005. F503 GIS: Methods and Applications I Notes – Part 7. University of
Montana, College of Forestry and Conservation.
http://www.forestry.umt.edu/academics/courses/for503/SPATIAL_Stats.htm
Xiao, C.L., Hao, J.J., and Subbarao, K.V., 1997. Spatial Patterns of Microsclerotia of
Verticillum dahliae in Soil and Verticillium Wilt of Cauliflower. The American
Phytopathological Society Vol. 87, No. 3: 325 - 331.

Geostatistics Without Tears 110

INDEX
censored data ................................37, 42
A
central tendency .................................31
A0 ......................................................53 Centre de Geostatistique ....................24
abnormal values.................................39 Centre de Morphologie Mathématique
accommodation.................................. 39 ........................................................24
additive constant.................................41 Channel Tunnel project ......................33
affine anisotropy................................61 CIAT ..................................................13
alternate variograms...........................63 class size.............................................50
angular tolerance................................53 Climatologically Aided Interpolation.13
anisotropic .........................................60 clustered..............................................37
anisotropic variogram ........................60 clustered data......................................63
anisotropies........................................ 72 coefficient of correlation..........27, 34, 35
anisotropy coefficient of determination...........27, 34
affine.............................................. 61 coefficient of skewness ......................27
geometric .......................................61 coefficient of variation...........27, 32, 33
zonal .............................................. 62 cokrige indicators...............................77
anisotropy ratio..................................61 cokriging ....................24, 40, 74, 76, 77
ANUDEM.......................................... 16 co-located samples..............................37
ANUSPIN..........................................12 conditional cosimulation ....................80
ANUSPLIN ........................... 13, 14, 16 conditional simulation .........................79
applying geostatistics.........................83 coordinate system...............................51
approximate .......................................11 correlation analysis ............................34
ArcInfo .............................................. 16 correlation coefficient ........................35
arcsine transform ................................ 40 correlogram ........................................63
ARCTIN ............................................ 16 covariable ...........................................40
arithmetic mean .................................31 covariance ..........................................63
assumptions of spatial prediction......... 9 CRES..................................................12
autocorrelation 9, 18, 27, 45, 51, 56, 58, cross validation ...................................38
60, 63, 65, 66, 68, 69 Cross-K ..............................................63
average...............................................31 cross-validation ..................................73
cubic .............................................20, 21
B
cumulative distribution.................27, 28
back transformation ................ 40, 41, 42 cumulative frequency .........................77
back-transform ................................... 43 cumulative frequency plot ..................32
bad data........................................ 39, 74 cumulative probability density...........81
barriers............................................... 11
D
Best Linear Unbiased Estimation ...... 10
bimodal .............................................. 31 data analysis .......................................27
binning ............................................... 46 data collection ....................................83
bivariate normal distribution ............... 74 data transformation .......................40, 75
bivariate standard normal distribution. 74 data trends...........................................43
block averages kriging........................72 data visualization................................84
block kriging ..........................70, 71, 72 Daymet...............................................14
BLUE................................................. 10 declustering ........................................37
BLUEPACK......................................24 definitions
Box-Cox....................................... 40, 42 EDA ...............................................27
breeding bird survey .................... 55, 59 exploratory data analysis ................27
bull’s eye...................................... 18, 19 geostatistics ......................................7
lag distance .....................................46
C
nugget .............................................54
CAI ....................................................13 regionalized variable........................18
calculating the variogram ..................46 spatial autocorration .........................9
Cartesian space .................................. 51 spatial independence ........................9
stope. ..............................................23

