This article was downloaded by: [Otterbein University]

On: 10 April 2013, At: 22:56

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Cartography and Geographic Information Science

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcag20

Assessment of regression kriging for spatial

interpolation – comparisons of seven GIS interpolation
methods
a c b c
Qingmin Meng , Zhijun Liu & Bruce E. Borders
a
Department of Geosciences, Mississippi State University, MS, 39762, USA
b
Department of Geography, the University of North Carolina at Greensboro, 1009 Spring
Garden Street, Greensboro, NC, 27412, USA
c
School of Forestry and Natural Resource, the University of Georgia, Athens, GA, 30602,
USA
Version of record first published: 29 Jan 2013.

To cite this article: Qingmin Meng , Zhijun Liu & Bruce E. Borders (2013): Assessment of regression kriging for spatial
interpolation – comparisons of seven GIS interpolation methods, Cartography and Geographic Information Science, 40:1,
28-39

To link to this article: http://dx.doi.org/10.1080/15230406.2013.762138

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
 Cartography and Geographic Information Science, 2013
Vol. 40, No. 1, 28–39, http://dx.doi.org/10.1080/15230406.2013.762138

Assessment of regression kriging for spatial interpolation – comparisons of seven GIS

interpolation methods
Qingmin Menga,c*†, Zhijun Liub and Bruce E. Bordersc
a
Department of Geosciences, Mississippi State University, MS 39762; bDepartment of Geography, the University of North Carolina at
Greensboro, 1009 Spring Garden Street; Greensboro, NC 27412; cSchool of Forestry and Natural Resource, the University of Georgia,
Athens, GA 30602
(Received 14 November 2008; accepted 30 September 2010)

As an important GIS function, spatial interpolation is one of the most often used geographic techniques for spatial query,
spatial data visualization, and spatial decision-making processes in GIS and environmental science. However, less attention
has been paid on the comparisons of available spatial interpolation methods, although a number of GIS models including
inverse distance weighting, spline, radial basis functions, and the typical geostatistical models (i.e. ordinary kriging, universal
Downloaded by [Otterbein University] at 22:56 10 April 2013

kriging, and cokriging) are already incorporated in GIS software packages. In this research, the conceptual and methodological
aspects of regression kriging and GIS built-in interpolation models and their interpolation performance are compared and
evaluated. Regression kriging is the combination of multivariate regression and kriging. It takes into consideration the spatial
autocorrelation of the variable of interest, the correlation between the variable of interest and auxiliary variables (e.g., remotely
sensed images are often relatively easy to obtain as auxiliary variables), and the unbiased spatial estimation with minimized
variance. To assess the efficiency of regression kriging and the difference between stochastic and deterministic interpolation
methods, three case studies with strong, medium, and weak correlation between the response and auxiliary variables are
compared to assess interpolation performances. Results indicate that regression kriging has the potential to significantly
improve spatial prediction accuracy even when using a weakly correlated auxiliary variable.
Keywords: Regression kriging; spatial interpolators; Ikonos; Landsat; climate data interpolation

Introduction been widely applied to the fields of geology, environmen-

Spatial interpolation is often used to predict a value of a tal science, geography, and ecology.
variable of interest at unmeasured locations with the avail- Over the past years, a number of methods for spatial
able measurements at sampled sites. Spatial interpolation interpolation have been developed and applied. Univariate
is one of the most often used geographic techniques for kriging methods including ordinary kriging and universal
spatial data visualization, spatial query of properties, and kriging have been widely applied to environmental science,
spatial decision-making processes in geography, earth natural resources, remote sensing, and GIS. Recently,
sciences, and environmental science. regression kriging as a spatial interpolation method has
Spatial interpolation theoretically depends on the first been applied to environmental problems, but it has not
law of geography, i.e. “Everything is related to everything obtained enough attention in the field of geographic infor-
else, but near things are more related than distant things” mation science (GIScience) and even less attention has been
(Tobler 1970). Geostatistics is a group of statistical tech- paid on comparing it to other available spatial interpolation
niques focusing on spatial correlation, spatial interpola- methods. Conceptual and methodological confusions about
tion, simulation, and visualization of values of random regression kriging exist. The theoretical and methodological
variables. Geostatistics is based on the theory of regiona- differences of available spatial interpolation methods have
lized variables (Journel and Huijbregts 1978; Goovaerts not received enough exploration in the field of GIScience,
1997, 1999a, 1999b) and is extensively applied because it although some primary comparisons of cokring, kriging,
allows GIS analysts to consider the spatial dependence inverse distance weighting (IDW), spline, or radial basis
between observations to predict values at unsampled loca- function (RBF) have already been conducted (Collins Jr.
tions. Based on the Master's thesis of Daniel Gerhardus 1996; Anderson 2001).
Krige, the theory using kriging for interpolation and extra- Regression kriging and universal kriging are not very
polation was developed by the French mathematician well understood in the field of GIScience. For example,
Georges Matheron (1963). Subsequently, kriging has some may think that regression kriging is nothing else but a

*Corresponding author. Email: [email protected]

†Current address: Department of Geosciences, Mississippi State University, MS 39762.

