1429
Land use and land cover (LULC) play a central role in fate and transport of water quality (WQ) parameters in watersheds. Developing relationships between LULC and WQ parameters is essential for evaluating the quality of water resources. In this paper, we present an artifi cial neural network (ANN)–based methodology to predict WQ parameters in watersheds with no prior WQ data. Th e model relies on LULC percentages, temperature, and stream discharge as inputs. Th e approach is applied to 18 watersheds in west Georgia, United States, having a LULC gradient and varying in size from 2.96 to 26.59 km2. Out of 18 watersheds, 12 were used for training, 3 for validation, and 3 for testing the ANN model. Th e WQ parameters tested are total dissolved solids (TDS), total suspended solids (TSS), chlorine (Cl), nitrate (NO
3), sulfate (SO
4), sodium (Na),
potassium (K), total phosphorus (TP), and dissolved organic carbon (DOC). Model performances are evaluated on the basis of a performance rating system whereby performances are categorized as unsatisfactory, satisfactory, good, or very good. Overall, the ANN models developed using the training data performed quite well in the independent test watersheds. Based on the rating system TDS, Cl, NO
3, SO
4, Na, K, and DOC
had a performance of at least “good” in all three test watersheds. Th e average performance for TSS and TP in the three test watersheds were “good.” Overall the model performed better in the pastoral and forested watersheds with an average rating of “very good.” Th e average model performance at the urban watershed was “good.” Th is study showed that if WQ and LULC data are available from multiple watersheds in an area with relatively similar physiographic properties, then one can successfully predict the impact of LULC changes on WQ in any nearby watershed.
Predicting Water Quality in Unmonitored Watersheds Using Artifi cial Neural Networks
Latif Kalin* and Sabahattin Isik Auburn University
Jon E. Schoonover Southern Illinois University
B. Graeme Lockaby Auburn University
Land use and land cover (LULC) play a crucial role in
driving hydrological processes in watersheds (Schoonover
et al., 2006). Th ey aff ect water quality (WQ) by altering sedi-
ment, chemical loads, and watershed hydrology. Due to land use
practices and rapid land use changes, nonpoint-source pollution
loading becomes a serious threat to WQ in streams (Basnyat et
al., 2000). Many studies have shown that agricultural land use
adversely impacts stream WQ by increasing nutrient levels, such
as nitrogen and phosphorus, and sediment loadings (Hill, 1981;
Arnheimer and Liden, 2000; Ahearn et al. 2005). Urban areas
have similar negative impacts on WQ (Osborne and Wiley,
1988; Arnold and Gibbons, 1996; Basnyat et al., 1999; Sliva
and Williams, 2001, Schoonover et al., 2006). Some research-
ers attributed these to point sources such as wastewater effl uents.
A study conducted in southern Ontario, for example, found no
correlation between urban land use and stream water phosphorus
levels originating from nonpoint sources once the contribution
from wastewater discharges were removed (Hill, 1981).
Ahearn et al. (2005) studied the impact of LULC on sedi-
ment and nitrate loadings in both dry and normal years in the
waterways of the Cosumnes River watershed in California. Th ey
found that geographic variables have the greatest control on WQ
in the Cosumnes watershed and population density does not have
a strong infl uence on stream nitrate loading until a wastewater
treatment plant is built within the basin. However, agriculture
had a signifi cant infl uence on both total suspended sediment and
nitrate loading. Basnyat et al. (1999) examined a methodology to
assess the relationships between multiple land use activities and
nitrate–sediment concentrations in streams in south Alabama.
Th eir results indicate that forests act as a sink or an active trans-
formation zone, and as the proportion of forest increases (or agri-
cultural land decreases), nitrate levels decrease. Th ey identifi ed
residential–urban–built-up areas as the strongest contributors of
nitrate. Sliva and Williams (2001) found that urban land use had
the greatest infl uence on river WQ within three local southern
Ontario watersheds.
Abbreviations: AIC, Akaike’s information criterion; ANN, artifi cial neural network;
BIC, Bayesian information criterion; DOC, dissolved organic carbon; EV, evergreen; IS,
impervious surfaces; LULC, land use and land cover; MI, mixed forest; MLR, multiple
linear regression; NMSE, normalized mean square error; PA, pasture; TDS, total
dissolved solids; TP, total phosphorus; TSS, total suspended solids; UG, urban grass;
WQ, water quality.
L. Kalin, S. Isik, and B.G. Lockaby, School of Forestry and Wildlife Sciences, Auburn Univ.,
602 Duncan Dr., Auburn, AL 36849-5126; J.E. Schoonover, Dep. of Forestry, Southern
Illinois Univ., Carbondale, IL 62901-4411. Assigned to Associate Editor Ying Ouyang.
Copyright © 2010 by the American Society of Agronomy, Crop Science
Society of America, and Soil Science Society of America. All rights
reserved. No part of this periodical may be reproduced or transmitted
in any form or by any means, electronic or mechanical, including pho-
tocopying, recording, or any information storage and retrieval system,
without permission in writing from the publisher.
J. Environ. Qual. 39:1429–1440 (2010)
doi:10.2134/jeq2009.0441
Published online 11 May 2010.
Received 4 Nov. 2009.
*Corresponding author ([email protected]).
© ASA, CSSA, SSSA
5585 Guilford Rd., Madison, WI 53711 USA
TECHNICAL REPORTS: SURFACE WATER QUALITY
1430 Journal of Environmental Quality • Volume 39 • July–August 2010
Th e eff ects of LULC on water quality and quantity can be
explored through various techniques varying from regression-
based methods, such as linear and multilinear regression, to
watershed models. Linear regression is an important tool for
the statistical analysis of water resources data (Helsel and
Hirsch, 2002). Multiple linear regression (MLR) is the exten-
sion of simple linear regression to the case of multiple explana-
tory variables. Th e MLR relates one dependent variable y to k
independent variables or predictors xi (i = 1, . , k). Th e result is
an equation that can be used for estimating y as a linear combi-
nation of the predictors xi. Th e main weakness of MLR models
is that transformations include a priori assumptions about the
type and consistency of the relation between two parameters
that may not be met completely (Brey et al., 1996). Many
researchers (e.g., Basnyat et al., 1999; Ahearn et al., 2005;
Schoonover and Lockaby, 2006; Schoonover et al., 2007) have
used regression analysis to study the LULC and WQ linkages.
Watershed models also are used in estimating the eff ects LULC
on water quality and quantity. Even though, at least in theory,
some watershed models can be relied on in the absence of mea-
sured WQ data, in practice even the physically based water-
sheds models are often calibrated or fi ne-tuned (Fohrer et al.,
2001; Di Luzio et al., 2002).
