————————————————–
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.