Environ. Sci. Technol.
1993, 27, 2035-2044
Sampling of Contaminated Soil: Sampling Error in Relation to Sample Size
and Segregation
Frank P. J. Lam6'vt and Peter R. Defize*
Institute of Environmental Sciences IMW-TNO and Institute of Applied Physics, Centre of Applied Statistics, The Netherlands
Organization for Applied Scientific Research TNO, P.O. Box 601 1, 2600 JA Delft, The Netherlands
When sampling a mixture of particles of different nature
like soil, there will always be a minimal amount of variation
between the samples due to the fact that the particle
composition of each sample will differ. This minimal
variation in sampling a particulate material is known as
the fundamental error. It was previously unknown to what
extent this fundamental error could influence the variance
in a series of soil samples taken for analysis. On the basis
of experiments with soil contaminated with cyanide, we
conclude that the fundamental error only influences the
variance when samples smaller than 10g are taken. When
using larger samples, the segregation error determines the
variance. However, there is no practically applicable
method available for estimating the segregation error
before sampling and analysis. Therefore, we developed a
method for estimating the segregation error using a
sampling board. This resulted in segregation plots which
can be compared to the lot to be sampled.
Introduction
In quality control of cleaned soil, both the mean
concentration and the variance in this concentration are
important testing criteria (1). In order to check on both
criteria, samples of the cleaned soil must be taken and
analyzed.
Publications on the sampling of particulate materials,
however, indicate that due to the fact that the samples
consist of individual particles with different characteristics
the variance is influenced by the size of the sample. As
the variance is one of the quality control criteria for cleaned
soil, it is essential to know to what extent the sample size
affects the variance in concentration between samples.
For this reason, we studied the effect of sample size on
sampling error.
In a strongly simplified model of the sampling of a
particulate material, namely, the sampling of a mixture
of red and white balls, the effect of sample size on sampling
error can be calculated using simple binomial statistics.
In the case of the sampling of a complex (and even partially
unknown) mixture of different kinds of particles, like soil,
the effect of sample size on sampling error cannot be
predicted with simple statistics. More precise theories
are needed to describe the relationship between sample
size and sampling error. In a review of literature in
Geophysical and Chemical Abstracts data bases, as well
as in literature on this subject which we already possessed,
we found only a relatively small number of publications
in which the effects of sample size on sampling error have
been described. The most relevant publications on this
subject are shown in the reference list to this paper (2-
14).
* Author to whom correspondence should be addressed.
t Institute of Environmental Sciences IMW-TNO.
t Institute of Applied Physics, Centre of Applied Statistics.
0013-936X/93/0927-2035$04.00/0 0 1993 American Chemlcal Society
When comparing these publications with the problem
of sampling contaminated soil, two major differences were
observed: (a) The concentrations are very high (% level)
compared with concentrations in contaminated soil (ppm
level) and (b) for most materials, the particles are much
larger than soil particles. Due to these two major
differences, it was unknown if these theories could be used
to predict sampling error in relation to sample size in
contaminated soil. Therefore, two experiments were
performed to investigate to what extent the existing
theories are applicable to contaminated soil.
In this paper we will give a short overview of the sampling
theories, focusing on the two basic causes of sampling error
when handling a particulate material: the fundamental
error and the segregation error. The experimental results
will be discussed and compared with the existing sampling
theories. Based on these results, a supplementary ap-
proach will be presented to tackle the problem of sampling
a particulate material, like contaminated soil, using an
appropriate sample size.
Theory
The theory of sampling particulate materials is vast and
complicated. It has been widely accepted that the most
complete theory has been developed by Gy (2). Other
references are Harnby (9), Ingamells (10,11),and Visman
(3). As our main point of interest was to investigate the
relation between sample size and sampling error, we
compared experimental results with theoretical ones. The
theories of Gy and Visman appeared to be the most useful.
Gy's Theory. As mentioned by Pitard (12,131, the
publications by Gy are hard to digest and difficult to read.
Gy introduces a whole set of new terminology and concepts
which are the building blocks of his sound and thorough
theory. Here we shall only discuss in short those parts of
his theory which are important for our study.
Cy defines two forms of heterogeneity: constitution
heterogeneity (CHL)and distribution heterogeneity (DHL).
CHL is the heterogeneity that is inherent to the compo-
sition of each particle making up the lot. The greater the
difference in compositionbetween each particle, the greater
the constitution heterogeneity. The DHL of a lot is the
heterogeneity that is inherent to the manner in which the
individual fragments are distributed over the lot. Dis-
tribution heterogeneity is directly connected with segre-
gation (a lot is segregated when the critical component is
concentrated in certain parts of the lot). The distribution
heterogeneity can be reduced by reducing the segregation,
e.g., by mixing the lot.
The mathematical definition of CHL is as follows:
Envlron. Scl. Technol., Vol. 27, No. 10, 1993 2095
with
