35-40
TI Journals
ISSN:
2306-7527
Copyright 2015. All rights reserved for TI Journals.
Keywords
Abstract
ANN
Evapotranspiration
FAO-Penman-Monteith,
GP
Evapotranspiration as an important element of hydrologic cycle has a main role in watersheds balance
assessment. In plant water requirements and Evapotranspiration volume calculation, firstly reference
evapotranspiration (ET0) have to be computed and then by using different methods for any plant water
requirements can be estimated. In this paper firstly the (ET0) factor calculated by standard FAO-PenmanMonteith formula by using climatic data of Tabriz station and then combinations of climatic parameters such
as mean, minimum and maximum of air temperature, mean, minimum and maximum of relative humidity
rainfall, wind speed and sunshine hours considered as an input of Multi-Layer Perceptrons (MLPs) Feed
Forward Back Propagation Neural Networks and Genetic Programming for estimate monthly (ET0). Results
showed that both approaches in modeling evapotranspiration showed exact results, but Genetic
Programming, can provides a applicable alternative mathematical relation to predict evapotranspiration.
Different scenario analysis with neural networks suggests that while rainfall data does not have any
significant influence in predicting evapotranspiration, but monthly time index could be improve modeling
results.
1.
Introduction
It is essential to accurately estimate crop water requirement in order to design irrigation system, irrigation scheduling, water management
systems and plant system modeling. Accurate estimates of crop water requirements in agriculture have a direct connection with the detailed
estimate of monthly evapotranspiration, which provide ground for precise design and implementation of hydraulic structures and optimal
utilization of the limited water resources within country. Evapotranspiration is considered as an important component of hydrological cycle,
which is significantly used to examine watershed balance. In order to calculate crop water requirement and evapotranspiration volume, (ET0) is
calculated at first. Then, evapotranspiration is calculated for each plant species. (ET0) can be directly measured either using Lysimeter such
empirical relationship as Penman-Monteith. The modified version of FAO-Penman-Monteithis used as the basis of such software as CropWat
used in the calculation of evapotranspiration. Although Lysimeter measures evapotranspiration accurately, placement of Lysimeter is costly and
not economically effective. So far, many approaches were developed to calculate ET0 based on meteorological data in various geographical and
climatic conditions. These methods require a range of complicated equations such as Penman-Monteith Allen et al. [1] dependent on many
meteorological data rather than less complicated equations such as Blanney and Criddle [2] and Hargreaves and Samani [3] dependent on small
data. Many researchers around the world argued that FAO-Penman-Monteith methods the most accurate method to estimate (ET0) in
comparison with Lysimeter . Therefore, FAO-Penman-Monteith method was used as the reference method in this study. Nowadays, such
intelligent systems as genetic programming and artificial neural network methods are used in hydrological processes modeling and water
engineering. These methods are known as valid methods to model the complex nonlinear processes. These methods can accurately estimate
(ET0) using the data recorded at meteorological stations. Evapotranspiration is essentially a complex and nonlinear process. Hence artificial
neural networks and genetic programming as novel methods with high reliability in modeling complex and nonlinear processes can be used for
modeling evapotranspiration process. Artificial neural networks are considered as black box techniques, which establish a relationship between
input and output data regardless of the physical processes governing the system. In recent years, artificial neural networks have been used in
various sectors of water engineering. Considering the continuous development trend, intelligent models of genetic programming are considered
as efficient tools to model water engineering processes. Many studies investigated modeling of evapotranspiration in abroad and within country.