If high-quality datasets of suffi cient duration exist, then arti-
fi cial neural networks (ANNs) could be eff ectively used in pre-
dicting the eff ects of LULC on WQ. Artifi cial neural networks
are parametric models that are generally considered lumped
(Dawson and Wilby, 2001). Neither a detailed understanding
of a watershed’s physical characteristics nor an extensive data
preprocessing is required for ANNs. Artifi cial neural networks
provide a novel and appealing solution to the problem of relat-
ing input and output variables in complex systems. (Dawson
and Wilby, 2001). Th e main advantage of using ANNs for
prediction purposes is that there are no a priori assumptions
about the relations between the independent and dependent
variables. However, those relations learned by an ANN are
hidden in its neural architecture and cannot be expressed in
traditional mathematical terms (Brey et al., 1996). A neural
network is more of a “black box” that delivers results without
an explanation of how the results were derived. Th us, it is dif-
fi cult or impossible to explain how decisions were made based
on the output of the network.
Th e use of ANNs in predicting WQ parameters is not
new (Maier and Dandy, 2000; Chau et al., 2002; Muttil and
Chau, 2006; Anctil et al., 2009; Amiri and Nakane 2009;
Dogan et al., 2009; Singh et al., 2009). Singh et al. (2009),
for instance, constructed an ANN-based WQ model for the
Gomti River (India) and demonstrated its application to pre-
dict WQ parameters. Th ey used 11 WQ parameters as inputs
to forecast dissolved oxygen and biochemical oxygen demand.
Similarly, Dogan et al. (2009) investigated the abilities of an
ANN model to improve the accuracy of biochemical oxygen
demand estimation in the Melen River (Turkey). Both stud-
ies relied on other measured WQ parameters to predict the
WQ parameters of interest. Anctil et al. (2009) applied ANNs
to simulate daily nitrate and suspended sediment fl uxes from
a small agricultural catchment. Th ey used hydroclimatic vari-
ables, such as streamfl ow, rainfall, and soil moisture index, and
historical mean nitrate and suspended sediment values to drive
their ANN model.
All of the aforementioned ANN-based studies were geared
toward predicting WQ parameters using input data such as
rainfall, streamfl ow, temperature, soil moisture index, and some
other WQ parameters. To the best of our knowledge, few studies
exist that incorporated the eff ect of LULC into ANNs to predict
WQ. Amiri and Nakane (2009) attempted to involve LULC per-
centages into an ANN model, while Ha and Stenstrom (2003)
used land use types as their target data. Amiri and Nakane (2009)
developed ANNs and MLR approaches to predict monthly aver-
age total nitrogen concentrations in Chugoku district of Japan
by using LULC percentages and human population density in
21 river basins as inputs. Th ey compared the performance of an
ANN-based model to that of the MLR modeling approach and
found better estimation with the ANN.
Th e main objective of this paper is to develop an ANN-based
approach to examine the relationship between LULC and various
WQ parameters and use it to predict WQ in nearby ungauged
and/or unmonitored watersheds with similar characteristics.
Similar to Amiri and Nakane (2009), we used LULC percentages
as one of the key model drivers supplementing temperature and
streamfl ow. A key diff erence between this study and Amiri and
Nakane’s study (2009) is that while our study totally relies on mea-
sured data, they generated most of their data through Monte Carlo
simulations. We applied the ANN model to 18 watersheds in the
Piedmont physiographic region of western Georgia. Th e WQ
parameters used in the study were total dissolved solids (TDS),
total suspended solids (TSS), chlorine (Cl), nitrate (NO3), sul-
fate (SO4), sodium (Na), potassium (K), total phosphorus (TP),
and dissolved organic carbon (DOC). Th e input variables (i.e.,
independent variables) were LULC percentages, temperature, and
streamfl ow. We limited the number of input parameters to the
ANN model since we want a model that can be used in predicting
WQ parameters in watersheds with no prior WQ measurements.
First, we explain the methodology used in developing the ANN
model, which is followed by description of the study area and data.
Next, the application of the ANN model to the study area is fol-
lowed by discussion of results.
Materials and Methods
Artifi cial Neural NetworksAn ANN is a machine (tool) designed to model the manner
in which the human brain performs a particular task or func-
tion of interest. To achieve good performance, neural net-
works use a massive interconnection of simple computing
cells referred to as neurons or processing units. Artifi cial neural
networks are capable of mapping input–output relationships
for natural complex problems and were developed to model
the brain’s interconnected system of neurons so that computers
could be used to imitate the brain’s ability to sort patterns and
learn from trial and error, thus observing relationships in data
(Haykin, 1999).
Artifi cial neural networks can be categorized on the basis
of the direction of information fl ow and processing. In a feed-
forward network, the nodes are generally arranged in layers,
starting from a fi rst input layer and ending at the fi nal output
layer. Information passes from the input to the output side.
Kalin et al.: Predicting Water Quality in Unmonitored Watersheds 1431
A synaptic weight is assigned to each link to repre-
sent the relative connection strength of two nodes at
both ends in predicting the input–output relationship
(ASCE Task Committee, 2000).
Artifi cial neural networks are highly data inten-
sive for training the network. Th e primary goal of
training is to minimize a predefi ned error function
by searching for a set of connection strengths and
threshold values so that the ANN can produce out-
puts that are equal or close to target values (ASCE
Task Committee, 2000). One of the commonly used
error function is the mean square error (MSE):
2
1
1MSE
n
i ii
S On
− [1]
where Si is the ANN output (simulated) and O
i is the
target (observation).
Since ANNs are the “black-box” class of models,
they do not require detailed knowledge of the internal
functions of a system to recognize relationships between
inputs and outputs (Ha and Stenstrom, 2003). Feed-
forward neural networks with back propagation are successfully
applied to hydrological and environmental problems. In this
study, three-layer feed-forward neural networks with Levenberg–
Marquardt back-propagation learning were constructed for the
relationship between LULC percentages and WQ parameters.
Th e proposed feed-forward neural network has three main
layers: input, hidden, and output layers. Th e hidden layer also has
multiple sublayers. Th e number of sublayers in the hidden layers
varies with WQ parameters. Th e architecture of neural network is
shown in Fig. 1. Th e percentages of fi ve dominant LULC types
(impervious surface [IS]; evergreen forest [EV]; mixed forest [MI];
pasture [PA]; and urban grass [UG]), temperature eff ect (Teff
), and
stream discharge (Q) constitute the neurons of the input layer. Th e
WQ parameters (TDS, TSS, Cl, NO3, SO
4, Na, K, TP, and DOC
loadings) are the output parameters.
Th e size of a hidden layer is one of the most important con-
siderations when solving actual problems using multilayer feed-
forward networks. No unifi ed theory exists for determining
such an optimal ANN architecture (ASCE Task Committee,
2000). Th e exact analysis of the issue is rather diffi cult because
of the complexity of the network mapping and due to the non-
deterministic nature of many successfully completed training
procedures (Zurada, 1992). Determination of the optimum
number of layers is usually a matter of experimentation. A
trial-and-error approach is the most commonly used method
to fi nd the number of hidden neurons and layers. In this study,
the number of hidden layers and hidden neurons were searched
from 1 to 2 and from 1 to 10, respectively. Th e commercial
software MATLAB (Th e MathWorks, Inc., Natick, MA) was
used in developing the ANN models.
