Published November, 1998

Scaling Water Retention Curves for Soils with Lognormal Pore-Size Distribution
K. Kosugi and J. W. Hopmans*
ABSTRACT erogeneous soils, or in field plots that are apparently
The scaling theory approach has been widely used as an effective homogeneous. The concept of similar media was intro-
method to describe the variation of soil hydraulic properties. In con- duced by Miller and Miller (1956) to develop scaling
ventional scaling, reference retention curves and scaling factors are theory for the analysis of such variations in field soils.
determined from minimization of residuals. Most previous studies Scaling provides a means to relate hydraulic properties
have shown that scaling factors are lognormally distributed. In this of different soils to those of a reference soil using scaling
study, we derived physically based scaling factors, assuming that soils factors. In recent years, the scaling theory approach has
are characterized by a lognormal pore-size distribution function. The been widely used as an effective method to describe the
theory was tested for three sets of retention data. Two data sets
variation of soil hydraulic properties.
included samples of a sandy loam soil, and one set included samples
of a loamy sand soil. Individual soil water retention data were fitted
Previous studies proposed different methods to deter-
to the retention model proposed by Kosugi (1996). The parameters mine scaling factors for soil hydraulic properties. While
of the model are the mean and variance of the log-transformed pore- Russo and Bresler (1980) suggested the scaling method
radius distribution. Scaling factors and parameters of the reference to compute reference hydraulic properties directly from
curve were computed directly from the parameters of individual soil observed hydraulic data, most previous studies em-
water retention functions. Assuming that (i) the soil pore radius of a ployed functional models for the reference retention
study area is lognormally distributed and (ii) soil samples are obtained and conductivity curves. Warrick et al. (1977) adopted
from random sampling of effective soil pore volume from the study polynomial functions to express reference h(6) and K(Q)
area, we have proposed a theoretical interpretation of the lognormal curves. Simmons et al. (1979) used a logarithmic func-
scaling factor distribution. Scaling results for all three data sets com- tion for the scaling of the water retention characteristic.
pared well with those obtained using the conventional scaling method.
The combined soil water retention-hydraulic conductiv-
ity model proposed by Brooks and Corey (1964) was
used by Ahuja and Williams (1991) in the scaling of
D ESCRIBING WATER FLOW in soils requires knowledge
of the soil hydraulic properties. The hydraulic
properties of unsaturated soil are represented by the
6(/z) and K(h) relationships. Clausnitzer et al. (1992)
proposed a method of simultaneous scaling of h(Q) and
K(Q) curves by employing the combined soil water re-
water retention characteristic (the relationship between tention-hydraulic conductivity model proposed by van-
the volumetric soil water content, 6, and the capillary Genuchten (1980). Moreover, Zhang et al. (1993)
pressure head, h, and the unsaturated hydraulic conduc- showed scaling results for four different soil water reten-
tivity, K, function. Both properties are variable in het- tion models. Most studies showed that scaling factors
are lognormally distributed (e.g., Warrick et al., 1977;
K. Kosugi, Graduate School of Agriculture, Kyoto Univ., Kyoto 606- Hopmans, 1987; Clausnitzer et al., 1992; Zhang et al.,
8502, Japan; J.W. Hopmans, Dep. of Land, Air, and Water Resources, 1993).
Hydrology Program, Univ. of California, Davis, CA 95616. Received
15 Aug. 1997. *Corresponding author (jwhopmans@ucdavis.edu). Abbreviations: C, conventional; PB, physically based; PDF, probabil-
ity density function; REV, representative elementary volume; RSS,
Published in Soil Sci. Soc. Am. J. 62:1496-1505 (1998). residual sum of squares.
 KOSUGI AND HOPMANS: SCALING WATER RETENTION CURVES FOR SOILS 1497

