Você está na página 1de 14

Geoderma 318 (2018) 64–77

Contents lists available at ScienceDirect

Geoderma
journal homepage: www.elsevier.com/locate/geoderma

Estimation of soil moisture using confined compression curve parameters T


a,⁎ a b
Hossein Bayat , Eisa Ebrahimi , Mahboubeh Fallah
a
Department of Soil Science, Faculty of Agriculture, Bu Ali Sina University, Hamedan, Iran
b
Tarbiat Modarres University, Tehran, Iran

A R T I C L E I N F O A B S T R A C T

Handling Editor: Morgan Cristine L.S. The confined compression curve (CCC) represents the relationship between the logarithm of the applied stress
Keywords: and the void ratio. There are several similarities between the soil water retention curve (SWRC) and the CCC.
Confined compression curve The aim of this study was to modify the Dexter SWRC model to fit the experimental CCC data by replacing the
Estimation matric suction and water content in the Dexter model with the normal stress and void ratio, respectively, as well
Pedotransfer function as using the CCC parameters and characteristics (pre-compression stress, compression, and swelling indices) as
Soil water retention curve predictors to estimate the SWRC. We collected 150 soil samples from five provinces of Iran. The SWRC, CCC, and
basic properties of the soil samples (clay, silt/sand, and bulk density) were measured. The Dexter model was
applied to the experimental data for both the CCC and SWRC, and their parameters were calculated. The CCC
parameters and basic soil properties were used to estimate the soil moisture at five input levels with the Dexter
model. The best results were obtained using the basic properties of soil as predictors, as well as with the
parameters of the Dexter model obtained by its fitting to the CCC data. The integral root mean squared error was
reduced from 0.059 and 0.061 (in the first step) to 0.053 and 0.056 g g− 1 in the training and testing steps,
respectively. The relative improvements in the SWRC estimates showed that improvements of 4.9% to 11.9%
were obtained by using the CCC parameters as predictors. These improved estimates can be attributed to the
apparent similarities between the two curves as well as the impacts of similar factors on these curves and the
correlation between them.

1. Introduction (Fig. 1). Pc is the stress indicating changes in the elastic behavior to
plastic behavior (Cavalieri et al., 2008). Thus, Pc is the maximum stress
Soil compaction is caused by a number of different factors that can that soils have ever experienced. Permanent soil compression will occur
destroy the soil structure, reduce soil porosity, decrease the infiltration when the applied stress exceeds Pc. However, whenever the applied
capacity of water, and also change the arrangement of soil particles. stress is lower than the Pc, the soil will be recompressed after com-
Compaction affects the soil water retention curve (SWRC) and me- pression.
chanical properties by changing the soil structure. In soil engineering, The absolute value of the slope of the virgin compression line is
the soil deformation is evaluated as elastic (instantaneous) or plastic called the compression index (Cc), which is considered to be an in-
(long term) depending on the intensity of the stress (Baumgartl and dicator of a soil's resistance to compaction, as well as being used as a
Koeck, 2004). Keller et al. (2011) showed that the graph of the log criterion for measuring the soil compressibility (Fig. 1). The resistance
stress (logσ) versus the void ratio (e) for an unsaturated soil can be used of the soil to compression will be lower when Cc is higher, and the
to describe the compression of the soil. This curve is called the confined compression will occur very rapidly after small changes in the applied
compression curve (CCC) and it has three important parameters: the stress. The slope of the virgin compression line is very important for
pre-compression stress (Pc), the swelling index (Cs), and the compres- analyzing and calculating soil settlement (Larson et al., 1980).
sion index (Cc) (Fig. 1). The absolute value of the elastic area slope of the CCC is the swel-
Pc is often obtained by plotting the void ratio or vertical strain of ling index (Cs), which is used as a measure of soil mechanical resilience.
soil versus the logarithm of the vertical pressure stress (Casagrande, Cs is significantly smaller than Cc and it can normally be obtained from
1936). The CCC has two separate areas that represent the elastic be- experimental results (Fig. 1). The soil resistance to compression will be
havior at low stresses (recompression or swelling line) and the per- higher when Cs is larger. Thus, larger values of Cs indicate that the soil
manent deformation (virgin compression line) at higher stresses will be recompressed after compression by removing the applied stress.


Corresponding author.
E-mail address: h.bayat@basu.ac.ir (H. Bayat).

https://doi.org/10.1016/j.geoderma.2017.12.034
Received 3 November 2016; Received in revised form 7 December 2017; Accepted 29 December 2017
0016-7061/ © 2018 Elsevier B.V. All rights reserved.
H. Bayat et al. Geoderma 318 (2018) 64–77

hydraulic conductivity (Vereecken et al., 1992), soil genetic data (Tietje


and Tapkenhinrichs, 1993), cation exchange capacity (Pachepsky and
Rawls, 1999), penetration resistance (Pachepsky et al., 1998), specific
surface area (Walczak et al., 2004), geometric specific surface (Walczak
et al., 2006), fractal parameters of particles and aggregates (Bayat et al.,
2011; Bayat et al., 2013), and micro-aggregate size distributions
(Ebrahimi et al., 2014b). Despite the use of different soil properties for
developing PTFs, it is still challenging to obtain accurate predictions of
the parameters for SWRC models in soil physics, and thus the water
contents. In addition, the selection of the input variables employed will
always depend on the specific data set and the region considered. In
arid countries where the soil organic matter contents are usually low,
this parameter will rarely be significant when estimating the SWRC
compared with data sets obtained from temperate or cold regions of the
world. Therefore, finding new input variables that are readily available
and correlated with the output variables in the study region would
facilitate the development of PTFs.
There are some close relationships between the SWRC and CCC, but
measuring the SWRC is far more expensive and time consuming than
measuring the CCC. In fact, the CCC cannot be measured as easily as
readily available soil properties such as the soil texture, but it can be
measured quickly. Expensive equipment such as a pressure plate and
sandbox are required to measure the SWRC, which are not available
Fig. 1. Confined compression curve and its components. Cc is the compression index, Cs is
the swelling index, Pc is the pre-compression stress, VCL is the virgin compression line,
everywhere, whereas the CCC can be measured using a relatively simple
and SL is the swelling line. single-axis apparatus, which is much cheaper than both a pressure plate
and sandbox. Thus, the CCC can be determined in less time and at a
lower cost.
The stress/void ratio graph or CCC is very similar to the SWRC.
Functions that only use very simple data as input variables, e.g., soil
These two curves start from an upper point and then decrease, before
texture, have lower precision when estimating the soil moisture. The
increasing on the horizontal axis. Furthermore, the shapes of the curves
moisture content is a function of many factors, and thus the moisture
are changed by variations in the pore size distribution (Baumgartl and
retention cannot be described fully by using features such as the soil
Koeck, 2004; Schoonover and Jackie, 2015). Dexter et al. (2008) re-
texture. Soil mechanical properties such as the CCC have strong impacts
ported that the soil structure is associated with the SWRC. Soane and
on soil moisture retention (Gregory et al., 2006; Imhoff et al., 2004;
Van Ouwerkerk (1994) and Chan et al. (2006) also stated that the CCC
Keller and Arvidsson, 2007; Lebert and Horn, 1991) because they are
is a function of the soil structure, whereas the SWRC is a function of the
correlated with the soil particle size distribution (Gregory et al., 2006;
three phases comprising solid, liquid, and air in the soil (Sillers et al.,
Imhoff et al., 2004; Lebert and Horn, 1991), organic matter (Kuan et al.,
2001). In addition, some studies have demonstrated that increasing the
2007), soil structure (Ekwue, 1990), BD, and pore size distribution
clay content of the soil causes increases in Pc and Cc (Gregory et al.,
(Culley and Larson, 1987; Horn, 2004). In previous studies, the CCC
2006; Saffih-Hdadi et al., 2009). In the SWRC, increasing the clay
and its coefficients have not been used to estimate the SWRC although
contents enhances the retention of soil water, as well as increasing the
they contain large amounts of information regarding the soil mechan-
air entry suction and reducing the slope of the curve in the middle area.
ical properties, so using them as predictors might improve estimates of
Baumgartl and Koeck (2004) reported that the suction and stress are
the SWRC.
two common parameters for both curves where they lie on the hor-
The Dexter model is a bimodal model and its parameters have
izontal axis in the graphs. Zhang et al. (2017) stated that there is a
physical meaning (Dexter et al., 2008), whereas the van Genuchten
relationship between the SWRC and soil compaction. Moreover, low
model is a unimodal model. Bimodal models are more accurate com-
compaction of soil leads to an increase in water retention at higher
pared with unimodal models for aggregated soils, including the soils
matric suctions because soil compression reduces the macroporosity,
used in this study where the pore size distribution of the soils contained
which is responsible for drainage and water withdrawal from the soil.
two peaks (the peaks indicate the pore size(s) with the maximum
In recent years, pedotransfer functions (PTFs) have become in-
density if the pore size distribution is represented in the form of density
creasingly popular for estimating the hydraulic properties of soil
values) (Seki, 2007). Thus, one of the advantages of the present study is
(Lamorski et al., 2008; Twarakavi et al., 2009). PTFs are regression
the use of the bimodal Dexter model instead of the unimodal van
equations or other models (e.g., artificial neural networks) that corre-
Genuchten–Mualem model. The objectives of the present study were:
late difficult to measure soil properties, such as the soil hydraulic
(1) to modify the Dexter et al. (2008) SWRC model to fit to the ex-
properties, with more readily available soil data, including the texture,
perimental CCC data; and (2) to use the CCC parameters and char-
organic matter content, and bulk density (BD) (Gupta and Larson,
acteristics (such as Pc, Cc, and Cs) as predictors to estimate the para-
1979). Thus, PTFs transform “what we have” into “what we need”
meters of the Dexter et al. (2008) SWRC model, and thus the soil
(Wösten et al., 2001). As an indirect approach, PTFs are among the
moisture content.
most widely used tools employed for predicting the SWRC. Obtaining
direct measurements of the SWRC is time consuming, expensive, and
1.1. Theory
highly laborious, so many attempts have been made to predict the
SWRC indirectly by using different methods based on both the physical
Dexter et al. (2008) presented a model based on the fact that the
and chemical properties of soil (Nguyen et al., 2014, 2017).
porosity of soil comprises four components, i.e., the residual porosity
In previous studies, various parameters have been used in PTFs to
representing very fine pores in soil, the porosity of the soil matrix, the
estimate the SWRC, including the soil texture (Lee and Ro, 2014), or-
structural porosity comprising the pore spaces among the micro-ag-
ganic matter, BD (Meskini-Vishkaee et al., 2014), soil saturated
gregates and among incipient aggregates, and the macroporosity due to

