Modeling and simulation provide essential knowledge on the aerial dispersion of gases and particles and the resulting environmental marking. This applies in particular to the releases that were generated by atmospheric nuclear tests carried out in the past by France in Polynesia. While regional-scale meteorological and dispersion calculations are reasonably reliable, their results have a degree of uncertainty and present discrepancies with heterogeneous measurements of activities or dose rates in the air, on the ground and in biological compartments. The thesis will aim to develop inversion methods, based on data assimilation, in order to reduce errors and uncertainties in simulations of regional dispersion of radionuclides. The application will concern certain nuclear tests in the atmosphere. However, the methods developed during the thesis, such as Monte Carlo sampling by Markov chains, will have a more general field of implementation. After a literature review on nuclear testing and data assimilation methods, original inverse modeling algorithms will be programmed, tested, and applied to the simulation of the dispersion of aerial releases from tests. This will allow us to estimate the anticipated important role of measurement assimilation in improving simulations.