See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/271880934

A review of spatial sampling

Article  in  Spatial Statistics · December 2012

DOI: 10.1016/j.spasta.2012.08.001

CITATIONS READS

114 1,278

4 authors:

Jinfeng Wang Alfred Stein

Chinese Academy of Sciences, LREIS University of Twente
263 PUBLICATIONS   3,538 CITATIONS    390 PUBLICATIONS   7,479 CITATIONS   

SEE PROFILE SEE PROFILE

Bingbo Gao Yong Ge

Chinese Academy of Sciences State Key Laboratory of Resources and Environmental Information System
48 PUBLICATIONS   302 CITATIONS    98 PUBLICATIONS   1,003 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Distributed GIS View project

cancer epidemiology View project

All content following this page was uploaded by Bingbo Gao on 16 May 2018.

The user has requested enhancement of the downloaded file.

Accepted Manuscript

A review of spatial sampling

Jin-Feng Wang, A. Stein, Bin-Bo Gao, Yong Ge

PII: S2211-6753(12)00025-5
DOI: 10.1016/j.spasta.2012.08.001
Reference: SPASTA 14

To appear in: Spatial Statistics

Received date: 5 August 2012

Accepted date: 6 August 2012

Please cite this article as: Wang, J.-F., Stein, A., Gao, B.-B., Ge, Y., A review of spatial
sampling. Spatial Statistics (2012), doi:10.1016/j.spasta.2012.08.001

This is a PDF file of an unedited manuscript that has been accepted for publication. As a
service to our customers we are providing this early version of the manuscript. The manuscript
will undergo copyediting, typesetting, and review of the resulting proof before it is published in
its final form. Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
 A Review of Spatial Sampling
Jin-Feng Wang1 , A Stein2 , Bin-Bo Gao1 , Yong Ge1

1
State Key Laboratory of Resources and Environmental Information Systems, Institute of Geographic
Sciences and Nature Resources Research, Chinese Academy of Sciences, A11 Datun Road, Beijing
100101, China
2
Twente University, Faculty of Geo-information Science and Earth Observation (ITC), Enschede, The
Netherlands

Abstract. Spatial sampling has as its major aim to collect samples in 1-, 2- or
3-dimensional space. It usually targets at the estimation of the total or mean of value
in an area, to optimize estimation of values at unsampled locations, or to predict the
location of a movable object. Some of the objectives are about populations,
representing the ―here and now‖, whereas other objectives concern superpopulations
that generate the populations. Data to be collected are usually spatially autocorrelated
and heterogeneous, whereas sampling is usually not repeatable. In various senses it is
distinct from the assumption of iid data from a population in conventional sampling.
The uncertainty of spatial sample estimation propagates along a chain from spatial
variation at the stochastic field, to distribution of a sample and a statistic to obtain an
estimate. This uncertainty is measured by either a design or a model based method.
Both methods can be used in population and superpopulation studies. The target to
obtain an unbiased estimate with the lowest variance is thus a common goal in spatial
sampling and inference. Reaching this objective can be addressed by sample
allocation in an area in order to reach a restricted objective function.

1 Introduction

Sampling concerns selection of a subset of individuals from within a population to

estimate characteristics of the whole population. The characteristics could be the total
or mean value of a random field (Haining, 2003; Christakos, 2005), or the values at
unsampled sites (Goovaerts, 1997), or locating target(s) (Rogerson et al., 2004).

1.1 A brief history of spatial sampling

Intuitive application of the principles of sampling in science has been taken place
for a long time from very early human history in Egypt, China and other places
throughout the world. The first known attempt to make statements about a population
using only information about part of it was made by the English merchant John
Graunt (1620 –1674). In his famous tract he describes a method to estimate the
population of London based on partial information. After that sampling theory has
1
developed separately from the mainstream of classical statistics (Neyman, 1934) and
has evolved into an extensive body of theory, method, and operation used daily all
over the world. Cochran (1946, 1977)‘s landmark book ―Sampling Techniques‖ is
well received and widely used in modern sampling practices. To deal with two
dimensional spatial sampling, 'Regionalized variable theory' referred to as
'geostatistics' was well built and is widely applied in geoscience (Matheron, 1963,
1971), in which spatial autocorrelation was introduced to improve sampling efficiency
such as estimator error variance in relation to sample design and sample size (Ripley,
1981; Haining, 1988; Christakos, 2005; Stein, 2003). Spatial stratified heterogeneity
was considered to get more efficiency of spatial sampling and inference (Goovaerts,
1997; Li and Wang, 2008; Wang et al., 2009a; Wang et al., 2010).

1.2 Contents of spatial sampling

Compared with complete enumeration, a statistically based sampling process has
large advantages. It leads to reduced cost, an increased speed, a greater scope and a
loss of accuracy that is negligible as compared to a complete enumeration (Cochran,
1977). Moreover, they result in designs that can be used repeatedly, hence leading to
small changes in estimates, that in turn provide measures of uncertainty. The aims of
spatial sampling methods are to get results of a higher quality at a lower cost.
Commonly applied constraints are a) a cost constraint that the costs do not exceed a
given budget and b) a quality constraint that the quality of the result meets a given
minimum requirement.
For a formal description we consider the triple ( , , ) consisting of a random
field , a design and an estimator . The triple is also called a ‗sampling strategy
trinity‘. The symbol refers to the characteristics of the random field such as the
random error, the trend, autocorrelation or stratification, being the determinants of the
efficiency of the sample estimation. The ‗Design‘ refers to the procedure
determining the sample selection of density and location, whereas the ‗estimator‘
defines the procedure to calculate an estimate of a population parameter from the
sample data. It is useless to discuss the merits of a sampling design without
considering the estimator to be used and the characteristics of the random field, or
vice versa. It is the combination of the three that determines the accuracy of the
estimates. (Wang et al., 2010).
Spatial sampling and inference is divided into the following six principle steps
(Cochran 1977, Wang et al., 2009b):
(1) Clarifying the objectives of sampling, to estimate the mean or total of population,
or to predict the values at un-sampled sites or mapping, or to identify the position
of a target, etc. It also should be clear that the objective is about the here and now
population (Haining, 2003) or about the superpopulation which generates the
cross section population.
(2) Defining the population to be sampled and sampling units. Population donates the
aggregate from which the sample is chosen, and the population is then divided
into non-overlapping parts called sampling units or elements. In spatial context,

