|
Activity Number:
|
190
- Contributed Poster Presentations: Section on Statistics and the Environment
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, July 29, 2019 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistics and the Environment
|
|
Abstract #306857
|
|
|
Title:
|
Scalable Smoother to Improve Particle Filtering of Spatially-Extended Data
|
|
Author(s):
|
Gregor Robinson* and Ian Grooms and William Kleiber
|
|
Companies:
|
University of Colorado Boulder and University of Colorado Boulder and University of Colorado
|
|
Keywords:
|
data assimilation;
surrogate models;
importance sampling;
spatial statistics;
random fields
|
|
Abstract:
|
Particle filtering (SIR) exhibits catastrophic ensemble under-dispersion (collapse) in high dimension. Previous work shows that collapse can be mitigated by smoothing observation data before proceeding as if observation errors are uncorrelated atop smoothed data. This is equivalent to modeling the errors as having a spectrum that grows in the progression to fine scales, and reduces the effective dimensionality of the filtering problem. Fine scales offer little predictive utility in geophysical fluids, justifying this approach in some geophysical applications where SIR has yet been unsuccessful due to collapse. Such applications often involve data measured at scattered locations, so we introduce a smoother that scales well to scattered data in high observational dimension and that allows prescribable attenuation across scales. The smoother is based on Gaussian approximations both of the data and of a smoothing kernel, in a special case where the data are interpolated by radial basis functions and the smoothing kernel is a multi-resolution approximation of the Green function for a bound-state Helmholtz operator. We demonstrate the smoother on scattered meteorological data.
|
Authors who are presenting talks have a * after their name.
