ANN-Based Reference Evapotranspiration Estimation: Effects of Data Normalization and Parameters Selection

Abstract

Accurate prediction of evapotranspiration (ET) is important in regions where agricultural plantations such as oil palm are abundant. This study was conducted in Peninsular Malaysia which has a large coverage of oil palm plantations that depend on rain-fed irrigation. This study attempts to improvise the estimation of reference evapotranspiration (ET0) to aid formulation of irrigation strategies. In order to obtain desirable estimations, data pre-treatment such as normalization and input selections are essential crucial steps that are needed. Therefore, it is an aim of this study to present the effect of normalization techniques on three specific ANN-based models for estimating ET0; namely the multilayer perceptron (MLP), radial basis function (RBF) and generalized regression neural network (GRNN). The case of different combinations of climatic parameters as input would be considered. Among the ANN models, the GRNN had the highest stability that could produce relatively stable performance regardless of the input combinations. Incorporation of normalization techniques prior to the training of the ANN-based models enabled diluting the effect of reduced input climatic parameters. For the MLP, the effect of normalization was minimal and insignificant. Selection of normalization technique of the RBF model should take the spread value of the model into consideration.