2
 the population where sample units to be drawn and the population which to be
estimated may be spatially inconsistent, the former is called sampling population
and the latter is called estimating population. Notice that these two are identical
for non-spatial sampling.
(3) Choosing the sampling methods and computing the sample size under specific
budget and precision constraints. A correct decision can be made under the
context of the sampling design trinity ( , , ).
(4) Designing the sampling plan that describes where, when and how sampling units
are to be taken.
(5) Drawing the sample and taking the measurements on the obtained sampling units.
The collection of measurements is commonly called the sample.
(6) Analyzing the sample together with prior knowledge to obtain the features of the
stochastic field, such as randomness, trend, autocorrelation, stratification, etc.
These features and the objectives of sampling provide the basis for choosing the
best statistic to make inference on population parameters.

In existing sampling and inference two theories are commonly distinguished:

design-based and model-based sampling. They differ in the element of randomness
they utilize in order to give stochastic structure to the inference (Särndal, 1978, 1992;
Gruijter and Ter Braak, 1990, Haining, 2003). In a spatial context, the ―here and now‖
characteristics of a random field are obtained once the entire area is exhaustively
surveyed. The uncertainty of the sample estimation comes from the randomness of the
sampling sites, which is reduced with the increase of sample size until it is fully
covered. Such an approach is called design based sampling, which can be
implemented by traditional sampling (Cochran, 1977) applying to spatial random
fields. Alternatively, the value is not yet obtained even the entire area was exhaustive
surveyed, which was actually once realization of an underlining stochastic process, i.e.
superpopulation. The approach is called model based sampling (Särndal, 1978), the
methods are mainly related to spatial statistics (Cliff and Ord, 1981; Haining, 2003)
and geostatistics (Matheron, 1963), where the observed random field is regarded to be
a single realization of a spatial process or a superpopulation. Some assumptions have
to be made and the process model has to be well defined in order to estimate the
superpopulation because only a cross section observation is available, which is often
not sufficient to support inference to the superpopulation.
Clearly, the two major distinct objectives of sampling result in two major
categories of techniques of estimation.

2 Design-based sampling

In design-based sampling, the population of values in the interested region are

regarded as unknown but fixed. The primary source of randomness arises by drawing
a sample randomly. Repeated sampling according to a given scheme such as random
sampling will generate a distribution of sample values, and parameters of the

3
population and their uncertainty are evaluated based on the distribution. For example,
when calculating weighted averages, the data are being assigned weights that are
determined by the selection probabilities of the sampling locations, not by their
geographical coordinates. Design-based sampling and statistics can be used both for
populations (a single realization of a stochastic field) and superpopulation (a
stochastic field). As an example of a population, we may consider spatial disease
incidence values in one year, whereas the disease prevalence, an accumulation of
cases divided by the susceptible population in a range years reflecting the risk of the
disease occurrence, may serve as an example of a superpopulation. In the population
example, incidence can be mapped as the cases in that single year, whereas in the
superpopulation example the prevalence can be mapped by averaging the cases in
many years.

2.1 Sampling i.i.d. populations

Classical sampling theory is based on probability sampling. It includes methods of
sample selection and estimation that provide, at the lowest costs, estimates that are
precise enough for our purpose (Cochran, 1977). In this way it makes sampling
efficient. Samples can be selected with an equal probability or with an unequal
probability. Equal probability sampling is often apparently feasible for sampling in
practical surveys. Simple random sampling (SRS) and systematic sampling (SY)
serves as the starting point to understand sampling. In SRS or SY sample units are
drawn independently from each other with equal probability. It is theoretically a
relatively simple approach, but it is rarely used in practical sampling work due to its
low efficiency. More cost-effective sampling methods are stratified sampling,
cluster sampling, multistage sampling, two-phase sampling, and sequential sampling
(Neyman, 1934, 1938; Cochran 1946; Horvitz and Thompson, 1952; Royall, 1968).

(1) Simple random sampling

Although this category is rarely applied in practice due to its low efficiency, all
design based sampling techniques originate from it. Those can only be fully
understood when proper knowledge of SRS is present. In SRS, the population is
assumed to be independent and identically distributed (i.i.d.). The required sample
size to reach a pre-specified precision is based upon the dispersion variance of the
population and the required survey precision. Then the sample units are chosen from
the population independently with equal probability, and inferences are conducted
using the sample. The mean of population is estimated by the mean of sample ,
the variance of the estimate is proportional to the dispersion variance of the sample
and inversely proportional to number of sampling units. The confidence interval of
2
the mean equals if the population is not only i.i.d. but also N(0, )

distributed, where 2 is dispersion variance of population, v is variance of and is

the level of significance. SRS is not often applied in spatial studies, as it completely
neglects prior knowledge to the object, and leads to relatively large gaps in the

4
sampled area, and access to locations is often prohibitive or difficult to find, despite
recent developments in Global Positioning Systems.
Horvitz and Thompson (1952) developed a general theory for constructing
unbiased estimates. Whatever the selection probabilities are, as long as they are
known and positive, it is always possible to construct a useful estimate. Horvitz and
Thompson completed the classical theory, and random sampling was almost
unanimously accepted. Most of the classical books about sampling were also
published by then (Cochran, 1946; Hansen et al., 1953;).

(2) Systematic sampling

In systematic sampling, the population is again assumed to be independent and
identically distributed (i.i.d.). Once the first sample unit is drawn from the population,
the other sampling units are drawn in a given and preset order relative to this first
point. Systematic sampling is relatively easy to implement. It has the advantage over
SRS that the sampling units are spread more evenly over the sampled area and that it
may produce more precise estimates than simple random sampling. Systematic
sampling requires the use of systematic sampling zones. The mean is the same

as that simple random sampling. The variance is ],

where s2 is the dispersion variance of pooled data,

, N is the number of units of the population, n is the sample size, is the value

of the j-th sample unit of the i-th systematic zone and is the mean of the i-th

systematic zone. If the population is not only i.i.d. but also N(0, 2) distributed and 2
is dispersion variance of population. The confidence interval of mean of population is
，where t is the t-statistic.

