Sensitivity to observations applied to FASTEX cases

Abstract

Abstract. The concept of targeted observations was implemented during field experiments such as FASTEX, NORPEX or WSRP in order to cope with some predictability problems. The techniques of targeting used at that moment (adjointbased or ensemble transform methods) lead to quite disappointing results: the efficiency of the additional observations deployed over sensitive areas did not turn out to remain consistent from one case to another. The influence of targeted observations on the forecasts could sometimes consist of strong improvements, or sometimes strong degradations. It turns out that the latter failure explains why the concept of optimal sampling arose. The efficiency of adaptive sampling appears to depend on the assimilation scheme that deals with the observations. It is then very useful to integrate the nature of the assimilation algorithm, as well as the deployment of the conventional network of observations (redundancy issues between targeted and conventional network) in the definition of the sensitive pattern to be sampled. Therefore, we chose the tool of the sensitivity to observations to allow us to test such an approach. The sensitivity to targeted observations (that utilizes the adjoint of the linearized NWP model and the adjoint of the assimilation operator) seems to be a suitable tool to obtain an insight into the tricky issue of the optimization of the sampling strategies.To understand better the intrinsic patterns and the influence of the 3D-Var assimilation scheme on the sensitive structures to be sampled, we present here some detailed results on a FASTEX targeting case. We focus on the dropsondes deployed by the Gulfstream IV (jet-aircraft) along its first flight during Intense Observing Period 17 that started on the 17 February 1997. The sensitivity to observation is used as a diagnostic tool for studing targeting from a critical point of view. It is shown that assimilation processes can have an important effect on the classical sensitivity fields, and particularly on their vertical extension. For example, in the studied case, the classical sensitivity fields remain at a lower level than 400 hPa, whereas the sensitivity to observationsCorrespondence to: T. Bergot (Thierry.Bergot@meteo.fr) stretches up to 250 hPa. However, the maximum values can be found at approximately 700 hPa in both sensitivity fields.The studied case shows that the efficiency of observations depends not only on the sensitivity but also on the deviations between the observations and the background field. An example of the use of this diagnosis for comparing the relative efficiency of different kinds of observations is also presented. This work points out that it is very complicated to optimize the efficiency of adaptive observations, and that the assimilation of an entire set of observations (both conventional and adaptive network) needs to be considered.

Adjoint sensitivity of the model forecast to data assimilation system error covariance parameters

“…The development of the adjoint of the forecast model and of the adjoint of the data assimilation system (adjoint-DAS) makes feasible the evaluation of the derivative-based, local sensitivity (Cacuci, 2003) of a scalar forecast aspect with respect to a large number of DAS input components. Adjoint-DAS estimation of the observation sensitivity was considered in the work of Baker and Daley (2000) and Doerenbecher and Bergot (2001) as a tool to design observation targeting strategies. Subsequently, Langland and Baker (2004) have shown that the combined information derived from the adjoint of the forecast model and the adjoint-DAS may be used to provide a detailed (all-at-once) assessment of the observation impact on reducing the forecast errors in variational data assimilation (VDA).…”

A Technical Description of the NAVDAS Adjoint System

“…It can also be noted that studies by Doerenbecher and Bergot (2001) and Fourrie et al (2002) use observation impact functions in observation space that are similar in form to Eq. (16) -e.g., they involve the innovation vector and an observation sensitivity gradient However, in those studies the observation sensitivity is not derived with an actual adjoint of the assimilation procedure.…”

Criteria of optimality for sensors' location based on adjoint transformation of observation data interpolation error

“…In data assimilation problems [1][2][3][6][7][8][9][10][11][12][13][14], usually, the cost functional (discrepancy between observations and model calculations) is minimized and assumes the following form:…”

“…In this approach, the forecast aspect gradients are estimated in metrics produced by the FIM resulting in a forecast aspect variance. The approach by Bergot and Doerenbecher [13,14] is very close to the V-optimality condition used in some works on the optimal experiment design [15].…”

“…In [13,14] the optimization of observations was conducted via direct computation of a reduction of the error variance of forecast score caused by additional observations. The variance of forecast score was expressed through adjoint sensitivity gradients and the covariance matrix of analysis (initial state) error.…”