where N is the number of particles in the lot, ai is the
concentration of the critical component in particle i, U L is
the concentration of the critical component in the lot, Mi
is the weight of the particle i, and is the mean weight
of the particles; hi is called the standardized deviation of
particle i.
The definition of DHL is similar to that of CHL.Instead
of dividing the lot into particles, the lot is divided into
groups of particles. The variance of the standardized
deviation of the groups is defined as DHL.The magnitude
of DHL for a lot is not a fixed value but depends on the
grouping that has been chosen.
The fundamental sampling error plays an important
role in the theory of Gy. This is the error that ‘remains’
under ideal sampling conditions due to the different
characteristics of the particles. These conditions are met
if the sampling procedure can be thought of as a Bernoulli
experiment where each particle of the lot has an equal
probability to be part of the sample.
Gy showed that the relative variance a 2 of~ the funda-
mental error is related to CHL as follows:
(3)
where P is the selection probability of a particle, and a,
is the concentration of the critical component in the
sample.
This theoretical formula cannot be applied in practice.
The formula, however, is a basis for the derivation of
formulas that can be used.
In fact, Gy did this and derived the following formula
under certain assumptions:
Cd3
~ ~ ~ (=a-with
, ) C = clfg (4)
M.9
where M, is the weight of the sample, c is the composition
factor, 1 is the liberation factor, f is the shape factor, g is
the size distribution factor, and d is the maximum diameter
of particles. These parameters and the maximum diameter
can be estimated from the characteristics of the particles.
The quantity C is called the sampling constant and is the
product of 4 parameters: c, 1,f, andg. Foreachparameter,
we give a short description.,
C o m p o s i t i o n Factor c. The following formula holds for
this factor:
(5)
where U L is the concentration of the critical component in
the lot, A, is the density of the critical component, and A,
is the density of the matrix.
Liberation Factor 1. The liberation factor has a value
between 0 and 1 and defines to what extent the critical
component is ‘liberated’ from the matrix (complete
liberation: 1 = 1). Gy (2) gave a rule of thumb on how 1
can be calculated from the values of the maximum diameter
d and the liberation diameter di. The latter parameter is
defined as ‘the particle size of a comminution that would
ensure the complete liberation of the constituent of
interest’ (this definition is not very precise and one of the
2038 Environ. Sci. Technol., Vol. 27, No. IO, 1993
bottlenecks in applications):
when d 5 dl then 1=1
when d > dl then 1 = (dl/d)l/2
S h a p e Factor f. This factor is related to the shape of
the particles. It is defined as the ratio of the volume of
a particle passing a certain sieve to the volume of a cube
passing the same sieve [Minkkinen (14)l. Gy stated that
a value of 0.5 for the shape factor is a good estimate for
most materials. For particles with a coin-shape, however,
avalue in the range of 0.1-0.2 should be taken. Pitard (ref
12, p 160) described a method to estimate f by means of
sieve experiments.
S i z e Distribution Factorg. Gy (2)gave a rule of thumb
on how to calculate g. Let d be the maximum diameter
and d’ give the minimum diameter of the particles:
if did‘ > 4 then g = 0.25
if 2 < did’ 5 4 then g = 0.50
if 1 < did’ 5 2 then g = 0.75
if d/d‘ = 1 then g = 1.00
Thus, as the variation in size diameter increases, the size
distribution factor decreases. From eq 4 it follows that
the relative variance of the fundamental error is inversely
proportional to the sample weight. It has been found by
others [Ingamells (IO), Harnby (91,Visman (311 that such
a relationship holds in many practical situations. It was
Gy’s merit that he gave a method to estimate the constant
C , purely on the basis of material properties and an initial
estimate of the concentration of the critical component.
Segregation. When sampling in the field or from a lot,
segregation is an important factor which affects the
sampling error. Usually, it cannot be assumed that lots
are perfectly mixed. This section is therefore concerned
with quantifying the effect of segregation on sampling
error in relation to sample size.
Consider a lot of N F particles to be divided into NG
subsamples. A sample is obtained by a Bernoulli exper-
iment with selection probability P for each subsample.
The selected subsamples are combined in one final sample.
Then the relative variance of the final sample can be shown
to be equal to
This relative variance can be compared with the relative
variance of the fundamental error as follows.
Define the grouping factor as
(7)
This factor is almost equal to the average number of
particles in a subsample when the sample size is small
relative to the size of the lot. Gy showed that the following
inequality between CHL and DHL holds:
By introducing a segregation factor [ with 0 5 5 5 1,one
can write
From eqs 3 and 6 it follows that
azo = (1 + y[)a$ (10)
Logically, 620 is always larger than U ~ F .How much larger
depends on the value of 75. Gy stated that ‘when designing
a sampling plant, it is always safe to assume that 75 = 1’
(ref 2, p 343). This was also confirmed by Pitard, who
mentioned: ‘Taking into account the large amount of
variographic experiments he performed between 1960 and
1975, Gy reached the pragmatic conclusion that the
product of yt stays slightly smaller than 1’ (ref 12, p 188).
In most cases, however, the sampling of contaminated
soil will deviate strongly from the ideal sampling that is
possible in a specially designed sampling plant. In the
majority of cases, soil contamination will have a strongly
heterogeneous nature, and therefore the amount of seg-
regation within a lot will be high. This is illustrated by