(3) Stratified random sampling

In stratified random sampling we assume that the population is spatially stratified
and that the subpopulation within each stratum is i.i.d. In stratified sampling the
population is divided into subpopulations called ‗strata‘, which are non-overlapping
and together comprise the whole of the population (Cochran 1977). Simple random
sampling or systematic sampling is then applied in each stratum. Through dividing the
population into subpopulations the heterogeneity in each stratum is reduced and it is
easier to collect samples of that are highly representative (Wang et al 2010). The
mean of each stratum is computed by , where is the sample size

of stratum h, is the value of the i-th sample unit of stratum h. The mean of the
sample is computed by , where L is the number of strata, is

5
the weight for stratum k, which can be computed as the fraction of stratum k within
the population. The dispersion variance of each stratum can be computed by
. The variance of the full sample equals

, where is the number of sampling units in

stratum h, and the standard deviation equals . The confidence interval

of the sampling mean equals ，where t is the t-statistic.

(4) Cluster sampling

Again, the population is assumed to be i.i.d. In each cluster the first sample unit is
selected randomly and the rest are drawn in an order relative to it. The mean of the
population is estimated by , where n is the number of clusters, is
the mean of the i-th cluster computed using the same method as systematic sampling
method. The variance equals , the standard deviation

equals . Notice that systematic sampling is a special case of cluster

sampling in which several clusters are chosen.

(5) Two-step random sampling

Again, the population is assumed to be i.i.d., and it is now assumed that the
population is divided into subpopulations (called primary units). The first step is to
decide upon subpopulations to be sampled. This decision is usually done using a
random selection. In the next step samples are selected, for example using simple
random sampling within each selected subpopulation. Two-step random sampling is
different from stratified sampling in the sense that only some of the subpopulations
are sampled. The mean of the population is estimated by , where n is

the number of primary units, is the mean of the i-th primary units computed using

the same method as simple random sampling. The variance equals

, and the standard deviation equals . If the

subpopulations are unequal in size, means and variances are to be calculated by
embedding a weighing factor proportional to the size of the subpopulation.

2.2 Sampling considering spatial autocorrelation

Spatial autocorrelation is ubiquitous in spatial populations. It is one possibility to

6
take the well-known geographical principle into account that observations close to
each other are more likely to be similar than observations further apart. This principle
is also known as Tobler‘s law in geography. Spatial autocorrelation violates the first
assumption of i.i.d. of the population, namely that of independence. Recognizing the
autocorrelation has implications for the efficiency with which sampling is carried out.
The variance of the estimation error may for example be incorrectly estimated with
respect to choices of sampling design and sample size (Ripley, 1981; Haining, 1988;
Christakos, 2005). If the objective is to estimate population parameters, then we do
not need as many sample size as in the survey population where the sampling units

are independent from each other. In fact,

, where R equals , r equals , v is the allowed error of spatial mean

estimated and is defined by the variance E( )2, 2 is dispersion variance of
the population and is defined by E(y - Ey)2 , y or yi refers to a sample element, is
mean value estimated by a spatial sample, the estimation is the same as the simple
random statistic but their variances are different as the above description. is
the sample size computed using classic sampling theory (Ripley, 1981; Dunn et al.,
1993; Griffith et al., 1994; Wang et al., 2009b). If the objective is to estimate the
superpopulation, then the sample size should be larger than that of sampling an i.i.d.
population, because the variance of the sample estimation is inflated by the spatial
autocorrelation (Haining, 1988). On the one hand, the number of sampling units
reduces with an increasing spatial autocorrelation, if the aim is to estimate the
population mean. On the other hand, more sampling units are needed to estimate
superpopulation mean, if sampling is constrained by a required variance of estimation
error.
In the case of stratified random sampling the error variance of estimating

population mean is given by (Ripley, 1981, Dunn et al., 1993): ,

where the second term inside the square brackets is the average autocovariance
between all possible pairs (k, l) within a stratum. Also the sample size in each stratum
can be computed like spatial simple random sampling. In the case of systematic
sampling, the error variance of estimating the population mean is given by (Ripley,

1981; Dunn et al., 1993): where is the average

autocovariance between members of the systematic sample.
Early work evaluating spatial sampling strategies in two dimensions include
contributions from Quenouille (1949), Das (1950), Zubrzycki (1958) and Matern
(1960). These theoretical studies were based on random fields that were the outcomes
of spatially homogeneous processes. Ripley (1981) showed that systematic sampling
outperforms other sampling schemes when no prior knowledge is available about
random fields except if there is periodicity in the attribute. These findings have been
generally endorsed by empirical studies (Mátern, 1986; Milne, 1959; Payandeh,
1970).

7
 The work by Dunn and Harrison (1993) showed that whilst systematic sampling
was the most efficient of the three methods of sampling (and random sampling
consistently the least efficient), the gains in efficiency, relative to stratified random
sampling, were highly variable. They also compared two different methods of
estimating the error variance of the mean from a single systematic sample (Ripley,
1981). Their work was based on sampling real land use maps with complex and
varied spatial autocorrelation structures and their findings suggested that the presence
of non-stationarity and of anisotropy in spatial data could have a severe effect on the
efficiency of systematic sampling.
Map stratification, whether applied to stratified random or systematic sampling,
usually involves the definition of strata and the sampler needs to decide on strata size
as well as the orientation and starting point for the overlaying grid. Typically the
starting point is chosen at random and the effect of the starting position on the
calculated sample variance is assumed to be small. This is also supposed to be true for
the grid orientation, unless there are strong directionalities in the random field as in
the case for example of a repetitive ridge and valley topography. The stronger the
spatial autocorrelation on the map, with high levels over long distances, the more
redundancy will be present in a sample if the inter-point sampling distance is small
(Griffith, 2005). Ripley (1981) remarks that the gain from stratification will be high
when spatial autocorrelation is large for all distances up to the scale of the strata. The
gain becomes negligible beyond that scale. This suggests that for monotonically
decreasing correlation functions we should take small strata; hence the number of
sample points per strata will be small (Ripley, 1981). If spatial autocorrelation
structures vary in different sections of the map, then presumably there will be
efficiency gains to be achieved by adapting strata size to such heterogeneity (Berry et
al., 1968).

