More details about the nonlinear SVR can be found in Smola and Schölkopf
(2004). It is well known that the kernel function and its
hyper-parameters have a great impact on the performance of nonlinear SVR
model. In our study, ε-SVR is used with the Gaussian radial basis
function as its kernel function. The relevant penalty coefficient and
gamma can be optimized by minimizing the model error. The SVR was
implemented in
R
“e1071” package (Meyer et al., 2015). Owing to the stratified
heterogeneity, the SVR models are established for different land cover
types, considering that different underlying surfaces might influence
the relationship among SSM and ancillary variables.
2.2 Area-to-area kriging
method
The area-to-area kriging is a case of areal interpolation, which changes
the supports before and after the interpolation (Kyriakidis,
2004).
A linear combination of areal data is used to predict other areal
values. The target areal value z over a given unit \(u_{\alpha}\)is estimated with the K neighboring observations at units \(u_{i}\):