Karl Pazdernik, Iowa State University

Efficient Kriging for Large Spatial Fields

Abstract: In spatial statistics, a common method for prediction over a Gaussian Random Field (GRF) is Kriging. Unfortunately, Kriging requires inverting a covariance matrix which, depending on the data set, can be extremely large and thus computationally expensive. Thus, I propose a new approach to estimation and prediction that uses a combination of concepts from reduced-rank Kriging and Ridge Regression. I will contrast the gains in run time versus the loss in precision, as well as explore the connection between the actual parameter values and the choice of basis functions in the model.

Presentation slides (pdf)