Intercomparision of ANN, regression and climate based models for estimation of reference evapotranspiration

摘要

The present study investigates the applicability of linear regression (LR) and artificial neural network (ANN) models for estimating reference evapotranspiration (ET0) and their intercomparison with climate-based models on the basis of limited data availability in semi-arid environment of Solapur, Maharashtra, India. The eight climate based methods viz., Soil Conservation Service Blaney-Criddle, Thomthwaite, Hargreaves-Samani, Pan evaporation, Jensen-Haise, Priestly-Taylor, Turc, and Radiation were compared with Penman-Monteith (P-M) method for estimation of ET_0. The input combinations for all LR and ANN models were decided on the basis of climatic parameters required for selected climate-based methods. These are viz., Model 1 (evaporation), Model2 (T_(max) and T_(mim)), Model 3 (T_(max) and T_(min) and Sun Shine Hours - SSH), Model 4 (T_(max), T_(min), RH_(max), RH_(min) and SSH). The accuracies of the models were evaluated by using statistical criteria such as: coefficient of determination (R2), index of agreement d(IA), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient efficiency (CE), and ranking were assigned of models. All LR and ANN models showed satisfactory performance in development and validation stage and can be accepted to predict ET_0 values. The overall comparison of climate-based, LR and ANN models were carried out using the data from the year 1980 to 2014. The average weekly ET_0 values were estimated using climate-based, LR and ANN models and compared with those of P-M method. It was observed that ANN4 secured first rank and exhibited overall best performance with R~2 = 0.895, d(IA) = 0.972, RMSE = 0.508, MAE = 0.391, MAPE = 7.931 and CE = 0.894 followed by ANN3, LR4, LR3, ANN2, LR2, ANN1, and LR1, while all climate-based methods showed poorer performance than ANN and LR models. It was inferred that all LR models showed satisfactory performance for estimation of ET_0, however the performance has improved marginally with corresponding ANN models. Based on the overall results it was recommended that all ANN models can be used for the prediction of ET_0 followed by all LR models as per data availability and simplicity of users for Solapur region.