Identification of internal structures in an active volcano is mandatory to quantify the physical processes preceding eruptions. We propose a fully unsupervised Bayesian inversion method that uses the point compound dislocation model as a complex source of deformation, to dynamically identify the substructures activated during magma migration. We applied this method at Piton de la Fournaise. Using 7-day moving trends of GNSS data preceding the June 2014 eruption, we compute a total of 15 inversion models of 2.5 million forward problems each, without a priori information. Obtained source shapes (dikes, prolate ellipsoids or pipes) exhibit a global migration from 7-8 km depth to the surface, drawing a “mechanical tomography”? of the plumbing system. Our results allow retrieving geometries compatible with observed eruptive fissures and seismicity distribution, and the retrieved source volume variations made this method a good proxy to anticipate erupted lava in case of no co-eruptive refilling.