Académique Documents
Professionnel Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: In this study, the ability of four different data-driven methods, multilayer perceptron artificial neural net-
Received 30 December 2014 works (ANN), adaptive neuro-fuzzy inference system (ANFIS) with grid partition (GP), ANFIS with sub-
Received in revised form 18 April 2015 tractive clustering (SC) and gene expression programming (GEP), was investigated in predicting
Accepted 29 April 2015
long-term monthly reference evapotranspiration (ET0) by using data from 50 stations in Iran. The period-
icity component, station latitude, longitude and altitude values were used as inputs to the applied models
to predict the long-term monthly ET0 values. The overall accuracies of the multilayer perceptron ANN,
Keywords:
ANFIS-GP and ANFIS-SC models were found to be similar to each other. The GEP model provided the
Neural networks
Adaptive neuro-fuzzy
worst estimates. The maximum determination coefficient (R2) values were found to be 0.997, 998 and
Genetic programming 0.994 for the ANN, ANFIS-GP and ANFIS-SC models in Karaj station, respectively. The highest R2 value
Geographical inputs (0.978) of GEP model was found for the Qom station. The minimum R2 values were respectively found
Reference evapotranspiration as 0.959 and 0.935 for the ANN and ANFIS-GP models in Bandar Abbas station while the ANFIS-SC and
GEP models gave the minimum R2 values of 0.937 and 0.677 in the Tabriz and Kerman stations, respec-
tively. The results indicated that the long-term monthly reference evapotranspiration of any site can be
successfully estimated by data-driven methods applied in this study without climatic measurements. The
interpolated maps of ET0 were also obtained by using the optimal ANFIS-GP model and evaluated in the
study. The ET0 maps showed that the highest amounts of reference evapotranspiration occurred in the
southern and especially southeastern parts of the Iran.
Ó 2015 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.compag.2015.04.015
0168-1699/Ó 2015 Elsevier B.V. All rights reserved.
O. Kisi et al. / Computers and Electronics in Agriculture 115 (2015) 66–77 67
Table 1
The geographical information of the stations used in the study.
No Station Long. (E) Lat. (N) Alt. (m) ET0 (mm/day) No Station Long. (E) Lat. (N) Alt. (m) ET0 (mm/day)
1 Tehran Merabad 51 19 35 41 1191 4.76 26 Ramsar 50 40 36 54 20 2.24
2 Sanandaj 47 0 36 20 1373 3.87 27 Babolsar 52 39 36 43 -21 2.53
3 Rasht 49 37 37 19 -8.6 2.32 28 Saqqez 46 21 36 15 1523 3.56
4 Ahwaz 48 40 31 20 22.5 6.20 29 Sabzevar 57 39 36 12 972 5.46
5 Zanjan 48 29 36 41 1663 4.04 30 Torbat-e Heydarieh 59 16 35 13 1451 4.15
6 Urmia 45 3 37 40 1328 3.10 31 Khoy 44 58 38 33 1103 2.88
7 Ardabil 48 17 38 15 1332 2.87 32 Kashan 51 27 33 59 982 3.39
8 Gorgan 54 24 36 54 0 2.84 33 Fasa 53 41 28 58 1288 4.69
9 Sari 53 0 36 33 23 2.78 34 Bandar Anzali 49 27 37 29 -23.6 2.41
10 Arak 49 55 34 6 1708 3.68 35 Bandar Lengeh 54 50 26 32 22.7 5.58
11 Kermanshah 47 9 34 21 1319 4.34 36 Jask 57 46 25 38 5.2 5.23
12 Ilam 46 26 33 38 1337 4.52 37 Bam 58 21 29 6 1067 6.47
13 Khorramabad 48 17 33 26 1148 4.11 38 Abadan 48 15 30 22 6.6 6.68
14 Shahrekord 50 51 32 17 2049 3.46 39 Dezful 48 23 32 24 143 5.00
15 Qazvin 50 3 36 15 1279 3.83 40 Zabol 61 29 31 2 489 8.42
16 Yasuj 51 33 30 41 1816 3.89 41 Mashhad 59 38 36 16 999 3.86
17 Bojnord 57 16 37 28 1112 3.59 42 Bandar Abbas 56 22 27 13 9.8 5.29
18 Birjand 59 12 32 52 1491 5.10 43 Semnan 53 25 35 35 1127 4.05
19 Zahedan 60 53 29 28 1370 5.84 44 Shiraz 52 36 29 32 1484 4.90
20 Yazd 54 17 31 54 1237 5.45 45 Kerman 56 58 30 15 1754 5.43
21 Bushehr 50 49 28 58 9 5.06 46 Tabriz 46 17 38 5 1361 4.11
22 Tabas 56 55 33 36 711 4.67 47 Hamedan 48 43 25 12 1679 3.93
23 Iranshahr 60 42 27 12 591 6.10 48 Qom 50 51 34 42 877 4.83
24 Chabahar 60 37 25 17 8 4.67 49 Isfahan 51 40 32 37 1550 4.42
25 Shahrud 54 57 36 25 1349 4.00 50 Karaj 50 54 35 55 1313 4.11
incomplete or not always available for many locations and even The main problem with modeling ET process is its nonlinear
