Use of Ensemble Methods for Improving Accuracy of Remotely Sensed-derived Actual Evapotranspiration of Global Databases Case Study: (Karkheh Dam Watershed)

Actual Evapotranspiration is one of the effective components in the hydrological cycle. Therefore, accurate estimation of this component at the watershed scale has an important role in the management of available water resources. In this study, the accuracy of actual evapotranspiration product values of five global terrestrial databases including MOD16, MYD16, SSEBOP, GLEAM, GLDAS was compared to the evapotranspiration values predicted from the SWAT model simulation in the Karkheh dam watershed in 2006, 2008 and 2011 which are low/high and medium rainfall respectively on a monthly scale. Then, the feasibility of improving the accuracy of evapotranspiration values obtained from these databases was investigated in eight different scenarios using simple averaging, M5 and SVR models as ensemble methods. The results showed that although actual evapotranspiration products are able to explain the trend of time changes of actual evapotranspiration in catchment, they are significantly different with the output values of the SWAT model as observation values at a significant level of 5%. The results also indicated the use of simple averaging has no effect on improving the results at the Karkheh dam watershed. However, the use of the other two ensemble methods improves the accuracy of actual evapotranspiration and the ensemble method explains 80% of the SWAT-derived actual evapotranspiration variations. Moreover, the ensemble model derived from SVR fed by the attributes of the superior data combination scenario, reduced the estimation error by about 44% compared to that derived from the best global terrestrial product which was GLEAM in this study.