Model SelectionNormalized mean square error (NMSE), Akaike’s information
criterion (AIC), and Bayesian information criterion (BIC) are
used as selection criteria in determining optimal input and hidden
neurons. We defi ne a revised form of MSE in this study due to
the nature of the problem. In this application, we need an error
measure that combines information from multiple watersheds.
Because we have multiple watersheds with varying size and vary-
ing number of measurements, MSE is not a suitable measure.
Th us, a NMSE was used for this purpose and is given by
[2a]
or
[2b]
where m is the total number of watersheds; nj is the total number
data in watershed j; Oj,i and S
j,i are the ith observed and simu-
lated values in watershed j, respectively; and jO is the average
of observed values in watershed j. Th ere are two reasons for the
use of NMSE for a given WQ parameter: to minimize the eff ect
of sample number and to minimize the eff ect of large and small
observations from watersheds. Note that we combined data from
several watersheds in training the ANN model. Th e number of
observed data from each watershed is not the same. If we simply
use MSE, then watersheds having more observed data will be
given more weight. Further, watersheds having high observed
values (e.g., due to diff erences in their size) will also carry higher
weights in the simple MSE formula.
Th e AIC and the BIC are commonly used in the literature
to fi nd optimal ANN architectures (Qi and Zhang, 2001; Ren
and Zhao, 2002; Zhao et al., 2008). Information-based criteria
such as AIC and BIC penalize large models that often tend to
overfi t (Qi and Zhang, 2001). Various forms of AIC and BIC
are used in the literature. We used the one proposed by Qi and
Zhang (2001):
Fig. 1. General architecture of artifi cial neural network (ANN) model. IS, impervi-ous surfaces; EV, evergreen; MI, mixed forest; PA, pasture; UG, urban grass; T
eff ,
temperature eff ect; Q, streamfl ow discharge; TDS, total dissolved solids; TSS, total suspended solids; TP, total phosphorus; DOC, dissolved organic carbon.
2
, ,
21 1
1NMSE jm n j i j i
j ijj
S O
On
−
1
2
1, 1,
2 111
2
, ,
2 1
1NMSE
1m
n i i
i
n m i m i
imm
S O
On
S O
On
−
−
1432 Journal of Environmental Quality • Volume 39 • July–August 2010
2MLEAIC log 2 if 1 40m n n m [3a]
2MLEAIC log 2 ( 1) if 1 40m n m n m− − [3b]
where n is the number of data and m is the number of param-
eters in the model. Th e term 2MLE denotes the maximum like-
lihood estimate of variance of the residual term or simply the
MSE between the observed and simulated data. Qi and Zhang
(2001) give BIC as
2MLEBIC log logm n n [4]
Performance MeasuresTh e performance of the model was measured with the coeffi cient
of determination (R2), Nash–Sutcliff e effi ciency (ENASH
), and
bias ratio (RBIAS
). Th e coeffi cient of determination is a measure
of linear correlation between two quantities and is given by
[5]
where O and S represent observed data and model outputs and n
is the number of data points. Th e Nash–Sutcliff e effi ciency sta-
tistic (ENASH
) is commonly used to assess the predictive power of
hydrological models (Nash and Sutcliff e, 1970). It is defi ned as
2
NASH 21
i i
i
O SE
O O
−−
− [6]
where O is the mean of the observed data. Th e effi ciency
statistic ENASH
theoretically varies from –∞ to 1 with 1 cor-
responding to a perfect model. It is a measure of how the plot
of observed versus simulated data deviates from a 1:1 line (i.e.,
perfect model). Th e bias ratio in percentage is expressed as
BIAS 100i i
i
S OR
O−
[7]
Th e bias ratio measures the degree to which the forecast is
under- or overpredicted. A negative bias ratio indicates under-
prediction, whereas a positive bias ratio refl ects overprediction
(Salas et al., 2000).
Study Area and DataWe applied the outlined ANN model to 18 small water-
sheds in western Georgia, near the city of Columbus (Fig. 2).
Th ese watersheds present a gradient of LULC. Th e southeast-
ern United States has experienced rapid urban development.
Consequently, Georgia’s streams have experienced hydrologic
alterations and WQ degradation from extensive development
and from other land use activities such as livestock grazing and
silviculture (Schoonover, 2005). Grab samples were collected
from May 2002 to January 2006 and analyzed for concentra-
tion and yields of TDS, TSS, Cl, NO3, SO
4, Na, K, TP, and
DOC at each watershed (Table 1). Details on sampling strate-
gies and chemical analysis are given in Schoonover (2005).
Watersheds ranged in size from 296 to 2659 ha and were
subbasins of the Middle Chattahoochee Watershed within the
Piedmont physiographic province. Dominant LULC within
the study area were classifi ed as mixed hardwood forest, ever-
green forest, urban, developing, and pastoral. One-meter
aerial photographs were taken during leaf-off in March 2003
to facilitate LULC classifi cation. Th e fi rst eff ort in the 1-m
image analyses was to generate an impervious (IS) percentage
for each watershed. Impervious surface is a widely accepted
and reliable indicator of urbanization due to its impacts on
natural resources, particularly for water resources (Arnold and
Gibbons, 1996). Th e remaining land cover classes were then
digitized using both unsupervised and supervised classifi ca-
tion methods. Th e overall accuracy was 91%. (Schoonover and
Lockaby, 2006). Th e image processing methods used in this
assessment are described in detail by Lockaby et al. (2005).
Fig 2. Watersheds used in this study: BLN, Blanton Creek; BR, Brookstone Branch; BU, Lindsey/Cooper Creek; CB, Clines Branch; FR, Flat Rock Creek; FS, Wildcat Creek; HC, House Creek; MU, Mulberry Creek; RB, Roaring Branch; SB, Standing Boy Creek; SC, Sand Creek.
2
2
2 22 2
i i i i
i i i i
n O S O SR
n O O n S S
−
− −
Kalin et al.: Predicting Water Quality in Unmonitored Watersheds 1433
Th e rates of most reactions in natural waters increase by
temperature (Chapra, 1997). Th erefore, we included tem-
perature as one of the input variables through the use of the
Arrhenius equation (Chapra, 1997):
0w w
0
20eff
20
T TKT
K−
[8]
where Tw is ambient water temperature (°C) and θ is a dimen-
sionless parameter typically within the range 1.0 to 1.1 but
assumed to be 1.05 in this study; Tw is computed from average
daily air temperature avT as given in Neitsch et al. (2005):
avw 5.0 0.75T T [9]
avT values were obtained from a nearby National Climatic
Data Center (NCDC) station in the city of Columbus
(COOPID:092159; 32°31′ N; 84°56′ W).
