Journal of Hydrology 540 (2016) 527–537

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

Detecting spatial structures in throughfall data: The effect of extent,

sample size, sampling design, and variogram estimation method
Sebastian Voss a, Beate Zimmermann b, Alexander Zimmermann a,⇑
a
University of Potsdam, Institute of Earth and Environmental Science, Potsdam, Germany
b
Research Institute for Post-Mining Landscapes, Finsterwalde, Germany

a r t i c l e i n f o a b s t r a c t

Article history: In the last decades, an increasing number of studies analyzed spatial patterns in throughfall by means of
Received 5 October 2015 variograms. The estimation of the variogram from sample data requires an appropriate sampling scheme:
Received in revised form 10 May 2016 most importantly, a large sample and a layout of sampling locations that often has to serve both vari-
Accepted 19 June 2016
ogram estimation and geostatistical prediction. While some recommendations on these aspects exist,
Available online 21 June 2016
This manuscript was handled by Andras
they focus on Gaussian data and high ratios of the variogram range to the extent of the study area.
Bardossy, Editor-in-Chief, with the However, many hydrological data, and throughfall data in particular, do not follow a Gaussian distribu-
assistance of Uwe Haberlandt, Associate tion. In this study, we examined the effect of extent, sample size, sampling design, and calculation
Editor method on variogram estimation of throughfall data. For our investigation, we first generated non-
Gaussian random fields based on throughfall data with large outliers. Subsequently, we sampled the
Keywords: fields with three extents (plots with edge lengths of 25 m, 50 m, and 100 m), four common sampling
Throughfall designs (two grid-based layouts, transect and random sampling) and five sample sizes (50, 100, 150,
Geostatistics 200, 400). We then estimated the variogram parameters by method-of-moments (non-robust and robust
Sampling estimators) and residual maximum likelihood. Our key findings are threefold. First, the choice of the
Variogram extent has a substantial influence on the estimation of the variogram. A comparatively small ratio of
Residual maximum likelihood
the extent to the correlation length is beneficial for variogram estimation. Second, a combination of a
minimum sample size of 150, a design that ensures the sampling of small distances and variogram esti-
mation by residual maximum likelihood offers a good compromise between accuracy and efficiency.
Third, studies relying on method-of-moments based variogram estimation may have to employ at least
200 sampling points for reliable variogram estimates. These suggested sample sizes exceed the number
recommended by studies dealing with Gaussian data by up to 100 %. Given that most previous through-
fall studies relied on method-of-moments variogram estimation and sample sizes
200, currently avail-
able data are prone to large uncertainties.
Ó 2016 Elsevier B.V. All rights reserved.

1. Introduction Zimmermann and Zimmermann, 2014; Zimmermann et al.,

In the last three decades, an increasing number of studies ana-
lyzed spatial patterns in throughfall to investigate the conse-
quences of rainfall redistribution for biogeochemical (Allen et al.,
2015; Hsueh et al., 2016; Möttönen et al., 1999; Whelan et al.,
1998) and hydrological processes in forests (Fathizadeh et al.,
2014; Gerrits et al., 2010; Hsueh et al., 2016; Keim et al., 2005;
Klos et al., 2014; Loescher et al., 2002; Shachnovich et al., 2008;
Staelens et al., 2006; Zimmermann et al., 2009). Other studies ana-
lyzed throughfall spatial patterns to optimize sampling schemes
for estimating mean throughfall (Ziegler et al., 2009;
⇑ Corresponding author.
E-mail address: alexander.zimmermann.ii@uni-potsdam.de (A. Zimmermann).
2010). In the majority of cases, variograms were used to character-
ize the spatial properties of the throughfall data.
The variogram is central to geostatistics (Webster and Oliver,
1992, 2007) because it describes spatial variation and provides
the parameters (nugget, sill and range) that are essential for spatial
prediction and the simulation of random fields. It is widely
accepted that estimates of the variogram are sensitive to the size
and spatial arrangement of the sample (e.g. Lark, 2002a; Russo
and Jury, 1987; Webster and Oliver, 1992). Furthermore, there is
ample evidence that the variogram range and sill depend on the
spatial scale of sampling (e.g. Blöschl, 1999; Western and Blöschl,
1999). Several studies investigated the influence of various aspects
of the sampling design on variogram estimation (Bogaert and
Russo, 1999; Blöschl, 1999; Corsten and Stein, 1994; Kerry et al.,

http://dx.doi.org/10.1016/j.jhydrol.2016.06.042
0022-1694/Ó 2016 Elsevier B.V. All rights reserved.
528 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537

