A Comparative Study of Diagnostic and Metrological Estimates of Spatial and Temporal Error Correlations for AMSU-A Type Observations

Large volumes of high-resolution satellite data are already available, but only 5-10% of these data are used in data assimilation (DA) due, in part, to a lack of knowledge of observation uncertainty structures. Operational centres may follow a pragmatic strategy that avoids treating spatial and temporal observation error correlations, with up to 95% of the data discarded by thinning spatially and temporally to avoid overfitting. This restricts the DA systems from providing an accurate analysis on the important small scale features associated with high impact weather such as tropical cyclones. In order to use larger volumes of data in DA, error correlations need to be explicitly accounted for. We compare and contrast two different approaches to estimating the spatial and temporal error correlations applied to observations from AMSU-A, which is a microwave sounder with tropospheric channels primarily sensitive to air temperature. These observations provide some of the biggest impacts on forecasts in the current ECMWF system, partly due to the number of working sensors in the orbit (6) and the large swath width (2000 km). Recently in 2021, the assimilation of AMSU-A radiances in all-sky conditions was introduced at ECMWF providing further benefits to forecast skill. It is expected that the error structures for AMSU-A will primarily consist of spatial and temporal correlations, with minimal inter-channel error correlations in clear-sky regions. The first estimate of the observation error covariance matrix will use the popular diagnosis approach where the full observation error covariance matrix is estimated by computing the sample covariances between observation minus background and observation minus analysis residuals that are routinely available from the DA system. The second estimate is based on metrological techniques that identify and accumulate individual sources of error correlations. A comparison between the two estimates will provide new confidence in our estimates and insight into the relative importance of the different sources of error. The estimates will help us to identify instances where error correlations can be reduced and where they cannot. This will inform the next step of the project, which is to modify the DA algorithms to explicitly account for the error correlations, allowing for a greater use of the available data.