Journal of Hydrology (2008) 362, 1– 18

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/jhydrol

On the relevance of using artificial neural networks

for estimating soil moisture content
Amin Elshorbagy *, K. Parasuraman

Centre for Advanced Numerical Simulation (CANSIM), Department of Civil and Geological Engineering,
University of Saskatchewan, 57 Campus Drive, Saskatoon, SK, Canada S7N 5A9

Received 6 March 2008; received in revised form 1 August 2008; accepted 11 August 2008

KEYWORDS Summary Soil moisture is a key variable that defines the land surface-atmosphere
Soil moisture content; (boundary layer) interactions, by contributing directly to the surface energy balance
Higher-order neural and water balance. This paper investigates the utility of the widely adopted data-driven
networks; model, namely artificial neural networks (ANNs), for modeling the complex soil moisture
Modeling; dynamics. Datasets from three experimental soil covers (D1, D2, and D3), with thickness
Reconstructed of 0.50 m, 0.35 m, and 1.0 m, comprising a thin layer of peat mineral mix over varying
watersheds; thickness of till, are considered in this study. Volumetric soil moisture contents at both
Conceptual models the peat and the till layers were modeled as a function of precipitation, air temperature,
net radiation, and ground temperature at different layers. Initial simulations illustrated
that, in the absence of time-lagged meteorological variables, the ground temperature
is the most influential state variable for characterizing the soil moisture, highlighting
the strong link between the soil thermal properties and the corresponding moisture status.
With the objective of extracting the maximum information from the most influential state
variables (ground temperature), a higher-order neural networks (HONNs) model was
developed to characterize the soil moisture dynamics. The HONNs resulted in relatively
higher correlation coefficient, than traditional ANNs, for some of the soil moisture simu-
lations. Time-lagged inputs were used to improve the model performance and obtain opti-
mum results. The ANN models performed better than a previously developed conceptual
model for estimating the depth-averaged soil moisture content. Results from the study
indicate that modeling of soil moisture using ANNs is challenging but achievable, and its
performance is largely influenced by the structure and formation of the soil covers, which
in turn governs the dynamics of soil moisture variability.
ª 2008 Elsevier B.V. All rights reserved.

* Corresponding author. Tel.: +1 306 9665414; fax: +1 306 9665427.

E-mail address: amin.elshorbagy@usask.ca (A. Elshorbagy).

0022-1694/$ - see front matter ª 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2008.08.012
2 A. Elshorbagy, K. Parasuraman
Introduction
Soil moisture content is a major control on several hydrolog-
ical processes. Soil moisture determines the partitioning of
available energy between latent and sensible heat (Entek-
habi et al., 1996), and the magnitude of net radiation ab-
sorbed by the surface (Eltahir, 1998). Kuczera (1983) and
Wooldridge (2003) have demonstrated that inclusion of soil
moisture data in hydrological modeling can yield substantial
reductions in the uncertainty of model parameters. Also,
accounting for soil water content on catchement scales
have proven beneficial for flow prediction (Kitanidis and
Bras, 1980; Georgakakos and Smith, 1990; Wooldridge et
al., 2003), and weather and climate studies (Georgakakos
and Bae, 1994; Georgakakos et al., 1995). As these studies
clearly highlight the role of soil moisture in hydrological
modeling, the importance of understanding the dynamics
of soil moisture cannot be over-emphasized.
Conventionally, field soil moisture estimates can be ob-
tained by the gravimetric method or by the Time Domain
Reflectometers (TDR). Although the field estimates of soil
moisture are reasonably reliable, the problems (labor inten-
sive, time consuming, expensive) associated with direct
measurement of soil moisture, make these techniques quite
impractical to characterize soil moisture patterns over an
area. Also, heterogeneity in topography, soil properties, cli-
mate patterns, vegetation, and boundary conditions instill
large amount of nonlinearity to the dynamics of soil mois-
ture movement and distribution. Over the past decade,
advancement in computer power and technology has made
measurement of spatial distribution of soil moisture feasible
using remote sensing techniques. However, soil moisture
measured using these techniques may suffer from poor tem-
poral resolution, and is representative of soil moisture con-
tent of the top few centimeters. Fully distributed,
continuous water balance models have also been used to
estimate soil moisture patterns (e.g. Wood et al., 1992;
Grayson et al., 1995). The accuracy of soil moisture esti-
mated by these models depends on the model physics and
the number and configuration of soil layers, as well as the
accuracy, and the temporal and spatial nature of the input
data. As evident from the large number of publications,
coupled modeling-remote sensing techniques (e.g. Grayson
et al., 1995; Makkeasorn et al., 2006) is the most widely
adopted procedure for characterizing soil moisture. Entek-
habi et al. (1994) showed that the ability of such models
to predict soil moisture averages over larger depths is se-
verely limited by the noise introduced by micro-topography
and roughness of vegetation cover. Since most of the meth-
ods have some limitations with regard to their ability in
characterizing the soil moisture dynamics, there is clearly
no single technique that can be adopted to estimate soil
moisture in an operational mode (Entekhabi, 1996).
In light of the problems associated with the measure-
ment and modeling of soil moisture using the above tech-
niques, this study investigates the utility of the most
widely adopted data-driven model, namely the artificial
neural networks (ANNs), for estimating soil moisture.
Although a plethora of studies have been carried out in
the past to evaluate the efficacy of ANNs in hydrological
modeling, for example, rainfall–runoff modeling (Hsu
et al. 1995; Shamseldin 1997; Dawson and Wilby, 1998;
Zhang and Govindaraju, 2000), streamflow forecasting (Kar-
unanithi et al. 1994; Thirumalaiah et al. 1998), and rainfall
forecasting (French et al. 1992; Zhang et al. 1997), the util-
ity of ANNs in modeling soil moisture has rarely been re-
ported in the literature. For modeling the rainfall–runoff
relationship using ANNs, Dawson and Wilby (1998) used
antecedent precipitation index based on earlier rainfall with
the objective of approximating the soil moisture fluctua-
tions. In modeling the runoff in a humid forest catchment
using ANNs, Gautam et al. (2000) showed that soil moisture
data obtained from 40 cm depth carry the integrated effect
of the upstream catchment area, and are important for esti-
mating stream discharge. These studies verify the impor-
tance of accounting for soil moisture in ANN modeling of
hydrological processes. Since soil moisture has been shown
to be an important variable in conceptual (e.g. Kitanidis
and Bras, 1980; Georgakakos and Smith, 1990) and data-dri-
ven (e.g. Dawson and Wilby, 1998; Gautam et al. 2000) mod-
eling of geophysical processes, the results from this study
would be of interest to the hydrological community in gaug-
ing the utility of ANNs in characterizing the most complex
and the most important hydrological variable, namely, soil
moisture.
As neural networks belong to the category of data-driven
models, proper input selection is a crucial step in any ANN
implementation. The lack of pertinent inputs or the pres-
ence of redundant inputs in neural network modeling se-
verely impairs the ability of the network to learn the
target patterns. Several methods have been proposed in
the literature to identify the effective inputs for neural net-
work modeling, and a comprehensive treatise on this issue
can be found in Bowden et al. (2005). The problem of iden-
tifying the effective inputs in neural network modeling is ill-
posed if the pertinent variables for modeling a particular
process are either not measured, or measured at different
spatial and temporal scales. Hence, neural network model-
ers are often faced with the task of extracting the maximum
information from the available (measured) independent
variables. Kim and Valdés (2003), and Anctil and Tape
(2004) used wavelet decomposed signals of the independent
variables as inputs to neural network models, with the idea
of providing the neural networks with information at differ-
ent frequencies. Bowden et al. (2006) used principal compo-
nents analysis (PCA) to reduce the dimensionality of the
input dataset and to remove the collinearity present be-
tween the independent variables. Modular neural networks
were also proposed with the aim of developing domain
dependent input–output relationships (e.g. Zhang and
Govindaraju, 2000; Parasuraman et al., 2006; Parasuraman
and Elshorbagy, 2007). While the above discussed methods
are promising, they are not straight-forward to implement.
This study investigates the ability of higher-order neural
networks (HONNs) (Gupta et al., 2003) to extract maximum
information (judged by improved prediction accuracy) from
the input data. The HONNs, in contrast to the widely
adopted first-order neural networks (FONNs), has the ability
to exploit self- and cross-correlation existing among inputs.
The nonlinear and higher-order properties, if any, in the in-
put space cannot be captured by the FONNs, as they employ
only the linear correlation between the input vector and the
On the relevance of using artificial neural networks for estimating soil moisture content
weight vector. For the same problem, the HONNs use
higher-order correlations to learn the higher-order proper-
ties in the input space. In general, the objectives of this
study can be stated as: (i) identifying the additional gains,
if any, of considering HONNs for modeling soil moisture con-
tents; (ii) evaluating the ability of neural networks in mod-
eling and characterizing the depth-averaged soil moisture
values of reconstructed watersheds (soil covers).
Study area and field data
Large scale mining in the Athabasca basin, Alberta, Canada,
involves stripping of large amount of organic and glacial
deposits and a layer of saline/sodic cretaceous shale to gain
access to the oil sands. Approximately 120,000 m3 of shale
overburden is removed daily and is placed in storage areas
that will eventually cover a combined area of about
80 km2 (Boese, 2003). Prior to mining, the shale overburden
is stable since it is over consolidated, confined, and is ex-
posed only to saline/sodic pore fluids. However, once the
shale is placed on the surface, it is exposed to fresh water
and oxygen and is susceptible to weathering, which in the
long run affects the stability of the pile (Barbour et al.,
2001). In order to overcome this problem, the piles are re-
contoured and capped with sufficient soil cover so that
the amount of precipitation percolating below the root zone
can be minimized while maintaining enough moisture for
vegetation. In this way, the overburden can be restored to
its natural state by supporting vegetation. Over the years,
several large scale soil cover (reconstructed watersheds)
experiments are being conducted to assess the performance
of different reclamation strategies by studying the basic
mechanisms that control the moisture movement within
these covers. In particular, three experimental soil covers
(D1, D2, and D3) were established in the year 1999. The
experimental covers were constructed over the saline-sodic
overburden with thickness of 0.50 m, 0.35 m, and 1.0 m,
comprising a thin layer of peat mineral mix over varying
thickness of secondary (glacial/till) soil. The layout of the
soil covers are shown in Fig. 1. Cover D1 consists of 20 cm
of peat overlying 30 cm of till; cover D2 consists of 15 cm
of peat overlying 20 cm of till, and cover D3 consists of
20 cm of peat overlying 80 cm of till (Fig. 1).
The three soil covers are contoured in such a way that
these soil covers can be considered as individual sub-water-
sheds. Each of the sub-watersheds has an area of 1 ha
(approximately 200 m long and 50 m wide), with a 5:1 slope
(5 horizontal to 1 vertical). These reconstructed watersheds,
compared to the natural watersheds, are not stable during
their initial stages, and hence evolve over time to achieve hy-
dro-sustainability. In order to track the evolution (hydrolog-
ical changes) of the watersheds, intensive instrumentations
were installed in these watersheds. Each watershed has an
individual soil station located at the middle of the slope,
which measures the volumetric soil moisture content of the
peat (SMp) and the till (SMt) layers, twice a day. Soil moisture
is measured using TDR principles with model CS615
(Boese, 2003). The TDR sensors were installed laterally
into the soil profile. Watershed D1 has eight TDR sensors
installed over a depth range of 0.05–1.00 m. Similarly, seven
TDR sensors were installed in watershed D2 over a depth
3
range of 0.05–0.80 m, and nine TDR sensors for
watershed D3 installed over a depth range of 0.05–1.70 m.
Hourly values of soil temperature of peat (STp) and till (STt)
layers are measured using thermisters buried in each of the
watersheds at the depth ranges corresponding to the TDR
sensors in each watersheds. Consequently, D1, D2, and D3
have eight, seven, and nine soil temperature sensors, respec-
tively. A weather station located in the mid-slope of D2 mea-
sures air temperature (AT), and precipitation (P). Similarly,
Bowen station located at the mid-slope of D2 measures net
radiation (NR) and energy fluxes. All the meteorological vari-
ables are measured in an hourly scale. More details on the
field instrumentation program and the data collected can
be found in Boese (2003) and Elshorbagy et al. (2007).
Average daily values of precipitation, air temperature,
soil temperature (STp and STt), net radiation (NR), and soil
moisture (SMp and SMt) are considered for simulation pur-
poses. The ground temperature and soil moisture contents
are depth-averaged. As the soil stratum is frozen during
the winter, only summer (May–September) time data are
considered. Neglecting the year 1999, when the dataset
was incomplete, soil moisture data between 2000 and
2005 are considered. As the reconstructed watersheds
evolve over time to achieve hydro-sustainability, the
freeze–thaw cycles and decomposition of highly organic
peat layer increases the porosity of the soil and conse-
quently increasing infiltration rates (Haigh, 2000). Hence,
modeling the moisture dynamics of such evolving water-
sheds would be adding to the already challenging task of
modeling soil moisture. The following section, in brief, out-
lines the ANN models adopted in this study.
Neural networks
Since a plethora of information on ANNs is available in the
literature (e.g. Haykin, 1999; ASCE Task Committee on
Application of Artificial Neural Networks in Hydrology,
2000), the description of ANNs herein is brief, and limited
to the needs of this study. A three layered feed-forward
neural networks (FFNNs) model is conventionally adopted
in modeling hydrological processes. The three layered
FFNNs consist of ‘j’ input neurons, ‘k’ hidden neurons,
and ‘l’ output neurons. These neurons are connected with
each other by means of connection weights and bias. The
hidden layer neurons usually employ sigmoidal activation
function and the output layer neurons use linear activation
function. These activation functions help in nonlinearly
transforming the inputs to the desired output. Symbolically,
the three layered FFNNs can be represented as ANN(j, k, l).
The development of a neural network model demands two
operations, namely, (i) training and (ii) testing. Training is
a process by which the connection weights between differ-
ent layers and the bias values of the neural networks are
optimized by minimizing the cost function. Once trained,
the neural network model can be tested on an independent
dataset that has not been used during the training process.
In ANN modeling of hydrological processes, the signifi-
cant input variables for characterizing a particular process
is usually determined by linear cross-correlation (Bowden
et al., 2005). Linear cross-correlation can only detect linear
dependence between two variables, and is not suited for
4 A. Elshorbagy, K. Parasuraman