2008; Lark, 2002a; Morris, 1991; Müller and Zimmerman, 1999; this knowledge gap, we sampled a set of unconditional simula-
Pardo-Igúzquiza and Dowd, 2013; Pettitt and McBratney, 1993; tions, which we obtained using real-world throughfall data, with
Russo and Jury, 1987; Skøien and Blöschl, 2006; Warrick and several extents, common spatial sampling designs, and a variety
Myers, 1987; Webster and Oliver, 1992; Western and Blöschl, of sample sizes. We then evaluated these sampling schemes in
1999). This work, however, has received little attention among for- terms of their ability to provide satisfactory estimates of the vari-
est hydrologists partly because of a missing common language ogram parameters. For our analysis we used both REML and MoM
between environmental statisticians and field hydrologists. variogram estimation.
A closer look at the studies that investigated the role of sam-
pling designs on variogram estimation reveals that they can be
2. Methods
divided into three groups. The first group (Bogaert and Russo,
1999; Lark, 2002a; Morris, 1991; Müller and Zimmerman, 1999;
2.1. Data
Pettitt and McBratney, 1993; Warrick and Myers, 1987) optimized
sampling designs based on various criteria linked to the variogram.
From a large throughfall data set (Zimmermann and
Early studies (Morris, 1991; Warrick and Myers, 1987) focused on
Zimmermann, 2014), we selected six events that showed distinct
the distribution of sampling points among the lags. For instance,
univariate distributions and autocorrelation structures, respec-
Warrick and Myers (1987) presented a criterion that aims on an
tively. While all events included outlying values (i.e. data points
equally distributed number of paired comparisons in each lag class.
which cannot be forced to the center of the distribution even after
Morris (1991) criticized this approach because it neglects the cor-
transformation), events 1 and 5 furthermore showed an underlying
relation of the spatial data and leads to a comparatively low effi-
asymmetry (cf. Kerry and Oliver, 2007). This type of asymmetry is
ciency of the sampling design (van Groenigen, 1999). Subsequent
not caused by outliers but by a skew of the underlying (or primary)
studies chose other, more complex criteria. For instance, Müller
distribution of the data, which can be statistically defined as the
and Zimmerman (1999) maximized the determinant of Fisher’s
region between the first and seventh octile (Zimmermann et al.,
information matrix and Lark (2002a) minimized the kriging vari-
2010). It is important to distinguish between underlying asymme-
ance to find an optimum configuration of sampling points. In his
try and skewness due to outliers because these deviations from the
comprehensive study, Lark (2002a) demonstrated that (i) a random
normal distribution require different treatments of the data.
process with a small spatial dependence is sampled best with scat-
Robust variogram estimators can deal with normally distributed
tered clusters of sampling points, (ii) for long range processes a
data that are contaminated with outliers. Robust estimators, how-
regular array is optimal, and (iii) if there is no prior information
ever, cannot deal with data that show an underlying skew because
about the spatial correlation, sampling in transects is the most
the estimators have a specific consistency correction for contami-
robust approach.
nated normal data (Lark, 2000a). We therefore had to transform
The second group of studies (Corsten and Stein, 1994; Kerry
data of events 1 and 5 before further processing.
et al., 2008; Pardo-Igúzquiza and Dowd, 2013; Webster and
For each event we constructed a Gaussian random field by
Oliver, 1992) sampled simulated random fields to assess the effect
unconditional simulation. The simulated values of fields 1 and 5
of different sampling designs and sample sizes on variogram esti-
were back transformed to ensure that the fields reflected the struc-
mation. This simple approach has the advantage that the experi-
ture of the original data. In a final step we contaminated the fields
mental variogram can be compared directly against the
with outlying values of the respective event (Fig. 1). For an in-
variogram which is based on all data of the simulated field.
depth description of the construction of the fields we refer to
Webster and Oliver (1992) sampled different random fields and
Zimmermann et al. (2010) and Zimmermann and Zimmermann
concluded that a sample size of 150 would be satisfactory for a pre-
(2014). The fields have an edge length of 100 m, a grid unit of
cise estimate of the variogram. Kerry et al. (2008) followed the
0.1 m and hence consist of 106 data points.
approach of Webster and Oliver (1992) and compared residual
A closer look at the simulated fields (Fig. 1, Table 1) reveals that
maximum likelihood (REML) with method of moment (MoM)
they comprise a large span of spatial structures which is reflected
based variogram estimation. They found that REML outperforms
in the variation of the nugget-to-sill ratio and the effective range,
MoM and that a sample size of 100 would be sufficient for vari-
respectively. Relatively strong spatial structures and short autocor-
ogram parameter estimation.
relation distances characterize fields 3, 4 and 5. Accordingly, these
The third group of studies (Blöschl, 1999; Skøien and Blöschl,
fields have nugget-to-sill ratios <25% and effective ranges of
2006; Western and Blöschl, 1999) investigated effects of scale on
around 3 m. In contrast, fields 1 and 2 feature somewhat weaker
variogram estimation. Although these studies did not primarily
spatial structures but a comparatively long autocorrelation dis-
focus on the influence of the sampling design on variogram estima-
tance. Finally, field 6 features a pure nugget structure and hence
tion, their work has important implications for sampling. For
displays no spatial correlation.
instance, Western and Blöschl (1999) showed that estimates of
the correlation length depend on the extent (i.e. on the size of
the sampling plot). 2.2. Sampling methods
Most of the studies that assessed the impact of the sampling
design on variogram estimation worked with normally distributed For our study we tested the influence of the extent, sampling
data. Furthermore, the data of previous studies often showed a design, sample size, and methodology on the estimation of the var-
comparatively strong autocorrelation and a long range. Throughfall iogram. In this section, we describe the selection of the extent,
data, however, usually do not follow a normal distribution; instead sampling design and sample size. In Sections 2.3 and 2.4 we deal
they often show skewed underlying distributions (Zimmermann with the variogram estimation methods.
and Zimmermann, 2014) and heavily outlying values (Lloyd and To assess the influence of the extent, we employed three plot
Marques, 1988; Zimmermann et al., 2009). Moreover, variograms sizes with edge lengths of 25 m, 50 m, and 100 m, respectively.
of throughfall data usually exhibit relatively small ranges com- The largest plots comprised the entire simulation fields. For the
pared to the extent of the research area (Möttönen et al., 1999; smaller plots we arbitrarily chose the lower left corner of the sim-
Zimmermann and Zimmermann, 2014; Zimmermann et al., ulated fields.
2009). Therefore, it is not clear if the results of previous studies For the analysis of the sampling design we tested random sam-
apply for the spatial analysis of data with these properties. To fill pling (R) in addition to three regular designs: a regular grid (G), a
 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537 529

field 1 field 2 field 3

100 [mm] [mm] [mm]

200
2.5

40
80

2.0

150

30
60

1.5
Y [m]

100

20
40

1.0

50

10
0.5
20

0.0

0
0
field 4 field 5 field 6
100 [mm] [mm] [mm]

40
80 150

30
60
100

6
Y [m]

20
40

4
50

10
2
20
0

0
0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
X [m] X [m] X [m]

Fig. 1. The six random fields used for re-sampling.

regular grid with additional observations at short distances (Gs),
and transects (T). Regular designs have been used frequently for
geostatistical analyses. Previous studies on throughfall spatial pat-
terns, however, also used random sampling (e.g. Bellot and Escarre,
1998; Hsueh et al., 2016; Keim et al., 2005), which is why we
tested this design too.
We combined the four sampling designs with five different
sample sizes (50, 100, 150, 200 and 400) to evaluate the influence
of the sample size on variogram estimation. Because of the regular
sampling pattern of grids and transects it was not always
possible to obtain the envisaged sample size. In these cases we
used the closest possible sample size (e.g. grid-based sampling:
grid of 7 by 7 = 49 points instead of n = 50, Supplementary
material A).
The sampling points were allocated as follows. For the regular
designs, we divided the simulated fields into square subplots. Their
size depended on the extent, sample size, and sampling design
(Fig. 2). The sampling points of designs G and Gs were randomly
arranged in the subplots, but the grid shape was maintained in
the whole plot area. For design Gs, the number of sampling points
in the grid was reduced, leading to a larger spacing and larger sub-
plots. Every second sampling point in the design Gs got an addi-
tional sampling point in 1 m distance (or 0.5 m distance for an
extent of 25 m and a sample size of 400). For design (T), we ran-
domly chose the position and orientation (horizontal or vertical)
of the transects. One transect consisted of six sampling points.
The distances of the first five sampling points in a transect were
fixed (Supplementary material A) and rose exponentially as recom-
mended by Pettitt and McBratney (1993). The distance to the last
sampling point depended on the size of the subplots and on the
sample size. Each of the six random fields (Fig. 1) was sampled
with all combinations of extents, sampling designs, and sample
sizes 1000 times. The obtained samples were used for variogram
estimation.
2.3. Variogram estimation by the method-of-moments (MoM)
The estimation of the variogram by the method-of-moments
(MoM) is widely used in geostatistics (de Gruijter et al., 2006, p.
173). With MoM the variogram parameters are estimated in two
steps. These consist of calculating an experimental variogram
and subsequently fitting a variogram model to it. Because of the
outliers in the simulated fields, we calculated experimental vari-
ograms using non-robust and robust estimators: the estimator of
Matheron (1962), Cressie and Hawkins (1980) and Dowd (1984).
Given normal data, Matheron’s estimator is the most efficient,
while the others can deal with outliers. Because the samples of
fields 1 and 5 showed underlying asymmetry (Section 2.1), we
transformed the data to square roots before variogram calculation.
Following Journel and Huijbregts (1978, p. 194), we calculated the
experimental variograms to a maximum lag distance of half the
diagonal of the plots. For all experimental variograms we chose
1 m lag distances.
In a second step, we fitted three models to the experimental
variogram that are most appropriate for throughfall data (expo-
nential, spherical, pure nugget; Zimmermann and Zimmermann,
2014). For variogram fitting we applied the R-function optim (R
Core Team, 2013) with the method ‘‘L-BFGS-B”. Subsequently, we
chose the model with the smallest sum of the weighted least
squares (Cressie, 1985). Lags with less than 5 pairs of observations
were not used for model fitting.
We performed this procedure for the three MoM variogram
estimators, which resulted in three different sets of variogram
parameters (variogram model, nugget, sill and range). To decide
which set of variogram parameters describes the sampled data
best, we predicted the throughfall volume at each location xi by
ordinary kriging using all data except the value at location xi (cross
validation). We then calculated the statistic hðxÞ for every point xi
(Lark, 2000a):
530 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537

