The main motivation of running ECMWF analyses using the IFS Ensemble of Data Assimilation (EDA) is to provide the best possible initial conditions for forecast initialisation. Atmospheric composition is as much a boundary condition problem as an initial condition problem, hence there is a need to infer surface fluxes in addition to the 3D representation of atmospheric variables. A first attempt to use the ECMWF’s IFS Ensemble of Data Assimilation (EDA) information to constrain the Copernicus Atmosphere Monitoring Service (CAMS) surface emissions is presented. With a focus on carbon monoxide, the methodology presented here aims to use the ensemble information and its flow dependency to derive a space and time varying relationship between assimilation increments on the atmospheric composition fields and surface emissions.
This study focused on carbon monoxide (CO) as it is well observed, well modelled and is a primary
pollutant mainly originating from anthropogenic and fire sources. The methodology is described and details how the emission ensemble perturbation are done. Using a statistical approach, the ensemble information is then used to derive Jacobians that provide a sensitivity between the increments to the 3D concentration fields and the surface emission or fluxes. We show how the Jacobians are computed and filtered to provide an update on the emission that is carried forward in time to the next assimilation
Results show impacts and comparisons of various experiments with assimilation solely (i.e. only the
initial conditions are modified by the observations) and assimilation plus inversions. In addition, different prior assumptions for the ensemble emission perturbations and different perturbation persistence times are investigated. Significant changes on the ensemble mean and spread are observed on the surface CO emissions but also in the near surface 3D CO concentrations.
Validation shows that this approach is sensitive to the design of the prior error used for the emissions
and how the information is propagated forward in time. Improvement is clearly seen from this approach on the bias scores but not so much in the root mean square errors (RMSE). Increasing the prior error perturbations and persistence time leads to significant RMSE degradation. Discussion about priorities for the next possible implementation phases of such an inversion capability for CAMS are also emphasised.