A worldwide analysis of spatiotemporal changes in water balance‐based evapotranspiration from 1982 to 2009

Abstract A satellite‐based water balance method is developed to model global evapotranspiration (ET) through coupling a water balance (WB) model with a machine‐learning algorithm (the model tree ensemble, MTE) (hereafter WB‐MTE). The WB‐MTE algorithm was firstly trained by combining monthly WB‐estimated basin ET with the potential drivers (e.g., radiation, temperature, precipitation, wind speed, and vegetation index) across 95 large river basins (5824 basin‐months) and then applied to establish global monthly ET maps at a spatial resolution of 0.5° from 1982 to 2009. The global land ET estimated from WB‐MTE has an annual mean of 593 ± 17 mm for 1982–2009, with a spatial distribution consistent with previous studies in all latitudes but the tropics. The ET estimated by WB‐MTE also shows significant linear trends in both annual and seasonal global ET during 1982–2009, though the trends seem to have stalled after 1998. Moreover, our study presents a striking difference from the previous ones primarily in the magnitude of ET estimates during the wet season particularly in the tropics, where ET is highly uncertain due to lack of direct measurements. This may be tied to their lack of proper consideration to solar radiation and/or the rainfall interception process. By contrast, in the dry season, our estimate of ET compares well with the previous ones, both for the mean state and the variability. If we are to reduce the uncertainties in estimating ET, these results emphasize the necessity of deploying more observations during the wet season, particularly in the tropics.