Figure 1 Layout of soil covers (sub-watersheds), all axes units in meters.

capturing the nonlinear dependence between the inputs and X

n X
n
z¼ w ij xi x j ð2Þ
the output. Moreover for most hydrological processes, sig-
i¼0 j¼i
nificant linear and higher-order correlations may exist be-
tween the independent and the dependent variables. Similarly for N = 3, the cubic synaptic operation (CSO) for
Nevertheless, the FONNs, which are the most widely the neural unit can be written as
adopted neural network architecture for modeling hydrolog- n X
X n X
n
ical process, do not have the ability to capture the higher- z¼ w ijl x i x j x l ð3Þ
order correlations, as they employ only linear correlation i¼0 j¼i l¼j
between the input vector and the weight vector. In this re-
gard, the robustness of the higher-order neural networks Accordingly, a general form for the Nth order neural units
(HONNs), which can extract the higher-order correlations can be written as follows (Gupta et al., 2003):
between the components of the input patterns, is evalu-
X
n X
n
ated. Taylor and Commbes (1993) showed that the HONNs z¼aþ w i1 x i1 þ w i1i2 x i1 x i2 þ   
have good computational, storage, pattern recognition, i1 i1;i2
and learning properties due to their ability to exploit the X
n
cross and self correlations between the inputs. This study þ w i1;...;iN xi1 ; . . . ; x iN ð4Þ
adopted the HONNs based on Gupta et al. (2003). The high- i1;...;iN
er-order neural units (HONUs) are the basic building blocks
for such a higher-order neural network. The HONUs contain Typically, only second-order (N = 2) networks are usually
all the linear and nonlinear correlation terms of the input employed in practice to give a tolerable number of weights
components to the order of N. A generalized structure of (Gupta et al., 2003). Hence, this study adopted the second-
the HONU is a polynomial network that includes the order networks. Considering x1 and x2 as the input variables
weighted sums of products of selected input components for characterizing a process, a second-order network can be
with an appropriate power Fig. 2 shows the HONNs adopted constructed by using both the actual inputs (x1 and x2), and
in this study, which is adopted from Gupta et al. (2003), the derived variables ðx 21 ; x 1 x 2 ; x 22 Þ required for the second-
where x1, x 2, . . . , xn, are the inputs, y is the corresponding order polynomial, as inputs to the higher-order neural unit.
output, and u(Æ) is a monotonic activation function. Consid- In this way, the self and cross-correlations existing between
ering a first-order network (N = 1), the linear synaptic oper- the input variables can be accounted for in ANN modeling.
ation (LSO) is given by Eq. (1) Symbolically, the output from the second-order neural unit
(e.g. the case of two inputs) can be given by Eq. (5). More
X
n information on HONNs can be found in Gupta et al. (2003).
z¼ w i xi ð1Þ
i¼0 z ¼ a þ w 1 x 1 þ w 11 x 1 x 1 þ w 12 x 1 x 2 þ w 2 x 2 þ w 22 x 2 x 2 ð5Þ
For N = 2, the quadratic synaptic operation (QSO) is given by HONNs can be programmed (e.g. in MATLAB) so that the de-
(Redlapalli, 2004) rived second-order inputs are computed internally, or they
On the relevance of using artificial neural networks for estimating soil moisture content 5

w1
x1
x2 w2
1st order
.
correlation .
xn .
x1
wn

x1
a
w11 w0
x1 2

x2 x2 x1x2
w12
2nd order z y
correlation
. ∑ ϕ ( .)
. .
. xn2 .
. .
wnn
. .
. .
xn .
. .
xn . .
. .
.
. W1...1
. x1N
x1N-1x2 W1...2
Nth order .
correlation .
x1N .
Wn...n

