Daisuke Hotta
Meteorological Research Institute, Japan Meteorological Agency
The 8th EnKF Data Assimilation Workshop
Toward improved LETKF assimilation of non-local and dense observation by direct covariance localization in model space
Talk:
EnKF2018_hotta.pdf
Covariance localization is an indispensable component of ensemble-based data assimilation systems with a limited member size. The benefit of localization is two-fold: (1) it suppresses spurious correlation due to sampling error, and (2) mitigates the rank-deficiency issue of the sample covariance matrix.
In Local Ensemble Transform Kalman Filter (LETKF), localization is typically implemented by domain localization (i.e., by performing analysis independently at each model grid) and by applying the so-called R-localization, in which the impact of an observation to the analyzed state variable is artificially damped by inflating the observation error variance by a factor that is a decreasing function of the physical distance between the observation and the analyzed grid.
Recent study at Japan Meteorological Agency using its operational global LETKF revealed that, while R-localization effectively suppresses spurious impact from an observation onto remote grids, it does not help to alleviate the rank deficiency issue within the local analysis, hindering the LETKF from extracting information from dense observations. R-localization also poses difficulty when assimilating non-local observations (e.g., ground-based GNSS observations, satellite radiances, or even in-situ surface pressure observations) whose physical locations are not clearly defined.
To resolve the above issues, we explore possibility of applying model-space localization within the framework of LETKF. The advantage of model-space localization over the R-localization will be discussed using an idealized one-dimensional toy system. The toy system is designed with assimilation of ground-based GNSS observations mind, but the methodology should also work for other non-local and dense observation types such as satellite radiances.
Covariance localization is an indispensable component of ensemble-based data assimilation systems with a limited member size. The benefit of localization is two-fold: (1) it suppresses spurious correlation due to sampling error, and (2) mitigates the rank-deficiency issue of the sample covariance matrix.
In Local Ensemble Transform Kalman Filter (LETKF), localization is typically implemented by domain localization (i.e., by performing analysis independently at each model grid) and by applying the so-called R-localization, in which the impact of an observation to the analyzed state variable is artificially damped by inflating the observation error variance by a factor that is a decreasing function of the physical distance between the observation and the analyzed grid.
Recent study at Japan Meteorological Agency using its operational global LETKF revealed that, while R-localization effectively suppresses spurious impact from an observation onto remote grids, it does not help to alleviate the rank deficiency issue within the local analysis, hindering the LETKF from extracting information from dense observations. R-localization also poses difficulty when assimilating non-local observations (e.g., ground-based GNSS observations, satellite radiances, or even in-situ surface pressure observations) whose physical locations are not clearly defined.
To resolve the above issues, we explore possibility of applying model-space localization within the framework of LETKF. The advantage of model-space localization over the R-localization will be discussed using an idealized one-dimensional toy system. The toy system is designed with assimilation of ground-based GNSS observations mind, but the methodology should also work for other non-local and dense observation types such as satellite radiances.
Thanks to McGill's Atmospheric and Oceanic Sciences department for hosting this web-site.