Sunteți pe pagina 1din 12

Adaptive Neurofuzzy Computing Technique

for Evapotranspiration Estimation


Özgür Kişi1 and Özgür Öztürk2
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

Abstract: The accuracy of an adaptive neurofuzzy computing technique in estimation of reference evapotranspiration 共ET0兲 is investi-
gated in this paper. The daily climatic data, solar radiation, air temperature, relative humidity, and wind speed from two stations, Pomona
and Santa Monica, in Los Angeles, Calif., are used as inputs to the neurofuzzy model to estimate ET0 obtained using the FAO-56
Penman–Monteith equation. In the first part of the study, a comparison is made between the estimates provided by the neurofuzzy model
and those of the following empirical models: The California Irrigation Management System, Penman, Hargreaves, and Ritchie. In this part
of the study, the empirical models are calibrated using the standard FAO-56 PM ET0 values. The estimates of the neurofuzzy technique
are also compared with those of the calibrated empirical models and artificial neural network 共ANN兲 technique. Mean-squared errors,
mean-absolute errors, and determination coefficient statistics are used as comparing criteria for the evaluation of the models’ perfor-
mances. The comparison results reveal that the neurofuzzy models could be employed successfully in modeling the ET0 process. In the
second part of the study, the potential of the neurofuzzy technique, ANN and the empirical methods in estimation ET0 using nearby station
data are investigated.
DOI: 10.1061/共ASCE兲0733-9437共2007兲133:4共368兲
CE Database subject headings: Fuzzy sets; Neural networks; Evapotranspiration; California; Computation; Models.

Introduction ships” 共Jensen et al. 1990兲. The Food and Agricultural Organiza-
tion of the United Nations 共FAO兲 assumed the ET definition from
Evaporation is the process whereby liquid water is converted to Smith et al. 共1997兲 and adopted the FAO Penman–Monteith as the
water vapor and removed from the evaporating surface. Transpi- standard equation for estimation of ET 共Allen et al. 1998; Naoum
ration consists of the vaporization of liquid water contained in and Tsanis 2003兲.
plant tissues and the vapor removal to the atmosphere. The com- Neural networks approaches have been successfully applied in
bination of the two separate processes whereby water is lost on a number of diverse fields, including water resources. In the hy-
the one hand from the soil surface by evaporation and on the drological context, recent experiments have reported that artificial
other hand from the crop by transpiration is referred to as evapo- neural networks 共ANN兲 may offer a promising alternative for
transpiration 共Allen et al. 1998兲. Knowledge of crop evapotrans- modeling hydrological variables 共e.g., rainfall–runoff, streamflow,
piration 共ET兲 is very important, because it allows optimization of suspended sediment兲 共Minnes and Hall 1996; Jain et al. 1999;
the irrigation water use in arid and semiarid regions where water ASCE Task Committee 2000a,b; Maier and Dandy 2000; Tayfur
shortage is a problem. The estimation of evapotranspiration is of 2002; Cancelliere et al. 2002; Supharatid 2003; Sudheer et al.
great importance for agricultural, hydrological, and climatic stud- 2003a; Cigizoglu 2003, 2004, 2005; Kumar et al. 2004; Kisi
ies, as it constitutes a major part of the hydrological cycle 2004a,b; Kisi 2005a,b; Kisi and Yıldırım 2005a,b; Cigizoglu and
共Sobrino et al. 2005兲. Kisi 2005, 2006; Sahoo and Ray 2006; Raghuwanshi et al. 2006;
Numerous methods have been proposed for modeling evapo- Chen and Adams 2006兲. However, the application of ANN to
transpiration as described by Brutsaert 共1982兲 and Jensen et al. evapotranspiration modeling is limited in the literature. Kumar et
共1990兲. In general, the combination of energy balance/ al. 共2002兲 used a multilayer perceptron 共MLP兲 with backpropaga-
aerodynamic equations “provides the most accurate results as a tion training algorithm for estimation of ET0. They used various
result of their foundation in physics and basis on rational relation- ANN architectures and found that the ANN gave accurate ET0
estimates. Sudheer et al. 共2003b兲 used radial basis ANN in mod-
1
Associate Professor, Engineering Faculty, Civil Engineering Dept., eling ET0 using limited climatic data. Trajkovic et al. 共2003兲 de-
Hydraulics Division, Erciyes Univ., 38039, Kayseri, Turkey. E-mail: veloped a radial basis ANN for forecasting ET0. Trajkovic 共2005兲
kisi@erciyes.edu.tr used temperature-based radial basis ANN for modeling FAO-56
2
Research Assistant, Engineering Faculty, Civil Engineering Dept., PM ET0. He compared the ANN results with those of the Har-
Hydraulics Division, Erciyes Univ., 38039, Kayseri, Turkey. E-mail: greaves, Thornthwaite, and reduced PM methods and concluded
ozturko@erciyes.edu.tr that the radial basis ANN generally performs better than the oth-
Note. Discussion open until January 1, 2008. Separate discussions ers in modeling the ET0 process. Kisi 共2006兲 investigated the
must be submitted for individual papers. To extend the closing date by
estimation of ET0 using MLP method and he compared ANN test
one month, a written request must be filed with the ASCE Managing
Editor. The manuscript for this paper was submitted for review and pos- results with those of the Penman and Hargreaves empirical mod-
sible publication on ; approved on May 3, 2007. This paper is part of the els. He found that the MLP model gave much better results than
Journal of Irrigation and Drainage Engineering, Vol. 133, No. 4, the empirical models. To the knowledge of the writers, no study
August 1, 2007. ©ASCE, ISSN 0733-9437/2007/4-368–379/$25.00. has been carried out to utilize the input–output mapping capabil-

368 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007

J. Irrig. Drain Eng. 2007.133:368-379.


Rule 1: If x is A1 and y is B1, then f 1 = p1x + q1y + r1 共1兲

Rule 2: If x is A2 and y is B2, then f 2 = p2x + q2y + r2 共2兲


The resulting Sugeno fuzzy reasoning system is shown in Fig.
1. Here, the output z⫽weighted average of the individual rule
outputs and is itself a crisp value. The corresponding equivalent
ANFIS architecture is shown in Fig. 2. Nodes at the same layer
have similar functions. The node function is described next. The
output of the ith node in layer l is denoted as Ol,i.
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

Layer 1
Every node i in this layer is an adaptive node with node function

Fig. 1. Two inputs first-order Sugeno fuzzy model with two rules Ol,i = ␮Ai共x兲 for i = 1,2
or

ity of neurofuzzy computing technique in ET0 modeling. This


Ol,i = ␮Bi−2共y兲 for i = 3,4
provided an impetus for the present investigation.
The purpose of this study is to investigate the accuracy of an where x 共or y兲⫽input to the ith node; and Ai 共or Bi−2兲 is a linguis-
adaptive neurofuzzy computing technique for modeling of ET0. tic label 共such as “low” or “high”兲 associated with this node. In
The ET0 values are obtained using the standard FAO-56 Penman– words, Ol,i is the membership grade of a fuzzy set A 共=A1, A2, B1,
Monteith equation. The performance of the neurofuzzy model is or B2兲 and it specifies the degree to which the given input x 共or y兲
compared with those of the ANN technique and commonly used satisfies the quantifier A. The membership functions for A and B
CIMIS Penman, Hargreaves, and Ritchie empirical methods. are generally described by generalized bell functions, e.g.

