Observation adjoint sensitivity and the adaptive
observation-targeting problem
Dr. Nancy Baker
NRL Monterey
Abstract:
This research introduces the adjoint of the data assimilation system,
which together with the classical adjoint sensitivity problem,
represents the two fundamental components of the complete forecast
adjoint sensitivity problem. This adjoint of the data assimilation
system is then used to investigate the sensitivity of the forecast
aspect to the observations and background for idealized analysis
problems, and finally a real-data case using the NAVDAS adjoint for a
situation with unusually large 72-h forecast errors over the
western United States during February 1999. The observation sensitivity
is largest when the observations are relatively isolated, assumed to be
more accurate than the background, and the analysis
sensitivity gradients are large in amplitude and have a spatial scale
similar to the background error covariances. The observation
sensitivity is considerably weaker for small-scale analysis sensitivity
gradients. The large observation sensitivities suggest that adaptive
observations near large-scale analysis sensitivity gradients have a
greater potential to change the forecast aspect than observations near
small-scale analysis sensitivity gradients. Therefore, targeting
decisions based on the adjoint of the data assimilation system may be
significantly different from targeting decisions based solely on
the analysis sensitivity gradients. These results emphasize the
importance of accounting for the data assimilation procedures in the
adaptive observation-targeting problem. While this research emphasizes
targeting applications, the research is also
relevant to non-targeting applications since the observation adjoint
sensitivity lends insight into understanding how the observation
information is used by the data assimilation system.