The task of the decision support in the case of a radiation accident is to provide up-to-date information on the radiation situation, prognosis of its future evolution and possible consequences. The reliability of predictions can be significantly improved using data assimilation, which refers to a group of mathematical methods allowing an efficient combination of observed data with a numerical model. The dissertation concerns application of the advanced data assimilation methods in the field of radiation protection. We focus on assessment of off-site consequences in the case of a radiation accident. The main contribution of this thesis is the development of sequential data assimilation methods for the early and the late phase of a radiation accident. Data assimilation is understood here as a particular case of recursive Bayesian estimation. Instead of using traditional estimation methods for state-space models based on Kalman filtering, we focused on sequential Monte Carlo methods, specifically particle filtering and marginalized particle filtering. Firstly, data assimilation methodology for the early phase of an accident was developed. It employs particle filtering with adaptive selection of proposal density for estimation of the most important variables describing the aerial propagation of radionuclides. The general methodology is applicable to all parametrized atmospheric dispersion models. It was demonstrated on a simulated release, where a bias of the basic meteorological inputs and the source term were corrected using available gamma dose measurements. Secondly, for the purpose of data assimilation in the late phase, we extended the idea of marginalized particle filtering to analytically intractable approximate filters. The result is a hybrid data assimilation methodology, where multiple ensemble filters are run in parallel. The methodology was applied to joint estimation of the spatial distribution of deposition on terrain and estimation of the speed of radionuclides removal due to environmental processes in a simulated release scenario. The proposed algorithms implemented in the decision support system HARP (HAzardous Radioactivity Propagation).