Numerous methods have been developed to geolocate fish from data storage tag data. Whereas demersal species have been tracked using tide-driven geolocation models, pelagic species which undertake extensive migrations have been mainly tracked using light-based models. Here, we present a new HMM-based model that can infer pelagic fish positions from the sole use of high-resolution temperature and depth histories. A key contribution of our model lies in model parameter inference (diffusion coefficient and errors on the reference geophysical fields - satellite SST and MARS3D hydrodynamic model), which improves model robustness. As a case study, we consider long time-at-liberty tagging data of European sea bass. For the first time, reconstructed individual trajectories are provided. They evidence strong site fidelity on summer feeding grounds and suggest site fidelity on winter spawning grounds. This result on the fine-scale spatial structuring of the population is a key issue poorly addressed by current management frameworks. Beyond seabass behavioral ecology, our model opens new avenues for the reconstruction and analysis of migratory patterns of other pelagic species in relatively contrasted geophysical environments.