© 2013 Cartography and Geographic Information Society

 Cartography and Geographic Information Science 29

re-expression of universal kriging (UK) with auxiliary vari- basis functions. Third, we then apply these stochastic and
ables. Thus, it is believed that UK and regression kriging deterministic interpolation methods to three case studies
are the same methods. Such thinking is misleading in with the correlation strength between response and auxili-
GIScience theory and methodology, particularly when we ary variables varying between strong, medium, and weak,
try to couple powerful spatial and geostatistical models with respectively, in order to address the above objectives.
GISystems. It can make unsophisticated users confuse spa- Fourth, the advantages and disadvantages of these inter-
tial heterogeneity, spatial dependence, and non-spatial asso- polation methods including cautionary notes on applying
ciation in spatial interpolation and overlook the conceptual them to GIScience are discussed and summarized.
and methodological development of the combination of
spatial association and non-spatial association into spatial
modeling. As Fotheringham and Wegener (1999) stated, the Conceptual understanding of regression kriging
danger of using GIS and spatial analysis models is that Odeh et al. (1994, 1995) first devised the so-called regression
unsophisticated users might be likely to use spatial analysis kriging (RK), and used it to explore the spatial correlation of
models without understanding the advantages and pitfalls. the response variable and the nonspatial correlation between
In addition, there is limited and conflicted understand- the response variable and auxiliary variables. Using Landsat
ing between stochastic interpolation methods such as cok- ETM+ as auxiliary data, Meng (2006) developed a regres-
riging, universal kriging, and ordinary kriging and sion kriging model to predict large area forest parameters.
Downloaded by [Otterbein University] at 22:56 10 April 2013

deterministic interpolation methods including inverse dis- Regression kriging was also compared with several fusion
tance weighting, spline, and radial basis functions. With methods regarding the spatial prediction efficiency for raster
the exception of regression kriging, these spatial interpola- data fusion (Meng, et al. 2009). In the past, regression kri-
tion methods have already been built into GIS software ging has often been applied to soil science (e.g.,Hengl, et al.
packages. However, it is still somewhat unclear what the 2004, 2007). The advantages of RK are that it does not suffer
performance differences between these methods are and from instability. Simple kriging with varying local means,
which interpolation method should be used in practice. explained by auxiliary data, is a very similar kriging process
Some may think that stochastic methods are more power- as regression kriging (Goovaerts 1997). RK can be easily
ful than deterministic methods, while others prefer deter- integrated with statistical computations like general additive
ministic methods over stochastic methods. The addition of modeling or regression trees (McBratney et al. 2000). Odeh
auxiliary variables is often believed to increase the perfor- et al. (1994, 1995) defined three basic types of RK: (1)
mance of spatial prediction, but does multivariate kriging “kriging combined with regression”, where regression is
(e.g., regression kriging and cokriging) perform better performed and followed by kriging of regressed values
than univariate kriging (e.g., ordinary kriging and univer- (Knotters et al. 1995); (2) “kriging with a guess field” by
sal kriging) with the inclusion of auxiliary variables? Ahmed and De Marsily (1987) which first performs regres-
In order to understand regression kriging and to compare it sion and residuals calculation, followed by separate kriging
with the available stochastic and deterministic interpolation of the predicted values and the regression residuals for the
methods, this article aims to present the conceptual and meth- final prediction; (3) the prediction is obtained using the
odological aspects of regression kriging and to compare those predicted values from regression models and residuals from
aspects with the widely used UK, ordinary kriging (OK), and kriging (Odeh et al. 1995). This third approach has been
cokriging (CoK) methods. Comparisons between regression widely accepted as regression kriging and in this study we
kriging and the popular deterministic models of inverse dis- explore this very concept of regression kriging.
tance weighting (IDW), spline, and radial basis functions are In practice regression kriging is often confused with
also explored in this paper to enhance the understanding of universal kriging. Universal kriging is kriging with an
advantages and disadvantages of these extensively used inter- internal trend that is purely dependent on the locations.
polation functions. Whether there are significant differences in This concept has been widely applied to geography and
model performance between stochastic and deterministic GIScience and it has already been integrated into many
interpolation methods and whether interpolation with auxili- GIS software packages. Universal kriging typically uses a
ary variables is better than approaches without using auxiliary polynomial function of locations (Goovarerts 1997),
variables are also addressed in this paper. which is widely used in geography and many other fields.
This study is organized as follows. First, we make a However, UK is not regression kriging. Specifically, UK is
conceptual exploration of regression kriging to clear some applied when a typical geographic trend exists in the data
of its misunderstanding and compare regression kriging and when a polynomial function of locations is fit to catch
with the often confused concept of universal kriging. the geographic trend. Universal kriging is the sum of a
Second, we describe the methodology of regression kri- smooth deterministic function, called the drift of a given
ging and briefly compare it with universal kriging, ordin- geospatial phenomenon, and a zero-mean random func-
ary kriging, cokriging, and three deterministic tion, called residual that captures this phenomenon’s erra-
interpolation methods, including IDW, spline, and radial tic fluctuations.
 30 Q. Meng et al.

