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 ).