Вы находитесь на странице: 1из 20

Journal of Hydrology 575 (2019) 1011–1030

Contents lists available at ScienceDirect

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

Research papers

Estimation of saturated hydraulic conductivity with pedotransfer functions: T


A review
⁎ ⁎
Yonggen Zhanga,b, , Marcel G. Schaapb,
a
Institute of Surface-Earth System Science, Tianjin University, Tianjin 300072, PR China
b
Department of Soil, Water and Environmental Science, University of Arizona, Tucson, AZ 85721, USA

A R T I C LE I N FO A B S T R A C T

This manuscript was handled by C. Corradini, Saturated hydraulic conductivity (Ks) is a singular parameter in earth system science. Ks not only governs the
Editor-in-Chief, with the assistance of Renato rate of flow of water under a hydraulic gradient as specified by the Darcy equation for saturated conditions, but
Morbidelli, Associate Editor also acts as a scaling factor in many unsaturated flow and transport applications that involve pore-size dis-
Keywords: tribution models. Without knowledge of saturated hydraulic conductivity, it would be difficult to accurately
Pedotransfer function describe the transport of water and dissolved or suspended constituents in soils and sediments, or calculate
Saturated hydraulic conductivity groundwater transport and recharge, and quantify the exchange between soils and the atmosphere. While the
Permeability determination of Ks is not especially difficult, it is expensive and (in many cases) infeasible to carry out field or
Kozeny-Carman
lab experiments for large-scale applications. Pedotransfer functions (PTFs) are a class of largely data-driven
Vadose zone
empirical models that aim to estimate Ks (and often other hydraulic quantities such as water retention char-
Soil
acteristics) from easily available data. In this review, we first briefly discuss the history of the development of the
concept of saturated hydraulic conductivity and its relation to the Kozeny-Carman (KC) equation. The KC
equation serves as a central point in this review because it determines which soil variables affect saturated flow
at the pore-scale, a domain which now can also be visited by computational fluid dynamics models. The KC
equation also provides us with a structure in which we can classify the large number of PTFs that have been
developed for estimating Ks. Datasets and statistical techniques available for PTF development are discussed, and
we also describe common metrics used to assess the accuracy and reliability of PTF estimates. The mutual
agreement of two main classes (i.e., an effective porosity KC-based and soil texture-based) of PTFs is analyzed
using a number of global maps of predicted Ks. Finally, we discuss challenges and perspectives that might lead to
PTFs with improved estimates of Ks. In particular, we suggest establishing and utilizing large and completely
independent databases to assess the accuracy and reliability of PTFs for global use, while also drawing in in-
formation from pedological and remote sensing sources.

1. Introduction characterization and prediction of these soil processes usually require


an accurate parameterization of soil hydraulic properties (i.e., the soil
Soil plays a fundamental role in the Earth’s terrestrial biosphere by water retention curve, the unsaturated hydraulic conductivity curve,
controlling the transfer of mass and energy between land surface and and saturated hydraulic conductivity). In many cases, however, direct
atmosphere (Amundson et al., 2015; Bittelli et al., 2015). Near-surface measurement of soil hydraulic properties is time-consuming and labor
soil is especially important because it regulates the local water balance intensive, and especially impractical for large-scale applications (Dai
through infiltration, evapotranspiration, surface runoff, groundwater et al., 2013; Vereecken et al., 2010). As an alternative, soil scientists
recharge and hence has a substantial effect on regional and global land have developed pedotransfer functions (PTFs) to estimate soil hydraulic
surface water and energy balances (Montzka et al., 2017; Vereecken properties from commonly available soil information such as soil tex-
et al., 2016; Verhoef and Egea, 2014). For example, nearly 60% of the ture and bulk density.
terrestrial precipitation is returned to the atmosphere through the soil- In the past few decades, numerous PTFs have been developed for a
plant-atmosphere continuum (Katul et al., 2012; Oki and Kanae, 2006), variety of purposes and objectives, ranging from theoretical studies to
while around half of the global biomass production and related carbon small plot-scale empirical relations to considerable efforts that meet the
cycle depend on soil processes (Cleveland et al., 2013). Accurate needs of regional and global-scale weather and climate modeling.


Corresponding authors at: Institute of Surface-Earth System Science, Tianjin University, Tianjin 300072, PR China (Y. Zhang).
E-mail addresses: ygzhang@tju.edu.cn (Y. Zhang), mschaap@email.arizona.edu (M.G. Schaap).

https://doi.org/10.1016/j.jhydrol.2019.05.058

Available online 21 May 2019


0022-1694/ © 2019 Elsevier B.V. All rights reserved.
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 1. An Excerpt of Darcy’s (1856) Appendix D in which he presents his empirical law.

Thorough reviews of the work in this field have previously been given hydrology. In order to provide a comprehensive overview, we will first
by Wösten et al. (2001), Pachepsky and Rawls (2004), Vereecken et al. provide a brief historical context of saturated hydraulic conductivity.
(2010), and Van Looy et al. (2017). Wösten et al. (2001) reviewed PTFs This overview starts with the empirical Darcy equation for saturated
to estimate soil hydraulic parameters, including soil water retention flow and proceeds to the physically-based Kozeny-Carman equation
and hydraulic conductivity characteristics. Vereecken et al. (2010) (see Kozeny, 1927; Carman, 1956, 1937). The Kozeny-Carman (KC)
specifically reviewed PTFs to estimate van Genuchten-Mualem equation subsequently provides us with a convenient framework to
(Mualem, 1976; van Genuchten, 1980) soil hydraulic properties. Van briefly discuss computational fluid dynamics (CFD) methods, which are
Looy et al. (2017) reviewed PTFs in different disciplines of the earth able to simulate saturated flow networks at the pore-scale. The flow
system science, including PTFs for water, solute, heat, and biogeo- fields computed with CFD allow us to compute saturated hydraulic
chemical soil processes. conductivity (and other quantities) in a mechanistic way from observed
Previous reviews did not focus primarily on the soil saturated hy- or specified pore-structure. The KC equation, as well as CFD methods,
draulic conductivity (Ks), which is a simple, yet critical, soil property. subsequently provide us with insight which soil variables should affect
Saturated hydraulic conductivity is needed to calculate flow rates in saturated hydraulic conductivity, which we will use as a guide to group
saturated soils using Darcy’s equation. However, Ks is also important in the many existing empirical PTFs according to their predictors. We will
the Richards equation for unsaturated flow where it is needed to scale subsequently discuss datasets and statistical techniques available for
relative hydraulic conductivity predicted by the pore-scale models of PTF development and evaluate a number of PTFs comprehensively.
Burdine (1953), Mualem (1976), and Alexander and Skaggs (1987) into Accuracy and reliability of PTFs are analyzed based on commonly used
unsaturated hydraulic conductivity. In a broader sense, our ability to criteria. Key challenges for the development and using PTFs are finally
estimate (un)saturated flow is crucial to many Earth Systems Science identified by evaluating a selection of widely-used PTFs on a global
applications, including oil and gas recovery from geological reservoirs, data set.
the protection of natural groundwater resources, the remediation of It is noted that throughout most of the review, we will use the term
polluted soil and groundwater, and the long-term stewardship of waste “saturated hydraulic conductivity” which is abbreviated by Ks with
disposal sites. units of length over time. However, it should be pointed out that other
The main thrust of this paper is to review the state-of-the-art PTFs disciplines such as civil engineering designates this quantity as per-
for estimating saturated hydraulic conductivity in soil science and meability (Sanchez-Vila et al., 2006), whereas soil sciences, geology,

1012
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

and petroleum engineering use permeability to indicate the (more (Richards, 1931). In the present, the “Richards” equation for vertical
fundamental) concept of intrinsic permeability k, which is related to Ks as unsaturated flow along axis z is commonly given as:
follows:
∂θ ∂ ⎡ ∂h
Ks = kρg μ (1) = Ks Kr ⎛ − 1⎞ ⎤
∂t ∂z ⎢
⎣ ⎝ ∂z ⎠⎥⎦ (4)
where ρ , g , and μ are fluid density, acceleration of gravity, and dynamic
where Ks scales the relative unsaturated hydraulic conductivity, Kr,
viscosity, respectively. This expression indicates that Ks depends on
with the latter varying orders of magnitude as a function of matric
both fluid properties (the density and dynamic viscosity) and the in-
potential or moisture content. Ks is therefore also an important quantity
trinsic, fluid-independent, permeability of the porous medium. The
for unsaturated flow (for example, see Mualem (1976), and van
units of intrinsic permeability are squared length units, with one
Genuchten (1980)), though there are some complexities if macropore
“darcy” unit being equivalent to a permeability of
flow is involved (for example, see Schaap and van Genuchten, 2006,
9.869233 × 10−13 m2 and defined in terms of standard atmospheric
and references therein).
pressure and unit viscosity and flow rate.
Motivated by an empirical “Übelstand” (roughly: “nuisance”) in
agricultural soil water balances, Kozeny (1927) was able to provide a
2. A brief history of saturated hydraulic conductivity and theoretical basis to hydraulic conductivity by relating intrinsic perme-
estimation from PTFs ability to pore volume and pore shape. This approach was later mod-
ified by Carman (1937, 1956) into what is now known as the Kozeny-
The concept of saturated hydraulic conductivity originated in the Carman equation, in which intrinsic permeability is written in terms of
middle of the 19′th century with Henry Darcy (1856). As an inspector particle diameter d (Kozeny, 1927; Carman, 1937):
for roads and bridges in the French city of Dijon, he wrote a publication
that mainly dealt with the distribution of water through the city’s d 2φ3
k=C
aqueducts. However, in his Appendix D, Darcy (1856) also reported on (1 − φ)2 (5)
a series of experiments on the flow of water through sand “filters”
(Fig. 1). Darcy stated that the flow (“ecoulement” in French) is pro- where φ is the porosity and C is a constant of proportionality. Alter-
portional to the pressure difference and inversely proportional to the natively, intrinsic permeability can also be defined in terms of hydraulic
thickness of the filter, which allowed him to establish the original radius, RH (Carman, 1937, see also Berryman, 1985; Berryman and
equation: Blair, 1986; Blair et al., 1996):

k×s φRH2
Q= [H + e] k=
e (2) GT 2 (6)

where Q is the flow (in volume over time); s and e are the area and where porosity (φ ) clearly provides the volume available to flow. The
thickness of the filter, respectively; H is the thickness of the layer of hydraulic radius RH can be defined as pore volume over pore-solid
water above the filter (the bottom of the filter is assumed to be at at- surface area, which means that RH is one-half of the physical radius of a
mospheric pressure). The variable k, is (translated) “a coefficient de- cylindrical pore. Tortuosity (T ) and the pore shape factor (G ) account
pendent on the nature of the sand” with material-specific values shown for the lengthening of the flow-path and non-cylindrical pores, re-
in the last column of the table contained in Fig. 1. spectively.
After eliminating the area s, and for the simplified case of vertical It is interesting to note here that the definitions by Eqs. (5) and (6)
flow, Darcy’s law is now commonly written in terms of flux (q, with are somewhat complementary: the first equation defines permeability
length over time units) and hydraulic gradient, ΔH Δz : in terms of particle phase (particle diameter), whereas the second
equation relies on a length scale of the pore space (hydraulic radius).
ΔH
q = −Ks Particle diameter is easy to measure, but hydraulic radius typically
Δz (3)
relies on analysis of two-dimensional or three-dimensional images ob-
where ΔH is the difference between water levels imposed on the inflow tained of the pore-space (e.g., Blair et al., 1996; Schaap and Lebron,
and outflow (Darcy’s H + e), along a column of length Δz (Darcy’s e), 2001). Both approaches are interesting, however, because they relate
all expressed in length units. Darcy’s k is now written as Ks, and in- intrinsic permeability to pore volume, shape and size, and the com-
terpreted as the saturated hydraulic conductivity (units: length over plexity of the flow path, all of which have a distinct physical inter-
time) for a homogeneous permeable porous medium that is fully per- pretation. We also note here that with Eq. (5), Carman established a
meated with water under laminar flow conditions. The negative sign is pedotransfer function avant-la-lettre in 1937.
needed to make the direction of the flow consistent with the coordinate Relying on measurements, Bear (1972) and later Freeze and Cherry
system. Saturated hydraulic conductivity in Darcy’s law is equivalent to (1979) summarized the values of hydraulic conductivity and intrinsic
thermal conductivity in Fourier’s law, diffusion coefficient in Fick’s law, permeability for aquifer, soil, and rocks, with saturated soil hydraulic
and electrical conductivity in Ohm’s low, all of which are conceptually conductivity ranging between 102 and 10−12 cm/s from textures ran-
similar and mathematically identical relationships. ging between clean gravel to unweathered clay (see Table 1). Campbell
The first application of saturated hydraulic conductivity to soil (1974) listed a table of saturated hydraulic conductivity for five types of
process modeling was proposed by Richardson (1922) in his book soils and derived an equation to estimate unsaturated hydraulic con-
“Weather Prediction by Numerical Process” (Fig. 2a, see the landscape ductivity. It is noted that soil water retention parameters and saturated
grids for velocity and pressure calculation in Fig. 2b; also refer to Dani hydraulic conductivity were provided together, because of the re-
Or: 2018 Walter B. Langbein Lecture), which was lost during the First quirement to derive unsaturated soil hydraulic conductivity for soil
World War, but was accidentally found and published in 1922. In this modeling. Clapp and Hornberger (1978) improved a soil water reten-
book, Richardson was attempting to solve the challenges of re- tion model and provided a lookup table of water retention parameters
presentation of land surface and water flux at the soil surface. Ri- and saturated hydraulic conductivity for 11 USDA soil classes using
chardson’s pioneering notion to combine the hydrological processes of 1446 soil samples from Holtan (1968). Cosby et al. (1984) derived re-
land surface and climate modeling not only marks the beginning of the tention parameters and saturated hydraulic conductivity by utilizing
numerical weather forecasting, but also permitted Richardson to derive 1448 samples from Holtan (1968) and Rawls et al. (1976). Both the
what is now known as the Richards equation in 1917 (see the equation lookup tables from Clapp and Hornberger (1978) and Cosby et al.
in Fig. 2c), long before Lorenzo A. Richards published his work in 1931 (1984) (see Table 2) have been extensively used in the Community

1013
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 2. (a) Lewis Frey Richardson (1881–1953). (b) Richardson’s map of meteorological weather stations and landscape grid for velocity (white) and pressure (dark)
calculations: frontispiece of Weather Prediction by Numerical Process by Richardson (Cambridge University, 1922). (c) The Richardson-Richards equation for
unsaturated soil on Page 108 from Richardson’s book. Adapted from Dani Or: 2018 Walter B. Langbein Lecture and revised based on Richardson’s book.

