One difficulty with coupled physical-biological ocean models is determining optimal values of poorly known model parameters. The variational adjoint assimilation method is a powerful tool for the automatic estimation of parameters. We used this method to incorporate remote-sensed chlorophyll-a data into a coupled physical-biological model developed for the Bohai Sea and the Northern Yellow Sea. A 3-D NPZD model of nutrients (N), phytoplankton (P), zooplankton (Z) and detritus (D) was coupled with a physical model, the Princeton Ocean Model. Sensitivity analysis was carried out to choose suitable control variables from the model parameters. Numerical twin experiments were then conducted to demonstrate whether the spatio-temporal resolutions of the observations were adequate for estimating values of the control variables. Finally, based on the success of the twin experiments, we included remote-sensed chlorophyll-a data in the NPZD model. With the adjoint assimilation of these chlorophyll-a data, the coupled model better describes spring and autumn phytoplankton blooms and the annual cycle of phytoplankton at the surface layer for the study area. The annual cycle of simulated surface nutrient concentrations also agreed well with field observations. The adjoint method greatly improves the modelling capability of coupled ocean models, helping us to better understand and model marine ecosystems.