Figure 2 Higher-order neural networks (Gupta et al., 2003).

can be computed externally and fed to the ANNs. Both ways
lead to the same results, however, if derived inputs are
computed internally, then the inputs will consist of all pos-
sible combinations. External computations allow for con-
structing partial HONNs (Input consists of some, not all,
second-order inputs). The neural network models adopted
in this study employ tan-sigmoidal activation function in
the hidden layer neurons and linear activation function in
the output layer neurons. The neural network models were
trained using the Bayesian-regularization back propagation
algorithm (Demuth and Beale, 2001). Unlike other training
algorithms, where the cost function involves minimizing
the mean sum of squares of the network errors (MSE) alone,
the cost function of the Bayesian-regularization back prop-
agation algorithm involves minimizing both the MSE and the
mean of sum of squares of the network bias and connection
weights (MSW). The cost function used in the Bayesian-reg-
ularization algorithm is presented in Eq. (6), where yi and y 0i
represent the measured and computed counterparts; a rep-
resents the regularization parameter; n, and N represents
the number of training instances and the number of network
parameters, respectively. By minimizing both the MSE and
MSW, the Bayesian-regularization algorithm improves the
generalization property of the ANN model by developing
networks with smaller weights and bias, and thus a smooth-
er response that is less likely to over fit the data (Demuth
and Beale, 2001). Conventionally, the optimal number of
hidden nodes and the network model parameters are deter-
mined by trial-and-error method. Hence, this study adopted
a similar approach to arrive at the optimal network config-
uration, by performing a systematic search on number of
hidden neurons and user adjustable parameters, with the
objective of minimizing the cost function (Eq. (6)).
!
1X n
 2 1 XN
MSE REG ¼ y i  y 0i þ ð1 þ aÞ W 2j ð6Þ
n i¼1 N j¼1
Soil moisture modeling
In order to account for the evolving nature of the water-
sheds in neural network modeling, the available dataset
(2000–2005) was apportioned in such a way that every other
data instance (alternating instances) appears in the training
dataset, while the remaining instances make up the testing
dataset. The number of instances considered for each of
training and testing were 384. The correlation statistics be-
tween the independent variables and the dependent vari-
ables for sub-watersheds D1, D2, and D3, are presented in
Table 1. Although there are no strong correlations between
each of the independent and the dependent variables (Table
1), the most correlated variables are not the same for all
the three sub-watersheds (soil covers). For sub-watershed
D1, NR has the highest correlation with SMp, and AT has
the highest correlation with SMt; for sub-watershed D2,
STt and STp have the highest correlation with SMp and SMt,
respectively, signifying the link between the moisture status
of one layer and the thermal properties of the other layer;
for sub-watershed D3, STt has the highest correlation with
both SMp and SMt, indicating the influence of the large stor-
age and the thermal properties of till layer on both SMp and
SMt of the thickest cover; D3 (Table 1).
6 A. Elshorbagy, K. Parasuraman

Table 1 Correlation statistics between the independent and the dependent variables for the three sub-watersheds during
training and testing
Sub-watershed Variable Training Testing
SMp SMt SMp SMt
D1 P 0.11 0.06 0.00 0.01
AT 0.05 0.22 0.03 0.21
NR 0.16 0.15 0.14 0.18
STp 0.02 0.16 0.02 0.15
STt 0.11 0.02 0.12 0.00
D2 P 0.09 0.01 0.04 0.00
AT 0.12 0.16 0.12 0.16
NR 0.03 0.09 0.09 0.20
STp 0.34 0.21 0.34 0.22
STt 0.35 0.05 0.36 0.06
D3 P 0.02 0.09 0.03 0.08
AT 0.18 0.16 0.18 0.15
NR 0.05 0.07 0.05 0.11
STp 0.19 0.20 0.17 0.20
STt 0.21 0.23 0.20 0.22