Table 1 easy-to-measure and/or widely available soil properties (e.g., soil tex-
The range of saturated hydraulic conductivity (Ks) for different geological ture and bulk density from soil survey). The development of PTFs
mediums. Modified from Bear (1972) and Freeze and Cherry (1979). therefore bridges the gap between the available soil data and required
Soil type Ks (cm/s) soil information and facilitates the extensive application of water flow,
energy transition, and solute transport at the field and large scales (Van
Unconsolidated Deposits Looy et al., 2017).
Gravel 1 × 10−1−1 × 102
Clean sand 5 × 10−4−1
Silty sand 1 × 10−5−1 × 10−1 3. Predictors of saturated hydraulic conductivity
Silt, loess 1 × 10−7−5 × 10−3
Glacial till 1 × 10−10−1 × 10−4 The Kozeny-Carman expression for permeability (Equation (6)) in-
Unweathered marine clay 1 × 10−10−1 × 10−7
dicates that soil saturated hydraulic conductivity, Ks, depends on par-
Rocks ticle size (i.e., soil constituents) and soil porosity (i.e., soil voids), hy-
Shale 1 × 10−11−1 × 10−7
draulic radius, and a pore shape factor. In Section 3.1, we use
Unfractured metamorphic and igneous rocks 1 × 10−12−1 × 10−8
Sandstone 1 × 10−8−5 × 10−4 computational fluid dynamics simulations on artificially generated
Limestone and dolomite 1 × 10−7−5 × 10−4 medium to show how these factors affect the permeability from a me-
Fractured metamorphic and igneous rocks 1 × 10−6−1 × 10−2 chanistic viewpoint. However, while insightful, the pore-scale imaging
process and the computational simulations are expensive and not
practical for most applications. Section 3.2 will therefore focus on
Land Model (CLM) and general circulation models (GCM) (see the land empirical PTFs that estimate permeability from readily measured
surface models in Chen and Dudhia, 2001; Dai et al., 2003; Kowalczyk quantities, such as soil particle size and soil porosity.
et al., 2006; Niu et al., 2011; Oleson et al., 2013).
To characterize many of the previous efforts, the term pedotransfer
3.1. Image analysis and computational fluid dynamics
function (PTF) was finally introduced by Bouma (1989). In general,
PTFs aim to estimate difficult to obtain soil hydraulic, thermal char-
Recent advances in both observational and computational techni-
acteristics, solute transport coefficients, or other soil properties, from
ques make it now possible to obtain the three-dimensional structure of

1014
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Table 2
Number of soil samples and Ks values of USDA soil textural classes for the class PTFs of selected publications (unit: cm/d).
Texture class Clapp and Hornberger (1978) Rawls et al. (1982) Cosby et al. (1984) Carsel and Parrish (1988) Zhang and Schaap (2017)*

Number Ks Number Ks Number Ks Number Ks Number Ks

Clay 140 11.09 291 1.44 148 8.41 1177 4.80 92 14.75
Silty clay 441 8.93 127 2.16 43 11.62 1002 0.48 29 9.61
Sandy clay 19 18.72 45 2.88 16 62.38 74 2.88 12 11.35
Clay loam 262 21.17 366 5.52 147 21.14 1317 6.24 147 7.06
Silty clay loam 147 14.69 689 3.60 325 17.58 1882 1.68 92 11.11
Sandy clay loam 80 54.43 498 10.32 104 38.46 610 31.44 179 13.23
Loam 125 60.05 383 16.32 103 29.18 1991 24.96 262 13.34
Silt loam 384 62.21 1206 31.68 394 24.27 3050 10.80 332 18.47
Sandy loam 204 299.52 666 62.16 124 45.19 2835 106.08 487 37.45
Silt – – – – – – 115 6.00 6 43.75
Loamy sand 30 1350.72 338 146.64 30 121.63 881 350.16 205 108.20
Sand 13 1520.64 762 504.00 14 402.76 803 712.80 308 642.95
Total 1845 5371 1448 15,737 2151

Note. All results are transformed into the same unit: cm/d.
* Several samples are located at the exact boundary of two classes and are therefore included in both classes, leading to the total number of soil samples that
exceed the number of soil samples in Rosetta, i.e., 2134.

the pore-space and conduct simulations of fluid flow at the pore-scale, are reached in the volume with the largest particles (top right). Fol-
allowing us to derive permeability numerically. Several studies have lowing Hilpert (2011), permeability for each of the volumes can be
carried out three-dimensional imaging of porous medium with computed from the pressure difference between inlet and outlet (top
Computed Tomography (CT, Ketcham and Carlson, 2001; Wildenschild and bottom), volume mean fluid velocity (along the pressure gradient),
and Sheppard, 2013), followed by segmentation of the gray-scale vo- the length of the flow domain (400 pixels), and the LB and physical
lumes into pore-space and particle space (Houston et al., 2013; Iassonov fluid viscosity along with some LB-topology constants. A physical voxel
et al., 2009; Iassonov and Tuller, 2010; Kaestner et al., 2008; Kulkarni size would subsequently provide a “real” permeability, but was arbi-
et al., 2012). The segmented pore-structure can subsequently be used trarily set to 1 in this example. The tortuosity of the flow field can be
for image analysis (Berryman and Blair, 1986; Blair et al., 1996; Schaap computed as:
and Lebron, 2001) or be used in Computational Fluid Dynamics (CFD)
〈 |U | 〉
to yield estimates of permeability (e.g., Hilpert, 2011; Spanne et al., T=
〈uz 〉 (7)
1994; Sukop et al., 2013; Zhou et al., 2018). Drawbacks of the CT
imaging-segmentation-CFD approach is that it is expensive (as it re- where |U | is the magnitude of the velocity, 〈 〉 is the volume mean, 〈uz 〉 is
quires CT imaging equipment, as well as substantial amounts of com- the volume mean velocity along the pressure gradient.
putation time), and that the quality of the simulations and derived Fig. 4 provides a plot of permeability versus relative particle size (by
permeability depends on the quality of the image data and the chosen mass, computed as 1- B/2L0, with L0 the initial number of large parti-
segmentation algorithm (Iassonov et al., 2009; Larsen, 2018). cles and B the number of particles broken up into smaller ones). The
As an example of the CFD approach, we provide here some results figure shows that the simulated permeability (red line with dots) gen-
obtained with a series of Lattice-Boltzmann simulations (e.g. Camargo erally increases with particle size. From the dimensional analysis of the
et al., 2012; Keehm et al., 2004; Martys and Chen, 1996; Sukop et al., units of permeability (square of length) and Eq. (5), it follows that
2013; Zhang et al., 2005) to obtain the pore-scale flow network in a permeability should increase with the square of a characteristic length
synthetic porous medium generated with a particle packing algorithm scale of the porous medium. This relation is indicated by the dashed
(Baranau and Tallarek, 2014). To this end, we initially packed 150 black line, which is referenced to the simulation at a particle size of 1.
particles of a nominal diameter of 100 pixels into a box of 5003 voxels However, the figure shows that it is not entirely correct to use particle
(volume elements) using the PackingGeneration software package size as a proxy for the length scale. Even though the permeability at a
(Baranov, 2017). The volume was subsequently cropped to a centered particle size of 0.5 is roughly a quarter of that at a particle size of 1.0 (as
400 × 400 × 400 voxel volume to eliminate boundary effects (as par- expected from Eq. (5)), the region in-between exhibits lower-than-ex-
ticles cannot be packed across the boundary of the volume). A single- pected permeability (i.e., the red line is under the dashed line). The
fluid Lattice Boltzmann (LB) simulation with pressure boundaries deviation is not caused by variation in the porosity or tortuosity; these
(Schaap et al., 2007) was subsequently run on the resulting pore values are almost constant as shown in the figure.
structure to obtain the pore-scale flow network and velocities. Twelve It turns out that hydraulic radius (as used in Eq. (6)) provides a
more simulations were run, in which we replaced single particles by better relation with permeability, as shown by the green line with solid
eight half-diameter ones of 50-pixel diameter, thus decreasing diamonds in Fig. 4. Hydraulic radius (RH) was here computed as the
“average” particle size while maintaining total particle volume. Speci- ratio of porosity and surface area, i.e., the ratio of pore voxels and pore-
fically, 5, 10, 15, 20, 25, 30, 45, 60, 75, 90, 120, and 150 ‘large’ par- solid vertices of the voxel-related volume. This implies that the hy-
ticles were thus replaced by smaller spheres, with the latter consisting draulic radius is a more appropriate length scale to characterize the
1200 smaller particles. Total in-house computation time amounted to porous medium than is particle size. This also means that -at least for
about 1500 ‘CPU’ hours, which would cost approximately $300 on an these synthetic samples- there is no direct relation between particle-size
equivalent Amazon EC2 system (https://aws.amazon.com/ec2/pricing/ and permeability, but that particle and pore arrangement is also im-
on-demand/, a1.2xlarge platform, retrieved on February 26, 2019). portant. Referring again to Fig. 3 it is clear that the particle packing and
Fig. 3 provides a depiction of packing of 1200 small particles (all flow pattern is rather “ordered” at a relative particle size at 1 (Fig. 3,
large particles converted, top left in Fig. 3), the midpoint (75 large top right) and 0.5 (Fig. 3, top left), but appears to be unordered at a
particles remaining, 75 particles converted into 600 smaller ones), and particle size of 0.75 (the middle column of Fig. 3). The random packing
packing of large particles only (top right in Fig. 3). The figure also apparently leads to a hydraulic radius that is smaller than expected
shows the resulting flow fields, indicating that the highest flow-rates from particle size.

1015
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 3. Top left: a volume with 1200 small


particles. Middle: 75 large and 600 small
particles. Top right: 150 large particles.
Bottom row: Lattice Boltzmann simulated
flow fields and relative speeds (blue for low
speed, yellow and red for mid and high
speeds, respectively). (For interpretation of
the references to color in this figure legend,
the reader is referred to the web version of
this article.)

Finally, Fig. 4 also shows an evaluation of the Kozeny-Carman arrangement, particle-size, pore topology (see Vogel et al., 2005) and
equation using data obtained from the generated volumes. For tortu- permeability. Evident here, and relevant for this review, is that the
osity, we assume the value of 1.25, which is the mean value obtained relation between particle-size and permeability is more complicated
with the LB simulations. It is interesting to note here that porosity- than a simple linear relation and that particle ordering and pore net-
based tortuosity estimates by Koponen et al. (1996), Yu et al. (2003) work effects play a role as well.
and other approaches (see Larsen, 2018), all yield much higher values
(1.5 to 2) than computed with Eq. (7). A possible explanation is that the
Koponen et al. (1996) and Yu et al. (2003) approaches rely on the total 3.2. Empirical approaches
porosity, whereas Eq. (7) relies on the velocity distribution in the actual
flow field. This approach accounts for the possibility that there is no (or A number of recent studies have endeavored to identify the factors
low) flow in some parts of the pore-space, which then does not con- controlling Ks in an empirical, data-driven way. For example, using the
tribute to tortuosity. We point out here that this is reminiscent of the HYPRES (Hydraulic Properties of European Soil) soil hydraulic data-
concept of effective porosity that is discussed in Section 3.1. The only base, Lilly et al. (2008) found that soil textural and structural in-
unknown that remains is the geometry factor, G (Eq. (4)); by setting G formation are the most useful input for deriving PTFs to estimate Ks.
equal to 1 , the Kozeny-Carman formulation according to Eq. (6) yields Papanicolaou et al. (2015) found that soil texture is the dominant factor
2
a near-exact match with the simulated LB data. in determining Ks values for high sand content (> 15%) soils and low
The above example shows that CFD can provide a detailed insight agricultural activities, while bulk density dominated low sand content
into the actual flow patterns inside the porous medium, provided that and high agricultural practices at three hillslopes in southeast Iowa.
the pore-structure can somehow be obtained with CT imaging, statis- Jarvis et al. (2013) and Jorda et al. (2015) found that Ks are strongly
tical reconstruction (Keehm et al., 2004; Yeong and Torquato, 1998), or depended on soil bulk density and land use instead of soil texture for
particle packing (as done here). Such analyses are computationally in- the top soils (< 0.3 m depth) based on published literature.
tensive, but provide detailed data about the relations between particle In order to prevent too much tedious enumeration in discussing the
empirical estimation of Ks, we have grouped available methods in three

6 6
Porosity, Tortuosity, Hydraulic Radius (length)

Hydraulic Radius
Square of particle size
5 Simulated k (from LBM) 5
Kozeny−Carman prediction
Permeability (length2)

Tortuosity
4 Porosity 4

3 3

2 2

1 1

0 0
0.5 0.6 0.7 0.8 0.9 1
Relative Particle Size (−)
Fig. 4. Permeability (values on the left axis), and porosity, tortuosity, hydraulic radius (values on the right axis) versus relative particle size (see text for details).
Simulated k is based on Lattice-Boltzmann method.

