2.4 Downscaling strategy

LST, Normalized Difference Vegetation Index (NDVI), land cover (LC), Blue Sky Albedo (BSA), Digital Elevation Model (DEM), aspect and slope were used as ancillary variables to downscale the CCI SSM product over thirty-six months (during May to October, 2010-2015). Considering the relatively low coverage of daily remotely-sensed observations, the 8-day composites of all variables were employed by using average aggregation to maintain stability and representativeness of each variable. A spatial-temporal prediction method (Gerber et al., 2018) was adopted to replace the missing values for LST and BSA due to cloud cover. Prior to performing the downscaling algorithm, a bias correction step (Djamai et al., 2016) was used for remotely-sensed SSM data to reduce the influence of the original SSM product. The downscaling procedure is shown in Figure 1, including the downscaling and validation processes.
After all data processing steps, including resampling aggregation, gap filling and bias correction, the 25-km and 1-km variables with full spatial coverage for each 8-day period was achieved, as well as the 8-day in-situ measurements within 1 km × 1 km grids. Three downscaling methods were implemented, involving the proposed SVATARK method and two benchmark methods. The two benchmark methods interpolate coarse regression residuals by applying kriging and bilinear interpolation, denoted by SVRK and SVRB respectively. In the trend prediction process, the SVR models were established for each land cover type. In the experiments, SSM values for water bodies and permanent snow and ice were not included. The downscaled SSM were validated by ground-measured SSM with four classical statistical metrics, including correlation coefficient (R ), mean absolute error (MAE ) (m3·m-3), root mean square error (RMSE ) (m3·m-3) and slope (SLOP ).