Evaluating an energy balance setting and random forest-based downscaling for the estimation of daily et at sub-kilometer spatial resolution

A large number of processes are impacted by the magnitude and timing of water fluxes. A proper monitoring of these processes would benefit from periodic estimation of evapo-transpiration (ET) with enough spatial detail to discriminate the different components of the landscape. However, high temporal and high/moderate spatial resolution are rarely present in the same ET product. In this study, we aimed at merging the strengths of geostationary and polar-orbit satellites to estimate daily ET at the native spatial resolution of Proba-V, 0.3 km. Two modelling paths were examined: the conduction of an energy balance at the target spatial resolution, and; energy balance at a coarser resolution followed by a statistical downscaling procedure based on building a Random Forest. The study area was the upper catchment of the Biebrza river in North Eastern Poland. The comparison of the modelled values with ET derived from in situ measurements (eddy covariance) indicate that both approaches performed good (Nash-Sutcliffe efficiency coefficient of 0.7589 and 0.7857 for first and second approach, respectively). The modelled spatial and temporal patterns were similar but the statistical downscaling path tended to generate lower ET estimates (bias equal to 0.0071 and -0.2254 for first and second approach, respectively).