On the basis of the lognormal distribution of scaling tion of physically based scaling factors is derived that, com-
factors, stochastic models have been proposed to ana- bined with the reference retention curve, characterizes the
lyze the effects of variable soil hydraulic properties on spatial variability of soil water retention data.
saturated and unsaturated soil water flow. By employing
Monte Carlo simulation and using the scaling factor as Lognormal Distribution Model for Soil Pore-Size
a lognormally distributed random variable, Clapp et al. Distribution and Water Retention Curve
(1983) concluded that heterogeneity in soil hydraulic The probability density function (PDF) of soil pore radius
properties may account for approximately 75% of the r, p(r), is defined as (Brutsaert, 1966)
observed standard deviation of field water content.
Ahuja et al. (1984) examined infiltration phenomena p(r) = dSJdr [1]
by using lognormally distributed scaling factors for the where 5e is the effective saturation
saturated hydraulic conductivity. Assuming that the
Se = (6 - er)/(6, - 8r) [2]
scaling factor distribution function is lognormal, the sta-
tistical properties of soil water regime (Hopmans and described by 6S and 6r, denoting 3 3
the saturated and residual
Stricker, 1989; Kim et al., 1997) and solute transport volumetric water content (L L~ ), respectively, and the di-
(van Ommen et al., 1989; Bresler and Dagan, 1979) for mension of p(r) is LT1. In Eq. [I], p(r)dr = dSe, represents
large soil domains (agricultural field, watershed) were the volume of pores of radius r > r + dr per unit effective
pore volume of soil. The effective pore volume is defined as
analyzed. Most recently, Nielsen et al. (1998) concluded the product of the total soil volume and the effective porosity,
that scaling opportunities to describe field soil water (6S - 6 r ), of the soil. Integrating Eq. [1] yields the cumulative
behavior continue to appear both promising and pro- pore-radius distribution function:
vocative.
Thus, many studies have demonstrated the potential Se(r) = p(r)dr [3]
of scaling to describe variability of soil hydraulic proper-
ties. However, these studies have not emphasized the Consequently, 5e(r) represents the effective saturation of the
statistical significance of the scaling factor distribution. soil when the pore volume fraction occupied by pores with a
No theoretical interpretation has been proposed for the radius equal to or smaller than r is filled with water. Here, r
apparent lognormal distribution of scaling factors. This is associated with the capillary pressure head, h (L), (h > 0
is most likely so because most previous studies used for unsaturated soil) by the capillary pressure function
empirical curve-fitting equations for soil hydraulic prop- h = 2y cosfi/pgr = Air or In h = In A In r [4]
erties, which do not address the physical significance of
their parameters. The objectives of this study were to where -y is the interfacial tension, p is the contact angle, p is
present a physically based method of scaling soil water the density of wetting fluid, and g is the acceleration of gravity.
The value of A for air-water-soil systems is 0.149 cm2. On
retention curves using the physically based retention the basis of the direct correspondence of r and /z, the function
model introduced by Kosugi (1996) and to propose a Se(r) is transformed into Se(h), which is the soil water reten-
theoretical interpretation for scaling factor distribu- tion curve.
tions. The Kosugi (1996) model assumes the soil pore The PDF of the natural logarithm of pore radius (Inr) is
radii to be lognormally distributed. Consequently, the expressed as
parameters of the retention model have physical signifi-
cance and are directly related to the statistical properties /(In r) = dSJd In r [5]
of the soil pore-size distribution. The lognormal soil Consequently, /(In r)d In r = dSe represents the volume of
pore-size distribution has been assumed in some previ- full pores of log-transformed radius In r > In r + d In r per
ous studies. Based on the fact that many soils show unit effective pore volume of medium. Similarly, the PDF of
a lognormal particle-size distribution, Brutsaert (1966) the natural logarithm of pore capillary pressure head (In h)
proposed the lognormal distribution to describe pore- can be defined as
size distribution. Gardner (1956) introduced the possi- g(ln h) = dSJd In h [6]
bility of characterizing soil structure using a lognormal Hence, g(ln h)d In h = dSe represents the volume of full
pore-size distribution, assuming a relationship between pores in which water is retained by log-transformed capillary
aggregate size and pore size. Most recently, Nimmo pressure In h to In h + d In h per unit effective pore volume
(1997) proposed such a relationship and subsequently of medium. The relationship between p(r), /(In r), and g(ln
derived a model to describe the soil structural influence h) is expressed by
of soil water retention using a lognormal aggregate-size
distribution model. Pachepsky et al. (1995) derived the p(r)dr = /(In r)d In r = g(ln h)d\nh = dSt [7]
fractal dimension of soil pores assuming a lognormal An analytical expression for the effective saturation, 5e, is
pore-size distribution. derived by integrating /(In r), or g(ln h):

THEORY S e (lnr) = f(lnr)dlnr [8]

After establishing the functional form of the soil water
retention model for soils with a lognormal pore-size distribu- S e (ln/0 = -[ g(In/Odin/i [9]
tion, we show how statistical theorems provide physically
based parameters for the soil water retention function of a Kosugi (1996) applied a lognormal distribution law to the
study field. Based on the similar media concept, the distribu- PDF of the soil pore radius to derive a combined soil water
1498 SOIL SCI. SOC. AM. J., VOL. 62, NOVEMBER-DECEMBER 1998

