Low-Rank Spatial Models for Big Global Data Sets

• Activity Number: 6
• Type: Invited
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Matthias Katzfuss*+
• Companies: Universität Heidelberg
• Keywords: Full-scale approximation ; Global satellite data ; Massive datasets ; Matern covariance ; Nonstationarity ; Spatial statistics

Abstract

In recent years, the global coverage of data collected by satellite instruments has made it possible to analyze environmental processes on a truly global scale. However, this requires spatial statistical models that are valid on a spherical domain, that are highly flexible to reflect the homogeneity of the globe, and that are computationally feasible to deal with the often massive satellite datasets.

For this purpose, we propose a parameterization of the nonstationary Matérn covariance function that is suitable for the sphere. This covariance function can then be used for the so-called parent process in the full-scale approximation, which combines a low-rank component and a tapered fine-scale component to obtain a computationally feasible model that is close to the parent process.

The methodology is illustrated using satellite measurements of CO2.