Main content area

Estimation of reference evapotranspiration in Brazil with limited meteorological data using ANN and SVM – A new approach

Ferreira, Lucas Borges, da Cunha, Fernando França, de Oliveira, Rubens Alves, Fernandes Filho, Elpídio Inácio
Journal of hydrology 2019 v.572 pp. 556-570
equations, evapotranspiration, irrigation scheduling, meteorological data, neural networks, relative humidity, solar radiation, support vector machines, temperature, weather stations, wind speed, Brazil
Reference evapotranspiration (ETo) is a variable of great importance for several purposes, such as hydrological studies and irrigation scheduling. The FAO-56 Penman-Monteith (FAO-56 PM) equation is recommended to estimate ETo given its good accuracy. However, the estimation of ETo poses a challenge when the availability of meteorological data is limited since the FAO-56 PM equation requires data on temperature, relative humidity, solar radiation and wind speed. This study evaluates, for the first time, the performance of alternative equations, artificial neural network (ANN) and support vector machine (SVM), for the estimation of daily ETo across the entirety of Brazil using measured data on temperature and relative humidity or only temperature. Two strategies, not yet used in Brazil, were used to develop the ANN and SVM models: (i) the definition of groups of weather stations with similar climatic characteristics, using the K-means clustering algorithm, to develop models specific for each group; and (ii) the addition of previous meteorological data as input for the models. Data from 203 weather stations distributed across Brazil were used. The ANN and SVM models showed higher performances than the equations that were studied, even when they were calibrated. The evaluated strategies (clustering and previous days) provided considerable performance gains. For the temperature-based models, the best performance was obtained by the ANN developed with the strategy of clustering and the use of data from two previous days as input; however, due to the similar performance and greater generalization capacity, the ANN developed without clustering and with the use of data from four previous days is recommended. For the temperature- and relative humidity-based models, the ANN developed with data from four previous days was the best option.