retention-hydraulic conductivity model. Based on this model, [13] performs well for observed water retention data sets of
/(In r) is expressed by the normal distribution A^ln rm, a2): many soil types (Kosugi, 1994, 1996, and 1997). It can be shown
that an alternative equivalent expression for the function in
1 (In r - In rm)2 Eq. [13] is
/(In r) = exp [10]
2a2
\n h - In
Soe/(ln
i h)=-
i\ ri C1
where In rm and a2 are the mean and variance of In r, respec- [15]
tively. The parameter rm is the geometric mean pore radius
and is equal to the median pore radius [5e(ln rm) = 0.5] based where erfc denotes the complementary error function.
on the assumption of the lognormal distribution. Figure la
shows a graphical representation of Eq. [10]. The correspond-
ing expression for g(ln h ) is Statistical Characterization of Soil Sampling
1 (In h - In hm) 2 We consider that soil samples are collected at different
g(ln h) = exp [11] locations within a study area. We treat the total effective soil
2o-2 pore volume in the study area as the population, whereas each
Equation [11] indicates that In h obeys the normal distribution soil sample is obtained from random sampling of effective soil
N(ln /zm, a2). The mean of g(\n h),\n hm, is related to the mean pore volume from the population. Based on previous studies
of/(In r), In rm, by In hm = In A - In rm, whereas the variance (Kosugi, 1996 and 1997), the PDF of In r for each soil sample
of g(ln h), a2, is identical to the variance of /(In r). The is expected to describe the normal distribution:
distribution described by Eq. [11] is shown in Fig. Ib. Substitut- 1 (In r - In rm,-):
ing Eq. [10] and [11] into Eq. [8] and [9], respectively, and /(In r} = exp [16]
integrating yields expressions for the cumulative distribution 2of
functions Se(ln r) and 5e(ln h): where In r m , and of denote the mean and variance of In r,
respectively, for soil sample i (i = 1, 2, ...,/; I is total number
5e(ln f) = [12] of soil samples). Soil sample size is assumed to be constant
a (i.e., each soil sample has the same effective pore volume). It
is assumed that the PDF of In r for the population (i.e., the
Se(ln h) = Fj [13] study area) is represented by the normal distribution:
a
where FJ*x) is the normal distribution function defined as (In r - In rm.)2
/(In r) = exp [17]
2ai
1
exp(-jc2/2)dx [14] where In /> and cr2 denotes the mean and variance of In r,
2<rr J_. respectively, for the population.
Graphs of Eq. [12] and [13] are presented in Fig. Ic and Id, In most soil sampling procedures, the sample size is small
respectively. From Fig. Id it is clear that hm is the capillary relative to the population size (i.e., the effective pore volume
pressure head for 5e = 0.5. The parameter <r is defined such of a soil sample is small relative to that of the study field). In
that 5e(ln hm - <r) - 5e(ln hm + a) = 0.68 (Fig. Id) and such cases, the sample mean is normally distributed with a
consequently controls the magnitude of changes in Se around mean value equal to the population mean (Spiegel, 1975)
the inflection point. The retention model expressed by Eq. only if the samples are drawn from a normally distributed

0.4 0.4

0.3 0.3
^ variance: a \variance: a2
mean: Inr mean: ln/z_.
1 0.2 0.2

0.1 0.1

0.0 0.0
-9 -8 -7 -6 -5 -4 -3 -28 6 5 4 3 2 1 l

Inr (r in cm) InA (A in cm)

1.0 1.0
0.84 0.84

J5 0.5 0.5
v

0.16 0.16
0.0 0.0
-9 -8 -7 -6 -5 -4 -3 -2 8 1 0

Inr (r in cm) InA (h in cm)

Fig. 1. Probability density functions (a)/(In r) and (b) g(ln h), and cumulative distribution functions (c) 5e(ln r) and (d) 5e(ln h) with the median
pore radius (In rm) = -5.9 (rm = 2.7 X 10~3 cm), the median capillary pressure head (In Am) = 4.0 (hm = 55 cm), and variance (or2) = 1.0.
 KOSUGI AND HOPMANS: SCALING WATER RETENTION CURVES FOR SOILS 1499

