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.

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 \(u_{i}\):