65
H. Bayat et al. Geoderma 318 (2018) 64–77

Fig. 2. Soil sampling locations.

macropores. The residual and macroporosity are too small and large, The parameter σ represents the stress (kPa) and the parameter e re-
respectively, to be characterized in standard water retention experi- presents the void ratio of soil (cm3 cm− 3). σ1⁎ and σ2⁎ are fitting
ments. The matrix and structural pores are in the range of the studied parameters (kPa).
suctions in the SWRC because they are where the water is mainly re- The relationship between volume change and the CCC is a common
tained; therefore, these components are more important than the other feature of the descriptions of the soil mechanical characteristics in both
two components. The Dexter bimodal equation is as follows: mechanical and agricultural engineering. A recompression curve at a
lower stress range can be obtained from an intact compression curve at
−h ⎞ −h ⎞
w = c + A1 exp ⎛ ⎜ + A2 exp ⎛
⎟ ⎜ , ⎟
higher stresses by using a semi-logarithmic plot of the normal stress vs.
⎝ h1 ⎠ ⎝ h2 ⎠ (1) strain (Casagrande, 1936; Terzaghi and Jelinek, 1954), where both
curves are separated by the pre-compression stress. Moreover, due to
where w is the gravimetric moisture content of the soil (g g− 1) and the
increased internal stresses, volume changes lead to a relationship that
first term, c, is the asymptote of the equation, which is the residual
behaves in a similar manner.
water content (g g− 1). The amounts of the matrix and structural pore
Smaller levels of shrinkage occur at lower capillary stresses (i.e.,
spaces are proportional to A1 (g g− 1) and A2 (g g− 1), respectively. h
lower matric suctions) but the rate of shrinkage increases with higher
(cm) is the matric suction corresponding to w. The values of h1 (cm)
matric suctions for soil water above a certain degree of dryness.
and h2 (cm) are the characteristic pore water suctions at which the
In order to compare the fitting accuracy of the Dexter CCC model
matrix and structural pore spaces empty, respectively (Dexter et al.,
with that of another CCC model, and to compare the possibility of es-
2008).
timating the SWRC using the parameters in the Dexter CCC model and
Volume changes in soils may be due to either external (mechanical)
another CCC model as predictors, we fitted the Gompertz (1825)
or internal (hydraulic) stresses or a combination of both. Therefore,
equation with four constant coefficients to the stress/void ratio ex-
both mechanical and hydraulic stresses must be included in the de-
perimental data. The form of the Gompertz equation is exponential:
scription of a soil to achieve a comprehensive depiction of the volume
change. By combining theories of the mechanical and hydraulic stress, a e = a + cd exp[−exp(b (log σ − m))], (3)
hydraulic function such as the Dexter model, which predicts the var-
where a, b, cd, and m are the parameters of the Gompertz model. Cc and
iation in the water volume as a function of a stress state parameter (i.e.,
soil water suction) (SWRC), can be used to model the volume change for Cs (Eqs. (4) and (5), respectively) were obtained according to Gregory
a soil sample in the confined compression test (CCT) (Baumgartl and et al. (2006) using the Gompertz equation coefficients, where the stated
Koeck, 2004). Hence, in this study, the Dexter SWRC model was mod- value of Cs was based on the average of the loading slope on the line at
ified to fit the CCC, as follows. 25 kPa (Cs loading–25 kPa):
bcd
−σ −σ Cc =
e = c∗ + A1∗ exp ⎛ ∗ ⎞ + A2∗ exp ⎛ ∗ ⎞
⎜ ⎟ ⎜ ⎟ exp(1) (4)
⎝ σ1 ⎠ ⎝ σ2 ⎠ (2)
e0 − e25kPa
In Eq. (2), the parameters marked with asterisks (c⁎, A1⁎, A2⁎, σ1⁎, Cs =
log(25kPa)
,
(5)
and σ2⁎) are obtained by fitting the Dexter et al. (2008) model to the
stress/void ratio curve data. The unit is cm3 cm− 3 for c⁎, A1⁎, and A2⁎. where b and cd are the Gompertz model coefficients, and e0 and e25