population. Consequently, under the conditions assumed in water retention curves. Hence, we define r m , as the microscopic
this study, the sample mean, In rmi, is normally distributed characteristic length scale rather than an arbitrary pore radius.
with its mean equal to the population mean, In rm., or Consequently, scaling factors are defined by
i = rmj/fm or In a, = In rmj - In [22]
In rm. = Mean(ln /,,,) = - 2f=i In V. [18]
A graphical representation of Eq. [22] is given in Fig. 2, where
The population variance, cri, can be written as the sum of the it is clearly shown how In a, can be inferred from the mean
mean of the sample variances, Mean(of) = (1/7) '=icr2, and values of the PDF for the sample i and the reference soil.
the variance of the sample means, Var(ln rml) (see Appendix Since the sample mean, In rm>,, is normally distributed under
A for its derivation), or the conditions assumed in this study, Eq. [22] also shows that
In a; is normally distributed. That is, the scaling factor obeys
al = Mean(of) + Var(ln rm,,-) [19] a lognormal distribution. From Eq. [22] the variance of In a,,
Var(ln a,), is equal to the variance of In r ml , Var(ln r ml ), so
As the volume of a soil sample approaches the representative that Eq. [19] becomes
elementary volume (REV) of the study area, the average of
In r, In r m/ , for all samples will coincide, since geometrical o-l = Mean (of) + Var(ln a;) [23]
aspects are preserved as the centroid of the REV is moved
from one place to another (Baveye and Sposito, 1984). In this Thus, the variance of log-transformed scaling factors is de-
situation, Var(ln rmj) will be zero and the value of <jl will be fined as the difference of the total variance of In r in the study
equal to Mean(of). area, ai, and the expected variance of In r within soil samples,
Mean(of).
Since by definition the scaling factor for the reference soil
Physically Based Scaling Approach is equal to one (Eq. [20]), earlier scaling studies assumed the
arithmetic average of scaling factors equal to unity (Peck et
Based on the similar media concept, the microscopic struc-
tures of soils are assumed to be identical (Warrick et al., 1977; al., 1977):
Kutilek and Nielsen, 1994), and the soils differ only by their
microscopic length scale, which is characterized by a scaling Mean(a,) = - [24]
factor. The scaling factor, a,, is defined as the ratio of a micro-
scopic characteristic length, \,, of soil sample i to the character- This constraint appears suitable for data with normally distrib-
istic length, \, of a reference soil: uted scaling factors. However, since scaling factors have been
found to be lognormally distributed, we use the constraint
( = \,/X
a, ~r [20]
l_ j that the geometric mean of scaling factors is unity:
Defining the pore radius r as the microscopic characteristic
length, Eq. [20] becomes '=1 a,
Y" = 1 so that Mean(ln a,) = i- S;=i In a, = 0
ct, = rjr or In a, = In r, In r [21]
[25]
where r\ and r are the largest water-filled pore radii at a specific
volumetric water content for sample ; and the reference soil, After substitution of Eq. [22] into Eq. [25], it can be shown
respectively. Based on the similar media concept, each soil that In fm is computed from
sample has identical saturated and residual water content val-
ues. Hence, Eq. [21] holds for all rt and r at equal effective In rm = In rm, = Mean(ln rm,() = - 2'=i In rm4 [26]
saturation values.
As an example, Fig. 2 shows the PDF /(In r) for a soil
Using the scaling theory approach as outlined above, the PDF
sample ;', /(In r), and for a reference soil, /(In r). Since the
of In r for the reference soil, /(In r), is defined by
similar media theory assumes that each soil sample is charac-
terized by a single scale factor only, of values for each sample 1 (lnr-lnrm)2
i and the variance of In r for the reference soil, cr2, are assumed /(In r) =
</2<7T<7 L 2CT2 J
identical. Moreover, In r m , and In rm represent the mean of
the PDF/(In r) and/(In r), respectively. Since rmt is the pore 2
where In rm is equal to Mean(ln rm4), and a is computed as
radius for which the effective saturation Se is equal to 0.5, rmj the mean of the sample variances, Mean(of) (Fig. 2).
is a suitable representative pore radius to characterize soil
Scaling of Soil Water Retention Curves
Based on the pore-radius distribution expressed by Eq. [16]
and [17], the respective functional expressions for the soil
water retention curves, representative for the individual soil
samples and the study area, are obtained:
_ /In hm) - In h
"Je.i ~ "n\ [28]
\ CT,

In hm. - In h
St. = FJ [29]
CT.
Meanynrmj)=lnrm, = In/;
where In hmi = In A In r m / and In hm = In A In rm.. The
In a, =lnrmj. -\nfm parameters In /zm/ and In hm, represent the mean In h values
Fig. 2. Schematic representation of the probability density function for sample i and the study area, respectively. Based on Eq.
(PDF) of the lognormal distribution of the pore radius (In r) for [27], the water retention function for the reference soil is
soil sample /./(In r), and for reference soil,/(In r). defined by
1500 SOIL SCI. SOC. AM. J., VOL. 62, NOVEMBER-DECEMBER 1998