Regression kriging, as used in this study, integrates X

p X
n
^zðs0 Þ ¼ ^βk  qk ðs0 Þ þ wi ðs0 Þ  eðsi Þ (2)
regression and kriging with varying local means mainly
k¼0 i
explained by auxiliary data. Universal kriging and regres-
sion kriging are different in theory and methodology.
where ^βk are the kth estimated regression coefficient, the kth
Regression kriging takes into consideration the spatial
external auxiliary variable at location s0 is qk (while,
autocorrelation in the variable of interest, the correlation
q0 ðs0 Þ ¼ 1), the number of auxiliary variables is indicated
between the variable of interest and auxiliary variables (e.
as p, the weights are wi ðs0 Þ, and the regression residuals are
g., remotely sensed images), and the unbiased spatial
eðsi Þ. We can represent RK in matrix notation as follows:
estimation with minimized variance (Meng et al. 2009).
Conceptually, regression kriging is only equivalent to UK
when a global search window is used, since in this case z ¼ qT  β þ e (3)
the regression coefficients are constant across the study
area and thus a drift is modeled in the regression kriging ^zðs0 Þ ¼ qT0  ^β þ λT0  e (4)
process (Goovaerts 1997). However, the chance of this
special circumstance of a drift happening is rare, because where e is the regression residuals, q0 is the vector of p
spatial dependence and spatial heterogeneity result in a auxiliary variables at s0 , ^β is the vector of p +1 estimated
local variation across space. model coefficients, λ0 is the vector of n kriging weights,
and e is the vector of n residuals. Accounting for the
Downloaded by [Otterbein University] at 22:56 10 April 2013

spatial correlation of residuals, we can solve the regression

Methodology coefficients using the following generalized least squares
estimation (Cressie 1993):
We may define spatial interpolation as the process to
estimate a real valued function of z(s) given the set of
values zðzðs1 Þ; zðs2 Þ; :::; zðsn ÞÞ at different points {s1, s2, β^ ¼ ðqT  C 1  qÞ1  qT  C 1  z (5)
… sn}. Deterministic models such as the IDW, spline, and
the RBF are widely applied to spatial interpolation. where q is the matrix of auxiliary variables at all observed
However, a deterministic approach may not be appropri- locations, z is the vector of sampled response observa-
ate because in geography, environmental science, and tions, and C is the n  n covariance matrix of residuals.
ecology we typically lack sufficient information of the An example to illustrate the regression kriging process
variation of properties in space. We need to allow for is to conduct a simple or multiple-linear regression, select
uncertainty in our description of the interpolation results an effective semivariogram model, and then use it to
because of this insufficient knowledge. Fortunately, geos- model the residuals, compute the regression coefficients,
tatistical modeling of the spatial variation of a property process the weight matrix, and finally employ the regres-
using a deterministic component and a stochastic (i.e. sion kriging model, and obtain the predictions at
random) component provide a stochastic approach to unsampled points (Meng, et al. 2009). In the process of
spatial interpolation. A concise methodological compar- RK, regression residuals (i.e. the model uncertainty) are
ison between regression kriging, other kriging methods, introduced into the interpolation system, which is then
and deterministic interpolation approaches is now pre- applied directly to the spatial prediction of the primary
sented below. variable.

Ordinary kriging and universal kriging

Regression kriging
Ordinary kriging and universal kriging (equation 6) are
In a spatial interpolation process, let the sampled values
two common univariate kriging methods, which are typi-
for the given locations be denoted as z(s1), z(s2), z(s3), …,
cally general linear regression models. ^zðs0 Þ is the value
z(sn), for example, where si = (xlatitude, ylongitude) is a
at location s0 to be interpolated. zðsi Þ are the sampled
location with the coordinates of xlatitude and ylongitude, and
values, and λi , determined by the semivariogram model-
location i = 1, 2, 3,…, n. The value at a new and
ing, are the weights to be assigned to each unsampled
unsampled location (s0) can be predicted using RK by
location.
adding the spatial trend and random components (i.e.
residuals) (Odeh et al. 1994) using the following equation: X
n
^zðs0 Þ ¼ λi zðsi Þ (6)
^ 0 Þ þ ^eðs0 Þ
^zðs0 Þ ¼ mðs (1) i¼1

where the residuals ^e are interpolated using ordinary kri- Universal kriging typically assumes that spatial varia-
ging and the trend is fit using linear regression as follows: tion in z values has a drift or structural component that can
 Cartography and Geographic Information Science 31
0 1
be captured by the variation in locations with the loca- w01 0 : : 0
tional information being incorporated into the kriging B 0 w02 : : 0 C
B C
process. Unlike ordinary kriging where the weights are wðs0 Þ ¼ B
B : : : : : CC (11)
determined by a stationary random function model, the @ : : : : : A
weights in the universal kriging are derived from a non- 0 0 : : w0n
stationary random function model.
where the prediction at location s0 is a function of the n
neighboring z(si), i= 1, 2,..., n, w(s0) is the distance matrix,
Cokriging (CoK) w0i is the distance between location s0 and si, α is an expo-
Cokriging is a versatile statistical approach for spatial point nent determining the weight (a value of 2 is often used) of
estimation, especially, when both primary and auxiliary each observation. As the exponent increases, predictions
attributes are available. If each component of z(s0) satisfies become more similar to the closer observations.
the intrinsic hypothesis that assumes that stationarity of the Splines, a deterministic interpolation method, fit the
differences between pairs of data points exists in the first data points sufficiently close while being as smooth as
and second moments (Journel and Huijbregts 1978), then possible (Hutchinson and Gessler 1994). It is not a quali-
cokriging is unbiased and defined by equations 7, 8, and 9. fied interpolation method for highly irregular surfaces. The
When CoK is methodologically compared to univariate spline method predicts z(s) with a smooth function g by
Downloaded by [Otterbein University] at 22:56 10 April 2013

kriging, CoK is more complex in notations without a new minimizing:

concept being added (Goovaerts 1997).
X
n

X
n ðzðsi Þ  gðsi ÞÞ2 þ ηHm ðgÞ (12)
^zðs0 Þ ¼ zðsj ÞΛj (7) i¼1
j¼1
where Hm ðgÞ measures the smoothness of the spline func-
X
n tion g defined by mth degree derivatives of g and η is a
Λj ¼ I (8) smooth parameter when it’s value is larger than 0. When η
j¼1 is 0, the spline function passes through all the data
points and as η becomes close to infinity the function g
X
v
approaches a least squares polynomial (Hutchinson 2000).
Γ ðsi ; sj Þ þ Ψ ¼ Γðsi ; s0 Þ i ¼ 1; :::; n (9)
f¼1
The function g is defined using equation 13. The mth
derivative and the degree of smoothness can be deter-
where I is an identity matrix = [1,0,…,0]T, T indicates a mined using generalized cross-validation.
transpose, and Λj are the weights associated with the
prediction. z(sj) is the vector z1(sj)…zm(sj). Γðsi ; sj Þand X
K X
n
gðsÞ ¼ ak fk ðsÞ þ λi Rðs  si Þ (13)
Γðsi ; s0 Þ are the cross variograms and Ψ is the Lagrange k¼0 i¼1
Multiplier for i from 1 to n.
where fk ðsÞ is the asset of K order monomials and
Rðs  si Þ is a scalar function of the Euclidean distance
Deterministic interpolation methods between s and si.
Inverse distance weighting (IDW, equation 10) is a typical Three typical layers compose the radial basis func-
deterministic GIS function for spatial interpolation. The tions, which include the input layer, the only hidden
contribution of the neighboring observations to the inter- layer, and the output layer (Figure 1). The activation
polation is determined by the geographic weight matrix function in the hidden layer is a type of a radial basis
using equation 11. With IDW we can only obtain pre- function. The Gaussian function is a typical activation
dicted values within the range of observations, but cannot function expressed by the equation below:
predict extreme values if observations do not include peak
ðscÞT ðscÞ
values. When some areas have dense sample locations hðsÞ ¼ ,ð r2
Þ
(14)
while other areas have sparse sample locations, IDW
may be a problematic function for spatial interpolation. where the vector c represents the center and the scalar r is
the radius of the function. The center c is a one-dimen-
P
n
zðsi Þ  wα
0i
sional shift if the input vector x is one-dimensional. The
^zðs0 Þ ¼ i¼1
(10) RBF has two different types of output: without activation
P
n
function the output is the total of the input; with a sigmoid
wα
0i
i¼1 type activation function with bias (i.e., with an unknown
 32 Q. Meng et al.

differences at the per-pixel level. (3) The Pearson correla-

tion coefficient was applied to check the similarity of the
distributions between the original and interpolated images.
Finally, (4) the general Bayes factor was applied to compare
how well the interpolation models predict values at the
unsampled locations. The Bayes factor (BF) is a robust
statistical criterion for comparing model performances
given the same dataset without the prerequisite that models
include the same dependent variables and the same sam-
ples. When comparing models using the BF, typically the
following four scenarios can be distinguished: weak evi-
dence if the BF is less than 3.16, positive evidence if the BF
is between 3.16 and 10, strong evidence if the BF is
Figure 1. Typical radial basis functions network. between 10 and 100, and very strong evidence if the BF
is larger than 100 (Raftery 1996).
shift) b is added to the input of the output neuron. After
the training is performed successfully, the independent Case Studies
Downloaded by [Otterbein University] at 22:56 10 April 2013

variables of the unknown sites are input and the output First case study
of the corresponding interpolated function values are
obtained. In summary, we can define an RBF as a An Ikonos image is first used to evaluate the performance
weighted sum of translations of a radial symmetric func- of regression kriging and to compare it with other inter-
tion augmented by a polynomial term. A general RBF can polation methods. Using this image as an example, it is
be represented as follows. relatively easy to conduct sampling techniques and to
obtain the sampled values for spatial interpolation. We
sc 2 applied the spatial interpolation models discussed above
hðsÞ ¼ ,ð r Þ
(15) using an Ikonos image with a spatial resolution of four
meters. Band 2 is used as the variable of interest and band
X
n
^zðsÞ ¼ pðsÞ þ λi hðjs  si jÞ (16) 3 is used as an auxiliary variable if it is necessary for
i¼1 interpolation. All interpolated values are recorded so that
it is convenient to assess the performance of the different
where zðsÞ is an RBF function, pðsÞ is a polynomial interpolation models.
function, λi is a real-valued weight, | . | indicates a A coastal area located in Camp Lejeune (34.57° N
Euclidean distance, and h is an activation function. One latitude, 77.28° W longitude, Figure 2), in southeastern
obvious advantage of the RBF interpolation is that it is North Carolina, was selected as the study area. Camp
suited to scattered points that are not distributed in a Lejeune is part of a coastal plain that changes in elevation
regular spatial pattern. from sea level to 19.2 m (63 feet) (Rootsweb 2007).
Different landscapes including hardwoods (H), mixed
softwoods (M), vegetated wetlands (W), and roads (R)
Evaluation cover this region (Figure 2).
Both qualitative and quantitative methods were used to Ikonos band 2 was used as the response variable. In
compare the various interpolation methods discussed order to make a comprehensive assessment of the inter-
above. Visual analysis was conducted to compare the polation results using regression kriging, 264 points gen-
quality of interpolated maps with the original maps. erated with simple random sampling were selected. Ikonos
While the visual analysis was subjective, the following band 3 was used as an auxiliary variable for cokriging and
statistical methods were performed to objectively quan- regression kriging to spatially interpolate the pixel values
tify the differences between the original and the interpo- based on the sampled pixel values.
lated maps: (1) basic statistics including the mean (z), Based on the randomly sampled 264 pixels, the spatial
minimum, maximum, standard deviation (SD), and a interpolation of pixel values was conducted first using
histogram were used to describe the basic distribution IDW, spline, RBF, OK, UK, CoK, and RK. The interpo-
of the interpolated and the original images and to com- lated images were then compared to the original image
pare their differences. (2) The root mean square error visually and quantitatively to fully understand the perfor-
(RMSE) was used to compare the differences between mance of different interpolation models.
the original data and the interpolated data, while the Interpolated Ikonos band 2 images using the randomly
mean absolute error (MAE) was used to compare the sampled 264 pixels and the original image were shown in
 Cartography and Geographic Information Science 33

