|
Abstract |
With the advent of remote sensing and GPS techniques, spatial data collection capacity increases dramatically. The growth in data size imposes challenges to classical spatial modeling methods and has driven the innovations of new modeling and computation tools scalable and parallelizable to handle large datasets. This work extends the state-of-the-art full scale covariance approximation approach that combines merits of reduced rank methods and sparse approximations, by accounting for the dependence across blocks of the residual covariance. We show that the proposed likelihood approximation approach induces a valid Gaussian process, which allows for a unified framework for model estimation and spatial prediction following standard kriging methods. |