Figure 2. The sampling of contaminated soil, therefore,
is one of the cases in which the assumption that 75 I 1
is not correct.
A source of inspiration for our own research was the
work of Visman (3). He is one of the few who did practical
research into the effect of segregation on sampling error.
Visman did hypothetical experiments with a ‘sampling
board’. This sampling board consisted of 10 000 positions,
2500 of which were occupied by lead pellets. The sample
taken from the board consisted of n neighboring positions.
The sample result was the fraction of positions containing
a pellet. Visman studied the variance of this fraction (a2)
in relation to the distribution of the pellets over the
sampling board. From these empirical experiments, he
concluded that the following relationship between variance
and sample size approximately holds:
a’= (A/M,) + B (11)
In this formula, A and B are parameters and M , is the
sample size. Visman called the quantity AIM, the random
variance and B the segregation variance. The parameters
A andB should be estimated from available data. Visman’s
formula was discussed by Duncan (41, who also suggested
some modifications.
Equations 10and 11are identical if the segregationfactor
5 is taken as a constant. The second term in eq 10 (75 a 2 ~ )
is then independent of sample size because y is directly
proportional to the sample size. By taking the product of
y and &, the influences cancel each other. The segre-
gation factor is not a constant by definition, but
simulation results of Duncan (4) and ourselves with a
sampling board give some confidence that it is a fairly
good approximation.
Experimental Section
Introduction. In view of the sampling theory by Gy
(2), a first experiment was carried out with soil contam-
inated with cyanide. Cyanide, mainly present as ferric
hexacyanoferrate(II), is a stable salt (15);it is present in
the soil in the form of small particles (ferric hexacyano-
ferrate(II1, Fe4[Fe(CN)&, better known as Prussian Blue;
it is a cyanide-iron complex which is formed during the
dry cleaning of coal gas and one of the major contaminants
found on former gasworks sites). It may therefore be
assumed that this type of contamination behaves as a
particulate material. Furthermore, the contaminated soil
can be dried at low temperature and, subsequently, be
ground and divided with the aid of a rotating divider,
without affecting the concentration. In this way it is
possible to make an analytical sample of the normal size
(50g) without allowing alarge error during sample division.
Of course, this assumption must be checked by duplicate
subsampling and analysis.
To confirm the results of the first experiment, it was
followed by a second one. In this second experiment, a
soil which was very heterogeneously contaminated with
cyanide was sampled and analyzed. The results of these
two experiments were compared with each other as well
as with the theoretical relation between sample size and
sampling error.
First Experiment. During the thermal cleaning process
of soil originating from a former gasworks site and
contaminated with cyanide and PAHs, 60 bulk samples
were taken from the conveyor belt which entered the plant.
It was preferred to sample the contaminated soil because
sampling of the cleaned soil would lead to a relatively
large number of samples with concentrations below the
detection limit, which could make it impossible to study
the effect of the sample size on sampling error.
From each bulk sample, subsamples of 0.1, 1, 50,400,
and 6500 g were taken without any further homogenizing
of the bulk samples. In this way, the subsamples taken
were comparable to samples of these sizes taken directly
from the conveyor belt.
The subsamples were dried at 40 “C and, for samples
larger than the normal analytical sample size of 50 g, ground
over a 1-mm sieve and divided with a rotating divider.
After the original sampling from bulk samples, duplicate
samples were taken during pretreatment and analysis. As
a result, it was possible to estimate the different variance
components due to sampling (a2(,)), due to sample pre-
treatment ( Q ~ ( ~ ) )and
, due to analysis (a2(*)), together
resulting in the total variance (a2(t)= a2(s)+ a2(p) +
(the number of duplicate samples and analyses was based
upon earlier studies with this type of contamination). By
means of an unbalanced hierarchical analysis of variance
(161,we were able to give unbiased estimates of the variance
components, without analyzing a very large number of
subsamples. An example of the resulting sampling
schemes is shown in Figure 1. Table I shows estimates of
the variance components and the total variance.
An intensive study of the soil and its cyanide contam-
ination was made in order to gain information about the
distribution of the cyanide in the soil. This included the
determination of the particle size distribution of the soil,
the distribution of cyanide over the particle size classes,
a microscopic study of the contaminated particles using
an optical microscope as well as a scanning electron
microscope (SEMI, astudyof the form in which the cyanide
is present by means of infrared microscopicanalysis, X-ray
microanalysis (XRMA), and specific analytical methods
to determine the amount of complex cyanide (Prussian
Blue), free cyanide (CN-), and thiocyanides (SCN-). As
a result of this part of the study, it was possible to obtain
estimates of the parameters which were used in the formula
for the fundamental error (eq 4). A direct comparison of
theoretical and practical results could therefore be made.
Second Experiment. To ensure a strongly heteroge-
neous distribution of cyanide, we used a contaminated
site for the second experiment. The top soil layer of a
small part of the site was carefullyremoved before sampling
in order to make the contamination visible. The exper-
Environ. Sci. Technol., Vol. 27, No. 10, 1993 2037
 Sampling etc. (60samples)