Figure 2. Study areas: (1) Ikonos band 2 image located in North Carolina with center coordinates 34.57º N/77.28º W and 264 randomly
sampled points; H denotes hardwoods, M denotes mixed softwoods, W denotes wetlands, and R denotes roads. (2) Central Big Sur
ecoregion, located in California with center coordinates 36.23º N/121.70º W and 50 randomly sampled points of annual maximum
temperature. (3) Landsat TM band 6 of the same study area as in (2) with 500 randomly sampled pixels.

Figure 3. It appears that the RK interpolated image is IDW, spline, RBF, OK, and UK. The CoK interpolation
superior to the other interpolated images for preserving results indicate significant spectral distortions (Figure 3)
Downloaded by [Otterbein University] at 22:56 10 April 2013

the spectral and spatial structure found in the original band along the road, along the left side of the road in the
2. The spectral and spatial characteristics in the original hardwoods, and in the upper boundary area of the mixed
band are highly distorted in the interpolated images using softwoods.

Figure 3. A visual comparison of the interpolation results of the response variable (i.e. Ikonos band 2) using 264 randomly sampled
points. (1) Inverse distance weighting, (2) spline, (3) radial basis function, (4) ordinary kriging, (5) universal kriging, (6) cokriging, (7)
regression kriging, (8) original Ikonos band 2 image, and (9) original Ikonos band 3 image.
 34 Q. Meng et al.

Table 1. Statistics of the interpolation using 264 randomly sampled points.

Mean Min Max SD MAE RMSE Corr BF

Band2 274 191 563 60
IDW 265 9 513 52 24.74 12.042 0.725 —
Spline 264 34 549 70 32.40 14.142 0.673 –85.3734
RBF 266 27 528 56 25.17 8.944 0.745 156.5285
OK 265 75 474 49 26.39 14.213 0.713 –88.0176
UK 265 75 470 49 26.00 14.213 0.720 –88.0175
CoK 264 93 477 50 23.49 14.142 0.782 –90.9494
RK 274 187 573 61 7.70 1.013 0.983 1300.982
Note: IDW, inverse distance weighting; RBF, radial basis functions; OK, ordinary kriging; UK, universal kriging; CoK, cokriging; and RK, regression
kriging. SD, standard deviation; MAE, mean absolute error; RMSE, root mean square error; Corr, Pearson correlation coefficient; and BF, Bayes factors
used for kriging model comparisons in which IDW was used as the base model.

Visual comparisons indicate that the RK interpolation those of the original band 2 compared to the interpolated
is the best spatial interpolation model. Quantitative com- images using CoK, UK, OK, IDW, spline, and RBF.
parisons were conducted next. Those comparisons show The Pearson correlation coefficient, MAE, and RMSE
Downloaded by [Otterbein University] at 22:56 10 April 2013

differences between the interpolated results using IDW, were used to further evaluate the performance of the RK
spline, RBF, OK, UK, and CoK and the original band 2 interpolation and the other interpolation models. The RK inter-
image (Table 1 and Figure 4). polation has the best performance with a much smaller MAE
Basic statistics, including the mean, maximum, mini- and a higher correlation coefficient than the CoK interpolation
mum, and the standard deviation (SD) show detailed dif- and the other interpolation models (Table 1). Compared with
ferences between the interpolated pixel values and the IDW, spline, RBF, OK, and UK, the CoK interpolation does
original pixel values (Table 1). The closer these basic not show an obvious better performance based on the MAE,
statistics between the interpolated and the original image RMSE and correlation coefficients (Table 1).
are, the closer the distributions between the original and Finally, the general Bayes factor (BF) (Raftery 1996)
interpolated images are. Results show that the RK inter- was applied to determine whether significant prediction
polation performed much better, showing much closer differences existed among the interpolation methods. In
values of the mean, minimum, maximum, and the SD to general, the BF supports the above results from the basic

Figure 4. Histograms of the interpolated Ikonos band 2 image using 264 randomly sampled points and the original Ikonos band 2
image. (1) Inverse distance weighting, (2) spline, (3) radial basis function, (4) ordinary kriging, (5) universal kriging, (6) cokriging, (7)
regression kriging, (8) original Ikonos band 2 image, and (9) original Ikonos band 3 image.
 Cartography and Geographic Information Science 35
Downloaded by [Otterbein University] at 22:56 10 April 2013