Table 1
The spatial properties of the simulation parameters used for the construction of the 6 non-Gaussian random fields and the spatial properties based on the exhaustive variograms
of the constructed fields (plot with an edge length of 100 m) and the subfields (plots with 25 m and 50 m edge lengths).

Field Method Edge length Nugget Sill Range [m] Effective range [m] Nugget:sill ratio [%] Variogram model
1 Simulation NA 0.02 0.06 2.84 8.53 24.80 Exponential
MoM 25 0.02 0.06 7.10 21.30 33.10 Exponential
50 0.02 0.06 3.05 9.14 35.06 Exponential
100 0.02 0.05 2.60 7.79 42.65 Exponential
REML 25 0.01 0.05 3.07 9.21 26.67 Exponential
50 0.01 0.05 3.13 9.40 27.03 Exponential
100 0.01 0.04 2.44 7.32 31.06 Exponential
2 Simulation NA 386.59 499.46 2.65 7.94 77.40 Exponential
MoM 25 437.97 551.07 1.88 5.65 79.48 Exponential
50 449.00 545.09 2.02 6.07 82.37 Exponential
100 444.77 545.42 2.52 7.57 81.55 Exponential
REML 25 380.42 482.79 1.93 5.78 78.80 Exponential
50 380.64 485.21 2.18 6.53 78.45 Exponential
100 382.29 480.69 2.52 7.55 79.53 Exponential
3 Simulation NA 2.69 11.30 3.39 3.39 23.77 Spherical
MoM 25 3.11 12.87 3.20 3.20 24.18 Spherical
50 2.73 13.09 3.08 3.08 20.82 Spherical
100 2.49 13.43 3.34 3.34 18.52 Spherical
REML 25 2.63 11.59 3.31 3.31 22.72 Spherical
50 2.48 11.30 3.14 3.14 21.93 Spherical
100 2.24 11.76 3.33 3.33 19.00 Spherical
4 Simulation NA 14.13 122.04 1.02 3.07 11.58 Exponential
MoM 25 30.90 133.32 0.94 2.82 23.18 Exponential
50 20.05 134.07 1.01 3.02 14.95 Exponential
100 31.51 136.71 1.00 3.00 23.05 Exponential
REML 25 12.30 114.01 0.86 2.59 10.79 Exponential
50 13.18 119.40 0.95 2.86 11.04 Exponential
100 14.01 123.04 0.99 2.97 11.38 Exponential
5 Simulation NA 0.02 0.21 0.99 2.96 7.82 Exponential
MoM 25 0.04 0.23 1.46 4.39 18.26 Exponential
50 0.05 0.21 1.14 3.42 22.29 Exponential
100 0.06 0.21 1.05 3.16 29.64 Exponential
REML 25 0.02 0.19 1.10 3.29 9.24 Exponential
50 0.02 0.19 1.07 3.20 10.49 Exponential
100 0.02 0.19 1.00 3.00 11.92 Exponential
6 Simulation NA 5.14 5.14 0.00 0.00 100.00 Nugget
MoM 25 5.87 5.87 0.00 0.00 100.00 Nugget
50 5.83 5.83 0.00 0.00 100.00 Nugget
100 5.72 5.72 0.00 0.00 100.00 Nugget
REML 25 5.13 5.13 0.00 0.00 100.00 Nugget
50 4.98 4.98 0.00 0.00 100.00 Nugget
100 5.00 5.00 0.00 0.00 100.00 Nugget

^ i Þ  zðxi Þg
fZðx
2
of the secondary process (outliers originating from drip points). In
hðxi Þ ¼ ð1Þ other cases, outlier removal will not be possible and hence, MoM
r 2
K;i
estimation of the variogram using robust estimators is the only
choice.
^ i Þ and the kriging
with the observed value zðxi Þ, the kriged value Zðx To identify the spatial outliers at locations xi in our throughfall
variance rK;i . A correct set of variogram parameters has a median
2
datasets, we used the standardized error of cross-validation, eS ðxÞ
hðxÞ of 0.455 (Lark, 2000a). Accordingly, we chose the MoM vari- (Bárdossy and Kundzewicz, 1990; Lark, 2002b) and the cross-
ogram estimator with hðxÞ closest to 0.455. validation result of the most accurate MoM estimator:

2.4. Variogram estimation by residual maximum likelihood (REML)
Maximum likelihood based estimation methods are state of the
art in variogram estimation (Kerry et al., 2008; Lark et al., 2006). In
this study, we used the residual maximum likelihood (REML) esti-
mation method (Patterson and Thompson, 1971) as described in
Diggle and Ribeiro (2007, p. 116). In contrast to MoM, REML esti-
mates the variogram parameters directly from the sample data
and not as a fit to the experimental variogram. Since REML is sen-
sitive to outliers, we removed all data points identified as spatial
outliers before variogram estimation (cf. Zimmermann et al.,
2010). This approach is sensible for throughfall data because in
most situations we wish to model the spatial variation of the pri-
mary process (throughfall data without outliers) not the variation
^ i Þ  zðxi Þ
Zðx
eS ðxi Þ ¼ ð2Þ
rK;i
^ i Þ, the observed value zðxi Þ and the kriging
with the kriged value Zðx
standard error rK;i . According to Bárdossy and Kundzewicz (1990),
it is plausible to assume that eS ðxÞ follows a standard normal distri-
bution. Consequently, we can use the z statistic to classify a value as
an outlier. In the present study, points in the sampled dataset with a
standardized error of cross-validation smaller than 2.576


a ¼ 0:005 were removed as spatial outliers before further calcula-
2
tions. Hence, with this statistical procedure we removed unusually
large values from our datasets.
After outlier removal, we applied REML to fit the same theoret-
ical models (exponential, spherical, pure nugget) as we did when
 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537 531

