In the event of a release of radioactive pollutants into the atmosphere, one of the missions of the authorities is to evaluate the consequences of this release in order to implement, if necessary, measures to protect the population. These may include evacuation or sheltering in the very short term, and restrictions on the consumption or marketing of contaminated foodstuffs in the longer term. For this purpose, numerical models are used to simulate the dispersion of radionuclides in the atmosphere.The accuracy of the results obtained from these models strongly depends on the knowledge of the source term, i.e. the location, duration and magnitude of the release as well as its distribution between radionuclides. However, knowledge of the source term is generally subject to significant uncertainties. In addition to the source term, other uncertainties arise from the transport model, meteorological fields, measurement data and the representativeness of the model with respect to the measurements.In this PhD, we have developed and applied inverse modelling methods to improve the evaluation of the source term and to quantify the uncertainties.Among the inverse modelling methods, variational deterministic approaches are effective in providing a rapid estimate of the source term, but the quantification of uncertainties associated with this estimate is generally difficult.We therefore propose to address the problem within the probabilistic framework of Bayesian inference, which is part of a formalism that allows a more complete assessment of uncertainties.Several Markov chain Monte Carlo (MCMC) sampling methods are implemented in order to reconstruct the variables describing the source: the Metropolis Hastings (MH) algorithm, the Parallel tempering algorithm and finally the Reversible-Jump MCMC.These algorithms are first applied and validated on the ruthenium 106 detection event that occurred in Europe in the fall of 2017. The probability densities of the source variables are reconstructed in order to identify the geographical origin of the detections as well as the quantities of ruthenium 106 released into the atmosphere.Then, in a second step, several methods are developed in order to incorporate and quantify different sources of errors within the Bayesian problem and thus enable us to obtain a better reconstruction of the distribution of the release.The second case study is dedicated to the Fukushima Daiichi plant accident which led to long releases associated with time-varying kinetics. The reconstruction of these releases with more complex characteristics required the development of a new MCMC algorithm, the Reversible-Jump MCMC, which was adapted from previous sampling methods.Applied to the Fukushima case, the Reversible-Jump MCMC shows its ability to sample more finely and efficiently the distribution of the source term and uncertainties.