1016
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 5. Relationship between bulk density


and soil OC content in NCSS (National
Cooperative Soil Survey, 2017) dataset. (a)
The scatter points of all dataset used in
NCSS; red represents high data density, blue
low density. (b) Average bulk density and
corresponding average soil OC contents
shown in (a); red line indicates 95% con-
fidence intervals. (For interpretation of the
references to color in this figure legend, the
reader is referred to the web version of this
article.)

general classes. Porosity based approaches, particle size-based ap- estimation of Ks by including the Brooks-Corey pore-size distribution
proaches and PTFs that use soil structural information. The first two index (Brooks and Corey, 1964) into Eq. (8) (Eq. A(2) in the Appendix).
classes parallel the KC formulations according to Eqs. (5) and (6), re- Timlin et al. (1999) found that further improvements were possible
spectively. The third class (soil structural effects), deals with PTFs when the Brooks-Corey pore-size distribution index was included in the
where the effect of porosity and particle size is less clear and the effects coefficient B (Appendix A3), although the improvement was not sig-
of soil inhomogeneity due to structure or macropores are apparent. nificant compared with the results from Eq. (8). Spychalski et al. (2007)
also modified the coefficients of Eq. (8) and obtained a model that
performed better than published PTFs when evaluated on independent
3.2.1. Estimation using soil porosity
data obtained from the UNSODA data set.
Eqs. (5) and (6) clearly show the dominating role of soil porosity in
Although estimation of Ks based on effective porosity from Eq. (8) is
permeability, not only because it sets the volume available to saturated
satisfying, accurate estimation of effective porosity is still difficult.
flow, but also because together with surface area it defines the hy-
Sobieraj et al. (2001) evaluated the performance of nine published PTFs
draulic radius. Several of the parameters in Eqs. (5) and (6) are difficult
and found estimates were not reliable for the La Cuenca catchment in
to measure which motivated Ahuja et al. (1984, 1989) to simplify the
Peru, which in turn made them unsuitable for modeling stormflow
KC equation (Carman, 1956) by utilizing the concept of effective por-
events. They suggested that an improved estimation of effective por-
osity, φe , as:
osity should be pursued. Revil & Cathles (1999) used the classical Ko-
Ks = Bφe n (8) zeny-Carman according to Eq. (6) and derived a new equation very
similar to Eq. (5) for pure shale and clay-free sands by including the
where B and n are empirically determined constants. Effective porosity electrical formation factor, separating pore throat from total porosity
is defined here as the difference between the total porosity and water and effective from total hydraulic radius. The resulting model captured
content at 330 cm pressure head. Effective porosity is therefore related the effective porosity and an effective hydraulic radius and was able to
to soil water retention and involves the volume fraction of large pores estimate permeability over 11 orders of magnitude for a wide variety of
that would contribute to saturated flow in a meaningful way. In this materials obtained from the literature; the model error was within one
context, it is interesting to note that the water content at the 330 cm order of magnitude.
pressure head value is (at least in the USA) considered to be field ca- Bulk density can also be used to provide information about soil
pacity (i.e., the pressure head where gravity drainage of unsaturated porosity, as well as soil compaction (Rabot et al., 2018). Generally, bulk
soil conceptually ceases). Moisture content at 330 cm pressure head is density is shown to be negatively correlated with Ks. High bulk density
also routinely measured in US soil surveys (National Cooperative Soil indicates tightly packed soils, with low porosities and presumably low
Survey, 2017). Ahuja et al. (1989) showed that there was good agree- hydraulic radii, which should result in low Ks values.
ment between estimated Ks from Eq. (8) and experimental data by Hierarchical PTFs such as Schaap et al. (2001), and Zhang et al.
utilizing independent validation data, considering the large uncertainty (2018) have shown that including bulk density as an additional pre-
in measured Ks values. Further details regarding the empirical constant dictor can increase R2 between estimated and measured Ks values from
B and n can be found in Appendix A1. Rawls et al. (1998) improved the

1017
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

around 0.45 by employing only soil textural percentages as predictors half an order of magnitude are not uncommon. The lookup tables of
to 0.55 based on the same calibrated Rosetta dataset. PTFs that use bulk Rawls et al. (1982) and Carsel and Parrish (1988) are widely used in
density as predictors include Rawls and Brakensiek (1985), Vereecken soil science and vadose zone hydrology.
et al. (1990), Campbell and Shiozawa (1992), Jabro (1992), Schaap Although class PTFs are straightforward and can be used even when
et al. (2001, 2004), Jarvis et al. (2002), Li et al. (2007), Aimrun and hand-classifying soil texture ‘in the field’, a disadvantage is that satu-
Amin (2009), Twarakavi et al. (2009), Merdun (2010), Tóth et al. rated hydraulic conductivity changes discontinuously across class
(2015), Yang et al. (2018), among others. boundaries (e.g., for sandy loam the Rawls et al., 1982 class PTF pro-
Bulk density is shown to have a negative correlation with soil OC vides 62 cm/day, but for the adjacent loam class it lists 16 cm/day in
content (Heuscher et al., 2005; Zacharias and Wessolek, 2007) and Table 2). When class PTFs are used, small changes in the actual sand,
therefore soil OC content has been employed to estimate bulk density in silt, and clay percentages may yield rather large changes in Ks, which
previous works, such as Adams (1973) and Rawls (1983). A scatter plot may not be appropriate for some applications.
between soil bulk density and OC content for 49,855 soil samples from The discontinuity problem of class PTFs can be alleviated by de-
the NCSS (National Cooperative Soil Survey, 2017) dataset is provided riving linear/nonlinear regression models that predict Ks from con-
in Fig. 5a. By averaging soil bulk density and OC content within ranges, tinuously varying variables that characterize the particle size distribu-
there is a clear decreasing relationship between these two quantities tion. The most commonly used variables to express the relationship
(Fig. 5b), with narrow 95% confidence intervals, which shows that between particle size distribution and Ks are sand, silt, and clay per-
there is indeed an inverse relationship between bulk density and soil OC centages (technically only two of these should be sufficient). Other
content. OC content as a predictor to estimate Ks includes work by quantities used to express particle size distribution include geometric
Vereecken et al. (1990), Tamari et al. (1996), Wösten (1997), Wösten mean particle diameter from Mishra et al. (1989), or the approach of
et al. (1999), Weynants et al. (2009), among others. Araya and Paris (1981) and Arya et al. (1999). In the latter approach,
Wagner et al. (2001) compared several PTFs by utilizing OC content the pore size distribution is estimated from particle size distribution.
and those PTFs without OC as predictors based on an independent 63 Saturated (and also unsaturated) hydraulic conductivity is subsequently
German soil horizons and found that Wösten (1997)’s work by in- computed using the Hagen-Poiseuille law. Using a small subset of the
cluding OC content as a predictor performed the best. However, Wagner UNSODA data set (Leij et al., 1996; Nemes et al., 2001), Arya et al.
et al. (2001) suggested that the evaluation dataset is too small, and the (1999) obtained R2 of 0.807 with an average root mean square error
conclusions might not be general enough, and therefore larger dataset is (RMSE, see Section 4.3) of 0.878.
suggested to be used for the comparison. Nemes et al. (2005) found a Because of the large amount of PTFs that are based on soil texture
negative correlation between soil organic carbon content and Ks based data information, we do not list them in detail here, but we provide
on four PTFs, including Vereecken et al. (1990), Wösten et al. (1999, some commonly used PTFs, as well as some recently developed ones in
2001), and their newly developed PTFs from effective porosity by uti- Appendix B. Also, Table 3 indicates that substantial improvements are
lizing Ahuja et al. (1984)’s approach. Based on three soil databases, possible when bulk density is included in the predictive model, showing
they found results that contradict findings by Mbagwu and Auerswald again the role of porosity. Here we also remind the readers that the
(1999) and Lado et al. (2004). An explanation may be that the latter Lattice Boltzmann simulations in Section 3.1 indicated an imperfect
authors used repacked soil columns for their experiments, whereas relation between particle size and Ks – even for an idealized medium.
Nemes et al. (2005) used undisturbed soil cores. Nemes et al. (2005) While important, soil texture data is likely an insufficient predictor for
suggested that soil OC content retains water well and allows less water Ks. This may become especially relevant for highly structured soil, as
to flow, thus reducing Ks values. In addition, larger cracks and clods will was suggested by Wösten et al. (2001) and some tropical soils in which
be replaced by more aggregated material with increasing OC content, stable clay aggregates tend to act as a sand fraction (Tomasella and
leading to more tortuous and thin pathways for water to go. Hodnett, 1997).
In summary, models that rely on effective porosity are promising,
but empirical studies discussed above also indicate that the effects of 3.2.3. Soil structure information
bulk density and organic carbon affect the estimation of Ks. Further Soil structure pertains to the spatial arrangement of both soil voids
work is needed to gauge whether these are local or study-specific effects and solid constituents, i.e., soil particles and organic matter (Dexter,
or whether they can be generalized to a global scale. 1988). Soil structure can be used to reflect biological activity (earth-
worms and roots), abiotic factors (wetting-drying and freezing-thawing
3.2.2. Soil textural information process), or the effects of tillage practices and farmland operations in
Soil textural information includes soil textural classes, soil textural the soil. Romero-Ruiz et al. (2018) suggested that traditionally mea-
percentages (sand, silt, clay portions), and soil particle size distribution. sured bulk soil properties, such as bulk density or porosity, provide very
Since soil texture is easy to measure, it is used in almost all PTFs that limited soil information regarding soil structure. Fig. 6 shows three
estimate Ks. Unfortunately, different countries or regions use different different soils with similar soil texture and bulk density but with major
soil textural classification systems, such that defined by the USDA (The differences in soil structure. Due to differences in soil structure, dif-
United States Department of Agriculture), or the European HYPRES ferent preferential flow pathways exist which may result in significantly
(Hydraulic Properties of European Soil), SSEW (Soil Survey of England different soil hydraulic conductivity for the three panels in Fig. 6.
and Wales) from UK, and AISNE (French texture triangle of the Aisne O’neal (1949) is probably the first to suggest to use soil structure to
region soil survey) from France. Currently, the USDA classification is predict Ks, by stating:
the most extensively used system in the literature, with soil particle size
“In general, it would seem that structure is the most significant factor in
of sand, silt, and clay being within 0.05 and 2.00 mm, 0.002 and
evaluating permeability. But permeability cannot be correctly evaluated
0.05 mm, and less than 0.002 mm, respectively.
on the basis of type of structure alone… In some section, heavy textures
Probably the simplest method to express the relation between soil
go along with slow permeability, and light textures with rapid perme-
texture and Ks is by averaging observed Ks values for each textural class
ability. But, in the main, texture alone not a reliable clue….Permeability
(e.g. “sandy loam”, “silty clay”, etc.). This leads to simple tables of
cannot be evaluated on the basis of one characteristic alone”.
values (and perhaps standard deviations) also known as “class PTFs”,
such as those published by Clapp and Hornberger (1978), Cosby et al. O’Neal (1949, 1952) further suggested that the most critical step to
(1984), and Rosetta3 (Zhang and Schaap, 2017). Several of these determine permeability is to determine the structural type, followed by
lookup tables are listed in Table 2 and indicate that there are sub- the relative size of the horizontal and vertical cracks and fissures and
stantial differences among estimates for the same class; differences of other factors. He also suggested that climate condition is like to be a

1018
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Table 3
Overview of the PTFs used to estimate Ks. Evaluated quantities, including the mean error (ME), root mean squared error (RMSE), coefficient of determination (R2),
are from corresponding literature based on validation dataset unless otherwise stated. These quantities are calculated based on log10(Ks) and the unit of is cm/day,
unless otherwise noted herein.
Sources Database/Location No. of samplesa Method Predictorsb RMSE ME R2

Cosby et al. (1984)c United States 1448 linear regression SSC 0.839 ∼ 0.872
Ahuja et al. (1989)c,d Pima clay loam, etc. 473 regression soil effective porosity 0.613 0.669
Vereecken et al. (1990) Belgium 127 linear regression PCA of texture, OC, BD, 0.262
Schaap and Leij (1998a) UNSODA 315 neural network SSC 0.840
Schaap and Leij (1998a) UNSODA 315 neural network SSC, BD 0.775
Schaap and Leij (1998a) UNSODA 315 neural network SSC, BD, θ33 0.720
Schaap and Leij (1998a) UNSODA 315 neural network SSC, BD, θ10, θ33 0.713
Wösten et al. (1999)c HYPRES 1139 linear regression SSC, BD, OC, T/S 0.190
Schaap et al. (2001)e ROSETTA 1306 neural network Textural class 0.739 −0.001 0.427
Schaap et al. (2001)e ROSETTA 1306 neural network SSC 0.717 −0.001 0.461
Schaap et al. (2001)e ROSETTA 1306 neural network SSC, BD 0.666 0.000 0.535
Schaap et al. (2001)e ROSETTA 1306 neural network SSC, BD, θ33 0.586 −0.004 0.640
Schaap et al. (2001)e ROSETTA 1306 neural network SSC, BD, θ33θ1500 0.581 −0.002 0.647
Li et al. (2007) Fengqiu, China 36 linear regression SSC, BD, OC 0.58 ∼ 0.74
Lilly et al. (2008) HYPRES 502 regression tree Textural class, ped size class, T/S 0.951
Weynants et al. (2009) Belgium 136 linear regression SSC, BD, OC 0.250
Wang et al. (2012) Loess Plateau, China 382 linear regression SSC, BD, OC, altitude 0.400 0.010 0.400
Jorda et al. (2015) 85 articles 353 regression tree Land use, BD, tension seq, textural class 0.390f
Jorda et al. (2015) 85 articles 353 regression tree Land use, BD, tension seq, textural class 0.150 g
Tóth et al. (2015) EU-HYDIh 3206 regression tree Modified FAO texture class, T/S 1.36 0.13
Tóth et al. (2015) EU-HYDI 2669 regression tree Modified FAO texture class, T/S, OC 1.05 0
Tóth et al. (2015) EU-HYDI 3206 regression tree Texture class, T/S 1.39 −0.10
Tóth et al. (2015) EU-HYDI 2628 regression tree SSC, T/S, OC 1.06 −0.01
Tóth et al. (2015) EU-HYDI 401 linear regression SSC, T/S, pH, CEC 0.90 −0.10
Zhang and Schaap (2017)i ROSETTA 1306 neural network Textural class 0.739 0.001 0.424
Zhang and Schaap (2017) ROSETTA 1306 neural network SSC 0.733 0.002 0.437
Zhang and Schaap (2017) ROSETTA 1306 neural network SSC, BD 0.681 0.001 0.514
Zhang and Schaap (2017) ROSETTA 1306 neural network SSC, BD, θ33 0.610 0.005 0.610
Zhang and Schaap (2017) ROSETTA 1306 neural network SSC, BD, θ33, θ1500 0.612 0.004 0.608
Zhang et al. (2018)i ROSETTA 1306 neural network Textural class 0.739 0.001 0.424
Zhang et al. (2018) ROSETTA 1306 neural network SSC 0.7336 0.0037 0.436
Zhang et al. (2018) ROSETTA 1306 neural network SSC, BD 0.6815 0.0030 0.513
Zhang et al. (2018) ROSETTA 1306 neural network SSC, BD, θ33 0.6134 0.0055 0.606
Zhang et al. (2018) ROSETTA 1306 neural network SSC, BD, θ33, θ1500 0.6152 0.0054 0.604

a
This is the total number of model development, including calibration and validation data.
b
Abbreviations of predictors: SSC: sand, silt, clay. BD: bulk density. OC: organic carbon content. PCA: principal component analysis. θ10: water content at 10 kPa.
θ33: water content at 33 kPa. θ1500: water content at 1500 kPa. T/S: a topsoil/subsoil distinction. Tension seq: direction of tension sequence, i.e., from dry to wet or
from wet to dry. Textural class: USDA (United States Department of Agriculture) soil textural class, unless otherwise stated. FAO: Food and Agriculture Organization
of the United Nations.
c
Only calibration quantity was found.
d
The unit is cm/hour.
e
Validation results are not provided, and model performance was based on the overall dataset because the bootstrap procedure ensured robust models that
provided similar results for calibration and validation
f
Result is from 10-fold cross validation; the unit is mm/h.
g
Result is from source-wise cross validation; the unit is mm/h.
h
EU-HYDI: European Hydropedological Data Inventory.
i
Model performance was based on the entire dataset, because bootstrapping was not applied to this PTF.

useful indicator of soil structure development, as also indicated by measure, although structures such as the plant roots and visible pores,
Tomasella and Hodnett (1997). However, soil structural effects on Ks are often documented in field descriptions (Vereecken et al., 2010).
have not received widespread attention and recognition in the estima- Quantification of this property and standardization of measurement
tion of Ks. McKeague et al. (1982) suggested that high Ks values were techniques will help to derive continuous PTFs by including soil
predominantly controlled by soil biopores based on 78 soil horizons structure as input, the same as the other quantitative predictors, such as
from North America. We note here that the Ks for the clay textural class soil texture (Lilly and Lin, 2004), which will be essential in quantifying
is higher than other classes for several of the class-PTFs listed in large-scale fluxes and climate in the application of land surface mod-
Table 2. This may point to the effects of soil structure causing the Ks of eling and global climate modeling.
fine textures (e.g., clay) to be higher than some textures with more We also point out that there are (at least) some conceptual simila-
coarser particles. Work that investigates the effects of soil structure on rities between the structured soils displayed in Fig. 6 and the CFD si-
Ks includes Wang et al. (1985), Coen and Wang (1989), Hollis and mulations in Fig. 3. Provided that relevant macro and micro-pore re-
Woods (1989), Griffiths et al. (1999), and Lilly (2000), and Gamie and presentations of structured soil can be obtained (see Larsen, 2018), it
De Smedt (2018). may be possible to also use CFD to resolve the effect of soil structure on
Predictors described in Sections 3.1 and 3.2 have been widely used saturated flow in soils and other dual porosity domains (e.g., see Sukop
to predict Ks. However, with the exception of Lilly et al. (2008), soil et al., 2013).
structure is usually missing in most PTFs. The main reason for this is
that soil structure (e.g., the size, shape, quantity, and continuity) are
generally less well recorded, difficult to quantify, and expensive to

1019
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 6. Schematic illustration of soil structure along a transect. Each subgraph indicates a different soil structure: (a) Homogeneous soil. (b) The same soil with
secondary structure from biological activity. (c) Soil structure from (b) but damaged by compaction. From Romero‐Ruiz et al. (2018).