Kouchakzadeh and Bahmani [4] examined artificial neural network performance in reducing the required parameters used in calculating the
(ET0) in Mehrabad Region in Tehran. Zanetti et al. [5] estimated evapotranspiration using meteorological data in Argentine with artificial neural
networks. Kisi and Ozturk [6] estimated evapotranspiration using FAO-Penman-Monteith method with Artificial Neural Fuzzy Inference System
(ANFIS). Rahimi Khoob [7] estimated evapotranspiration with high precision using data from meteorological stations in Khuzestan Region in
Iran with artificial neural network methods. Mogadamnia et al. [8] used artificial neural networks and neuro-fuzzy inference system to model
daily evaporation in Chahnime Region located in Sistan and Baluchestan Province. They finally stated that neural network could more precisely
predict the evaporation than empirical models and neural-inference system Fuzzy system. Piri et al. [9] estimated evapotranspiration for hot and
dry areas using artificial neural networks. Assari et al. [10] estimated (ET0) using artificial neural networks in greenhouse conditions. Sayyadi et
al. [11] examined the performance of two types of artificial neural network models for estimating (ET0) in Tabriz. Ghabaei Sough et al. [12]
estimated evapotranspiration in daily scale using data from synoptic meteorological stations in Shiraz with artificial neural networks. Several
researchers also used genetic programming to estimate evapotranspiration. Guven et al. [13] estimated (ET0) in California Region in America
using genetic programming. They compared genetic programming results with the observed data and showed that the former has high precision
in estimating ET0. Kisi and Guven [14] used different combinations of meteorological parameters for California Data and used artificial neural
networks and linear genetic programming method to model (ET0). Terzi [15] used genetic programming to estimate daily evapotranspiration
from free water surface level in Egirdir Lake in Turkey. Samadian et al. [16] considered 30 different scenarios and used genetic programming
method to predict reference crop potential evapotranspiration in Tabriz Region. They compared the obtained results using the former method
with Hargreaves and Jensen - Hayes experimental model results. They concluded that reference crop potential evapotranspiration could be
36
predicted with acceptable precision using genetic programming with air temperature and wind speed data. Shiri and Kisi [17] used intelligent
neural network, neural fuzzy inference system and genetic programming methods to estimate and compare daily evaporation in pans. Sattari et
al. [18] used two methods of artificial neural networks and M5 tree model to predict daily (ET0) in Bonab region, northwest of Iran. They
concluded that artificial neural networks provide more accurate results compared to M5 tree model. The present study aimed to model and
compare monthly (ET0) using such intelligent systems as neural networks and genetic programming in Tabriz. In this study, different scenarios
in relation to effective combination of meteorological parameters such as temperature, precipitation, wind speed, relative humidity, sunshine
hours, etc. are provided as input to the model in order to calculate evapotranspiration. The best scenarios with the minimum number of
parameters for more accurate prediction of monthly evapotranspiration are selected.
2.
Input layer
Hidden layer
Output layer
37
denotes net radiation at the surface vegetation, T refers to mean air temperature at a height of two meters above ground level, U2 denotes
average wind speed at a height of two meters above ground leveles-earefers to vapor pressure deficit at the height of two meters, denotes the
slope of vapor pressure curve, G refers to the heat flux into the soil and denotes moisture ratio (Allen et al.) [1].
890
]U 2 (es ea )
T 273
(1 0.34U 2 )
0.408 ( Rn G ) [
ET0
3.
(1)
Tabriz is the capital of Eastern Azerbaijan located in the West of the province in the Far East and South of Tabriz Plain. The city is located at l46
degrees, 26-minute longitude, and 38 degrees and 6 minutes latitude with an altitude of 1350 m to 1550 m above sea level. Tabriz City is
restricted from the North, South and East to the mountains and from the West to flat lands and salt marshes in Talkherood. Tabriz City's
geographical location is shown in Figure 2.
In this study, data from 19 years from January 1982 to October 2000 (75% of the data) was selected as training data while data from 7 years
from November 2000 to December 2006 (25% of the data) was selected as testing data. Performance of the models used in this study is
evaluated based on calculating the root mean square error (RMSE) and correlation coefficient (R 2). The formulas for calculation of the above
parameters are shown in the equations 2 and 3.
