Evaluation and Comparison of GRNN, MLP and RBF Neural Networks for Estimating Cucumber, Tomato and Reference Crop's Evapotranspiration in Greenhouse Condition

In this study, the performance of multi-layer perceptron (MLP), radial basis function (RBF) and generalize regression neural networks (GRNN), were evaluated to estimate cucumber, tomato and reference crops’ evapotranspiration (ET) in greenhouse environment. For this purpose, the lysimetric ET values along with the effective meteorological factors on evapotranspiration process, including air temperature, humidity, air vapor pressure, and incoming radiation were measured. The results indicated that introducing the all meteorological factors as artificial neural network models’ inputs increased the models accuracy. Based on the results, the GRNN model estimated the reference evapotranspiration with the highest accuracy in comparison to the RBF and MLP models. The average estimation errors of MLP, RBF and GRNN networks for the test phase were 9.4, 13.3 and 9 percent corresponding to the values of 0.24, 0.27 and 0.20 mm d-1, respectively. Performance of GRNN model in estimating cucumber and tomato evapotranspiration values also was appropriate. The average values of the errors for the estimated amounts of the cucumber and tomato evapotranspiration by the GRNN model, were 11.0 (0.21 mm d-1) and 10.1 (0.22 mm d-1) percent, while they were obtained by MLP and RBF models 11.4 (0.22 mm d-1) and 10.9 (0.26 mm d-1); and also 12.3 (0.23 mm d-1) and 13.8 (0.28 mm d-1) percent for MLP and RBF models, respectively. Based on ideal point error index, the GRNN model showed accurate performance in estimating crop evapotranspiration in greenhouse.