————————————————–
In the proposed SVATARK downscaling method, the values of MAE andRMSE at all five stations range from 0.033
m3·m-3 to 0.065
m3·m-3 and from 0.041
m3·m-3 to 0.076
m3·m-3, respectively, where SVATARK
is found to be more accurate than the other two approaches. From the
time series comparisons at different ground stations (Figure 10(a-e)),
the downscaled SSM predictions, especially of the SVATARK method, show
temporal consistency with the in-situ observations, although there is a
significant bias between them. This indicates that the downscaled SSM of
SVATARK can describe the temporal changes of the in-situ SSM. The
discrepancies are mainly because of the large scale differences between
1-km predictions and point observations. The best performance was
obtained at Station D, which has the lowest MAE of all stations.
The variation in performance of various stations might be the result of
the station’s location, which would affect the soil type and have
different accuracy of the input variables. The range of values for the
downscaled SSM are almost all less than the ground measurements’ range.
This matches well with the fact that the range of SSM decreases
dynamically from fine to coarse scales (Abbaszadeh et al., 2019).
Areal-averaged downscaled SSM agrees well with the ground-based SSM in
Figure 10(f). However, the discrepancies between three downscaled SSM
and ground observations in the network area seem smaller than those of
the stations, particularly for the SVRK and SVRB methods, perhaps due to
the comparisons at the same scale. The errors associated with upscaling
the SSM from 1-km and point scale to network scale require further
research and exploration. Note that a better performance of the
downscaled SSM is also obtained by using SVATARK.
Soil moisture is a direct indicator of agricultural drought. In-situ
observations of SSM may not be able to assess drought conditions in a
region, whereas the 1-km downscaled predictions could provide powerful
data support. A simple relative drought analysis was attempted using the
downscaled SSM of SVATARK. The pixels with anomalously low values in the
downscaled images were counted by comparing the pixel values on the same
date every year. The mean and standard deviation were calculated for
each pixel. Pixels with a larger absolute value than the standard
deviation were considered to be in a drought condition. The main idea
behind this assumption is to find the pixel which has a low value and
relatively large variation in SSM over the same period. Figure 11 shows
the proportion of pixels with relatively smaller SSM in the downscaled
images using SVATARK during study period. Several proportions are larger
than 0.30, meaning that thirty percent of the 1 × 1 km pixels in the
corresponding date have abnormally low values. The proportion values
suddenly increase in mid-July 2015, indicating relative drought
conditions. These results are consistent with Zhu et al. (2016).
Although the SSM at 0-5 cm depth might have a limited ability to reflect
soil drought without deep soil moisture, this preliminary attempt
demonstrates that the proposed downscaling method could be used in
drought remote sensing monitoring applications for a large area. The
downscaled SSM could also help understand how often and where these
droughts occur. In this study only four types of ancillary variables
were employed, but rainfall (including its infiltration and runoff) also
affects SSM variations, and should be explored as an ancillary variable
in the downscaling process in future work.