4. Datasets and statistical techniques available for PTF are available (e.g. UNSODA 2.0, Nemes et al., 2001; HYPRES, Wösten,
development 2000; and Grenoble soil catalog (GRIZZLY, Haverkamp et al., 1997))
were assembled from literature searches or other data-gathering efforts.
4.1. Calibration and evaluation data As a result, the large databases often contain data that “happen to be
available” without a guarantee that the data are representative for a
PTF development typically aims to estimate difficult-to-measure soil target population of soils. In many cases, the availability of hydraulic
properties by establishing relations between accessible input data and properties is the limiting factor with, in the experience of the authors,
estimands. In order to achieve this, the PTF developer must have access an overrepresentation of sandy soils, presumably because hydraulic
to data sets that contain both observed hydraulic data as well as the measurements are easier in these soils. Large pedological databases
predictors. The requirements that these data sets need to meet depends (such as NCSS and WoSIS) are more representative of the global po-
on the objectives being pursued. If the aim is merely to identify which pulation of soils but are of somewhat limited use for PTF development
variables affect Ks (e.g., in our numerical example in Section 3.1), dif- because of limited retention data and the unavailability of observed Ks
ferent data may be needed than in an effort where the primary objective values. Rahmati et al. (2018) recently published soil water infiltration
is to develop a PTF that is representative for a population of soils. In the global database (SWIG), containing 5023 infiltration curves with a
former case, a small data set with limited variability may suffice, but in subset of 1896 Ks dataset from 54 different countries with broad cov-
the latter case, a much larger dataset would be needed to cover the erage of soil types and climate regions. These datasets are valuable for
variability relevant for a population of soils. This is especially important the development and evaluation of PTFs for saturated hydraulic con-
when data-driven approaches (e.g., artificial neural networks, as well as ductivity. However, as was pointed out by Wösten et al. (2001),
some more traditional statistical techniques) are used to develop PTFs. Pachepsky and Rawls (2004), and Vereecken et al. (2010) that soil
In a hypothesis-driven approach (e.g., the Kozeny-Carman model as structural properties, support scale, and other soil information are also
modified by Ahuja et al., 1984), the model and its input variables are critical for the development of PTFs and this information is usually
known a priori and model coefficients are obtained in a calibration missing from most databases.
procedure. In a data-driven approach, however, the model, its coeffi- The set of data used for PTF development is often much smaller than
cients as well as the relevant input variables are found in the calibration the independent national datasets or worldwide collections of soil
(or ‘training’) procedure. This requires that the calibration data set is survey data, such as the NCSS (National Cooperative Soil Survey, 2017)
representative (and covers all the variability) of the soils to which the and WoSIS (Batjes et al., 2017) that are useful for PTF evaluation. This
model will be applied. can be problematic because different calibration databases lead to
The authors know of few large databases that were established with systematically different PTFs (Schaap and Leij, 1998b). Although the
the primary objective of PTF development. Most large databases that combined dataset had a larger support and resulted in the Rosetta PTF

1020
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 7. Soil dataset in USDA soil textural triangle. (a) Rosetta dataset. (b) NCSS dataset. (c) WoSIS dataset. (d) Second layer (5–15 cm) of SoilGrid 1 km dataset. Red
color represents high density, blue low density. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this
article.)

(Schaap et al., 2001), there is no guarantee that this model is “uni- organic carbon content) and estimand (Ks). Because of their general
versal” in a global sense. Evidence for this appears in Fig. 7a which empirical nature, the methods available to derive PTFs are extremely
shows that the textural distribution of Rosetta calibration dataset is diverse. For a detailed overview, we refer to Van Looy et al. (2017),
significantly different from the distribution of NCSS and WoSIS datasets who reviewed a variety of the available methods that have been
(Fig. 7b and c, respectively). Sandy soils are well represented in Rosetta (widely) used to establish PTFs, including lookup table, regression
database, while silty clay loam soils have higher densities in the NCSS techniques, artificial neural networks, support vector machines, k-
dataset, and the frequencies of silty clay loam, sandy soils are high in nearest neighbor methods, decision/regression trees, and the random
WoSIS dataset. forest method. There is still a demand to develop and apply new sta-
A related problem is that, although they are georeferenced, the tistical techniques based on existing or newly obtained databases. Many
comprehensive NCSS and WoSIS databases cannot provide the soil statistical packages, such as the open-source programming language R
properties at any arbitrary global location. Additional geospatial tech- and open source software libraries such as Google’s TensorFlow, allow
niques are needed to produce gridded maps of soil properties, such as the relationship between input and output data to be mined with in-
SoilGrids (Hengl et al., 2014) and Global Soil Dataset for ESMs (GSDE, creased efficiency.
Shangguan et al., 2014). Even though SoilGrids relies on data from the Advanced techniques also bring along the temptation to increase
NCSS and WoSIS datasets, Fig. 7d indicates that its textural distribution model complexity, which may be able to find patterns in data that are
of the 1 km2 data is more limited than either of those databases: ex- hidden to simple models. However, more complex models usually re-
treme silt and clay soil samples are missing. This effect is found for all quire more parameters, which may lead to overfitting problems. In
seven layers of SoilGrids data set. The possible explanation for the these cases, model calibration latches on to noise in the calibration data
discrepancy is that the kriging interpolation that was used by Hengl (e.g., see Fig. 4 in, Schaap and Bouten, 1996). The resultant model may
et al. (2014) smoothed out the extreme values of the original point perform surprisingly well on the calibration data, but poorly on in-
dataset. A less likely competing explanation is that loam soils are more dependent data (because noise or other data artifacts were incorporated
common globally than represented by the NCSS and WoSIS data. in the model). Splitting the dataset into calibration and validation is
The implication of differences in the distribution among develop- useful for determining whether the model is overfitted by using cali-
ment and evaluation datasets is that PTF estimates are not necessarily bration dataset for fitting and validation dataset for testing; most sta-
representative of the true global distribution of soil texture, and other tistical packages now have facilities to ensure proper model general-
relevant variables. This may produce a bias for practical applications at ization. In addition, advanced model selection criteria, such as AIC,
both pedon and global scales. Care must be taken when using PTFs in a AICc, BIC, and KIC (e.g. Ye et al., 2008) can be used to rank and select
different environment compared with the conditions of PTF develop- models that provide an optimal balance between predictive capabilities
ment (Schaap and Leij, 1998a). and model complexity (Hastie et al., 2001; Ye et al., 2008).
Finally, Wösten et al. (2001) and Minasny and Hartemink (2011)
4.2. Statistical techniques used to develop PTFs suggest the importance of reliable data rather than more complex
models, since most PTFs were developed using a data-driven approach
A wide variety of statistical or mathematical techniques is available and therefore are heavily dependent on high-quality and reliable da-
to establish the relationship between predictors (texture, density, tasets. Accurate measurement of soil hydraulic conductivity is probably

1021
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

N
the most important factor that controls the accuracy and reliability of
∑ [log10 (Ki ) − log10 (K ′i )]2
PTFs. The authors would also like to stress that all data that were used i=1
R2 = 1 −
for the development should be published along with the PTFs. Although N

open data-repositories are currently more or less a requirement for ∑ [log10 (Ki ) − log10 (K )]2
i=1 (11)
publication in many journals and data used for PTF development is
available, this was not the case for some models that were established where K is the average of K .
several decades ago. Several PTFs that remain in widespread use have a Table 3 provides the error metrics for a selection of PTFs found in
data provenance that is either uncertain or hard or impossible to trace. the literature for which these are available. Here, reliability results
(validation values) are shown if both accuracy (calibration values) and
4.3. Accuracy and reliability of PTFs reliability (validation values) quantities are listed in the original lit-
erature. We found that most publications listed both calibration and
When considering the use of PTFs, several straightforward but cri- validation results and few showed only calibration results. The number
tical questions can be raised: should a new PTF specific for a population of soil samples used to develop PTFs is also listed and ideally should be
of soils be developed or should one (or more) PTFs be selected from the large and representative for the soils to which the PTFs is being applied.
literature? If so, which ones? What is the accuracy of the PTFs? Are Overall, the RMSE values vary considerably among PTFs shown in
predictions by PTFs reliable? Table 3, and range from 0.53 to 1.39 from PTFs respectively from Li
To answer these questions, the differences between estimated and et al. (2007) and Tóth et al. (2015). However, in general, we cannot
measured values must be quantified. The term accuracy in this regard is conclude that the PTF in Li et al. (2007) will outperform the PTF of Tóth
used to quantify the correspondence between estimated and measured et al. (2015). This is because Li et al. (2007) only used 36 samples from
values during PTF calibration, whereas reliability is used to determine a small area in China (Fengqui county) for calibration and validation.
the correspondence between estimates and measurements on in- By contrast, Tóth et al. (2015) used a significantly larger number of soil
dependent data (i.e., data not used for PTF calibration). This latter step samples (3206, from the European Hydropedological Data Inventory,
is often called validation which can be done independently from cali- EU-HYDI), with a much wider soil coverage in continental scales. This
bration or be integrated into the calibration procedure using techniques assessment is confirmed by Wang et al. (2012), who validated pub-
such as k-fold cross-validation, bootstrap, leave-one-out cross-valida- lished PTFs (Rosetta, Schaap et al., 2001; Wösten et al., 1999; Li et al.,
tion, and the Jackknife (Hastie et al., 2001). 2007), and site-specific PTFs (based 252 soil samples) developed for the
The accuracy of PTFs varies considerably; the reason for the varia- Loess Plateau of China. By utilizing 130 additional samples, they found
bility may result from the data and methodology used for PTF devel- that R2 for the site-specific PTF, Rosetta, Wösten et al. (1999) were
opment. A PTF can make accurate estimates for calibration data, but 0.400, 0.305, 0.272, respectively, while the value for Li et al. (2007)
estimates may not be reliable when the PTF is applied to independent was 0.001, indicating no correlation between estimated Ks and mea-
data, especially when the characteristics of the input data differ sig- sured data, although the data used to develop the Li et al. (2007) PTF
nificantly from the calibration data. The accuracy of estimation can were also from China.
serve as benchmark values for reliability assessment because it is un- ME values were provided only in Schaap et al. (2001), Tóth et al.
likely that a PTF will perform better on data that was not used for ca- (2015), Zhang and Schaap (2017) and Zhang et al. (2018). ME values
libration. Apart from the fact that the distribution of input data may range from −0.1 to 0.13 in the work of Tóth et al. (2015), while Schaap
differ from that of the calibration data, the reliability of PTFs may fur- et al. (2001), Zhang and Schaap (2017) and Zhang et al. (2018) showed
ther depend on factors that were not considered as predictors, such as that ME values range from −0.004 to 0.0055, all of which are based on
soil characteristics, climate conditions, landscape features, and geo- the Rosetta database. The differences between the ME values might
graphical regions of the soils. For example, PTFs developed for soils in result from the database they used. It is also noted that positive and
temperate climate regions may produce significant errors for tropical negative residuals will cancel out, and the results can be misleading. A
soils (Minasny and Hartemink, 2011). PTF may have substantial positive and negative ME for sub-populations
Donatelli et al. (2004) and Schaap (2004) have comprehensively of the data. To properly characterize bias in PTF estimates, it may
reviewed various criteria for assessing the accuracy and reliability of therefore be better to stratify the evaluation data in relevant classes
PTFs. Among them, the most commonly used indicators for evaluating (e.g., by textural class) and calculate ME values for each.
model performance are probably root mean square error (RMSE), mean In terms of R2, the values range from 0.150 to 0.872, which are
error (ME), and coefficient of determination (R2). The RMSE is defined respectively from the work of Jorda et al. (2015) and Cosby et al.
as (1984). Cosby et al. (1984) employed a large number of soil samples
N
(with a total of 1448) with a wide coverage from 23 states in the United
1 States. However, Cosby et al. (1984) provided only calibration results
RMSE =
N
∑ [log10 (Ki′) − log10 (Ki)]2
i=1 (9) and the performance outside the calibrated dataset should be in-
vestigated. The lowest R2 value in Table 3 (0.150) is from Jorda et al.
where Ki and Ki′ indicate measured and estimated saturated hydraulic (2015), which utilized 353 soil samples from 85 different peer-reviewed
conductivity Ks. The subscript s in Ks was removed for brevity; N is the articles, with diverse land use and soil samples being measured through
number of measurement for Ks. different methods and devices. However, it is worth noting that Jorda
The ME is used to assess systematic bias of estimations relative to et al. (2015) used a different validation method called “source-wise
the measurements of Ks, defined as cross validation”, which indicated that the dataset used for validation
N are not from any data in the same studies from calibration dataset
1
ME =
N
∑ [log10 (Ki′) − log10 (Ki)] within the same sampling loop. Since no single location was included in
i=1 (10)
more than one study from the peer-reviewed articles collected in Jorda
ME is negative value when the PTF underestimates Ks on average, et al. (2015), this work guaranteed the soil samples from the same study
positive if overestimation. Log-10 transformations are used here be- will not be included in calibration and validation dataset simulta-
cause Ks can vary several orders of magnitude (e.g., Table 2). Un- neously, which is stricter than the commonly used K-fold cross-vali-
transformed use of Ks would yield error metrics that are biased to high- dation and bootstrapping methods. The latter methods split the data set
conductivity soils. Because of the log-10 transformations, the RMSE and into calibration and validation, with the possibility of correlated data in
ME values are dimensionless (since log(a) − log(b) = log(a b) ). calibration and validation datasets. When tested using 10-fold cross
R2 is defined as validation in Jorda et al. (2015), R2 was increased to 0.39, which