Pretreatment etc. (105 su bsamples)

Analysis

Flgure 1. Samplingscheme for the sample pretreatment of 400-g samples. The original number of samples (60) was based on previousexperlence
with the analysls of cyanide and the variance components involved (sampling, sample pretreatment, analyses). Knowing the order of magnitude
of the different variance components,a hierarchicalanalysis of variance could be designed which would give the deslred power of dlscrlmlnatlon.
For each series of sample sizes a dlfferent pretreatment was necessary, resulting from the size of the original sample and the amount of soil
used for the analysis (=50 9). Samples smaller than 50 g were dried and, excludlngthe 0.1- and 1-g samples, ground over a 1-mm sleve; samples
larger than 50 g were also dlvlded In subsamples by means of rotating dividers. Therefore, each sample slze had its own pretreatment pattern
and analysls of variance. For the 4 0 0 9 samples, this resulted in 105 subsamples and 120 analyses.

Table I. Mean Concentration, Total Variance, and Variance Contributions for First Experiment
sample size (9)
0.1 1 50 400 6500
no. of analysed 60 72 90 120 50
mean concentration (mg/kg) 1194 501. 480 506 529
total variance (&)) 3.09 X 10’ 2.87 x 104 8.04 x 103 2.73 x 104 5.46 X 109
Variance Components
absolute
sampling ( u 2 d 2.72 x 104 7.93 x 103 2.48 x 104 5.16 x 103
pretreatment ( U Z ( ~ ) ) 2.37 x 103 Ob
analysis 1.51 x 103 110 90 300
relative ( % )
sampling (&) 94.7 98.6 91.0 94.5
pretreatment ( u ~ ( ~ ) ) 8.7 Ob
analysis (a!$*)) 5.3 1.6 0.3 5.5
The number of analyses depends upon the pretreatment scheme used; 0.1 g, no pretreatment, no duplicate analyses (no analysis of
variance); 1 g, no pretreatment, some duplicate analyses; 50 g, drying and grinding, duplicates of pretreatment and analyses; 400 g, see Figure
1;6500 g, drying, grinding, subsampling, duplicates of pretreatment, and analyses. In the 6500-gseries, only half of the original bulk samples
were used. Negative estimates, which are theoretically possible due to random variation, have been replaced by zero estimates.

imental area was then divided into 60 squares of 0.25 m2.
Starting from the center of each square, samples were
taken from a surface of 0.0001, 0.01,0.09, and 0.25 m2and
a depth of approximately 1cm. With each sampling step,
the center which had already been sampled was not
sampled again. This resulted in 60 samples of approxi-
mately 1, 400, 2500, and 6000 g. The sampling of 0.1-g
samples took place in the laboratory. Subsampling from
both the 2500- and 6000-g samples resulted in two series
of 0.1-g samples.
Sample pretreatment and analysis comparable to the
first experiment were carried out. Because of the heter-
ogeneity within the lot, see Figure 2, the resulting
distribution of the analytical data was strongly skewed, as
shown in Table 11. The unbalanced hierarchical analysis
of variance used in the first experiment resulted therefore
in highly unreliable estimates of the variance components.
Therefore, only an estimate of the totalvariance was made.
By neglecting the variance caused by sample pretreatment
and analysis (these components proved to be less than
10% of the total variance in the first experiment), the
total variance slightly overestimates the sampling variance.
In contrast to the first experiment, the distribution and
particle characteristics of the cyanide contamination were
not investigated in the second experiment. Therefore, it
was not possible to give estimates for the parameters in
the formula of the fundamental error (eq 4)and to compare
the experimental results with theoretical estimates. Any-
how, the results of this second experiment made a
comparison irrelevant, as will be shown in the discussion.
Discussion of the Results. By knowing the distri-
bution and the nature of the cyanide contamination in the
first experiment, we were able to make good estimates of
the shape factor f (=0.5), the size distribution factor g
(=0.25), the composition factor c (=3600 g/cm3), and the
maximum particle size d (=0.8 cm). As already mentioned,
the estimate of the liberation factor 1 is more difficult.
Calculation of 1 from d and dl gave a very low value for 1.
In line with Gy’s recommendation (ref 2, p 2621, a minimum
value of 0.03 for 1 was chosen to calculate the fundamental
error.
In Figure 3, the coefficient of variation according to
Gy’s formula (= CJF) is plotted against the sample size,
together with the coefficient of variation due to the
sampling calculated from the data of the first experiment.
When the experimental variance for the 0.1-g samples is
not taken into account, there seems to be no significant
difference between the variances found for the different
sample sizes. Furthermore, the shape of the experimental
relationship does not support the theoretical relationship
based upon Gy’s theory with respect to the fundamental
error.
The fundamental error is, for at least part of the
investigated sample size range, larger than the experi-
mental sampling error. As the fundamental error is the
smallest possible sampling error, it is expected that the

