Abstract
Process of evapotranspiration (ETo) is a major component of the hydrologic cycle that its accurate estimation plays an important role to achieve sustainable development in water balance, irrigation system design and planning and management of water resources. Being a function of different metrological parameters and their interactions, evapotranspiration is a complex, nonlinear phenomenon. Preprocessing input parameters to select appropriate combinations is complex when modeling nonlinear systems. Using these methods reduce steps by trial and error by recognizing most important parameters for modeling by intelligent methods. In this study, two methods of stepwise regression (FS) and gamma test (GT) were used for preprocessing input parameters in multi layer perceptron neural network (MLP) to estimate daily estimation of ETo at Shiraz synoptic station. To evaluate the effect of Input parameters preprocessing in artificial neural networks using different statistical error criteria, ANN-FS and ANN-GT both with pre-processed parameters were compared against each other and also with ANN model without pre-processed parameters. The results showed that all three models have a high degree of accuracy to estimate daily ETo. ANN-GT model represented a determination coefficient (R2) of 0.9995 and the root mean square error (RMSE) of 0.0483. For ANN-FS and ANN models R2 values were 0.9984 and 0.9994 respectively and RMSEs were 0.0874 and 0.0548 respectively. Although the accuracy of ANN-GT model was slightly greater than ANN, but the ability of determination of importance of input parameters, education and recognition of the best combination of input parameters with 800 data in this study, makes this model a useful tool for fast preprocessing input parameters to model evapotranspiration.

Keywords: Potential evapotranspiration, Artificial neural networks, Gamma test, Stepwise regression, Shiraz synoptic station