Optimal deployment of targeted observations requires a systematic assessment of the information provided by the observational network to a specific data assimilation and forecast system. The four dimensional variational data assimilation (4D-Var) allows for multiple time varying targeting areas in the assimilation window following the flow regime. The interaction between time and space distributed adaptive observations and other observing systems in the vicinity is not properly incorporated into current objective targeting methods such as singular vectors and analysis sensitivity.
In this work a method to assess the information provided by the routine observational network and to account for data interactions within the targeting procedure is considered in the 4D-Var context. The method is based on the dynamical interaction between the forecast sensitivity field,as computed with an objective targeting method, and a sensitivity field associated to all additional data available to the assimilation procedure. At each targeting instant forward/backward integrations of the tangent linear/adjoint model are used to account for the impact of data from other observing systems. A dynamical update of the forecast sensitivity field provides an updated field that is inversely proportional to the information provided by all observations already located. Adaptive observations are then deployed in regions where the forecast sensitivity to analysis errors is large and little additional information may be extracted from all other observational resources. Using this procedure the redundancy between time and space distributed targeted and routine observations is minimized.
Numerical results are presented with a 2D global shallow-water model using the Lin-Rood flux-form semi-Lagrangian scheme and its adjoint. Initial conditions are specified from the ECMWF 500mb ERA-40 dataset and 4D-Var data assimilation is setup in an idealized twin experiments framework. Implementation in a comparative analysis with targeted observations using singular vectors and gradient sensitivity shows that the novel approach is effective in providing improved forecasts at a low additional computational effort. Applications to optimal flight path design that accounts for multiple observing times and data interaction are discussed.