2
R
( X i X )(Yi Y )
i 1
N
N
2
2
(
X
X
)
*
(
Y
Y
)
i
i
i 1
i 1
(2)
(y
RMSE
xi ) 2
i 1
(3)
The statistical properties of the data used in this study are presented in Table 1. According to nature of the problem, various scenarios of input,
hidden and output layers were introduced for artificial neural network in this study. After implementation of the model, the model prediction
accuracy was examined based on such statistics as RMSE and R2.
4.
To determine the best combination, 13 different scenarios of combinations of input data were entered to the neural network models. These
scenarios and the results for each scenario based on two statistics such as RMSE and R2are shown in Table 2. Monthly numbers from 1 to 12
were considered as monthly indices in order to consider the impact of each month in modelling artificial neural network. In this section,
hyperbolic tangent function was considered as the stimulus function while momentum learning rule was used to rapidly converge the network.
As Table 2 shows, Scenario 6 with six input parameters (with four hidden layers) delivered the best results (R2=0.992, RMSE=0.053), which is
considered as the best scenario for introduction to genetic programming model. The results relevant to the observed values of (ET0) obtained
from the artificial neural network model, genetic programming in contrast to the values obtained using FAO-Penman-Monteith equation for the
input parameter in Scenario 6 are respectively shown in Figures 3 and 4.
38
Statistics
Mean
Standard deviation
Max
Min
MaxT
( C)
18.63
10.91
35.76
-3.55
SunH
(hr)
236.05
83.93
379.50
51.80
Wind 2
(m/s)
3.00
1.91
31.00
0.78
Minimum relative
humidity (%)
34.56
14.56
75.48
12.17
Maximum relative
humidity (%)
69.50
13.73
94.16
31.00
Precipitation
(mm)
20.98
21.33
123.40
0.00
Evapotranspiration
FPM (mm)
25.11
18.32
65.85
2.03
Table 2. Results of Artificial Neural Network Model in different scenarios of combination of input parameters
Scenarios
Input parameters
The monthly index, AvT, MinT, MaxT, Precipitation, AvH, MinH, MaxH, Wind 2, SunH
The monthly index, AvT, MinT, MaxT, AvH, MinH, MaxH, Wind 2, SunH
The monthly index, AvT, MinT, MaxT, MinH, MaxH, Wind 2, SunH
The monthly index, MinT, MaxT, MinH, MaxH, Wind 2, SunH
The monthly index, MinT, MaxT, Precipitation, AvH, Wind 2, SunH
The monthly index, AvT, MinH, MaxH, Wind 2, SunH
The monthly index, MinT, MaxT, MaxH, MinH
The monthly index, MinT, MaxT, Wind 2
The monthly index, MinT, MaxT, SunH
The monthly index, MinT, MaxT, precipitation
The monthly index, MinT, MaxT, AvH
The monthly index, MinT, MaxT
The monthly index, AvT
Y = 0.9792X + 0.9414
R = 0.96
80
70
60
50
40
30
20
10
0
70
GP ET0 (mm)
GP ET0 (mm)
1
2
3
4
5
6
7
8
9
10
11
12
13
R2
RMSE
(mm)
0.992
0.986
0.990
0.984
0.990
0.992
0.960
0.992
0.990
0.964
0.956
0.968
0.956
0.057
0.070
0.063
0.077
0.063
0.053
0.122
0.057
0.063
0.114
0.126
0.105
0.126
Y = 0.9799X + 0.4859
R = 0.99
60
50
40
30
20
10
20
40
60
80
20
40
60
80
High values of R2 in these figures show high power of the models in predicting (ET0) in the study area. Figure 5 also shows high similarity
between three methods of ANN (artificial neural network), genetic programming and FAO Penman equation for testing data.
Various statistics were used to compare the performance of neural network and genetic programming methods. Table 3 shows values of these
parameters for three methods of FAO-Penman-Monteith, artificial neural networks, genetic programming using testing data in Scenario 6.
