Introduction to Data Assimilation
Weather forecasts are obtained by integrating atmospheric models from initial conditions given by the current weather. Even if these weather models perfectly described atmospheric dynamics, the chaotic nature of the model amplifies small errors in the intial conditions over extended forecasts. So, improvements in the initial conditions for the model will reduce errors in the resulting forecasts.
Such improved estimates of the current weather are obtained by using data assimilation techniques to combine information from previous forecasts with information from atmospheric observations. Operational data assimilation techniques must be computationally efficient, incorporate satellite observations, and account for model dynamics. These goals met by the Local Ensemble Transform Kalman Filter (LETKF). This scheme was first proposed by Ott et al (2004), verified by Szunyogh et al (2005), and reformulated for efficiency by Hunt (2005) (publications available on the Applied Dynamics - Numerical Weather Forecasting group site). The group is continuing research to show the operational potential of LETKF.
Relevant Paper
Klein, E. (2005): Scholarly Paper for Masters Degree
Applying LETKF to the NASA fvGCM
Collaborators: Hong Li, Junjie Liu
Advisors: Brian Hunt, Eugenia Kalnay, Eric Kostelich (1), Ed Ott, Istvan Szunyogh, Ricardo Todling (2)
To show the potential of LETKF as an operational scheme, it must be applied to operational models and compared to operational data assimilation schemes. Accordingly, in this study, we apply LETKF to the NASA fvGCM weather model and compare the forecasts resulting from this data assimilation scheme to those obtained by the operational NASA PSAS scheme. Initially, perfect model experiments assimilating simulated rawinsonde observations reveal the superiority of the intial conditions obtained from LETKF. Currently, we are working to assimilate real rawinsonde observations and AIRS satellite observations.
Relevant Paper:
(In Progress) Liu, J., Fertig, E., Li, H., Szuynogh, I., Hunt, B., Kalnay, E., Kostelich, E., and Todling, R., Application of Local Ensemble Kalman Filter: perfect model experiments with NASA fvGCM model. (Extended Abstract)
Assimilating Satellite Observations
Advisors: Brian Hunt, Ed Ott, Istvan Szuynogh
Many Ensemble Kalman Filter data assimilation schemes benefit from employing vertical localization (e.g., to avoid spurious cross correlations and improve computational efficiency). On the other hand, satellite observations are often sensitive to the dynamics over broad layers of the atmosphere. This nonlocality can present problems for assimilating these observations. We study ways to assimilate nonlocal radiances and retrievals with local Ensemble Kalman Filter schemes.
Relevant Paper:
(In Progress) Fertig, E., Hunt, B., Szunyogh, I., and Ott, E., Assimilating Nonlocal Observations with a Local Ensemble Kalman Filter
A Comparative Study of 4D-Var and a 4D Ensemble Kalman Filter
Collaborator: John Harlim
Advisor: Brian Hunt
Historically, data assimilation schemes have been applied assuming all observations are taken at the time of the analysis. However, as satellite observations begin to dominate observing networks, this assumption is no longer valid. To handle this challenge, two competing methods are currently being developed and implemented for the next generation of data assimilation, 4D-VAR and Ensemble Kalman Filter Techniques. 4D-VAR obtains an analysis by seeking the best model trajectory fitting the available atmospheric observations. This analysis is obtained by minimizing an appropriate cost function. On the other hand, 4D Ensemble Kalman Filter techniques, such as 4D-LETKF, run an ensemble of model forecasts from a perturbed set of initial conditions. The Kalman Filter Equations are applied to obtain the best linear combination of model trajectories given the available atmospheric observations.
Relevant Paper:
(In Review, Submitted to Tellus) Fertig, E., Hunt, B., and Harlim, J., A Comparitive Study of 4D-VAR and a 4D Ensemble Kalman Filter: Perfect Model Simulations with Lorenz-96
Statistical Topography of Noisy Self-Affine Surfaces
Advisors: Jane Kondev (3), Greg Huber (4)
Analysis of Mayall 4m Environment
Advisor: Nigel Sharp (5)
Special thanks to NASA Goddard and AIRS Science Team for funding my graduate research
Collaborators and advisors affiliated with University of Maryland, College Park MD, unless otherwise noted.
(1) Arizona State University, Tempe AZ
(2) NASA Goddard Space Flight Center, Greenbelt MD
(3) Brandeis University, Waltham MA
(4) University of Massachusetts, Boston MA
(5) National Optical Astronomy Observatory, Tucson AZ
Site last updated: Nov 15, 2005