University of Tabriz Water and Soil Science 2008-5133 19 1 2009 06 22 Evaluation of Two Artificial Neural Network Software in Predict of Crop Reference Evapotranspiration Evaluation of Two Artificial Neural Network Software in Predict of Crop Reference Evapotranspiration 201 212 1487 FA H Zare Abyaneh A Gasemi M Bayat Varkeshi K Mohammadi AA Sabziparvar Journal Article 2014 06 15 In this study, the performance of two different artificial neural network software's named neuro solution (NS) and neural works professional II (NW) in estimation of crop reference evapotranspiration (ET₀) were evaluated. For models evaluation, some statistical parameters such as root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R²) were calculated for different arrays, learning rules and transfer functions. For the NS software the best fitted array characterizing with lowest values of RMSE, MAE and highest R² were found to be 0.08, 0.07 (mm day^-1) and 0.87, respectively. Results showed that the NS software with the best fitted network array of: learning rule of conjugate gradient and transfer function of sigmoid type, which required shorter computational time and less iteration loops, can perform better prediction. The results indicated that using two hidden layers did not improve the accuracy of ET₀ predictions, in comparison with the results obtained by one hidden layer layout. The sensitivity analysis of neural network model revealed that ET₀ is very sensitive to maximum air temperature (T_max). In contrast, the estimated daily ET₀ showed the lowest sensitivity to minimum relative humidity (RH_min).  In this study, the performance of two different artificial neural network software's named neuro solution (NS) and neural works professional II (NW) in estimation of crop reference evapotranspiration (ET₀) were evaluated. For models evaluation, some statistical parameters such as root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R²) were calculated for different arrays, learning rules and transfer functions. For the NS software the best fitted array characterizing with lowest values of RMSE, MAE and highest R² were found to be 0.08, 0.07 (mm day^-1) and 0.87, respectively. Results showed that the NS software with the best fitted network array of: learning rule of conjugate gradient and transfer function of sigmoid type, which required shorter computational time and less iteration loops, can perform better prediction. The results indicated that using two hidden layers did not improve the accuracy of ET₀ predictions, in comparison with the results obtained by one hidden layer layout. The sensitivity analysis of neural network model revealed that ET₀ is very sensitive to maximum air temperature (T_max). In contrast, the estimated daily ET₀ showed the lowest sensitivity to minimum relative humidity (RH_min).  https://water-soil.tabrizu.ac.ir/article_1487_5c62a6870af818d7403d62f08860de17.pdf