————————————————–

4.2 Validation with the in-situ SSM measurements

The downscaled 1-km SSM of each algorithm were validated using the in-situ observations from 57 ground stations over the Naqu region within the available days from 2010 to 2014. Figure 6 shows the comparisons between ground observations and 1-km predictions of SVATARK, SVRK and SVRB. The SVRK model produces more accurate predictions than those of the SVRB-based SSM data with an RMSE value of 0.10 m3·m-3, a MAE value of 0.07 m3·m-3 and a SLOP value of 0.70, but with slightly smaller R value of 0.64. The comparison illustrates that the proposed SVATARK approach significantly outperforms the other two downscaling approaches with the smallest RMSE andMAE values of 0.08 m3·m-3 and 0.06 m3·m-3, the largest Rand SLOP values of 0.72 and 0.71. The scatterplot from the SVATARK approach visually gathers along the 1:1 line and has the lowest dispersion. Because of the model prediction error, errors in input variables and the representativeness errors of different supports, there are some discrepancies between the 1-km downscaled results and in-situ measurements. Although the spatiotemporal prediction approach can help fill the missing values of remote-sensed data, the errors from this process can be propagated into the final results. In future research, more error analyses, especially before downscaling, should be performed to improve the downscaling accuracy. The improvements made by SVATARK are illustrated by an increase in R (0.06 or 10.7% on average) and a decrease in RMSE (0.03 m3·m-3 or 23.6% on average) and slightly better MAE and SLOP values. A general improvement can be seen in Figure 7-9.