G Gs 2.6. The exhaustive variogram as comparison

100
In order to assess the accuracy of the variogram estimation, we
80
calculated exhaustive variograms using all data of the respective
60 extent and field (Table 1). It was technically not feasible to calcu-
Y [m]

late a variogram model with all data points. Hence, both for
40 MoM and REML, we estimated variogram models as described
20
above using 1000 randomly chosen data points and made sure that
no data point was used more than once. This procedure allowed
0 the estimation of 1000, 250 and 62 variogram models for the
extents with a 100 m, 50 m and 25 m edge length, respectively.
T R
100 By averaging the parameters of the respective variograms, we
received one exhaustive variogram for the three extents of every
80 field. The exhaustive variogram represents an approximation of
60 the real variogram for the respective extent and simulated field.
Y [m]

To assess the accuracy of the variogram estimation with varying

40 sample sizes and sampling designs, we compared the results of the
estimated variogram parameters with the parameter values of the
20
exhaustive variograms. The individual steps of this procedure are
0 summarized in Fig. 3. For comparability of variogram parameters
0 20 40 60 80 100 0 20 40 60 80 100 and estimation methods (MoM and REML), we defined a span for
X [m] X [m]

Fig. 2. The four sampling designs: grid (G), regular grid with additional observa-
tions at short distances (Gs), transects (T) and random (R). Plots show a sample size
of 100. The gray lines divide the plots into subplots, additional observations of
Unconditional simulation of a Gaussian random variable
parameter: mean, sill, nugget, range, variogram model
design Gs are marked in gray.

using MoM estimation, and chose the one with the highest maxi- 6 simulated throughfall fields
mized log-likelihood.
Installation of outliers

2.5. Automatic methods for the estimation of input parameters 6 contaminated throughfall fields

up to here from Zimmermann & Zimmermann (2014)

The two variogram estimation methods described above require
initial estimates of the variogram parameters nugget, sill, and
range, which are based on the experimental variogram. The calcu- Sampling:
3 extents, 4 sampling designs, 5 sample sizes, 1000 iterations
lation of an experimental variogram is indispensable for MoM but
not required for REML. However, for finding initial estimates for
REML, too, we calculated the Matheron’s experimental variogram

Residual Maximum Likelihood (REML)

after outlier removal. Remove spatial outliers
Calculation of the
The initial estimates are usually obtained by visual inspection of with < -2.576
Method-of-Moments (MoM)

experimental variogram
the experimental variogram (Diggle and Ribeiro, 2007, p. 107). For with 3 estimators
this study, however, visual inspection was not feasible because of (Cressie and Hawkins,
1980; Dowd, 1984; Calculation of the
the vast number of experimental variograms that had to be ana-
Matheron, 1962) experimental variogram
lyzed. Consequently, we had to use automatic methods for the (Matheron, 1962)
derivation of input parameters.
Two automatic methods for the estimation of input parameters Automatic estimation of initial parameters
based on the experimental variogram were found in the literature
and used in this study. The first method (Jian et al., 1996) calcu-
lates the initial sill as the mean of the semivariance values of the Estimation of the
Estimation of the
variogram model by
last three lags. The initial nugget is defined as the intercept of a lin- variogram model by
REML (Patterson and
MoM (Cressie, 1985)
ear model with the semivariance values of the first two lags, and Thompson, 1971)
the initial range is defined as half of the mean distance of the
research area. The second method (Hiemstra et al., 2009) calculates
the initial sill as the mean of the maximum and median semivari- Calculate statistic
ance, the initial nugget as the minimum semivariance and the ini- and choose estimator
with median [ ]
tial range as 0.10 times the diagonal of the bounding box of the closest to 0.455
data (this method is implemented in their R-package automap).
We varied the initial parameters of the two methods by ±10%,
which produced six parameter sets for MoM and REML-based var- Comparison of the variogram parameters
iogram estimation, respectively. With these sets of initial parame- with the parameters of the simulated fields
ters we calculated three different variogram models (spherical,
exponential, nugget) using both MoM and REML. For each vari-
ogram estimation method we finally chose the model with the best Fig. 3. Summary of steps involved in estimating the variogram parameters based
fit, as described in Sections 2.3 (MoM) and 2.4 (REML). on sampling of the simulated throughfall fields.
532 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537

each field and parameter, within which we deemed parameter val-
ues to be accurately estimated. The nugget and sill of an exhaustive
variogram ±25% of the sill value defined the tolerable span for the
semivariance-axis. The range of the exhaustive variogram ±25%
defined the tolerable span for the distance-axis.
All calculations were processed with the free software environ-
ment R (R Core Team, 2013). Particularly, we used the packages
automap (Hiemstra et al., 2009), geoR (Diggle and Ribeiro, 2007),
and georob (Papritz, 2014).
3. Results and discussion
3.1. The role of the extent
Our results indicate that the choice of the extent has a substan-
tial influence on the estimation of the experimental variogram
(Fig. 4, Supplementary material B). Particularly for sample sizes
of 150 and less, semivariance estimates from the 25 m by 25 m plot
deviate much less from the exhaustive variogram than the esti-
mates from the larger plots. As expected, the choice of the extent
also influences the estimation of the variogram parameters, albeit
to a varying degree (Fig. 5, Supplementary material C). Most esti-
mates of the sill display only minor differences between the three
extents. Estimates of nugget and range, in contrast, clearly improve
with a decreasing extent which is reflected in an increasing num-
ber of estimates within the tolerable span (Fig. 5; for a definition
of the tolerable span see Section 2.6). Consequently, our results
seem to suggest that a comparatively small ratio of the plot extent
to the range of the data, and hence a relatively small spacing
between sampling points, facilitates the estimation of variograms
with a short range (cf. Blöschl, 1999; Skøien and Blöschl, 2006;
Western and Blöschl, 1999).

25 50 100
n = 50 n = 50 n = 50
100
semivariance
60
0 20

n = 100 n = 100 n = 100

100
semivariance
60
0 20

n = 150 n = 150 n = 150

100
semivariance
60
0 20

n = 200 n = 200 n = 200

100
semivariance
60
0 20

n = 400 n = 400 n = 400

100
semivariance

field 3
G Gs T R
60
0 20

0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
distance [m] distance [m] distance [m]

Fig. 4. Comparison between the 5th and 95th percentiles of the experimental variograms from field 3 sampled with three extents (plot edge lengths of 25 m, 50 m, and
100 m), four sampling designs (G, Gs, T, R), and various sample sizes (50, 100, 150, 200, 400). The exhaustive variograms are plotted as solid lines. For results of all fields we
refer to Supplementary material B.
 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537 533

25 50 100
100

80

60
sill

40
percentage of variogram parameters within the tolerable span [%]

20

100

80

60
nugget

40

20

100
G REML
Gs MoM
T
80
R

60
range

40

20

50 100 150 200 400 50 100 150 200 400 50 100 150 200 400
sample size sample size sample size

Fig. 5. The percentage of variogram parameters within the tolerable span as a function of extent (plot edge lengths of 25 m, 50 m, and 100 m), sample size (50, 100, 150, 200,
400), sampling design (G, Gs, T, R), and estimation method (MoM, REML). The plots show averages of the 6 random fields. For results of the individual fields we refer to
Supplementary material C.