Geostatistics Without Tears 111

 surface interpolation ........................9 GEO-EAS ..........................................24
Delaunay............................................ 22 geometric anisotropy ..........................61
DEM ..................................................16 geostatistical conditional simulation.....79
dependency rules ................................ 18 geostatistical process ...........................18
dependent variable.............................45 geostatistical simulation .....................79
depiction versus prediction................39 geostatistics .........................................18
design storm depth.............................12 geostatistics quality ............................18
detection limit ....................................37 GIDS ..................................................14
deterministic .......................... 10, 18, 71 global..................................................10
digital elevation model ......................16 global trend.........................................70
digital terrain models (DTM)............. 17 Gradient plus inverse distance square 14
directional variograms .......................60 graphing the semivariogram...............50
Dirichlet............................................. 22 grid based DEM .................................16
discretization......................................77 GSLIB................................................24
disjunctive kriging .............................. 74 gstat ....................................................25
distance tolerance ........................52, 53
H
DL...................................................... 37
drift .............................................. 64, 70 h 46, 50, 53, 57, 63, 64
DTM ..................................................17 Hermite polynomials ..........................76
duplicate samples................................ 37 hermitian polynomials.........................74
heteroscedasticity................................40
E
histogram................................27, 28, 40
effective range ...................................58 histogram visualization ......................84
error statistics .....................................38 history of kriging................................23
estimated qualifier .............................37 hole effect...........................................59
estimated values.................................37 holesin ................................................59
evaluation of model ........................... 84 hyperbolic model ................................58
exact................................................... 11
I
exact interpolator ..........................55, 71
exceedence...................................74, 81 IDF .....................................................12
experimental variogram...................... 57 IDW....................................................19
exploratory data analysis .............27, 37 inclusion.............................................39
exponential...................................57, 58 independent variable ..........................45
extreme values..............................38, 39 indicator geostatistics ..........................38
indicator kriging ...........................74, 77
F
indicator transform..............................75
first degree polynomial ......................70 indicator values ..................................77
first law of geography 39, 45, 56, 60, 63, interpolation .........................................9
66 interquartile range ........................32, 33
first quartile.................................. 32, 33 InvCov................................................65
Fourier series ......................................18 inverse distance to a power..................19
fractal dimension ...............................64 inverse distance weighted....................19
fractals .........................................64, 78 inverse distance weighted least squares
fractogram.......................................... 64 ........................................................19
frequency distribution........................27 inverted covariance ............................65
full indicator cokriging .......................74 irregular triangular mesh ....................21
G ISATIS ...............................................24
isofactorial kriging .............................76
Gandin, L.S........................................23 isotropic..............................................60
Gaussian ...................................... 57, 58 isotropic model...................................61
Gaussian disjunctive kriging .............76
K
Gaussian simulation.....................80, 82
Geary’s C.....................................65, 66 KED ...................................................71
GEMS ................................................ 25 Krige, Danie........................................23
general relative variograms................. 67 kriging ................................................69

Geostatistics Without Tears 112

 block .................................. 70, 71, 72 lognormal probability plot...................38
block averages................................ 72 lognormal transform............................41
cokriging................ 24, 40, 74, 76, 77 lognormal transformation....................39
disjunctive......................................74 long tails.............................................30
external drift ..................................71 longitude/latitude ...............................51
full indicator cokriging ................... 74
M
Gaussian disjunctive ......................76
history ............................................ 23 madogram...........................................65
indicator ................. 38, 39, 43, 74, 77 Mandelbrot.........................................64
isofactorial .....................................76 mass points.........................................21
kriging with external drift ...............70 Matern, B. ..........................................24
lognormal....................................... 75 mathematical morphology ..................24
median indicator ............................ 76 Matheron ................7, 23, 24, 37, 55, 58
multi-Gaussian................................ 41 Matheron model .................................58
multiple indicator...........................76 mean ...................................................31
ordinary.................. 24, 70, 71, 74, 76 mean compared to the standard
parameters ......................................51 deviation.........................................33
point............................................... 71 measure of similarity..........................46
probability................................ 77, 97 measure of spread...............................32
pronunciation of............................. 23 measures of correlation........................27
punctual.................................... 70, 71 measures of location.....................27, 31
quality ............................................ 18 measures of shape ........................27, 28
regression .......................................77 measures of spread .............................27
residual indicator ...........................77 median................................................31
search neighborhood......................72 median indicator kriging ....................76
simple.................................24, 70, 71 mesokurtic..........................................29
transGaussian ................................. 40 missing values ....................................39
uniform conditioning ..................... 76 mixed model .......................................38