not always reliable (Rahimikhoob, 2010). Regarding the above con- dynamic and high complexity, in this respect, using data-driven
texts, there are so many factors affecting ET estimation based on methods (e.g. ANN, ANFIS, SVM ...) based on soft computing tech-
indirect methods, thereafter it is extremely difficult to formulate niques could be considered as proper methods for estimating
the ET0 equation that can produce reliable estimates. ET0. These methods are known for their ability in dealing with
68 O. Kisi et al. / Computers and Electronics in Agriculture 115 (2015) 66–77
and Ahani, 2012). Iran is a vulnerable country in regard to climate Here, ðpi ; qi ; r i Þ is the linear combination of all inputs and needs
changes and global warming because it has low amounts of annual to be adjusted as consequent parameters.
precipitation; highly spatial and temporal variation in precipita- – Layer VI (summation): in this layer the summation of all
tion in contrast to the tremendous potential of evapotranspiration difuzzified values produce the output of the ANFIS model.
(Abbaspour et al., 2009). X
The data of 50 stations were compiled from the Iran OutputVI ¼ W if i ð6Þ
Meteorological Organization (Tehran). The location of the stations
in Iran is represented in Fig. 1. The geographical positions of these
In Takagi–Sugeno ANFIS, a hybrid-learning algorithm is applied
stations are given in Table 1. Data cover the long-term averaged
for training the network consisting of the combination of gradient
evapotranspirations between the periods of 1956–2010. The
descent (for identifying the premise parameters) and least-squares
long-term monthly ET0 ranges from as low as 2.24 mm/day in
methods (to assign the linear consequent parameters). The learn-
Ramsar station to as high as 8.42 mm/day in Zabol station.
ing algorithm tunes the premise (MF parameters) and consequence
parameters (the response of the network), according to the training
3. Methods data. In the forward part of hybrid learning algorithm, the conse-
quent parameters are identified by the least squares method until
3.1. ANFIS layer IV. In the backward part, the premise parameters are updated
Layer V (deffuzification): the adjustable nodes of this layer cal- Gene recombinaon
culate the weighted output of the rth fuzzy rule by multiplica-
tion of the normalized firing strength (Output IV) and
consequent function (pi + qi + ri). New chromosomes of next generaon
Output V;i ¼ W i f i ¼ W i ðpi þ qi þ r i Þ ð5Þ Fig. 4. The flowchart of the gene expression algorithm (Ferreira, 2006).
70 O. Kisi et al. / Computers and Electronics in Agriculture 115 (2015) 66–77
by gradient descent using the back-propagation algorithm (Jang ANFIS-SC model with a radii value of 0.90. Similar to the
et al., 1997). More information about ANFIS can be found in Jang ANFIS-GP models, training of the ANFIS-SC models were stopped
(1993). after 100 epochs.
Generally, two ANFIS types can be established by using GP and
SC methods according to the input vector and output variables. In 3.2. ANN
GP method, the input and output space are partitioned into fuzzy
grids. Using these fuzzy grids the location of MFs can be optimized. Artificial neural networks are biologically inspired computa-
In SC method, each input parameter in the first layer is clustered tional model constructed of many simple interconnected elements
into several class values to build up fuzzy rules. More details about called neurons (processing elements) connected with coefficients
GP and SC can be found in (Kisi and Zounemat-Kermani, 2014). In
the present study, different membership functions, triangular
(trimf), Gaussian (gaussmf), trapezoidal (trapmf), two Gaussian
(gauss2mf), pi-shaped (pimf) and generalized bell (gbellmf) and
100 iterations were used for the ANFIS-GP models. Different num-
bers of membership functions were tested and the best one that
gave the minimum RMSE (root mean square errors) was selected
for the ANFIS-GP models. For the ANFIS-SC model, different radii
values were tried and optimal results were obtained from an
Table 2
Test results of ANN training algorithms.
