Performance evaluation of numerical and machine learning methods in estimating reference evapotranspiration in a Brazilian agricultural frontier

Abstract

The reference evapotranspiration (ET0) estimates is important for water resources and irrigation management. The Penman-Monteith equation is known for its accuracy but requires a high number of climatic parameters that are not always available. Thus, this study aimed to evaluate the performance of machine learning techniques (cubist regression, artificial neural network with Bayesian regularization, support vector machine with linear kernel function) and stepwise multiple linear regression method to estimate daily ET0 with limited weather data in a Brazilian agricultural frontier (MATOPIBA). Climatic data from 2000 to 2016 obtained from 23 weather stations were used. Five data scenarios were evaluated: (i) all variables, (ii) radiation and temperature, (iii) temperature and relative humidity, (iv) wind speed and temperature, and (v) temperature. The results showed that the machine learning methods are robust in estimating ET0, even in the absence of some variables. Among the methods evaluated using only temperature data, the cubist regression showed better performance. When estimating water demand for soybean and maize crops using only temperature, the cubist regression and calibrated Hargreaves-Samani equation showed the smallest errors.