Estimation of daily evapotranspiration using machine learning models in China: based on ChinaFLUX sites

ABSTRACT Evapotranspiration (ET) is an important component of the hydrological cycle and its accurate estimation is very necessary for water resource management and agricultural precision irrigation. The direct measurement of ET is difficult and the observed data are very limited. Therefore, this study explores the possibility of a back-propagation (BP) neural network for simulating ET based on eight flux sites in China. The input variables consist of land surface temperature, the water vapor pressure of the land surface, net radiation, and photosynthetically active radiation. Four input combination categories are determined, including one input variable, two input variables, three input variables, and four input variables. The results demonstrated the following: (1) Adding more input variables generally improves model performance, with a significant gain from one to two variables, but only slight improvements from two to three. (2) While some stations show better performance with four variables, others perform best with three. (3) Daily and monthly ET estimates are achieved across all stations, with summer estimates consistently outperforming winter ones. (4) The variability in the best input combination for the stations indicates that factors such as climate zones and land cover influence ET accuracy.