Several experiments (models) were considered to esti- evolution of the reconstructed watersheds to reach sustain-
mate the soil moisture content of the peat (SMp) and the till ability constantly alters the hydraulic conductivity of the
(SMt) layers, using different combinations of the indepen- soil strata, which in turn influences the phase lag. Because
dent variables, with the objective of finding the optimal in- of the soil effect as a low-pass filter, it is hypothesized that
put combination for modeling SMp and SMt. The independent the cumulative effect of the meteorological variables is
variables considered in this study are P, AT, NR, STp, and more important than the incremental daily effect. In other
STt. Since precipitation is the main source of soil moisture, words, for example, instead of considering the preceding
it would have been coherent to use the precipitation of pre- week daily precipitation as seven input variables
vious days (Pt1, Pt2, . . . , Ptn) in estimating the soil mois- (Pt, Pt1, . . . , Pt6), the summation of all values (R Pt7) can
ture of a particular day. However, the first category of be considered as one input. The favorable impact of reduc-
experiments in this study does not consider the preceding ing the number of inputs on the model parameter estima-
precipitation for estimating soil moisture. It is hypothe- tion and robustness cannot be overemphasized. The
sized, and will be proved subsequently, that the effect of performances of models developed in this study were eval-
this phase lag can be implicitly accounted for by considering uated based on the root-mean squared error (RMSE), the
soil temperature as one of the inputs to the ANN model. mean absolute relative error (MARE), and correlation coef-
Eltahir (1998) indicated that a positive feedback mechanism ficient (R). RMSE, MARE, and R are calculated using Eqs.
exists between soil moisture and rainfall as soil moisture (7)–(9), respectively, where N represents the number of in-
conditions themselves reflect past occurrence of rainfall, stances presented to the model; yi and y 0i represent mea-
and the feedback mechanism driven by the control of soil sured and computed counterparts; and y represents the
moisture on surface albedo and Bowen ratio. This implies mean of the corresponding variable.
that an increase in soil moisture decreases the surface albe-
do and Bowen ratio, which in turn increases the net radia- " #0:5
1X N
tion, and decreases the ground and surface temperature RMSE ¼ ðy  y 0i Þ2 ð7Þ
(Eltahir, 1998). Due to the link between rainfall, soil mois- N i¼1 i
ture, and soil temperature, the phase lag between the
N  
occurrence of precipitation and its manifestation as soil 1 X 0
y i  y i 
MARE ¼ ð8Þ
moisture would be reflected in the phase of soil tempera- N i¼1  y i 
ture. The phase of soil moisture and soil temperature would
be nearly identical due to the strong link between the soil P
N  
thermal properties and soil moisture. ðy i  y i Þ y 0i  y 0i
i¼1
The second category of experiments considers preceding R ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð9Þ
PN PN  2
meteorological variables for estimating soil moisture. The
ðy i  y i Þ2 y 0i  y 0i
soil acts as a low-pass filter where there is a phase lag be- i¼1 i¼1
tween the occurrence of precipitation and other variables
and their manifestation as soil moisture. This phase lag is
influenced by the atmospheric forcing on the soil strata, Results and analysis
and as a result, keeps shifting over time. This shifting may
be more pronounced in the case of reconstructed water- To understand the relationship between soil moisture con-
sheds, where in addition to the atmospheric forcing, the tents at different layers (depths), and the arrangement
On the relevance of using artificial neural networks for estimating soil moisture content
(thickness) of the corresponding soil layers, scatter plots
between SMp and SMt for the three watersheds are pre-
sented in Fig. 3. As indicated earlier, D3 is the thickest cov-
er, followed by D1 and D2. The correlation between SMp and
SMt for D1, D2, and D3, are 0.7, 0.6, and 0.2, respectively.
As expected, the correlation between SMp and SMt was the
least for the thickest cover. Nevertheless, interestingly, in-
stead of D2, which is the thinnest cover, D1 resulted in the
highest correlation between SMp and SMt. This illustrates
that the relationship between the moisture dynamics of
the peat and till layers is not just driven by the thicknesses
of the two layers, but by a more complex, and probably
poorly understood, mechanism.
Experiment category I
For each sub-watershed, various neural network models
(M1, M2, M3, and M4) were developed, as the first category
of experiments, with the objective of estimating the soil
moisture (SMp and SMt). These models differed from each
other with regards to the input combinations considered
for modeling the soil moisture patterns. While model M1
(FONNs) uses only climatic data (P, AT, and NR) as the input
vector, models M2, M3, and M4 use soil temperature
(STp, STt) or/and its variants (ST2p ; ST2t , STpSTt) as or part of
the input vector. Models M2, M3, and M4 can be used to
adjudicate the degree of relevance of using soil tempera-
ture data in soil moisture modeling. Model M2 (FONNs) con-
siders P, AT, NR, STp and STt in the input vector; Model M3
(FONNs) uses STp and STt alone as inputs; and Model M4
(HONNs) employs STp, STt, ST2p , ST2t , STpSTt as inputs. For
the three watersheds (D1, D2, and D3), the optimal number
of hidden neurons for each of the models (M1, M2, M3, and
M4) varied between four and six. For brevity, the number of
hidden neurons for individual models is not presented.
Sub-watershed D1
For sub-watershed D1, the performance of different models
in terms of RMSE, MARE, and R in estimating the SMp is pre-
sented in Table 2. The correlation between the measured
and the simulated soil moisture indicate that the models
were able to account for only a moderate proportion of
the variability in the soil moisture data. Models M2, M3,
and M4 performed significantly better than M1 during both
a 0.6 b 0.6
0.5 0.5
SMt
SMt
0.4 0.4
0.3 0.3
0.2 0.2
0.2 0.3 0.4 0 .5 0.6 0.2
SMp
Figure 3 Scatter plots between SMp and SMt for sub-watersheds: (a) D1, (b) D2, and (c) D3.
7
training and testing (Table 2), signifying the strong correla-
tion between soil thermal properties and the moisture sta-
tus. For D1, the model M3 performed better than M2
(Table 2) indicating that the STp and STt alone can explain
the variance in SMp. Comparing different models, during
testing, M4 outperformed the other models in estimating
SMp. This illustrates the ability of the HONNs in extracting
the maximum information from the most significant vari-
ables in estimating SMp for sub-watershed D1 Fig. 4 shows
the correlation plot between the measured and simulated
SMp by different models during testing, for sub-watershed
D1. Relatively, M4 has less scatter than the other models
(Fig. 4). Table 3 indicates the performance of different
models in estimating SMt for watershed D1. In general, the
performance of models in estimating SMt (Table 3) is similar
to the performance of models in estimating SMp (Table 2).
Inclusion of STp and STt as inputs to the ANN models helped
in significantly improving the performance of the models.
During training, M2 performed better than other models.
However, during testing, though M2, M3, and M4 performed
on par in terms of RMSE and MARE, M4 resulted in relatively
better correlation statistics (Table 3). Correlation plots be-
tween the measured and simulated till moisture for sub-wa-
tershed D1 are presented in Fig. 5. To a less extent, the
same phenomenon could be observed with SMp (Fig. 4). This
could be attributed to the possibility of lack of precipitation
for extended periods. During such dry periods, the evapo-
transpiration dynamics becomes the driving variable that af-
fects the available moisture. Most probably, these will be
the most challenging periods for the ANNs models to learn
the soil moisture patterns.
Sub-watershed D2
For sub-watershed D2, Table 4 shows the performance of
different models in estimating SMp; and Table 5 shows the
performance of models in estimating SMt. Similar to sub-wa-
tershed D1, the performance of models in estimating SMp
and SMt improved significantly when STp and STt were used
as inputs to the ANN models. However, in estimating SMp for
sub-watershed D2, contrary to sub-watershed D1, the per-
formance of models M3 and M4 were inferior to the perfor-
mance of model M2 (Table 4). This could be attributed to
the thinner depth of the peat layer in D2, and also the thin-
ner depth of the entire D2 soil cover. Hence, the peat layer
c 0.6
0.5
SMt
0.4
0.3
0.2
0.3 0.4 0 .5 0.6 0.2 0.3 0.4 0 .5 0.6
SMp SMp
8 A. Elshorbagy, K. Parasuraman

Table 2 Performance measure of different models in estimating peat moisture content (cm/cm) of sub-watershed D1
Model Training Testing
RMSE MARE R RMSE MARE R
M1 0.051 0.098 0.18 0.051 0.099 0.128
M2 0.040 0.073 0.644 0.046 0.083 0.507
M3 0.043 0.081 0.547 0.043 0.081 0.552
M4 0.042 0.076 0.597 0.042 0.077 0.581
M1, M2, and M3 are FONN models, whereas M4 is an HONN model.

a 0.6 b 0.60

0.55

0.5 0.50
Simulated

Simulated
0.45

0.4 0.40

0.35

0.3 0.30

0.25
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Measured Measured

c 0.6 d 0.60

0.55

0.5 0.50
Simulated
Simulated

0.45

0.4 0.40

0.35

0.3 0.30

0.25
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Measured Measured

Figure 4 Comparison of measured vs. simulated peat moisture of cover D1 by (a) M1; (b) M2; (c) M3; and (d) M4 – testing.

Table 3 Performance measure of different models in estimating till moisture content (cm/cm) of sub-watershed D1
Model Training Testing
RMSE MARE R RMSE MARE R
M1 0.018 0.050 0.236 0.018 0.050 0.209
M2 0.015 0.0430 0.624 0.016 0.043 0.532
M3 0.016 0.043 0.552 0.016 0.043 0.539
M4 0.016 0.043 0.573 0.016 0.043 0.556

of D2 is more dynamic and responsive to the meteorological 0.061, 0.138, and 0.575, respectively, M2 outperformed the
variables. During testing, with RMSE, MARE, and R values of other models. The better performance of the M2 reiterates
On the relevance of using artificial neural networks for estimating soil moisture content 9

a 0.32 b 0.32

0.30 0.30

Simulated
Simulated
0.28 0.28

0.26 0.26

0.24 0.24

0.22 0.22
0.22 0.24 0.26 0.28 0.30 0.32 0.22 0.24 0.26 0.28 0.30 0.32
Measured Measured

c 0.32
d 0.32

0.30 0.30
Simulated

Simulated
0.28 0.28

0.26 0.26

0.24 0.24

0.22 0.22
0.22 0.24 0.26 0.28 0.30 0.32 0.22 0.24 0.26 0.28 0.30 0.32
Measured Measured

Figure 5 Comparison of measured vs. simulated till moisture of cover D1 by (a) M1; (b) M2; (c) M3; and (d) M4 – testing.

Table 4 Performance measure of different models in estimating peat moisture content (cm/cm) of sub-watershed D2
Model Training Testing
RMSE MARE R RMSE MARE R
M1 0.072 0.173 0.265 0.1 0.171 0.297
M2 0.062 0.139 0.566 0.061 0.138 0.575
M3 0.064 0.146 0.519 0.063 0.146 0.525
M4 0.063 0.144 0.529 0.063 0.144 0.537

