IMAGe Seminar- Spatial Interpolation via Bayesian Multiscale Clustering

09/30/2014 - 12:00pm
Mesa Laboratory - Chapman Room
Zachary Thomas
 
Zachary Thomas
The Ohio State University
Tuesday, September 30, 2014
Mesa Laboratory - Chapman Room
12:00 PM
 

Classical techniques for spatial interpolation have relied on the specification a stationary/isotropic covariance model for the characterization of spatial coherence. However, there is much interest in the development of more flexible methodologies capable of approximating the complex, nonstationary spatial dependence which often manifests itself in geophysical (and other) processes. Current techniques for achieving this have focused on either geometric deformations of the spatial domain or the construction of spatially varying covariance models. We present an alternative approach in which covariance is not directly modeled. Rather, dependence structure is learned via the construction of a multiresolution Bayesian spatial clustering process. Induced covariance is then naturally data-driven, nonstationary, and anisotropic. Example applications for several geostatistical problems will be presented.