2.3 Sampling considering spatial heterogeneity

Spatial heterogeneity is another characteristic of natural phenomena and is
becoming increasingly important in different areas of geographical data analysis and
environmental sciences (Csillag et al., 1996, Goodchild and Haining, 2004, Green and
Plotkin, 2007), i.e. at the more practical side of spatial statistics. Heterogeneity of
attributes of a geographical random field comprises two elements of second-order
variation: a global variance (population variance) and the spatial structure of that
variation (population spatial autocorrelation) (Wang et al., 2010). Both elements of
geographical variation need to be recognized in designing and evaluating sample
designs including the sampling plan and the choice of the estimator (Cochran, 1977;
Haining, 1988; Griffith, 2005).
Although pervasiveness and importance of heterogeneity has been realized and
studied, there are only a few methods on how to deal with it in the design-based
sampling family. Purposive sampling is generally more efficient than probability
sampling (Brus and Gruijter, 1997; Gruijter et al., 2006). However, purposive
sampling cannot satisfy the sampling randomness required by design based theory.

8
Two main methods dealing with the spatial heterogeneity are spatial stratified
sampling in design based sampling and the MSN sampling (Means of Surfaces with
Nonhomogeneity) in model based sampling (Wang et al., 2009a).
In spatial stratified sampling, the heterogeneous area is divided into several
subareas that are each more homogeneous than the area as a whole, in order to reduce
the total variance. The division may be based on prior knowledge, pre-sampling, an
effective proxy variable (Rodeghiero and Cescatti, 2008), or on the distribution of
other ancillary variables that are known to affect the value of the target variable. After
the division, the number of sampling sites is assigned to each subarea according to the
area proportion or variance proportion or both, followed by simple random sampling
applied to each subarea. Wang et al. (2010) suggested that there was usually no ‗true‘
or ‗correct‘ divisions of the whole interested area, but at any given scale of
geographical detail some divisions would be better than others. The authors
investigated the effect of the first-order heterogeneous on the performance of different
methods for estimating the population mean and its error variance, pointed out that
zoning improves estimator efficiency when sampling a heterogeneous random field
and compared the efficiency of different stratified methods.

3 Model-based sampling

In model-based sampling, the values in a region of interest are regarded as unfixed

and a set of values observed across the region represents a single realization of a
stochastic model of the variation in the universe. The randomness in this method is
introduced by a set of stochastic models also called the superpopulations. The
sampling sites are fixed and inference is based on assuming the validity of the
stochastic model. To compute an estimator or a predictor, the weights of the sample
data are determined by the covariances between the observations, which are given by
the model as a function of the coordinates of the sampling locations (van Groenigen
and Stein, 1998). In contrast to design based sampling, model-based sampling is
featured by an objective function predefined and that is minimized and thus leads to
the (single) optimal sampling scheme for that specific objective. Different objective
functions may lead to different optimal sampling schemes. A model-based sampling
scheme and model-based statistics can be applied to both estimating population
parameters (a single realization of a stochastic field) and superpopulation (a stochastic
field) parameters. Kriging, for example, can be considered as a model based method.
The incidences in a single year are considered as a population and a map can be
produced by kriging by interpolation of incidences at the nodes of a fine-meshed grid.
A prevalence map for a range of years is a superpopulation and can similarly be
produced by kriging interpolation of prevalences at the grid nodes, as well.
In this paper we distinguish three specific objectives that have shown to be
relevant within a spatial context: Minimizing the estimation error variance, achieving
an equal spatial coverage for irregular polygons and achieving an equal coverage in
feature space.

9
3.1 Objective 1: Minimizing the estimation error variance
Kriging is the main estimation method of geostatistics which was first used by D.J.
Krige and H.S. Sichel in mining theory in the nineteen fifties and sixties and was
further developed and applied by G. Matheron as a full-fledged scientific theory
(Matheron, 1963). It provides unbiased estimation of values of a variable at unvisited
locations with a minimum mean squared estimation error (Isaaks and Srivastava, 1989;
Pardo-Iguzquiza and Dowd, 1998; Olea, 1999; Bayraktar and Turalioglu, 2005; Saito
et al 2005; Wang et al., 2009a). It puts requirements on the random field, assuming it
to be for example second order stationary, and that thus can be represented as a
regionalized variable. Further, the random field is potentially autocorrelated in space,
which is then modeled by the covariance function or the variogram. Kriging is a
generic name and consists of many variants such as normal kriging, universal kriging,
block kriging, co-kriging, indicator kriging, factorial kriging, regression kriging and
other variants. MSN is another example of model based approach which takes both
spatial autocorrelation and spatial stratified non-homogeneity into account to
minimizing estimation error (Wang et al., 2009).
In order to minimize the estimation error variance a sample of a given size is to
be optimally distributed over space (Hughes and Lettenmaier, 1981; Russo, 1984;
Warrick and Myers, 1987; Sacks and Schiller, 1988; van Groenigen and Stein, 1998;
Bertolino, 1983; Wang et al., 2009). To achieve this purpose it is often required that
the spatial variation as provided by the variogram is known beforehand. If that is the
case, then the optimal model which is a function of the Kriging variance and sampling
sites are allocated initially at random in a region that are then shifted using an
optimization technique. Examples are spatial simulated annealing and particle swarm
optimization (Hu and Wang, 2011). The technique has been applied in many areas,
such as in soil surveys, environmental pollution monitoring and agricultural modeling
(van Groenigen et al., 2000; Zimmerman et al., 2005).