In practice, however, it is not straightforward to choose an opti- typical correlation lengths between three and eight meters
mum extent because the covariance structure of throughfall at a (Table 1), a sampling area of 25 m by 25 m is already too small.
particular site is usually unknown. Furthermore, throughfall auto-
correlation structures vary with rainfall magnitude, i.e. small 3.2. The role of the sample size
events often display a larger range than large events
(Zimmermann and Zimmermann, 2014). Lastly, and probably most The sample size has a considerable influence on the estimation
important, the extent itself influences estimates of the range of both the experimental variogram (Fig. 4, Supplementary mate-
(Blöschl, 1999; Skøien and Blöschl, 2006; Western and Blöschl, rial B) and the variogram parameters (Fig. 5, Supplementary mate-
1999). Previous studies on soil moisture and snow cover patterns rial C). Generally, with rising sample size the spread of the
(Blöschl, 1999; Western and Blöschl, 1999) indicate that the semivariance estimates narrows towards the values of the exhaus-
observed autocorrelation distance decreases with a decreasing tive variogram. Furthermore, the semivariance estimates become
extent. Our simulations do not always show this pattern (Table 1). less erratic (Fig. 4, Supplementary material B). An effect, which
However, the ranges of the exhaustive variograms deviated most was already documented in previous studies (Webster and
from the range of the variogram that was used to build the simu- Oliver, 1992). However, in contrast to previous studies (Gascuel-
lation if the calculations were based on the smallest extent (25 m Odoux and Boivin, 1994; Kerry et al., 2008; Webster and Oliver,
by 25 m plot, Table 1). We therefore suspect that for our data with 1992), our results indicate that a sample size of 200 is not
534 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537

25 50 100
G G G
number of pairs
600
200
0

Gs Gs Gs
number of pairs
600
200
0

T T T
number of pairs
600
200
0

R R R
number of pairs
600
200
0

0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30

distance [m] distance [m] distance [m]

Fig. 6. The distribution of the number of point pairs per lag class for a sample size of 150. The plot shows results for four sampling designs (G, Gs, T, R) and three extents (plot
edge lengths of 25 m, 50 m, and 100 m).

necessarily the upper end. In fact, if the ratio of the extent to the
range is relatively large, 200 and more sampling points seem to
be necessary to avoid erratic semivariance estimates of the exper-
imental variogram (Fig. 4; Supplementary material B).
A rising sample size also goes along with a higher accuracy of
the variogram parameter estimates (Fig. 5, Supplementary mate-
rial C). Particularly for small samples (e.g. n = 50), an increase in
sample size is reflected in a large gain in accuracy, whereas
increasing the sample size of an already large sample (n = 200)
often results in a comparatively small increase in accuracy. This
result is in accordance with previous studies (Gascuel-Odoux and
Boivin, 1994; Kerry et al., 2008), although the point where the ben-
efit of an increase in sample size levels off seems to be case-
specific. For our throughfall data this point is reached only at a
large sample size of P200 (Fig. 5).
3.3. The role of the sampling design
The performance of the tested sampling designs strongly
depends on the extent, sample size, and variogram estimation
method (Fig. 4, Supplementary material B; Fig. 5, Supplementary
material C). Therefore, an assessment of a particular design always
requires to consider the entire sampling setup. For instance,
transect-based sampling (T) and random sampling (R) produce
particularly erratic semivariance estimates for plots with an edge
length P50 m and sample sizes 6100 (Fig. 4, Supplementary mate-
rial B). This pattern emerges because certain lag classes contain too
few data. For the same setup, grid sampling (G) and optimized grid
sampling (Gs) cannot capture the spatial structure of the simulated
fields either due to missing semivariance estimates at short lags.
Considering all combinations of extents and sample sizes (Fig. 4,
Supplementary material B) provides further evidence that simple
grid sampling (G) is inferior to all other designs. Again, this is
because of missing data at short lags (Fig. 6).
A closer look at the variogram parameter estimates (Fig. 5, Sup-
plementary material B) allows to further assess the sampling
designs. Considering a sample size P150, designs G, Gs and R esti-
mated the sill slightly better than the design T. The nugget and the
range, however, were estimated best with design T. The superior
performance of design T is related to the sampling of several short
lag distances, which allowed capturing the small scale variation
typical for throughfall data (cf. Keim et al., 2005; Möttönen et al.,
1999; Zimmermann and Zimmermann, 2014). Somewhat surpris-
ing, random sampling (R) often estimated the nugget and sill as
good as or even better than design Gs. This is because few data
pairs in multiple short lag distances (design R, Fig. 6) seem to be
better than comparatively many pairs in just one short distance
(design Gs, Fig. 6). In practice, however, simple random sampling
has the disadvantage that the distribution of points among lag
classes cannot be controlled.
Our results (Fig. 5, Supplementary material C) and data of
previous studies (Corsten and Stein, 1994) indicate that the
 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537 535

Table 2
Comparison of sampling designs and throughfall autocorrelation ranges from studies conducted in a variety of forest ecosystems.

