Spatial interpolation via GWR, a plausible alternative?

Abstract

Spatial interpolation can be done through either univariate methods that rely solely on the spatial structure of the data or by combining the spatial information and attribute information. Geographically weighted regression, although is used primarily in modeling the spatially varying relationships, falls within the category of combining both spatial and attribute information to interpolate unknown values. Using both artificially generated data with predefined parameters and actual house data from the City of Milwaukee, this study evaluates the interpolation accuracy of the univariate interpolation method represented by ordinary Kriging and multivariate interpolation represented by regression Kriging and GWR interpolation. It is found that by including relevant auxiliary variable(s), RK and GWR interpolations yield more accurate results than the univariate interpolation method, though the subtlety of how the spatial structure is assumed produces slight difference between RK and GWR. This study suggests GWR can serve as a useful alternative interpolation method in data analysis in addition to providing more detailed understanding of the spatially varying relationships between target and auxiliary variables.