Table 3
Test results of different ANFIS-GP structures.
Table 4
Test results of different GEP basic and linking functions for the parse tree.
(weights and biases) which constitute the neural structure. ANNs biases in the descending gradient direction. This algorithm consists
are capable of recognizing underlying relationships between input of different versions such as: (1) simple gradient descent BP, (2)
and output procedures. gradient descent with adaptive learning rate, (3) gradient descent
Several types of ANNs have been designed and reported with momentum, (4) gradient descent with momentum and adap-
(Zounemat-Kermani et al., 2013). One of the most comprehensive tive learning rate, and (5) scaled conjugate gradient (Mukherjee
types of ANNs is multilayer perceptron models which have been and Routroy, 2012).
proven to be a flexible and universal approximator (Fig. 3). In mul- The LM algorithm approximates to the Gauss–Newton method
tilayer perceptron ANN networks the flow of computation is from which approximates the error of the network with a second order
the input layer to the output layer. The arriving variables (inputs), expression. Hence, LM can be considered a combination of steepest
multiplied by the connection weights and added by the biases are descent and the Gauss–Newton method. It is proved that LM out-
first combined (summation). Each neuron has an adjustable weight performs simple gradient descent and other conjugate gradient
factor (w) and bias (b). Thereafter, the adjusted parameters passed methods in a wide variety of problems (Zounemat-Kermani et al.,
through an activation (transfer) function to make the output for 2013). LM algorithm uses an approximation to the Hessian matrix
that neuron. The outputs produced from the first hidden layer in the following Gauss–Newton weight update:
become the inputs to the second layer and so on to the output
layer. wkþ1 ¼ wk ½H þ kI1 J T e ð8Þ
In supervised learning of ANNs, there are several applicable
learning algorithms but the most often used is back-propagation where w is the weight of neural network, H is the Hessian matrix
(BP) rule. In BP, optimization of the weights and biases is made evaluated at wk. The Jacobian matrix (J) can be computed through
by backward propagation of the error during training phase. The a standard back-propagation technique which is much less complex
ANN compares the input and output values in the training data than computing the Hessian matrix. Thereafter, the Hessian matrix
set and alters the values of the processing units to reduce the dif-
can be approximated as H ¼ JT J. e implies the vector of network
ference between the predicted and target values. The difference
errors. k is a scalar which can adjusted thorough the training proce-
between predicted and target values makes the error vector, which
dure. When the scalar k is zero, this is just Gauss–Newton’s method,
is used to adjust the network weights and biases.
using the approximated Hessian matrix. However considering large
The error performance measure is generally mean-square error
values for k, the algorithm becomes gradient descent with a small
(MSE), which can be calculated as:
step size. It should be mentioned that Newton’s method is faster
and more accurate therefore the goal is to shift toward Newton’s
1 XN
E¼ e2 ð7Þ method as fast as possible. However, the main shortcoming of LM
2N i¼1 i
algorithm is its increased memory requirement to calculate the
where N is total number of training patterns and ei ¼ yi t i equals Jacobian matrix of the error function. In the present paper, multi-
error for the ith training pattern. Y is the network predicted output layer perceptron ANN models were used. For the ANN models, dif-
and t is the observed target. ferent training algorithms, Levenberg–Marquardt (lm), gradient
Considering BP, there are several algorithms which can be used descent (gd), gradient descent with adaptive learning rate (gda),
for training ANNs such as gradient descent and Levenberg– gradient descent with momentum (gdm), gradient descent with
Marquadt. Gradient descent which is one of the most popular momentum and adaptive learning rate (gdx) and scaled conjugate
training algorithms works by measuring the output error, calculat- gradient (scg), were used for calculating the weights and results
ing the gradient of this error, and then adjusting the weights and were compared.
5
Table 6
ANN ANFIS-GP ANFIS-SC GEP Summary of the test process of ANFIS and GEP models.
