Ensemble evapotranspiration estimates and uncertainties: EVASPA

Quantifying evapotranspiration (ET) beyond the local scale is essential for many water-related studies. Compared to in-situ instruments, Remote sensing (RS) has allowed the continuous monitoring of ET at larger spatial scales. By exploiting the physical relationship between remotely sensed surface biophysical parameters and the Earth’s thermal emission, continuous ET at such spatial scales can be obtained. In this study, we applied EVASPA, a tool that provides an ensemble of ET estimates, among other surface energy balance (SEB) variables, from various sources of data and several algorithms. Here, we applied MODIS data, which included: Land Surface Temperature/Emissivity (LST/E), NDVI, albedo, among others. Landsat data was separately applied for estimates at relatively high spatial resolution. Our multi-data multi-method approach resulted in 1215 ET estimates for the MODIS-based ETs (i.e., 5 LST/E (MYD/MOD 11/21 and VIIRS 21); 3 radiation sources (ERA5Land, MSG, MERRA); 9 Evaporative Fraction methods (5 S-SEBI based, 4 T-VI based), and 9 Ground heat flux methods (based on NDVI and LAI)). Evaluations using in-situ flux data yielded reasonable results even when a simple average was used (for example, RMSE of ~0.9 mm/d over the forested Puechabon site), with a broad absolute and performance range between the member estimates being observed (for instance, an ensemble RMSE range of ~0.6 to ~1.2 mm/d for the best-to-worst performing EVASPA members over the Puechabon site). Uncertainty analyses were also performed where we analysed how each of the distinct variables (i.e. radiation, LST, EF and G methods) influenced the modelled ET. Irrespective of the combination criteria selected, LST and EF were observed to be the main uncertainty drivers; this was despite instances where radiation resulted in higher uncertainties that were dependent on the combination selected and/or the period of simulation. G flux methods exhibited the least influence on the ensemble simulations. Overall, we showed that ensemble-based contextual modelling can provide enough spread for better flux simulations. This work aims to guide the establishment of an optimal weighting criteria of the members for improved ET estimates. The EVASPA algorithms will be used for providing ET estimates in the frame of the Indo/French future mission TRISHNA to be launched by the end of 2026.

Keywords: ET, SEB, contextual ET, multi-method multi-data, ensemble modeling.