Estimation of global land surface evapotranspiration and its trend using a surface energy balance constrained deep learning model

Estimating global land surface evapotranspiration (ET) is of great significance for assessing the impact of climate change on the global hydrological cycle and energy balance. In this study, we propose a surface energy balance constrained deep learning (DL-SEB) model for simulating global land surface evapotranspiration (ET). The accuracy of the DL-SEB model in estimating ET was tested using FLUXNET observations. The results suggested that the proposed DL-SEB model significantly enhanced the simulation capability of extreme ET events compared with the original deep learning model (without being coupled with the energy balance equation). The DL-SEB model was further applied to reconstruct global ET changes during 2000–2019 based on meteorological, soil, vegetation, and flux data sets. The annual average global land surface ET was 613 mm/yr during the period 2000–2019 (exclude Antarctica and deserts). The global land surface ET exhibited a significant upward trend with average increase rate of 1.16 mm/yr during the past two decades, which corresponds to approximately 3.8% increase above the mean global ET during 2000–2019. The positive trend of global land surface ET was driven by the combined effect of air temperature (Ta), soil moisture (SM), net radiation flux (Rn) and leaf area index (LAI). The natural climatic events such as El Niño events significantly altered short-term global ET variation, but did not changed the long-term increase trend of global ET. This study enhanced the understanding of the impact of climate change on the global land surface ET. The proposed DL-SEB model achieved a physics-based, smart and reliable ET simulation at global and regional scales.