Artificial Neural Network Model of Soil Heat Flux over Multiple Land Covers in South America

Soil heat flux (G) is an important component for the closure of surface energy balance (SEB) and estimation evapotranspiration (ET) by remote sensing algorithms. Over last decades, efforts have been focused on parameterizing empirical models G prediction, based biophysical parameters estimated sensing. However, due to existing models’ nature restricted conditions in which they were developed, using these large-scale applications may lead significant errors. Thus, objective this study was assess ability artificial neural network (ANN) predict mid-morning extensive meteorological reanalysis data over a broad range climates land covers South America. Surface temperature (Ts), albedo (α), enhanced vegetation index (EVI), obtained from moderate resolution imaging spectroradiometer (MODIS), net radiation (Rn) global assimilation system 2.1 (GLDAS 2.1) product, used as inputs. The ANN’s predictions validated against measurements 23 towers multiple cover types America, their performance compared that commonly models. Jackson et al. (1987) Bastiaanssen (1995) prediction calibrated tower quadratic errors minimization. ANN outperformed models, with mean absolute error (MAE) reductions 43% 36%, respectively. Additionally, inclusion information input reduced MAE 22%. This indicates structure more suited than can potentially refine SEB fluxes ET estimates