Jonathan Poterjoy
NOAA Atlantic Oceanographic and Meteorological Laboratory ; University of Maryland
The 8th EnKF Data Assimilation Workshop
Toward the application of particle filters for numerical weather prediction and research
Talk:
JPoterjoy_EnKF_2018.pdf
The success of ensemble Kalman filters in oceanography, meteorology, and other fields of geoscience is remarkable, considering the relatively small ensemble sizes permitted by high-performance computing resources available at research and operational centers. Dimension-reduction procedures and methods for treating sampling errors in Gaussian approximated probability densities, via covariance localization, variance inflation, etc., make this achievement possible. As moderate to large ensemble sizes becsome regularly available for geophysical data assimilation, a natural progression from current Kalman filter-based strategies is to attempt more general (Bayesian) probability density estimation. For the case of numerical weather prediction, advancements of this sort may be necessary to extract information from underused observing systems such as cloud- and precipitation-affected satellite measurements, or better constrain dynamical processes responsible for severe convective storms and tropical cyclones. Nevertheless, Bayesian filters constructed for these applications will likely require tactics for overcoming dimensionality constraints that are analogous to those used by ensemble Kalman filters. This presentation will summarize recent efforts to develop Monte Carlo filters that bridge between Gaussian and Bayesian data assimilation methods for numerical weather prediction. Strategies adopted for this purpose typically rely on sequential importance resampling techniques used for particle filters, but their practical application for high-dimensional systems follow from ideas that emerged from decades of ensemble Kalman filtering research. Using a recently developed particle filter approach that operates effectively for high-dimensional applications, this presentation will also present examples that motivate future development of non-Gaussian filters.
The success of ensemble Kalman filters in oceanography, meteorology, and other fields of geoscience is remarkable, considering the relatively small ensemble sizes permitted by high-performance computing resources available at research and operational centers. Dimension-reduction procedures and methods for treating sampling errors in Gaussian approximated probability densities, via covariance localization, variance inflation, etc., make this achievement possible. As moderate to large ensemble sizes becsome regularly available for geophysical data assimilation, a natural progression from current Kalman filter-based strategies is to attempt more general (Bayesian) probability density estimation. For the case of numerical weather prediction, advancements of this sort may be necessary to extract information from underused observing systems such as cloud- and precipitation-affected satellite measurements, or better constrain dynamical processes responsible for severe convective storms and tropical cyclones. Nevertheless, Bayesian filters constructed for these applications will likely require tactics for overcoming dimensionality constraints that are analogous to those used by ensemble Kalman filters. This presentation will summarize recent efforts to develop Monte Carlo filters that bridge between Gaussian and Bayesian data assimilation methods for numerical weather prediction. Strategies adopted for this purpose typically rely on sequential importance resampling techniques used for particle filters, but their practical application for high-dimensional systems follow from ideas that emerged from decades of ensemble Kalman filtering research. Using a recently developed particle filter approach that operates effectively for high-dimensional applications, this presentation will also present examples that motivate future development of non-Gaussian filters.
Thanks to McGill's Atmospheric and Oceanic Sciences department for hosting this web-site.