|
Abstract:
|
Traditional spatial modeling approaches assume that data are second-order stationary, which is rarely true over large geographical areas. A simple way to model nonstationary data is to partition the space and build models for each region in the partition, but this has the side effect of creating discontinuities in the prediction surface at region borders. The R package remap deploys a regional border smoothing approach that ensures continuous predictions by using a weighted average of predictions from regional models. The weights are based on the shortest distance from the observation to each region boundary, with greater weight given to observations that are within or near the boundaries of a region. Special consideration is given to distance calculations that make the remap package scalable to large problems. Improvements in accuracy using the remap package, as opposed to global spatial models, are illustrated using a national-level ground snow load dataset. These accuracy improvements, coupled with their computational feasibility, illustrate the efficacy of the remap approach to modeling nonstationary data.
|