1
␮Ai共x兲 = 共3兲
Adaptive Neurofuzzy Inference System 1 + 关共x − ci兲/ai兴2bi
The adaptive neurofuzzy inference system 共ANFIS兲, first intro- where 兵ai, bi, ci其 is the parameter set. As the values of these
duced by Jang 共1993兲, is a universal approximator and as such is parameters change, the bell-shaped function varies accordingly,
capable of approximating any real continuous function on a com- thus exhibiting various forms of membership functions on
pact set to any degree of accuracy 共Jang et al. 1997兲. ANFIS is linguistic label Ai. In fact, any continuous and piecewise differ-
functionally equivalent to fuzzy inference systems 共Jang et al. entiable functions, such as commonly used triangular-shaped
1997兲. Specifically the ANFIS system of interest here is function- membership functions, are also qualified candidates for node
ally equivalent to the Sugeno first-order fuzzy model 共Jang et al. functions in this layer 共Jang 1993兲. Parameters in this layer are
1997; Drake 2000兲. Below, the hybrid learning algorithm, which referred to as premise parameters. The outputs of this layer are the
combines gradient descent and the least-squares method, is membership values of the premise part.
introduced.
As a simple example we assume a fuzzy inference system with
Layer 2
two inputs x and y and one output z. The first-order Sugeno fuzzy
model, a typical rule set with two fuzzy If–Then rules can be This layer consists of the nodes labeled ⌸, which multiply incom-
expressed as ing signals and sending the product out. For instance

Fig. 2. Equivalent ANFIS architecture

JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007 / 369

J. Irrig. Drain Eng. 2007.133:368-379.


O2,i = wi = ␮Ai共x兲␮Bi共y兲, i = 1,2 共4兲 fixed, the over all output can be given as a linear combination of
the consequent parameters. The output f can be written as
Each node output represents the firing strength of a rule.
w1 w2
f= f1 + f2
Layer 3 w1 + w2 w1 + w2

In this layer, the nodes labeled N calculate the ratio of the ith =w̄1共p1x + q1y + r1兲 + w̄2共p2x + q2y + r2兲
rule’s firing strength to the sum of all rules’ firing strengths
=共w̄1x兲p1 + 共w̄1y兲q1 + 共w̄1兲r1 + 共w̄2x兲p2
wi
O3,i = w̄i = i = 1,2 共5兲
w1 + w2 + 共w̄2y兲q2 + 共w̄2兲r2 共9兲
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

The outputs of this layer are called normalized firing strengths. which is linear in the consequent parameters p1, q1, r1, p2, q2,
and r2. Then we have S⫽set of total parameters; S1⫽set of
premise 共nonlinear兲 parameters; and S2⫽set of consequent 共lin-
Layer 4 ear兲 parameters.
Given values of S1, we can plug P training data into the Eq.
This layer’s nodes are adaptive with node functions 共9兲 and obtain the matrix equation
A␪ = y 共10兲
O4,i = w̄i f i = w̄i共pix + qiy + ri兲 共6兲
where ␪⫽unknown vector whose elements are parameters in S2,
where w̄i⫽output of Layer 3; and 兵pi , qi , ri其⫽parameter set. Pa- the set of consequent 共linear兲 parameters.
rameters of this layer are referred to as consequent parameters. Then the set S2 of consequent parameters can be identified
with the standard least-squares estimator 共LSE兲
␪* = 共ATA兲−1ATy 共11兲
Layer 5
where AT⫽transpose of A; and 共ATA兲−1AT⫽pseudoinverse of A if
This layer’s single fixed node labeled 兺 computes the final output ATA is nonsingular. The recursive least-square estimator 共RLS兲
as the summation of all incoming signals can also be used to calculate ␪* 共Jang 1993兲.
Backpropagation and LSE can now be combined to update the
兺wi f i parameters of the adaptive network. For hybrid learning applied
in batch mode 共off-line learning兲, each epoch is composed of a

i
O5,i = w̄i f i = 共7兲
i=1 兺wi forward pass and a backward pass as summarized in Table 1.
i In the forward pass of the hybrid learning algorithm, node
outputs go forward until the final layer 共Layer 4 in Fig. 2兲 and the
Thus, an adaptive network is functionally equivalent to a Sugeno
consequent parameters are identified by the least-squares method.
first-order fuzzy inference system.
In the backward pass, the error signals propagate backward and
The learning rule specifies how the premise parameters 共see
the premise parameters are updated by gradient descent. More
Layer 1兲 and consequent parameters 共see Layer 4兲 should be up-
information about ANFIS can be found in Jang 共1993兲.
dated to minimize a prescribed error measure E. The error mea-
sure is a mathematical expression that measures the difference
between the networks actual output and the desired output, such
as the squared error. The steepest descent method is used as the
Artificial Neural Networks
basic learning rule of the adaptive network. In this method the
Artificial neural networks 共ANNs兲 are based on the present un-
gradient is derived by repeated application of the chain rule. Cal-
derstanding of biological nervous system, though much of the
culation of the gradient in a network structure requires use of the
biological detail is neglected. ANNs are massively parallel sys-
ordered derivative denoted as ⳵+ as opposed to the ordinary partial
tems composed of many processing elements connected by links
derivative ⳵. This technique is called the backpropagation rule
of variable weights. Of the many ANN paradigms, the backpropa-
共Jang 1993; Drake 2000兲. The core of this learning rule involves
gation network is by far the most popular 共Lippman 1987兲. The
how to recursively obtain a gradient vector in which each element
network consists of layers of parallel processing elements, called
is defined as the derivative of an error measure with respect to a
neurons. Each layer is fully connected to the backflow layer by
parameter 共Haykin 1998兲. The update formula for simple steepest
interconnection which is fully connected to the proceeding layer
descent for the generic parameter ⬀ is
by interconnection strengths, or weights, W. Fig. 3 illustrates a
three-layer neural network consisting of layers i, j, and k, with the
⳵ +E interconnection weights Wij and W jk between layers of neurons.
⌬␣ = − ␩ 共8兲 Initial estimated weight values are progressively corrected during
⳵␣
a training process that compares predicted outputs to known out-
where ␩⫽learning rate. puts, and backpropagates any errors 共from right to left in Fig. 3兲
While the backpropagation learning rule can be used to iden- to determine the appropriate weight adjustments necessary to
tify the parameters in an adaptive network, this method is slow to minimize the errors. The methodology used here for adjusting the
converge. The hybrid learning algorithm 共Jang 1993兲, which com- weights is called “backpropagation,” and is based on the “gener-
bines backpropagation and the least-squares method can be used alized delta rule,” as presented by Rumelhart et al. 共1986兲.
to rapidly train and adapt the equivalent fuzzy inference system. It Throughout all ANN simulations the adaptive learning rates were
can be seen from the Fig. 2 that if the premise parameters are used for increasing the convergence velocity. For each epoch, if