Areas and LULC percentages of the 18 study watersheds are
given in Table 1. Percentage of IS ranges from 1.2 to 41.9%.
Forest occupies a major fraction of each watershed. Th e range
for percentage of EV is 20.9 to 48.3%. Percentage of MI varies
from 7.0 to 37.1%. Percentage of land in PA was quite vari-
able, with a range of 5.5 to 44.5%. Urban grass percentage
was usually small with a range 0.1 to 18%. Other LULC types
constitute only minor fractions of the watersheds and therefore
were not included in the analyses.
Th e total number of data points for each WQ parameter
was 801, ranging from 15 to 54 among watersheds. Out of 18
watersheds, 12, which contained 66% of the total data, were
used for training the ANN model; 3 watersheds were used for
validation, and the remaining 3 for testing purposes (Table 1).
Th e validation and testing watersheds contained about 16 and
18% of the total data set, respectively. Each set of validation
and testing data consisted of 1 forested, 1 pastoral, and 1 urban
watershed, while training data consisted of 7 forested, 3 pasto-
ral, and 2 urban watersheds. Land use–based classifi cations of
the watersheds were based on Schoonover (2005).
Nutrient yields (kg ha−1 d−1) were calculated and used in
the ANN network for each parameter. Summary statistics such
as arithmetic mean, minimum, maximum, median, standard
deviation, and coeffi cient of variation of training, testing, and
validation data are given for each WQ parameter in Table 2.
Total suspended solids shows the largest variation among all
parameters as evidenced by its large coeffi cient of variation
values in training, validation, and testing data, which were
8.76, 4.85, and 7.44, respectively.
Natural logarithms of WQ parameters were used in the net-
work to avoid zero outputs since we have very low target values.
Before the training of the network, all data were normalized
within the range 0.1 to 0.9 as follows:
min
max min
0.1 0.8i
ix x
zx x
−
− [10]
where zi is the normalized value of x
i, which is the log-trans-
formed observed value of a certain parameter, and xmin
and xmax
are the minimum and maximum values in the database for this
parameter, respectively. Th e observed data and model output
values are transformed back to their original domains before
evaluating model performances.
Results and DiscussionTh e LULC percentages of IS, EV, MI, PA, and UG, temperature
eff ect (Teff
), and streamfl ow discharge (Q) were used as inputs to
the ANN network. We experimented with various combinations
of these input parameters to identify the optimal input layer.
Table 1. Land use/land cover (LULC) classes, land use percentages, and watershed areas.
Basin no.
Basin ID†
LULC class‡
Number of data
Purpose Area (ha)LULC percentages (%)§
IS EV MI PA UG Other
1 CB F 49 Training 897 1.5 48.3 33 11.8 0.1 5.3
2 HC F 50 665 1.3 47.9 26.7 18 0.3 5.8
3 MU2 F 52 606 2.6 42.4 25 14.4 1.2 14.4
4 MU3 F 46 1044 1.9 41.5 37.1 13 0.8 5.7
5 SB1 F 52 2009 1.8 38.6 35 18.8 0.6 5.2
6 SB2 F 52 634 3.4 37.3 35.4 16.3 1.5 6.1
7 SB4 F 54 2659 3.3 41.1 22.7 25.5 2.2 5.2
8 FR P 15 2396 13 31 7 35.6 4.9 8.5
9 HC2 P 37 1395 1.6 30.5 22.2 44.5 0.6 0.6
10 MU1 P 53 1178 3.7 29.3 24.3 35 2.8 5
11 BR U 15 471 23.0 29 14 10.9 16.1 7
12 RB U 54 367 30.3 28.4 11.1 10.9 16.9 2.4
13 BLN F 39 Validation 364 1.2 48.1 28.3 18.4 0.2 3.8
14 FS3 P 36 296 2.6 32 29.9 33.1 0.5 1.9
15 BU2 U 50 2469 24.9 30.5 15.9 7.6 18 3.1
16 SC F 53 Testing 896 1.2 44.8 28.8 20.3 0.2 4.7
17 FS2 P 36 1449 2.7 30.7 28.2 35.2 0.8 2.4
18 BU1 U 54 2548 41.9 20.9 12.3 5.5 17.6 1.8
† See Fig. 2 caption for full names.
‡ F, forest; P, pasture; U, urban.
§ IS, impervious surfaces; EV, evergreen; MI, mixed forest; PA, pasture; UG, urban grass.
1434 Journal of Environmental Quality • Volume 39 • July–August 2010
We tried the combinations LULC, Q, LULC + Q, LULC + Teff
,
Teff
+ Q, and LULC + Teff
+ Q. Th e AIC, BIC, and NMSE error
criteria were used in determining optimal input layers. Results
are given in Table 3. Mostly, all three error measures consistently
picked the same combination. At least two criteria picked the
same input layer in all WQ parameters. Th e LULC + Teff
+ Q
combination for all WQ parameters was determined to be gen-
erating better results than other combinations. Th e WQ param-
eters TDS, TSS, Cl, NO3, SO
4, Na, K, TP, and DOC were the
dependent variables in the proposed ANN models. We devel-
oped a separate ANN model for each WQ parameter.
Table 3 also provides useful information on parameter sensi-
tivities. Normalized mean square error can be used as a sensitivity
measure. If the model is insensitive to a parameter, then adjust-
ing that parameter would not improve the model performance
(low NMSE in this case). Only calibration of sensitive param-
eters could yield improved model performances. From Table 3,
it is evident that the ANN model is more sensitive to Q for TDS,
Cl, Na, K, and TP and to LULC for TSS, SO4, NO
3, and DOC.
In this study, the number of hidden neurons was searched
from 1 to 10, and the number of hidden layers was searched
from 1 to 2. We limited the size of hidden layers to 10 nodes
in each hidden layer as networks over 10 nodes did not result
in better performance based on NMSE, AIC, and BIC. For all
WQ parameters, model performance peaked before reaching
10 nodes and steadily decreased after that. Th e highest number
for optimum number of nodes was 7, which was obtained for
DOC. Table 4 presents the optimum number of hidden neu-
rons for each WQ parameter. As an example, for TSS there
were two neurons in each of the two hidden layers with a total
of four neurons. Th e optimum number of hidden layers was
1 for NO3, SO
4, Na, and DOC and 2 for TDS, TSS, Cl, K,
and TP. Th e optimal number of neurons in these hidden layers
varied from 1 to 7 (Table 4). A trial-and-error procedure was
used to determine the learning rate and momentum parameter.
Th eir values were 0.01 and 0.5, respectively. Th e log-sigmoid
transfer function is adopted for both hidden and output layers.
Th e network training stops as soon as any of these conditions
occur: (i) model performance in validation dataset decreases in
10 successive iterations; (ii) the maximum number of epochs,
which is predetermined at 1000, is reached.
