B1776
Title: A multi-resolution approximation by linear projection and covariance tapering for large spatial datasets
Authors: Toshihiro Hirano - Kanto Gakuin University (Japan) [presenting]
Abstract: Estimation and prediction for large spatial datasets, such as maximum likelihood estimation and kriging, are impractically time-consuming. A multi-resolution approximation via linear projection has been developed to deal with this computational burden. However, this method can cause the partly mismatched fitting around the origin for the covariance function if the two locations are not in the same subregion at the low resolution. Additionally, in this case, there is a possibility of producing artificiality in the predictive surface. To solve this problem, we propose an algorithm that approximates the covariance function by iteratively applying the covariance tapering instead of dividing the region at each resolution. We also elicit fast computation algorithms for estimation and prediction by using the approximated covariance function. A real data analysis for air dose rates demonstrates that our proposed method works well and avoids artificiality in the predictive surface.