66
H. Bayat et al. Geoderma 318 (2018) 64–77

represent the initial and void ratio at a stress of 25 kPa, respectively. suctions of 10, 30, 100, 200, 400, 800, and 1500 kPa using a pressure
plate apparatus (Or and Wraith, 2002). The measurements were con-
2. Materials and methods ducted using the undisturbed soil samples up to matric suctions of
100 kPa, but with the disturbed soil samples for matric suctions over
2.1. Sampling 100 kPa. The Dexter et al. (2008) model was then fitted to the experi-
mental data for the SWRC using Solver Excel 2010 and the parameters
In this study, 150 disturbed and undisturbed soil samples were were obtained, where its fitting capacity was compared with that of the
collected from 150 sampling locations in five provinces of Iran (30 van Genuchten (1980) SWRC model.
samples from each province), i.e., Kermanshah, Mazandaran, West
Azarbaijan, East Azarbaijan, and Hamedan. Disturbed and undisturbed 2.2.2. Confined compression test
soil samples were collected from each location by using stainless steel The undisturbed samples were used in this experiment after re-
cylinders (5.3 cm internal diameter and 4.5 cm height) to obtain the moving the samples from the pressure plate at a suction of 100 kPa. The
undisturbed soil samples. These samples were collected from the sur- compression curve was measured with the undisturbed samples.
face (0–30 cm depth) soils. The soil sampling locations are shown in According to Koolen and Kuipers (1989), undisturbed samples should
Fig. 2. The sampling process was based on the principle that the max- be used to measure CCC because they exhibit minimal changes in the
imum difference existed among the soils because the suitability of the soil structure and porosity, as well as simulating the field conditions
models would be enhanced by increasing the difference among the soils more accurately. In addition, the aim of this study was to estimate the
(Tomasella et al., 2003). Therefore, information regarding the soil SWRC from CCC, and the soil structure is one of the main factors that
texture, structure, color, and series for the sampling sites was acquired affects moisture retention (Dexter et al., 2008) as well as soil compac-
from the agricultural research center in each province. Subsequently, tion (Chan et al., 2006; Soane and Van Ouwerkerk, 1994), so the un-
samples were collected from the dominant soil series in each province, disturbed samples were used to measure the CCC. It should also be
preferably with high textural and structural diversity, during the spring noted that application of matric suction at 100 kPa affected the CCC of
of 2011 when the soil moisture was at the field capacity. the soil samples because matric suction affects the pre-compression
All five provinces contain impassable mountainous areas so soil stress and soil moisture content (Horn et al., 1998).
sampling from these areas (mountainous area) was not possible and In the CCT, the soil was compacted inside a steel cylinder via the
there were also long distance between the sampling sites in some cases. downward movement of a piston. Moreover, the stress on the piston
According to the land use types, West Azerbaijan province included and changes in the soil volume were recorded gradually (Koolen and
orchard as well as farmland under wheat, barley, and hay cultivation. Kuipers, 1989). The CCT was performed using a uniaxial device (Cali-
East Azarbaijan province included more agricultural land with the same fornia Bearing Ratio, CBR). This experiment was conducted in loading
cultivation types as West Azerbaijan province. Mazandaran province and unloading stages, where the height reduction reading was obtained
contained forest lands, orchards, and farmlands with dominant culti- at intervals of 10 kPa to approximately 1000 kPa stress in the loading
vation of rice. Hamedan and Kermanshah included a limited area of stage and at intervals of 30 kPa in the unloading stage (Koolen, 1974).
orchard and agricultural lands with dominant cultivation of wheat, The height reduction (ΔH) in the soil samples caused by the applied
barley, alfalfa, and cucurbits. Some characteristics of the soil sampling stress was measured using another gauge (this device provided a visual
sites are shown in Table 1. display of the measured level or the height) simultaneously with im-
posed stresses at intervals of 10 kPa. The void ratio (or volume change)
2.2. Measurements corresponding to each stress was obtained using Eq. (6). The soil
samples were oven dried at 105 °C for 48 h after performing the CCT.
2.2.1. Soil physical properties After performing this test, the stress/void ratio curve or CCC in the
Hydrometer and dry-sieving methods were used to determine the unsaturated soil was obtained where the decrease in the soil volume
particle size distribution in the range of clay (0–0.002 mm), silt after increasing the stress was due to the outflow of air from the soil
(0.002–0.05 mm), and sand (0.05–2.0 mm), according to the United (Koolen and Kuipers, 1989):
States Department of Agriculture (USDA) system (Gee and Or, 2002).
ΔH Δe
The core method was used to measure the bulk density (BD) (Grossman = e = e0 − Δe,
H 1 + e0 (6)
and Reinsch, 2002).
To obtain the SWRC, the moisture contents were measured at matric where H (cm) is the initial height of the soil sample, e0 (cm3 cm− 3) is
suctions of 0, 1, 2, 4, and 6 kPa using a sand box, as well as at matric the initial void ratio, e (cm3 cm− 3) is the void ratio equivalent to any

Table 1
Characteristics of the soil sampling sites.

Kermanshah Hamedan Mazandaran East Azarbaijan West Azarbaijan

Temperature (°C) 13.4 9.74 18 13.5 10.4


Total rainfall 438 300 977 315.5 373.3
(mm year− 1)
Temperature regime Mesic Mesic Thermic Mesic Mesic
Moisture regime Xeric Xeric Udic/Xeric Xeric Xeric

Taxonomy Entisols, Vertisols, Inceptisols, Mollisols Inceptisols, Entisols Mollisols, Inceptisols, Entisols, Alfisols Inceptisols, Entisols Inceptisols, Entisols

Longitude Min 45°24' 47°48' 50°21' 45°05' 44°03'


Max 48°06' 49°28' 58°08' 48°22' 47°23'
Latitude Min 33°41' 34°00' 35°46' 36°45' 35°08'
Max 35°17' 35°44' 36°58' 39°26' 39°46'

67
H. Bayat et al. Geoderma 318 (2018) 64–77

Table 2 hidden and output layers.


Input variables at each level used to build PTFs. The uncertainty of the predicted parameters in the Dexter SWRC
model was evaluated using a combination of artificial neural networks
Level Predictor parameters
and the Delete – d method (Good, 2001). In the Delete – d method, a
1 Clay, silt/sand and BDa random percentage d of the observations is set aside for validation
2 Clay, silt/sand, BD + Pc, Cc, Cs purposes (i.e., cross-validation and/or testing), and the remaining 100 –
3 Clay, silt/sand, BD + parameters of Gompertz model (a, b, cd and m)
d% are used as the development set. The input data were selected
4 Clay, silt/sand, BD + parameters of Dexter CCC model (A1⁎, A2⁎, σ1⁎ and
σ2⁎) randomly from all of the samples 100 different times in order to obtain
5 Clay, silt/sand and BD as well as peer to peer parameters of Dexter CCC 100 data sets with the same size (96 samples) for use as the training
model data set, and the remaining samples were employed in the cross-vali-
a
dation and testing steps. A network was trained for each data set and
BD is the soil bulk density, Cc is the compression index, Cs is the swelling index, Pc is
the SWRC parameters in the Dexter SWRC model were estimated. The
the pre-compression stress, and a, b, cd, and m are parameters in the Gompertz model.

Asterisks denote parameters obtained from fitting the Dexter CCC model to the stress/
mean of all 100 predictions was treated as the final estimate (Good,
void ratio data. 2001).

applied stress, Δe (cm3 cm− 3) is the void ratio change, and ΔH (cm) is 2.4. Evaluation criteria
the sample height change. The Gompertz (1825) equation and the
Dexter model modified according to Eq. (2) in order to fit to the CCC Akaike's information criterion (AIC), relative improvement (RI), and
(stress/void ratio) were fitted to the experimental CCC data using root mean squared error (RMSE) were used to evaluate the errors made
Solver Excel 2010, and their parameters were obtained. Cc and Cs were by the PTFs when estimating the parameters for the Dexter SWRC
calculated using Eqs. (4) and (5), respectively. Pc was obtained using model. After estimating the parameters for the Dexter SWRC model in
the graphical method described by Casagrande (1936). each step, they were used to simulate the SWRC for the soil samples in
the related step in order to evaluate the errors made by the PTFs in that
step. The simulated curves were then compared with the measured (or
2.3. Development of PTFs
fitted) curves (curve by curve) based on the AIC, coefficient of de-
termination (R2), RI, integral root mean square error (IRMSE) (g g− 1),
Table 2 shows the input variables for five levels (stages or steps). All
and integral mean error (IME) (g g− 1). The AIC (Akaike, 1974) and the
of the inputs for each level were used to predict each parameter sepa-
other metrics were calculated as follows:
rately in the Dexter et al. (2008) SWRC model. These parameters were
obtained by fitting the Dexter et al. (2008) SWRC model to the SWRC ∑ (θp − θm )2 ⎞
experimental data. Thus, multiple PTFs were developed to estimate AIC = N1n ⎛⎜ ⎟ + 2q
⎝ N ⎠ (7)
each parameter individually (i.e., a PTF based on artificial neural net-
works with one output neuron). The soil texture and BD are the most
common predictors used in PTFs to estimate the SWRC (Husz, 1967). ∑ (θm − θp )2
R2 = 1 −
Therefore, at the first level, these variables were used as predictors and ∑ (θm − θmean )2 (8)
the other levels were compared with them. The input variables at the
other levels were selected to test the utility of the compression curve ∑ (pf − pp ) 2
parameters for improving the water content estimates. RMSE =
N (9)
At the fifth level, every parameter in the modified Dexter CCC
model fitted to the CCC data was used to predict the same parameter in b
the Dexter SWRC model. Thus, the same parameters were used as inputs 1
IME = ∫ (θp − θm )d log |h|
(from CCC)/outputs (from SWRC) in the PTFs developed for the fifth b−a a (10)
level. Therefore, we used the CCC data to predict the SWRC.
The coefficient c (Baumgartl and Koeck, 2004) in the Dexter model b
1
2
was 0.27 (cm3 cm− 3) for the CCC in all of the samples but a fixed ⎡ 1 ⎤
number cannot be used as a predictor, so the other four parameters in
IRMSE = ⎢
b−a
∫ (θp − θm )2d log |h| ⎥
⎣ a ⎦ (11)
the Dexter CCC model were used as inputs to estimate the coefficient c
for the Dexter SWRC model. The detailed process employed for pre- IRMSE1 − IRMSEi
dicting the SWRC using the CCC parameters as predictors is shown in RI = × 100 i = 2…,5,
IRMSE1 (12)
Fig. 3.
First, the normality of the variables was confirmed with the where IRMSE1 is the IRMSE for the first step and IRMSEi is the IRMSE
Kolmogorov–Smirnov test using Minitab (Minitab, 2000) and non- for the second to fifth steps.
normal variables were transformed into normal variables, where the In the equations above, pf and pp are the fitted and estimated
different transformations included Log x, ln x, x2, and x0.5, as well as parameters, respectively, q is the number of parameters that should be
some regression equations. Normality was also tested for the trans- estimated in the model, N is the number of observation points for the
formed forms of the variables and the transformation with the highest SWRC for every soil sample when calculating the overall evaluation
p-value in the normality test was selected as the transformation for each criteria for each step, and N is the number of soil samples used for
non-normal variable. All of the variables were standardized with a zero calculating the evaluation criteria to estimate the Dexter SWRC model
mean and unit variance. parameters, θm, θp and θmean, are the measured, estimated, and mean
MATLAB (MathWorks, 2005) was used to develop the PTFs based on value of measured gravimetric/volumetric water content, respectively,
artificial neural networks, where 64% (96 samples), 9% (14 samples), and h is the matric suction (kPa).
and 27% (40 samples) of the samples were selected randomly for the The AIC was calculated to evaluate the accuracy of the estimated
training, cross-validation, and testing steps, respectively. Multilayer Dexter SWRC model parameters, whereas the fitted and estimated va-
perceptron artificial neural networks were used with one hidden layer, lues of the parameters were used instead of θm and θp, respectively, in
10 hidden units, and sigmoid tangent activation functions in both the Eq. (7).