3.2 Objective 2: Equal spatial coverage

On various grounds it can be important to have a good coverage over a spatial
area. For once, it may serve as a general purpose type of optimality, there may be
good reasons not to miss out any sub-region, and there could be a good economic
argument to have sparse data equally distributed. Moreover, the first objective may
not be realistic as no variogram may be available. Spatial coverage is an alternative
in this case, which distributes sampling sites to cover the space as uniformly as
possible over a region. This is analog to the Horvitz-Thompson sampling, which
weights sample by its inclusion probability to cover unsampled units.
The criterion to apply is then the MMSD (Minimization of the Mean of the
Shortest Distances) criterion. The target quantity to be minimized is the mean of the
distances of un-sampled sites to the nearest sampling site. In order to achieve this, the
area is often discretized into a finite grid and the center of each grid cell is used to

10
compute this criterion. (van Groenigen et al., 1999). As an alternative one may use the
WMSD (Weighted Mean of the Shortest Distances): the mean of distances multiplied
with weights to give varying emphasis to different sub-regions. One other criterion is
Mean Squared Distance based on Thiessen polygons. The basic idea of this criterion
is that if the sampling sites are evenly distributed, the variance of the area constructed
by Thiessen polygons should equal zero, whereas for un-evenly distributed sampling
sites, the variance should be large. As before, sampling locations are moved at
random to reach the objective. More methods have been proposed to disperse the
distribution of the sample points. Typical examples are repulsion simulation that
distributes sampling sites evenly within a given irregular polygon via simulating the
movements of some ideal homogeneous point charges (Chen et al., 2012). For other
purposes, grid sampling, transect sampling, sequential sampling and nested sampling
are often used in practical sampling (Pettit and McBratney, 1993; Oliver, 1986).

3.3 Objective 3: Equal coverage in feature space

The objective to reach an equal coverage in feature space aims to distribute
the sampling units such that the samples reveal the population distribution as
good as possible. If the distribution of population is unknown, a postulated
distribution from experience or ancillary data is used. Two main criteria that
occur in the literature are the WM (Warrick-Myers) and LHS (Latin Hypercube
sampling) criteria.
The WM criterion optimizes sample locations for variogram estimation (Warrick
and Myers, 1987). It approximates the distribution of sampling units such that pairs of
the units follow a prespecified distribution as closely as possible. The LHS criterion is
a stratified random sampling procedure that provides an efficient way of sampling
variables from their multivariate distributions. It fully utilizes ancillary variables to
ensure each of the variables to be represented in a fully stratified feature space
(McKay et al., 1979; Iman and Conover, 1980). In a recent study, A stratified
conditional Latin hypercube sampling (scLHS) based on variance quadtree technique
(VQT) and LHS using ancillary variables was combined to coordinate both spatial
coverage and feature space coverage (Lin et al., 2011).

4 Population or Superpopulation ?

Originally, design based and model based methods were developed to estimate
population and superpopulation respectively (Cochran, 1977; Särndal, 1978). A major
characteristic of design-based sampling is that it is a combination of probability
sampling and design-based inference, whereas model-based sampling is a
combination of purposive sampling and interpolation (Gruijter et al., 2006).
Design-based sampling is more suitable for tackling the ‗how much‘ questions –
estimating global quantities, including frequency distribution of the target variable
and parameters of this distribution, such as the mean, the standard deviation or a

11
quantile. Model-based sampling is more suitable for tackling ‗where‘ questions - for
example to predict values at unvisited locations, mapping and for estimating the
parameters of the underlying stochastic model. If at least a medium sample size can
be afforded, strong autocorrelations exist and a reliable model of the spatial variation
is available, model-based sampling should be considered (Haining., 2003; Wang et al.,
2010). If parameters of a small domain or within a local area are to be estimated, both
methods may be favorable. Gruijter et al. (2006) pointed out the suitability of the two
methods as expressed by functions of spatial resolution at which estimates are
required. A decreasing function points to the appropriateness of design-based
sampling, i.e. the higher the spatial resolution, the more suitable design-based
sampling is. On the contrary the relative suitability of model-based sampling is an
increasing function of spatial resolution of estimator.
Whether a sampling objective concerns population or superpopulation can be
easily justified. Sampling concerns a population if the attribute to be estimated is
about ―here and now‖. Sampling concerns a superpopulation if the enumeration of the
sampling units is a single realization of a stochastic process, and parameters of the
latter is of interest. Population is one realization of a superpopulation. Table 1
summarizes the objective of sampling and inference, as well as several methods and
examples.

Table 1 Population and Superpopulation

Population (Definition: here and now distribution, Superpopulation (Definition: a process which
Objective
once realization of a process or superpopulation) generates a time series of spatial distributions)
areal mean distribution areal mean distribution
Assume iid iid iid iid
Model Cochran 1977 Cochran 1977 Sum of iid Sum of iid at each site
Daily average temp Maps of daily temp in
M1
Examples Average temp on 1 June Map of temp on 1 June in 2012 in a plain area
2012 in a plain area 2012 in a plain area 2012 in a plain area
Yearly average temp Maps of yearly temp on
on 1 June in NC 1 June in a plain area
Assume 2nd order stationary 2nd order stationary Slow Slow
Model Block Kriging with Kriging with spatial Block Kriging with Kriging with
spatial semivariogram semivariogram S-T semivariogram S-T semivariogram
M2
Maps of yearly annual
Average temperature Map of annual temp. in Yearly annual temp temp in in a plain area
Examples in 2012 in a plain area 2012 in a plain area in in a plain area Maps of daily temp in
2012 in a plain area
Assume Strata& Autocorrelation Strata & Autocorrelation temporal process* ST process**
Model MSN (Wang 2009)# Sandwich (Wang 2002)## Time series BME (Christakos 1992) ‡
Soil antibiotic residue in Map of soil antibiotic Yearly net Map of yearly net
a town surveyed in 2011 residue in a town surveyed accumulation of accumulation of
(Xie et al 2012) in 2011 antibiotic in soil antibiotic in soil
M3
$
Examples Incidence of NTD in Map of incidence of NTD Poisson process Maps of determinants
12
 2005 in Heshun county in 2005 in Heshun county Map of prevalence
An cross section Physical environment Map of the physical
observation of all cases induced NTD environmental NTD
Assume etc.
etc.
Model etc.
Method Design based (enumeration or statistics) Model based (spatiotemporal process)
or Model based (spatial process) or Design based (summation over long term)
Notes:
*
temporal process: the attribute changes with time
**
ST process: the attribute change in both spatial and temporal dimensions
#
MSN: a model to estimate the Mean of Stratified Nonhomogeneous Surface (Wang et al., 2009);
##
Sandwich: a model to make multi units reporting for stratified spatial attribute (Wang et al., 2002).
‡
BME: Bayesian Maximum Entropy, a model to make spatiotemporal interpolation (Christakos, 1992).
$
NTD: neural tube defect (Gu et al., 2007)