Figure 5. A visual comparison of the interpolation results of temperature with 50 randomly sampled points. (1) inverse distance
weighting; (2) spline, (3) radial basis functions, (4) ordinary kriging, (5) universal kriging, (6) cokriging, (7) regression kriging, and (8)
actual temperature values.

statistics summarized in Table 1. There is very strong similar to the true temperature values. For the other inter-
evidence that RK is the best interpolation method, fol- polation methods this does not seem to be the case.
lowed by RBF. The OK, UK, and CoK did not show any The histograms show that RK is probably the best
improvement for spatial prediction compared to the deter- interpolation method, since it resulted in a very similar
ministic interpolation models. distribution to the true values. However, IDW, spline,
RBF have a very similar data range but very different
frequencies (Figure 6). Ordinary kriging, universal kri-
Second case study ging, and cokriging resulted in rather different range and
The second case study involved the spatial interpolation of frequency patterns. These results coincide with the
annual maximum temperature at the central Big Sur ecor- results from the basic statistics, where regression kriging
egion, located in California (Figure 2). The study area has very similar mean and standard deviation values as
extends about 33km from north to south and 27km from the true observations. Regression kriging is the best
east to west. The temperature data were collected for cell interpolation approach, since it possesses the smallest
sizes with a resolution of 800m. Elevation was used as the mean absolute error, smallest root mean square error,
auxiliary variable for the spatial interpolation of tempera- and the highest Pearson correlation coefficient of all
ture. Elevation is weakly correlated with maximum tem- interpolation methods (Table 2). Cokriging results in
perature, based on a Pearson correlation coefficient of the worst interpolation method with the largest error
–0.24 and an R square of 0.06, when temperature is and the lowest correlation.
regressed on elevation (p < 0.0001). The goal of the Similar to case study one, the Bayes factor is again
second case study was to evaluate whether regression applied to evaluate differences in the prediction performance
kriging can significantly improve spatial predictions com- between all interpolation methods (Table 2). Compared to
pared with other interpolation methods even if only a the IDW base model, there is not any strong evidence that
weak correlated predictor variable was used. OK, UK, and CoK are superior. However, the large positive
Following a procedure similar to the first case study, BF value indicates that the performance of RK is signifi-
IDW, spline, radial basis function, OK, UK, CoK, and RK cantly better compared to all other interpolation models.
were applied to spatially predict the maximum tempera-
ture across the central Big Sur ecoregion. The results are
displayed in Figure 5, in which different interpolation Third case study
methods result in somewhat different temperature patterns. The correlation between the response variable and the
It seems that RK, IDW, and RBF have spatial patterns auxiliary variable was very strong in the first and was
 36 Q. Meng et al.
Downloaded by [Otterbein University] at 22:56 10 April 2013

Figure 6. Histograms of the interpolated maximum temperature and the actual temperature values. (1) Inverse distance weighting, (2)
spline, (3) radial basis functions, (4) ordinary kriging, (5) universal kriging, (6) cokriging, (7) regression kriging, and (8) actual
temperature values.

Table 2. Statistics of the interpolation using 50 random samples at Big Sur, California.

Mean Min Max SD MAE RMSE Corr BF

True values 19.32 17.22 21.33 0.83
IDW 19.43 17.64 20.7 0.51 0.52 0.671 0.61 —
Spline 19.41 17.65 21.04 0.57 0.49 0.635 0.66 5.5155
RBF 19.41 17.64 21.04 0.56 0.49 0.636 0.66 5.3916
OK 19.39 18.47 20.36 0.44 0.53 0.663 0.63 1.1236
UK 19.41 18.5 20.31 0.44 0.54 0.671 0.61 1.01
CoK 19.39 18.53 20.02 0.37 0.58 0.735 0.49 –13.0281
RK 19.39 17.8 21.4 0.65 0.47 0.583 0.72 10.1031
Note: IDW, inverse distance weighting; RBF, radial basis functions; OK, ordinary kriging; UK, universal kriging; CoK, cokriging; and RK, regression
kriging. SD, standard deviation; MAE, mean absolute error; RMSE, root mean square error; Corr, Pearson correlation coefficient; and BF, Bayes factors
used for kriging model comparisons in which IDW was used as the base model.

very weak in the second case study. In the third case study, The interpolated image resulting from RK is closest to
Landsat TM band 1 is applied as the auxiliary variable for the original image, while IDW, spline, RBF, OK, UK, and
the interpolation of the Landsat TM band 6 (Figure 2). CoK show apparent color distortions in different locations
The correlation between the two TM bands is moderate at of their images compared to the original image (Figure 7).
0.52 (p < 0.0001). When histograms are compared, RK is the best interpola-
IDW, RBF, spline, OK, UK, CoK, and regression tion method, resulting in a very similar distribution of
kriging are used to interpolate the Landsat TM band 6. pixel values compared to the original image (Figure 8).
Results for the basic statistics are summarized in Table 3.
Much smaller MAE and RMSE errors and a higher corre-
lation coefficient indicate that RK performs much better Conclusion and Discussion
than OK, UK, CoK, IDW, spline, and RBF. The correla- Regression kriging is a powerful GIS function for spatial
tion coefficients further indicate that CoK, UK, and OK interpolation. It needs auxiliary variables that can be rela-
perform somewhat worse than IDW, spline, and RBF. The tively easily obtained in practice. For example, remote
Bayes factor shows strong evidence that RK performs sensing images are typically treated as auxiliary variables
much better than OK, UK, CoK, and the other three for spatial prediction and data fusion (Meng 2006; Meng,
deterministic interpolation methods. et al. 2009). In this research, a concise methodological
 Cartography and Geographic Information Science 37

Table 3. Case study 3: Statistics of the interpolation.

