Modern and ancient genomes are not necessarily drawn from homogeneous populations, as they may have been collected from different places and at different times. This heterogeneous sampling can be an issue for demographic inferences and results in biased demographic parameters and incorrect model choice if not properly considered. When explicitly accounted for, it can result in very complex models and high data dimensionality that are difficult to analyse. In this paper, we formally study the impact of such spatial and temporal sampling heterogeneity on demographic inference, and we introduce a way to circumvent this problem. To deal with structured samples without increasing the dimensionality of the site frequency spectrum (SFS), we introduce a new structured approach to the existing program fastsimcoal2. We assess the efficiency and relevance of this methodological update with simulated and modern human genomic data. We particularly focus on spatial and temporal heterogeneities to evidence the interest of this new SFS-based approach, which can be especially useful when handling scattered and ancient DNA samples, as in conservation genetics or archaeogenetics.
