————————————————–
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.