In hm - In h Se, rather than volumetric water content (Clausnitzer et al.,

<>
<Je = rF-
t\ [30] 1992).
cr The third data set (data set H) included samples of a loamy
where In hm (= In A - In rra) represents the mean In h sand soil of 100 cm3 collected from six locations at three depths
value for the reference soil. Since both In fm and In rm (30, 60, and 90 cm) in a 2-m2 area in the Hupselse Beek
are equal to the mean of In r m , (Eq. [26]), In hm is watershed (Hopmans and Stricker, 1987). A total of 43 soil
computed from samples were analyzed. The parameters hm<i and cr, were fitted
to the Kosugi Se(h) model (Eq. [28]) using the observed reten-
tion data. Following Hopmans and Stricker (1987), 0S was
In hm = In hm, = Mean(ln &,,-) = - 2'=i In hm,, [31] considered a fitting parameter and 6r was assumed to be zero.
Hence, Se was numerically equal to the degree of saturation.
The variance for the reference soil, cr2, is computed as the Mean and standard deviation of 9S were 0.338 and 0.060 m3
mean of the sample variances, Mean(o-j2). Moreover, scaling m~ 3 , respectively.
factors defined by Eq. [22] can be computed directly from The parameters for the reference soil water retention curve
in Eq. [30] for each of the three data sets were directly com-
a; = hjhm4 or In a; = In hm - In [32] puted from the mean values of In hmj (Eq. [31]) and cr2 =
The mean In h value, In hmj, for sample / is related to the Mean(a2). The scaling factors a, were directly computed from
mean In r value, In rm,, by In hm ,, = In A - In rm ,. Since In rm , Eq. [32].
is normally distributed under the assumed conditions, In hmj Reference soil water retention curves and scaling factors
is also normally distributed. Analogous to Eq. [19], the total were also computed from a "brute force" scaling approach,
variance of In h for the study area, cr2, can be written in terms minimizing the residual sum of squares (RSS) computed from
of the expected variance of In h within soil samples, Mean(cr2), the effective saturation values of the scaled vs. reference reten-
and the variance of mean In h values between soil samples, tion curves for I soil samples and /(;') retention data (conven-
Var(ln hmj): tional method):
= Mean(cr2) + Var(ln hmj) [33] RSS = [34]
where S'j is the y'th effective saturation value of soil sample i
MATERIALS AND METHODS with a capillary pressure head of h'J, and Se(a,/z'J) is the effective
saturation obtained from substituting h = a/zy into the refer-
The proposed method adopts the assumption that each soil ence soil water retention function that is expressed by Eq.
sample is obtained from random sampling of effective soil [30]. This method of scaling is similar to the method used by
pore volume from a study area comprising the population. Hills et al. (1992), except that the Kosugi retention model was
This assumption is likely to be valid if soil samples are collected used instead of the van Genuchten (1980) retention function.
from a study area with no apparent abrupt changes in soil Moreover, we imposed the same constraint expressed by Eq.
type. Three sets of laboratory-measured soil water retention [25], requiring the geometric mean of scaling factors to be
data that appeared to satisfy this requirement were subjected unity, while Hills et al. (1992) imposed no constraint. Thus,
to the proposed scaling approach. For each data set, parame- in the conventional scaling method, the parameters of the
ters ft m , and cr, of the soil water retention model in Eq. [28] reference retention curve, In hm and <r2, and scaling factors
were optimized to fit the soil water retention data using the are optimized while minimizing RSS, whereas the physically
nonlinear least squares optimization procedure that is pro- based approach determines these parameter values directly
vided in the spreadsheet program Excel (Wraith and Or, 1998). from hmj and <r,.
Values of hmi and cr/ were determined uniquely for every data
set, even if various initial parameter estimates were used.
The first two data sets (A and B) included soil water reten- RESULTS AND DISCUSSION
tion curves reported by Carvallo et al. (1976) for 35 soil sam-
ples. Soil samples of a Maddock sandy loam (sandy, mixed Soil Water Retention Parameters
Udorthentic Haploborolls) of 345 cm3 were collected at seven Values for In hm;i and cr2, which together describe the
depths from five plots in a study area of approximately 100 soil water retention function of Eq. [28] for each sample
m2 in size. Particle-size distribution data clearly showed differ- i, are presented in Fig. 3 for all three data sets. The top
ences in soil texture between the surface and subsurface soil.
horizontal scale in Fig. 3 represents the median pore
Therefore, the soil samples were categorized into two data sets
based on their clay content. Data set A (>8% clay content) radius rm,,, which is related to /zm,, by Eq. [4]. The corre-
contained soil water retention data for 17 soil samples of the sponding mean parameter values, Mean(ln hmj) and
surface soil, and data set B (<8% clay content) contained Mean(o-2), are presented in Table 1. The value of
retention data for 18 samples of the subsurface soil. Mean(ln /zm,,) is equal to both In hm, and In hm, represent-
In the fitting of soil water retention data, the saturated ing the mean of In h for the study area and the reference
water content, 6S, was fixed to the observed value and the soil, respectively. Moreover, Mean(a2) describes the
residual water content, 6r, was considered a fitting parameter. variance of In h for the reference soil. From Fig. 3 and
Optimized values of 0r were constrained such that their values Table 1 it can be seen that the soil water retention curves
were smaller than the minimal observed 6-value, but always of data set B are described by the smallest Mean(ln /z mi )
aO. Mean values of 6S and 6r for data set A were 0.390 and and Mean(cr2) values among the three data sets. Conse-
0.173 m3 m~3, respectively, with corresponding standard devia-
tions of 0.021 and 0.036 m3 m~3. Mean 6S and er values for quently, data set B has the largest median pore radius
data set B were 0.396 and 0.088 m3 m~3, respectively, with and the smallest pore-radius distribution width among
corresponding standard deviations of 0.024 and 0.022 m3 m~3. the three data sets. The data and Fig. la in Carvallo et
Since values of both 0S and Or varied between soil samples, al. (1976) support the distinct difference in pore-radius
soil water retention data were scaled using effective saturation, distribution between data sets A and B. Indeed, it is
 KOSUGI AND HOPMANS: SCALING WATER RETENTION CURVES FOR SOILS 1501