70
ET0 (mm)
60
50
40
30
20
10
0
1
9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 67 69 71 73
Month
39
Table 3. Characteristics of the statistical results obtained from different methods of calculating (ET0) for testing data
Statistics
FAO-PenmanMonteith
Mean
Standard deviation
Min
Max
Correlated with one lag delay
25.50
18.89
2.93
62.48
0.793
Artificial neural
networks
Genetic
programming
25.48
18.59
2.88
59.47
0.784
25.91
18.89
1.99
70.71
0.785
As shown in Table 3, the results from both models fit well with the observed data (obtained from FAO-Penman-Monteith method). The equation
and function of genetic programming is shown in Equation 6.
ET
SunH
AvT -
AvT
AvT 3
SunH - AvT MinH
MinH
6.077 Wind 2. MaxH
(6)
In the above relation, SunH refers to sunshine hours, MaxH denotes maximum relative humidity, MinH refers to minimum relative humidity,
AvT denotes average temperature, Wind2 refers to wind speed at a height of 2 m above the ground. R2 and RMSE values for comparison of both
artificial neural networks and genetic programming methods are shown in Table 4. According to these results, it can be observed that artificial
neural network and genetic programming techniques can predict the potential evapotranspiration with high and ideal precision in meteorological
station in Tabriz.
Table 4. Results obtained from three different methods for calculating (ET0) in the 6. scenario
Method
ANN
GP
5.
R2
0.99
0.96
RMSE (mm)
1.69
3.84
Conclusion
In this study, data from meteorological station in Tabriz was used as input to predict monthly (ET0) using the artificial neural network and
genetic programming methods. The results from various scenarios in both models indicated that a scenario in which the model inputs include
average temperature, minimum and maximum relative humidity, wind speed, sunshine hours and the desired month deliver acceptable and high
precision. Due to the high value of the correlation coefficient obtained from ANN and genetic programming methods, it is concluded that both
methods estimate monthly (ET0) with high precision. Although genetic programming method is slightly less precise than neural network method
in estimating daily evapotranspiration, the former can be introduced as a competitive alternative to artificial neural networks in terms of
providing a mathematical relationship (Sayyadi et al.) [11]. Using data from the meteorological station of Tabriz, monthly (ET0) was estimated
with acceptable precision (R2=0.88) using the networks with radial basis function (RBF) and multilayer perceptron (MLP) networks. The
precision value obtained by the two methods is much higher than 0.88 in this study. Analysis of multiple scenarios showed that monthly index
improved the results in modeling (ET0) with artificial neural networks in contrast to perspiration. The latter had no tangible impact on
calculation of (ET0).
Acknowledgement
This project was performed with funds provided in the research budget of the Islamic Azad University of Tabriz, and the researchers conducting
the project would like to use this opportunity to express their gratitude and appreciation to university officials who supported them in carrying
out this study
References
[1]
Allen, R.G., Pereira, L.S., Raes, D., and Smith, M. (1998), Crop evapotranspiration. Guidelines for computing crop water requirements. Irrigation and
Drainage Paper No. 56, FAO, Rome, Italy, 300 pp.
[2] Blaney, H.F., and Criddle, W.D. (1950), Determining water requirements in irrigated areas from climatological and irrigation data. Soil Conservation Service
Technical Paper 96. 44 pp.
[3] Hargreaves, G.H., and Samani, Z.A. (1982), Estimating potential evapotranspiration. J. Irrig. and Drain. Engrg., ASCE, 108(3) 223230.
[4] Kouchakzadeh, M., and Bahmani A. (2005), Assessment of Artificial Neural Networks Revenue in Reducing Required Parameters for Estimation of
Reference Evapotranspiration, Journal of Agricultural Sciences (4) 87-97. Koza, J.R. (1992), Genetic Programming: On the Programming of Computers by
Means of Natural Selection. Cambridge, MA: The MIT Press.