Th e R2 and NMSE for the training and validation data sets
are given in Table 5. Th e training dataset was only used for
training the ANN model to identify the ANN model param-
eters (i.e., weights and biases); it was not used to measure the
performance of the models. Indeed, the independent valida-
tion dataset (see Table 1) is used in selecting the best models.
Th is was also done to prevent the overtraining of the model
(Srivastava et al., 2006). Except for TSS, all WQ parameters
have R2 values at or above 0.7 in the validation dataset. Th e
R2 for TSS is 0.49. However, one should note that it is dif-
fi cult to make real comparisons between model performances
for diff erent WQ parameters based on R2. As stated earlier, R2 is
merely an indication of the degree of linear correlation between
two datasets. Normalized mean square error is a better metric
for interparameter comparisons. It is in a sense similar to bias
or mass balance error. Th e parameter TSS has higher NMSE
values than all other WQ parameters, about 0.1; TDS, K, Na,
and Cl all have very low NMSE values.
Table 6 presents the R2, ENASH
, and RBIAS
model performance
criteria at the three test watersheds for each WQ parameter.
Simulated and observed values of each WQ parameter are shown
Table 2. Summary statistics of input data (eff ective temperature, fl ow discharge, and water quality) used for training, validation, and testing the artifi cial neural network (ANN) model.†
Teff
Q TDS TSS Cl NO3
SO4
Na K TP DOC
L s−1 ha−1 ——————————————————————— kg ha−1 d−1 ———————————————————————
Training
Min. 0.005 0.0001 0.0003 0 0.00002 0 0.00002 0.0002 0 0 0.00002
Max. 0.014 9.336 20.073 183.57 3.740 1.039 2.743 2.278 1.961 0.127 6.081
Mean 0.009 0.231 0.687 1.29 0.085 0.019 0.079 0.092 0.043 0.003 0.140
Median 0.009 0.069 0.230 0.02 0.025 0.003 0.016 0.033 0.013 0.0005 0.026
SD 0.003 0.633 1.785 11.28 0.279 0.069 0.234 0.206 0.119 0.010 0.428
CV 0.275 2.747 2.597 8.76 3.276 3.686 2.955 2.236 2.774 3.257 3.053
Validation
Min. 0.005 0.001 0.007 0 0.001 0 0.001 0.001 0.001 0 0.001
Max. 0.014 1.505 4.147 21.06 0.563 0.276 0.606 0.639 0.440 0.045 0.987
Mean 0.009 0.179 0.496 0.46 0.065 0.029 0.060 0.054 0.041 0.002 0.066
Median 0.009 0.107 0.238 0.03 0.030 0.011 0.016 0.029 0.020 0.001 0.024
SD 0.003 0.243 0.750 2.22 0.099 0.049 0.118 0.084 0.067 0.005 0.140
CV 0.274 1.355 1.513 4.85 1.529 1.675 1.987 1.545 1.638 2.591 2.118
Testing
Min. 0.005 0.004 0.019 0 0.004 0.00005 0.001 0.004 0.003 0 0.001
Max. 0.014 5.611 10.423 172.28 0.815 0.695 1.175 0.709 1.016 0.257 3.147
Mean 0.009 0.251 0.585 2.23 0.078 0.045 0.071 0.061 0.048 0.004 0.093
Median 0.009 0.082 0.216 0.03 0.031 0.014 0.017 0.029 0.017 0.001 0.022
SD 0.003 0.611 1.209 16.61 0.137 0.095 0.153 0.102 0.106 0.022 0.292
CV 0.275 2.437 2.066 7.44 1.760 2.126 2.163 1.665 2.212 5.234 3.150
† Teff
, temperature eff ect; Q, streamfl ow discharge; TDS, total dissolved solids; TSS, total suspended solids; TP, total phosphorus; DOC, dissolved
organic carbon.
Kalin et al.: Predicting Water Quality in Unmonitored Watersheds 1435
Table 3. The best performances for input layers of water quality parameters.
Parameter† Input layer‡ NMSE§ AIC§ BIC§
TDS LULC + Teff
+ Q 0.0026 −3.8 −3.7
LULC + Q 0.0029 −3.6 −3.6
Q 0.0084 −2.7 −2.7
Q + Teff
0.0132 −2.7 −2.6
LULC 0.0284 −0.7 −0.6
LULC + Teff
0.0272 −0.6 −0.5
TSS LULC + Teff
+ Q 0.0804 1.0 1.2
LULC + Q 0.0947 1.1 1.1
Q 0.1578 1.1 1.2
Q + Teff
0.1036 1.2 1.3
LULC 0.1530 1.7 1.7
LULC + Teff
0.1549 1.7 1.7
Cl LULC + Teff
+ Q 0.0064 -6.4 −6.2
LULC + Q 0.0068 −6.4 -6.3
Q 0.0237 −5.8 −5.8
Q + Teff
0.0247 −5.8 −5.7
LULC 0.0276 −4.7 −4.7
LULC + Teff
0.0260 −4.6 −4.6
NO3
LULC + Teff
+ Q 0.0106 −7.6 −7.5
LULC + Q 0.0204 −7.4 −7.3
Q 0.2997 −7.0 −7.0
Q + Teff
0.0561 −6.3 −6.2
LULC + Teff
0.0399 −6.1 −6.0
LULC 0.0388 −6.1 −6.0
SO4
LULC +Teff
+ Q 0.0256 -6.4 -6.4
LULC + Q 0.0462 −6.3 −6.2
Q 0.1697 −5.4 −5.3
Q + Teff
0.2076 −5.3 −5.2
LULC 0.0468 −4.4 −4.4
LULC + Teff
0.0388 −4.3 −4.3
Na LULC + Teff
+ Q 0.0069 -6.5 -6.4
Q + Teff
0.0126 −6.1 −6.0
LULC + Q 0.0077 −6.1 −6.0
Q 0.0101 −5.9 −5.9
LULC 0.0249 −5.0 −5.0
LULC + Teff
0.0236 −5.0 −4.9
K LULC + Teff
+ Q 0.0041 -7.2 -7.1
LULC + Q 0.0039 −7.1 −7.1
Q + Teff
0.0053 −6.8 −6.7
Q 0.0069 −6.7 −6.7
LULC + Teff
0.0286 −5.5 −5.4
LULC 0.0290 −5.4 −5.4
TP LULC + Teff
+ Q 0.0399 −11.6 −11.5
Q + Teff
0.0446 −11.4 −11.3
LULC + Q 0.0515 −11.1 −11.0
Q 0.0554 −11.0 −11.0
LULC 0.0880 −10.3 −10.3
LULC + Teff
0.0760 −10.3 −10.2
DOC LULC + Teff
+ Q 0.0262 −5.8 −5.8
LULC + Q 0.0347 −5.8 −5.6
Q 0.0634 −5.6 −5.6
Q + Teff
0.1189 −5.5 −5.4
LULC 0.0308 −3.9 −3.9
LULC + Teff
0.0315 −3.9 −3.8
† TDS, total dissolved solids; TSS, total suspended solids; TP, total phosphorus; DOC, dissolved organic carbon.