4
Stations ANN ANFIS-GP ANFIS-SC GEP
RMSE, (mm/day)
from all the evolutionary advantages this brings about (Ferreira, In this research, different functions were utilized, including basic
2006). arithmetic operators ðþ; ; ; Þ as well as some of the other basic
GEP is an example of a full-fledged replicator/phenotype system mathematical functions as blow:
where the chromosomes/expression trees form a truly functional,
indivisible whole (Ferreira, 2001). The fundamental steps of the pffiffiffi pffiffiffi
ln x; e x ; x2 ; x3 ; x; 3 x; sin x; cos x; arctanðxÞ
gene expression algorithm (GEA) are schematically showed in
Fig. 4. The third major step is to choose the chromosomal architecture.
The procedure for ET0 prediction by GEP consists five major The fourth major step is to choose the kind of linking function. And
steps. The first step is to choose the fitness function. Root relative finally, the fifth major step in preparing to use GEP is to choose the
squared error (RRSE) is selected as the best fitness function in this set of genetic operators and their rates (Ferreira, 2006).
study. The second major step is to choose the set of terminals (T) Soft computing program called GeneXpro Tools 4.0 was used to
and the set of functions (F). In current work, the terminal set con- predict ET0 using GEP. The parameters used are summarized as fol-
sists of the geographical information of stations (latitude, longi- lows: Number of chromosomes: 30, head size: 8, number of genes:
tude and altitude) and the number corresponding each month. 3, mutation rate: 0.044, inversion rate: 0.1, one point recombina-
The set of functions is selected based on the nature of the problem. tion rate: 0.3, two point recombination rate: 0.3, gene
PM-FAO56 ANN
PM-FAO56 ANN
10 ANFIS-GP ANFIS-SC 10 ANFIS-GP ANFIS-SC
GEP GEP
8 8
ET0 (mm/day)
ET0 (mm/day)
6 6
4 4
2 2
(a) (b)
0 0
JAN
APR
JUN
AUG
NOV
FEB
MAR
JUL
OCT
DEC
MAY
SEP
FEB
MAR
OCT
JUL
SEP
DEC
JAN
APR
MAY
JUN
AUG
NOV
month month
Fig. 7. Penman–Monteith (FAO-56) versus predicted ET0 values in (a) Mashhad and (b) Bandar Abbas weather stations.
ET0 (mm/day)
6 6
4 4
2 2
(a) (b)
0 0
DEC
FEB
MAR
JUL
OCT
MAY
SEP
JAN
APR
JUN
AUG
NOV
JAN
APR
JUN
AUG
NOV
MAR
OCT
FEB
MAY
JUL
SEP
DEC
month month
Fig. 8. Penman–Monteith (FAO-56) versus predicted ET0 values in (a) Semnan and (b) Shiraz weather stations.
6 6
4 4
2 2
(a) (b)
0 0
FEB
MAR
JUL
SEP
OCT
DEC
APR
JAN
JUN
AUG
NOV
JAN
FEB
MAR
APR
JUN
JUL
AUG
OCT
SEP
NOV
DEC
MAY
MAY
month month
Fig. 9. Penman–Monteith (FAO-56) versus predicted ET0 values in (a) Kerman and (b) Tabriz weather stations.
O. Kisi et al. / Computers and Electronics in Agriculture 115 (2015) 66–77 73
ET0 (mm/day)
ET0 (mm/day)
6 6
4 4
2 2
(a) (b)
0 0
FEB
MAR
OCT
DEC
JUL
SEP
APR
JAN
JUN
AUG
NOV
APR
JAN
MAR
FEB
JUN
JUL
AUG
SEP
OCT
NOV
DEC
MAY
MAY
month month
Fig. 10. Penman–Monteith (FAO-56) versus predicted ET0 values in (a) Hamedan and (b) Qom weather stations.
ET0 (mm/day)
ET0 (mm/day)
6 6
4 4
2 2
(a) (b)
0 0
FEB
MAR
JUL
OCT
SEP
DEC
JAN
APR
MAY
JUN
AUG
NOV
JAN
APR
JUN
AUG
NOV
FEB
MAR
JUL
OCT
DEC
MAY
SEP
month month
Fig. 11. Penman- Monteith (FAO-56) versus predicted ET0 values in (a) Isfahan and (b) Karaj weather stations.
