Generation of spatio-temporally continuous evapotranspiration and its components by coupling a two-source energy balance model and a deep neural network over the Heihe River Basin

Evapotranspiration (ET) and its components of soil evaporation (E) and vegetation transpiration (T), as key variables for the water-energy exchange between the land surface and the atmosphere, are widely used in hydrological and agricultural applications. The land surface temperature based two-source energy balance (TSEB) model can provide high accuracy E, T and ET, which are spatio-temporally discontinuous, whereas the spatio-temporally continuous daily ET is more helpful in water resources management. In this study, to improve the continuity of estimates from the TSEB model, we developed a new combined model coupling the TSEB model and deep neural network (DNN) (TSEB_DNN). First, spatio-temporally continuous reference data was prepared based on the remote sensing and meteorological data as input, and E from soil and T from vegetation were obtained from the TSEB model under clear-sky condition as outputs. Then, the DNN was trained under clear-sky condition to obtain the relationship between E and T estimates from TSEB and reference data. Finally, the trained DNN was driven by the spatio-temporally continuous reference data to obtain spatio-temporally continuous E, T and ET. Compared with the ET estimates from the original TSEB model, the continuity was significantly improved for the TSEB_DNN model. The TSEB_DNN model was well consistent with the in situ measurements and had the overall correlation coefficient (R), root-mean-square-error (RMSE), and bias values of 0.88, 0.88 mm d(-1), and 0.37 mm d(-1), respectively. The ratio of T/ET estimates from the TSEB_DNN model had high accuracy against in situ measurements with RMSE and bias values of 7.49% and -2.22%, respectively. The combined model and the maps of E, T and ET will help improve water resource management.