Table 1 lists several aspects of sampling in the spatial domain. In principle, these
could be expanded as well to the spatiotemporal domain (for monitoring) but at this
stage we restrict ourselves to spatial sampling, as is the topic of the paper. From the
above table, a population might be consisting of several superpopulations. For
example (see M1 in Table 1), observations of temperatures across a plain at a single
day (spatial population) consist of a single realization of the meteorological process
(spatial superpopulation) of daily variation within a year; alternatively, the population
could consist of temperature variations at that day during many years. Similar remarks
apply to disease observations (see M3 in Table 1), where the prevalence could be the
human population. The number of cases of a disease within a cross section (human
population) in China is a summation of incidences over previous years that have not
recovered yet, and hence there is a combination of prevalence and population.
Prevalence, however, could also be a superpopulation. For example, Neural Tube
Defects (NTD) prevalence in Heshun county is stable during decades or during
hundred years (Gu et al., 2007), reflecting the physical environment risk. NTD is a
rare disease, i.e. a disease of a small probability, and single or several years
incidences (within the human population) do not reflect the baseline risk in the area.
Both spatial (design based) and temporal (model based) sampling highly varied. The
boundary of the two methods vanishes when time becomes longer. In the same way,
incidence could be either population or superpopulation.
The design-based or model-based sampling methods actually provide the full
sampling strategy, i.e. the components of the triple ( , , ).
Design-based or model-based ideas also apply to non-sampling aspects as well,
and kriging is a good example. Kriging assumes a form of stationarity, such as 2 nd
order stationarity, intrinsic stationarity or strict stationarity. The data are assumed to
originate from a random field and because of this assumption sampling within a
kriging context could be categorized into model based method (Särndal, 1978; Brus
and Gruijter, 1997). When spatially interpolating with kriging values at unsampled
sites are predicted, and in that sense is closer to a population technique than to a
13
superpopulation based technique. In that sense it is also a design based method. It
shows, that model does not need to be categorized as a specific type, i.e. being either
design based or model based, as that may lead to unnecessary logical confusion.
This may have an effect on the sampling strategy as well. It is best to first judge
whether the objective of using a technique concerns a population or a superpopulation.
Next, appropriate models should be identified that match the objective and a strategy
should be decided on how to reach the objective. For example, both incidence and
prevalence could be both considered from a population and from a superpopulation
point of view. Before designing where to sample, samplers must decide which
sampling strategy is to be adopted. The appropriate strategy then depends upon the
objective and relevant constraints.

5. Biased Sampling and Remedy

An estimate of a parameter is biased if its expectation differs from the true value
of the parameter, whatever the true value is, and the estimate is unbiased if its
expectation is identical to the true value. The expectation is a value for population or
superpopulation parameter, such as its total, mean, the value of an individual
unsampled unit, etc. An estimate is obtained by applying a statistic to a sample y.
= (y)
Therefore, the bias of an estimate depends upon the interaction between the
sample y and the statistic . A sample y is often said to be unbiased if it is drawn
randomly (thus each sample has an equal probability of being obtained) from its
population. But this would lead to a conflict to common sense, as elaborated in the
following section. Herewith a sample is unbiased if it results in an estimate unbiased
to the population under a simple random statistic, a so called SRS unbiased sample.
An unbiased estimation can be achieved either by a sample randomly drawn (with
equal probability) from its population, or by a sample biased to its population under
SRS or remedied by a particular statistic.
Figure 1 assumes that the population is the area within the outermost circle, the
parameter of the population to be estimated is the innermost circle, and four samples
from the population are investigated where the sampling units are indicated by points.
An SRS unbiased estimate can be obtained either by drawing randomly from a
population (Figure 1d), or by non-randomly drawing (unequal probability) from a
population (Figure 1c). A biased sample, however, does not necessarily lead to a
biased estimate, because several statistics like e.g. the Heckman method (Heckman,
1979) and the B-shade model (Wang et al., 2011) can remedy the bias of a sample in
some circumstances and result in an unbiased estimate. In another words, an estimate
(y) may be biased under SRS, but may be unbiased under some other statistics, such
as the Heckman method and the B-shade method . Figures 1a,b show a biased sample
with low and high variation, whereas Figures 1c,d show unbiased sampling of a low
and of a high variation. Figure 1c is an SRS unbiased sample while the sample is
biased to population. The above discussion was summarized in Table 2.

14
 a b

c d

Figure 1 Biased Sampling

Table 2 A combination between sampling and inference for Figure 1

Way of statistics
Bias of an estimate
SRS Remedy
Biased to both population and SRS*: Fig.1a, b A. Biased B. Unbiased
Sample Biased to population: Fig.1c
Unbiased to SRS C. Unbiased D. N/A
Unbiased to population: Fig.1d
* SRS stands for simple random sampling

