Hybrid of Artificial Neural Network-Genetic Algorithm for Prediction of Reference Evapotranspiration ( ET 0 ) in Arid and Semiarid Regions

Evapotranspiration is a principal requirement in designing any irrigation project, especially in arid and semiarid regions. Precise prediction of Evapotranspiration would reduce the squandering of huge quantities of water. Feedforward Backpropagation Neural Network (FFBPNN) model is employed in this study to evaluate the performance of Artificial Neural Networks (ANNs) in comparison with Empirical FAO Penman-Monteith (P-M) Equation in predicting reference evapotranspiration (ET0); later, a hybrid model of ANN-Genetic Algorithm (GA) is proposed for the same evaluation function. Daily averages of maximum air temperature (Tmax), minimum air temperature (Tmin), relative humidity (Rh), radiation hours (R), and wind speed (U2) from Mosul station (Nineveh, Iraq) are used as inputs to the ANN simulation model to predict ET0 values obtained using P-M Equation. The main performance evaluation functions for both models are the Mean Square Errors (MSE) and the Correlation Coefficient (R). Both models yield promising results, but the hybrid model shows a higher efficiency in prediction of Evapotranspiration and could be recommended for modeling ET0 in arid and semiarid regions.


Introduction
Increasing demand for water and scarcity of water supply are growing concerns in both arid and semiarid regions of the world.Iraq, which consists of both arid and semiarid climates, suffers from a low rainfall rate of about 160 mm per year and tremendous amount of water loss from evaporation and transpiration (Evapotranspiration).The condition is further aggravated by other climate events such as drought and salinity (Kerr, 1998).Native date palm, cotton, barley, and wheat are common products in Iraq that depend on irrigation because of scarcity of rain water.An understanding of the precise prediction of Evapotranspiration would allow for optimization of water use in irrigation projects.
The specific concept of Evapotranspiration (ET) is influenced by alteration of weather parameters, crop types, stage of growth, and other environmental conditions.The precise determination of ET produced the need for another comprehensive concept, namely, Reference Evapotranspiration (ET₀), which can be defined as "the rate of Evapotranspiration from an extensive surface of 8 to 15 cm tall, green grass cover of uniform height, actively growing, completely shading the ground and not short of water" (Doorenbos & Pruitt, 1977).The extensive surface resembles the reference surface indicated by FAO experts as the "A hypothetical reference crop with an assumed crop height of 0.12 m, a fixed surface resistance of 70 s m -1 and an albedo of 0.23" (Allen et al., 1998).Methods for predicting evapotranspiration are either Direct or Indirect methods.Direct Methods, such as Lysimeters (Wright, 1988), or field experiments, are mostly for scientific researches.Indirect Methods are usually empirical and depend on weather parameters such as solar radiation; air temperature; air humidity; and wind speed, measured or estimated at meteorological stations.Many empirical methods have been studied, issued, and applied in the prediction of ET, but the Food and Agriculture Organization (FAO) of the United Nations assumes that the combination of energy balance/aerodynamic equations provides the most accurate results for prediction of ET₀, and it adopted the FAO Penman-Monteith (P-M) Equation as the only standard equation for estimation of ET₀ (Allen et al., 1998).
Artificial Intelligence (AI) applications distinguished itself in many scientific fields during the last decades; however, researchers in hydrology, among many other fields, are still attracted to apply Artificial Neural Networks (ANNs) in modeling stream flow, rainfall, suspended sediment and ET.Kumar et al. (2002) found that application of Multiple Layers Perception (MLP) of ANN, with backpropagation algorithm; gives an accurate estimation of ET₀.In 2005, Trajkovic approved the possibility of achieving reliable results of ET₀ on the basis of temperature data only; FAO P-M Equation is used in comparison with another three temperature-based empirical equations and Radial-Based Function (RBF) Network.The RBF model better predicted FAO P-M ET₀ than the other calibrated empirical methods.Kisi (2006) investigated the possibility of modeling ET₀ using the technique of Generalized Regression Neural Network.The results of the intelligent model were successful.In 2007, Kisi and Öztürk investigated the accuracy of Adaptive Neurofuzzy Inference System Models (ANFIS) in modeling ET₀.Values of ET₀ were obtained using FAO P-M Equation with four years records of daily climate parameters of Pomona Station and Santa Monika Station, operated by the California Irrigation Management Information System (CIMIS), the results were compared with ANFIS and ANN, and also in comparison with Hargreaves and Ritchie empirical methods.The comparative results proved the superiority of ANFIS with inputs of T, U 2 , RH and R in modeling daily ET₀ over the ANN and empirical methods.Kisi (2008) examined the accuracy of three ANN techniques, namely, the Generalized Regression (GR), MLP and Radial Basis Neural Networks (RBNNs) in a model of P-M Evapotranspiration.Results proved that both MLP and RBNN techniques could be successfully used in modeling ET₀.Landeras et al. (2008) implemented Seven ANN techniques with different inputs and compared the results to ten empirical and semi-empirical ET₀ equations calibrated to FAO P-M equation using meteorological data as inputs.The comparisons criteria are the statistical error techniques; using PM56 daily ET₀ values as a reference.ANN techniques have obtained better results than the calibrated equations.El-Baroudy et al. ( 2010) compared the Evolutionary Polynomial Regression (EPR) to ANN and Genetic Programming (GP).The EPR model provided a performance comparable to that of GP and ANN models.Tabari and Talaee (2012) indicated that the main obstacle in application of FAO P-M Equation is the wide range of meteorological data essential as an input for calculation of ET₀, and because of the nonlinearity of ET phenomenon.They used a multi-layer neural network (MLNN) with variable inputs of data sets for modeling of ET₀ in the semiarid region in Hamedan, the model with all required climate parameters used as inputs performed best among other MLP models.Khoshhal and Mokarram (2012) evaluated different structures of MLP for estimation of ET₀, using meteorological data of Eghlid station in Iran for the period 2000-2010 as inputs and P-M ET₀, obtained using the same meteorological data, as output.The performance of 10 ANN models with different inputs were evaluated, the functions used for evaluation is root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R 2 ).The model with T min , T max , R h , R, U 2 is more accurate in predicting ET₀ than other models.
This study investigates the performance of ANN and a hybrid of Artificial Neural Network-Genetic Algorithm (ANN-GA) techniques in predicting ET₀ in comparison with values obtained using P-M Equation with historical data from Mosul meteorological stations in Iraq.Evaluation of the proposed Artificial Intelligence techniques will be carried out against estimations done using the empirical method.