Table 1. Mean parameter values of Eq. [28], the variance of sam-

ple mean values [Var(ln ft m .,)J, and the variance for the study
-4.5 -5.0 -5.5 -6.0 -6.5 -7.0 -7.5 area (cr).
8 Var(ln
Data set Mean(ln hmj) Mean(o'J)
7 - Data set A A 3.914 1.317 0.058 1.375
o Data set B B 3.393 0.268 0.064 0.332
6 - D Data set H H 4.612 1.904 0.112 2.016
t Values of <r; are computed as the sum of Mean(of) and Var(In /imj) (see
5 Eq. [33]).
V 41
every data set, the mean value for the reference soil, In
hm, optimized by the C method is close to its calculated
value using the PB method, whereas the variance for
the reference soil, d2, optimized by the C method is
slightly smaller than the variance calculated by the PB
method (Table 2).
Figure 8 presents the probability plots of In a, using
2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 both the PB and C scaling methods. It can be seen that
a lognormal distribution describes the distribution of
scaling factors for all three data sets and both scaling
Fig. 3. Relationship between the median capillary pressure head (In
/imj.) and the variance (a?) for soil samples of all three data sets. methods. The Kolmogorov-Smirnov test did not reject
The median pore-radius, rmj, is related to hmj by Eq. [4]. the null hypothesis, i.e., In a, is normally distributed, at
the 0.20 level of significance for all cases. The statistical
shown in Fig. 3 that the soil samples of data set A (>8% properties (mean and variance) of the lognormal scaling
clay content) have large In /zm), and of values relative to factor distribution, as computed from the slope and in-
those of data set B (<8% clay content). In contrast, the terception of the linear relationship, are summarized in
retention curves of data set H have the smallest median Table 2. As expected, the variance of In CT, obtained by
pore radii and that set's pore-radius distribution width the PB method is equal to the variance of In /zmj, (Table
is the largest (Table 1). 1) for all three data sets. From Fig. 8 and Table 2 it can
According to the probability plots of In /z m , shown in be seen that the scaling factor distribution obtained by
Fig. 4, In hm_, values are normally distributed for all three the PB method is quite similar to that obtained by the
data sets, supporting the assumptions with regard to the C method for every data set. The scaling requires that
mean properties of the sampling distribution. Using the the mean of In a, be equal to zero in all cases (Eq. [25]),
Kolmogorov-Smirnov test, the null (normal) hypothesis and the scaling factor distributions are characterized by
was not rejected at the 0.20 level of significance for any the variance of In CT, only.
of the three data sets. The variance values of In hmj, Most studies (e.g., Warrick et al., 1977; Vachaud et
Var(ln /zm,,), presented in Table 1 were computed from al., 1985; Hopmans and Overmars, 1987; Clausnitzer et
the slopes of the probability plots in Fig. 4. The variance al., 1992) have shown that scaling factors are approxi-
of the sample mean values was largest for data set H; mately lognormally distributed. The same was true for
i.e., the variability of soil water retention was largest all three soil water retention data sets analyzed in this
for this data set. study. Moreover, it has been demonstrated many times