References Forest type, location Extenta Nb Supportc Min. Sampling design Effective
[m2] [cm2] distance [m] ranged [m]
Bellot and Escarre (1998) Holm-oak fores, NE Spain 950 50 ? ? Random Pure nugget
Fathizadeh et al. (2014) Persian oak trees, W Iran 60 16 64 0.5 Radial 5–6
Gerrits et al. (2010) 120-year-old beech forest, 596 81 400 2 Grid 6–7
Luxembourg
Gómez et al. (2002) Olive trees, Spain 36 36 113 0.5 Grid 1.5–4.5
Hsueh et al. (2016) 42-year-old mixed-hardwood forest 484 100 145 ? Random 2–6
Keim et al. (2005) 60-year-old conifer forest, W USA 225 94 9 ? Random 5e
Keim et al. (2005) Old conifer forest, W USA 900 94 9 ? Random 5e
Keim et al. (2005) 60-year-old deciduous forest, W USA 304 94 9 ? Random 10e
Loescher et al. (2002) Tropical rain forest, Costa Rica 13,273 56 95 5 Radial and short distances 43
Loustau et al. (1992) 18-year-old pine stand, S France 2500 52 707 ? Stratified random Pure nugget
Möttönen et al. (1999) 150 to 200-year-old Scots pine 10,000 181 50 1.25 Grids with different 9
forest, E Finland distances
Ritter and Regalado (2010) Evergreen tree heath-laurel forest, ? 22 214 1 Transect Pure nugget
Canary islands
Staelens et al. (2006) 85-year-old beech tree, Belgium 180 48 460 0.75 Grid and transect 3–4
Staelens et al. (2006) 85-year-old beech tree, Belgium 180 50 158 0.25 Transect No stable sill
Ziegler et al. (2009) Montane forest, Thailand 500 20 540 2 Stratified random sill not
reached
Zimmermann et al. (2009) Tropical rain forest, Panama 10,000 220 113 1 Stratified random and short Mostly pure
distances nuggetf
Zimmermann and Teak plantation 10,000 350 113 0.1 Stratified random and short 2.6
Zimmermann (2014) distances
Zimmermann and Young secondary forest 10,000 350 113 0.1 Stratified random and short 5.3
Zimmermann (2014) distances
Zimmermann and Old secondary forest 10,000 350 113 0.1 Stratified random and short 3.9
Zimmermann (2014) distances
a
Size of sampling area.
b
Sample size.
c
Receiving area of collector.
d
The effective range Reff is calculated as exponential model: Reff = range ⁄ 3. Spherical model: Reff = range.
e
Keim et al. (2005) estimated the range visually from the experimental variogram, no variogram model was fitted.
f
Zimmermann et al. (2009) detected a pure nugget structure in 82% of the analyzed events, the remaining events showed weak structures (i.e. high nugget/sill ratios) with
highly variable effective ranges.

