5 Conclusions

In this study, we proposed a machine learning-based geostatistical downscaling method. The proposed SVATARK relies on SVR that expresses the nonlinear relationship between target (i.e., SSM) and ancillary variables (i.e., LC, LST, NDVI, BSA and terrain factors), and utilizes ATAK to achieve the predictions on changed supports. SVATARK was compared to the benchmark methods SVRK and SVRB to obtain 1-km predictions from a 25-km SSM product over the Naqu region during a thirty-six month period from 2010-2015. The downscaled predictions were validated using ground stations. In general, the comparison results indicate that the SVATARK downscaling approach obtained the greatest accuracy, and the dynamic analysis of 1-km SSM reached the same conclusion. The downscaled predictions were used to capture the abnormally low SSM by using a simple count analysis, which reveals the capability of monitoring the relative drought and could be further generalized for large areas with systematic analysis methods. The proposed SVATARK method is entirely general, and it can be employed to downscale or even upscale other continuous variables owing to the changes in the supports in ATAK. Other machine learning or deep learning methods such as such as random forest or neural network algorithms could be applied in trend predictions and could be integrated with ATAK for spatial scaling. The comparisons among different artificial intelligence algorithm-based ATAK models will be explored in the future work.