Model predictions for a rapid assessment and prognosis of possible radiological consequences after an accidental release of radionuclides play an important role in nuclear emergency management. Radiological observations, e.g. dose rate measurements, can be used to improve such model predictions. The process of combining model predictions and observations, usually referred to as data assimilation, is described in this article within the framework of the real time on-line decision support system (RODOS) for off-site nuclear emergency management in Europe. Data assimilation capabilities, based on Kalman filters, are under development for several modules of the RODOS system, including the atmospheric dispersion, deposition, food chain and hydrological models. The use of such a generic data assimilation methodology enables the propagation of uncertainties throughout the various modules of the system. This would in turn provide decision makers with uncertainty estimates taking into account both model and observation errors. This paper describes the methodology employed as well as results of some preliminary studies based on simulated data.