————————————————–
In addition to the representativeness errors, there is a bias between
the coarse SSM product and the in-situ observations. To reduce the
negative effect of the bias in the downscaled prediction accuracy, in
this experiment the mean difference between the 25-km SSM values and the
point supports was removed before downscaling. Considering that all the
ground stations were installed on grasslands, this bias step was used
without consideration of the differences resulting from topography and
LC types. However, the remotely-sensed product might perform differently
over various surfaces and therefore incorporation of the impact of LC
types may be beneficial for improving the bias correction accuracy. In
addition, how to determine whether bias needs to be applied to all
coarse grids is still a problem, if the SSM can be effectively observed
at some grids.
4.3 Dynamic analysis of downscaled
SSM
The downscaled maps and validation analyses described in this study
illustrate that the downscaled SSM results generally show a good
performance compared with ground-based measurements and their spatial
pattern follows those of the coarse SSM. In this section, we investigate
whether the fine-resolution SSM predictions from the three downscaled
methods also capture the temporal dynamics of ground-based SSM
observations during the study period. Figure 10 shows the temporal
variations of 1-km downscaled SSM derived from all three downscaling
methods and in-situ observations at five ground stations (in Figure 2
and Figure 9) and at network scale (i.e., network area in Figure 2).
There is a significant seasonal variation in the time series, generally
reaching its highest value in August. By using average aggregation
within the network domain, the aggregated values of ground measurements
and 1-km SSM were obtained. Two statistical metrics, MAE andRMSE, were used to evaluate their performance.