REFERENCE EVAPOTRANSPIRATION ESTIMATION WITH NEURAL NETWORKS AND ERA5 REANALYSIS IN DATA-SPARSE REGIONS

Abstract

Reference evapotranspiration (ET0) is a fundamental hydrological variable for quantifying crop water requirements, conducting watershed water balance assessments, and optimizing irrigation management. However, the limited availability of high-quality meteorological data hinders the accurate determination of ET0 across most regions of Brazil. Therefore, this study aimed to estimate ET0 in areas lacking meteorological observations by employing ERA5 reanalysis climate data and artificial neural network (ANN) models, as well as to evaluate its spatial and temporal variability. Observational data from 32 automated weather stations in the State of Mato Grosso, Brazil, were utilized to estimate ET0 using the Penman-Monteith FAO56 method. These estimates were linked to ERA5 meteorological data, including global solar radiation and top-of-atmosphere radiation, using ANN models. The Mann-Kendall test was applied to detect trends in the ET0 time series for the period 1980–2019. In Mato Grosso, the Pantanal biome exhibited the highest ET0 values from October to March, followed by the Cerrado (Brazilian Savanna) and Amazon biomes. Statewide, the peak evapotranspirative demand occurred in August, September, and October, whereas April, May, and June recorded the lowest values. All biomes in Mato Grosso contained areas with statistically significant increasing trends in ET0. The proposed models demonstrated satisfactory error metrics for ET0 estimation, and the methodological approach enabled a robust assessment of the spatiotemporal dynamics of this hydrological variable, facilitating ET0 estimation even in data-scarce regions.