Mapping by spatial predictors exploiting remotely sensed and ground data: A comparative design-based perspective

Abstract

This study was designed to compare the performance – in terms of bias and accuracy – of four different parametric, semiparametric and nonparametric methods in spatially predicting a forest response variable using auxiliary information from remote sensing. The comparison was carried out in simulated and real populations where the value of response variable was known for each pixel of the study region. Sampling was simulated through a tessellation stratified design. Universal kriging and cokriging were considered among parametric methods based on the spatial autocorrelation of the forest response variable. Locally weighted regression and k-nearest neighbor predictors were considered among semiparametric and nonparametric methods based on the information from neighboring sites in the auxiliary variable space. The study was performed from a design-based perspective, taking the populations as fixed and replicating the sampling procedure with 1000 Monte Carlo simulation runs. On the basis of the empirical values of relative bias and relative root mean squared error it was concluded that universal kriging and cokriging were more suitable in the presence of strong spatial autocorrelation of the forest variable, while locally weighted regression and k-nearest neighbors were more suitable when the auxiliary variables were well correlated with the response variable. Results of the study advise that attention should be paid when mapping forest variables characterized by highly heterogeneous structures. The guidelines of this study can be adopted even for mapping environmental attributes beside forestry.