1022
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

suggested that K-fold cross-validation and bootstrapping sampling 1980).


might over-estimate the predictive performance of developed PTFs, if
calibration and validation are not completely independent. These re-
sults indicate that we might be overly-optimistic about the predictive 5.2. Texture-based PTFs
reliability of PTFs. Likewise, Vereecken et al. (2010) suggested that the
validation of PTFs is best carried out by utilizing totally independent Six soil texture-based PTFs were used to produce global maps of Ks
dataset rather than validation based on splitting the dataset into cali- using the Soilgrids 1 km dataset from Hengl et al. (2014). Contrary to
bration and validation. the previous section, which used the Soilgrids data set indirectly
Hierarchical PTFs shown in Table 3, such as the work of Schaap (through effective porosities estimates validated on NCSS data), this
et al. (2001) with RMSE values decreased from 0.739 to 0.581 and R2 section applies six PTFs directly to the global Soilgrids data. Four of the
increased from 0.427 to 0.647, suggested that the performance of PTFs six PTFs used only soil texture data (i.e., textural class or sand, silt and
can be increased with the increasing of predictors, which will facilitate clay percentages: Clapp and Hornberger (1978), Rawls et al. (1982),
the practical applications in data-poor to data-rich situations. Similar Cosby et al. (1984) lookup table, Carsel and Parrish (1988)), while the
hierarchical PTFs were published by Schaap and Leij (1998a), Tóth Rosetta3 model (Zhang and Schaap, 2017) also used bulk density as
et al. (2015), Zhang and Schaap (2017) and Zhang et al. (2018). input. The Weynants et al. (2009) PTF used bulk density and organic
In summary, it is problematic to compare different PTFs based on carbon content. Some well-known PTFs were excluded from this ana-
published error metrics. Small and homogeneous databases tend to lysis because they produced unreasonable values. For example, Wösten
yield better error metrics than large datasets that include significant et al. (1999) utilized the terms of 1/silt, log(silt), and 1/organic matter
variability in soil types, porosity, texture, and structure. Given the in- in the equations to calculate Ks values, and we found that silt and or-
creased need for detailed global distribution of soil hydraulic properties ganic matter (OM) data can be zero in the Hengl et al. (2014) data,
to meet a variety of modeling and evaluation goals, comprehensive leading to infinite values for the estimated Ks.
evaluation of PTFs on a consistent and well-documented database with Generally, there are considerable variations among Ks values esti-
PTF input and observed Ks values would be both timely and useful. mated from different PTFs. Ks values estimated by Weynants et al.
(2009) PTF (Fig. 9f) are significantly smaller than the estimates from
5. A comparison of global estimates of saturated hydraulic the other PTFs. The dataset used to develop the Weynants et al. (2009)
conductivity PTF was limited to Belgium and is therefore not representative of the
global distribution of soils. Ks values obtained from PTFs from Clapp
It is currently difficult to objectively evaluate how reliable Ks esti- and Hornberger (1978), Rawls et al. (1982), Cosby et al. (1984), Carsel
mates by PTFs are at a global scale because global-scale data currently and Parrish (1988) all used only soil textural information as input,
are lacking. While large databases with substantial global coverage while soil bulk density information are not included, which might
such as NCSS and WoSIS contain observed retention data, they lack produce biased results because bulk density is related to porosity as
observed Ks values. No definitive statements can be made which PTF (or discussed in Section 3.2. Arya et al. (1999) suggested that similarities of
combination of PTFs) provides the best estimates at a global scale. soil textural data do not necessarily indicate similarities of soil hy-
However, the relative agreement among PTFs for estimating Ks can be draulic properties, because significant variability in bulk density and
tested by evaluating these on a global data set, for which we will use the organic matter can be present for the same texture.
previously discussed gridded data set of Hengl et al. (2014). Two ap- Compared with the effective porosity-based ensemble model shown
proaches are followed: an indirect approach which relies on estimated in Fig. 8a, the Rosetta3 model (which uses texture and bulk density as
effective porosity, and a second approach which evaluates six PTFs predictors) produces much higher Ks values in the mid- to high latitude
directly from particle size, and other soil data. of the Northern Hemisphere (Fig. 9a). PTFs that use only soil textural
information produce lower Ks values in these regions. Compared with
5.1. Effective porosity Cosby et al. (1984) PTF, Clapp and Hornberger (1978), and Rawls et al.
(1982) PTFs, the work of Carsel and Parrish (1988) produced higher Ks
In work that is currently under review, Zhang et al. (2019) (see the values (warm color) in Eurasia and North America. The Carsel and
preprint version in arXiv) provided a high-resolution global map of Parrish (1988) PTF generated the highest Ks values in the Southern
effective porosity by utilizing multi-model ensemble of 13 widely-used Hemisphere, especially in Australia and South America.
water retention PTFs based on NCSS dataset and using the Soilgrids 10 Mean values of Ks obtained by averaging the six PTFs produce
km dataset from Hengl et al. (2014). An evaluation of NCSS data in- smoothed-out results (Fig. 9g), with consequently less severe and ex-
dicates that the estimated effective porosity by the ensemble is more treme values. High CV values for Ks (Fig. 9h) are found in west India,
accurate than individual PTFs that are currently being used in earth Gabon and southeast of Ethiopia in the African continent, southeast of
system models. By utilizing the Ahuja PTF (1984, Eq. (8) in this study), Brazil, Iowa and Missouri states in the US, and some areas in Canada
global maps of saturated hydraulic conductivity and the corresponding and Siberia. High CV values suggest significant different estimates ob-
coefficient of variation (CV) values are produced in Fig. 8. These CV tained from different PTFs and might serve as a reference to conduct
values represent the agreement among estimated Ks values derived future PTF improvement.
from effective porosity estimated from the 13 PTFs.
High Ks values in Fig. 8a are mainly found in desert regions, such as
the Sahara and Arabian deserts, and high latitude north of 50°N, such as 6. Challenges and perspectives for PTFs to estimate saturated
northeast of North America, which is likely due to high sand content in hydraulic conductivity
the desert and high soil OC content at high latitudes. Low Ks values can
be found in India, southeast to the Sahara, where clay soil prevails. In this review, we have introduced the history, predictors, and
However, CV values at the low Ks regions turn out to be high, with the statistical techniques for PTF development of saturated hydraulic con-
values in India exceeding the color scale limits used in Fig. 8b. High CV ductivity, and summarized the accuracy and reliability of existing PTFs.
values suggest significant disagreement of effective porosity estimated Some issues and areas that might improve the PTF development have
among individual PTFs, which might help for future PTF development been highlighted. Here we summarize the challenges and perspectives
that more attention should be focused on clay soils. This global map is for further improving PTFs to estimate saturated hydraulic con-
promising and provides a way of moving forward to estimate un- ductivity.
saturated hydraulic conductivity (Mualem, 1976; van Genuchten,

1023
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 8. Global maps of (a) mean values and (b) coefficients of variation (CV) of log10(Ks) estimated from effect porosity (macroporosity) in 10 km resolution from
overall 13-PTF ensemble model employed in Zhang et al. (2019). Weights for different PTFs are based on the average values of 100 bootstrap replicas of optimized
weights. Calculations are based on the surface soil of SoilGrids 10 km data set (Hengl et al., 2014).

6.1. Increasing the reliability of measurement of saturated hydraulic magnetic resonance imaging (Rabot et al., 2018). However, there is no
conductivity unified protocol to characterize soil structure, and it is critical to
identify agents and find proxies to quantify soil structure.
There are diverse methods and devices for measuring saturated Remote sensing products are revolutionary and provide valuable
hydraulic conductivity in laboratory and in-situ, such as constant-head information for PTF development. Employing the wealth of remote
well permeameters, pressure and tension infiltrometers, single/double sensing products, such as Moderate Resolution Imaging
ring infiltrometers, undisturbed soil core methods (Morbidelli et al., Spectroradiometer (MODIS), Advanced Land Observing Satellite
2017). These methods have different systematic errors and biases and (ALOS), sentinel-2, Landsat, MEdium Resolution Imaging Spectrometer
produce inconsistent and sometimes conflicting results, which should (MERIS), will provide information, such as the vegetation types and
be considered in PTF development. Standard methods, procedures, and dynamics, leaf area index (LAI). With further processing, these ob-
protocols for measuring saturated hydraulic conductivity will be ben- servations may yield further soil and landscape information by which
eficial and improve the development of PTFs. PTFs can be improved or validated.
Other useful information and possible predictors include land use,
6.2. Including relevant predictors tillage activity, soil horizon, and direction of tension sequence (dry-to-
wet or wet-to-dry), which will be beneficial to build new data infra-
We have summarized the predictors that have been extensively used structure and improve PTF development.
in PTF development in Section 3. For increasing the predictive power
and estimation accuracy, more materially relevant predictors should be 6.3. New statistical techniques and ensemble model
included. However, this requires that such predictors are available
when the PTFs are used, which may limit their general applicability due With the increasing accessibility of “big data” relevant to the soil
to lack of required input data. sciences and the breakthrough advances being developed in machine
Considering the importance of soil structure as predictors, imaging learning, we see great opportunities in enhancing our physical under-
techniques are promising and attractive non-invasive methods to in- standing and insights, new scientific discovery, and predictive power.
vestigate the details of soil structure. These techniques may also yield Statistical techniques, such as linear regression, neural networks,
additional insight when combined with CFD methods. Commonly used support vector machines, regression trees, and random forest, have
approaches include optical or electron thin-section microscopy, X-ray been widely used for the development of PTFs. Other statistical tech-
tomography, gamma-ray and neutron tomography, and nuclear niques, such as Gaussian Process (Ju et al., 2018; Raissi et al., 2017),

1024
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Fig. 9. Global map of saturated hydraulic conductivity (log10(Ks)) in 1 km resolution estimated from (a) Rosetta3 H3w model (Zhang and Schaap, 2017) by em-
ploying soil textural percentage and bulk density data as input. (b) Clapp and Hornberger (1978). (c) Rawls et al. (1982). (d) Cosby et al. (1984) lookup table. (e)
Carsel and Parrish (1988). (f) Weynants et al. (2009) models. (g) Mean values. (h) Coefficient of variation of log10(Ks) are also shown. Calculations are based on the
surface soil of SoilGrids 1 km data set (Hengl et al., 2014).

Naive Bayes Classifiers (Frank et al., 2000; Tsangaratos and Ilia, 2016), many stochastic applications but also point out soil types or geographic
have -to our knowledge- not been explored in PTF development and areas that deserve more detailed attention from future PTF develop-
might be worthwhile investigating. Recently developed deep learning ment.
algorithms such as convolutional neural network (Krizhevsky et al.,
2012; Mo et al., 2019), deep neural network (Silver et al., 2016), a
currently hot topic in artificial intelligence, seem to be very promising 6.4. Data assimilation and inverse modeling
in this field.
Another technique that needs to be explored is multi-model en- Despite the fact that extensive development of PTFs using various
semble prediction. Ensemble model combines different independent predictors and more advanced machine learning methods, we should be
models, which are consists of various data, theory, methodology, and aware that the accuracy and reliability of Ks estimates is modest. Data
model structure. Hagedorn et al. (2005) suggested that a single model assimilation and inverse modeling are critical and valuable to improve
may outperform in some cases, but the overall ensemble model will be the estimation of Ks and moisture content by conditioning on field
more reliable in the long term. The overall accuracy of the ensemble observed system variables, such as moisture content and capillary
model depends much more on how individual models are combined, pressure (Pachepsky et al., 2014; Pan et al., 2012; Shi et al., 2015;
i.e., by assigning proper weights of each model (Figs. 9g and h provided Zhang et al., 2016; Man et al., 2016; Ju et al., 2018; Yu et al., 2018;
unweighted ensemble values for Ks and its CV). New national and in- Zheng et al., 2019). We suggest performing inverse modeling or data
ternational computing and data infrastructures facilitate the develop- assimilation if observed state variables are available.
ment and evaluation of ensemble models at a global scale. In turn,
uncertainty estimates produced by ensemble models can be useful in

1025
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

6.5. Application in land surface models and global climate models simulations. Different predictors were systematically reviewed in the
context of the Kozeny-Carman equation in soil column scales, whereas
Traditional measurement and representation of Ks are carried out at we emphasized soil structural information for future PTF development.
laboratory soil column scale and field scales. However, land surface By utilizing readily available soil mapping products of basic soil prop-
models and global climate models require the parameters in kilometer erties, global maps (and associated uncertainty) of saturated hydraulic
and even larger scales, which poses challenges for the estimation by conductivity were obtained. We highlighted that accuracy and relia-
PTFs for these applications. bility of PTFs for estimating Ks should be carried out based on large and
Saturated hydraulic conductivity is known to vary with scales completely independent databases. PTF development on small and
(Pachepsky and Hill, 2017; Pachepsky and Rawls, 2004; Vereecken correlated databases often over-estimate the reliability of PTFs.
et al., 2010). Neuman (1994) suggested that log permeability of porous To improve data quality and PTF development, we suggest con-
medium can be considered as multiscale random functions over a sistent methods and protocols for measuring saturated hydraulic con-
continuum. He demonstrated that the logarithm of effective perme- ductivity. Given the many factors that may influence the Ks estimation,
ability varies as a power of the characteristic length, with the latter we suggest including relevant predictors (including soil structure)
being treated as the support scale of a permeability test for geologic should be selected; employing new statistical techniques and ensemble
media. The data used to develop PTFs for saturated hydraulic con- models for PTF development should be employed. Data assimilation
ductivity, therefore, should be scale-consistent, e.g., all at the labora- and inverse modeling are valuable and promising to improve the esti-
tory soil-column scale. Because Ks is scale dependent by nature, the mation of Ks by conditioning on available soil state variables, such as
scaling effects of saturated hydraulic conductivity should be considered soil moisture content. Saturated hydraulic conductivity in large scale
in the application of PTFs used in catchment, regional, and global simulations should be evaluated and accessed by utilizing numerical
scales. experiments such as in land surface models and global climate models
It is difficult to measure Ks at kilometer-scales and even harder to via assessing observed state variables.
quantify the associated error and bias. The parameters, therefore,
should be evaluated and accessed by utilizing numerical experiments Declaration of Competing Interest
such as land surface models, numerical weather forecast models, and
global climate models by quantifying the hydrological processes, ve- The authors declare that they have no known competing financial
getation feedbacks, surface energy, and global climate responses. interests or personal relationships that could have appeared to influ-
ence the work reported in this paper.
7. Conclusions The authors declare the following financial interests/personal re-
lationships which may be considered as potential competing interests:
Saturated hydraulic conductivity (Ks) is crucial to many Earth sys-
tems science applications. Obtaining this soil property heavily depends Acknowledgments
on soil pedotransfer functions (PTFs) by utilizing easily measurable
basic soil properties. In this review, we have given an overview of the Yonggen Zhang thanks the National Natural Science Foundation of
historical background of deriving and utilizing saturated hydraulic China (grant 41807181); Marcel G. Schaap was supported by the USDA
conductivity. By using computational fluid dynamics model based on National Institute of Food and Agriculture under Multistate Research
synthetic data, we showed how permeability (extendable to saturated Project W3188, entitled “Soil, Water, and Environmental Physics Across
hydraulic conductivity) is affected by different predictors in pore-scale Scales”.

