Predicting Water Quality in Unmonitored Watersheds Using Artificial ...

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

−

−


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

−

− −


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.


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.


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.


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.


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.


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.


Fig. 3. Continued.


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.

Predicting Water Quality in Unmonitored Watersheds Using Artificial ...

Documents

Transcript of Predicting Water Quality in Unmonitored Watersheds Using Artificial ...