Mean Min Max SD MAE RMSE Corr BF

Band6 148 119 190 7.83
IDW 148 127 183 5.45 4.9314 11.3743 0.5156 —
Spline 148 106 192 8.48 4.9057 11.3456 0.5136 2.5263
RBF 148 127 182 5.27 5.2277 9.6809 0.5128 161.2015
OK 148 134 159 3.99 5.613 11.7753 0.4461 –34.6478
UK 148 134 158 3.89 5.7237 11.8571 0.4359 –41.5705
CoK 148 132 164 4.83 4.8373 8.9089 0.5039 238.091
RK 148 81 195 7.08 3.3991 6.2555 0.8014 591.6807
Note: IDW, inverse distance weighting; RBF, radial basis functions; OK, ordinary kriging; UK, universal kriging; CoK, cokriging; and RK, regression
kriging. SD, standard deviation; MAE, mean absolute error; RMSE, root mean square error; Corr, Pearson correlation coefficient; and BF, Bayes factors
used for kriging model comparisons in which IDW was used as the base model.
Downloaded by [Otterbein University] at 22:56 10 April 2013

Figure 7. Case study 3: A visual comparison of the interpolation results. (1) inverse distance weighting, (2) spline, (3) radial basis
functions, (4) ordinary kriging, (5) universal kriging, (6) cokriging, (7) regression kriging, and (8) the original Landsat TM band 6.

Figure 8. Case study 3: Histograms of the interpolated pixel values and the actual values. (1) inverse distance weighting, (2) spline,
(3) radial basis functions, (4) ordinary kriging, (5) universal kriging, (6) cokriging, (7) regression kriging, and (8) the original Landsat
TM band 6.
 38 Q. Meng et al.

comparison was presented between regression kriging and since the BF does not have the prerequisites of nested
six other standard interpolation approaches (i.e., cokriging, models, non-nested models, or does not require the same
universal kriging, ordinary kriging, and three deterministic auxiliary variables.
models, including IDW, spline, and RBF) to evaluate the This research does not include the so-called kriging
performance between stochastic and deterministic interpo- with an external drift method. Kriging with an external
lation methods and specifically between regression kriging drift (KED) can be treated as a special case of regression
and universal kriging. A basic framework of spatial inter- kriging, where the external drift is typically a linear trend
polation using regression kriging was presented. From a fitted between the response and auxiliary variables that
theoretical point of view, incorporating spatial autocorrela- smoothly varies across the whole study area. However,
tion of the variable of interest and the correlation between the assumption of an external drift between response and
this variable and an auxiliary variable into the interpolation auxiliary variables across the study area limits the effi-
process, regression kriging should produce much better ciency of kriging with an external drift for spatial inter-
spatial interpolation results than any other interpolation polation. The main reason for this is that spatial
method used in this research. One prerequisite for regres- dependence changes locally and spatial heterogeneity is
sion kriging to use an auxiliary variable is a significant often dominant, which can be modeled and captured by
correlation between the auxiliary variable and the variable regression kriging described in this study. In kriging with
of interest. Remote sensing supplies a relatively easy way to an external drift, “the trend is modeled as a linear func-
Downloaded by [Otterbein University] at 22:56 10 April 2013

collect auxiliary variables, which typically are significantly tion of smoothly varying secondary variables instead of
correlated to geographic phenomena. as a function of the spatial coordinates” (Goovaerts
The results from three case studies, in which correlations 1997). This is a limitation of KED’s instability of the
between response and auxiliary variables were strong, weak, extended matrix (Hengl, et al. 2007). If the trend is a
and medium, respectively, showed that regression kriging function of the spatial coordinates, kriging is similar to
performed much better than ordinary kriging, universal kri- universal kriging as discussed above. Kriging with an
ging, cokriging, and the three deterministic interpolation external trend requires that the secondary data vary
methods (inverse distance weighting, spline, and radial smoothly in space, but regression kriging does not
basis functions). The first two case studies showed that require this prerequisite.
there was little evidence that OK, UK, and even cokriging Regression kriging is a very powerful spatial interpola-
were superior to the simple deterministic methods spline, tion method, which is supported by similar conclusions from
IDW, and RBF; however, there was very strong evidence other related studies (e.g., Minasny and McBratney 2007).
that RK was superior to all other interpolation methods. Results are inconclusive, whether stochastic cokriging, uni-
Incorporating auxiliary variables into cokriging cannot guar- versal kriging, or ordinary kriging are superior in practice
antee improved interpolation results although, in the third compared to the relatively simple deterministic methods
case study the Bayes factor indicated a strong evidence that such as inverse distance weighting, radial basis functions,
CoK was better than OK, UK, and the other deterministic and spline. Cokriging could provide better results in some
interpolation methods. Cokriging has the following three circumstances (Goovaerts 1997). Our first two case studies
characteristics: (1) some weights of the auxiliary variables showed that simple deterministic interpolation methods pro-
are negative thereby increasing the risk of outliers in inter- vided better interpolation results compared to OK, UK, and
polations; (2) the interpolation contribution from auxiliary CoK, but CoK was superior for the data used in the third case
variables is reduced because most of the weights are small; study. An advantage of the three standard kriging interpola-
and (3) it may deliver superior results in some situations tion methods (e.g., ordinary kriging, universal kriging, and
(Goovaerts 1997). In other words, there are sometimes cri- cokriging) is that they provide the standard deviation of
tical uncertainties in cokriging interpolation. This also means errors, while deterministic interpolation methods cannot sup-
that it is important to compare CoK with ordinary kriging, ply this type of uncertainty analysis. Compared with kriging,
universal kriging, and other deterministic interpolation meth- the deterministic interpolation methods are simple in com-
ods in order to determine which of the methods should be putation and are easy to understand and to implement.
applied in practice.
Regression kriging must be compared with ordinary
kriging and cokriging (Hengl, et al. 2007). It is necessary References
to compare both deterministic and stochastic interpolation Ahmed, S., G. De Marsily. 1987. “Comparison of geostatistical
methods with each other in practice. The advantages of methods for estimating transmissivity using data on trans-
deterministic interpolation methods are that they are less missivity and specific capacity”. Water Resources Research
complicated in methodology, are relatively easy to imple- 23: 1717–1737.
Anderson, S. 2001. “An Evaluation of Spatial Interpolation
ment, and are widely included in or are coupled with many Methods on Air Temperature in Phoenix, AZ”. http://www.
GIS software packages. The Bayes factor is a robust statis- cobblestoneconcepts.com/ucgis2summer/anderson/anderson.
tical criterion for comparing different interpolation models, htm. Accessed on November 15 2012.
 Cartography and Geographic Information Science 39

