Machine learning models applied in the estimation of reference evapotranspiration from the Western Plateau of Paulista

Abstract

Evapotranspiration depends on the interaction between meteorological variables (solar radiation, air temperature, precipitation, relative humidity and wind speed) and phytosanitary conditions of agricultural crops. It is complex to build reliable evapotranspiration measurements due to the high costs of implementing micrometeorological techniques, in addition to difficulties in the operation and maintenance of the necessary equipment. The purpose of this research was to model the reference evapotranspiration through machine learning techniques in climatic data from 30 automatic weather stations in the Planalto Ocidental Paulista, State of São Paulo, Brazil, in the period 2013-2017. A comparison of the statistical performance between the techniques used was carried out, where the best performance of the EToMLP4 model (rRMSE = 0.62%), followed by EToANFIS4 (rRMSE = 0.75%), EToSVM4 (rRMSE = 1.19%) and EToGRNN4 (rRMSE = 11.05 %). Performance measures of the validation base show that the proposed models are able to estimate the reference evapotranspiration, with emphasis on the MPL technique.