1 Introduction
Land surface soil moisture (SSM) is an essential hydro-ecological
parameter for monitoring energy, water, and carbon cycles (Seneviratne
et al., 2010; Bateni and Entekhabi, 2012). Continuous SSM at fine
spatial resolutions provides crucial information for hydrological
models, precipitation forecasting models, land-atmosphere models,
drought and flood forecasting, and vegetation growth monitoring
(Krishnan et al., 2006; Wang et al., 2016; Dorigo et al., 2017). In
general, soil moisture is acquired by using in-situ measurements
(Dobriyal et al., 2012), including wireless sensor networks (Kerkez et
al., 2012) and Cosmic-ray Soil Moisture Observing System (Zreda et al.,
2012), which have helped overcome the sparse sampling and poor dynamic
limitations of traditional in-situ methods. These ground-based
measurements methods require continuous financial support and suitable
ground conditions and are limited to small monitoring areas. With the
development of active and passive microwave remote sensing techniques
(Petropoulos et al., 2015), it becomes possible and more convenient to
acquire SSM information dynamically at different spatiotemporal
resolutions over large areas. A series of SSM products derived from
various satellite-based microwave sensors has been released (Njoku et
al., 2003; Parinussa et al., 2014; Meissner et al., 2018). However, with
spatial resolutions of tens of kilometers, the current microwave-based
SSM products are limited to large-scale monitoring applications.
Many approaches have been developed for downscaling these coarse scale
SSM products. Some of these benefit from ancillary information that
captures the variations of SSM at fine resolution, combined with
correlated variables (Ge et al., 2019). There are two main sources of
ancillary variables: active microwave data and visible/infrared data.
Change detection based downscaling algorithms (Piles et al., 2009; van
der Velde et al., 2015) and Bayesian merging methods (Zhan et al., 2006;
Wu et al., 2017) have been proposed to downscale the coarse SSM by using
active microwave data. The active microwave technique is highly
sensitive to SSM and can even penetrate clouds, however it is greatly
affected by soil roughness and vegetation. An alternative downscaling
approach is to use fine resolution optical/thermal data. A number of
downscaling algorithms have been developed to generate fine-resolution
SSM, such as Disaggregation based on Physical And Theoretical scale
Change (Merlin et al., 2015; Malbéteau et al., 2016), trapezoid-based
methods (Yang et al., 2015; Babaeian et al., 2018), regression-based
approaches (Duan et al., 2016; Liu et al., 2018) and
geostatistical
methods (Mukherjee, 2015; Jin et al., 2018). For downscaling with
optical/thermal data, the statistical correlation between SSM and
ancillary variables or physically based models have been explored (Peng
et al., 2017).
Chauhan et al. (2003) proposed an empirical polynomial fitting
downscaling approach using a polynomial regression at coarse spatial
resolution to obtain the fine-spatial-resolution SSM. Since then,
further polynomial fitting downscaling methods have been presented by
employing multiple data sources or different ancillary parameters (Piles
et al., 2014; Knipper et al., 2017), such as land surface temperature
(LST), vegetation information, brightness temperature, albedo,
evapotranspiration and terrain indices. Meanwhile, geographically
weighted regression, which takes into consideration local
characteristics (Song et al, 2019), and machine learning algorithms have
been introduced into downscaling. Machine learning algorithms such as
random forest and support vector regression (SVR) perform better in
capturing the nonlinear relationships among variables and have been
widely applied to downscaling SSM (Zhao et al., 2018; Abbaszadeh et al.,
2019). Some studies directly combined the fine-resolution trend and
coarse-resolution residual to predict the fine-resolution SSM (Im et
al., 2016; Wei et al., 2019). Interpolation techniques such as
bilinear
interpolation and kriging interpolation have been generally used in
residual analysis for approximating the actual fluctuations (Song and
Jia, 2016; Chen et al., 2019). Geostatistical methods with a focus on
the spatial correlation between variables have been increasingly applied
in downscaling (Kaheil et al., 2008; Djamai et al., 2016). However,
these downscaling approaches ignore the change in supports before and
after downscaling. Due to the stratified heterogeneity in geographical
variables, downscaling models for various stratifications established in
the scaling process (Ge et al., 2019) are limited to the smaller samples
captured by the model. The SVR approach, benefiting from its high
generalization ability, could provide a solution to the small sample
size problem (Srivastava et al., 2013).
Considering all the previous machine learning algorithms, this paper
proposes a new machine learning-based geostatistical model that
integrates SVR and area-to-area kriging (ATAK) to achieve spatial
downscaling by fusing various ancillary variables. The proposed support
vector area-to-area regression kriging (SVATARK) can tackle the
modifiable areal unit problem, as well as model the complex nonlinear
relationship among variables in the downscaling process. The downscaling
approach was employed to predict 1-km-resolution SSM data by downscaling
ESA’s 25-km-resolution SSM product, Climate Change Initiative (CCI),
with consideration of land cover types. Downscaled SSM predictions were
produced every eight days over the Naqu region in the central Tibetan
Plateau (TP), and were evaluated using in-situ SSM measurements.
A
comparison of the SSM residuals obtained from the ATAK method versus the
residuals from bilinear interpolation and kriging interpolation
indicated advantages of the SVATARK downscaling approach.
The remainder of the paper is organized as follows. Section 2 describes
the downscaling methodology, including the downscaling strategy during
the experiment. Both the study area and the data sets are introduced in
Section 3. Section 4 validates the downscaled predictions and discusses
the comparison results. Finally, some conclusions are summarized in
Section 5.