Quantifying Observation Impact for a Limited Area Atmospheric Forecast Model - Data Assimilation for Atmospheric, Oceanic and Hydrologic Applications - page 127

The adjoint observation impact framework has also been derived in terms of a
Taylor series expansion and shown to be a third-order metric by Errico ( 2007 ). The
nonlinear nature of the metric means that cross terms with other observations may
exist, which may make subsets of impacts difficult to interpret. However, the cross
term effect was found to be small in a global system for the major observation
networks ( Gelaro et al. 2007 ). Although the cross terms will not be considered in
this chapter, they may be important for smaller subsets of observations in a limited
area model.
Another consideration not considered in this chapter is redundancy of infor-
mation. Removing an observation from the DA process may result in a pre-
viously unimportant observation becoming important. Finally, the observation
impact framework in this chapter applies to sequential DA, like three-dimensional
variational systems. Approximations are needed to apply the framework to four-
dimensional variational systems ( Tremolet 2008 ).
The adjoint observation impact framework ( Langland and Baker 2004 )isa
powerful tool for monitoring data assimilation performance and observation quality.
Some of the subsequent applications (mainly to global NWP systems) of this
framework will be discussed below (Sect. 6.1.4 ) and its application to a limited area
model will be presented in Sect. 6.3 .

6.1.4 Applications

In the seminal work by Langland and Baker ( 2004 ), the framework was applied
to the Naval Research Laboratory's (NRL) global atmospheric modeling system.
The NWP model was the Navy's Operational Global Atmospheric Predicition
System (NOGAPS) and the accompanying DA component was the NRL Variational
Data Assimilation System (NAVDAS). In the Northern Hemisphere, the largest
error reductions were due to the assimilation of rawindsondes, satellite wind data,
and aircraft observations, while in the Southern Hemisphere, satellite retrieved
temperature profiles were important along with rawindsondes and satellite wind
data. The framework has been implemented at a number of operational NWP center
as a diagnostic monitoring system for the DA process ( Langland 2005 ; Gelaro and
Zhu 2009 ; Cardinali 2009 ). The system will indicate if a particular observation type
or physical area is repeatedly increasing forecast error. These problem observations
or areas can then be further investigated to find the cause of forecast degradation.
A comparison of observation impact systems for three distinct global modeling
systems showed that observations have similar impact no matter the system on
aglobalscale( Gelaro et al. 2010 ). Satellite sounding radiances provided the
largest total impact in each of the forecast systems. Satellite winds, radiosondes,
aircraft observations were also important contributors and are important components
of the global atmospheric observing network. Slightly more than half of all the
measurements (between 50 and 55 %) for each observation type are actually
beneficial to forecast error reduction, meaning that a large number (between 45 and
50 %) are degrading the forecast.