Stations ANN ANFIS-GP ANFIS-SC GEP Four different data-driven methods, ANN, ANFIS-GP, ANFIS-SC
Mashhad 4 1 2 3 and GEP, were applied for estimating long-term monthly mean
Bandar abbas 2 1 4 3 ET0. Latitude, longitude, altitude and periodicity component
Semnan 3 2 1 4 (month of the year) data were used as inputs to the each model.
Shiraz 2 4 1 3 Monthly data of 40 weather stations (40 sta-
Kerman 2 1 3 4
Tabriz 4 2 3 1
tions 12 months = 480 data) were used for training and 10 sta-
Hamedan 4 1 2 3 tions’ data (10 stations 12 months = 120 data) were used for
Qom 1 4 2 3 testing. The stations used for the testing procedure are Mashahd,
Isfahan 1 4 2 3 Bandar abbas, Semnan, Shiraz, Kerman, Tabriz, Hamedan, Qom,
Karaj 1 3 2 4
Isfahan and Karaj. The models were compared with each other
Total 24 23 22 31 with respect to RMSE (root mean square errors) and determination
coefficient (R2) statistics. The RMSE can be defined as
recombination rate: 0.1, gene transposition rate: 0.1, IS transposi- vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
tion rate: 0.1, RIS transposition: 0.1. Different basic functions and u1 X
linking functions were used for the GEP models and optimal one RMSE ¼ t ðET m ET e Þ2 ð9Þ
N i¼1
was selected with respect to RMSE.
Table 8
Comparison of the applied models with respect to fitting line equations.
where ETo and ETe denote the measured and estimated evapotran- ET0ðGEPÞ ¼ B þ exp ½arctg ðarctg ðarctg ð2 Longitude ðAltitude 2:057ÞÞÞÞ
spiration values, respectively and N is the number of data sets. ð10Þ
The comparison of different ANN training algorithms is made in
Table 2. In this table, 4-2-1 indicates an ANN model respectively 6
tan1 ðLongitude Þ
comprising 4 inputs, 2 hidden and 1 output nodes at 1,000 epochs. where B ¼ A þ Latitude
and
The hidden nodes’ number was determined using trial and error monthþLongitude
method. For each ANN model, hidden node number increased from 2=3
A ¼ 3:061COS month ðmonth Þ ð11Þ
1 to 10 and the optimal ones were selected with respect to RMSE
criterion. Training of the lm, gd, gda, gdmgdx and scg algorithms Training and test results of the optimal data-driven models
was respectively stopped after 1000, 50,000, 50,000, 50,000, which obtained from 50 stations are given in Table 5. It is clear
50,000 and 1000 iterations because the decrease in accuracy was from the table that the ANFIS-GP model has the lowest RMSE
too small after these epochs. The tangent sigmoid and linear acti- and the highest R2 in both training and test periods. The ranks of
vation functions were used for the hidden and output nodes, the applied models in the test period are: ANFIS-GP, GEP, ANN
respectively. It is clear from Table 2 that gradient descent with and ANFIS-SC. The RMSE accuracy of data-driven models in the
momentum and adaptive learning rate back-propagation algo- training period is compared for each station in Fig. 6. It is clear that
rithm has a better accuracy than the other algorithms in the test the ANFIS-GP model generally has lower RMSE values than the
period. other models. In some cases (Rasht, Yasuj, Shahrud, Babolsar,
Different ANFIS-GP models having different types of MFs is Torbat-e Heydarieh, Bandar Anzali, Bandar Lengeh, Jask and
compared in Table 3. In this table, 4-2-2-2 indicates an ANFIS-GP Zabol), ANFIS-SC model performs better than the ANFIS-GP model.
model having 4, 2, 2 and 2 MFs for the month of the year, longi- Test results of the optimal data-driven models for each station
tude, latitude and altitude inputs. It is apparent from the table that are provided in Table 6. It is clear from the table that the RMSE
ANFIS-GP model with pi-shaped MFs has the best accuracy with values range 0.238–1.474, 0.337–1.090, 0.261–1.290, 0.420–
respect to RMSE and R2. 0.930 mm/day for the ANN, ANFIS-GP, ANFIS-SC and GEP models,
Table 4 gives the test results of the GEP models. It is clear from respectively. For the ANN, ANFIS-GP and GEP models, the
the table that the best accuracy was obtained from the GEP model maximum RMSE (1.474, 1.090, 0.930 mm/day) values were
with addition linking function and F5 operators. The expression found for the Tabriz station, respectively. For the ANFIS-SC
tree for the optimal GEP model is shown in Fig. 5. The equation model, however, the maximum RMSE value was found to be
of the optimal GEP model is 1.290 mm/day in the Bandar abbas station. The best ANNFIS-SC
Fig. 12. ET0 maps (mm) in January, February, March and April of Iran based on ANFIS-GP model.
