1218 Ocean Dynamics (2019) 69:1217–1237 sea level, are frequently available measurements of the sea surface. The assimilation of physical observations to con- strain the physical ocean model is common practice. How- ever, it has been found that the assimilation of these obser- vations to constrain the physical ocean state can deteriorate the biogeochemical (BGC) fields. For the North Atlantic, Berline et al. (2007) found that the assimilation of sea sur- face temperature (SST) and sea surface height (SSH) data changed the mixed layer so that much higher vertical nutri- ent fluxes appeared in the mid-latitudes and sub-tropics, which caused deteriorated phytoplankton concentrations. Also, While et al. (2010) reported increased nutrients and in consequence overestimated primary production and chloro- phyll concentrations in the subtropical gyres and at the equator. Similar increased upward flux of nutrients and cor- responding increased production was found by Raghukumar et al. (2015) in the California Current System. To correct for spurious changes by the data assimilation, corrections to the nutrient fields have been proposed (While et al. 2010; Shul- man et al. 2013) while (Park et al. 2018) suggests to reduce the assimilation effect around the Equator. There are also observations of the ocean colour, from which, e.g. concentrations of chlorophyll or diffuse attenuation rates are derived. In particular, chlorophyll concentrations have been used to directly influence the BGC model state (e.g. Nerger and Gregg 2007, 2008; Gregg 2008; Ciavatta et al. 2011; Ford et al. 2012, 2017). However, the data errors are higher for chlorophyll than for physical quantities like SST. Further, satellite chlorophyll observations have particularly high uncertainties in coastal waters, because the standard processing, like the ocean colour algorithm by Hu et al. (2012) commonly used in the processing of MODIS data, is only valid for clear case-1 waters and the availability of data sets processed for the coastal regions is very limited. Another data source on BGC quantities are in situ data, e.g. of nitrate. While these data are also available below the surface, they are much more sparse than satellite data, which strongly limits their applicability for data assimilation. In a coupled data assimilation, one can classify the data assimilation approach depending on which model fields are influenced by which data type. The studies mentioned above performed a so-called ‘weakly coupled’ assimilation, by assimilating observations of the ocean physics into the physical model component or assimilating observations of BGC variables into the ecosystem component of the coupled model. A more sophisticated approach is the ‘strongly coupled’ data assimilation. In this case, one uses cross-covariances between the physical and BGC model components to let the assimilation algorithm utilise physical observations to directly update also BGC model variables. Strongly coupled data assimilation is challenging because it depends on the quality of the estimated cross-covariances and requires that compatible assimilation methods are used in the different model components. This appears to be a particular issue for the assimilation into coupled atmosphere-ocean models as the recent review by Penny et al. (2017) shows. Only a limited number of studies have so far considered the combined assimilation of physical and BGC obser- vations. However, while assimilating both physical and BGC observations, the published studies (Anderson et al. 2000; Ourmie`res et al. 2009; Song et al. 2016a, b; Mattern et al. 2017) all set the cross-covariances between differ- ent variables to zero. Thus, in terminology of coupled data assimilation, only the weakly coupled data assimilation was performed, in which the direct assimilation influence of the physical observation was only on the physical model fields, while the BGC observations had only a direct influ- ence on the modelled BGC concentrations. Only during the subsequent model forecast, or in iterations of a variational minimisation method, the changed model fields interacted. Nonetheless, the studies find that the combined weakly coupled assimilation of physical and BGC observations improved the overall consistency of the coupled model state. Until now, the strongly coupled assimilation into a cou- pled ocean-BGC model was only studied by Yu et al. (2018). The study used an idealised configuration of a channel with wind-induced upwelling and synthetically generated obser- vations, i.e. a twin experiment. Different combinations of the weakly and strongly coupled assimilation assimilating either physical (SSH, SST and temperature profiles) or BGC data (surface chlorophyll and nitrogen profiles) or assimilat- ing both data types were conducted. The experiments showed that in this idealised case, the cross-covariances between the physical and BGC model variables contain useful informa- tion that can be used in the strongly coupled assimilation. In this study, the effect of the strongly coupled assimilation in a realistic ocean-BGC model is assessed. For this purpose, the data assimilation is performed on the coastal coupled ocean-BGC model HBM-ERGOM configured for the North and Baltic Seas using two nested meshes. An earlier model version of the physical circulation model (BSHcmod, Dick et al. 2001; Kleine 2003) with a simpler model configuration without nesting was used in previous studies (Losa et al. 2012, 2014; Nerger et al. 2016) to assess the influence of SST assimilation. Only satellite SST data is assimilated here and the effect of both weakly and strongly coupled assimilation is assessed. A particular focus is on the question whether the strongly coupled assimilation of SST data, i.e. direct joint update of both the physical and BGC model fields, improves the model state in this coastal setup. A further aspect examined here is the different effect when treating the BGC model fields in the assimilation using the actual concentrations or the logarithm of them. Based on the fact that the chlorophyll concentrations can be well described as log-normally distributed (Campbell 1995), many studies employing ensemble Kalman filters