Scaling Results
Figures 5 through 7 present the scaling results for all -4.5 -5.0 -5.5 -6.0 -6.5 -7.0 -7.5
three data sets, with each of the three figures showing
the unsealed and scaled retention data, and each using
the proposed physically based (PB in Table 2) and con-
ventional (C in Table 2) scaling methods, respectively.
o
Values of RSS and the obtained parameters of the refer-
ence soil water retention curves are summarized in Ta-
ble 2. Despite the fact that the RSS value for the PB
method is slightly larger than that for the C method, a
the PB method produced excellent scaling results for O Data set A
C
IS o Data set B
every data set. The reductions in RSS were 57.8 to 89.2% tn a Data set H
and 68.9 to 89.5% for the PB and C methods, respec-
tively. In the physically based approach, all retention
data are scaled using /zm,, data only. Consequently, the 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
scaled soil water retention data obtained by the PB
method match the reference retention curve at 5e = 0.5 Fig. 4. Fractile diagram of the median capillary pressure head (In /i,,,j)
for all data sets (Fig. 5b, 6b, and 7b). In the conventional for all three data sets. The median pore radius, / is related to
approach, all retention data receive equal weight. For *,, by Eq. [4].
1502 SOIL SCI. SOC. AM. J., VOL. 62, NOVEMBER-DECEMBER 1998

a: Unsealed b: Physically-Based c: Conventional

1000
RSS = 0.101

jl 100

0.0 0.5 1.0 0.0

S S S
Fig. 5. (a) Unsealed soil water retention data, and scaled soil water retention data using (b) physically based (PB) scaling and (c) conventional
(C) scaling for data set A.

a: Unsealed b: Physically-Based c: Conventional

1000
RSS = 0.109

. 100 :

1.0 0.0 0.5 1.0 0.0

S.
Fig. 6. (a) Unsealed soil water retention data, and scaled soil water retention data using (b) physically based (PB) scaling and (c) conventional
(C) scaling for data set B.

a: Unsealed b: Physically-Based c: Conventional

1000 j
RSS = 1.496

100 :

1.0 0.0 0.5 1.0 0.0

S.
Fig. 7. (a) Unsealed soil water retention data, and scaled soil water retention data using (b) physically based (PB) scaling and (c) conventional
(C) scaling for data set H.
 KOSUGI AND HOPMANS: SCALING WATER RETENTION CURVES FOR SOILS 1503

Table 2. Residual sum of squares (RSS), parameters for reference retention curves, and statistical properties of scaling factor sets
obtained by the proposed physically based (PB) and conventional (C) scaling methods.
RSS In hm *2 Mean(ln a,) Var(ln a,)
Data set PB C PB C PB C PB c PB C
A 0.119 0.101 3.914 3.913 1.317 1.286 0 0 0.058 0.067
B 0.109 0.107 3.393 3.397 0.268 0.257 0 0 0.064 0.065
H 1.496 1.101 4.612 4.601 1.904 1.390 0 0 0.112 0.107

that the distribution of the saturated hydraulic conduc- dom sampling of effective soil pore volume from the
tivity (Ks) obeys a lognormal distribution (e.g., Kutilek study area.
and Nielsen, 1994; Bierkens, 1996). This result would
also suggest a lognormal distribution of scaling factors, Upscaling of Soil Water Retention Curves
since the hydraulic conductivity is expected to be pro-
portional to the square of the scaling factor (Jury et al., According to Eq. [23], the variance of In h for the
1987). This study proposes an interpretation of these study area, al, is equal to the sum of the expected vari-
empirically derived results. That is, a lognormal scaling ance of In h within soil samples, Mean(of), and the
factor distribution was derived by assuming that the soil variance of scaling factors, Var(ln a,), which is equal to
pore radius of a study area is lognormally distributed the variance of mean In h values between soil samples,
and that individual soil samples are obtained from ran- Var(ln hmjl) (see Eq. [33]). In Fig. 9, the contributions
of Mean(of) and Var(ln a,) to al value for all three data
sets are shown. The al values computed as the sum of
Mean(of) and Var(ln hmii) are summarized in Table 1.
Data sets H and B have the largest and smallest al
values, respectively. In all cases, however, the expected
variance within samples is much larger (81-96% of total
variance) than the variance between samples.
The retention curve for the whole study area ex-
pressed by Eq. [29] can be regarded as the upscaled soil
water retention curve, of which the parameters hm. and
a, are computed from Eq. [31] and [33], respectively.
Others have attempted to derive upscaled parameters
from mostly empirical parameters of the soil water re-
tention curves of individual soil samples (e.g., Yeh and
Harvey, 1990; Green et al., 1996) using different averag-
ing techniques (arithmetic, harmonic, or geometric). In
the proposed physically based approach of Eq. [31], the
parameter hm* is computed from the geometric mean of
/im>i, whereas the variance parameter of the upscaled
retention function, al, is derived from the arithmetic
mean of a] and the variance of In /zmj, (see Eq. [33]). In
contrast to describing the effect of soil heterogeneity
on the water regime in a study area using the Monte
Carlo approach from scaled soil hydraulic properties,
the upscaled soil water retention function expressed by