Table 5 Performance measure of different models in estimating till moisture content (cm/cm) of sub-watershed D2
Model Training Testing
RMSE MARE R RMSE MARE R
M1 0.023 0.067 0.167 0.023 0.067 0.179
M2 0.014 0.039 0.807 0.015 0.041 0.774
M3 0.014 0.0140 0.792 0.014 0.041 0.783
M4 0.014 0.039 0.805 0.014 0.040 0.789

the strong influence of atmospheric forcing on the peat
moisture dynamics of sub-watershed D2. This underscores
the need for considering P, AT, NR, STp, and STt for charac-
terizing SMp of watershed D2 Fig. 6 shows the correlation
plot between the measured and simulated values of SMp dur-
ing testing, for watershed D2. From the large scatter in the
low-ranges of SMp (Fig. 6), it can be concluded that most of
the models were unable to learn the moisture dynamics at
low-ranges. For watershed D2, the models were able to esti-
mate SMt more accurately than SMp (Tables 4 and 5). This
may be attributed to the reason that the upper peat layer
acts like a buffer zone that trims the effect of the meteoro-
10
a 0.6
0.5
Simulated
0.4
0.3
0.2
0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Measured
c 0.6
0.5
Simulated
0.4
0.3
0.2
0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Measured
Figure 6 Comparison of measured vs. simulated peat moisture of cover D2 by (a) M1; (b) M2; (c) M3; and (d) M4 – testing.
logical variables on the till layer, and thereby rendering SMt
less vulnerable to atmospheric forcing. During testing,
compared to other models, the HONN model (M4) resulted
in the highest R statistics. Nevertheless, models M2, M3,
and M4 performed on par in terms of RMSE and MARE (Table
5) Fig. 7 shows the correlation plots between the measured
and simulated SMt by different models. Disregarding M1, all
the other models were able to estimate SMt with
reasonable accuracy over the entire modeling range
(Fig. 7).
Sub-watershed D3
The performance of different models in estimating SMp and
SMt for sub-watershed D3 is presented in Tables 6 and 7,
respectively. For sub-watershed D3, similar to sub-water-
sheds D1 and D2, inclusion of STp and STt as inputs to the
ANN models improved the performance of the models in
estimating soil moisture (SMp and SMt). However, the
improvement in the performance of the models for sub-wa-
tershed D3 is not as significant as compared to sub-water-
sheds D1 and D2. This may be attributed to the weak
correlation (R = 0.2) existing between SMp and SMt. Due to
the association between soil thermal properties and mois-
ture status, the weak correlation between SMp and SMt im-
plies that the information content between STp and STt
would also be relatively weak. Hence, inclusion of STp and
STt as inputs to the ANN models was not able to appreciably
improve the performance of the models in estimating the
A. Elshorbagy, K. Parasuraman
b 0.6
0.5
Simulated
0.4
0.3
0.2
0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Measured
d 0.6
0.5
Simulated
0.4
0.3
0.2
0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Measured
soil moisture of the thick cover D3. Due to the relative
incoherency between STp and STt, the effect is more pro-
nounced in M4, where higher-order derivatives of the above
variables were adopted for estimating SMp (Table 6). The
correlation plots between measured and simulated SMp for
sub-watershed D3, are presented in Fig. 8. From the plots,
it is evident that none of the models was able to learn the
dynamics of SMp. In modeling SMt for sub-watershed D3, it
is observed that inclusion of STp and STt as inputs to the
ANN models was able to improve the performance of the
models in terms of R statistics. However, the performances
of the models are comparable in terms of RMSE and MARE
statistics (Table 7) Fig. 9 shows the correlation between
the measured and the simulated SMt by different models.
As can be seen from the plots, though M3 and M4 resulted
in less scatter in a particular modeling range (Fig. 9), the
overall performance of the models are not convincing. In
general, it can be concluded that all ANN models, within
this first category of experiments, encountered difficulties
in learning the moisture patterns of sub-watershed D3.
Since D3 is the thickest cover among the three covers con-
sidered in this study, it implies that sub-watershed D3 re-
sults in the highest phase-lag between the occurrence of
precipitation and its manifestation as soil moisture. In other
words, sub-watershed D3 has more storage effects than the
other two sub-watersheds, and thereby is prone to more
topographic forcing. This makes modeling the soil moisture
dynamics of sub-watershed D3, more intricate than model-
On the relevance of using artificial neural networks for estimating soil moisture content 11

a 0.34
b 0.34

0.32 0.32

0.30 0.30
Simulated

Simulated
0.28 0.28

0.26 0.26

0.24 0.24

0.22 0.22

0.20 0.20
0.20 0.22 0.24 0.26 0.28 0.30 0.32 0.34 0.20 0.22 0.24 0.26 0.28 0.30 0.32 0.34
Measured Measured

c 0.34
d 0.34

0.32 0.32

0.30 0.30
Simulated
Simulated

0.28 0.28

0.26 0.26

0.24 0.24

0.22 0.22

0.20 0.20
0.20 0.22 0.24 0.26 0.28 0.30 0.32 0.34 0.20 0.22 0.24 0.26 0.28 0.30 0.32 0.34
Measured Measured

Figure 7 Comparison of measured vs. simulated till moisture of cover D2 by (a) M1; (b) M2; (c) M3; and (d) M4 – testing.

Table 6 Performance measure of different models in estimating peat moisture content (cm/cm) of sub-watershed D3
Model Training Testing
RMSE MARE R RMSE MARE R
M1 0.044 0.092 0.275 0.044 0.093 0.279
M2 0.044 0.091 0.293 0.044 0.091 0.296
M3 0.043 0.087 0.328 0.044 0.089 0.326
M4 0.044 0.092 0.215 0.045 0.093 0.211

Table 7 Performance measure of different models in estimating till moisture content (cm/cm) of sub-watershed D3
Model Training Testing
RMSE MARE R RMSE MARE R
M1 0.013 0.024 0.177 0.013 0.024 0.165
M2 0.013 0.024 0.237 0.013 0.025 0.228
M3 0.012 0.023 0.400 0.012 0.023 0.383
M4 0.012 0.023 0.405 0.012 0.023 0.386

ing the soil moisture dynamics of sub-watersheds D1 and D2. averages over larger depths is severely limited by the noise
This finding is in line with Entekhabi et al. (1994), where it is introduced by the micro-topography and roughness of vege-
shown that the ability of models to predict soil moisture tation cover.
12 A. Elshorbagy, K. Parasuraman

a 0.50 b 0.50

0.45 0.45

Simulated

Simulated
0.40 0.40

0.35 0.35

0.30 0.30

0.25 0.25
0.25 0.30 0.35 0.40 0.45 0.50 0.25 0.30 0.35 0.40 0.45 0.50
Measured Measured

c 0.50
d 0.50

0.45 0.45
Simulated

Simulated
0.40 0.40

0.35 0.35

0.30 0.30

0.25 0.25
0.25 0.30 0.35 0.40 0.45 0.50 0.25 0.30 0.35 0.40 0.45 0.50
Measured Measured

Figure 8 Comparison of measured vs. simulated peat moisture of cover D3 by (a) M1; (b) M2; (c) M3; and (d) M4 – testing.

Experiment category II For each sub-watershed, additional experiments were

In order to prove the hypothesis that the phase lag between
the occurrence of precipitation, and its manifestation as
soil moisture can be implicitly accounted by considering
the soil temperature as one of the inputs for characterizing
soil moisture, model M5 was constructed. Model M5 uses P
t1, ATt1, AT, NR, STp, and STt, as inputs to estimate SMp
and SMt. Pt1 and ATt1 indicates the preceding day precip-
itation and air temperature, respectively. For brevity, only
the testing statistics of model M5 is presented for the three
sub-watersheds. For estimating the SMp of sub-watershed
D1, M5 resulted in an RMSE of 0.048; MARE of 0.09; and R
of 0.39. The corresponding values for estimating SMt are
0.016, 0.043, and 0.56. For estimating the SMp of sub-wa-
tershed D2, M5 resulted in an RMSE of 0.062; MARE of
0.14; and R of 0.56. The corresponding values for estimating
SMt are 0.015, 0.042, and 0.75. Similarly, for estimating the
SMp of sub-watershed D3, M5 resulted in an RMSE of 0.042;
MARE of 0.087; and R of 0.40. The corresponding values
for estimating SMt are 0.013, 0.024, and 0.23. Except for
simulating the SMp of sub-watershed D3, the performance
of M5 is either on par or poorer than the models, which do
not consider the preceding climatic information as part of
the input vector. These results clearly highlight the impor-
tance of considering the soil temperature as an input for
estimating the soil moisture patterns, except for signifi-
cantly large depths (e.g. D3).
conducted to improve the accuracy of predicting the SMp
and SMt using the ANN models by incorporating more infor-
mation about two processes that affect soil moisture;
namely, the cumulative precipitation and evapotranspira-
tion. The experiment M2, from the previous category, was
used as a base experiment to build on for improving the pre-
diction accuracy. In addition to the current day’s precipita-
tion (Pt), the daily precipitation of the previous six days can
be summed (RP6) and used as an input. Similarly, weekly
evapotranspiration may be represented using the cumula-
tive air temperature (RAT6). Two weeks of precipitation
and evapotranspiration can be represented by RP13 and
RAT13, respectively, in addition to the current Pt and ATt.
There is literally an infinite possible combination of inputs,
and experiments conducted in this study cannot be exhaus-
tive. However, Tables 8 and 9 present the results of four dif-
ferent experiments (M6, M7, M8, and M9) as well as the best
experiment from the previous section, which is referred to
as the reference ANNs. Model M6 uses the same inputs as
M2 (P, AT, NR, STp, and STt) plus RP6; M7 is the same as
M2 plus RP13; M8 is M2 plus RP6 plus RAT6; and M9 is M2 plus
RP13 plus RAT6.
For brevity, only testing results for the three sub-water-
sheds are presented in Tables 8 and 9 for predicting the SMp
and SMt, respectively. Significant improvement in the pre-
diction accuracy of the soil moisture of the upper peat layer
was achieved with all D1, D2, and D3 (Table 8). The R
On the relevance of using artificial neural networks for estimating soil moisture content 13