Study Area
The study area, named Mosul, is located at the center of Nineveh Governorate, northern Iraq, between latitude 36° 22' 00" N, longitude 43° 07' 00" E, and altitude 222.6 m above sea level, with an overall area of 37,323 km 2 (Figure 1) (Statistical report, 2009).The weather data used in this study are obtained from the main meteorological station in Mosul (Global Station Code 608), which includes the daily averages of: maximum air temperature (T max ), minimum air temperature (T min ), relative humidity (R h ), radiation hours (R), and wind speed (U 2 ) from 1980-2005.www.ccsen

Penma
The

The mean
The binary the stored interval.

Results
The nonlin latitude an dependabl period 198 in Figure 3 system sho water acco     Figure 8 shows that the Hybrid ANN-GA simulation has modify the result of the Correlation Coefficient (R 2 ).It exhibited a higher consistency rather than using ANN by itself.
The main obstacle in the application of P-M Equation is the comprehensive and rich weather data needed for specific location or project in the design and operation stages.Latitude and altitude are to be specified as well; however, as the location specifications have only slight effect on implementation of intelligent models; these models would allow using historical records of weather data available in any alternative location for training the system and using the limited records available of the exact location for testing the system and even predicting actual values of ET₀ for any period.

Conclusion
ET₀ is the key parameter in designing any irrigation project.Building an intelligent model of ANN and Hybrid ANN-GA would provide an effective alternative to the empirical method used in the estimation of ET.The proposed Artificial Intelligence models have significant superiority as compared to the limitations embedded in the traditional empirical FAO P-M Equation method.This study shows that both ANN and Hybrid ANN-GA models can be used for prediction of ET₀ in Iraq, taking into consideration the unstable and sudden changes of weather conditions.
The normalization of input and output parameters would give the researchers a chance to overcome the limitations of application of relevant data to a specific location.In P-M Equation, both latitude and altitude values are considered properties of the study location and have great effect on ET₀ values.Normalization of input and output parameters in the FFBPNN and FFBPNN-GA models will eliminate the effect of those two parameters as they are constants; shared in all time intervals; the AI model could be recommended to be used without restrictions of location specifications.

Figure
Figure 8. T