6. Conclusion and Discussion

Spatial sampling can raise its efficiency by embedding prior knowledge about the
random fields. From this perspective, spatial sampling theories is developing along
the path: sampling population which is iid (Cochran, 1977), spatially autocorrelated
(Haining, 2003; Stein and Ettema, 2000; Christakos, 2005), spatial heterogeneous as
well as spatially autocorrelated (Goovaerts, 1997; Wang et al., 2009b, 2010). The
properties of the latest methods needs to be studied in more detail; the relationship
between distinct random fields, ways of sampling and inference, and the precision of
estimation deserve to be investigated thoroughly. Although a division between design
based and model based methods are relevant, some methods are seemly difficult to
categorize into any of the two classes. Actually, population and superpopulation
approaches are easily justified, and a research question can be defined flexibly.
Finally, spatial sampling and inference methods can be chosen from a broad spectrum
of available methods and techniques quickly and appropriately and taking all relevant
constraints into consideration.
The uncertainty of the final estimation originates, propagates and accumulates in
the trinity ( , , ) of spatial sampling and statistical inference. It consists of (1) the
stochastic field defined by its distribution such as variability aspects; (2) the spatial
sampling , including its unbiasedness and density; and (3) the spatial statistic for
statistical inference. We notice that bias of may be overcome by a proper choice of
. In practice, the determinants X of the sampling attribute may be helpful to reduce
the uncertainty, especially when the sample is small or biased. (4) Determinants X:
15
when the original sample is small a large sample of direct determinants of the original
sampling attribute or the proxy variables strongly associated to the direct determinants
can be used to estimate the parameters in spatial statistics, such as stratification, auto-
and cross- correlation, and a ratio between sample and population values. This was
illustrated by a small sample survey of occupational diseases in China, where data for
the number and size of coal mining factories in each county in China, available from
the statistic year book of local government served as a large sample of determinants.
In addition, production of the mining industry in each county was available from the
statistic year book of central government and served as proxy variables.
This brings us finally to the role of a spatial statistician. S/he may contribute to the
implementation of a spatial statistic if the sample was already collected, or to both
spatial sampling and the spatial statistic if the data (at least partly) still have to be
collected. The role of disciplinary scientists like epidemiologists or other experts in
the domain represented by the field could be responsible for choosing determinants
X mostly related to the sampling attribute.

Acknowledgments
This study was supported by NSFC (41023010), MOST (2012CB955503;
2012ZX10004-201; 2011AA120305) and CAS (XDA05090102) grants.

Reference

Bayraktar, H, Turalioglu, F.S., 2005. A Kriging-based approach for locating a sampling site—In
the assessment of air quality. Stochastic Environmental Research and Risk Assessment 19(4),
301–305.
Berry, B.L., Baker, A.M., 1968. Geographic sampling. In: B.L. Berry and D.F. Marbles, eds.
Spatial Analysis. Englewood Cliffs, NJ: Prentice-Hall, 91–100.
Bertolino, F., Luciano, A., Racugno, W., 1983. Some aspects of detection networks optimization
with the kriging procedure. Metron, 41(3), 91-107.
Brus, D.J., de Gruijter, J.J., 1997. Random sampling or geostatistical modelling? Choosing
between design-based and model-based sampling strategies for soil (with discussion).
Geoderma 80, 1–44.
Chen, B.S., Pan. Y.C., Wang, J.H., 2012. Even sampling designs generation by charges repulsion
simulation. Environmental Monitoring and Assessment 184, 3545-3556.
Christakos, G., 2005. Random field models in earth sciences. Dover Publications, New York.
Cliff, A. D., Ord, J. K., 1981. Spatial Processes: Models and Applications. Pion, London.
Csillag, F., Kertesz, M., Kummert, A., 1996. Sampling and mapping of heterogeneous surfaces:
multi-resolution tiling adjusted to spatial variability. International Journal of Geographical
Information Systems 10, 851–875.
Cochran,W.G., 1946. Relative accuracy of systematic and stratified random samples for a certain
class of populations. Annals of Mathematical Statistics 17, 164–177.
Cochran, W. G. 1977. Sampling Techniques. 3rd ed. Wiley, New York.

16
Das, A.C., 1950. Two-dimensional systematic sampling and the associated stratified and random
sampling. Sankhya 10, 95–108.
Dunn, R. and Harrison, A.R., 1993. Two-dimensional systematic sampling of land use. Applied
Statistician 42, 585–601.
Goodchild, M.F., Haining, R.P., 2004. GIS and spatial data analysis: converging perspectives.
Papers Regional Science 83, 363–385.
Goovaerts, P., 1997. Geostatistics for Natural Resources Evaluation. Oxford University Press
Green, J.L., Plotkin, J.B., 2007. A statistical theory for sampling species abundances. Ecology
Letters 10, 1037–1045.
Griffith, D.A., 2005. Effective geographic sample size in the presence of spatial autocorrelation.
Annals of the Association of American Geographers 95, 740–760.
de Gruijter, J. J., ter Braak, C. J. F., 1990. Model-Free Estimation from Spatial Samples: A
Reappraisal of Classical Sampling Theory. Mathematical Geology 22(4), 407-415.
de Gruijter, J. J., Brus, D..J, Bierkens, M.F.P., Knotters, M., 2006. Sampling for Natural Resource
Monitoring. Springer-Verlag, New York.
Gu, X., Lin, L.M., Zheng, X.Y., Zhang, T., Song, X.M., Wang, J.F., Li, X.H., Li, P.Z., Chen, G.,
Wu, J.L., Wu, L.H., Liu, J.F., 2007. High prevalence of NTDs in Shanxi Province: A
combined epidemiological approach. Birth Defects Research Part A – Clinical and Molecular
Teratology 79 (10), 702-707.
Haining, R.P., 1988. Estimating spatial means with an application to remote sensing data.
Communication Statistics - Theory and methodology 17, 537–597.
Haining, R.P., 2003. Spatial Data Analysis: Theory and Practice. Cambridge University Press,
Cambridge.
Hansen, M.H., Hurwitz, W.N., Madow, W.G., 1953. Sample Survey Methods and Theory. Wiley,
New York.
Heckman, J.J., 1979. Sample selection bias as a specification error. Econometrica 47(1), 153-161.
Horvitz, D.G., Thompson, D.J., 1952. A generalization of sampling without replacement from a
finite universe. Journal of the American Statistical Association 47, 663-685.
Hu, M.G, Wang, J.F., 2011. A spatial sampling optimization package using MSN theory.
Environmental Modeling & Software 26(4), 546-548.
Hughes, J.P., Lettenmaier, D.P., 1981. Data requirements for Kriging: Estimation and network
design. Water Resources Research 17, 1641-1650.
Iman, R.L., Conover, W.J., 1980. Small sample sensitivity analysis techniques for computer
models, with an application to risk assessment. Communications in Statistics A9, 1749–1874.
Isaaks, E.H., Srivastava, R. M., 1989. Applied Geostatistics. Oxford University Press, New York.
Li, L.F., Wang, J.F., Cao, Z.D., Feng, X.L., Zhang, L.L., Zhong, E.S., 2008. An
information-fusion method to regionalize spatial heterogeneity for improving the accuracy of
spatial sampling estimation. Stochastic Environmental Research and Risk Assessment 22(6),
689-704.
Lin, Y.P, Chu, H.J., Huang, Y.L., 2011. Monitoring and identification of spatiotemporal landscape
changes in multiple remote sensing images by using a stratified conditional Latin hypercube
sampling approach and geostatistical simulation. Environmental Monitoring and Assessment
177, 353–373.
Matheron, G., 1963. Principles of geostatistics. Economic Geology 58, 1246-1266.

