Spatial interpolation in massively parallel computing environments

Prediction of environmental phenomena at non-observed locations is a fundamental task in geographic information science. Often, samples are taken at a limited number of sensor locations and spatial and spatio-temporal interpolation is used to generate continuous maps. The computational cost of the underlying algorithms usually grows with the number of data entering the interpolation and the number of locations for which interpolated values are needed. Thus, real-time provision and processing of large spatio-temporal datasets call for scalable computing solutions. This requires re-thinking of established (sequential) programming paradigms. In this paper, we study the implementation and behavior of inverse distance weighted interpolation (IDW) on a single graphics processing unit (GPU), as an example for spatial interpolation algorithms in massively parallel computing environments. We argue that the underlying ideas can be expanded to a framework providing highly parallel functions for geostatistics.