————————————————–
To explore the spatial distribution of the estimation errors, theMAE values from 2010 to 2014 of the 57 ground stations were
calculated. Figure 9 visualizes the MAE of the downscaled results
using SVATARK, SVRK and SVRB for each ground station with color bars.
The MAE values tend to be higher in the upper left and middle
part, likely due to higher topographic relief and the lack of the
corresponding original remotely-sensed observations, which could
introduce errors from filling gaps. The SVATARK model has the smallestMAE values at each station, suggesting a better performance than
SVRK and SVRB. Further SSM analyses at each station are shown in section
4.3. Although the above validations were all taken at stations of
grasslands, which is the main vegetation type in the Naqu region, the
SVATARK method could theoretically result in accurate downscaling
predictions from other areas given its ability to learn for small
samples and the strong generalization of SVR, as well as the coherence
of ATAK. The proposed method should be validated and applied to other
land cover types in future work.