17
Mátern, B., 1986. Spatial Variation. Lecture Notes in Statistics 36. Springer-Verlag, Berlin.
McKay, M.D., Beckman, R.J., Conover, W.J., 1979. A comparison of three methods for selecting
values of input variables in the analysis of output from a computer code. Technometrics 21,
239–245.
Milne, A., 1959. The centric systematic area-sample treated as a random sample. Biometrics 15,
270–297.
Neyman, J., 1934. On the two different aspects of the representative method: the method of
stratified sampling and the method of purposive selection. Journal of the Royal Statistical
Society 97, 558–606.
Neyman, J., 1938. Contribution to the theory of sampling human populations. Journal of the
American Statistical Association 33,101-106
Olea, R. A., 1999. Geostatistics for Engineers and Earth Scientists. Boston. Kluwer, MA.
Oliver, M. A., Webster, R., 1986. Combining nested and linear sampling for determining the scale
and form of spatial variation of regionalized variables, Geographical Analysis 18, 227-242.
Pardo-Iguzquiza, E., Dowd, P.A., 1998. The second-order stationary universal Kriging model
revisited. Mathematical Geology 30(4), 347–378.
Payandeh, B., 1970. Relative efficiency of two-dimensional systematic sampling. Forestry Science
16, 271–276.
Pettitt, A.N., McBratney, A.B., 1993. Sampling designs for estimating spatial variance
components. Applied Statistics 42, 185-209.
Rogerson, P.A., Delmelle, E., Batta, R., Akella, M., Blatt, A., Wilson, G., 2004. Optimal sampling
design for variables with varying spatial importance. Geographical Analysis 36(2), 177-194.
Quenouille, M.H., 1949. Problems in plane sampling. Annals of Mathematical Statistics 20, 355–
375.
Ripley, B., 1981. Spatial Statistics. Wiley, New York.
Royall, R.M., 1968. An old approach to finite population sampling theory. Journal of the
American Statistical Association 63, 1269-1279.
Rodeghiero, M., Cescatti, A., 2008. Spatial variability and optimal sampling strategy of soil
respiration. Forest Ecology and Management 255(1), 106–112.
Russo, D., 1984. Design of an optimal sampling network for estimating the variogram. Soil
Science Society of America Journal 48, 708-716.
Sacks, J., Schiller, S., 1988. Spatial designs. In: Gupta, S.S., Berger, J.O. (Eds.), Statistical
Decision Theory and Related Topics IV, vol. 2. Springer Verlag, New York. 385–399.
Saito H., McKenna S. A. and Zimmerman D. A. et al., 2005. Geostatistical interpolation of object
counts collected from multiple strip transects: Ordinary Kriging versus finite domain Kriging.
Stochastic Environmental Research and Risk Assessment 19(1), 71–85.
Särndal, C.E., 1978. Design-based and model-based inference in survey sampling. Scandinavian
Journal of Statistics 5, 27–52.
Särndal, C.E., Swensson, B., Wretman, J., 1992. Model Assisted Survey Sampling. Springer, New
York.
Stein, A., Ettema, C., 2003. An overview of spatial sampling procedures and experimental design
of spatial studies for ecosystem comparisons. Agriculture, Ecosystems and Environment 94(1),
31-47.
van Groenigen, J.W., Stein, A., 1998. Constrained optimization of spatial sampling using

18
 continuous simulated annealing. Journal of Environmental Quality 27(5), 1078–1086.
van Groenigen, J. W., Pieters, G,. Stein, A., 2000. Optimizing spatial sampling for multivariate
contamination in urban areas. Environmetrics 11(2), 227-244.
van Groenigen, J.W., Siderius, W., Stein, A., 1999. Constrained optimisation of soil sampling for
minimisation of the kriging variance. Geoderma 87, 239–259.
Wang, J.F., Christakos, G, Hu, M.G., 2009a. Modeling spatial means of surfaces with stratified
nonhomogeneity. IEEE Transactions on Geoscience and Remote Sensing 47(12), 4167-4174.
Wang, J.F., Jiang, C.S., Li, L.F., Hu, M.G., 2009b. Spatial Sampling and Inference. Science Press,
Beijing.
Wang, J.F., Haining, R.P., Cao, Z.D., 2010. Sample surveying to estimate the mean of a
heterogeneous surface: reducing the error variance through zoning International Journal of
Geographical Information Science 24(4), 523–543.
Wang, J.F., Reis, B.Y., Hu, M.G., Christakos, G., Yang, W.Z., Sun, Q., Li, Z.J., Li, X.Z., Lai, S.J.,
Chen, H.Y., Wang, D.C., 2011. Area disease estimation based on sentinel hospital records.
PLoS ONE 6(8), e23428.
Warrick, A.W, Myers, D.E., 1987. Optimization of sampling locations for variogram. Calculations
Water Resources Research 23(3), 496-500.
Zimmerman, D. L., Holland, D.M., 2005. Complementary co-kriging: spatial prediction using data
combined from several environmental monitoring networks. Environmetrics 16(3), 219-234.
Zubrzycki, S., 1958. Remarks on random, stratified and systematic sampling on a plane.
Colloquium Mathematicum 6, 251-264.

View publication stats

19