2038 Envlron. Scl. Technol., Vol. 27, No. 10, 1993

 8
4
R
C
0
.J

4
a3
c
8
h

Q)
-0
.J
8
u,
c
a3

0 m
e

Flgure 2. Cyanide distribution on the sampling location for the second experiment.

Table 11. Mean Concentration (mg/kg) and Total Variance for Second Experiment
samde size (a)
0.1 0.1 1 400 2500 6000
no. of analysesn 72 72 72 84 84 84
mean 99.6 118 267 116 126 200
total variance 1.16x 105 2.70 X lo6 5.55 x 105 0.64 x 104 0.81 x 104 3.01 X 106
median 19.5 21.5 72.0 28.8 30.0 36.0
lower quartile 6.5 3.0 27.5 13.8 16.0 19.0
upper quartile 92.0 57.0 217 88.5 96.5 147
minimum 2.0 2.0 1.0 2.0 5.0 7
maximum 1690 6460 2770 3320 2430 2280
Variance is estimated on the basis of 60 samples. When more than one concentration was available due to duplicate sampling, the variance
was calculated using the mean concentration for a sample which was analyzed in duplicate.

fundamental error is always smaller than the experimental
sampling error. Therefore, neither the shape nor the
position of the experimental relationship supports the
relationship based upon the fundamental error.
In the first experiment, the sampling error seemed to
be more or less independent of the sample size for samples
larger than approximately 1g. For samples smaller than
1g, the experimental variance increased with decreasing
sample size. In the second experiment, with a much more
strongly segregated lot, there did not seem to be a direct
relationship between sample size and sampling error for
the whole range of investigated sample sizes (0.1-6000 g).
Therefore, it must be concluded that the fundamental
error alone is of no use for predicting the sampling error.