O. Kisi et al. / Computers and Electronics in Agriculture 115 (2015) 66–77 75
(RMSE = 0.261 mm/day) and GEP (RMSE = 0.420 mm/day) models to October months in this station. The applied models cannot catch
results were found for the Shiraz station while the ANN and these sudden changes. The accuracy ranks of the data-driven mod-
ANFIS-GP models provided best accuracy in the Karaj and els are given in Table 7. It is clear from the table that the overall
Kerman stations, respectively. It is clearly seen from Table 6 that rank of the methods’ accuracies is: ANFIS-SC, ANFIS-GP, ANN and
the accuracy of the ANFIS-GP model is generally better than the GEP. The overall accuracies of the ANFIS-SC, ANFIS-GP and ANN
other models in long-term monthly ET0 prediction. In four of ten models are similar to each other. The GEP model seems to be the
stations, the ANFIS-GP model has the best accuracy. The ANN worst model. However, the main advantage of the GEP model is
and ANFIS-SC models respectively performed the best in three that it has an explicit formulation can be simply used in practical
and two stations while the GEP was found to be the best model applications. Recently, Kisi and Sanikhani (2015) applied
in only one station. For the ANN, ANFIS-GP and ANFIS-SC models, ANFIS-GP, ANFIS-SC and ANN models for precipitation estimation
the maximum determination coefficient values were found to be and they found the overall rank of the methods’ accuracies as
0.997, 998 and 0.994 in Karaj station, respectively. The GEP model ANFIS-GP, ANN and ANFIS-SC. This indicates that the models’ accu-
has the highest R2 value of 0.978 for the Qom station. The mini- racies change according to the investigated phenomenon.
mum determination coefficient values were respectively found as Test results of the applied models are compared with respect to
0.959 and 0.935 for the ANN and ANFIS-GP models in Bandar abbas fit line equations (assume that the equation is y = ax + b) in Table 8.
station while the ANFIS-SC and GEP models have the minimum R2 In this table, a values smaller than 1 indicate underestimation
values of 0.937 and 0.677 in the Tabriz and Kerman stations, while higher values show overestimation. From the table, it is clear
respectively. that the all the models give underestimations for the Tabriz,
The test results of the data-driven models are compared in Hamedan, Qom and Karaj stations. It is clear from the Fig. 1 that
Figs. 7–11. The monthly ET0 estimates given in the graphs gener- these stations are located close to each other and thus have same
ally confirm the test results given in Table 6. From the Figs. 8 climatic conditions. It is clear from the Table 8 that the models
and 11, it is clear that all the models provided good accuracy for generally underestimates in long-term monthly ET0 prediction.
the Semnan, Isfahan and Karaj stations. ANFIS-SC and GEP models The reason of this may be the difference between the training
significantly overestimated the ET0 values of Bandar abbas. All the and validation data characteristics. Higher values used in training
models gave underestimations for the summer months of Tabriz period may lead data-driven models underestimations in the vali-
station. The reason of this may be the fact that a sudden increase dation period. In Table 8, a and b values respectively close to 1 and
is seen from May to June and from June to July months and sudden 0 indicate that fitting line close to exact line (y = x). From Table 8, it
decrease happens from August to September and from September is apparent that the fitting lines of the ANFIS-SC models are
Fig. 13. ET0 maps (mm) in May, June, July and August of Iran based on ANFIS-GP model.
76 O. Kisi et al. / Computers and Electronics in Agriculture 115 (2015) 66–77
Fig. 14. ET0 maps (mm) in September, October, November and December of Iran based on ANFIS-GP model.
determination coefficient statistics. Different training algorithms Karimaldini, F., Shui, L.T., Mohamed, A.T., Abdollahi, M., Khalili, N., 2012. Daily
evapotranspiration modeling from limited weather data by using neuro-fuzzy
were used for ANN models and gradient descent with momentum
computing technique. J. Irrig. Drain Eng. 138 (1), 21–34.
and adaptive learning rate back-propagation algorithm was found Kim, S., Kim, H.S., 2008. Neural networks and genetic algorithm approach for
to be better than the other algorithms. Various membership func- nonlinear evaporation and evapotranspiration modeling. J. Hydrol. 351, 299–
tions were tried in ANFIS-GP models and pi-shaped MFs gave the 317.