a 0.38
b 0.38

Simulated 0.36 0.36

Simulated
0.34 0.34

0.32 0.32

0.30 0.30
0.30 0.31 0.32 0.33 0.34 0.35 0.36 0.37 0.38 0.30 0.31 0.32 0.33 0.34 0.35 0.36 0.37 0.38
Measured Measured

c 0.38
d 0.38

0.36 0.36

Simulated
Simulated

0.34 0.34

0.32 0.32

0.30 0.30
0.30 0.31 0.32 0.33 0.34 0.35 0.36 0.37 0.38 0.30 0.31 0.32 0.33 0.34 0.35 0.36 0.37 0.38
Measured Measured

Figure 9 Comparison of measured vs. simulated till moisture of cover D3 by (a) M1; (b) M2; (c) M3; and (d) M4 – testing.

Table 8 Performance measure of different models in estimating peat moisture content (cm/cm) of all sub-watersheds – testing
D1 (50 cm cover) D2 (35 cm cover) D3 (100 cm cover)
RMSE MARE R RMSE MARE R RMSE MARE R
Ref. ANNs 0.04 0.08 0.58 0.06 0.14 0.58 0.042 0.09 0.40
M6 0.05 0.09 0.42 0.06 0.13 0.64 0.04 0.08 0.54
M7 0.05 0.08 0.49 0.06 0.12 0.66 0.03 0.065 0.73
M8 0.03 0.07 0.69 0.05 0.13 0.68 0.04 0.075 0.62
M9 0.03 0.07 0.65 0.05 0.11 0.72 0.03 0.064 0.74
SDW model 0.06 0.13 0.60 0.08 0.14 0.42 0.04 0.08 0.68
M6, M7, M8, and M9 are all FONN models.
Ref. ANNs represent the best model from Experiment category I.

statistic is the best indication in this case to reflect the
improvement compared to the reference ANNs. For the
thickest cover, D3, it is interesting to note that when RP6
was included (M6), a good improvement was achieved. How-
ever, even more significant improvement was achieved by
replacing RP6 with RP13 (M7) compared to adding RAT6
(M8). Similar conclusions can be drawn regarding the till
layer of D3 (Table 9). This reflects the fact that the thick
cover has large storage potential, which is captured by
the ANN models through the precipitation history. The cover
can store enough moisture during the wet days so that
evapotranspiration (represented by air temperature) does
not become the most important control of the soil moisture.
In case of the thinnest cover D2, one can observe that
including RAT6 helps achieve reasonable improvement in
the prediction accuracy; M9 is better than M7 in both Table
14
Table 9 Performance measure of different models in estimating till moisture content (cm/cm) of all sub-watersheds – testing
D1 (50 cm cover) D2 (35 cm cover)
RMSE MARE R RMSE
Ref. ANNs 0.016 0.04 0.56 0.014
M6 0.016 0.043 0.56 0.014
M7 0.016 0.042 0.56 0.014
M8 0.013 0.037 0.70 0.011
M9 0.012 0.036 0.70 0.011
SDW model 0.021 0.08 0.67 0.10
8 and Table 9. Unlike D3, moving from M6 to M8 is better
than moving from M6 to M7. Even though precipitation is
still an important factor, evapotranspiration is a major con-
trol of moisture on D2. The cover is thin and lacks the stor-
age ability; therefore, it cannot store moisture from far in
the past. Evapotranspiration could really cause D2 to dry
out. In case of the intermediate cover D1, both the peat
(Table 8) and the till layers (Table 9) behaves in a way that
show a balancing effect of both the storage and the precip-
itation. Models M8 and M9 are significantly better than M6
and M7; reflecting the need of presenting balanced memory
of both the storage and the precipitation.
Noting that all experiments of category II (M5–M9) are
based on FONN models, an additional HONN model (M10)
was tested in this category. M10 was designed to be fully
second-order ANN model based on the best performance re-
ported in Table 8 and Table 9; i.e., M10 is a second-order
ANNs of M8 for the peat layer of D1 cover, and a second-or-
der ANNs of M9 for the peat layer of D2 and D3, as well as for
the till layer of all three covers. Other possibilities and sce-
narios of fully and partial HONNs are endless, however,
experiment M10 is considered sufficient for the purpose of
this study. The only noticeable improvement achieved by
the use of HONNs (M10) over the FONNs is the increase in
the R statistic in the case of the peat layer of the thick cover
D3 (slight increase from 0.74 to 0.77) as well as the till layer
(considerable increase from 0.59 to 0.70). It can be con-
cluded that second-order ANNs played a positive role in
improving the prediction of soil moisture in thicker cover.
It might be possible that even higher-order ANNs could im-
prove the results further. This should be subject to future
investigation.
Comparison between the ANNs and the conceptual
SDW model
Even though the second category of experiments signifi-
cantly improved the ANN model performance with regard
to the estimation of the soil moisture content, it is desirable
to measure such performance against a validated concep-
tual model. Elshorbagy et al. (2005, 2007) developed a
site-specific system dynamics watershed (SDW) model that
simulates the various hydrological processes occurring in
the reconstructed watersheds; D1, D2, and D3. The model
is a lumped mechanistic model that was conceptualized as
a control volume that simulates the water balance compo-
nents among the two soil layers (peat and till) as well as
the underlying shale layer including evapotranspiration
A. Elshorbagy, K. Parasuraman
D3 (100 cm cover)
MARE R RMSE MARE R
0.04 0.79 0.012 0.02 0.39
0.04 0.78 0.011 0.023 0.45
0.04 0.80 0.01 0.022 0.56
0.036 0.84 0.011 0.023 0.52
0.036 0.84 0.01 0.022 0.59
0.29 0.08 0.02 0.05 0.50
and runoff on a daily basis. A combination of physically
based formulations (e.g. Green-Ampt for infiltration and soil
moisture redistribution) and fitted-parameter formulations
(e.g. actual evapotranspiration based on simulated soil
moisture index) was employed in formulating the SDW mod-
el. More details on the model can be found in Elshorbagy
et al. (2007). The model was extensively evaluated for the
purposes of design and hydrologic performance assessment
of reconstructed watersheds (Elshorbagy and Barbour,
2007) as well as for management decision making (Elshor-
bagy, 2006).
Even though it is infeasible to compare the input struc-
ture of mechanistic and data-driven models, the input vari-
ables and data needed for the SDW model are provided here
to facilitate the evaluation of the performances of the ANN
models relative to the SDW model. The current version of
the SDW model requires daily precipitation (P), average dai-
ly air temperature (AT), daily net radiation (NR), relative
humidity (RH), average daily temperatures of both peat
and till layers (STp and STt), and soil physical properties that
allow for the estimation of the soil water characteristic
curve (SWCC). However, it should be noted that by defini-
tion of mechanistic models, the SDW model has the ability
to keep ‘‘indefinite’’ memory of past values (time-lagged)
of input and simulated parameters through is tank (stock)
component. At the same time, the SDW model can provide
soil moisture content, actual evapotranspiration, and runoff
as outputs. The SDW model has nine different calibration
parameters related to the processes of infiltration, evapo-
transpiration, snowmelt, and runoff (Elshorbagy et al.,
2007).
The performance of the SDW model in estimating the soil
moisture in the peat layer of the three sub-watersheds is
presented in Table 8. Clearly, the best ANN models per-
formed reasonably better, and sometimes much better as
in the case of D2, than the SDW model. The R statistic
should be noted in all covers. Similar conclusions can be
drawn with regard to the till layer (Table 9). It is interesting
to note that the poor performance of the SDW model ap-
plied to the thinnest cover D2 was significantly improved
and overcome by the ANN models. Elshorbagy et al. (2007)
reported that the presence of macro pores and the steep
slope of the covers make the influence of the surface runoff
and the interflow dynamics more prominent with the thin
cover (D2). Such dynamics were difficult to capture within
the conceptual SDW model, whereas the data-driven ANNs
were less affected by such dynamics and phenomena. For vi-
sual assessment of the second category of ANN experiments,
On the relevance of using artificial neural networks for estimating soil moisture content 15

