Evaluation of pan coefficients from ANN, ANFIS, and empirical methods, for estimation of daily reference evapotranspiration

Introduction: Evaporation and evapotranspiration are two major components of hydrological cycle which are very important for agricultural studies as well as water resources management. So far, various methods have been addressed for the estimation of daily reference crop evapotranspiration (ET0). Among them, FAO Penman-Montieth 56 (PMF-56) (Allen et al., 1998) is widely used as a standard method, particularly for arid and semi-arid regions. The major drawback to this method is the fact that the required weather data are not usually available for majority of the study sites. Pan coefficient (Kpan) method (Eq.1) is an alternative procedure which can be used for such conditions.
ET0 = KPan (1)
To estimate the pan coefficient (KPan), many works have been performed by various researchers (e.g. Cuenca, 1989; Allen and Pruitt, 1991; Snyder, 1992; Orang , 1998; Ranghuwanshi and Wallender, 1998). In a humid region, Imark et al (2002) used Frevert et al (1983) and Snynder (1992) methods for estimation of KPan and ET0. They suggested that Frevert method performs more reliable ET0 estimation close to PMF-56 than other Kpan estimators. Sabziparvar et.al (2009) introduced Orang (1998) and Ranghuwanshi and Wallender (1998) methods as the most accurate Kpan estimators for warm arid and cold semi-arid regions of Iran.
Another research work conducted for north Spain suggested that the daily ET0 which estimated by Artificial Neural Network (ANN) method performs more accurate results than empirical and semi-empirical relations.
Aims and Scope: The main aim of this study is to assess Kpan values as estimated by ANN and ANFIS (Adaptive Neuro-Fuzzy Interface System) predictors against empirical estimators such as Cuenca, Orang , Pereira , Snyder , Raghuwanshi and Wallender methods. The comparison is made by using the statistics such as coefficient of determination (R2), Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).
Summary and conclusions: In this research, the performance of different pan models (Cuenca, 1989; Allen and Pruitt, 1991; Snyder, 1992; Orang , 1998; Ranghuwanshi and Wallender, 1998) for better estimation of pan coefficients for the selected sites in warm arid climate of Iran is compared with Artificial Neural Network (ANN) and Adaptive Neuro-Fuzzy Interface System (ANFIS) results. For this purpose, a ten-year daily measured pan data (1996-2005) were used. Having considered the shortage of lysimeter data, we applied the FAO recommended alternative approach of Penman-Montieth FAO-56 for the determination of daily ET0 at the study sites. In ANN and ANFIS methods, wind speed, relative humidity and fetch distance were applied as the input of the verified networks. The pan coefficients as estimated by PMF-56 method were also used as the input of the intelligent networks. Model validation was presented by using Root Mean square Error (RMSE), coefficient of determination (R2), and Mean Absolute Error (MAE) criteria. The results showed that the ANFIS method performs more accurate pan coefficient and reference daily evapotranspiration values compare to other approaches.
For the selected ANFIS method, the mean values of R2 , RMSE and MAE were 0.83, 0.97 (mm day-1) and 0.74 (mm day-1) respectively. Among the empirical pan models, Cuenca and Snyder methods are recommended for prediction of Kpan in warm arid climates.
In this work, we assumed that the synoptic weather sites of Shiraz and Kerman are representatives of agricultural fields in warm arid climate. This assumption might affect the estimated Kpan values. To remove some weaknesses, of the fetch distances, using evaporation pan and lysimeter instruments inside the agricultural fields, in addition to other weather instruments such as wind recorder (anemometer) are recommended for more reliable results.