performance of sampling designs strongly depends on the ratio of
the range to the extent of the study area. For data with a compar-
atively large range, random-, grid- and transect-based sampling
designs differ little in their performance (Corsten and Stein,
1994). In contrast, the analysis of short-ranging autocorrelation
structures may greatly benefit from sampling approaches that
include multiple small distances (cf. Fig. 5, Supplementary material
C). Because sampling with transects performs best for a range of
extents and sample sizes (Fig. 5, Supplementary material C), this
design is particularly advantageous when there is no prior knowl-
edge on autocorrelation structures (cf. Lark, 2002a).
3.4. The role of the estimation method
Of the estimation methods, REML proved superior to MoM
(Fig. 5, Supplementary material C). For instance, MoM-based esti-
mates of the range often required 50–100 sampling points more
to reach the same level of accuracy as obtained with REML (cf.
Fig. 5, transect-based sampling). The superior performance of max-
imum likelihood-based variogram estimation has been found in
previous studies too (Kerry et al., 2008; Pardo-Igúzquiza, 1998;
Pardo-Igúzquiza and Dowd, 2013). Our data further indicate that
REML seems to be particularly advantageous for spatial fields with
a short autocorrelation distance (cf. Supplementary material C,
transect based sampling, fields 3, 4, 5). This observation corrobo-
rates findings of an earlier simulation study by Lark (2000b) who
noted that variograms with weak spatial structures (short ranges
and comparatively large nugget-to-sill ratios) were significantly
better estimated by REML than with the MoM-based approach.
Consequently, these results hint at the usefulness of REML for
the spatial analysis of throughfall- and other weakly structured
data. The drawback of REML, however, is that it cannot deal with
outlying values. If the removal of outliers is not justifiable or not
effective, MoM is still an appropriate approach to analyze autocor-
relation structures in throughfall data.
3.5. Implications for throughfall studies
A closer look on previous studies that analyzed throughfall spa-
tial patterns reveals that sample sizes, sampling designs and
extents of research areas differed widely in the past (Table 2). Most
studies applied a sample size of less than 100, some studies even
worked with less than 50 data. Furthermore, all but one study
applied MoM estimation. According to our data (Figs. 4 and 5),
the results of most of the studies on throughfall spatial patterns
are associated with large uncertainties. In fact, it is likely that some
of the detected autocorrelation patterns rather reflect shortcom-
ings of the sampling than true structures in throughfall. Conse-
quently, the data of these studies have to be interpreted very
carefully. Therefore, claims that the range can be linked to crown
structure (Keim et al., 2005; Loescher et al., 2002) or to phenolog-
ical dynamics in the canopy (Gerrits et al., 2010; Keim et al., 2005)
need to be verified by other case studies and not taken as common
knowledge.
From our point of view, understanding spatio-temporal pat-
terns of hydrological processes in forests requires a much higher
sampling effort than previously implemented. A combination of a
sample size of 150–200, a sampling design that ensures the sam-
pling of small distances and variogram estimation by REML seems
to be a good compromise between accuracy and efficiency. Studies
that rely on the MoM-based approach have to use at least 200 sam-
ples. Furthermore, the choice of the extent offers yet another pos-
536 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537
sibility to improve variogram estimation. Very large ratios of the
extent to the autocorrelation length should be avoided. At the
same time, plots should extend over several times of the ‘‘true”
range to avoid biased estimates of the range (cf. Western and
Blöschl, 1999).
4. Conclusions
In this study we employed three extents (plots with an edge
length of 25 m, 50 m, and 100 m), four common sampling designs
(grid, grid with additional sampling points, transect, and random
sampling), a wide range of sample sizes (50, 100, 150, 200, 400),
and two estimation methods (residual maximum likelihood
(REML) vs. method-of-moments (MoM)) to investigate which sam-
pling setup and methodology allows an optimum variogram esti-
mation of non-Gaussian and weakly structured throughfall data.
Our calculations are based on six simulated throughfall fields that
we repeatedly sampled. Subsequently, we compared the obtained
variogram parameter estimates against the true field parameters.
Our results provide evidence that the choice of the extent
strongly influences variogram estimation. A comparatively small
ratio of the extent to the correlation length is beneficial for vari-
ogram estimation. Yet, the extent should be larger than several
times the ‘‘true” range to avoid biased estimates of the range
parameter. Our results further indicate that transect-based sam-
pling is superior to both grid-based designs and to random sam-
pling. Random sampling, in turn, outperformed the grid-based
designs due to the inclusion of multiple small distances. Lastly,
REML-based variogram estimation is superior to the widely used
MoM approach.
For an optimum spatial analysis of throughfall data we recom-
mend the following setup and methodology: an extent exceeding
several times the ‘‘true” range, a minimum of 150 samples, tran-
sects covering several small (0.1–1.5 m) distances, and REML-
based variogram estimation. In situations that do not allow the
application of REML, MoM can be used but a minimum of 200 data
is required. The suggested data requirements double previous rec-
ommendations which were based on normally distributed data
with comparatively strong and long ranging autocorrelation struc-
tures. Our calculations further indicate that most previous studies
on throughfall autocorrelation structures are likely associated with
large uncertainties. Therefore, we suggest that new case studies
may verify previous results regarding the autocorrelation in
throughfall data and its link to forest characteristics.
Acknowledgements
This research was supported by the German Research Founda-
tion (ZI 1300/1-1). Field data collection was further supported by
the Agua Salud Project, a research initiative sponsored by the HSBC
Climate Partnership. We thank two anonymous reviewers and the
associate editor for their constructive comments.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.jhydrol.2016.06.
042.
References
Allen, S.T., Keim, R.F., McDonnell, J.J., 2015. Spatial patterns of throughfall isotopic
composition at the event and seasonal timescales. J. Hydrol. 522, 58–66. http://
dx.doi.org/10.1016/j.jhydrol.2014.12.029.
Bárdossy, A., Kundzewicz, Z.W., 1990. Geostatistical methods for detection of
outliers in groundwater quality spatial fields. J. Hydrol. 115, 343–359. http://dx.
doi.org/10.1016/0022- 1694(90)90213-H.
Bellot, J., Escarre, A., 1998. Stemflow and throughfall determination in a resprouted
Mediterranean holm-oak forest. Ann. Sci. For. 55, 847–865. http://dx.doi.org/
10.1051/forest:19980708.
Blöschl, G., 1999. Scaling issues in snow hydrology. Hydrol. Process. 13, 2149–2175.
http://dx.doi.org/10.1002/(SICI)1099-1085(199910)13:14/15<2149::AID-
HYP847>3.0.CO;2-8.
Bogaert, P., Russo, D., 1999. Optimal spatial sampling design for the estimation of
the variogram based on a least squares approach. Water Resour. Res. 35, 1275–
1289. http://dx.doi.org/10.1029/1998WR900078.
Corsten, L.C.A., Stein, A., 1994. Nested sampling for estimating spatial
semivariograms compared to other designs. Appl. Stoch. Models Data Anal.
10, 103–122. http://dx.doi.org/10.1002/asm.3150100205.
Cressie, N., 1985. Fitting variogram models by weighted least squares. J. Int. Assoc.
Math. Geol. 17, 563–586. http://dx.doi.org/10.1007/BF01032109.
Cressie, N., Hawkins, D.M., 1980. Robust estimation of the variogram: I. Math. Geol.
12, 115–125. http://dx.doi.org/10.1007/BF01035243.
De Gruijter, J.J., Brus, D.J., Bierkens, M., Knotters, M., 2006. Sampling for
Natural Resource Monitoring, first ed. Springer-Verlag, Berlin Heidelberg
New York.
Diggle, P.J., Ribeiro, P.J., 2007. Model-based Geostatistics. Springer, New York.
Dowd, P.A., 1984. The variogram and kriging: robust and resistant estimators. In:
Verly, G., David, M., Journel, A.G., Marechal, A. (Eds.), Geostatistics for Natural
Resources Characterization. Springer, Netherlands, pp. 91–106.
Fathizadeh, O., Attarod, P., Keim, R.F., Stein, A., Amiri, G.Z., Darvishsefat, A.A., 2014.
Spatial heterogeneity and temporal stability of throughfall under individual
Quercus brantii trees. Hydrol. Process. 28, 1124–1136. http://dx.doi.org/