68
H. Bayat et al. Geoderma 318 (2018) 64–77

Fig. 3. Detailed process used to predict the SWRC with the CCC parameters as predictors. BD is the soil bulk density, Si/S is the silt to sand ratio, C is the clay content, Cc is the
compression index, Cs is the swelling index, Pc is the pre-compression stress, and a, b, cd, and m are the parameters in the Gompertz model. c is the residual water content, where the
matrix and structural pore space amounts are proportional to A1 and A2, respectively. The values of h1 and h2 are the characteristic pore water suctions at which the matrix and structural
pore spaces are empty, respectively. c⁎, σ1⁎, σ2⁎, A1⁎ and A2⁎ are the fitting parameters for the Dexter CCC model. CCC and SWRC are the confined compression curve and soil water
retention curve, respectively. PTFs are the pedotransfer functions.

3. Results and discussion considering the variability and number of soil samples used in this
study, the developed PTFs should have adequate reliability.
3.1. Statistical properties of the variables considered The statistical properties of the variables studied for the 150 soil
samples are shown in Table 3. The textural distribution of the soil
The reliability and robustness of the PTFs will be higher when the samples in the USDA texture triangle (Fig. 4) demonstrates the high
number of soil samples used to develop them is larger. In addition, variation in the textures of the soils used in this study.
when estimating less accessible properties, the suitability of PTFs for At the beginning of this study, both the van Genuchten and Dexter
different conditions increases as the ranges of the variables used for models were fitted to the SWRC experimental data, but the results de-
developing the PTFs increase (Tomasella et al., 2003). In this study, soil monstrated the better performance of the Dexter model (higher R2 and
samples were collected from provinces with very different soil types lower RMSE values) compared with the van Genuchten model
and climatic conditions in order to enhance the generalizability of the (Table 4).
developed PTFs. Table 4 compares the fitting accuracy of the Dexter model to the
Mayr and Jarvis (1999) reported that the use of small, relevant, and stress/void ratio (CCC) and the moisture/suction (SWRC) data. The
appropriate samples (if available) is better than using excessive fitting accuracy of the Dexter SWRC model was relatively better than
amounts of general data. In addition, Hillel (1980) suggested that the that of the modified Dexter CCC model (Table 4). Moreover, the
appropriate number of soil samples is 98 for estimating the moisture at Gompertz model was more accurate when fitted to the stress/void ratio
a matric suction of 1500 kPa (which has the highest variability) in the data than the modified Dexter CCC model. Thirteen previously reported
range of 10% of the mean (p < 0.05). Pachepsky and Rawls (1999) SWRC models were fitted by Bayat et al. (2013) and their results
also reported that PTFs developed using the group method for data showed that the Dexter model has the highest precision when fitted to
handling with > 100 samples exhibited high reliability. Thus, the SWRC data. The mean value of R2 obtained by Bayat et al. (2013)

69
H. Bayat et al. Geoderma 318 (2018) 64–77

Table 3 Table 4
Statistical properties of the input and output variables. Comparison of the fitting accuracy for the different models according to the moisture/
suction and stress/void ratio data.
Parameter Max Mean Min SD
Model R2 RMSE
Sand (%) 90.0 32.8 5.9 16.9
Silt (%) 79.5 43.5 2.5 15.8 Mean Max Min SD Mean Max Min SD
Clay (%) 65.6 23.1 3.5 13.8
BD (g cm− 3)a 1.95 1.4 1.0 0.2 Dexter SWRC 0.980 0.997 0.777 0.024 0.012 0.178 0.003 0.015
Pc (kPa) 660.7 339.9 70.8 131.9 model-Solver
Cc 0.858 0.469 0.117 0.167 Van Genuchten 0.936 0.999 0.673 0.085 0.028 0.170 0.001 0.025
Cs 0.028 0.005 0.0007 0.004 SWRC model
a 0.987 0.212 0.000 0.223 Modified Dexter 0.959 0.999 0.773 0.045 0.039 0.656 0.001 0.093
b 7.737 2.173 0.186 0.733 CCC model
cd 3.160 0.642 0.160 0.399 Gompertz model 0.981 0.999 0.228 0.075 0.022 0.604 0.001 0.086
m 8.354 3.258 1.213 1.001
c (g g− 1) 0.290 0.141 0.018 0.048
A1 (g g− 1) 0.451 0.146 0.023 0.064 (Table 5). Fig. 5 shows how the distribution of each attribute was
h1 (cm) 6046 1954 1039 754
normalized.
A2 (g g− 1) 0.303 0.105 0.019 0.050
h2 (cm) 182.90 52.18 9.86 28.26 Dexter et al. (2008) stated that the parameters of their model have a
A1⁎ (cm3 cm− 3) 1.086 0.505 0.074 0.197 physical basis.
σ1⁎ (cm) 46,248 12,885 2303 8527 We found that the compression index had a significant (p < 0.05)
A2⁎ (cm3 cm− 3) 1.438 0.134 0.000 0.276
positive correlation with the parameter c (Table 5). The significant
σ2⁎ (cm) 8375.6 1009.8 9.186 1702.3
correlation between Cc and parameter c can be interpreted as indicating
a
BD is the soil bulk density, Cc is the compression index, Cs is the swelling index, Pc is that increasing the clay content in the soil increased the value of Cc
the pre-compression stress, and a, b, cd, and m are parameters in the Gompertz model. c is (Gregory et al., 2006), and thus the parameter c would also increase.
the residual water content, where the matrix and structural pore space amounts are Therefore, there was a significant positive correlation between the two
proportional to A1 and A2, respectively. The values of h1 and h2 are the characteristic pore parameters c and Cc. There were significant correlations between the
water suctions at which the matrix and structural pore spaces are empty, respectively. σ1⁎,
CCC parameters (or characteristics) as inputs and the parameters in the
σ2⁎, A1⁎, and A2⁎ are the fitting parameters for the Dexter CCC model.
Dexter SWRC model as outputs. These correlations may be effective for
improving estimations of the water content. There was a physical re-
lationship between the parameters obtained from the Dexter SWRC and
was equal to 0.968, which was less accurate compared with the results CCC models, but the similarity between the two curves probably ex-
obtained in the present study. plained these correlations.

3.2. Correlations between the parameters in the Dexter SWRC model and 3.3. Estimations of the parameters in the Dexter SWRC model
the input variables
3.3.1. Parameter c
Correlation analysis was performed between the input and output Table 6 shows the estimated results obtained for the parameter c in
parameters using the normal and/or normalized forms of the variables the Dexter SWRC model in the training and testing steps. The average

Fig. 4. Textural distribution of the soil samples on the USDA soil


texture triangle.