370 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007

J. Irrig. Drain Eng. 2007.133:368-379.


Table 1. Hybrid Learning Procedure for ANFIS 共for One Epoch兲
Layer Procedure
共a兲 Forward pass
1 Set premise parameters of the membership functions, arbitrarily 关e.g., 共ai, bi, ci兲兴 and compute the membership values
关e.g., ␮Ai共x兲, ␮Bi共y兲兴
2 Compute the firing strength of the ith rule by multiplying the membership values obtained in Layer 1 关e.g., wi = ␮Ai共x兲 ⫻ ␮Bi共y兲兴
3 Compute the normalized firing strengths by calculating the ratio of the ith rule’s firing strength to the sum of all rules’ firing
strengths 共e.g., w̄i = wi / 兺i=1 2
w i兲
4 Obtain the weighted node function by multiplying the normalized firing strengths with the node function for each rule
关e.g., w̄i f i = w̄i共pix + qiy + ri兲, herepi, qi, and ri⫽consequent parameters兴.
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

5 Obtain the matrix equation A · ␪ = y using the Eq. 共9兲, where ␪⫽unknown vector comprising the consequent parameters.
Determine the consequent parameters with the standard LSE given in the Eq. 共11兲.
共b兲 Backward pass
1 Determine the premise parameters using the update formula of simple gradient descent given by Eq. 共8兲.

performance decreases toward the goal, then the learning rate is coefficient, skewness, minimum, and maximum, respectively. The
increased by the factor learning increment. If performance in- wind speed shows a skewed distribution for both stations 共see Csx
creases, the learning rate is adjusted by the factor learning decre- values for entire data in Table 2兲. As can be seen from the corre-
ment. The sigmoid and linear activation functions are used for the lation coefficients between the Rs and ET0 in Table 2, the solar
hidden and output node共s兲, respectively. The hidden layer node radiation is closely correlated with evapotranspiration for both
numbers of each model were determined after trying various net- stations. The air temperature seems to be the second best param-
work structures since there is no theory yet to tell how many eter correlated with ET0. In the training ET0 data of the Pomona
hidden units are needed to approximate any given function. The Station, minimum and maximum values fall in the ranges 0–
ANN networks training were stopped after 50,000 epochs. 7.55 mm. However, the maximum of the testing flow data of the
Pomona Station is 8.51 mm, which is higher than the correspond-
ing training sets’ value. The other station also has the same case.
Case Study This may cause some extrapolation difficulties in prediction of
high flow values.
The daily climatic data of two automated weather stations,
Pomona Station 共latitude 34° 03⬘ N, longitude 117° 48⬘ W兲 and
Santa Monica Station 共latitude 34° 02⬘ N, longitude 118° 28⬘ W兲 Application and Results
operated by the California Irrigation Management Information
System 共CIMIS兲 are used in the study. The elevations are 222 and
First, the ET0 values for the Pomona and Santa Monica Stations
104 m for the Pomona and Santa Monica Stations, respectively.
were calculated using the FAO-56 PM method as described in
The location of the stations is shown in Fig. 4. The CIMIS is a
Allen et al. 共1998兲
program developed in 1982 by the California Department of
Water Resource and the University of California at Davis to assist
California’s irrigators manage their water resources efficiently.
The CIMIS manages a network of over 120 automated weather
stations in the state of California. The following weather data
used in the present study are measured at the CIMIS weather
stations. The total incoming solar radiation is measured using
pyranometers at height of 2.0 m above the ground. Air tempera-
ture is measured at a height of 1.5 m above the ground using a
thermistor. Relative humidity is the ratio of the actual amount of
water vapor, to the amount the atmosphere can potentially hold at
the given air temperature. It is expressed as a percentage. The
relative humidity sensor is sheltered in the same enclosure as the
air temperature sensor at 1.5 m above the ground. Wind speed is
measured using three-cup anemometers at 2.0 m above the
ground. These measured daily climatic data and the ET0 values
calculated using the CIMIS Penman were downloaded from the
CIMIS web server 共http://www.ipm.ucdavis.edu/WEATHER/
wxretrieve.html兲.
The data sample consisted of daily records of four years
共2001–2004兲 of solar radiation, air temperature, relative humidity,
and wind speed. For each station, the first three years 共2001–
2003兲 data were used to train the ANN models and the remaining
data were used for testing. The daily statistical parameters of the
climatic data are given in Table 2. In Table 2, the xmean, Sx, Cv,
Csx, xmin, and xmax denote the mean, standard deviation, variation Fig. 3. An ANN architecture used for ET0 estimation

JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007 / 371

J. Irrig. Drain Eng. 2007.133:368-379.


Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

Fig. 4. Location of the 共1兲 Pomona and 共2兲 Santa Monica Stations in Los Angeles

900 stations are estimated, separately and the ANFIS estimates are
0.408⌬共Rn − G兲 + ␥ U2共ea − ed兲 compared with those of the ANN and empirical models. In the
T + 273
ET0 = 共12兲 second application, the ET0 data of one station are estimated
⌬ + ␥共1 + 0.34U2兲
using the climatic data from the other station and the ANFIS test
where ET0⫽reference evapotranspiration 共mm day−1兲; ⌬⫽slope results are compared with those of the ANN and multilinear re-
of the saturation vapor pressure function 共kPa ° C−1兲; Rn⫽net ra- gression 共MLR兲.
diation 共MJ m−2 day−1兲; G⫽soil heat flux density 共MJ m−2 day−1兲;
␥⫽psychometric constant 共kPa ° C−1兲; T⫽mean air temperature
共 0C−1兲; U2⫽average 24 h wind speed at 2 m height 共m s−1兲;
ea⫽saturation vapor pressure 共kPa兲; and ed⫽actual vapor pressure Estimation of ET0 Data of Pomona Station
共kPa兲.
Then, these calculated FAO-56 PM ET0 values were used for The ANFIS and ANN techniques are first used for the Pomona
the calibration of ANFIS and ANN models. Two different pro- Station. The weather parameters considered in the study are solar
gram codes were written in MATLAB language for the neuro- radiation 共Rs兲, air temperature 共T兲, relative humidity 共RH兲, and
fuzzy and ANN model simulations. Different ANFIS and ANN wind speed 共U2兲. Two different models were developed using the
architectures were tried using these codes and the appropriate neurofuzzy and ANN techniques. The ANFIS2 and ANN2 models
models’ structures were determined. The ANFIS and ANN mod- comprising two inputs, T and Rs, were developed for the valid
els were tested and the results were compared with those of the comparison with the Hargreaves and Ritchie models with two
CIMIS Penman, Hargreaves, and Ritchie empirical methods. The input parameters. The MSE, MAE and R2 statistics of each
mean-squared errors 共MSE兲, mean-absolute errors 共MAE兲, and ANFIS model in test period are compared with the ANN and
determination coefficient 共R2兲 statistics were used as the compar- empirical models in Table 3. The final architectures of the ANFIS
ing criteria. The R2 measures the degree to which two variables and ANN models that were found after many trials for the
are linearly related. MSE and MAE provide different types of Pomona Station are also provided in Table 3. Fuzzy membership
information about the predictive capabilities of the model. The functions can take many forms, but simple straight line 共triangu-
MSE measures the goodness of fit relevant to high ET0 values, lar兲 and Gaussian functions are often preferred 共Russel and
whereas the MAE yields a more balanced perspective of the Campbell 1996; Vernieuwe et al. 2005兲. In the present study, the
goodness of fit at moderate ET0 共Karunanithi et al. 1994兲. The triangular and Gaussian membership functions were tried. The
MSE and MAE are defined as Gaussian functions were found to be more adequate for the ET0
estimation. In each application, different number of membership
N
1 functions was tried and the best one that gives the minimum MSE
MSE = 兺
N i=1
共ETiobserved − ETipredicted兲2 共13兲 was selected. In Table 3, the ANFIS2共5,2兲 denotes the adaptive
neurofuzzy model comprising 5 and 2 membership functions for
the inputs Rs and T, respectively. The consequent parameters of
N
1 the ANFIS2 model are provided in Table 5. From Table 5, the bias
MAE = 兺
N i=1
兩ETiobserved − ETipredicted兩100 共14兲 term 共r兲 seems to be mostly effective in ET0 estimation. In Table
3, the ANN2 共2,4,1兲 denotes and ANN model comprising 2 inputs
in which N denotes the number of test data sets. Rs and T, 4 hidden, and 1 output nodes.
Next, the potential of the ANFIS technique is tested for two The CIMIS Penman 共Snyder and Pruitt 1985兲, Hargreaves, and
different applications. In the first application, the ET0 data of two Ritchie methods are considered for the comparison in the present

372 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007

J. Irrig. Drain Eng. 2007.133:368-379.


Table 2. Daily Statistical Parameters of Each Data Set
Cv Correlation
Data set Variable 共Sx / xmean兲 xmean Sx C sx xmin xmax with ET0
共a兲 Pomona
Training Rs 17.1 7.09 0.41 −0.06 1.20 29.7 0.938
T 16.9 5.00 0.30 −0.02 5.00 30.0 0.759
RH 71.2 11.9 0.17 −0.88 20.0 100 −0.399
U2 1.54 0.36 0.23 0.46 0.40 3.10 0.350
ET0 3.58 1.78 0.50 0.01 0.00 7.55 1.000
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

Test Rs 1.62 0.34 0.21 0.25 0.90 2.70 0.952


T 17.9 7.52 0.42 −0.18 1.20 29.9 0.841
RH 17.1 4.85 0.28 −0.02 7.20 27.5 −0.582
U2 17.5 5.82 0.33 −0.12 4.90 27.5 0.332
ET0 3.86 2.00 0.52 0.09 0.00 8.51 1.000
Entire Rs 13.2 9.09 0.69 0.10 0.90 29.7 0.942
T 17.2 5.75 0.33 −0.01 1.20 30.0 0.778
RH 57.6 25.7 0.45 −0.73 7.20 100 −0.452
U2 5.54 7.52 1.36 1.60 0.40 27.5 0.349
ET0 3.65 1.84 0.50 0.06 0.00 8.51 1.000
共b兲 Santa Monica
Training Rs 18.9 7.47 0.40 −0.08 1.10 33.6 0.956
T 16.4 3.34 0.20 0.09 7.20 26.9 0.489
RH 17.6 4.61 0.26 −0.15 5.70 26.9 −0.303
U2 1.72 0.46 0.27 1.57 0.40 5.80 0.255
ET0 3.28 1.50 0.46 −0.03 0.00 6.86 1.000
Test Rs 17.7 7.01 0.40 −0.21 0.90 30.1 0.898
T 16.7 3.32 0.20 0.17 8.60 27.8 0.695
RH 17.2 4.73 0.28 −0.24 4.90 26.0 −0.332
U2 1.81 0.38 0.21 0.63 0.40 3.60 0.205
ET0 3.16 1.54 0.49 −0.05 0.00 7.11 1.000
Entire Rs 18.6 7.37 0.40 −0.10 0.90 33.6 0.941
T 16.5 3.34 0.20 0.11 7.20 27.8 0.539
RH 17.5 4.64 0.27 −0.18 4.90 26.9 −0.312
U2 1.74 0.44 0.25 1.37 0.40 5.80 0.238
ET0 3.25 1.51 0.46 −0.04 0.00 7.11 1.000

study. The CIMIS Penman equation employs the modified Pen- midnight兲 are then summed to produce estimates of daily ET0.
man equation 共Pruitt and Doorenbos 1977兲 with a wind function The hourly PM equation that CIMIS uses to estimate hourly Pen-
that was developed at the University of California at Davis. The man method 共PM兲 ET0 is the Food and Agricultural Organiza-
method uses hourly average weather data as an input to calculate tion’s version that is described in Irrigation and Drainage Paper
hourly ET0. The 24 hourly ET0 values for the day 共midnight to No. 56 共Allen et al. 1998兲. The CIMIS Penman equation is also

