Abstract
Many satellite-derived evapotranspiration (ET) estimates rely on coarse resolution (CR) land surface temperature (LST) from 1-km thermal infrared bands offered by NASA and ESA instruments, like Terra MODIS and Sentinel-3 SLSTR. This affects prediction performance of ET, especially in complex regions, such as the Alps. Since most two-source energy balance (TSEB) models assume no thermal variability within a pixel, a major challenge in ET modelling is related to cell grid heterogeneity. Given this limitation, we investigate the potential of kernel-driven downscaling to obtain sub-kilometer LST products based on fine resolution (FR) sensors, i.e., Sentinel-2 MSI and MODIS VNIR, for estimating TSEB-based ET over South Tyrol, in the South-Eastern Alps. To this aim, we exploit relationships between CR LST and FR predictors using trees-based algorithms. Due to reduced capabilities of univariate models in complex ecosystems, multi-source predictors are considered, including multispectral reflectances, spectral indices, solar radiation, and topography. The performance of the TSEB model driven by disaggregated outputs is evaluated against original 1-km LST and ground-based fluxes from two eddy covariance towers. In general, turbulent fluxes forced with downscaled LST resulted in RMSE of 86 Wm-2 and mean bias of 55 Wm-2, which translated to 8% and 15% decrease in the respective estimates when compared to TSEB results with 1-km LST. Despite some limitations related to small-scale heterogeneity in landcover and topography, the enhanced LST showed potential for providing ET at finer spatial resolution in heterogenous ecosystems.