70
H. Bayat et al. Geoderma 318 (2018) 64–77

Table 5 soil texture (Dexter et al., 2008), and the parameters used as predictors
Correlations between the parameters in the Dexter SWRC model and the input variables. in this study are directly or indirectly correlated with the soil texture
and structure. For instance, Bronick and Lal (2005) stated that in-
c A1 h1 A2 h2
(g g− 1) (g g− 1) (cm) (g g− 1) (cm) creasing the clay content of the soil improves the soil structure. The
relationships between the clay content, Pc, Cc, and Cs have also been
Sand (%) −0.430⁎⁎ − 0.256⁎⁎ − 0.006 − 0.078 −0.038 widely reported (Imhoff et al., 2004). Therefore, there was generally a
Silt (%) −0.008 0.283⁎⁎ 0.013 0.107 0.100
physical correlation between the predictors and the parameter h1.
Log (Clay, %) 0.534⁎⁎ 0.018 0.021 − 0.014 −0.023
BD (g cm− 3)a −0.533⁎⁎ − 0.457⁎⁎ − 0.043 0.399⁎⁎ −0.086
1.9364ln(Pc, −0.046 − 0.090 − 0.129 − 0.004 0.284⁎⁎ 3.3.4. Parameter A2
kPa)–10.637 The second level obtained the most accurate estimates of the
Cc 0.156⁎ 0.166⁎ − 0.050 0.205⁎⁎ 0.096 parameter A2 and there were improvements of 15% and 10% in the
0.6783lnCs + 3.21 −0.090 − 0.045 0.007 0.053 0.244⁎⁎
training and testing steps, respectively (Table 6). At the second level,
a 0.260⁎⁎ 0.078 − 0.077 0.015 −0.133
b 0.055 − 0.018 0.162⁎ 0.057 0.180⁎ clay, silt/sand, BD, Pc, Cc, and Cs were used as predictors. The sig-
cd 0.060 0.257⁎⁎ 0.392⁎⁎ 0.052 −0.129 nificant correlations between the parameter A2 with Cc and BD prob-
m 0.062 − 0.023 0.059 0.030 0.103 ably explained the appropriate estimates obtained at the second level.
A1⁎ (cm3 cm− 3) 0.164⁎ 0.144 0.080 0.136 −0.055
The addition of the parameters in the Gompertz model at the third level
σ1⁎ (cm) 0.165⁎ 0.029 − 0.115 − 0.023 −0.001
A2⁎ (cm3 cm− 3) −0.031 − 0.003 0.064 − 0.001 0.072 improved the testing step results by 5.6%. However, this improvement
σ2⁎ (cm) −0.003 − 0.209⁎ 0.224⁎ − 0.175⁎ −0.192⁎ was not significant according to the AIC. The mechanical properties of
soil and the parameter A2 are quantitative soil structure and soil sta-
⁎ ⁎⁎
, Indicate significant correlations at 5% and 1% levels, respectively. bility parameters. Thus, there are physical relationships among them
a
BD is the soil bulk density, Cc is the compression index, Cs is the swelling index, Pc is and using the mechanical properties as predictors improved the esti-
the pre-compression stress, and a, b, cd, and m are the parameters in the Gompertz model.
mates of A2.
c is the residual water content, where the matrix and structural pore space amounts are
proportional to A1 and A2, respectively. The values of h1 and h2 are the characteristic pore
water suctions at which the matrix and structural pore spaces are empty, respectively. σ1⁎, 3.3.5. Parameter h2
σ2⁎, A1⁎, and A2⁎ are the fitting parameters for the Dexter CCC model. At the second, third, and fourth levels, using the input parameters
obtained directly or indirectly from the CCC yielded significant im-
value of the parameter c, which is equal to the residual moisture con- provements in the estimates of parameter h2 in both the training and
tent, was estimated as 0.150 for all of the levels except the third level. testing steps compared with the first level (Table 6). At the fifth level,
At the third level, the value of this parameter was considerably lower, the accuracy of the estimates decreased and the RI values were nega-
where it was estimated as less than the usual values, and the second tive.
level obtained the best results for the parameter c. At this level, the Dexter (2004) stated that the pore size distribution in a soil is a
estimation was performed using clay, silt/sand, BD, and the CCC vari- function of the soil texture and structure. The slope of the SWRC re-
ables (Pc, Cc, and Cs) as the predictors. The appropriate accuracy of the presenting the soil physical quality parameters will vary due to soil
estimated values of the parameter c obtained at the second level may be stress. Richard et al. (2001) reported the same results as Dexter (2004).
explained by its correlation with the parameter Cc (Table 5). The findings obtained in these studies illustrate the relationships be-
Moreover, the presence of the parameter Pc among the estimators tween the parameter h2 and the predictors obtained from the CCC.
had a considerable impact by increasing the accuracy at the second
level. This is because the two parameters c and Pc are influenced sub- 3.4. Curve-by-curve comparison of the water content estimates obtained
stantially by the clay content (Dexter et al., 2008; Lebert and Horn, using the Dexter SWRC model
1991). The improvement in the estimation of the parameter c was
significant at the second and fourth levels in the testing step (Table 6). In Tables 7 and 8 and Fig. 6, we compare the estimated moisture
Our results indicated that there was a physical association between the obtained by the Dexter SWRC model at five levels. We performed curve-
parameter c and indices derived from the CCC, and the appropriate by-curve comparisons of the water contents at different matric suctions
accuracy of the estimations obtained with the SWRC model parameters for the estimated curves at various levels together with the fitted
can be explained based on this physical relationship. curves.
We also developed van Genuchten-based parametric PTFs to esti-
3.3.2. Parameter A1 mate the SWRC. However, the developed PTFs had higher IRMSE and
The first level estimations of the parameter A1 were less accurate IME values but lower R2 values compared with the Dexter et al. model-
than those at the other levels in both the training and testing steps based parametric PTFs when estimating the SWRC. The Dexter et al.
(Table 6), possibly due to the weaker capacity of the predictors at the (2008) model is based on the soil aggregate and soil structure formation
first level (clay, silt/sand, and BD) compared with the other levels when process, which yields a bimodal distribution for the soil pore size. This
estimating A1. In the testing steps at the second, third, and fifth levels, model described the retention of water well in the agricultural soils
there were significant improvements in the accuracy of the estimations considered in this study, which had a bimodal structure. Therefore, the
compared with the first level. The correlation between A1 and Cc Dexter et al. (2008) model could obtain better estimates of the moisture
(Table 5) probably explained the significant improvement at the second retention curve (Tables 6, 7 and 8) and it described the CCC char-
level in the testing step. The third level used the four coefficients from acteristics well (Table 4). The van Genuchten-based parametric PTFs
the Gompertz model and the first level inputs as predictors, and the performed poorly so their results are not reported.
greatest improvement was obtained (significant at p < 0.05) in the
accuracy at this level compared with the first level in the testing step. 3.4.1. First level
At the first level, we used the texture information and bulk density
3.3.3. Parameter h1 (clay, silt/sand, and BD) as predictors and satisfactory results were
In the testing step, there were significant improvements in the ac- obtained. Dexter (2004) stated that an increase in the percentage of
curacy of the estimates obtained for the parameter h1 at the fourth and clay in the soil causes a decrease in the SWRC slope. Furthermore,
fifth levels (Table 6). The fourth and fifth levels obtained very similar Sheng and Zhou (2011) showed that the BD has various effects on the
estimated results for the parameter h1 and the accuracy was highest SWRC by changing the soil porosity. Patil et al. (2012) estimated the
compared with the other levels. The parameter h1 is correlated with the water content using the Campbell model with similar predictors at the

71
H. Bayat et al. Geoderma 318 (2018) 64–77

Fig. 5. Transformation of the non-normal variables


comprising silt/sand, clay, pre-compression stress (Pc)
and swelling index (CS). The normalization functions
are shown in the graphs.

72
H. Bayat et al. Geoderma 318 (2018) 64–77

Table 6
Estimated results obtained for the Dexter SWRC model parameters.

Parameter Level Train Test

Mean RMSE AIC RI Mean RMSE AIC RI


(%) (%)

c (g g− 1)a 1 0.148 0.055 − 426 0.142 0.053 − 432


2 0.142 0.038 − 484⁎ 32.2 0.145 0.037 − 485⁎ 30.1
3 0.003 0.152 − 277 −174.5 0.002 0.144 − 285 −169.4
4 0.150 0.048 − 448⁎ 14 0.152 0.040 − 475⁎ 25.4
5 0.156 0.048 − 447⁎ 13.3 0.157 0.052 − 437 3.1
A1 (g g− 1) 1 0.146 0.062 − 410 0.145 0.063 − 407
2 0.146 0.059 − 417 5 0.153 0.054 − 430⁎ 14.2
3 0.145 0.055 − 427 11.3 0.154 0.049 − 444⁎ 21.7
4 0.147 0.052 − 436⁎ 16.4 0.144 0.060 − 414 4.4
5 0.149 0.059 − 418 5.5 0.240 0.050 − 441⁎ 20.1
h1 (cm) 1 1739 0.992 0.07 1739 1.035 3.835
2 1939 0.973 − 5.2⁎ 2 1935 1.156 9.123 −11.6
3 1971 0.924 − 16.6⁎ 6.8 1972 1.113 7.642 −7.5
4 1908 0.746 − 41.8⁎ 24.8 1912 0.839 − 11.018⁎ 18.9
5 1888 0.746 − 44.4⁎ 24.8 1883 0.942 − 5.008⁎ 8.9
A2 (g g− 1) 1 0.103 0.038 − 483 0.100 0.039 − 477
2 0.105 0.032 − 508⁎ 15.1 0.0103 0.036 − 492 9.9
3 0.106 0.032 − 506 14.2 0.104 0.037 − 485 5.6
4 0.109 0.036 − 490 4.4 0.110 0.038 − 481 2.6
5 0.110 0.032 − 507⁎ 14.9 0.195 0.037 − 485 5.2
h2 (cm) 1 53 0.989 − 1.2 53 1.102 5.645
2 50 0.961 − 3.4⁎ 2.8 50 1.008 1.118⁎ 8.5
3 50 0.917 − 16⁎ 7.3 49 1.061 5.035⁎ 3.7
4 62 0.909 − 13.4⁎ 8.1 63 1.038 3.332⁎ 5.8
5 62 1.068 8 −7.9 64 1.124 6.101 −1.9