[5] Zanetti, S.S., Sousa, E.F., Oliveira, V.P.S., Almeida, F.T., and Bernardo, S. (2007), Journal of Irrigation and Drainage Engineering. 133: 2(83).
[6] Kisi, O., and Ozturk, O. (2007), Adaptive neurofuzzy computing technique for vapotranspiration estimation. Journal of Irrigation and Drainage Engineering.
ASCE (133) 368-379.
[7] Rahimi Khoob, A. (2008), Artificial neural network estimation of reference evapotranspiration from pan evaporation in a semi-arid environment. Journal of
Irrigation Science, 27(1) 3539.
[8] Moghaddamnia, A., Ghafari Gousheh M., Piri J., Amin S., and Han, D. (2009), Evaporation estimation using artificial neural networks and adaptive neurofuzzy inference system techniques. Advances in Water Resources (32) 8897.
[9] Piri, J., Amin, S., Moghaddamnia, A., Keshavarz, A., Han, D., andRemesan, R. (2009), Daily Pan Evaporation Modeling in a Hot and Dry Climate. J.
Hydrol. Eng. 14, 803.
[10] Assari, M., Kouchakzadeh, M., Shahabifar, M., and Bayat, K. (2009), Estimation of Reference Evapotranspiration in Greenhouse by Artificial Neural
Network, Journal of Water and Soil Conservation (1) 107-121.
[11] Sayyadi, H., Oladghaffari, A., Faalian A., and Sadraddini A.A. (2009), Comparison of RBF and MLP Neural Networks Performance of Reference Crop
Evapotranspiration, Water And Soil knowledge (1) 1-12.
40
[12] Ghabaei Sough, M., Mosaedi, A., Hesam, M., and Hezarjaribi, A. (2010), Evaluation Effect of Input Parameters Preprocessing in Artificial Neural Networks
(ANNs) by Using Stepwise Regression and Gamma TestTechniques for Fast Estimation of Daily Evapotranspiration, Journal of Water and Soil (3) 610-624.
[13] Guven, K., Aytek, A., Yuce, M.I., and Aksoy, H. (2008), Genetic Programming based empirical model for daily reference evapotranspiration estimation.
Clean 36 (10-11): 905-912.
[14] Kisi, O., and Guven, A. (2010), Evapotranspiration modeling using linear genetic programming technique. Journal of Irrigation and Drainage 136(10): 715723.
[15] Terzi, O. (2010), Modeling of daily pan evaporation of lake Egirdir using data-driven techniques. International Symposium on Innovations in Intelligent
Systems and Applications. 320-324. Istanbul. Turkey
[16] Samadianfard, S., Sadraddini, AA., and Trajkovic, S. (2011), Estimating Reference Evapotranspiration in a case of Limited Weather Data Using Genetic
Programming. International conference "Innovation as a Function of Engineering Development". University of Nis, Faculty of Civil Engineering and
Architecture, Serbia.
[17] Shiri, J., and Kisi, O. (2011), Application of artificial intelligence to estimate daily pan evaporation using available and estimated climatic data in the
Khozestan Province (Southwestern Iran). Journal of irrigation and drainage engineering. 137 (7), 412425.
[18] Sattari, M.T., Nahrein, F., and Azimi, V. (2013), M5 Model Trees and Neural Networks Based Prediction of Daily ET0 (Case Study: Bonab Station). Iranian
Journal of lrrigation and Drainage, 7(1), 104-113.
[19] Koza, John R. (1992), Genetic Programming: On the Programming of Computers by Means of Natural Selection. Cambridge, MA: The MIT Press.
[20] Soltani, A., Gorbani, M.A., Fakheri Fard A., Darbandi, S., Farsadizadeh, D. (2011), Genetic Programming and Its Application in Rainfall-Runoff Modeling,
Water and Soil Science, 20(4). 61-71.