C/3

-1.0

H
-a- Physically-Based Conventional Data set
Fig. 8. Fractile diagrams for scale factor distributions using the physi- Fig. 9. Contributions of the expected variance of In h within soil
cally based and conventional methods for data sets (a) A, (b) B, samples [Mean(a/)] and the variance of scaling factors | Var(ln
and (c) H. a,)] to the study area variance (oi) value for all three data sets.
1504 SOIL SCI. SOC. AM. J., VOL. 62, NOVEMBER-DECEMBER 1998

Eq. [29] can potentially characterize field effective soil

variability using a single soil water retention curve. Of (lnr-mrm,)2/(lnr)dlnr
course, this would require additional investigations, in-
cluding the upscaling of unsaturated hydraulic conduc- + 2(ln rm. - In rm,;) (In r
tivity data and the analysis of spatial correlation be-
tween samples. - In rm.) /(In r)d In r + (In rm. - In rm,,)2 [A2]
The second term in the right hand side of Eq. [A2] can be
SUMMARY AND CONCLUSIONS written as
A physically based scaling method is proposed that 2(ln rm. - In rm,,) (In r) /(In r)d In r - In rm.
uses a lognormal pore-size distribution. By using this LJ 00 J

method, scaling factors and parameters of the reference = -2(ln /v - In rmji) 2

[A3]
curve are computed directly from the parameters of
individual soil water retention functions, which are the Substituting Eq. [A3] into [A2] leads to
mean and variance of the log-transformed pore-radius a2 = (In r - In rm.)2 /(In r)d In r - (In rm, - In rmj,)2
distribution. The proposed method was compared with
a conventional scaling method for three sets of soil water [A4]
retention data and was shown to provide excellent scal- Hence
ing results. The scaling factor distribution obtained by
the proposed method was quite similar to that obtained 1 1 r /"
by the conventional method. 7 2'=i CT? = - 2Li (In r - In rm,)2 /(In r)d In r
-* -* LJ oo

In contrast to conventional scaling, the physically

based scaling provides a theoretical interpretation of - In rm,,)2 [A5]
the lognormal scaling factor distribution with the as-
sumptions that the soil pore radius of a study area is or
lognormally distributed and that individual soil samples
are obtained from random sampling of effective soil Mean(oi) = oi Var(ln rm,) [A6]
pore volume from this study area. The statistical descrip- Consequently, Eq. [A6] is identical to Eq. [19].
tion of a set of soil water retention data is characterized
by the variance of log-transformed scaling factors, which
is defined by the difference of the total variance of the APPENDIX B
log-transformed pore radius in the study area and the A constant in the capillary pressure function given by
expected variance of the log-transformed pore radius Eq. [4], cm2
within soil samples. Moreover, physically based scaling Fa the normal distribution function
allows upscaling of the physically based parameters of /(In r) probability density function of log-transformed soil
the soil water retention function, thereby providing a pore radius
g(\n h) probability density function of log-transformed soil
soil hydraulic parameterization representative of the capillary pressure head
whole study area from individual soil samples. As a h soil capillary pressure head, cm
result, the physically based scaling can be effectively hm geometric mean and median in the lognormal distri-
used for theoretical studies on spatial variability and bution of soil capillary pressure head, cm
upscaling of soil water retention characteristics. / total number of soil samples
In hmi mean of In h for soil sample i
In hm, mean of In h for study area (= Mean(ln hmj))
ACKNOWLEDGMENTS
ln/im mean of In h for reference soil
The first author sincerely thanks Dr. T. Mizuyama, Dr. S. In rmj mean of In r for soil sample i
Kobashi, and Dr. M. Fujita (all of Kyoto University) for giving In rm, mean of In r for study area (= Mean(ln r m ,))
him an opportunity to study in the USA. This research was In rm mean of In r for reference soil
partly supported by International Scientific Research Program Mean arithmetic mean operator
of the Ministry of Education, Science, Sports, and Culture of p(r) probability density function of soil pore radius, cm"1
the Japanese Government. r soil pore radius, cm
rm geometric mean and median of the lognormal distri-
bution of soil pore radius, cm
APPENDIX A Se effective saturation
The variance of the PDF of In r for soil sample i is defined Var variance operator
a, scaling factor of soil sample i
6 volumetric soil water content, m3 m~ 3
of = J (In r - In rm,,)2 /(In r)d In r [Al] 0r residual soil water content, m3 m"3
6S saturated soil water content, m3 m~ 3
which is transformed to cr standard deviation in the normal distribution of both
log-transformed soil pore radius and capillary pres-
a] = \ (In r - In rm. + In rm - rm,,)2 /(In r)d In r sure head
of variance of both In r and In h for soil sample i
 KOSUGI AND HOPMANS: SCALING WATER RETENTION CURVES FOR SOILS 1505

oi variance of both In r and In h for study area [=

Mean(<r,2) + Var(ln /zm,;)]
a2 variance of both In r and In h for reference soil