universal...................................43, 70
weights...........................................72 mode...................................................31
model evaluation ................................84
kriging algorithm................................69 Monte Carlo simulation .......................79
kriging and remotely sensed images.. 77 Monte-Carlo.......................................63
kriging quality....................................18 Moran’s I......................................65, 66
kriging with external drift.................. 71 multi-Gaussian kriging........................41
kurtosis ........................................ 27, 29 multi-indicator kriging .......................76
multimodal .........................................32
L
multiple samples .................................37
lag ......................................................51
N
lag class..............................................52
lag class variable................................52 ND......................................................37
lag distance ............................ 45, 50, 51 nearest neighbor .................................20
lag distance tolerance...................52, 53 negative skew.....................................28
lag increment................................ 46, 52 normal.................................................41
lag spacing..........................................52 normal distribution.............................28
lag tolerance.......................................52 normal probability plot.................27, 29
latitude/longitude............................... 51 normal QQ plots .................................40
left skewed...................................28, 31 normally distributed ............................40
leptokurtic.......................................... 29 normally distributed data....................30
less than detection.............................. 37 not detected ........................................37
linear ............................................ 21, 57 nugget effect.......................................54
linear/sill ............................................ 57 nugget effect in different directions.....55
local.................................................... 10 nugget variance ..................................60
local relative variograms.....................67
O
local trend........................................... 70
lognormal kriging .............................. 75 objective analysis .........................15, 24

Geostatistics Without Tears 113

offset transform .................................. 42 replicates.............................................37
OI.......................................................15 residual indicator kriging ...................77
omindriectional..................................60 right skewed .................................28, 31
optimal interpolation....................15, 24 Ripley’s K ....................................63, 68
optimal prediction..............................10 RK ......................................................77
ordinary kriging...........................24, 70 rodogram ............................................68
outliers .... 19, 20, 21, 27, 30, 31, 38, 39, root mean squared deviation................32
56, 63, 67, 68, 73, 74, 75, 84 rubber sheet ........................................20
P S
pairwise relative variograms ...............67 sample size..........................................52
Parameter-evaluation Regressions on scale transform....................................42
Independent Slopes Model ............15 scatter plot ..........................................56
parameters for kriging.........................51 scattergram.........................................56
Pearson’s coefficient of correlation ... 34 SE.......................................................35
p-field simulation............................... 80 search neighborhood ..........................72
platykurtic.......................................... 29 second quartile ...................................32
Plurigaussian simulation.................... 80 secondary populations.........................43
point kriging ......................................71 semivariance.................................45, 50
polynomial trend surface ................... 21 definition ........................................45
polynomials ........................................20 semivariogram..............................45, 50
population variance ............................53 models ............................................57
positive skew .....................................28 nugget effect...................................54
power transformation.......................... 40 range...............................................53
PRISM .........................................14, 15 shape ..............................................57
Probability Field simulation............... 80 sill...................................................53
probability intervals...........................32 smoothing.......................................53
probability kriging ............................. 77 semivariogram models

probability of exceedence..................74
process................................................18 aniistrophic.....................................60
correlogram ....................................63
projection........................................... 51 covariance ......................................63
pronunciation of kriging ..................... 23 cross-K ...........................................63
punctual kriging............................ 70, 71 directional.......................................60
drift.................................................64
Q
exponential...................57, 58, 66, 82
QQ plot .............................................. 31 fractal .............................................64
quadratic ............................................ 21 Gaussian .............................57, 58, 66
quality of geostatistics/kriging...........18 Geary’s C .................................65, 66
quantile .............................................. 32 general relative variograms .............67
quantile-quantile plot ............. 27, 29, 31 hole effect.......................................59
holesin ............................................59
R
hyperbolic.......................................58
r 34, 35 InvCov............................................65
r2 34, 35 inverted covariance ........................65
Radarsat..............................................17 istrophic....................................60, 61
range ......................................32, 53, 57 linear ........................................57, 65
rank order transform ..........................77 linear/sill.........................................57
regionalized variable...........................18 local relative variograms .................67
regression kriging ............................... 77 madogram.......................................65
regression line....................................35 Matheron ........................................58
regularized spline................................ 20 Moran’s I ..................................65, 66
rejection ............................................. 39 omnidirectional ..............................60
relative variograms ............................. 67 pairwise relative variograms............67
remotely sensed images .....................77 relative variograms..........................67
replacement........................................39 Ripley’s K ................................63, 68

