Accuracy Evaluation And Uncertainty Analysis On Five Global Remote Sensing Evapotranspiration Products

Evapotranspiration(ET)is a major component of terrestrial water cycle,carbon cycle and energy balance.It also has a profound impact on global climate change and terrestrial ecosystem.Therefore,accurate estimation of large-scale ET is crucial for climate change prediction,hydrological process research and drought warning.Remote sensing(RS)technology has the ability to obtain continuous large-scale underlying surface information,which greatly promote the research of large-scale ET model.Several global RS-based ET products have been invented.Comprehensive evaluation of these products at multi-scale can not only provide reference for scientific researchers to select suitable ET datasets,but also facilitate the further development of ET models.Both the accuracy and uncertainty of five global RS-based ET products(BESS,GLEAM,AVHRR,MOD16,SSEBop)were evaluated at multi-scale combining with flux observations,water balance analysis method and threecornered hat method.Mann-Kendall test was employed to analyze the trend of global ET and the differences among products.At point scale,the ET simulated by all products has good correlation with the observations of flux sites,and the correlation coefficient is between 0.59 and 0.79.The mean absolute error(MAE)and root mean square error(RMSE)of BESS product are the lowest among all products.However,it has obvious systematic deviation.By comparison,the overall accuracy of AVHRR and GLEAM performs better,MOD16 and SSEBop have the lowest accuracy.All ET products are able to describe general characteristics of seasonal variations of ET and perform similarly over same vegetation types,while there are great discrepancies in precision when estimating ET of different vegetation types.Moreover,BESS,GLEAM and AVHRR have higher accuracy in low elevation and relatively humid areas due to the influence of dry and wet conditions and elevation.At basin scale,the accuracy of ET estimation is higher in huge basins compared with small basins.The ET simulated by each product is highly correlated with ET derived by water balance analysis and the correlation coefficient values are between 0.88 and 0.98,among which BESS product has the highest accuracy,followed by SSEBop,AVHRR,GLEAM and MOD16.The results of uncertainty quantification indicate that BESS,GLEAM and AVHRR have lower uncertainties of interannual variations when estimating ET of huge river basins,with an average value of 11.48,9.99 and 14.94 mm/year,respectively.Their uncertainties of seasonal variations are less than 7 mm/month.MOD16 and SSEBop have higher uncertainties for interannual variations,which are 25.38 and 24.65 mm/year,respectively.Their uncertainties of seasonal variations are between 0 and 22 mm/month.Additionally,the seasonal and interannual variations of ET estimated by all products in the southern hemisphere basins have higher uncertainties.At global scale,the global average ET estimated by the five products for many years(2003?2013)is between 450 and 530 mm/year.The MK trend test illustrates that the global annual average ET of GLEAM and AVHRR has a significant increasing trend,while the other products have no obvious variations with the 0.01 level of significance.Furthermore,the spatial distribution of multi-year ET trend for all products indicates that there is no significant increasing or decreasing trend in most regions of the world.The discrepancies of ET trend for these products are mainly distributed in northern South America,central and northern Africa,high latitudes of Russia,Western Central Asia and parts of Eastern Europe.The results of uncertainty quantification demonstrate that BESS,GLEAM and AVHRR have lower relative uncertainties when simulating interannual and seasonal variations of global ET.The MOD16 product performs the worst when simulating the seasonal variations of southern hemisphere ET,and SSEBop product has the highest relative uncertainty among all products.Affected by elevation and land cover,the areas with large relative uncertainties are mainly distributed in plateau,mountainous area,bare land,grassland,farmland.