‡LULC, land use and land cover; Teff
, temperature eff ect; Q, streamfl ow discharge.
§ NMSE, normalized mean square error; AIC, Akaike’s information criterion; BIC, Bayesian information criterion.
1436 Journal of Environmental Quality • Volume 39 • July–August 2010
on scatterplots for each of the three test watersheds in Fig. 3.
Overall, the ANN model performs quite well and exception-
ally well for some WQ parameters regardless of the watershed.
Th ere are no established criteria in the literature for good–bad
model performance based on any of these three metrics. Moriasi
et al. (2007) proposed performance ratings based on some rec-
ommended statistics that include ENASH
, and RBIAS
in watershed
modeling at monthly time scale. Our time scale is much smaller
(instantaneous). Models are known to perform better at coarser
scales. Taking Moriasi et al. (2007) as the base and relaxing some
of the constraints, we developed the following performance
rating in evaluating the ANN model performance:
Very Good: ENASH
≥ 0.7; |RBIAS
| ≤ 0.25
Good: 0.5 ≤ ENASH
< 0.7; 0.25 < |RBIAS
| ≤ 0.5
Satisfactory: 0.3 ≤ ENASH
< 0.5; 0.50 < |RBIAS
| ≤ 0.7
Unsatisfactory: ENASH
< 0.3; |RBIAS
| > 0.7
Total Dissolved Solids
Both R2 and ENASH
values were quite high in all three water-
sheds. Th e lowest ENASH
was at the pastoral watershed FS2,
with a value of 0.95. Th e model also overestimated TDS in
this watershed by 26%. Based on the criteria we set, the devel-
oped ANN model performance can be considered “very good”
to “good,” with an average rating of “very good.” Overall, the
ANN model developed for TDS had one of the best perfor-
mances compared with other WQ parameters.
Total Suspended Solids
Although observed TSS data contained more basefl ow data
than storm data, the developed ANN model performed strik-
ingly well at all three watersheds. Based on our rating system the
model performance is “very good/good” at the forested and pas-
toral watersheds SC and FS2, respectively. Th e urban watershed
BU1 received a “satisfactory” rating. Th e overall rating based on
the average rating from the three watersheds was “good.”
Chloride
Model performance for Cl varied from “good” to “very good.”
It produced best results at the pastoral watershed, which was
surprising. We expected better model performance at the urban
watershed as Cl is often found in potable water and on roads
during winter months as deicing material. Although chlorine
is added to water at the water treatment plants, it is sometime
added to irrigation water also. Some of the areas classifi ed as pas-
ture in the study watersheds could potentially be agricultural.
For instance, it is almost impossible to distinguish between hay
and soybeans from aerial photos or remote
sensing, unless there is ground-truthing.
Nitrate
Th e developed model predicted NO3
levels quite well in each watershed based
on ENASH
values. Bias ratios were higher
than expected given the ENASH
values.
Although nitrate level was overpredicted
in the urban watersheds, it was underpre-
dicted in the forested and pastoral water-
sheds. Model performances were “good/
very good” at all three watersheds.
Sulfate
Th e developed ANN model performs quite
well at each watershed for SO4 with model
performance varying from “good” to “very
good.” Th ere were no distinct diff erences in
model performances between watersheds.
Sources of sulfate could be atmospheric
or from groundwater. Sulfates also occur
naturally in minerals and in some rock for-
mations and thus may be present due to
weathering processes.
Sodium and Potassium
Th e ANN models developed for Na and
K both worked exceptionally well with
performance ratings of “very good” at
forested and pastoral watersheds for both
WQ parameters. Th e performance at
urban watershed was also “very good” for
K, but the performance at urban water-
shed varied from “good” to “very good”
for Na.
Table 4. Number of neurons in input, hidden, and output layers for each water quality parameter.
Parameter†
Number of neurons Best performances‡
1st hidden layer
2nd hidden layer
NMSE AIC BIC
TDS 3 3 0.0026 −3.9 −3.7
TSS 2 2 0.0943 1.1 1.2
Cl 2 1 0.0069 −6.2 −6.1
NO3
5 – 0.0120 −7.6 −7.4
SO4
4 – 0.0110 −6.1 −6.0
Na 6 – 0.0089 −5.9 −5.7
K 4 1 0.0048 −6.9 −6.8
TP 3 3 0.0533 −11.0 −10.8
DOC 7 – 0.0139 −6.6 −6.5
† TDS, total dissolved solids; TSS, total suspended solids; TP, total phosphorus; DOC, dissolved
organic carbon.
‡ NMSE, normalized mean square error; AIC, Akaike’s information criterion; BIC, Bayesian informa-
tion criterion.
Table 5. R2 and normalized mean square error (NMSE) values obtained for each water quality parameter during training and validation of the artifi cial neural network (ANN) models.
Parameter†Training Validation
R2 NMSE‡ R2 NMSE
TDS 0.93 0.0074 0.97 0.0030
TSS 0.56 0.0290 0.49 0.1010
Cl 0.74 0.0286 0.81 0.0062
NO3
0.85 0.0181 0.85 0.0131
SO4
0.87 0.0430 0.73 0.0170
Na 0.92 0.0058 0.79 0.0051
K 0.89 0.0089 0.85 0.0038
TP 0.62 0.0350 0.71 0.0427
DOC 0.96 0.0084 0.93 0.0131
† TDS, total dissolved solids; TSS, total suspended solids; TP, total phosphorus; DOC, dissolved
organic carbon.
‡ NMSE, normalized mean square error.
Kalin et al.: Predicting Water Quality in Unmonitored Watersheds 1437
Total Phosphorus
All three watersheds had quite similar ENASH
values varying
between 0.52 and 0.58. Th e model under predicted TP loadings
at each watershed. Th e largest underprediction was at the urban
watershed BU1. Th e forested watershed had the lowest under-
prediction at 16%. Model performance varied from “very good/
good” to “good/satisfactory.” Th e extremely high R2 value of 0.99
at the urban watershed BU1 indicated a systematic over/under
prediction of the model, where we know from RBIAS
that it under-
predicted observed TP loadings by almost 60%. Th is is fortunate
since systematic errors are easier to fi x. Systematic errors are related
to model structure and could be stemming from ignoring some of
the processes or due to use of some redundant variables.
Dissolved Organic Carbon
Th e ANN model performance for DOC was “very good” in
all three watersheds. It overpredicted DOC loadings in all by
15 to 25%. Model performance in the urban watershed was
superior to model performance in the other two watersheds.