Table 3. Performance Statistics of the Models in the Test Period: Pomona Station
MSE MAE
Model Model input 共mm2 day2兲 共mm day−1兲 R2
ANFIS 1 共2,2,2,2兲 Rs, T, RH, U2 0.041 0.147 0.992
ANFIS2共5,2兲 R s, T 0.065 0.186 0.984
ANN1共4,5,1兲 Rs, T, RH, U2 0.098 0.225 0.976
ANN2共2,4,1兲 R s, T 0.124 0.255 0.970
CIMIS Penman Rs, T, RH, U2 0.439 0.361 0.970
Hargreaves Rs,T 27.43 4.62 0.981
Ritchie R s, T 1.381 0.756 0.985
C_CIMIS Penman Rs, T, RH, U2 0.109 0.253 0.970
C_Hargreaves R s, T 0.078 0.216 0.981
C_Ritchie R s, T 0.071 0.205 0.985

JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007 / 373

J. Irrig. Drain Eng. 2007.133:368-379.


Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

Fig. 5. FAO-56 PM and estimated ET0 values of the Pomona Station in the test period

described in detail in Hidelgo et al. 共2005兲, 共see CIMIS website:


http://wwwcimis.water.ca.gov/cimis/infoEtoCimisEquation.jsp兲 ET0 = 0.0023Ra 冉 Tmax + Tmin
2
+ 17.8 冊冑
Tmax − Tmin 共16兲

ET0 = 冉 冊 冉

⌬+␥
Rn + 1 −

⌬+␥

共ea − ed兲f U 共15兲
where ET0⫽reference evapotranspiration 共mm day−1兲; Tmax and
Tmin⫽maximum and minimum temperature 共0C兲; and
Ra⫽extraterrestrial radiation 共mm day−1兲.
The Ritchie method, as described by Jones and Ritchie 共1990兲
where ET0⫽mean hourly reference evapotranspiration
共mm day−1兲; ⌬⫽slope of the saturation vapor pressure function
共kPa ° C−1兲; Rn⫽mean hourly net radiation 共W m−2兲;
␥⫽psychometric constant 共kPa ° C−1兲; ea⫽saturation vapor pres- ET0 = ␣1 . 关3.87 ⫻ 10−3 . Rs . 共0.6Tmax + 0.4Tmin + 29兲兴 共17兲
sure 共kPa兲; ed⫽actual vapour pressure 共kPa兲; and the f U⫽wind
function 共m s−1兲. Daily ET0 equals to the sum of 24 h ET0 共mm兲.
The Hargreaves empirical formula is one of the simplest equa- where ET0⫽reference evapotranspiration 共mm day−1兲; Tmax and
tions used to estimate ET0. It is expressed as 共Hargreaves and Tmin⫽maximum and minimum temperature 共0C兲; and Rs⫽solar
Samani 1985兲 radiation 共MJ m−2 day−1兲. When

374 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007

J. Irrig. Drain Eng. 2007.133:368-379.


Table 4. Performance Statistics of the Models in the Test Period: Santa Monica Station
MSE MAE
Model Model input 共mm2 day2兲 共mm day−1兲 R2
ANFIS 1 共2,2,2,5兲 Rs, T, RH, U2 0.005 0.045 0.998
ANFIS2共5,2兲 R s, T 0.290 0.388 0.860
ANN1共4,8,1兲 Rs, T, RH, U2 0.066 0.179 0.969
ANN2共2,6,1兲 R s, T 0.400 0.432 0.814
CIMIS Penman Rs, T, RH, U2 0.410 0.392 0.830
Hargreaves R s, T 23.57 4.35 0.818
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

Ritchie R s, T 0.641 0.650 0.780


C_CIMIS Penman Rs, T, RH, U2 0.371 0.415 0.830
C_Hargreaves R s, T 0.381 0.471 0.818
C_Ritchie R s, T 0.456 0.474 0.780

5 ⬍ Tmax ⱕ 35 ° C ␣1 = 1.1 greaves’s performance 共see Table 3 and Fig. 5兲. The estimation of
total ET0 was also considered for comparison due to its impor-
Tmax ⬎ 35 ° C ␣1 = 1.1 + 0.05共Tmax − 35兲 tance in irrigation management. The ANFIS2, ANN2, C_Har-
greaves, and C_Ritchie computed the total FAO-56 PM ET0 of
1 , 288.80 mm as 1,271, 1,272, 1,273, and 1,283 mm with under-
Tmax ⬍ 5 ° C ␣1 = 0.01 exp关0.18共Tmax + 20兲兴 共18兲
estimations of 1.4, 1.3, 1.2, and 0.4% while the ANFIS1, ANN1,
The temperature-based methods mostly underestimate or over- CIMIS Penman, Hargreaves, Ritchie, and C_CIMIS Penman
estimate ET0 obtained by the FAO-56 PM method. In those cases, methods resulted in 1,317, 1,291, 1,411, 3,443, 1,377 and
Allen et al. 共1994兲 recommended that empirical methods be cali- 1,295 mm, with overestimations of 2.2, 0.2, 9.5, 167, 6.8, and
brated using the standard PM method. ET0 is calculated as 0.5%, respectively. The ANN1 model had the closest estimate.
The C_Ritchie estimate was almost the same as the total
ET0 = a + bETeq 共19兲 C_CIMIS Penman ET0 value and they ranked as the second best.
where ET0⫽grass reference ET defined by the FAO-56 PM equa- The calibration significantly increases the performance of the em-
tion; ETeq⫽ET estimated by the temperature-based methods; and pirical models in total ET0 estimation.
a and b⫽calibration factors, respectively.
The data used for the training ANFIS and ANN models were Estimation of ET0 Data of Santa Monica Station
used for calibration of temperature-based methods. The CIMIS
Penman method was also calibrated. Thus, the calibrated empiri- For the Santa Monica Station, the performances of the ANFIS,
cal methods were also included for the comparison in the study. ANN, and empirical models in the test period are summarized in
The ANFIS and ANN models are compared with the empirical Table 4. Here, also the trend of the results is same as that of the
CIMIS Penman, Hargreaves, and Ritchie methods, and their cali- Pomona Station. The best performance criteria are obtained by the
brated versions denoted as C_CIMIS Penman, C_ Hargreaves, ANFIS1 model whose inputs are the Rs, T, RH, and U2. Results
and C_Ritchie in respect of the MSE, MAE, and R2 statistics for from Table 4 clearly indicate that the ANFIS1 model provides the
the Pomona Station in Table 3. The input variables used for each most accurate ET0 estimates as found for the Pomona Station. The
model are also given in Table 3. The ANFIS2, ANN2, Har- ANFIS2 model also performs better than the empirical models.
greaves, and Ritchie models use the same input variables. It can The consequent parameters of the ANFIS2 model are given in
be seen from the Table 3 that the neurofuzzy models outperform Table 5. Here, also the bias term 共r兲 is the most effective in ET0
all other models in terms of various performance criteria. The estimation. Unlike the Pomona Station, The CIMIS Penman and
C_ Hargreaves and C_Ritchie models perform better than the
ANN models. It can be obviously seen that the calibration of the
empirical models significantly increases the estimation accuracy. Table 5. Consequent Parameters of the ANFIS2 Models
The ANFIS2, Hargreaves, Ritchie, C_Hargreaves, and C_Ritchie
models are rather simple and consider only T and Rs data. Com- Station
pared with the empirical models, the ANFIS2 model slightly per- Pomona Santa Monica
forms better than the C_Hargreaves and C_Ritchie models. The
comparison of the models that used minimum data, i.e., ANFIS2, Rule p q r p q r
ANN2, Hargreaves, Ritchie, C_Hargreaves, and C_Ritchie 1 0.0035 0.3235 −12.27 0.0132 −0.111 5.004
model, reveals that the results are good for the neurofuzzy and 2 0.0009 0.8973 −65.12 0.0662 −0.648 39.42
calibrated empirical models. Among the empirical methods, the 3 0.0063 0.0415 −2.059 −0.0016 0.034 1.300
C_Hargreaves and C_Ritchie models have similar results and they 4 0.0185 0.0687 −6.804 0.0413 0.302 −28.12
perform better than the CIMIS Penman. The ET0 estimates of 5 0.0087 0.0958 −4.837 0.0009 0.094 1.384
each model are represented in Fig. 5 in the form of a scatterplot. 6 0.0125 0.1617 −14.17 0.0186 0.199 −19.35
It is seen from the scatterplots that the ANFIS estimates are closer 7 0.0069 0.1141 −4.865 0.0036 0.036 0.574
to the corresponding FAO-56 PM ET0 values than those of the
8 0.0009 0.1652 −12.91 0.0152 0.157 −15.39
other models. The Hargreaves significantly overestimates ET0 ob-
9 0.0012 0.4439 −16.83 0.0037 0.731 −32.92
tained by the FAO-56 PM method as reported by the Allen et al.
10 0.0109 0.4107 −34.18 0.0243 0.949 −82.87
共1994兲. However, the calibration considerably increased the Har-

JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007 / 375

J. Irrig. Drain Eng. 2007.133:368-379.


Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

Fig. 6. FAO-56 PM and estimated ET0 values of the Santa Monica Station in the test period

its calibrated version show better accuracy than the Ritchie and with underestimations of 0.2, 0.1, 3.4, and 2.1%, the ANFIS2,
C_Ritchie methods, respectively. The extended match of FAO-56 Hargreaves, Ritchie, C_CIMIS Penman, C_ Hargreaves, and
PM ET0, and the estimates of each model are plotted in Fig. 6, in C_Ritchie models resulted in 1,191, 2,772, 1,301, 1,213, 1,206
term of scatter diagram type of comparison. As seen from Fig. 6, and 1,194 mm, with overestimations of 0.9, 135, 10, 2.7, 2.1, and
the ANFIS1 has less scattered estimates than the other models. 1.1%, respectively. The ANFIS estimates are closest to the ob-
There is also considerable overestimation seen for the Hargreaves served one. Here, also the calibration significantly increased the
model 共also see the errors in Table 4兲. For this station, the perfor- performance of the temperature-based models, Hargreaves
mance differences between the two- and four-parameter models and Ritchie, in total ET0 estimation as found in the preceding
are much more than those of the Pomona Station. This implies application.
that the relative humidity and wind speed data are more effective
for estimating ET0 estimates for the Santa Monica Station than
Estimation of ET0 Data of Pomona Station Using Data
those of the Pomona Station. This may be due to the fact that the
of Santa Monica
Santa Monica Station is located in a coastal area 共see Fig. 4兲.
While the ANFIS1, ANN1, ANN2, and CIMIS Penman estimated The ET0 estimation using nearby station data is an important issue
the total FAO-56 PM ET0 as 1,179, 1,179, 1,140 and 1,155 mm, since the data of some stations are missing. In this case the me-
compared to the computed FAO-56 PM ET0 value of 1,181 mm, teorological data from the nearby station can be used to estimate

376 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007

J. Irrig. Drain Eng. 2007.133:368-379.


Table 6. Performance Statistics of the Models in ET0 Estimation of Pomona Station Using Data of nearly Santa Monica Station
MSE MAE
Model Model input 共mm2 day2兲 共mm day−1兲 R2
ANFIS 1 共3,2,3,5兲 Rs, T, RH, U2 0.615 0.592 0.871
ANFIS2共2,4兲 R s, T 0.666 0.630 0.861
ANN1共4,8,1兲 Rs, T, RH, U2 0.682 0.643 0.847
ANN2共2,5,1兲 R s, T 0.712 0.695 0.843
MLR 1 Rs, T, RH, U2 1.051 0.826 0.828
MLR 2 Rs,T 1.129 0.850 0.811
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

missing ET0 values. To solve this problem, the regression tech- ET0 estimation. The ANFIS1, ANFIS2, ANN1, ANN2, MLR1,
niques are frequently used. This part of study focused on the and MLR2 estimated the total FAO-56 PM ET0 value of
investigation of neurofuzzy technique performances to solve this 1,289 mm as 1,163, 1,166, 1,210, 1,201, 1,099, and 1,133 mm,
problem. The climatic data of Santa Monica Station are used to with underestimations of 9.8, 9.6, 6.1, 6.8, 14.7, and 12.1%, re-
estimate FAO-56 PM ET0 of Pomona Station. In this application spectively. The ANN models have similar estimates and their ac-
also three years 共2001–2003兲 data are used for the calibration of curacies seems to be better than the other models in estimation of
ANFIS models and remaining one year data are used for testing. total ET0 value of the Pomona using the data of nearby station.
The optimum number of membership functions was given in The neurofuzzy models also have much better accuracy than the
Table 6 for the neurofuzzy models. The ANN and multilinear MLR. The results imply that the ANFIS models may be used
regression are also applied to the same data. The performance instead of ANN and MLR in cross-station applications.
statistics of the models in test period are provided in Table 6. The
ANFIS1 model has the lowest MSE 共0.615 mm2 day−2兲 and MAE
共0.592 mm day−1兲 and the highest R2 statistic 共0.871兲. The
ANFIS2 model is ranked as the second best. The ANN models Concluding Remarks
also perform much better than the MLR. The FAO-56 PM ET0
and estimated ET0 of the Pomona Station using the data of Santa The accuracy of an adaptive neurofuzzy technique in estimation
Monica Station are shown in Fig. 7. As seen from the fit line of reference evapotranspiration using climatic variables was pre-
equations 共assume that the equation is y = aox + a1兲 in the scatter- sented in this study. The study indicated that modeling of daily
plots that the ao and a1 coefficients for the ANFIS1 model are, reference evapotranspiration is possible through the use of ANFIS
respectively, closer to the 1 and 0 with a higher R2 value of 0.871 technique. The neurofuzzy model 共ANFIS1兲 whose inputs are the
than those of the other models. This confirms the MSE and MAE Rs, T, RH, and U2 was found to perform better than the other
statistics evaluated in Table 6. Four-parameter models 共ANFIS1, neurofuzzy, ANN and empirical models in estimation of FAO-56
ANN1, and MLR1兲 perform better than the corresponding two- PM ET0. However, in some areas 共e.g., developing countries兲 the
parameter models 共ANFIS2, ANN2, and MLR2兲. This indicates available data may be the solar radiation, Rs, and air temperature,
that all these variables 共Rs, T, RH, and U2兲 are needed for better T, due to the difficulty in obtaining the data of other two param-