10.1002/hyp.9638.
Gascuel-Odoux, C., Boivin, P., 1994. Variability of variograms and spatial estimates
due to soil sampling: a case study. Geoderma 62, 165–182. http://dx.doi.org/
10.1016/0016-7061(94)90034-5.
Gerrits, A.M.J., Pfister, L., Savenije, H.H.G., 2010. Spatial and temporal variability of
canopy and forest floor interception in a beech forest. Hydrol. Process. 24,
3011–3025. http://dx.doi.org/10.1002/hyp.7712.
Gómez, J., Vanderlinden, K., Giráldez, J., Fereres, E., 2002. Rainfall concentration
under olive trees. Agric. Water Manag. 55, 53–70. http://dx.doi.org/10.1016/
S0378-3774(01)00181-0.
Hiemstra, P.H., Pebesma, E.J., Twenhöfel, C.J.W., Heuvelink, G.B.M., 2009. Real-time
automatic interpolation of ambient gamma dose rates from the Dutch
radioactivity monitoring network. Comput. Geosci. 35, 1711–1721. http://dx.
doi.org/10.1016/j.cageo.2008.10.011.
Hsueh, Y., Allen, S.T., Keim, R.F., 2016. Fine-scale spatial variability of throughfall
amount and isotopic composition under a hardwood forest canopy. Hydrol.
Process. 30, 1796–1803. http://dx.doi.org/10.1002/hyp.10772.
Jian, X., Olea, R.A., Yu, Y.-S., 1996. Semivariogram modeling by weighted least
squares. Comput. Geosci. 22, 387–397. http://dx.doi.org/10.1016/0098-3004
(95)00095-X.
Journel, A.G., Huijbregts, C.J., 1978. Mining Geostatistics. Academic Press, London,
New York, San Francisco.
Keim, R.F., Skaugset, A.E., Weiler, M., 2005. Temporal persistence of spatial patterns
in throughfall. J. Hydrol. 314, 263–274. http://dx.doi.org/10.1016/j.
jhydrol.2005.03.021.
Kerry, R., Ingram, B.R., Goovaerts, P., Oliver, M.A., 2008. How many samples are
required to estimate a reliable REML variogram. In: Geostats 2008. Proceedings
of the Eighth International Geostatistics Congress, pp. 1155–1160.
Kerry, R., Oliver, M.A., 2007. Determining the effect of asymmetric data on the
variogram. I. Underlying asymmetry. Comput. Geosci. 33, 1212–1232. http://dx.
doi.org/10.1016/j.cageo.2007.05.008.
Klos, P.Z., Chain-Guadarrama, A., Link, T.E., Finegan, B., Vierling, L.A., Chazdon, R.,
2014. Throughfall heterogeneity in tropical forested landscapes as a focal
mechanism for deep percolation. J. Hydrol. 519 (Part B), 2180–2188. http://dx.
doi.org/10.1016/j.jhydrol.2014.10.004.
Lark, R.M., 2000a. A comparison of some robust estimators of the variogram for use
in soil survey. Eur. J. Soil Sci. 51, 137–157. http://dx.doi.org/10.1046/j.1365-
2389.2000.00280.x.
Lark, R.M., 2000b. Estimating variograms of soil properties by the method-of-
moments and maximum likelihood. Eur. J. Soil Sci. 51, 717–728.
Lark, R.M., 2002a. Optimized spatial sampling of soil for estimation of the variogram
by maximum likelihood. Geoderma 105, 49–80. http://dx.doi.org/10.1016/
S0016-7061(01)00092-1.
Lark, R.M., 2002b. Modelling complex soil properties as contaminated regionalized
variables. Geoderma 106, 173–190. http://dx.doi.org/10.1016/S0016-7061(01)
00123-9.
Lark, R.M., Cullis, B.R., Welham, S.J., 2006. On spatial prediction of soil properties in
the presence of a spatial trend: the empirical best linear unbiased predictor (E-
BLUP) with REML. Eur. J. Soil Sci. 57, 787–799. http://dx.doi.org/10.1111/j.1365-
2389.2005.00768.x.
Lloyd, C., Marques, A.D.O., 1988. Spatial variability of throughfall and stemflow
measurements in Amazonian rainforest. Agric. For. Meteorol. 42, 63–73. http://
dx.doi.org/10.1016/0168-1923(88)90067-6.
Loescher, H.W., Powers, J.S., Oberbauer, S.F., 2002. Spatial variation of throughfall
volume in an old-growth tropical wet forest, Costa Rica. J. Trop. Ecol. 18, 397–
407. http://dx.doi.org/10.1017/S0266467402002274.
Loustau, D., Berbigier, P., Granier, A., Moussa, F.E.H., 1992. Interception loss,
throughfall and stemflow in a maritime pine stand. I. Variability of throughfall
and stemflow beneath the pine canopy. J. Hydrol. 138, 449–467. http://dx.doi.
org/10.1016/0022-1694(92)90130-N.
 S. Voss et al. / Journal of Hydrology 540 (2016) 527–537
Matheron, G., 1962. Traité de géostatistique appliquée, tome i: Mémoires du bureau
de recherches géologiques et minières. Pairs Ed. Tech. 14.
Morris, M.D., 1991. On counting the number of data pairs for semivariogram
estimation. Math. Geol. 23, 929–943. http://dx.doi.org/10.1007/BF02066733.
Möttönen, M., Järvinen, E., Hokkanen, T.J., Kuuluvainen, T., Ohtonen, R., 1999.
Spatial distribution of soil ergosterol in the organic layer of a mature Scots pine
(Pinus sylvestris L.) forest. Soil Biol. Biochem. 31, 503–516. http://dx.doi.org/
10.1016/S0038-0717(98)00122-9.
Müller, W.G., Zimmerman, D.L., 1999. Optimal designs for variogram estimation.
Environmetrics 10, 23–37. http://dx.doi.org/10.1002/(SICI)1099-095X(199901/
02)10:1<23::AID-ENV333>3.0.CO;2-P.
Papritz, A., 2014. Georob: Robust Geostatistical Analysis of Spatial Data. Version
0.1-5.
Pardo-Igúzquiza, E., 1998. Maximum likelihood estimation of spatial covariance
parameters. Math. Geol. 30, 95–108. http://dx.doi.org/10.1023/
A:1021765405952.
Pardo-Igúzquiza, E., Dowd, P.A., 2013. Comparison of inference methods for
estimating semivariogram model parameters and their uncertainty: the case
of small data sets. Comput. Geosci. 50, 154–164. http://dx.doi.org/10.1016/j.
cageo.2012.06.002.
Patterson, H.D., Thompson, R., 1971. Recovery of inter-block information when
block sizes are unequal. Biometrika 58, 545–554. http://dx.doi.org/10.1093/
biomet/58.3.545.
Pettitt, A.N., McBratney, A.B., 1993. Sampling designs for estimating spatial variance
components. Appl. Stat. 42, 185. http://dx.doi.org/10.2307/2347420.
R Core Team, 2013. R: A language and environment for statistical computing. R
Foundation for Statistical Computing, Version 3.0.1, Vienna, Austria.
Ritter, A., Regalado, C.M., 2010. Investigating the random relocation of gauges below
the canopy by means of numerical experiments. Agric. For. Meteorol. 150,
1102–1114. http://dx.doi.org/10.1016/j.agrformet.2010.04.010.
Russo, D., Jury, W.A., 1987. A theoretical study of the estimation of the correlation
scale in spatially variable fields. 1. Stationary fields. Water Resour. Res. 23,
1257–1268. http://dx.doi.org/10.1029/WR023i007p01257.
Shachnovich, Y., Berliner, P.R., Bar, P., 2008. Rainfall interception and spatial
distribution of throughfall in a pine forest planted in an arid zone. J. Hydrol. 349,
168–177. http://dx.doi.org/10.1016/j.jhydrol.2007.10.051.
537
Skøien, J.O., Blöschl, G., 2006. Scale effects in estimating the variogram and
implications for soil hydrology. Vadose Zone J. 5, 153–167. http://dx.doi.org/
10.2136/vzj2005.0069.
Staelens, J., De Schrijver, A., Verheyen, K., Verhoest, N.E.C., 2006. Spatial variability
and temporal stability of throughfall water under a dominant beech (Fagus
sylvatica L.) tree in relationship to canopy cover. J. Hydrol. 330, 651–662. http://
dx.doi.org/10.1016/j.jhydrol.2006.04.032.
Van Groenigen, J.W., 1999. Constrained Optimisation of Spatial Sampling: A
Geostatistical Approach. Wageningen.
Warrick, A.W., Myers, D.E., 1987. Optimization of sampling locations for variogram
calculations. Water Resour. Res. 23, 496–500. http://dx.doi.org/10.1029/
WR023i003p00496.
Webster, R., Oliver, M.A., 2007. Geostatistics for Environmental Scientists (Statistics
in Practice), second ed. Wiley, Chichester.
Webster, R., Oliver, M.A., 1992. Sample adequately to estimate variograms of soil
properties. J. Soil Sci. 43, 177–192. http://dx.doi.org/10.1111/j.1365-2389.1992.
tb00128.x.
Western, A.W., Blöschl, G., 1999. On the spatial scaling of soil moisture. J. Hydrol.
217, 203–224. http://dx.doi.org/10.1016/S0022-1694(98)00232-7.
Whelan, M.J., Sanger, L.J., Baker, M., Anderson, J.M., 1998. Spatial patterns of
throughfall and mineral ion deposition in a lowland Norway spruce (Picea abies)
plantation at the plot scale. Atmos. Environ. 32, 3493–3501. http://dx.doi.org/
10.1016/S1352-2310(98)00054-5.
Ziegler, A.D., Giambelluca, T.W., Nullet, M.A., Sutherland, R.A., Tantasarin, C., Vogler,
J.B., Negishi, J.N., 2009. Throughfall in an evergreen-dominated forest stand in
northern Thailand: comparison of mobile and stationary methods. Agric. For.
Meteorol. 149, 373–384. http://dx.doi.org/10.1016/j.agrformet.2008.09.002.
Zimmermann, A., Zimmermann, B., 2014. Requirements for throughfall monitoring:
the roles of temporal scale and canopy complexity. Agric. For. Meteorol. 189–
190, 125–139. http://dx.doi.org/10.1016/j.agrformet.2014.01.014.
Zimmermann, A., Zimmermann, B., Elsenbeer, H., 2009. Rainfall redistribution in a
tropical forest: spatial and temporal patterns. Water Resour. Res. 45, W11413.
http://dx.doi.org/10.1029/2008WR007470.
Zimmermann, B., Zimmermann, A., Lark, R.M., Elsenbeer, H., 2010. Sampling
procedures for throughfall monitoring: a simulation study. Water Resour. Res.
46, W01503. http://dx.doi.org/10.1029/2009WR007776.