Kisi, O., Ozturk, O., 2007. Adaptive neurofuzzy computing technique for
best accuracy. In the test period, the ANFIS-GP model generally evapotranspiration estimation. J. Irrig. Drain. Eng. ASCE 133 (4), 368–379.
performed better than the other models in long-term monthly Kisi, O., Sanikhani, H., 2015. Prediction of long-term monthly precipitation using
ET0 prediction. Test results of the optimal data-driven models were several soft computing methods without climatic data. Int. J. Climatol. http://
dx.doi.org/10.1002/joc.4273.
also compared for each station. The ANFIS-GP model had the best Kisi, O., Zounemat-Kermani, M., 2014. Comparison of two different adaptive neuro-
accuracy in four out of ten stations. The ANN and ANFIS-SC models fuzzy inference systems in modelling daily reference evapotranspiration. Water
respectively have the best accuracy in three and two stations while Resour. Manage. 28 (9), 2655–2675.
Kousari, M.R., Ahani, H., 2012. An investigation on reference crop
the GEP was found to be the best model in only one station. It evapotranspiration trend from 1975 to 2005 in Iran. Int. J. Climatol. 32 (15),
should be noted that all the models gave R2 value higher than 2387–2402.
0.900 except the one case (GEP model in Kerman station). As a con- Kumar, M., Raghuwanshi, N.S., Singh, R., Wallender, W.W., Pruitt, W.O., 2002.
Estimating evapotranspiration using artificial neural network. J. Irrig. Drain.
clusion, it can be said that the all the data-driven techniques can be Eng. 128 (4), 224–233.
successfully used to predict the long-term reference evapotranspi- Kumar, M., Raghuwanshi, N.S., Singh, R., 2011. Artificial neural networks approach
ration of any site without climatic measurements. in evapotranspiration modeling: a review. Irrig. Sci. 29 (1), 11–25.
Landeras, G., Ortiz-Barredo, A., Lopez, J.J., 2008. Comparison of artificial neural
Important information on various water resources issues such
network models and empirical and semi-empirical equations for daily reference
as examining water use by different regions, water rights, water evapotranspiration estimation in the Basque Country (Northern Spain). Agric.
allocations, water consumption and planning and management of Water Manage. 95, 553–565.
ground and surface water resources can be provided by ET0 maps. Marti, P., González-Altozano, P., López-Urrea, R., Mancha, L.A., Shiri, J., 2015.
Modeling reference evapotranspiration with calculated targets. Assessment and
The ET0 maps in different months revealed that the highest implications. Agric. Water Manage. 149 (2), 81–90.
amounts of reference evapotranspiration occurred in the southern Monteith, J.L., (1965). The state and movement of water in living organisms. In:
and especially southeastern parts of the Iran. More attention Proc., Evaporation and Environment, XIXth Symp., Soc. for Exp. Biol., Swansea.
Cambridge University Press, NY, pp. 205–234.
should be focused on water resource planning of these parts. Mukherjee, I., Routroy, S., 2012. Comparing the performance of neural networks
developed by using Levenberg–Marquardt and Quasi-Newton with the gradient
descent algorithm for modelling a multiple response grinding process. Exp.
Acknowledgement Syst. Appl. 39, 2397–2407.
Penman, H.L., 1963. Vegetation and hydrology. Tech. Communication No. 53,
Commonwealth Bureau of Soils. Harpenden, England.
This study was partly supported by The Turkish Academy of Rahimikhoob, A., 2010. Estimation of evapotranspiration based on only air
Sciences (TUBA). The first author would like to thank TUBA for temperature data using artificial neural networks for a subtropical climate in
Iran. Theor. Appl. Climatol. 101, 83–91.
their support of this study. Shiri, J., Nazemi, A.H., Sadraddini, A.A., Landeras, G., Kisi, O., Fard, Fakheri., Martí, P.,
2014a. Comparison of heuristic and empirical approaches for estimating
reference evapotranspiration from limited inputs in Iran. Comput. Electron.
References 108 (10), 230–241.
