1222 Ocean Dynamics (2019) 69:1217–1237 Fig. 3 Satellite SST observations on both model grids. Shown are two extremes of data coverage. On the 10th of April, the North and Baltic Seas were nearly fully covered by clouds. On the coarse grid data is only available on 7 % of the grid points, while for the fine mesh there are zero observations over the 12-h time window. For the 25th of May, the domains were nearly cloud free so that there are only small data-void regions seadata.bsh.de/csr/retrieve/dod index.html) operated by the BSH. Apart from water temperature and salinity, the data set also includes measured concentrations of oxygen, nitrate, ammonium, phosphate, silicate, and chlorophyll, which can be used to assess the corresponding concentrations in the ERGOM model. The validation of the assimilation experiments will focus on the surface and will be conducted for both the fine and coarse model grids. 4 Experimental setup The assimilation experiments are conducted over the time period from April to July 2012 with an analysis update after each 12 h. An ensemble of 40 model states is used. The initial physical ocean state (i.e. ensemble mean) is provided by the operational run of the HBM model at the BSH. The BGC model state was initialised on 1st November 2011 using for the Baltic Sea an initial state provided by the Danish Technical University (generated by the model of Maar et al. (2011) by M. Maar, personal communication) and for the North Sea an initial state generated by the model of Lorkowski et al. (2012). The ensemble perturbations were computed using second-order exact sampling (Pham et al. 1998) using the variability of the model state in a forecast run of the HBM-ERGOM model for April 2012. The state vector for the assimilation jointly includes the model fields on both nested model grids (similar to Barth et al. 2007) and consists of physical and BGC parts on both nested model grids. For the physical part, the state vector includes the SSH and the 3-dimensional temperature, salinity and horizontal velocities. For ERGOM, all 13 prognostic pelagic and 2 benthic variables as well as the Secchi depth and chlorophyll concentration are included in the state vector. The two latter diagnostic variables are, however, only included to access their ensemble values, but they are not directly updated by the analysis step of the LESTKF. For the localisation of the analysis step an influence radius for the observations of 50 km is used for the coarse grid, while 9 km areused for the fine grid. An inflation of the ensemble variance with a forgetting factor of ? = 0.95 is used. For the assimilation of the SST observations, an observation error standard deviation of 0.8 ?C is assumed as in Losa et al. (2014) for both model grids. Two assimilation experiments are performed to assess the different effects of the weakly and strongly coupled assimilation. The experiment WEAK assimilates the SST observations so that only the physical model fields in the state vector are directly updated. BGC model fields react only dynamically to the changed physical conditions during the next forecast phase of 12 h. In contrast, in the experiment STRONG both the physical as well as BGC model fields are directly updated. Thus, the strongly coupled assimilation uses the multivariate ensemble-estimated cross-covariances between the SST and the BGC variables to compute an update of the biogeochemistry. Further, the experiment FREE was performed in which the ensemble was integrated without assimilating observations. The experiment STRONG is performed in two variants. STRONG-lin performs the assimilation using the actual