Appendix

Equations used to calculate saturated hydraulic conductivity (KS)

This section provides equations selected from literature to calculate Ks. The symbols for sand, silt, clay are textural percentages (%) based on
USDA soil textural classification; OC and OM are soil organic carbon content and soil organic matter content, respectively; BD is soil bulk density (g/
cm3); ϕs is soil effective porosity, which is defined as the total porosity minus field capacity (volumetric soil water content at 330 cm pressure head).
Dai et al. (2013) and the report of Guber and Pachepsky (2010) are extensively referenced in this work.

A. Estimation of Ks based on effective porosity

1) Ahuja et al. (1989)


Ks = 764.5φe 3.288 (A1)
The unit for Ks is cm/hour in Ahuja et al. (1989). If unit cm/day is used, the coefficient 746.5 will be 18,348.

2) Rawls et al. (1998)


Ks = 1930φe 3 − λ (A2)
where λ is pore size distribution index in Brooks and Corey (1964) water retention model. Eq. (A2) was originally given in Rawls et al. (1998) with
the unit of mm/hour. If unit cm/day is used, the coefficient 1930 will be 4632.

3) Timlin et al. (1999)


Ks = 2.59 × 10−4 × 100.6λφe 2.54 (A3)
Eq. (A3) was originally given with the unit of m/s. The coefficient 2.59 × 10−4 will be 2237.76 if unit cm/day is used.

1026
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

4) Suleiman and Ritchie (2001)


Ks = 37.15φer 1.92 (A4)

Ks = 12302φe 3.63 (A5)


where φer is relative effective porosity, which is defined as the ratio of soil effective porosity (φe ) and field capacity. Suleiman and Ritchie (2001)
suggested that Eqs. (A4) outperformed (A5) when tested on field measured Ks.

B. Estimation of Ks based on soil texture, bulk density, etc.

(1) One independent variable to estimate Ks in Cosby et al. (1984)


Ks = 100.0153 × sand − 0.884 (B1)
The unit for Ks is inches/hour in Cosby et al. (1984). If unit cm/day is used, the equation will be
Ks = 60.96 × 100.0153 × sand − 0.884 (B2)

(2) Two independent variables to estimate Ks in Cosby et al. (1984)


Ks = 60.96 × 100.0126 × sand − 0.0064 × clay − 0.6 (B3)

(3) Wösten et al. (1999)

Ks
= exp[7.755 + 0.0352 × silt + 0.93 × topsoil − 0.967 × BD 2 − 0.000484 × clay 2 − 0.000322 × silt 2 + 0.001 silt − 0.0748 OM − 0.643 × ln(silt )
− 0.01398 × BD × clay − 0.1673 × BD × OM + 0.02986 × topsoil × clay − 0.03305 × topsoil × silt ] (B4)
where topsoil and subsoil are qualitative variables having the value of 1 or 0, respectively.

(4) Saxton and Rawls (2006)


θ33 = θ33t + [1.283(θ33t )2 − 0.374θ33t − 0.015]

θ33t = −0.251 × sand − 0.195 × clay + 0.011 × OM + 0.006(sand × OM ) − 0.027(clay × OM ) + 0.452(sand × clay ) + 0.299

θ1500 = θ1500t + (0.14 × θ1500t − 0.02)

θ1500t = −0.024 × sand + 0.487 × clay + 0.006 × OM + 0.005(sand × OM ) − 0.013(clay × OM ) + 0.068(sand × clay ) + 0.031

θS - 33 = θ(S − 33) t + (0.636θ(S − 33) t − 0.107)

θ(S − 33) t = 0.278 × sand + 0.034 × clay + 0.022 × OM − 0.018(sand × OM ) − 0.027(clay × OM ) − 0.584(sand × clay ) + 0.078

θS = θ33 + θS − 33 − 0.097 × sand + 0.043

λ = [ln(θ33) − ln(θ1500)] [ln(1500) − ln(33)]

Ks = 1930(θs − θ33 )3 − λ (B5)


It is noted that sand and clay data should use a decimal value, and OM uses percentage value. For example, the input for sand, clay, and OM is
0.88, 0.05, and 2.5 if sand, clay, and OM content is 88%, 5%, 2.5%, respectively. The output unit for Ks is mm/hour; the coefficient 1930 will be
4632 if unit cm/day is used.

(5) Weynants et al. (2009)


Ks = exp(1.9582 + 0.0308 × sand − 0.6142 × BD − 0.01566 × OC ) (B6)

(6) Schaap et al. (2001), Zhang and Schaap (2017), and Zhang et al. (2018)’s work to estimate Ks based on artificial neural networks combined with
the Bootstrap method (Efron and Tibshirani, 1993). These models contain a large number of coefficients that cannot reasonably be represented
here. The models (in python scripting language) can be downloaded from http://www.u.arizona.edu/~ygzhang/download.html.

References Alexander, L., Skaggs, R.W., 1987. Predicting unsaturated hydraulic conductivity from
soil texture. J. Irrig. Drain. Eng. 113, 184–197.
Amundson, R., Berhe, A.A., Hopmans, J.W., Olson, C., Sztein, A.E., Sparks, D.L., 2015.
Adams, W.A., 1973. The effect of organic matter on the bulk and true densities of some Soil and human security in the 21st century. Science 348, 1261071.
uncultivated podzolic soils. J. Soil Sci. 24, 10–17. Araya, L.M., Paris, J.F., 1981. A physicoempirical model to predict the soil moisture
Ahuja, L.R., Naney, J.W., Green, R.E., Nielsen, D.R., 1984. Macroporosity to characterize characteristic from particle-size distribution and bulk density data. Soil Sci. Soc. Am.
spatial variability of hydraulic conductivity and effects of land management. Soil Sci. J. 45, 1023–1030. https://doi.org/10.2136/sssaj1981.03615995004500060004x.
Soc. Am. J. 48, 699–702. Arya, L.M., Leij, F.J., Shouse, P.J., van Genuchten, M.T., 1999. Relationship between the
Ahuja, L.R., Cassel, D.K., Bruce, R.R., Barnes, B.B., 1989. Evaluation of spatial distribution hydraulic conductivity function and the particle-size distribution. Soil Sci. Soc. Am. J.
of hydraulic conductivity using effective porosity data. Soil Sci. 148, 404–411. 63, 1063–1070.
Aimrun, W., Amin, M.S.M., 2009. Pedo-transfer function for saturated hydraulic con- Baranau, V., Tallarek, U., 2014. Random-close packing limits for monodisperse and
ductivity of lowland paddy soils. Paddy Water Environ. 7, 217–225. polydisperse hard spheres. Soft Matter 10, 3826–3841.