Fig. 7. FAO-56 PM and estimated ET0 values of Pomona Station using data of the Santa Monica Station in the test period

JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007 / 377

J. Irrig. Drain Eng. 2007.133:368-379.


eters, relative humidity and wind speed. Therefore, ANFIS2 and Cigizoglu, H. K. 共2004兲. “Estimation and forecasting of daily suspended
ANN2 models containing only two inputs, Rs and T were devel- sediment data by multi layer perceptrons.” Adv. Water Resour., 27,
oped and compared with the two-parameter Hargreaves and 185–195.
Ritchie models. The comparison results revealed that the ANFIS2 Cigizoglu, H. K. 共2005兲. “Application of the generalized regression neu-
model was superior to the two-parameter empirical models ral networks to intermittent flow forecasting and estimation.” J. Hy-
drol. Eng., 10共4兲, 336–341.
共temperature-based models兲. It was found that the calibration sug-
Cigizoglu, H. K., and Kisi, O. 共2005兲. “Flow prediction by three back
gested in Allen et al. 共1994兲 is very relevant now, which greatly propagation techniques using k-fold partitioning of neural network
improved the underestimation of the Hargreaves formula, in esti- training data.” Nord. Hydrol., 36共1兲, 49–64.
mation of ET0. The results indicate that when only data of radia- Cigizoglu, H. K., and Kisi, O. 共2006兲. “Methods to improve the neural
tion and temperature are available, the Hargreaves method yields network performance in suspended sediment estimation.” J. Hydrol.,
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

dependable results on calibration. The ANFIS2 model can be used 317, 221–238.
in estimation of FAO-56 PM ET0 where there exist only the Rs Drake, J. T. 共2000兲. “Communications phase synchronization using the
and T data. In such a situation, the Hargreaves and Ritchie models adaptive network fuzzy inference system.” Ph.D. thesis, New Mexico
should be used after calibration. Among the empirical methods, State Univ., Las Cruces, N.M.
the Hargreaves method showed much less accuracy than the oth- Hargreaves, G. H., and Samani, Z. A. 共1985兲. “Reference crop evapo-
ers. The bad performance of the Hargreaves method may be pe- transpiration from temperature.” Appl. Eng. Agric., 1共2兲, 96–99.
culiar to the stations considered in the study. This standard Haykin, S. 共1998兲. Neural networks—A comprehensive foundation, 2nd
method is found to be unsuitable for the case studies considered, Ed., Prentice-Hall, Upper Saddle River, N.J., 26–32.
but not for all stations in general. The comparison results of the Hidalgo, H. G., Cayan, D. R., and Dettinger, M. D. 共2005兲. “Sources of
cross-station applications indicated that the ANFIS models could variability of evapotranspiration in California.” J. Hydroeteorol., 6,
3–19.
be more adequate than the ANN and MLR models in estimation
Jain, S. K., Das, D., and Srivastava, D. K. 共1999兲. “Application of ANN
of ET0 using nearby station data.
for reservoir inflow prediction and operation.” J. Water Resour. Plann.
The ANFIS technique could be of use in design of reservoirs Manage., 125共5兲, 263–271.
and various other hydrological analyses where other models may Jang, J.-S. R. 共1993兲. “ANFIS: Adaptive-network-based fuzzy inference
be inappropriate. These neurofuzzy models can be embedded as a system.” IEEE Trans. Syst. Man Cybern., 23共3兲, 665–685.
module for estimating ET0 data in hydrological modeling studies. Jang, J.-S. R., Sun, C.-T., and Mizutani, E. 共1997兲. Neurofuzzy and soft
The study used data from only two stations and further studies computing: A computational approach to learning and machine intel-
using more data from various areas may be required to reinforce ligence, Prentice-Hall, Upper Saddle River, N.J.
the conclusions drawn from this study. Jensen, M. E., Burman, R. D., and Allen, R. G. 共1990兲. Evapotranspira-
tion and Irrigation Water Requirements. ASCE Manuals and Reports
on Engineering Practices No. 70, ASCE, New York.
Jones, J. W., and Ritchie, J. T. 共1990兲. “Crop growth models.” Manage-
Acknowledgments ment of farm irrigation system, G. J. Hoffman, T. A. Howel, and K. H.
Solomon, eds., ASAE Monograph No. 9, ASAE, St. Joseph, Mich.,
The data used in this study were downloaded from the California 63–89.
Irrigation Management Information System 共CIMIS兲 web server. Karunanithi, N., Grenney, W. J., Whitley, D., and Bovee, K. 共1994兲.
The author wishes to thank the staff of the CIMIS who are asso- “Neural networks for river flow prediction.” J. Comput. Civ. Eng.,
ciated with data observation, processing, and management of 8共2兲, 201–220.
CIMIS Web sites. Kisi, O. 共2004a兲. “River flow modeling using artificial neural networks.”
J. Hydrol. Eng., 9共1兲, 60–63.
Kisi, O. 共2004b兲. “Multilayer perceptrons with Levenberg–Marquardt
training algorithm for suspended sediment concentration prediction
References and estimation.” Hydrol. Sci. J., 49共6兲, 1025–1040.
Kisi, O. 共2005a兲. “Suspended sediment estimation using neurofuzzy and
Allen, R. G., Pereira, L. S., Raes, D., and Smith, M. 共1998兲. “Crop evapo- neural network approaches.” Hydrol. Sci. J., 50共4兲, 683–696.
transpiration guidelines for computing crop water requirements.” Kisi, O. 共2005b兲. “Daily river flow forecasting using artificial neural net-
Proc., FAO Irrigation and Drainage, Paper No. 56, Food and Agri- works and auto regressive models.” Turk. J. Eng. Environ. Sci., 29,
culture Organization of the United Nations, Rome. 9–20.
Allen, R. G., Smith, M., Perrier, A., and Pereira, L. S. 共1994兲. “An update Kisi, O. 共2006兲. “Evapotranspiration estimation using feed-forward neu-
for the calculation of reference evapotranspiration.” ICID Bull., ral networks.” Nord. Hydrol., 37共3兲, 247–260.
43共2兲, 35–92. Kisi, O., and Yıldırım, G. 共2005a兲. “Discussion of ‘Estimating actual
ASCE Task Committee. 共2000a兲. “Artificial neural networks in hydrology evapotranspiration from limited climatic data using neural computing
I.” J. Hydrol. Eng., 5共2兲, 115–123. technique.’” J. Irrig. Drain. Eng., 131共2兲, 219–220.
ASCE Task Committee. 共2000b兲. “Artificial neural networks in hydrology Kisi, O., and Yıldırım, G. 共2005b兲. “Discussion of ‘Forecasting of refer-
II.” J. Hydrol. Eng., 5共2兲, 124–132. ence evapotranspiration by artificial neural networks.’” J. Irrig.
Brutsaert, W. H. 共1982兲. Evaporation into the atmosphere, Reidel, Drain. Eng., 131共4兲, 390–391.
Dordrecht, Holland. Kumar, D. N., Raju, K. S., and Sathish, T. 共2004兲. “River flow forecasting
Cancelliere, A., Giuşiano, G., Ancarani, A., and Rossi, G. 共2002兲. “A using recurrent neural networks.” Water Resour. Manage., 18, 143–
neural networks for deriving irrigation reservoir operating rules.” 161.
Water Resour. Manage., 16, 71–88. Kumar, M., Raghuwanshi, N. S., Singh, R., Wallender, W. W., and Pruitt,
Chen, J., and Adams, B. J. 共2006兲. “Integration of artificial neural net- W. O. 共2002兲. “Estimating evapotranspiration using artificial neural
works with conceptual models in rainfall–runoff modeling.” J. Hy- network.” J. Irrig. Drain. Eng., 128共4兲, 224–233.
drol., 318, 232–249. Lippman, R. 共1987兲. “An introduction to computing with neural nets.”
Cigizoglu, H. K. 共2003兲. “Estimation, forecasting, and extrapolation of IEEE ASSP Mag.., 4, 4–22.
flow data by artificial neural networks.” Hydrol. Sci. J., 48共3兲, 349– Maier, H. R., and Dandy, G. C. 共2000兲. “Neural network for the predic-
361. tion and forecasting of water resources variables: A review of model-