Envlron. Scl. Technol., Vol. 27, No. 10, 1993 2099

 - mxpmrinrntrl
+ rsoording Oy
1000
‘+\ i
‘t
0.1 1 10 ice ieee ieeee
rmmplm miz. (grrmr)
Figure 3. Coefficient of variation in relation to sample size. Plotted
are the experimentalresultsand the relationshipbetweenthe coefficient
of variation and the sample size based on the formula of the fundamen-
tal error given by Gy. Experimental results: [(experimental
variance)l/*/mean concentration]. According to fundamental error:
fJF .
The experimental results clearly showed that the
sampling error is dominated by the heterogeneity within
the lot. Only for very small samples is the fundamental
error a substantial part of the sampling error. The results
of the first experiment made clear that a minimum sample
size of approximately 10g must be used for the investigated
lot to obtain a good estimate of the heterogeneity within
the lot. Using smaller samples, the particulate hetsro-
geneity started to have an important effect on total
variance, and the sampling error increased due to the
fundamental error. No significant increase of the sampling
error was found in the sample size range of the second
experiment. Compared to the first experiment, the
heterogeneity of the second lot was much greater, even for
small samples, resulting in a relatively small contribution
of the fundamental error to the total sampling error.
Although Gy’s sampling theory (2) also introduced the
segregation error, which describes the variance caused by
heterogeneity within the lot, Gy himself gave no practical
method for calculating this sampling error. In fact, due
to the use of composite samples, Gy stated that the
segregationerror is normally smaller than the fundamental
error (ref 2, p 343). In environmental sampling and
specifically in quality control, the use of composite samples
is, however, limited, and mostly individual samples are
analyzed. Therefore, the segregation error is one of the
main sources of sampling error.
When the importance of the segregation error is rec-
ognized, it is important to know that Visman (3) did take
this sampling error into account. In order to apply eq 11,
it is necessary to know the variances which occur when
sampling and analyzing large and small samples. By
knowing these variances, it is possible to calculate the
values of A and B which are used in eq 11. Based on the
2040 Environ. Sci. Technol., Vol. 27, No. 10, 1993
- rxpmriinrntrl
+ ass. Uirmmn
8.1 1 ie $00 ieee iewe
m r m p l m r i z m (grmrnm)
Figure 4. Coefficient of variation In relation to sample size. Plotted
are the experimentalresultsand the relationship betweenthe coefficient
of variation and the sample size based on Visman’s theory. Exper-
imental results: [(experimental variance)”2/mean concentration],
According to Visman’s theory: ( A M , i-
1-and 6500-g samples, this resulted in A = 2.2035 X 10-8
and B = 5.1536 X (M, in grams). The relationship is
shown in Figure 4 . With these estimates, and the
assumption that f is constant, the product -yf in eq 10 can
be estimated. For 1-g samples the product equals 0.23,
and for 6500-g samples -yf = 1520. The last value is much
greater than l!
It must be stated that by using the 0.1-gsamples instead
of the 1-g samples the ‘Visman curve’ would, within the
investigated sample size range, resemble much more the
‘Gy curve’ which is shown in Figure 3. Only for samples
larger than approximately 1000 g would the relationship
flatten. We used the estimated variance of the 1-gsamples
because this estimate is more accurate than the estimated
variance of the 0.1-g samples (because the variance itself
is larger).
When comparing the sampling error in the first exper-
iment with the theoretical prediction according to Visman,
there seems to be a large amount of similarity between the
two. The more or less constant variance found for the
highly segregated second lot can be explained by the fact
that the segregation error is independent of the sample
size. Therefore, Visman’s formula could be useful in
predicting the minimum sample size to be used. However,
the practical applicability of this formula is limited because
the variance between samples must be known for two
sample sizes. Therefore, we used the basic ideas of Visman
to develop an alternative method to predict the sampling
error in relation to sample size and segregation patterns.
Application of Geostatistics. Although based on a
different concept, segregation can also be analyzed by using
geostatistics. Geostatistics is based on the assumption
that data are correlated in space and/or time. When this
correlation exists, it can be used to predict data at locations
 gamma
322
12000 *
10500 178 288 284
w s
216
90000 *
146
75000 * 118
*
60000

45000

30000
80
15000 * 32
*
0
I I I I I I 1 I I I I
0 0.500 1.000 1.500 2.000 2.500 3.000 3.500 4.000 4.500 5.000

d istance

mode 1twpe : spher ica 1

direction: total
cO= 8.39€+0004 e = 1.72€+0004 a = 1.07€+0000
S S D / S S T = 7.494