1027
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Baranov, V., 2017. VasiliBaranov/packing-generation: PackingGeneration 1.0.1.28. Samuel-Rosa, A., Kempen, B., Leenaars, J.G.B., Walsh, M.G., Gonzalez, M.R., 2014.
Batjes, N.H., Ribeiro, E., van Oostrum, A., Leenaars, J., Hengl, T., Mendes de Jesus, J., SoilGrids1km - global soil information based on automated mapping. PLoS One 9,
2017. WoSIS: providing standardised soil profile data for the world. Earth Syst. Sci. 1–17.
Data 9, 1–14. Heuscher, S.A., Brandt, C.C., Jardine, P.M., 2005. Using soil physical and chemical
Bear, J., 1972. Dynamics of Fluids in Porous Materials. Dover Publication Inc., New York. properties to estimate bulk density. Soil Sci. Soc. Am. J. 69, 51–56.
Berryman, J.G., 1985. Measurement of spatial correlation functions using image pro- Hilpert, M., 2011. Determination of dimensional flow fields in hydrogeological settings
cessing techniques. J. Appl. Phys. 57, 2374–2384. via the MRT lattice-Boltzmann method. Water Resour. Res. 47, 1–14.
Berryman, J.G., Blair, S.C., 1986. Use of digital image analysis to estimate fluid perme- Hollis, J.M., Woods, S.M., 1989. The Measurement and Estimation of Saturated Soil
ability of porous materials: application of two-point correlation functions. J. Appl. Hydraulic Conductivity. Research Report for the Ministry of Agriculture, Fisheries
Phys. 60, 1930–1938. and Food. Soil Survey and Land Research Centre, Silsoe, England.
Bittelli, M., Campbell, G.S., Tomei, F., 2015. Soil Physics with Python: Transport in the Holtan, H.N., 1968. Moisture-tension data for selected soils on experimental watersheds.
Soil-Plant-Atmosphere System. Oxford University Press, New York, USA. USDA ARS 609.
Blair, S.C., Berge, P.A., Berryman, J.G., 1996. Using two-point correlation functions to Houston, A.N., Schmidt, S., Tarquis, A.M., Otten, W., Baveye, P.C., Hapca, S.M., 2013.
characterize microgeometry and estimate permeabilities of sandstones and porous Effect of scanning and image reconstruction settings in X-ray computed micro-
glass. J. Geophys. Res. Solid Earth 101, 20359–20375. tomography on quality and segmentation of 3D soil images. Geoderma 207, 154–165.
Bouma, J., 1989. Using Soil Survey Data for Quantitative Land Evaluation, in: Stewart, B. Iassonov, P., Tuller, M., 2010. Application of segmentation for correction of intensity bias
A. (Ed.), Springer, New York, NY, pp. 177–213. in X-ray Computed Tomography ImagesAll rights reserved. No part of this periodical
Brooks, R., Corey, A., 1964. Hydraulic Properties of Porous Media. Hydrology Papers. may be reproduced or transmitted in any form or by any means, electronic or me-
Colorado State University, pp. 1–27. chanical, including photocopying, recording, or any information storage and retrieval
Burdine, N.T., 1953. Relative permeability calculations from pore size distribution data. system, without permission in writing from the publisher. Vadose Zone J. 9, 187–191.
J. Petrol. Technol. 5, 71–78. Iassonov, P., Gebrenegus, T., Tuller, M., 2009. Segmentation of X-ray computed tomo-
Camargo, M.A., Facin, P.C., Pires, L.F., 2012. Lattice boltzmann method for evaluating graphy images of porous materials: a crucial step for characterization and quantita-
hydraulic conductivity of finite array of spheres. Sci. World J. tive analysis of pore structures. Water Resour. Res. 45, 1–12.
Campbell, G.S., 1974. A simple method for determining unsaturated conductivity from Jabro, J.D., 1992. Estimation of saturated hydraulic conductivity of soils from particle
moisture retention data. Soil Sci. 117, 311–314. size distribution and bulk density data. Trans. ASAE 35, 557–560.
Campbell, G.S., Shiozawa, S., 1992. Prediction of Hydraulic Properties of Soils using Jarvis, N., Koestel, J., Messing, I., Moeys, J., Lindahl, A., 2013. Influence of soil, land use
Particle-Size Distribution and Bulk Density Data. Indirect Methods for Estimating the and climatic factors on the hydraulic conductivity of soil. Hydrol. Earth Syst. Sci. 17,
Hydraulic Properties of Unsaturated Soils. University of California, Riverside, pp. 5185–5195.
317–328. Jarvis, N.J., Zavattaro, L., Rajkai, K., Reynolds, W.D., Olsen, P.-A., McGechan, M., Mecke,
Carman, P.C., 1937. Fluid flow through granular beds. Trans. Inst. Chem. Eng. 15, M., Mohanty, B., Leeds-Harrison, P.B., Jacques, D., 2002. Indirect estimation of near-
150–166. saturated hydraulic conductivity from readily available soil information. Geoderma
Carman, P.C., 1956. Flow of Gases through Porous Media. Academic Press Inc., New York. 108, 1–17.
Carsel, R.F., Parrish, R.S., 1988. Developing joint probability distributions of soil water Jorda, H., Bechtold, M., Jarvis, N., Koestel, J., 2015. Using boosted regression trees to
retention characteristics. Water Resour. Res. 24, 755–769. explore key factors controlling saturated and near-saturated hydraulic conductivity.
Chen, F., Dudhia, J., 2001. Coupling an advanced land surface-hydrology model with the Eur. J. Soil Sci. 66, 744–756.
Penn state–NCAR MM5 modeling system. Part I: model implementation and sensi- Ju, L., Zhang, J., Meng, L., Wu, L., Zeng, L., 2018. An adaptive Gaussian process-based
tivity. Monthly Weather Review. iterative ensemble smoother for data assimilation. Adv. Water Resour. 115, 125–135.
Clapp, R.B., Hornberger, G.M., 1978. Empirical equations for some soil hydraulic prop- Kaestner, A., Lehmann, E., Stampanoni, M., 2008. Imaging and image processing in
erties. Water Resour. Res. 14, 601–604. porous media research. Adv. Water Resour. 31, 1174–1187.
Cleveland, C.C., Houlton, B.Z., Smith, W.K., Marklein, A.R., Reed, S.C., Parton, W., Del Katul, G.G., Oren, R., Manzoni, S., Higgins, C., Parlange, M.B., 2012. Evapotranspiration:
Grosso, S.J., Running, S.W., 2013. Patterns of new versus recycled primary produc- a process driving mass transport and energy exchange in the soil-plant-atmosphere-
tion in the terrestrial biosphere. Proc. Natl. Acad. Sci. 110, 12733–12737. climate system. Rev. Geophys. 50.
Coen, G.M., Wang, C., 1989. Estimating vertical saturated hydraulic conductivity from Keehm, Y., Mukerji, T., Nur, A., 2004. Permeability prediction from thin sections: 3D
soil morphology in Alberta. Can. J. Soil Sci. 69, 1–16. reconstruction and Lattice-Boltzmann flow simulation. Geophys. Res. Lett. 31, 1–4.
Cosby, B.J., Hornberger, G.M., Clapp, R.B., Ginn, T.R., 1984. A statistical exploration of Ketcham, R.A., Carlson, W.D., 2001. Acquisition, optimization and interpretation of X-ray
the relationships of soil moisture characteristics to the physical properties of soils. computed tomographic imagery: applications to the geosciences. Comput. Geosci. 27,
Water Resour. Res. 20, 682–690. 381–400.
Dai, Y., Zeng, X., Dickinson, R.E., Baker, I., Bonan, G.B., Bosilovich, M.G., Denning, A.S., Koponen, A., Kataja, M., Timonen, J., v,, 1996. Tortuous flow in porous media. Phys. Rev.
Dirmeyer, P.A., Houser, P.R., Niu, G., Oleson, K.W., Schlosser, C.A., Yang, Z.L., 2003. E 54, 406.
The common land model. Bull. Am. Meteorol. Soc. 84, 1013–1024. Kowalczyk, E.A., Wang, Y.P., Law, R.M., Davies, H.L., McGregor, J.L., Abramowitz, G.,
Dai, Y., Shangguan, W., Duan, Q., Liu, B., Fu, S., Niu, G., 2013. Development of a China 2006. The CSIRO Atmosphere Biosphere Land Exchange (CABLE) model for use in
dataset of soil hydraulic parameters using pedotransfer functions for land surface climate models and as an offline model. CSIRO Marine Atmos. Res. Paper 13, 42.
modeling. J. Hydrometeorol. 14, 869–887. Kozeny, J., 1927. Uber kapillare leitung der wasser in boden. Royal Academy of Science,
Darcy, H.P.G., 1856. Les Fontaines publiques de la ville de Dijon. Exposition et applica- Vienna. Proc. Class I 136, 271–306.
tion des principes à suivre et des formules à employer dans les questions de dis- Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with deep con-
tribution d’eau, etc. V. Dalamont. volutional neural networks. Adv. Neural Inf. Process. Syst. 1097–1105.
Dexter, A.R., 1988. Advances in characterization of soil structure. Soil Tillage Res. 11, Kulkarni, R., Tuller, M., Fink, W., Wildenschild, D., 2012. Three-dimensional multiphase
199–238. segmentation of X-ray CT data of porous materials using a Bayesian Markov random
Donatelli, M., Wösten, J.H.M., Belocchi, G., 2004. Methods to evaluate pedotransfer field framework. Vadose Zone J. 11.
functions. Dev. Soil Sci. 30, 357–411. Lado, M., Paz, A., Ben-Hur, M., 2004. Organic matter and aggregate-size interactions in
Efron, B., Tibshirani, R.J., 1993. An Introduction to the Bootstrap. Chaoman & Hall/CRC, saturated hydraulic conductivity. Soil Sci. Soc. Am. J. 68, 234–242.
Boca Raton, Florida. Larsen, J.D., 2018. Computational Fluid Dynamic Modeling: Approaches For Permeability
Frank, E., Trigg, L., Holmes, G., Witten, I.H., 2000. Naive Bayes for regression. Mach. Modeling An Particle Tracking Using Lattice Boltzmann Methods. University of
Learn. 41, 5–25. Arizona, Tucson.
Freeze, R.A., Cherry, J.A., 1979. Groundwater. Prentice-Hall, Englewood Cliffs, NJ. Leij, F.J., Alves, W.J., Van Genuchten, M.T., Williams, J.R., 1996. Unsaturated soil hy-
Gamie, R., De Smedt, F., 2018. Experimental and statistical study of saturated hydraulic draulic database, UNSODA 1.0 User’s Manual. Office of Research and Development,
conductivity and relations with other soil properties of a desert soil. Eur. J. Soil Sci. United States Environmental Protection Agency, Cincinnati, Ohio.
69, 256–264. Li, Y., Chen, D., White, R.E., Zhu, A., Zhang, J., 2007. Estimating soil hydraulic properties
Griffiths, E., Webb, T.H., Watt, J.P.C., Singleton, P.L., 1999. Development of soil mor- of Fengqiu County soils in the North China Plain using pedo-transfer functions.
phological descriptors to improve field estimation of hydraulic conductivity. Soil Res. Geoderma 138, 261–271.
37, 971–982. Lilly, A., 2000. The relationship between field-saturated hydraulic conductivity and soil
Guber, A.K., Pachepsky, Y.A., 2010. Multimodeling with Pedotransfer Functions: structure: development of class pedotransfer functions. Soil Use Manage. 16, 56–60.
Documentation and User Manual for PTF Calculator (CalcPTF), version 3.0, Lilly, A., Lin, H., 2004. Using soil morphological attributes and soil structure in pedo-
Environmental Microbial and Food Safety. Laboratory Beltsville Agricultural transfer functions. In: Pachepsky, Y., Rawls, W.J. (Eds.), Developments in Soil
Research Center, USDA-ARS. Science: Development of Pedotransfer Functions in Soil Hydrology. Elsevier,
Hagedorn, R., Doblas-Reyes, F.J., Palmer, T.N., 2005. The rationale behind the success of Amsterdam, pp. 115–141.
multi-model ensembles in seasonal forecasting-I. Basic concept. Tellus A: Dyn. Lilly, A., Nemes, A., Rawls, W.J., Pachepsky, Y.A., 2008. Probabilistic approach to the
Meteorol. Oceanogr. 57, 219–233. identification of input variables to estimate hydraulic conductivity. Soil Sci. Soc. Am.
Hastie, T., Tibshirani, R., Friedman, J., 2001. The Elements of Statistical Learning Data J. 72, 16–24.
Mining, Inference, and Prediction, Springer. Springer New York, New York. Man, J., Li, W., Zeng, L., Wu, L., 2016. Data assimilation for unsaturated flow models with
Haverkamp, R., Zammit, C., Boubkraoui, F., Rajkai, K., Arrúe, J.L., Heckmann, N., 1997. restart adaptive probabilistic collocation based Kalman filter. Adv. Water Resour. 92,
GRIZZLY, Grenoble Soil Catalogue: Soil Survey of Field Data and Description of 258–270.
Particle-Size, Soil Water Retention and Hydraulic Conductivity Functions. Martys, N.S., Chen, H., 1996. Simulation of multicomponent fluids in complex three-
Laboratoire d’Etude des Transferts en Hydrologie et Environnement (LTHE), dimensional geometries by the lattice Boltzmann method. Phys. Rev. E 53, 743.
Grenoble Cedex 9. Mbagwu, J.S.C., Auerswald, K., 1999. Relationship of percolation stability of soil ag-
Hengl, T., de Jesus, J.M., MacMillan, R.A., Batjes, N.H., Heuvelink, G.B.M., Ribeiro, E., gregates to land use, selected properties, structural indices and simulated rainfall