the performance of the best ANN models during testing is

presented in Figs. 10–12.
a 0.6

Discussion
0.5

Even though the performance of the ANN models, for esti-

Simulated
mating the peat and till layers’ moisture content, was bet-
ter than the SDW model in all cases, one can observe that 0.4
the ANN model superiority is less in case of D1 (50 cm) than
in case of D2. The least improvement of the ANN models
over the SDW model was achieved with the thickest cover
0.3
D3 (100 cm) as shown in Tables 8 and 9. Clearly, the storage
effect, which is more prominent in thicker soil layers, was
the most difficult phenomenon to be captured by the ANNs.
Considering meteorological and hydrological parameters 0.2
representing a longer memory from the past could help even 0.2 0.3 0.4 0.5 0.6

better. However, in this study, experimenting was stopped Measured

once reasonable results were obtained. One should note
0.36
the possible effect of HONNs (Table 10) on the prediction b
accuracy in case of thicker soil layers. This work can be ex-
tended by conducting comprehensive input analysis to iden-
tify the optimum input for each sub-watershed. A technique 0.32

Simulated
a 0.60
0.28

0.55

0.50 0.24
S i mulated

0.45

0.20
0.40 0.20 0.24 0.28 0.32 0.36
Measured
0.35
Figure 11 Comparison of measured vs. simulated moisture of
0.30 cover D2 by M9 (a) peat layer; and (b) till layer – testing.

0.25
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 such as wavelet analysis could play a major role in this task
as well as in evaluating the localized performance of various
Measured
models.
0.32 The classical root mean squared error (RMSE) and the
b mean absolute relative error (MARE) were used to test the
overall performance of the models. However, in cases like
0.30 predicting the soil moisture content (cm/cm or percent),
the average error value could be deceivingly low (i.e., good
performance) but the predicted values do not match well
Simulated

0.28
with the observed values in terms of pattern and trend. In
this case, the correlation coefficient (R) is strongly recom-
mended as a model performance statistic. This study shows
0.26
that the R statistic was the most important error measure,
among the three adopted, in revealing the improvement in
0.24 soil moisture prediction. The ANN models and experiments
conducted in this study will help the authors develop a gen-
eric ANN-based model that can be used by the oil sands min-
0.22 ing industry to obtain estimates of the soil moisture
0.22 0.24 0.26 0.28 0.30 0.32 contents of various reclamation soil covers using available
Measured meteorological and hydrological parameters.
As seen from the above discussions, modeling soil mois-
Figure 10 Comparison of measured vs. simulated moisture of ture using neural networks is challenging, and their perfor-
cover D1 by M8 (a) peat layer; and (b) till layer – testing. mance is largely influenced by the structure and formation
16 A. Elshorbagy, K. Parasuraman

that D3 was always providing significantly more moisture

a 0.50
for evapotranspiration than D2 under all climatic condi-
tions; indicating its ability to ‘‘store’’ water and release it
0.45 for vegetation at the time of need. D2 was more responsive
to climatic conditions and running short of water at times of
high evapotranspiration. This argument shows that the ANN
Measured

0.40 models, developed in this study, are not only good from a
predictive point of view but also conceptually meaningful.

0.35
Summary and conclusions

0.30 As soil moisture is one of the important factors, which influ-

ence the boundary-layer climate, significant efforts have
been made in the past to improve the accuracy of soil mois-
0.25 ture estimation using statistical and remote-sensing tech-
0.25 0.30 0.35 0.40 0.45 0.50 niques. The scarcity of field-scale models to estimate soil
Measured moisture, and the success of artificial neural networks
(ANNs) in modeling different hydrological processes warrant
0.38
b their application to modeling the most complex and the
most important hydrological variable, namely, soil mois-
ture, at field scales. In this study, an attempt has been
0.36 made to estimate soil moisture from readily available cli-
matic data. Datasets from three reconstructed sub-water-
sheds (D1, D2, and D3), located in northern Alberta,
Simulated

Canada, with thickness of 0.50 m, 0.35 m, and 1.0 m, com-

0.34
0.32
0.30
0.30 0.32 0.34
Measured
Figure 12 Comparison of measured vs. simulated moisture of
cover D3 by M9 (a) peat layer; and (b) till layer – testing.
of the soil strata. As expected and partially proven in this
study, the poorest performance of ANN models might be
encountered during modeling the soil moisture phenomenon
that is influenced largely by the storage effect. The results
presented in this study supports those presented by Elshor-
bagy and Barbour (2007), in which the SDW model was used
to simulate the three sub-watersheds (D1, D2, and D3) over
a period of 60 years. Elshorbagy and Barbour (2007) show
prising a thin layer of peat mineral over varying thickness of
glacial till, was considered in this study. Depth-averaged
soil moisture contents at both the upper peat (SMp) and
the lower till (SMt) layers were modeled as function of pre-
cipitation (P), air temperature (AT), net radiation (NR), and
soil temperature of peat (STp) and the till (STt) layers. For
the three sub-watersheds, investigating the relationship be-
0.36 0.38 tween the soil moisture dynamics of the peat and till layers,
it is found that the moisture dynamics is not just driven by
the thickness of the soil layers, but by a more complex, and
probably poorly understood, mechanism. ANN models with
different input configurations were developed to estimate
the SMp and the SMt. In the absence of time-lagged meteo-
rological variables, which reflect the history of the bound-
ary conditions, STp and STt were found to be the most
important variables for modeling soil moisture. A higher-or-
der neural network (HONN) model was developed using
these variables, with the objective of extracting the maxi-
mum information from the input vector. The performances
of the models were evaluated based on the root mean
squared error (RMSE), the mean absolute relative error
(MARE) and the correlation coefficient (R). Although, in

