Modelling the daily reference evapotranspiration in semi-arid region of South India: A case study comparing ANFIS and empirical models

Abstract

The estimation of evapotranspiration (ETo) is one of the main tools for the control of crop growth and to make a rational use of water resources. To estimate this parameter accurately, it is necessary to have a daily measurement of four meteorological variables, these are: temperature, solar radiation, relative humidity and wind speed. It is not always possible to count on all the variables, that is why there are empirical methods that use a limited number of variables that make an approximate estimate of the ETo value. Each of these models are applicable to different regions with completely different climates. In this paper, a study has been carried out to define the model of ETo estimation that best adapts to the semi-arid region in South India. Two different datasets for the same period from different meteorological stations were used. In addition to the empirical methods for estimating ETo, computer models ANFIS (Adaptive Neuro-Fuzzy Inference System) were implemented. These models consist in the future estimation of a certain parameter by using current variables and a history of variables and past results. The results of this work show that ANFIS 11 model makes the best estimate with RMSD = 0.002 and r = 0.999. The RITCHIE method is the most suitable empirical model for this region, which reaches RMSD = 0.507 and r = 0.851. In addition, ranking of equations is elaborated for both datasets for daily estimates of ETo. Finally, comparison is made with the results for each case and thus confirm or reject the convenience of one model over the rest. To achieve this, a series of statistical indicators were used: Index of agreement (d), MAE (Mean absolute error), SEE (Standard error of estimate) and RMSD (Root mean square difference). Moreover, a sensitivity analysis was performed in order to compare and show the stability of the best models when an error is introduced within the input parameters. In this case, the empirical models demonstrate a better performance than the ANFIS models. This work demonstrates that the Ritchie method is a good estimator of the ETo value for a semi-arid region in southern India. In addition, the results of the ANFIS models are promising and could be used as estimation methods.