1028
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

erosion. Soil Tillage Res. 50, 197–206. Raissi, M., Perdikaris, P., Karniadakis, G.E., 2017. Inferring solutions of differential
McKeague, J.A., Wang, C., Topp, G.C., 1982. Estimating saturated hydraulic conductivity equations using noisy multi-fidelity data. J. Comput. Phys. 335, 736–746.
from soil morphology. Soil Sci. Soc. Am. J. 46, 1239–1244. Rawls, W.J., 1983. Estimating soil bulk density from particle size analysis and organic
Merdun, H., 2010. Alternative methods in the development of pedotransfer functions for matter content. Soil Sci. 135, 123–125.
soil hydraulic characteristics. Eur. Soil Sci. 43, 62–71. Rawls, W.J., Brakensiek, D.L., Saxton, K.E., 1982. Estimation of soil water properties.
Minasny, B., Hartemink, A.E., 2011. Predicting soil properties in the tropics. Earth Sci. Trans. ASAE.
Rev. 106, 52–62. Rawls, W.J., Brakensiek, D.L., 1985. Prediction of soil water properties for hydrologic
Mishra, S., Parker, J.C., Singhal, N., 1989. Estimation of soil hydraulic properties and modeling. In: Watershed Management in the Eighties. American Society of Civil
their uncertainty from particle size distribution data. J. Hydrol. 108, 1–18. Engineers, pp. 293–299.
Mo, S., Zhu, Y., Zabaras, N., Shi, X., Wu, J., 2019. Deep convolutional encoder-decoder Rawls, W., Yates, P., Asmussen, L., 1976. Calibration of selected infiltration equations for
networks for uncertainty quantification of dynamic multiphase flow in heterogeneous the Georgia Coastal Plain. Report ARS-S-113.
media. Water Resour. Res. 55, 703–728. Rawls, W.J., Gimenez, D., Grossman, R., 1998. Use of soil texture, bulk density, and slope
Montzka, C., Herbst, M., Weihermüller, L., Verhoef, A., Vereecken, H., 2017. A global of the water retention curve to predict saturated hydraulic conductivity. Trans. ASAE
data set of soil hydraulic properties and sub-grid variability of soil water retention 41, 983.
and hydraulic conductivity curves. Earth Syst. Sci. Data 9, 529–543. Revil, A., Cathles, L.M., 1999. Permeability of shaly sands. Water Resour. Res. 35,
Morbidelli, R., Saltalippi, C., Flammini, A., Cifrodelli, M., Picciafuoco, T., Corradini, C., 651–662.
Govindaraju, R.S., 2017. In situ measurements of soil saturated hydraulic con- Richards, L.A., 1931. Capillary conduction of liquids through porous mediums. Physics 1,
ductivity: assessment of reliability through rainfall–runoff experiments. Hydrol. 318–333.
Process. 31, 3084–3094. Richardson, L.F., 1922. Weather Prediction by Numerical Process. Cambridge University
Mualem, Y., 1976. A new model for predicting the hydraulic conductivity of unsaturated Press, Cambridge, UK.
porous media. Water Resour. Res. 12, 513–522. Romero-Ruiz, A., Linde, N., Keller, T., Or, D., 2018. A review of geophysical methods for
National Cooperative Soil Survey, 2017. National cooperative soil survey characterization soil structure characterization. Rev. Geophys. 56, 1–26.
database [WWW Document]. URL: http://ncsslabdatamart.sc.egov.usda.gov (ac- Sanchez-Vila, X., Guadagnini, A., Carrera, J., 2006. Representative hydraulic con-
cessed 01.01.17). ductivities in saturated groundwater flow. Rev. Geophys. 44, 1–46.
Nemes, A., Schaap, M.G., Leij, F.J., Wösten, J.H.M., 2001. Description of the unsaturated Saxton, K.E., Rawls, W.J., 2006. Soil water characteristic estimates by texture and organic
soil hydraulic database UNSODA version 2.0. J. Hydrol. 251, 151–162. matter for hydrologic solutions. Soil Sci. Soc. Am. J. 70, 1569–1578.
Nemes, A., Rawls, W.J., Pachepsky, Y.A., 2005. Influence of organic matter on the esti- Schaap, M.G., 2004. Accuracy and uncertainty in PTF predictions. In: Pachepsky, Y.A.,
mation of saturated hydraulic conductivity. Soil Sci. Soc. Am. J. 69, 1330–1337. Rawls, W.J. (Eds.), Developments in Soil Science. Elsevier, Amsterdam, pp. 33–43.
Neuman, S.P., 1994. Generalized scaling of permeabilities: validation and effect of sup- Schaap, M.G., Bouten, W., 1996. Modeling water retention curves of sandy soils using
port scale. Geophys. Res. Lett. 21, 349–352. neural networks. Water Resour. Res. 32, 3033–3040.
Niu, G.Y., Yang, Z.L., Mitchell, K.E., Chen, F., Ek, M.B., Barlage, M., Kumar, A., Manning, Schaap, M.G., Lebron, I., 2001. Using microscope observations of thin sections to estimate
K., Niyogi, D., Rosero, E., Tewari, M., Xia, Y., 2011. The community Noah land soil permeability with the Kozeny-Carman equation. J. Hydrol. 251, 186–201.
surface model with multiparameterization options (Noah-MP): 1. Model description Schaap, M.G., Leij, F.J., 1998b. Database-related accuracy and uncertainty of pedo-
and evaluation with local-scale measurements. J. Geophys. Res. Atmos. 116, 1–19. transfer functions. Soil Sci. 163, 765–779.
O’Neal, A.M., 1949. Soil characteristics significant in evaluating permeability. Soil Sci. Schaap, M.G., Leij, F.J., 1998a. Using neural networks to predict soil water retention and
67, 403–410. soil hydraulic conductivity. Soil Tillage Res. 47, 37–42.
O’Neal, A.M., 1952. A key for evaluating soil permeability by means of certain field clues. Schaap, M.G., Nemes, A., van Genuchten, M.T., 2004. Comparison of models for indirect
Soil Sci. Soc. Am. J. 16, 312–315. estimation of water retention and available water in surface soils. Vadose Zone J. 3,
Oki, T., Kanae, S., 2006. Global hydrological cycles and world water resources. Science 1455–1463.
313, 1068–1072. Schaap, M.G., Porter, M.L., Christensen, B.S.B., Wildenschild, D., 2007. Comparison of
Oleson, K., Lawrence, D.M., Bonan, G.B., Drewniak, B., Huang, M., Koven, C.D., Levis, S., pressure-saturation characteristics derived from computed tomography and lattice
Li, F., Riley, W.J., Subin, Z.M., Swenson, S., Thornton, P.E., Bozbiyik, A., Fisher, R., Boltzmann simulations. Water Resour. Res.
Heald, C.L., Kluzek, E., Lamarque, J.-F., Lawrence, P.J., Leung, L.R., Lipscomb, W., Schaap, M.G., Leij, F.J., van Genuchten, M.T., 2001. ROSETTA: a computer program for
Muszala, S.P., Ricciuto, D.M., Sacks, W.J., Sun, Y., Tang, J., Yang, Z.-L., 2013. estimating soil hydraulic parameters with hierarchical pedotransfer functions. J.
Technical description of version 4.5 of the Community Land Model (CLM), NCAR Hydrol. 251, 163–176.
Technical Note NCAR/TN-503+STR (422 pp). Schaap, M.G., van Genuchten, M.T., 2006. A modified Mualem–van Genuchten for-
Pachepsky, Y.a., Guber, A.K., Yakirevich, A.M., McKee, L., Cady, R.E., Nicholson, T.J., mulation for improved description of the hydraulic conductivity near saturation.
2014. Scaling and pedotransfer in numerical simulations of flow and transport in Vadose Zone J. 5, 27–34.
soils. Vadose Zone J. 13. Shangguan, W., Dai, Y., Duan, Q., Liu, B., Yuan, H., 2014. A global soil data set for earth
Pachepsky, Y., Hill, R.L., 2017. Scale and scaling in soils. Geoderma 287, 4–30. system modeling. J. Adv. Model. Earth Syst. 6, 249–263.
Pachepsky, Y.A., Rawls, W.J., 2004. Development of Pedotransfer Functions in Soil Shi, L., Song, X., Tong, J., Zhu, Y., Zhang, Q., 2015. Impacts of different types of mea-
Hydrology, Developments in Soil Science. Elsevier, Amsterdam, the Netherlands. surements on estimating unsaturated flow parameters. J. Hydrol. 524, 549–561.
Pan, F., Pachepsky, Y., Jacques, D., Guber, A., Hill, R.L., 2012. Data assimilation with soil Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G.,
water content sensors and pedotransfer functions in soil water flow modeling. Soil Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., 2016. Mastering
Sci. Soc. Am. J. 76, 829–844. the game of Go with deep neural networks and tree search. Nature 529, 484.
Papanicolaou, A.T.N., Elhakeem, M., Wilson, C.G., Burras, C.L., West, L.T., Lin, H.H., Sobieraj, J.A., Elsenbeer, H., Vertessy, R.A., 2001. Pedotransfer functions for estimating
Clark, B., Oneal, B.E., 2015. Spatial variability of saturated hydraulic conductivity at saturated hydraulic conductivity: implications for modeling storm flow generation. J.
the hillslope scale: understanding the role of land management and erosional effect. Hydrol. 251, 202–220.
Geoderma 243, 58–68. Spanne, P., Thovert, J.F., Jacquin, C.J., Lindquist, W.B., Jones, K.W., Adler, P.M., 1994.
Rabot, E., Wiesmeier, M., Schlüter, S., Vogel, H.-J., 2018. Soil structure as an indicator of Synchrotron computed microtomography of porous media: topology and transports.
soil functions: a review. Geoderma 314, 122–137. Phys. Rev. Lett. 73, 2001.
Rahmati, M., Weihermüller, L., Vanderborght, J., Pachepsky, Y.A., Mao, L., Sadeghi, S.H., Spychalski, M., Kaźmierowski, C., Kaczmarek, Z., 2007. Estimation of saturated hydraulic
Moosavi, N., Kheirfam, H., Montzka, C., Van Looy, K., Toth, B., Hazbavi, Z., Al conductivity on the basis of drainage porosity. Electron. J. Polish Agric. Univ. 10.
Yamani, W., Albalasmeh, A.A., Alghzawi, M.Z., Angulo-Jaramillo, R., Antonino, Sukop, M.C., Huang, H., Alvarez, P.F., Variano, E.A., Cunningham, K.J., 2013. Evaluation
A.C.D., Arampatzis, G., Armindo, R.A., Asadi, H., Bamutaze, Y., Batlle-Aguilar, J., of permeability and non-Darcy flow in vuggy macroporous limestone aquifer samples
Béchet, B., Becker, F., Blöschl, G., Bohne, K., Braud, I., Castellano, C., Cerdà, A., with lattice Boltzmann methods. Water Resour. Res. 49, 216–230.
Chalhoub, M., Cichota, R., Císlerová, M., Clothier, B., Coquet, Y., Cornelis, W., Suleiman, A.A., Ritchie, J.T., 2001. Estimating saturated hydraulic conductivity from soil
Corradini, C., Coutinho, A.P., de Oliveira, M.B., de Macedo, J.R., Durães, M.F., porosity. Trans. ASAE 44, 235.
Emami, H., Eskandari, I., Farajnia, A., Flammini, A., Fodor, N., Gharaibeh, M., Tamari, S., Wösten, J.H.M., Ruiz-Suarez, J.C., 1996. Testing an artificial neural network
Ghavimipanah, M.H., Ghezzehei, T.A., Giertz, S., Hatzigiannakis, E.G., Horn, R., for predicting soil hydraulic conductivity. Soil Sci. Soc. Am. J. 60, 1732–1741.
Jiménez, J.J., Jacques, D., Keesstra, S.D., Kelishadi, H., Kiani-Harchegani, M., Timlin, D.J., Ahuja, L.R., Pachepsky, Y., Williams, R.D., Gimenez, D., Rawls, W., 1999.
Kouselou, M., Kumar Jha, M., Lassabatere, L., Li, X., Liebig, M.A., Lichner, L., López, Use of Brooks-Corey parameters to improve estimates of saturated conductivity from
M.V., Machiwal, D., Mallants, D., Mallmann, M.S., de Oliveira Marques, J.D., effective porosity. Soil Sci. Soc. Am. J. 63, 1086–1092.
Marshall, M.R., Mertens, J., Meunier, F., Mohammadi, M.H., Mohanty, B.P., Pulido- Tomasella, J., Hodnett, M.G., 1997. Estimating unsaturated hydraulic conductivity of
Moncada, M., Montenegro, S., Morbidelli, R., Moret-Fernández, D., Moosavi, A.A., Brazilian soils using soil-water retention data. Soil Sci. 162, 703–712.
Mosaddeghi, M.R., Mousavi, S.B., Mozaffari, H., Nabiollahi, K., Neyshabouri, M.R., Tóth, B., Weynants, M., Nemes, A., Makó, A., Bilas, G., Tóth, G., 2015. New generation of
Ottoni, M.V., Ottoni Filho, T.B., Pahlavan-Rad, M.R., Panagopoulos, A., Peth, S., hydraulic pedotransfer functions for Europe. Eur. J. Soil Sci. 66, 226–238.
Peyneau, P.-E., Picciafuoco, T., Poesen, J., Pulido, M., Reinert, D.J., Reinsch, S., Tsangaratos, P., Ilia, I., 2016. Comparison of a logistic regression and Naïve Bayes clas-
Rezaei, M., Roberts, F.P., Robinson, D., Rodrigo-Comino, J., Rotunno Filho, O.C., sifier in landslide susceptibility assessments: the influence of models complexity and
Saito, T., Suganuma, H., Saltalippi, C., Sándor, R., Schütt, B., Seeger, M., Sepehrnia, training dataset size. Catena 145, 164–179.
N., Sharifi Moghaddam, E., Shukla, M., Shutaro, S., Sorando, R., Stanley, A.A., Twarakavi, N.K.C., Šimůnek, J., Schaap, M.G., 2009. Development of pedotransfer
Strauss, P., Su, Z., Taghizadeh-Mehrjardi, R., Taguas, E., Teixeira, W.G., Vaezi, A.R., functions for estimation of soil hydraulic parameters using support vector machines.
Vafakhah, M., Vogel, T., Vogeler, I., Votrubova, J., Werner, S., Winarski, T., Yilmaz, Soil Sci. Soc. Am. J. 73, 1443–1452.
D., Young, M.H., Zacharias, S., Zeng, Y., Zhao, Y., Zhao, H., Vereecken, H., 2018. van Genuchten, M.T., 1980. A closed-form equation for predicting the hydraulic con-
Development and analysis of the Soil Water Infiltration Global database. Earth Syst. ductivity of unsaturated soils. Soil Sci. Soc. Am. J. 44, 892–898.
Sci. Data 10, 1237–1263. Van Looy, K., Bouma, J., Herbst, M., Koestel, J., Minasny, B., Mishra, U., Montzka, C.,

1029
Y. Zhang and M.G. Schaap Journal of Hydrology 575 (2019) 1011–1030

Nemes, A., Pachepsky, Y.A., Padarian, J., Schaap, M.G., Tóth, B., Verhoef, A., Wösten, J.H.M., 2000. The HYPRES database of hydraulic properties of European soils.
Vanderborght, J., van der Ploeg, M.J., Weihermüller, L., Zacharias, S., Zhang, Y., Adv. GeoEcol. 135–143.
Vereecken, H., 2017. Pedotransfer functions in earth system science: challenges and Wösten, J.H.M., Lilly, A., Nemes, A., Le Bas, C., 1999. Development and use of a database
perspectives. Rev. Geophys. 55, 1199–1256. of hydraulic properties of European soils. Geoderma 90, 169–185.
Vereecken, H., Maes, J., Feyen, J., 1990. Estimating unsaturated hydraulic conductivity Wösten, J.H.M., Pachepsky, Y.A., Rawls, W.J., 2001. Pedotransfer functions: bridging the
from easily measured soil properties. Soil Sci. 149, 1–12. gap between available basic soil data and missing soil hydraulic characteristics. J.
Vereecken, H., Weynants, M., Javaux, M., Pachepsky, Y., Schaap, M.G., Genuchten, M.T. Hydrol. 251, 123–150.
Van, 2010. Using pedotransfer functions to estimate the van Genuchten-Mualem soil Yang, Y., Jia, X., Wendroth, O., Liu, B., 2018. Estimating saturated hydraulic conductivity
hydraulic properties: a review. Vadose Zone J. 9, 795–820. along a south-north transect in the loess plateau of China. Soil Sci. Soc. Am. J. 108,
Vereecken, H., Schnepf, A., Hopmans, J.W.W., Javaux, M., Or, D., Roose, T., 1–17.
Vanderborght, J., Young, M.H., Amelung, W., Aitkenhead, M., Allison, S.D.D., Ye, M., Meyer, P.D., Neuman, S.P., 2008. On model selection criteria in multimodel
Assouline, S., Baveye, P., Berli, M., Brüggemann, N., Finke, P., Flury, M., Gaiser, T., analysis. Water Resour. Res. 44.
Govers, G., Ghezzehei, T., Hallett, P., Lamorski, K., Hendricks Franssen, H.J., Heppell, Yeong, C.L.Y., Torquato, S., 1998. Reconstructing random media. Phys. Rev. E 57, 495.
J., Horn, R., Huisman, J.A., Jacques, D., Jonard, F., Kollet, S., Lafolie, F., Lamorski, Yu, B., Li, J., Li, Z., Zou, M., 2003. Permeabilities of unsaturated fractal porous media. Int.
K., Leitner, D., McBratney, A., Minasny, B., Montzka, C., Nowak, W., Pachepsky, Y., J. Multiph. Flow 29, 1625–1642.
Padarian, J., Romano, N., Roth, K., Rothfuss, Y., Rowe, E.C., Schwen, A., Šimůnek, J., Yu, D., Yang, J., Shi, L., Zhang, Q., Huang, K., Fang, Y., Zha, Y., 2018. On the uncertainty
Tiktak, A., Van Dam, J., van der Zee, S.E.A.T.M., Vogel, H.J., Vrugt, J.A., Wöhling, T., of initial condition and initialization approaches in variably saturated flow modeling.
Young, I.M., 2016. Modeling soil processes: review, key challenges, and new per- Hydrol. Earth Syst. Sci. Discuss. 1–42. https://doi.org/10.5194/hess-2018-557.
spectives. Vadose Zone J. 9, 795–820. Zacharias, S., Wessolek, G., 2007. Excluding organic matter content from pedotransfer
Verhoef, A., Egea, G., 2014. Modeling plant transpiration under limited soil water: predictors of soil water retention. Soil Sci. Soc. Am. J. 71, 43–50.
comparison of different plant and soil hydraulic parameterizations and preliminary Zhang, X., Deeks, L.K., Bengough, A.G., Crawford, J.W., Young, I.M., 2005. Determination
implications for their use in land surface models. Agric. For. Meteorol. 191, 22–32. of soil hydraulic conductivity with the lattice Boltzmann method and soil thin-section
Vogel, H.-J., Tölke, J., Schulz, V.P., Krafczyk, M., Roth, K., 2005. Comparison of a lattice- technique. J. Hydrol. 306, 59–70.
Boltzmann model, a full-morphology model, and a pore network model for de- Zhang, Y., Schaap, M.G., 2017. Weighted recalibration of the rosetta pedotransfer model
termining capillary pressure–saturation relationships. Vadose Zone J. 4, 380–388. with improved estimates of hydraulic parameter distributions and summary statistics
Wagner, B., Tarnawski, V.R., Hennings, V., Müller, U., Wessolek, G., Plagge, R., 2001. (Rosetta3). J. Hydrol. 547, 39–53.
Evaluation of pedo-transfer functions for unsaturated soil hydraulic conductivity Zhang, Y., Schaap, M.G., Guadagnini, A., Neuman, S.P., 2016. Inverse modeling of un-
using an independent data set. Geoderma 102, 275–297. saturated flow using clusters of soil texture and pedotransfer functions. Water Resour.
Wang, C., McKeague, J.A., Topp, G.C., 1985. Comparison of estimated and measured Res. 52, 1–14.
horizontal Ksat values. Can. J. Soil Sci. 65, 707–715. Zhang, Y., Schaap, M.G., Zha, Y., 2018. A High-resolution global map of soil hydraulic
Wang, Y., Shao, M., Liu, Z., 2012. Pedotransfer functions for predicting soil hydraulic properties produced by a hierarchical parameterization of a physically-based water
properties of the chinese loess plateau. Soil Sci. 177, 424–432. retention model. Water Resour. Res. 54, 9774–9790.
Weynants, M., Vereecken, H., Javaux, M., 2009. Revisiting Vereecken pedotransfer Zhang, Y., Schaap, M.G., Wei, Z., 2019. Hierarchical multimodel ensemble estimates of
functions: introducing a closed-form hydraulic model. Vadose Zone J. 8, 86–95. soil water retention with global coverage. arXiv:1906.03182 [stat.AP].
Wildenschild, D., Sheppard, A.P., 2013. X-ray imaging and analysis techniques for Zheng, Q., Zhang, J., Xu, W., Wu, L., Zeng, L., 2019. Adaptive multifidelity data assim-
quantifying pore-scale structure and processes in subsurface porous medium systems. ilation for nonlinear subsurface flow problems. Water Resour. Res. 55, 203–217.
Adv. Water Resour. 51, 217–246. Zhou, Hongxiang, Yu, Xiuling, Chen, Cheng, Zeng, Lingzao, Lu, Shenggao, Wu, Laosheng,
Wösten, J.H.M., 1997. Pedotransfer functions to evaluate soil quality. In: Gregorich, E.G., 2018. Evaluating hydraulic properties of biochar-amended soil aggregates by high-
Carter, M.R. (Eds.), Soil Quality for Crop Production and Ecosystem Health. performance pore-scale simulations. Soil Sci. Soc. Am. J. 82, 1–9.
Developments in Soil Science, Elsevier, Amsterdam, pp. 221–246.

1030

Вам также может понравиться