Aplicação de imagens satélitárias na modelagem espacial da evapotranspiração, mediante redes neurais artificiais, em áreas da Bacia do Rio Parnaíba Uruçuí–Preto, Brasil.

Abstract

This study developed an artificial neural network (ANN) capable of spatially estimating daily evapotranspiration (ETd) in the Uruçuí-Preto River basin, located in the northeastern region of Brazil. Input variables were derived from Landsat 8 satellite data, including reflectance from bands 2, 3, 4, 6, and 7, land surface temperature, albedo, instantaneous net radiation, and daily net radiation. The input variables exhibited correlation coefficients with ETd with absolute values of up to ±0.848, indicating the presence of dependency relationships that support its estimation through neural network processing. The ANN was trained using a dataset of ETd values obtained from the SEBAL algorithm, representing both agricultural and natural land covers of the Cerrado biome (savanna, grassland, and forest formations). Model performance evaluation resulted in a root mean square error (RMSE) of 0.02352 mm day⁻¹, a mean square error (MSE) of 0.018623 mm day⁻¹, and a coefficient of determination (R²) of 0.94, indicating a reliable estimation provided by the network. Residual errors did not exhibit abrupt or outlier behavior, being predominantly within ±5.0% and, to a lesser extent, within ±15.0%. Error trends indicate that higher ETd values—particularly in natural Cerrado areas—are associated with lower estimation uncertainty, whereas lower ETd values, typically found in agricultural areas, present higher estimation errors. In anthropogenic areas, the presence of bare soil or low vegetation density results in minimal or negligible evapotranspiration processes, which may not be fully captured by the ANN given the available input variables. The ANN approach reduces the need for iterative application of the SEBAL algorithm, enabling more practical and automated ETd estimation for similar environments and across different time periods.