Indicates a significant improvement (p < 0.05) compared with the first model.
a
c is the residual water content, where the matrix and structural pore space amounts are proportional to A1 and A2, respectively. The values of h1 and h2 are the characteristic pore
water suctions at which the matrix and structural pore spaces are empty, respectively.

first level and reported an R2 value of 0.95, which is inferior to the The appropriate estimates obtained at the second level may be ex-
result obtained using the Dexter SWRC model in the present study. The plained by the accurate estimations of the parameters at this level
superior performance of the Dexter SWRC model compared with the (Table 6), or by the incorporation of soil structure information in the
Campbell model was probably due to the physical parameters con- model inputs. During the estimation of most of the Dexter SWRC model
sidered in the Dexter SWRC model. Ebrahimi et al. (2014a) estimated parameters, the second level obtained significant increases in accuracy,
the Dexter SWRC model using sand, clay, and BD, and the IRMSE value thereby indicating the utility of the parameters derived from the com-
obtained (0.335) was higher than that in the present study. pression curve (Pc, Cc, and Cs) when estimating moisture using the
Dexter SWRC model. In addition, the significant correlations between
the parameters in the Dexter SWRC model and the indices derived from
3.4.2. Second level
the CCC (Table 5) may also explain the appropriate estimates obtained.
The use of three parameters derived from the compression curve (Pc,
Figs. 6 and 7 show the good results obtained at the second level.
Cc, and Cs) as well as clay, silt/sand, and BD as predictors at the second
Sun et al. (2014) reported that the moisture content near saturation
level obtained a significant improvement. This improvement was re-
increases in a linear manner with the void ratio. Previous studies have
lated to the accuracy of the moisture estimates produced by the Dexter
shown that there is a close relationship between the soil moisture
SWRC model in both the training and testing steps (Table 7 and Fig. 6)
content and the soil mechanical properties. In addition, the stress on
compared with the first level.

Table 7
Estimated results obtained for the gravimetric water content using the Dexter SWRC model at five levels.

Level IRMSE IME R2 AIC RI


(g g− 1) (g g− 1) (%)

Training 1 0.059 (0.04)a − 0.001(0.07) 0.978(0.03) − 1567


2 0.052(0.05) − 0.003(0.07) 0.977(0.02) − 1582 11.9⁎
3 0.157(0.07) − 0.160(0.08) 0.977(0.04) − 1048 − 166.1⁎
4 0.053(0.04) 0.005(0.06) 0.980(0.03) − 1612 10.2⁎
5 0.055(0.04) 0.007(0.06) 0.979(0.03) − 1617 6.8⁎
Testing 1 0.061(0.05) − 0.006(0.07) 0.981(0.02) − 1590
2 0.058(0.05) 0.004(0.07) 0.977(0.02) − 1603 4.9⁎
3 0.155(0.08) − 0.160(0.08) 0.977(0.02) − 1100 − 154.1⁎
4 0.056(0.04) 0.010(0.07) 0.979(0.03) − 1667 8.2⁎
5 0.065(0.05) 0.005(0.08) 0.975(0.02) − 1573 − 6.6⁎


Indicates a significant difference (p < 0.05) compared with the first model.
a
Values in parentheses are the standard deviations.

73
H. Bayat et al. Geoderma 318 (2018) 64–77

Table 8
Estimated results obtained for the volumetric water content using the Dexter SWRC model at five levels.

Level IRMSE IME R2 AIC RI


(cm3 cm− 3) (cm3 cm− 3) (%)

Training 1 0.086(0.066)a 0.007(0.10) 0.978(0.03) − 1356


2 0.073(0.057) 0.003(0.09) 0.977(0.03) − 1391 15.11⁎
3 0.214(0.083) − 0.218(0.10) 0.977(0.03) − 881 − 148.80⁎
4 0.075(0.057) 0.009(0.09) 0.980(0.02) − 1408 12.79⁎
5 0.075(0.069) 0.012(0.11) 0.979(0.02) − 1411 12.79⁎
Testing 1 0.085(0.069) − 0.005(0.10) 0.981(0.02) − 1406
2 0.083(0.066) 0.007(0.10) 0.977(0.02) − 1397 2.30
3 0.219(0.105) − 0.226(0.11) 0.977(0.02) − 893 − 157.6⁎
4 0.079(0.060) 0.018(0.09) 0.943(0.03) − 1452 7.05⁎
5 0.093(0.054) 0.011(0.09) 0.940(0.03) − 1365 − 9.41⁎


Indicates a significant difference (p < 0.05) compared with the first model.
a
Values in parentheses are the standard deviations.

soil affects the void ratio (Baumgartl and Koeck, 2004; Sheng and Zhou, Many studies have shown that there is a relationship between the CCC
2011). Sun et al. (2007) reported that soil moisture is a function of both and SWRC. For example, Sheng and Zhou (2011) demonstrated that a
the soil suction and soil void ratio. Therefore, using these properties change in the initial status of the void ratio led to a change in the air entry
that affect the void ratio as predictors will increase the accuracy of the suction, and thus the shape and position of the SWRC. Therefore, using
soil moisture content estimates. the CCC indices can enhance the accuracy of SWRC estimates.

Fig. 6. Scatter plots of the measured versus estimated volu-


metric water contents for the PTFs.

74
H. Bayat et al. Geoderma 318 (2018) 64–77

Fig. 7. Comparison of the measured water contents, the fitted SWRCs based on the Dexter SWRC model, and the estimated SWRCs obtained at five levels for four soil samples.

3.4.3. Third level using different predictors, and large error values were found in no
The use of the four coefficients in the Gompertz equation as pre- cases. Thus, further research is needed in this area.
dictors as well as the soil basic properties (clay, silt/sand, and BD) at
the third level did not improve the accuracy of the moisture content
3.4.4. Fourth level
estimates obtained by the Dexter SWRC model in both the training and
At the fourth level, the use of the modified Dexter CCC model
testing steps compared with the first level (Tables 7 and 8, Fig. 6 and 7),
parameters as well as the soil basic properties as predictors significantly
possibly due to underestimation of the parameter c (Table 6). At this
improved the accuracy of the moisture content estimates obtained by
level, the estimate of parameter c was not appropriate or sufficiently
the Dexter SWRC model compared with the first level. The high accu-
accurate and its values were much less than the fitted values. Moreover,
racy of this level can probably be explained by the similarity of the
according to the Dexter SWRC model formula, this parameter has a
equations fitted to the CCC and SWRC, the effect of the pore size dis-
great impact on estimates of the moisture content. If the value of this
tribution in soils (Sillers et al., 2001), and the soil structure (Hamza and
parameter is less than the usual value, the estimates obtained from the
Anderson, 2005) on both the SWRC and CCC, and the proximity of the
entire curve shift to a lower moisture level and the curve will be dif-
two curves. Ng and Pang (2000) and Lee et al. (2005) demonstrated
ferent from the measured curve, as shown by the graphs plotted for the
that the degree of saturation is a function of the suction and average
Dexter SWRC model (Fig. 6 and 7).
stress in soil.
Dexter et al. (2008) reported that an average value of 0.085 was
obtained for the parameter c by fitting. However, the mean value es-
timated for parameter c at the third level was 0.002. Therefore, the use 3.4.5. Fifth level
of applied predictors (soil basic properties and the four coefficients in At the fifth level, the peer-to-peer coefficients of the Dexter CCC
the Gompertz model) at the third level is not recommended for esti- model with the inputs of the first level were used as predictors to es-
mating this parameter. After studying the soils in Guilan province, Iran, timate the coefficients of the Dexter SWRC model. According to the
Ebrahimi et al. (2014a) evaluated the estimation of this parameter coefficient for the RI, the accuracy of the estimates decreased in the
testing step. Therefore, using only one parameter in the Dexter CCC