Geostatistics Without Tears 114

 rodogram........................................68 sum of square deviations of prediction
spherical....................... 57, 58, 66, 82 ........................................................35
standardized variograms ................. 68 sum of squares of error.......................35
trend ............................................... 64 summary analysis................................37
semivariogram visualization .............. 84 summary data analysis .......................27
sequential conditional simulation ...... 79 summary statistics ........................27, 37
sequential gaussian simulation........... 81 summary statistics visualization.........84
sequential indicator simulation ..........81 support..........................................69, 77
Shepard’s Method..............................19 surface interpolation.............................9
short tails............................................30 symmetric distributions ......................31
Sichel, Herbert.................................... 23 symmetrical distribtution....................28
sill .............................. 45, 52, 53, 57, 62 T
simple kriging.............................. 24, 70
simularity measure.............................. 46 temporal dependency .........................45
simulation tension spline.......................................20
conditional......................................79 tessellation..........................................22
conditional cosimulation................80 The Centro Internacional de Agricultura
Gaussian...................................80, 81 ........................................................13
p-field ............................................ 80 Theissen .............................................22
Plurigaussian..................................80 thin-plate spline..................................16
probability field ............................. 80 third quartile .................................32, 33
sequential indicator........................81 three parameter lognormal...................41
truncated gaussian..........................82 time.....................................................40
truncated plurigaussian .................. 82 TIN...............................................16, 21
turning bands ................................. 82 Tobler's law of geography...................45
skewed ....28, 31, 38, 41, 62, 63, 67, 74, TOPOGRID .......................................16
75, 84 TPS.....................................................16
skewed data with zeroes .....................38 transformation...................38, 40, 62, 75

skewed distributions
skewed left ..........................31
.........................................31 arcsine.............................................40
back ................................................43
skewed right.......................................31 Box-Cox ...................................40, 42
skewness ............................................ 28 lognormal........................................41
social phenomena ................................9 normal score ...................................41
spatial autocorrelation. 9, 18, 45, 56, 58, offset...............................................42
60, 65, 66, 68, 69 power..............................................40
spatial dependency.............................45 scale................................................42
spatial independence............................9 square root ......................................42
spatial outliers..............................39, 74 standard back ..................................43
spatial pattern .....................................39 weighted back.................................43
sphere of influence.............................58 transGaussian......................................40
spherical....................................... 57, 58 trend............................43, 62, 64, 70, 74
splines ................................................ 20 global..............................................70
square root transform..........................42 local................................................70
standard back-transform ..................... 43 trend removal......................................40
standard deviation........................ 27, 32 trend surface .................................21, 43
standard error of estimate ........ 27, 34, 35 cubic...............................................21
standarized variograms.......................68 linear ..............................................21
stationarity .................................... 27, 40 quadratic.........................................21
stochastic ........................................... 10 trend surface analysis .........................70
stochastic simulation ...........................79 trend versus anomalies .......................39
storm depth ........................................12 triangular irregular network .........16, 21
strata...................................................39 triangular mesh...................................21
stratification.................................38, 39 trimodal ..............................................31
subset ................................................. 39 truncated gaussian simulation ............82
truncated plurigaussian simulation.....82

Geostatistics Without Tears 115

turning bands simulation..............79, 82 omnidirectional ..............................53
parameters.......................................51
U
VARIOWIN.......................................25
uniform conditioning kriging............. 76 Voronoi ..............................................22
universal kriging...........................43, 70
W
V
weighted back-transform.....................43
variable lag class................................ 52 weights ...............................................72
variance........................................27, 32 Witwatersrand goldfields ....................23
variance cloud ....................................56
Z
variogram...........................................45
calculation......................................46 zonal anisotropy .................................62
cloud .............................................. 56
directional ......................................60
experimental................................... 57 γ(0) .........................................37,
53, 55
map ................................................ 60 γ(h) .........................................46,
50, 53
model ............................................. 57 γ(h) formula........................................46
modeling ........................................50

Geostatistics Without Tears 116