Actual evapotranspiration by machine learning and remote sensing without the thermal spectrum

The objectives of this study were to use machine learning algorithms to establish a model for estimating the evapotranspiration fraction (ET(f)) using two data input scenarios from the spectral information of the Sentinel-2 constellation, and to analyze the temporal and spatial applicability of the models to estimate the actual evapotranspiration (ET(r)) in agricultural crops irrigated by center pivots. The spectral bands of Sentinel 2A and 2B satellite and vegetation indices formed the first scenario. The second scenario was formed by performing the normalized ratio procedure between bands (NRPB) and joining the variables applied in the first scenario. The models were generated to predict the ET(f) using six regression algorithms and then compared with ET(f) calculated by the Simple Algorithm For Evapotranspiration Retrieving (SAFER) algorithm, was considered as the standard. The results possible to select the best model, which in both scenarios was Cubist. Subsequently, ET(f) was estimated only for the center pivots present in the study area and the classification of land use and cover was accessed through the MapBiomas product. Land use was necessary to enable the calculation of ET(r) in each scenario, in the center pivots with sugarcane and soybean crops. ET(r) was estimated using two ET(o) approaches (ET(o)Brazil and Hargreaves-Samani). It was found that the Hargreaves-Samani equation overestimated ET(r) with higher errors mainly for center pivots with sugarcane, where systematic error (MBE) ranged from 0.89 to 2.02 mm d(-1). The ET(o)Brazil product, on the other hand, presented statistical errors with MBE values ranging from 0.00 to 1.26 mm d(-1) for both agricultural crops. Based on the results obtained, it is observed that the ET(r) can be monitored spatially and temporally without the use of the thermal band, which causes the estimation of this parameter to be performed with greater temporal frequency.