Shiri, J., Marti, P., Singh, V.P., 2014b. Evaluation of gene expression programming
approaches for estimating daily pan evaporation through spatial and temporal
Abbaspour, K.C., Faramarzi, M., Ghasemi, S.S., Yang, H., 2009. Assessing the impact
data scanning. Hydrol. Processes 28 (3), 1215–1225.
of climate change on water resources in Iran. Water Resour. Res. 45, W10434.
Shiri, J., Sadraddini, A.A., Nazemi, A.H., Kisi, O., Landeras, G., Fakheri Fard, A., Marti,
http://dx.doi.org/10.1029/2008WR007615.
P., 2014c. Generalizability of Gene Expression Programming-based approaches
Ahrens, C., 1998. Essentials of Meteorology: An Introduction to the Atmosphere,
for estimating daily reference evapotranspiration in coastal stations of Iran. J.
2nd ed. Publishing Company, Wadsworth.
Hydrol. 508, 1–11.
Aytek, A., 2009. Co-active neurofuzzy inference system for evapotranspiration
Sudheer, K.P., Gosain, A.K., Ramasastri, K.S., 2003. Estimating actual
Modeling. Soft Comput. 13 (7), 691–700.
evapotranspiration from limited climatic data using neural computing
Azamathulla, H.M., Ghani, A.A., Leow, C.S., Chang, C.K., Zakaria, N.A., 2011. Gene-
technique. J. Irrig. Drain. Eng. ASCE 129 (3), 214–218.
expression programming for the development of a stage-discharge curve of the
Tabari, H., Martinez, C., Ezani, A., Talaee, P.H., 2013. Applicability of support vector
Pahang River. Water Resour. Manage. 25 (11), 2901–2916.
machines and adaptive neurofuzzy inference system for modeling potato crop
Cobaner, M., 2011. Evapotranspiration estimation by two different neuro-fuzzy
evapotranspiration. Irrig. Sci. 31 (4), 575–588.
inference systems. J. Hydrol. 398 (3–4), 292–302.
Trajkovic, S., 2005. Temperature-based approaches for estimating reference
Dinpashoh, Y., Jhajharia, D., Fakheri-Fard, A., Singh, V.P., Kahya, E., 2011. Trends in
evapotranspiration. J. Irrig. Drain. Eng. 131 (4), 316–323.
reference crop evapotranspiration over Iran. J. Hydrol. 399 (3–4), 422–433.
Traore, S., Guven, A., 2012. Regional-specific numerical models of
Dogan, E., 2009. Reference evapotranspiration estimation using adaptive neuro-
evapotranspiration using gene-expression programming interface in Sahel.
fuzzy inference systems. Irrig. Drain. 58, 617–628.
Water Resour. Manage. 26, 4367–4380.
Ferreira, C., 2001. Gene expression programming: a new adaptive algorithm for
Traore, S., Wang, Y.M., Kerh, T., 2010. Artificial neural network for modeling
solving problems. Complex Syst. 13 (2), 87–129.
reference evapotranspiration complex process in Sudano-Sahelian zone. Agric.
Ferreira, C., 2006. Gene Expression Programming: Mathematical Modeling by an
Water Manage. 97, 707–714.
Artificial Intelligence. Springer, Berlin: Heidelberg New York, 478.
Zounemat-Kermani, M., Scholz, M., 2013. Computing air demand using the Takagi-
Guven, A., Aytek, A., Ishak Yuce, M., Aksoy, H., 2008. Genetic programming-based
Sugeno model for dam outlets. Water 5 (3), 1441–1456.
empirical model for daily reference evapotranspiration estimation. Clean 36
Zounemat-Kermani, M., Teshnehlab, M., 2008. Using adaptive neuro-fuzzy
(10–11), 905–912.
inference system for hydrological time series prediction. Appl. Soft Comput. 8
Jang, J., 1993. ANFIS: adaptive-network-based fuzzy inference system. IEEE
(2), 928–936.
Transact. Syst. Man. Cybern. 23 (3), 665–685.
Zounemat-Kermani, M., Kisi, O., Rajaee, T., 2013. Performance of radial basis and
Jang, J.S.R., Sun, C.T., Mizutani, E., 1997. Neuro-Fuzzy and Soft Computing: A
LM-feed forward artificial neural networks for predicting daily watershed
Computational Approach to Learning and Machine Intelligence. Prentice-Hall,
runoff. Appl. Soft Comput. 13 (12), 4633–4644.
New Jersey.