Accounting for correlated errors in Swarm magnetic data within the CHAOS field modelling framework

Modern geomagnetic field models can successfully represent many details of the observed large-scale field and its slow time changes. However, the obtained model uncertainty is often underestimated, which limits our ability to evaluate the reliability of signals recovered in the field models. The increasing amount of globally distributed, high-quality magnetic data observed by low-Earth orbit satellites, such as Swarm, MSS-1 and the planned NanoMagSat mission, present an opportunity to improve the model uncertainty by providing important statistical information on the expected errors of the input magnetic data used in field modelling.  

During the field model estimation, data errors are usually assumed to be uncorrelated in time and independent of position. However, limitations in the parameterization of the models regarding magnetospheric and ionospheric sources lead to residuals between model predictions and magnetic observations that are not only larger than the expected measurement noise but are also correlated and varying with position. Not adequately describing these correlations during the model estimation leads to unrealistic model uncertainties, which hinders, for example, their use in applications such as assimilation into numerical Geodynamo simulations.

Here, the statistics of vector residuals between magnetic observations from the Swarm satellites and the CHAOS-7 geomagnetic field model predictions are studied by computing sample means and covariances for the field components as a function of time and magnetic coordinates. This analysis reveals significant covariances, particularly at mid-to-high latitudes. The sample covariances are used to construct non-diagonal data error covariance matrices, which can be used in field modelling.

Finally, test field models built using the non-diagonal data error covariances matrices within the CHAOS modelling framework are discussed, illustrating the effect of correlated data errors on the recovered fields and the associated model uncertainties.