Adaptive Weighting of Innovations in Atmospheric Data Assimilation

This describes a new adjoint-based method for adaptive weighting of innovations in atmospheric data assimilation, and results of testing the method for improved numerical weather prediction (NWP). The study uses the Navy Global Environmental Forecast System (NAVGEM) and its 4d-Var data assimilation procedure (NAVDAS-AR). The method is general and can be used for assimilation of any type of in-situ, satellite or remotely-sensed observation. In this study, the main objective is to improve assimilation of satellite radiance observations in NAVDAS-AR and thereby increase the accuracy of NAVDAS-AR atmospheric analyses and NAVGEM forecasts.

The procedure to obtain the innovation weight uses the following steps. Adjoint versions of NAVGEM and NAVDAS-AR are used to obtain the sensitivity of short-range forecast error (measured by a total moist energy norm) to all assimilated innovations for the most-recent 30 days. The sensitivity is then partitioned into categories of observation types, including sensors (AMSU-A, AIRS, SSMI/S, etc), channels (IR, MW, visible, etc.), automated wind vectors, and in-situ observations such as radiosondes, aircraft, etc. In practice, the sensitivity can be examined for any user-selected sub-set of observations. The categorized sensitivity results are then examined to determine if short-range forecast error would be reduced in a time-average and global-average measure by either increasing or decreasing the influence of the innovations associated with each observation category.

The innovation weight is represented by scalar values (S) applied to subsets of innovations. The default value of S = 1.0, and S can vary from 0.0 to values greater than 1.0. Optimal values of S will be obtained through a process of experimentation, tuning, and updating, as is done for specification of observation error and background error in the data assimilation process. However, since S directly weights the innovation, the process of optimizing S is simpler than attempting to optimize observation and background error, which have opposing effects on innovation influence. The computational cost of adding an innovation weight is negligible, and the required sensitivity gradients are already available from operational production of observation sensitivity.

Conceptually, the weight S mitigates systematic inaccuracies that exist in specification of observation error and background error, as defined in a particular data assimilation procedure. These inaccuracies may be caused by sub-optimal methods of error estimation or from failure to properly account for biases in observations and background. By definition, inaccurate specification of observation or background error reduces in an average sense the amount of useful information extracted from assimilation of observations, and may contribute to intermittent, but significant, forecast failures. It will be shown that the innovation weighting method is a practical and effective way to improve the balance of information obtained from observations and background in operational data assimilation.