Table 10 Performance measure of HONNs (experiment category II) models in estimating peat and till moisture content (cm/cm)
of all sub-watersheds – testing
D1 (50 cm cover) D2 (35 cm cover) D3 (100 cm cover)
RMSE MARE R RMSE MARE R RMSE MARE R
Peat layer
M8/M9 0.03 0.07 0.69 0.05 0.11 0.72 0.03 0.064 0.74
M10 0.03 0.06 0.70 0.05 0.10 0.74 0.03 0.06 0.77
Till layer
M8/M9 0.012 0.036 0.70 0.011 0.036 0.84 0.01 0.022 0.59
M10 0.01 0.037 0.70 0.01 0.036 0.82 0.01 0.02 0.70
On the relevance of using artificial neural networks for estimating soil moisture content
some cases, the correlations between the measured and the
estimated soil moisture for both the peat and the till layers
are not very strong, the results are encouraging considering
the large variance that exists between the independent and
the dependent variables. The better performance of the
ANN models considering soil temperature as one of the in-
puts reiterates the strong link between the soil thermal
properties and the soil moisture status. This finding would
be of great importance to orient the future monitoring pro-
gram of the reconstructed watersheds, where it is recom-
mended to increase the intensity of soil temperature
measurement network system, and thereby possibly im-
prove the accuracy of soil moisture estimates.
A second category of experiments were time-lagged in-
puts were considered revealed the need to consider variable
time window to obtain optimum inputs and results for each
sub-watershed. The optimum inputs were found to match
with the conceptual understanding of the significance/insig-
nificance of the storage effect, which is based on the soil
layering and thickness. For the three sub-watersheds, it is
hard to generalize the best combination of inputs for char-
acterizing soil moisture. This illustrates that the atmo-
spheric forcing and the soil moisture response to such
forcing is influenced by the type, depth, and structure of
the soil strata. In general, the results from this study indi-
cate that modeling the soil moisture using ANNs is challeng-
ing but achievable, and the model performance is largely
influenced by the structure of the soil strata. The results re-
ported in this study are unique for three reasons: (i) they
established a link between readily available climatic vari-
ables and soil-moisture, using ANNs; (ii) they evaluated
the potential of adopting the ANNs for estimating soil mois-
ture of reconstructed watersheds, where the dynamics of
atmospheric forcing and the soil moisture response to such
forcing change based on the soil cover layering and thick-
ness; and (iii) they show that ANNs can perform as good
as, and possibly better, than conceptual watershed model
in estimating depth-averaged soil moisture content.
Acknowledgements
The authors gratefully acknowledge the financial support of
the National Science and Engineering Research Council
(NSERC) of Canada, the CEMA consortium, and the Depart-
ment of Civil and Geological Engineering, University of
Saskatchewan.
References
Anctil, F., Tape, D.G., 2004. An exploration of artificial neural
network rainfall–runoff forecasting combined with wavelet
decomposition. J. Environ. Eng. Sci., S121–S128. doi:10.1139/
S03-07.
ASCE Task Committee on Application of Artificial Neural Networks in
Hydrology, 2000. Artificial neural networks in hydrology. I:
Preliminary concepts. J. Hydrol. Eng. 5 (2), 115–123.
Barbour, S.L., Boese, C., Stolte, B., 2001. Water balance for
reclamation covers on oilsands mining overburden piles. In:
Proceedings of the 54th Canadian Geotechnical Conference,
Calgary, Alta., pp. 16–19 September 2001. Canadian Geotech-
nical Society, Alliston, Ont. pp. 313–319.
17
Boese, K., 2003. The Design and Installation of a Field Instrumen-
tation Program for the Evaluation of Soil-atmosphere Water
Fluxes in a Vegetated Cover Over Saline/Sodic Shale Overbur-
den, M.Sc. Thesis, University of Saskatchewan, Saskatoon,
Sask.
Bowden, G.J., Dandy, G.C., Maier, H.R., 2005. Input determina-
tion for neural network models in water resources applica-
tions. Part 1 – background and methodology. J. Hydrol. 301,
75–92.
Bowden, G.J., Dandy, G.C., Maier, H.R., 2006. An evaluation of
methods for the selection of inputs for an artificial network
based river model. In: Recknagel, F. (Ed.), Ecological Informat-
ics. Springer, Berlin, Heidelberg, pp. 275–292.
Dawson, C.W., Wilby, R., 1998. An artificial neural network
approach to rainfall–runoff modeling. Hydrol. Sci. J. 43, 47–
66.
Demuth, H., Beale, M., 2001. Neural Network Toolbox Learning For
Use with MATLAB. The Math Works Inc, Natick, Mass.
Elshorbagy, A., 2006. Multicriterion decision analysis approach to
assess the utility of watershed modeling for management
decisions. Water Resour. Res. 42, W09407. doi:10.1029/
2005WR00426.
Elshorbagy, A., Barbour, L., 2007. Probabilistic approach for design
and hydrologic performance assessment of reconstructed water-
sheds. J. Geotech. Geoenv. Eng. ASCE 133 (9), 1110–1118.
Elshorbagy, A., Jutla, A., Kells, J., 2007. Simulation of the
hydrological processes on reconstructed watersheds using sys-
tem dynamics. Hydrol. Sci. J. 52, 538–562.
Eltahir, E.A.B., 1998. A soil moisture-rainfall feedback mechanism,
1: Theory and observations. Water Resour. Res. 34 (4), 765–
776.
Entekhabi, D., Nakamura, H., Njoku, E.G., 1994. Solving the inverse
problem for soil moisture and temperature profiles by sequential
assimilation of multifrequency remotely sensed observations.
IEEE Trans. Geosci. Remote Sensing 32 (2), 438–448.
Entekhabi, D., Rodriguez-Iturbe, I., Castelli, F., 1996. Mutual
interaction of soil moisture state and atmospheric processes.
J. Hydrol. 184, 3–17.
French, M.N., Krajewski, W.F., Cuykendall, R.R., 1992. Rainfall
forecasting in space and time using a neural network. J. Hydrol.
137, 1–31.
Gautam, M.R., Watanabe, K., Saegusa, H., 2000. Runoff analysis in
humid forest catchment with artificial neural network. J.
Hydrol. 235, 117–136.
Georgakakos, K.P., Bae, D.H., 1994. Climatic variability of soil
water in the American midwest: 1, Spatio-temporal analysis. J.
Hydrol. 162, 379–390.
Georgakakos, K.P., Smith, G.F., 1990. On improved operational
hydrologic forecasting: results from a WMO real-time forecasting
experiment. J. Hydrol. 114, 17–45.
Georgakakos, K.P., Bae, D.H., Cayan, D.R., 1995. Hydroclimatology
of continental watersheds, temporal analyses. Water Resour.
Res. 31 (3), 655–675.
Grayson, R.B., Bloschl, G., Moore, I.D., 1995. Distributed param-
eter hydrologic modelling using vector elevation data: Thales
and TAPES-C. In: Singh, V.P. (Ed.), Computer Models of
Watershed Hydrology. Water Resources Pub. Highlands Ranch,
Colorado, pp. 669–695.
Gupta, M.M., Jin, L., Homma, N., 2003. Static and Dynamic Neural
Networks. John Wiley and Sons, Inc., Hoboken, New Jersey.
Haigh, M.J., 2000. The Aims of Land Reclamation, Land Recon-
struction and Management, vol. 1. A.A. Balkema Publishers,
Rotterdam, The Netherlands, pp. 1–20.
Haykin, S., 1999. Neural Networks: A Comprehensive Foundation,
second ed. MacMillan, New York.
Hsu, K.L., Gupta, H.V., Sorooshian, S., 1995. Artificial neural
network modeling of the rainfall–runoff process. Water Resour.
Res. 31 (10), 2517–2530.
18
Karunanithi, N., Grenney, W.J., Whitley, D., Bovee, K., 1994.
Neural networks for river flow prediction. J. Comput. Civil Eng. 8
(2), 201–220.
Kim, T.W., Valdés, J.B., 2003. Nonlinear model for drought
forecasting based on a conjunction of wavelet transforms and
neural networks. J. Hydrol. Eng. 8 (6), 319–328.
Kitanidis, P.K., Bras, R.L., 1980. Real time forecasting with a
conceptual hydrologic model, 2. Applications and results. Water
Resour. Res. 16 (6), 1034–1044.
Kuczera, G., 1983. Improved parameter inference in catchment
models; 2. Combining different kinds of hydrologic data and
testing their compatibility. Water Resour. Res. 19 (5), 1163–1172.
Makkeasorn, A., Chang, N.B., Beaman, M., Wyatt, C., Slater, C.,
2006. Soil moisture estimation in a semiarid watershed using
RADARSAT-1 satellite imagery and genetic programming. Water
Resour. Res. 42, W09401. doi:10.1029/2005WR00403.
Parasuraman, K., Elshorbagy, A., 2007. Cluster-based hydrologic
prediction using genetic algorithm-trained neural networks. J.
Hydrol. Eng., ASCE 12 (1), 52–62.
Parasuraman, K., Elshorbagy, A., Carey, S., 2006. Spiking-modular
neural networks: A neural network modeling approach for
hydrological processes. Water Resour. Res. 42, W05412.
doi:10.1029/2005WR00431.
A. Elshorbagy, K. Parasuraman
Redlapalli, S.K., 2004. Development of Neural Units with Higher-
order Synaptic Weights Operations and Their Applications to
Logic Circuits and Control Problems, M.Sc. Thesis, University of
Saskatchewan, Saskatoon, SK, Canada.
Shamseldin, A.Y., 1997. Application of a neural network technique
to rainfall–runoff modeling. J. Hydrol. 199, 272–294.
Taylor, J.G., Commbes, S., 1993. Learning higher order correla-
tions. Neural Networks 6 (3), 423–428.
Thirumalaiah, K., Deo, M.C., 1998. River stage forecasting using
artificial neural networks. J. Hydrol. Eng. 3 (1), 26–32.
Wood, E.F., Lettenmaier, D.P., Zartarian, V.G., 1992. A land-
surface hydrology parameterization with subgrid variability for
general circulation models. J. Geophys. Res. 97, 2717–2728.
Wooldridge, S.A., Kalma, J.D., Walker, J.P., 2003. Importance of
soil moisture measurements for inferring parameters in hydro-
logic models of low-yielding ephemeral catchments. Environ.
Model. Software 18 (1), 35–48.
Zhang, B., Govindaraju, S., 2000. Prediction of watershed runoff
using Bayesian concepts and modular neural networks. Water
Resour. Res. 36 (3), 753–762.
Zhang, M., Fulcher, J., Scofield, R.A., 1997. Rainfall Estimation
Using Artificial Neural Network Group Neurocomputing, vol. 16.
pp. 97–115.