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.