In data assimilation systems (DASs), the effect of each assimilated observation dataset on an analysis field is one of the main factors in determining analysis and succeeding forecast accuracy. We call the effect the linear observation impact, which is fully determined by the Kalman gain. However, the Kalman gain is never constructed explicitly in variational DASs. Therefore, the estimation of the linear observation impact is a difficult problem. In this study, we analyze linear observation impacts using two methods, the adjoint-based method (Baker and Daley 2000, Langland and Baker 2004, and Errico 2007) and the tangent linear approximation based method (Ishibashi 2011).
First, the adjoint-based method is implemented to the JMA global 4D-Var DAS. One-month long experiments of the observation impact estimation (using a dry total energy norm and 15 or 27 hour forecasts) show that almost all observation data contribute forecast error reduction in average, and this result is consistent with those of past OSEs in JMA. However, the experiments show that the impact of a total satellite radiance data is about the same magnitude as that of radio sonde data. This result implies that there is still many room for improvement of the forecast accuracy by improving usage of the radiance data, since, in previous studies in other NWP centers shows larger impacts from radiances (Cardinali 2009). We also find impacts of water vapor channel radiances increase in case of using a norm including humidity energy, and the method can detect wrong observation data which are given artificially deflated (too small) observation error settings.
Second, we construct the tangent linear approximation based method. The method estimates the linear observation impacts as a partial analysis increment vector that is generated by each observation dataset. The method enables us to see how the Kalman gain transforms information of observations into analysis increments.
In the seminar, applications of these methods for a design of an optimal observation system and an optimization of covariance matrix (Daescu 2008, Daescu and Todling 2010) will also be discussed.