75
H. Bayat et al. Geoderma 318 (2018) 64–77

model as well as the soil basic properties as predictors to estimate the same factors. The parameters related to the soil texture and structures
parameters of the Dexter SWRC model failed to enter sufficient in- were effective for controlling both of the curves. Furthermore, the re-
formation in the PTFs to estimate the soil water content. sults plotted in the figures indicated that the SWRC estimates obtained
Fig. 7 shows the measured water contents, the fitted SWRCs based using the predictors of the CCC agreed well with the measured SWRC as
on the Dexter SWRC model, and the estimated SWRCs obtained from well as confirming the results in the tables.
the five levels for six soil samples, which were selected randomly from
150 soil samples. Fig. 6 also shows the measured versus the estimated Acknowledgements
water contents obtained at the five levels. These figures indicate the
high accuracy of the water content estimates obtained by the Dexter The authors are grateful to the anonymous reviewers and the editors
SWRC model using the CCC parameters as predictors, except for the for their helpful comments regarding the manuscript.
third level where the underestimation of parameter c shifted the esti-
mated curve at the third level to a low water content. However, the References
other estimated curves agreed well compared to the fitted SWRCs.
In general, the accuracy of the water content estimates was im- Akaike, H., 1974. A new look at the statistical model identification. IEEE Trans. Autom.
proved by using the mechanical properties of soils as predictors. Ohu Control 19 (6), 716–723.
Baumgartl, T., Koeck, B., 2004. Modeling volume change and mechanical properties with
et al. (1987) reported that the soil mechanical properties had various hydraulic models. Soil Sci. Soc. Am. J. 68 (1), 57–65.
impacts on the SWRC, especially compression. Ohu et al. (1987) also Bayat, H., Neyshabouri, M., Mohammadi, K., Nariman-Zadeh, N., 2011. Estimating water
showed that the use of parameters derived from compression is useful retention with pedotransfer functions using multi-objective group method of data
handling and ANNs. Pedosphere 21 (1), 107–114.
for estimating water retention, where they can enhance the accuracy of Bayat, H., Neyshaburi, M.R., Mohammadi, K., Nariman-Zadeh, N., Irannejad, M., 2013.
the estimates. Romero et al. (2001) and Chateau and Viet (2009) found Improving water content estimations using penetration resistance and principal
that the soil mechanical properties can affect the pore size distribution, component analysis. Soil Tillage Res. 129, 83–92.
Bronick, C.J., Lal, R., 2005. Soil structure and management: a review. Geoderma 124 (1),
pore shape, and other chemical and physical properties of soil, thereby
3–22.
affecting the SWRC. Moreover, Horn et al. (1998) reported that the Casagrande, A., 1936. The determination of the pre-consolidation load and its practical
mechanical properties, such as the CCC, are strongly influenced by the significance. In: Proceedings of the International Conference on Soil Mechanics and
Foundation Engineering. Harvard University Cambridge, pp. 60–64.
moisture content, where the soil compressibility decreases as the soil
Cavalieri, K.M.V., Arvidsson, J., da Silva, A.P., Keller, T., 2008. Determination of pre-
moisture content declines. compression stress from uniaxial compression tests. Soil Tillage Res. 98 (1), 17–26.
Overall, the estimations of the Dexter model parameters were im- Chan, K., Oates, A., Swan, A., Hayes, R., Dear, B., Peoples, M., 2006. Agronomic con-
proved in most functions by using the mechanical properties of soils sequences of tractor wheel compaction on a clay soil. Soil Tillage Res. 89 (1), 13–21.
Chateau, X., Viet, T., 2009. Influence of the temperature on the water content curves: a
(Table 6). The improvements in the accuracy of the estimates of the micromechanical approach. In: Buzzi, O., Fityus, S.G., Sheng, D. (Eds.), Unsaturated
Dexter model coefficients ranged from 2% to 32.2% in the training step Soils: Theoretical and Numerical Advances in Unsaturated Soil Mechanics. CRC Press,
and from 2.6% to 30.1% in the testing step, although the accuracy of Taylor and Francis Group, London, UK, pp. 849–854.
Culley, J., Larson, W., 1987. Susceptibility to compression of a clay loam Haplaquoll. Soil
the estimates was reduced with some functions. The use of CCC prop- Sci. Soc. Am. J. 51 (3), 562–567.
erties for estimating the Dexter model coefficients at levels 2–5 yielded Dexter, A., 2004. Soil physical quality: part I. Theory, effects of soil texture, density, and
improvements in the accuracy of 18 and 16 functions (from a total of 20 organic matter, and effects on root growth. Geoderma 120 (3), 201–214.
Dexter, A., Czyż, E., Richard, G., Reszkowska, A., 2008. A user-friendly water retention
models developed to estimate five coefficients for the Dexter et al. function that takes account of the textural and structural pore spaces in soil.
(2008) model) in the training and testing steps, respectively. Therefore, Geoderma 143 (3), 243–253.
the rate of improvement (RI coefficient) and the number of functions Ebrahimi, E., Bayat, H., Neyshaburi, M.R., Zare Abyaneh, H., 2014a. Prediction capability
of different soil water retention curve models using artificial neural networks. Arch.
improved were significant. Agron. Soil Sci. 60 (6), 859–879.
In addition, a comparison of the RI factors for the estimated Ebrahimi, E., Bayat, H., Zare Abyaneh, H., 2014b. Effect of input variables on predict-
moisture content (Table 7) showed that the increase in the accuracy ability of soil water content through different soil water retention curve models.
Appl. Soil Res. 2, 26–42 (In Persian).
ranged from 4.9% to 11.9%. However, the results were not satisfactory
Ekwue, E., 1990. Organic-matter effects on soil strength properties. Soil Tillage Res. 16
in the training and testing steps for the third function, and in the testing (3), 289–297.
step for the fifth function. In general, the results showed that using the Gee, G.W., Or, D., 2002. 2.4 Particle-size analysis. In: Methods of Soil Analysis. Part 4, pp.
CCC characteristics is an appropriate approach for estimating the 255–293.
Gompertz, B., 1825. On the nature of the function expressive of the law of human mor-
parameters of the Dexter model and the soil moisture content, where tality, and on a new mode of determining the value of life contingencies. Philos.
they increased the accuracy significantly. Trans. R. Soc. Lond. 513–583.
Good, P.I., 2001. Resampling Methods.
Gregory, A., Whalley, W., Watts, C., Bird, N., Hallett, P., Whitmore, A., 2006. Calculation
4. Conclusion of the compression index and precompression stress from soil compression test data.
Soil Tillage Res. 89 (1), 45–57.
There are some similarities between the CCC and SWRC. These si- Grossman, R., Reinsch, T., 2002. 2.1 Bulk density and linear extensibility. In: Methods of
Soil Analysis: Part 4 Physical Methods (methods of soil an4), pp. 201–228.
milarities allow the CCC to be used to estimate the SWRC. In this study, Gupta, S., Larson, W., 1979. Estimating soil water retention characteristics from particle
modified Dexter and Gompertz models were fitted to CCC experimental size distribution, organic matter percent, and bulk density. Water Resour. Res. 15 (6),
data. The CCC parameters and characteristics were then used as pre- 1633–1635.
Hamza, M., Anderson, W., 2005. Soil compaction in cropping systems: a review of the
dictors for estimating the water content at five levels with the Dexter nature, causes and possible solutions. Soil Tillage Res. 82 (2), 121–145.
SWRC model. The results showed that the CCC is useful for estimating Hillel, D., 1980. Fundamentals of Soil Physics. Academic Press, Inc. (London) Ltd.
the SWRC and there were significant improvements in the accuracy of Horn, R., 2004. Time dependence of soil mechanical properties and pore functions for
arable soils. Soil Sci. Soc. Am. J. 68 (4), 1131–1137.
the estimated water contents in most cases compared with the first
Horn, R., Richards, B.G., Gräsle, W., Baumgartl, T., Wiermann, C., 1998. Theoretical
level. Overall, the fourth level obtained the best results by using the principles for modelling soil strength and wheeling effects—a review—. Z.
parameters in the Dexter CCC model as well as basic soil properties Pflanzenernähr. Bodenkd. 161 (4), 333–346.
(clay, silt/sand, and BD) as inputs. The improvements in the water Husz, G., 1967. The determination of pF-curves from texture using multiple regressions.
German. Z. Pflanzenernähr. Düng. Bodenkd. 116 (2), 23–29.
content estimates obtained when using the CCC parameters as pre- Imhoff, S., Da Silva, A.P., Fallow, D., 2004. Susceptibility to compaction, load support
dictors have several explanations, including the similarity of the two capacity, and soil compressibility of Hapludox. Soil Sci. Soc. Am. J. 68 (1), 17–24.
curves (SWRC and CCC), correlations between the parameters in the Keller, T., Arvidsson, J., 2007. Compressive properties of some Swedish and Danish
structured agricultural soils measured in uniaxial compression tests. Eur. J. Soil Sci.
Dexter SWRC model and the parameters in the CCC models, the soil 58 (6), 1373–1381.
physical parameters, and controlling both the CCC and SWRC using the