378 / JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007

J. Irrig. Drain Eng. 2007.133:368-379.


ing issues and applications.” Environ. Modell. Software, 15, 101–124. tion with hourly data.” California Irrigation Management Information
Minnes, A. W., and Hall, M. J. 共1996兲. “Artificial neural networks as System Final Rep., Land, Air and Water Resources Paper #10013,
rainfall–runoff models.” Hydrol. Sci. J., 41共3兲, 399–416. Univ. of California at Davis, Davis, Calif.
Naoum, S., and Tsanis, I. K. 共2003兲. “Hydroinformatics in evapotranspi- Sobrino, J. A., Gomez, M., Munoz, J. C. J., Olioso, A., and Chehbouni,
ration estimation.” Environ. Modell. Software, 18, 261–271. G. 共2005兲. “A simple algorithm to estimate evapotranspiration from
Pruitt, W. O., and Doorenbos, J. 共1977兲. “Empirical calibration, a requi- DAIS data: Application to the DAISEX campaigns.” J. Hydrol., 315,
site for evapotranspiration formulae based on daily or longer mean 117–125.
climatic data.” Proc., Int. Round Table Conf. on Evapotranspiration, Sudheer, K. P., Gosain, A. K., and Ramasastri, K. S. 共2003b兲. “Estimating
Budapest, Hungary, International Commission on Irrigation and actual evapotranspiration from limited climatic data using neural com-
Drainage. puting technique.” J. Irrig. Drain. Eng., 129共3兲, 214–218.
Raghuwanshi, N. S., Singh, R., and Reddy, L. S. 共2006兲. “Runoff and Sudheer, K. P., Nayak, P. C., and Ramasastri, K. S. 共2003a兲. “Improving
Downloaded from ascelibrary.org by SOUTHERN CALIFORNIA UNIVERSITY on 04/04/14. Copyright ASCE. For personal use only; all rights reserved.

sediment yield modeling using artificial neural networks: Upper Si- peak flow estimates in artificial neural network.” Hydrolog. Process.,
17, 677–686.
wane River, India.” J. Hydrol. Eng., 11共1兲, 71–79.
Supharatid, S. 共2003兲. “Application of a neural network model in estab-
Rumelhart, D. E., Hinton, G. E., and Williams, R. J. 共1986兲. “Learning
lishing a stage–discharge relationship for a tidal river.” Hydrolog.
internal representation by error propagation.” Parallel distributed pro-
Process., 17, 3085–3099.
cessing, D. E. Rumelhart and J. L. McClelland, eds., Vol. 1, MIT, Tayfur, G. 共2002兲. “Artificial neural networks for sheet sediment trans-
Cambridge, Mass. port.” Hydrol. Sci. J., 47共6兲, 879–892.
Russel, S. O., and Campbell, P. F. 共1996兲. “Reservoir operating rules with Trajkovic, S. 共2005兲. “Temperature-based approaches for estimating ref-
fuzzy programming.” J. Water Resour. Plann. Manage., 122共3兲, 165– erence evapotranspiration.” J. Irrig. Drain. Eng., 131共4兲, 316–323.
170. Trajkovic, S., Todorovic, B., and Stankovic, M. 共2003兲. “Forecasting ref-
Sahoo, G. B., and Ray, C. 共2006兲. “Flow forecasting for a Hawaii stream erence evapotranspiration by artificial neural networks.” J. Irrig.
using rating curves and neural networks.” J. Hydrol., 317, 63–80. Drain. Eng., 129共6兲, 454–457.
Smith, M., Allen, R., and Pereira, L. 共1997兲. Revised FAO methodology Vernieuwe, H., Georgieva, O., Baets, B. D., Pauwels, V. R. N., Verhoest,
for crop water requirements, Land and Water Development Division, N. E. C., and Troch, F. P. D. 共2005兲. “Comparison of data-driven
FAO, Rome. Takagi–Sugeno models of rainfall–discharge dynamics.” J. Hydrol.,
Snyder, R., and Pruitt, W. 共1985兲. “Estimating reference evapotranspira- 302, 173–186.

JOURNAL OF IRRIGATION AND DRAINAGE ENGINEERING © ASCE / JULY/AUGUST 2007 / 379

J. Irrig. Drain Eng. 2007.133:368-379.

S-ar putea să vă placă și