Collins Jr., F.C. 1996. “A Comparison of Spatial Interpolation Matheron, G. 1963. “Principles of geostatistics”. Economic
Techniques in Temperature Estimation”. http://www. ncgia. Geology, 58: 1246–1266.
ucsb.edu/conf/SANTA_FE_CD-ROM/sf_papers/collins_ McBratney, A., L. Odeh, T. Bishop, M. Dunbar, and T. Shatar.
fred/collins.html. Accessed on November 15 2012. 2000. “An overview of pedometric techniques of use in soil
Cressie, N. 1993. Statistics for Spatial Data, Revised Ed. New survey”. Geoderma, 97 (3–4): 293–327.
York: Wiley. Meng, Q. 2006. “Geostatistical prediction and mapping for large
Fotheringham, A.S. and M. Wegener. 1999. Spatial Models and area forest inventory using remote sensing data”. 2006
GIS: New and Potential Models. CRC press. UCGIS Summer Symposium. www.ucgis.org/summer2006/
Goovaerts, P. 1997. Geostatistics for Natural Resources studentpapers/Mengqm_July03_2006.pdf. Accessed on
Evaluation. New York: Oxford University Press. January 8 2010.
Goovaerts, P. 1999a. “Geostatistics in soil science: state-of-the- Meng, Q., B. Borders, M. Madden. 2009. “High-resolution
art and perspectives”. Geoderma, 89: 1–46. image fusion using regression kriging”. International
Goovaerts, P. 1999b. “Using elevation to aid the geostatistical Journal of Remote Sensing (in press).
mapping of rainfall erosivity”. Catena, 34 (3–4): 227–242. Minasny, B. and A. McBratney. 2007. “Spatial prediction of soil
Hengl, T., G.B.M. Heuvelink, A. Stein. 2004. “A generic frame- properties using EBLUP with the Matérn covariance func-
work for spatial prediction of soil variables based on regres- tion”. Geoderma, 140(4): 324–336.
sion-kriging”. Geoderma. 120: 75–93. Odeh, I., A. McBratney, and D. Chittleborough. 1994. “Spatial
Hengl, T., G.B.M. Heuvelink, D.G. Rossiter. 2007. “About prediction of soil properties from landform attributes derived
Regression-kriging: from equations to case studies”. from a digital elevation model”. Geoderma, 63 (3–4): 197–214.
Computers & Geosciences. 33: 1301–1315. Odeh, I., A. McBratney, and D. Chittleborough. 1995. “Further
Hutchinson, M.F. and P.E. Gessler. 1994. “Splines-more than just results on prediction of soil properties from terrain attributes:
Downloaded by [Otterbein University] at 22:56 10 April 2013

a smooth interpolator”. Geoderma, 62: 45–67. heterotopic cokriging and regression-kriging”. Geoderma, 67
Hutchinson, M.F. 2000. ANUSPLIN Version 4.1 User Guide. (3–4): 215–226.
Canberra: The Australian National Uversity Center for Raftery, A.E. 1996. “Approximate Bayes factors and accounting
Resource and Environmental Studies. for modle uncertainty in generalized linear models”.
Journel, A.G. and C.J. Huijbregts. 1978. Mining Geostatistics. Biometrika, 83: 251–266.
New York: Academic Press. Rootsweb, http://rootsweb.com/~nconslow/onsloww.htm.
Knotters, M., D.J. Brus, J.H. Oude Voshaar. 1995. “A compar- Accessed: November 15 2012.
ison of kriging, co-kriging and kriging combined with Tobler, W.R. 1970. “A computer model simulation of urban
regression for spatial interpolation of horizon depth with growth in the Detroit region”. Economic Geography, 46
censored observations”. Geoderma, 67: 227–246. (2): 234–240.