76
H. Bayat et al. Geoderma 318 (2018) 64–77

Keller, T., Lamandé, M., Schjønning, P., Dexter, A.R., 2011. Analysis of soil compression approaches to estimate AWC. J. Irrig. Drain. Eng. 138 (2), 177–184.
curves from uniaxial confined compression tests. Geoderma 163 (1), 13–23. Richard, G., Cousin, I., Sillon, J., Bruand, A., Guérif, J., 2001. Effect of compaction on the
Koolen, A., 1974. A method for soil compactibility determination. J. Agric. Eng. Res. 19 porosity of a silty soil: influence on unsaturated hydraulic properties. Eur. J. Soil Sci.
(3), 271–278. 52 (1), 49–58.
Koolen, A., Kuipers, H., 1989. Soil deformation under compressive forces. In: Mechanics Romero, E., Gens, A., Lloret, A., 2001. Temperature effects on the hydraulic behaviour of
and Related Processes in Structured Agricultural Soils. Springer, pp. 37–52. an unsaturated clay. Geotech. Geol. Eng. 19 (3–4), 311–332.
Kuan, H., Hallett, P., Griffiths, B., Gregory, A., Watts, C., Whitmore, A., 2007. The bio- Saffih-Hdadi, K., Défossez, P., Richard, G., Cui, Y.-J., Tang, A.-M., Chaplain, V., 2009. A
logical and physical stability and resilience of a selection of Scottish soils to stresses. method for predicting soil susceptibility to the compaction of surface layers as a
Eur. J. Soil Sci. 58 (3), 811–821. function of water content and bulk density. Soil Tillage Res. 105, 96–103.
Lamorski, K., Pachepsky, Y., Sławiński, C., Walczak, R., 2008. Using support vector ma- Schoonover, J.E., Jackie, F.C., 2015. An introduction to soil concepts and the role of soils
chines to develop pedotransfer functions for water retention of soils in Poland. Soil in watershed management. J. Contemp. Water Res. Educ. 154 (1), 21–47.
Sci. Soc. Am. J. 72 (5), 1243–1247. Seki, K., 2007. SWRC fit? A nonlinear fitting program with a water retention curve for
Larson, W., Gupta, S., Useche, R., 1980. Compression of agricultural soils from eight soil soils having unimodal and bimodal pore structure. Hydrol. Earth Syst. Sci. Discuss. 4
orders. Soil Sci. Soc. Am. J. 44 (3), 450–457. (1), 407–437.
Lebert, M., Horn, R., 1991. A method to predict the mechanical strength of agricultural Sheng, D., Zhou, A.-N., 2011. Coupling hydraulic with mechanical models for unsaturated
soils. Soil Tillage Res. 19 (2), 275–286. soils. Can. Geotech. J. 48 (5), 826–840.
Lee, T.K., Ro, H.M., 2014. Estimating soil water retention function from its particle-size Sillers, W.S., Fredlund, D.G., Zakerzaheh, N., 2001. Mathematical attributes of some
distribution. Geosci. J. 18 (2), 219–230. soil–water characteristic curve models. Geotech. Geol. Eng. 19 (3–4), 243–283.
Lee, I.-M., Sung, S.-G., Cho, G.-C., 2005. Effect of stress state on the unsaturated shear Soane, B., Van Ouwerkerk, C., 1994. Soil Compaction Problems in World Agriculture, Soil
strength of a weathered granite. Can. Geotech. J. 42 (2), 624–631. Compaction in Crop Production. Elsevier, Amsterdam, pp. 1–26.
MathWorks, I, 2005. MATLAB: the language of technical computing. In: Desktop Tools Sun, D.a., Sheng, D., Sloan, S.W., 2007. Elastoplastic modelling of hydraulic and
and Development Environment, Version 7, 9. MathWorks. stress–strain behaviour of unsaturated soils. Mech. Mater. 39 (3), 212–221.
Mayr, T., Jarvis, N., 1999. Pedotransfer functions to estimate soil water retention para- Sun, W., Sun, D.a., Fang, L., Liu, S., 2014. Soil-water characteristics of Gaomiaozi ben-
meters for a modified Brooks–Corey type model. Geoderma 91 (1), 1–9. tonite by vapour equilibrium technique. J. Rock Mech. Geotech. Eng. 6 (1), 48–54.
Meskini-Vishkaee, F., Mohammadi, M.H., Vanclooster, M., 2014. Predicting the soil Terzaghi, K., Jelinek, R., 1954. Theoretische Bodenmechanik: Übers. u. bearb. nach d.
moisture retention curve, from soil particle size distribution and bulk density data fünften amerikan. Aufl. Springer.
using a packing density scaling factor. Hydrol. Earth Syst. Sci. 18, 4053–4063. Tietje, O., Tapkenhinrichs, M., 1993. Evaluation of pedo-transfer functions. Soil Sci. Soc.
Minitab, I, 2000. MINITAB statistical software. In: Minitab Release 13. Am. J. 57 (4), 1088–1095.
Ng, C.W., Pang, Y., 2000. Influence of stress state on soil-water characteristics and slope Tomasella, J., Pachepsky, Y., Crestana, S., Rawls, W., 2003. Comparison of two techni-
stability. J. Geotech. Geoenviron. 126 (2), 157–166. ques to develop pedotransfer functions for water retention. Soil Sci. Soc. Am. J. 67
Nguyen, P.M., Van Le, K., Cornelis, W.M., 2014. Using categorical soil structure in- (4), 1085–1092.
formation to improve soil water retention estimates of tropical delta soils. Soil Res. Twarakavi, N.K., Šimůnek, J., Schaap, M., 2009. Development of pedotransfer functions
52, 443–452. for estimation of soil hydraulic parameters using support vector machines. Soil Sci.
Nguyen, P.M., Haghverdi, A., de Pue, J., Botula, Y.-D., Le, K.V., Waegeman, W., Cornelis, Soc. Am. J. 73 (5), 1443–1452.
W.M., 2017. Comparison of statistical regression and data-mining techniques in es- van Genuchten, M.T., 1980. A closed-form equation for predicting the hydraulic con-
timating soil water retention of tropical delta soils. Biosyst. Eng. 153, 12–27. ductivity of unsaturated soils. Soil Sci. Soc. Am. J. 44 (5), 892–898.
Ohu, J., Raghavan, G., Prasher, S., Mehuys, G., 1987. Prediction of water retention Vereecken, H., Diels, J., Van Orshoven, J., Feyen, J., Bouma, J., 1992. Functional eva-
characteristics from soil compaction data and organic matter content. J. Agric. Eng. luation of pedotransfer functions for the estimation of soil hydraulic properties. Soil
Res. 38 (1), 27–35. Sci. Soc. Am. J. 56 (5), 1371–1378.
Or, D., Wraith, J.M., 2002. Soil water content and water potential relationships. In: Soil Walczak, R., Witkowska-Walczak, B., Sławiński, C., 2004. Pedotransfer studies in Poland.
Physics Companion. 1. pp. 49–84. Dev. Soil Sci. 30, 449–463.
Pachepsky, Y.A., Rawls, W., 1999. Accuracy and reliability of pedotransfer functions as Walczak, R., Moreno, F., Sławiński, C., Fernandez, E., Arrue, J., 2006. Modeling of soil
affected by grouping soils. Soil Sci. Soc. Am. J. 63 (6), 1748–1757. water retention curve using soil solid phase parameters. J. Hydrol. 329 (3), 527–533.
Pachepsky, Y., Rawls, W., Gimenez, D., Watt, J., 1998. Use of soil penetration resistance Wösten, J.H.M., Pachepsky, Y., Rawls, W.J., 2001. Pedotransfer functions: bridging the
and group method of data handling to improve soil water retention estimates. Soil gap between available basic soil data and missing soil hydraulic characteristics. J.
Tillage Res. 49 (1), 117–126. Hydrol. 251, 123–150.
Patil, N., Pal, D., Mandal, C., Mandal, D., 2012. Soil water retention characteristics of Zhang, X., Mavroulidou, M., Michael, J.G., 2017. A study of the water retention curve of
Vertisols and pedotransfer functions based on nearest neighbor and neural networks lime-treated London clay. Acta Geotech. 12, 23–45.

77

Você também pode gostar