————————————————–
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.