Comparison of neural network models for estimation of reference crop evapotranspiration

Abstract

Reference evapotranspiration (ETo) is important component in calculation of crop water requirements, climatological and hydrological studies. However, it is a nonlinear dynamic and complex process. Artificial neural network (ANN) techniques are powerful to accurately map complex and nonlinear input output relationships. Penman-Monteith (P-M) model is sole standard method for estimation of ETo but it requires many kinds of weather data. Hence it was planned to test an ANN technique for estimating the ETo with limited and full data for Kolhapur region, Maharashtra, India. The five ANN models were developed by different combinations of meteorological parameters as neurons in input layer with varying neurons in middle layer and one neuron in the output layer i.e. ETo by P-M method. These are ANN1 (Pan evaporation); ANN2 (Tmax. and Tmin.); ANN3 (Tmax.,Tmin and SSH); ANN4 (Tmax., Tmin., RHmax., RHmin., and SSH) and ANN5 (Tmax., Tmin., RHmax., RHmin., SSH and WS). During training mode, it was observed that ANN5 model showed the best values of all performance measures i.e. correlation coefficient (0.997), index of agreement (0.998), efficiency coefficient (0.993),root mean square error (0.100), mean absolute error (0.059), mean absolute percentage error (2.076) due to full data. Under limited data condition, it was observed that the results of all performance measure (ANN1 to ANN4) models varied in the range as correlation coefficient (0.929 to 0.967), index of agreement (0.962 to 0.983), root mean square error (0.317 to 0.460), mean absolute error (0.194 to 0.297), mean absolute percentage error (6.027 to 9.679) and efficiency coefficient (0.863 to 0.934) in training mode. It can be seen that all limited data models demonstrate relatively very close performances based on statistical criteria. Similar kind of close differences for each performance measure were observed during validation stage of all ANN models. Results indicated that all ANN models satisfy performance criteria well in training and validation mode and can be generalized for prediction of ETo values. Overall, the performance suggests that ANN models can be an accepted for accurate prediction of ETo values for Kolhapur region as per data availability.