Summary and ConclusionsWe presented a methodology based on artifi cial neural networks
to predict water quality parameters in unmonitored basins. Th e
model relied on LULC percentages, temperature, and fl ow dis-
charge as inputs. Th e developed model made use of WQ and
fl ow data from nearby watersheds with similar physical char-
acteristics. Th e only required measurements at the watershed
where WQ parameters were needed are fl ow and temperature.
Th e model was applied to several watersheds in west Georgia
varying in size and LULC. Th e WQ parameters used in this
application were TDS, TSS, Cl, NO3, SO
4, Na, K, TP, and
DOC. Out of the total 18 watersheds, 12 were used in training
model parameters, 3 in model validation, and 3 for testing. Each
set of validation and testing data consists of 1 forested, 1 pasto-
ral, and 1 urban watershed, while the training dataset consisted
of 7 forested, 3 pastoral, and 2 urban watersheds. Th e model
developed using the training data set has successfully predicted
the WQ parameters in the independent testing watersheds.
To better compare interparameter and interwatershed model
performances, we developed a qualitative performance rating
system. According to this rating system model performances
were categorized as unsatisfactory, satisfactory, good, or very
good. Th e statistical measures Nash–Sutcliff e effi ciency (ENASH
)
and bias ratio (RBIAS
) were used in determining the performance
ratings. Based on this rating system, TDS, Cl, NO3, SO
4, Na,
K, and DOC had a performance of at least “good” in all three
Table 6. Performance statistics (R2, ENASH
, and RBIAS
) of the developed artifi cial neural network (ANN) models at each testing watersheds for the selected water quality parameters.
WSD† R2 ENASH
† RBIAS
† Performance‡
TDS§ SC 0.97 0.97 0.8 VG
FS2 0.99 0.95 26.2 VG/G
BU1 0.99 0.99 −5.2 VG
TSS SC 0.69 0.66 −11.4 VG/G
FS2 0.80 0.76 −28.8 VG/G
BU1 0.42 0.31 −54.9 S
Cl SC 0.61 0.61 −12.7 VG/G
FS2 0.96 0.94 15.9 VG
BU1 0.81 0.78 −24.0 VG
NO3
SC 0.84 0.79 −34.5 VG/G
FS2 0.91 0.84 −41.3 VG/G
BU1 0.86 0.77 28.9 VG/G
SO4
SC 0.90 0.87 25.4 VG/G
FS2 0.98 0.98 7.2 VG
BU1 0.83 0.79 −13.7 VG
Na SC 0.92 0.89 16.9 VG
FS2 0.98 0.98 9.4 VG
BU1 0.95 0.92 29.5 VG/G
K SC 0.97 0.96 −6.2 VG
FS2 0.97 0.97 3.4 VG
BU1 0.98 0.91 −20.4 VG
TP SC 0.60 0.54 −16.2 VG/G
FS2 0.71 0.58 −38.0 G
BU1 0.99 0.52 −58.9 G/S
DOC SC 0.75 0.75 14.6 VG
FS2 0.91 0.88 24.7 VG
BU1 0.98 0.95 18.9 VG
† WSD, watersheds; ENASH
, Nash–Sutcliff e effi ciency; RBIAS
, bias ratio.
‡ VG, very good; G, good; S, satisfactory.
§ TDS, total dissolved solids; TSS, total suspended solids; TP, total phosphorus; DOC, dissolved organic carbon.
1438 Journal of Environmental Quality • Volume 39 • July–August 2010
test watersheds. Th e average performance for TP in the three
test watersheds was “good,” with the lowest being “good/satisfac-
tory.” Total suspended solids had the lowest average performance
among all WQ parameters. It had a performance of “satisfac-
tory” at the urban watershed, whereas the forested and pastoral
watershed had performance rating of “good/very good.”
Th e average of the ENASH
for all WQ parameters was higher
at the pastoral watershed than in the forested and urban water-
sheds, with a value of 0.88. Th e average ENASH
values from all
WQ parameters for the urban and forested watersheds were 0.77
and 0.78, respectively. In addition to having the smallest aver-
age ENASH
values, the urban watershed also had a larger varia-
tion in ENASH
values compared with the forested and pastoral
watersheds, implying larger uncertainties associated with the
urban watersheds. Based on RBIAS
values, however, the ANN
model worked much better in the forested watershed. Averages
of the absolute values of RBIAS
were 15.4, 21.7, and 28.3% for the
forested, pastoral and urban watersheds, respectively. Standard
Fig. 3. Scatter plots of ANN generated and measured loadings for the water quality parameters total dissolved solids (TDS), total suspended solids (TSS), total phosphorus (TP), and dissolved organic carbon (DOC). The abbreviations on the upper left corner of each fi gure refer to watershed names: SC, Sand Creek; FS2, Wildcat Creek; BU1, Lindsey Creek.
Kalin et al.: Predicting Water Quality in Unmonitored Watersheds 1439
Fig. 3. Continued.
1440 Journal of Environmental Quality • Volume 39 • July–August 2010
deviation of absolute RBIAS
values was also lower at the forested
watershed. It was 9.4% at the forested watershed and 12.7 and
16.9% at the pastoral and urban watersheds, respectively. If we
had applied the rating system to the combined performances of
diff erent WQ parameters, the forested and pastoral watersheds
would have received a “very good” performance. Th e perfor-
mance of the urban watershed was “good/very good.”
Results from this study indicate that if WQ and LULC data
are available from multiple watersheds in an area with relatively
similar physiographic properties, then one can successfully pre-
dict the impact of LULC changes on WQ in any nearby water-
shed if streamfl ow data are available or can be estimated. In
this study, we did not attempt to predict fl ow discharge, which
is one of the limitations of the study. Since all the WQ data
were “snapshots” in time, taken at irregular time intervals, fl ow
discharge data were needed at the times of those WQ measure-
ments. It is extremely diffi cult to predict instantaneous fl ows.
Because the study watersheds are quite small, rainfall data with
high temporal resolution (in addition to soil characteristics, and
topographic and morphologic parameters) are needed to develop
an ANN model for prediction of fl ow discharges. Note that it is
not only WQ parameters varying with LULC; fl ow would also
change as a function of LULC. Th is complicates the problem if
one wants to explore the impacts of various LULC change sce-
narios on WQ, for existing fl ow data cannot be used with those
LULC scenarios.
ReferencesAhearn, D.S., R.W. Sheibley, R.A. Dahlgren, M. Anderson, J. Johnson, and
K.W. Tate. 2005. Land use and land cover infl uence on water quality in the last free-fl owing river draining the western Sierra Nevada, California. J. Hydrol. 313:234–247.
Amiri, B.J., and K. Nakane. 2009. Comparative prediction of stream water to-tal nitrogen from land cover using artifi cial neural network and multiple linear regression approaches. Pol. J. Environ. Stud. 18:151–160.