gamma
322

24000 284
*
*
20000

216
16000
*
146
&

12000

80000 /
/
118
40000 *
80 32
0
I I I I I I I I
* 1
* I

d istance
inputfile: cn2500s.rem
node 1 type : spher ica 1
direction: total
cO= 5.43€+0004 e = 1.55E+0005 a= 1.82€+0000
SSD/SST=469.416
~~~~ ~~

Flgure 5. Semlvarlogramsfor the first series of 0. Ig samples of the second experiment (a, top) and the 25004 samples of the second experiment
(b , bottom).

that have not been sampled (17). The existence of
segregation implies the existence of spatial correlation.
The fundamental error, as described in this paper, has no
relation with spatial or temporal correlation. So in
principle geostatistics is a useful method to investigate
whether the segregation error is predominant over the
fundamental error.
Spatial and/or temporal correlation can be measured
by means of a semivariogram. In a semivariogram, the
variance of differences between samples is plotted against
the distance (in time or space) between the sample
locations.
The semivariograms of the first experiment (measure-
mente in time) show no temporal correlation. Apparently,

Envlron. Scl. Technol., Vol. 27, No. 10, lQQ3 2041

 64.14

54.18 64.19

0:

54.21 64.2

SEGREGATION PLOTS

Coefficient of variation
Segregation factor : given in upper right-hand corner

Flgure 8. Segregation plots, N = 6561, n = 81, u y = 0.4.

the scale of segregation is much smaller than the scale on iments (approximately 10-15 kg each), the sampling
which the temporal correlation can be studied. This can interval (between 10 and 60 min), and the amount of soil
be explained by the amount of soil taken for the exper- passing through the soil cleaning plant in this period (20-

2042 Environ. Scl. Technol.. Vol. 27, No. 10, 1993

25 tons per hour).
For the second experiment, two semivariograms are
given in Figure 5a,b. Figure 5a is the semivariogram of
the first series of 0.1-g samples which were taken in the
laboratory from the 2500-g samples. Figure 5b shows the
semivariogram of the 2500-g samples. In both plots the
semivariogram values drop for distances larger than
approximately 2.5 m. This is caused by the fact that high
values are concentrated in a relative small area of the
investigation site, see Figure 2. Only Figure 5b does show
some correlation for data at distances less than 1.5 m.
Semivariograms of other sample sizes are comparable to
the ones given, showing no spatial correlation for the 0.1-g
samples and some correlation for larger sample sizes.These
results are in line with our conclusion that the segregation
error is of much more importance than the fundamental
error in sampling (contaminated) soil.
Segregation Plots
We used Visman’s sampling board to apply Gy’s theory.
Suppose one takes a sample of n neighboring locations
from a sampling board with N locations on which a fraction
of p pellets are placed at random. The relative variance
of the fraction locations filled with pellets ( & I ) in the
sample, disregarding the final population correction, is
or
1
P= (13)
(nUzv+ 1)
With the aid of this relationship, it is possible to visualize
the constitution heterogeneity (CHL). Suppose N = 6561
(812),n = 81, and UV =.0.4. Then it follows from eq 13 that
p = 0.0717. A picture of a sampling board with 6561
locations with a fraction of 0.0717 pellets randomly
distributed over the sampling board is shown in the top
left corner of Figure 6.
The heterogeneity of the picture is completely due to
the random distribution of the pellets. It cannot be
reduced. It can be enlarged, however, by distributing the
pellets not randomly over the board but in a segregated
pattern, as shown in the other pictures of Figure 6. The
pellets are concentrated around certain locations (hot
spots). Tne sampling board can be divided into 81
nonoverlapping squares of 81 positions. Then, by calcu-
lating the relative variance of these squares and using eq
10 with y = 81, one can estimate the segregation factor 4.
The estimates of 5 are printed at the top right corner of
each picture. The value of 5 is 0.18 in the third picture
of Figure 6. Since y = 81, it follows from eq 10 that UG
= 3.95UF in this case.
The numerical exercises as performed above can help
to get a feeling for the sampling error in relation to sample
size and degree of segregation. The idea is first to calculate
the relative variance of the fundamental error with eq 4;
to then simulate different kinds of segregation plots using
Visman’s sampling board, with the first plot representing
only constitution heterogeneity; and then to calculate the
variance in relation to sample size. An important point
of consideration is the choice of the sample size on the
sampling board (we used 81 in our example) and the size
of the sampling board itself (we used 6561). All these
values can be changed and will lead to different plots which
in fact visualize the lot and samples at different scales.
Appropriate scales should be chosen for specific appli-
cations.
We attempted to perform such an exercise for contam-
inated soil. The sampling constant C in eq 4 was estimated
from the first experiment: C = 13.5 g/cm3 cf = 0.5, g =
0.25, 1 = 0.03, c = 3600 g/cm3,aL = O.O005g/g). For samples
of 400 g, this resulted in a coefficient of variation for the
.
fundamentalerror of 13.1% Suppose that this is the ‘true’
value of UF in the second experiment. From the data of
experiment 2, a coefficient of variation of 268% has been
calculated.
The size of the experimental area in experiment 2 was
15 m2. A sample of 400 g took about 0.01 m2. In terms
of the sampling board, this gave a value of 15/0.01= 1500
for the ratio N / n . Substituting an arbitrary value of 0.1
for p in eq 13 gave n = 524 and consequently N = 786 000.
From eq 10 with y = 524, UF = 0.131, and UG = 2.68, it
follows that 4 = 0.80. Unfortunately such a large degree
