Examination of observation impacts derived from observing system experiments (OSEs) and adjoint models

With the adjoint of a data assimilation system, the impact of any or all assimilated observations on measures of forecast skill can be estimated accurately and efficiently. The approach allows aggregation of results in terms of individual data types, channels or locations, all computed simultaneously. In this study, adjoint-based estimates of observation impact are compared with results from standard observing system experiments (OSEs) using forward and adjoint versions of the NASA GEOS-5 atmospheric data assimilation system. Despite important underlying differences in the way observation impacts are measured in the two approaches, the results show that they provide consistent estimates of the overall impact of most of the major observing systems in reducing a dry total-energy metric of 24-h forecast error over the globe and extratropics and, to a lesser extent, over the tropics. Just as importantly, however, it is argued that the two approaches provide unique, but complementary, information about the impact of observations on numerical weather forecasts. Moreover, when used together, they reveal both redundancies and dependencies between observing system impacts as observations are added or removed from the data assimilation system. Understanding these dependencies appears to pose an important challenge in making optimal use of the global observing system for numerical weather prediction.