Anctil, F.O., M. Filion, and J. Tournebize. 2009. A neural network experiment on the simulation of daily nitrate-nitrogen and suspended sediment fl uxes from a small agricultural catchment. Ecol. Modell. 220:879–887.
Arnheimer, B., and R. Liden. 2000. Nitrogen and phosphorus concentrations for agricultural catchments; infl uence of spatial and temporal variables. J. Hydrol. 227:140–159.
Arnold, C.L., Jr., and G.C. Gibbons. 1996. Impervious surface coverage: Th e emergence of a key environmental indicator. J. Am. Plann. Assoc. 62:243–259.
ASCE Task Committee. 2000. Artifi cial neural network in hydrology: I. Pre-liminary concepts. J. Hydrol. Eng. 5:115–123.
Basnyat, P., L.D. Teeter, K.M. Flynn, and B.G. Lockaby. 1999. Relationships between landscape characteristics and nonpoint source pollution inputs to coastal estuaries. Environ. Manage. 23:539–549.
Basnyat, P., L.D. Teeter, B.G. Lockaby, and K.M. Flynn. 2000. Land use char-acteristics and water quality: A methodology for valuing of forested buf-fers. Environ. Manage. 26:153–161.
Brey, T., A. Jarre-Teichmann, and O. Borlich. 1996. Artifi cial neural network versus multiple linear regression: Predicting P/B ratios from empirical data. Mar. Ecol. Prog. Ser. 140:251–256.
Chapra, S.C. 1997. Surface water-quality modeling. McGraw-Hill, New York.
Chau, K.W., C. Chuntian, and C.W. Li. 2002. Knowledge management sys-tem on fl ow and water quality modeling. Expert Syst. Appl. 22:321–330.
Dawson, C.W., and R.L. Wilby. 2001. Hydrologic modeling using artifi cial neural networks. Prog. Phys. Geogr. 25:80–108.
Di Luzio, L., R. Srinivasan, and J.G. Arnold. 2002. Integration of watershed tools and SWAT model into BASINS. J. Am. Water Resour. Assoc. 38:1127–1141.
Dogan, E., B. Sengorur, and R. Koklu. 2009. Modeling biochemical oxygen demand of the Melen River in Turkey using an artifi cial neural network technique. J. Environ. Manage. 90:1229–1235.
Fohrer, N., S. Haverkamp, K. Eckhardt, and H.G. Frede. 2001. Hydrologic response to land use changes on the catchment scale. Phys. Chem. Earth B 26:577–582.
Ha, H., and M.K. Stenstrom. 2003. Identifi cation of land use with water quali-ty data in stormwater using a neural network. Water Res. 37:4222–4230.
Haykin, S. 1999. Neural networks: A comprehensive foundation. Prentice Hall, New Jersey.
Helsel, D.R., and R.M. Hirsch. 2002. Statistical methods in water resources. Chapter A3. In Techniques of water-resources investigations of the United States Geological Survey. Book 4. Available at http://water.usgs.gov/pubs/twri/twri4a3 (verifi ed 22 Apr. 2010). U.S. Geological Survey, Denver, CO.
Hill, A.R. 1981. Stream phosphorus exports from watersheds with contrasting land uses in southern Ontario. Water Resour. Bull. 17:627–634.
Lockaby, B.G., D. Zhang, J. McDaniels, H. Tian, and S. Pan. 2005. Inter-disciplinary research at the urban-rural interface: Th e WestGA project. Urban Ecosyst. 8:7–21.
Maier, H.R., and G.C. Dandy. 2000. Neural network for the prediction and forecasting of water resources variables: A review of modeling issues and applications. Environ. Model. Softw. 15:101–124.
Moriasi, D.N., J.G. Arnold, M.W. Van Liew, R.L. Bingner, R.D. Harmel, and T.L. Veith. 2007. Model evaluation guidelines for systematic quantifi ca-tion of accuracy in watershed simulations. Trans. ASABE 50:885–900.
Muttil, N., and K.W. Chau. 2006. Neural network and genetic programming for modelling coastal algal blooms. Int. J. Environ. Pollut. 28:223–238.
Nash, J.E., and J.V. Sutcliff e. 1970. River fl ow forecasting through conceptual models: Part I. A discussion of principles. J. Hydrol. 10:282–290.
Neitsch, S.L., J.G. Arnold, J.R. Kiniry, J.R. Williams, and K.W. King. 2005. Soil and water assessment tool theoretical documentation. Version 2005. Temple, TX.
Osborne, L.L., and M.J. Wiley. 1988. Empirical relationships between land-use cover and stream water-quality in an agricultural watershed. J. Envi-ron. Manage. 26:9–27.
Qi, M., and G.P. Zhang. 2001. An investigation of model selection criteria for neural network time series forecasting. Eur. J. Oper. Res. 132:666–680.
Ren, L., and M.Z. Zhao. 2002. An optimal neural network and concrete strength modeling. Adv. Eng. Software 33:117–130.
Salas, J.D., M. Markus, and A.S. Tokar. 2000. Streamfl ow forecasting based on artifi cial neural networks. p. 23–51. In R.S. Govindaraju and A.R. Rao (ed.) Artifi cial neural networks in hydrology. Kluwer Academic, Dordre-cht, the Netherlands.
Schoonover, J.S. 2005. Hydrology, water quality, and channel morphology across an urban-rural land use gradient in the Georgia piedmont, USA. Ph.D. diss. Auburn Univ., Auburn, AL.
Schoonover, J.S., and B.G. Lockaby. 2006. Land cover impacts on stream nutrients and fecal coliform in the lower piedmont of west Georgia. J. Hydrol. 331:371–382.
Schoonover, J.S., B.G. Lockaby, and B. Helms. 2006. Eff ects of watershed land use on perennial streams of west Georgia, USA: 1. Infl uence of urban development on stream hydrology. J. Environ. Qual. 35:2123–2131.
Schoonover, J.S., B.G. Lockaby, and J.N. Shaw. 2007. Channel morphology and sediment origin in streams draining the Georgia piedmont. J. Hy-drol. 342:110–123.
Singh, K.P., A. Basant, A. Malik, and G. Jain. 2009. Artifi cial neural net-work modeling of the river water quality: A case study. Ecol. Modell. 220:888–895.
Sliva, L., and D.D. Williams. 2001. Buff er zone versus whole catchment ap-proaches to studying land use impact on river water quality. Water Res. 35:3462–3472.
Srivastava, P., J.N. McNair, and T.E. Johnson. 2006. Comparison of process-based and artifi cial neural network approaches for streamfl ow modeling in an agricultural watershed. J. Am. Water Resour. Assoc. 42:545–563.
Zhao, Z., Y. Zhang, and H. Liao. 2008. Design of ensemble neural net-work using the Akaike information criterion. Eng. Appl. Artif. Intell. 21:1182–1188.
Zurada, J.M. 1992. Introduction to artifi cial neural systems. PWS Publishing, Boston, MA.