of segregation can only be visualized by means of the
sampling board by simulating 78 600 pellets (10%)con-
centrated in a subarea of a size just over 10% of the size
of the totalarea (say 12-15 % ). Suchapicture willstrongly
deviate from a picture with 78 600 pellets distributed
completely at random over the complete area.
Exercising with data, as has been done above, does give
a rough idea of how much the pattern of contamination
in experiment 2 is due to segregation, compared with the
pattern that would arise if the distribution was purely
random and only heterogeneous because the material
under study consists of particles.
Conclusions
In the experimental part of this study, two lots of
contaminated soil have been sampled. The first lot was
sampled from a conveyor belt, while the second lot was a
highly segregated part of a former gasworks site.
From both experiments it was difficult to reach quan-
titative conclusions due to the uncertain assumptions
which had to be made. Nevertheless, we can still obviously
conclude that the distributional pattern of experiment 2
is highly due to segregation and that the influence of the
fundamental error can be ignored. This can also be
concluded for the first experiment in spite of the fact that
the sampled lot of the first experiment was much more
homogeneous.
Sample sizes for the analysis of cyanide in soil should
not be smaller than approximately 10 g to avoid a
significant contribution of the fundamental error on the
sampling error. By making segregation plots, it is possible
to get an idea of the segregation in a lot to be sampled.
When the estimated segregation of a lot is compared to
a segregation plot with the same kind of segregation, an
estimation of the segregation factor is available prior to
sampling. This can be used to estimate a minimum sample
size, which results in a negligible contribution of the
fundamental error on sampling error.
Acknowledgments
We performed part of this study for the Netherlands
Integrated Soil Research Programme and were financially
supported by the Netherlands Ministry of Housing,
Physical Planning, and the Environment. We thank
Envlron. Sci. Technol., Vol. 27. No. 10, 1993 2049
Afvalstoffen Terminal Moerdijk B.V. and t h e City of T h e
Hague for their cooperation in t h e sampling, a n d Mr. M.
M. Albert for carrying out t h e larger part of t h e practical
work.
Nomenclature
concentration of the critical component in particle
i (kg/kg)
concentration of the critical component in the lot
(kg/kg)
concentration of the critical component in the
sample (kg/kg)
sampling constant determined by random variance
sampling constant determined by segregation
variance
composition factor (g/cm3)
sampling constant (g/cm3)
constitution heterogeneity; the heterogeneity that
is inherent to the composition of each particle
making up the lot (dimensionless)
maximum diameter of particles (m)
liberation diameter (m)
distribution heterogeneity; the heterogeneity that
is inherent to the manner in which the individual
fragments are distributed over the lot (dimen-
sionless)
shape factor (dimensionless)
size distribution factor (dimensionless)
standardized deviation of particle i (76)
mean weight of the particles (kg)
liberation factor (dimensionless)
weight of particle i (kg)
weight of sample (kg)
number of neighboring locations in one sample
(dimensionless)
number of particles in the lot; number of locations
on a sampling board (dimensionless)
number of particles in a lot
number of subsamples in a lot
selection probability of a particle or subsample
(dimensionless)
grouping factor (dimensionless)
density of the critical component (g/cm3)
density of the matrix or gangue (g/cm3)
2044 Envlron. Scl. Technol., Vol. 27, No. 10, 1993
segregation factor (dimensionless)
relative variance of the fundamental error (% )
relative variance of the final sample (% )
variance due to analysis
variance due to sample pretreatment
variance due to sampling
total variance
relative variance of the fraction locations (% )
Literature Cited
(1) Bosman, R.; De Kwaadsteniet, J. W.; Lamb, F. P. J.;
Versluijs, C. W. A quality control method for decontam-
inated soil; Soil Protection Series, Part 79; Sdu Uitgenerij:
Den Haag, 1989; ISBN 90 120 6466 x (in Dutch).
(2) Gy, P. M. Sampling of particulate materials, Theory and
practice; Developmentsin Geomathematics 4; Amsterdam,
1982; ISBN 0 444 42079 7.
(3) Visman, J. Mater. Res. Stands. 1969, 9, 8-66.
(4) Duncan, A. J. Mater. Res. Stands. 1971, 11, 25.
(5) Visman, J.; Lerner, M. Mater. Res. Stands. 1971,11,32-37.
( 6 ) Stange, K. Chem.-Ing.-Tech. 1954,26,150-155 (inGerman).
(7) Stange, K. Chem-Ing.-Tech. 1954,26,331-337 (inGerman).
(8) Stange, K. Chem.-Zng.-Tech. 1963,35,580-582 (inGerman).
(9) Harnby, N.; Edwards, M. F.; Nienow, A. W. Mixing i n the
process industries; Butterworths Series in Chemical En-
gineering;Butterworths: London, 1985; ISBN 0 408 11574
2.
(10) Ingamells, C. 0.;Switzer, P. Talanta 1973, 20, 547-568.
(11) Ingarnells, C. 0. Talanta 1978,25,731-732.
(12) Pitard, F. F. Pierre Gy’s sampling theory and sampling
practice, volume Z Heterogeneity and sampling; CRC
Press: Boca Raton, 1989; ISBN 0 8493 6658 5.
(13) Pitard, F. F. Pierre Gy’s sampling theory and sampling
practice, volume IZ Sampling Correctness and sampling
practice; CRC Press: Boca Raton, 1989; ISBN 0 8493 6659
3.
(14) Minkkinen, P. Anal. Chim. Acta 1987, 196, 237-245.
(15) Meeussen,J. C. L.; Keizer,M. G.; de Haan, F. A.M. Environ.
Sci. Technol. 1992, 26, 511-516.
(16) Gower, J. C. Biometrics 1962, 18, 537.
(17) Journell, A. G.; Huijbregts, Ch. J. Mining geostatistics;
Academic Press: London, 1978.
Received for review September 15, 1992. Revised manuscript
received June 2, 1993.’
~~ ~~~~
0 Abstract published in Advance ACS Abstracts, August 15,1993.