International audienceThis work considers a radio-astronomy inverse problem of physical parameters inference from multispectral images. The forward model is a numerical simulation, and the observation model mixes different sources of noise. This results in a non-explicit non-log-concave likelihood function. We introduce a likelihood approximation with controlled error that allow the conception of a Monte Carlo Markov Chain (MCMC) method. The obtained sampler provides credibility intervals along with point estimates. We believe that the proposed approach is sufficiently generic to be applied to similar inverse problems.Ce travail considère un problème inverse en astrophysique, qui consiste à estimer un ensemble de paramètres physiques à part...