Direct numerical simulations of pore-scale flow and transport in porous media require the knowledge of pore-space topology. Real porous systems, e.g. rocks, pose additional challenges since they usually exhibit multi-modal distributions in physical and chemical properties. These hierarchical media cannot be approached by a single continuum formulation. Moreover, multimodal distribution in physical properties also affect the accuracy of X-ray tomographic (XCT) images of natural rocks, whose resolution is often insufficient to fully characterize pore-space structure within fine-grained regions. In the first part of the talk, we focus on sequential homogenization techniques, that build a hierarchy of effective equations that sequentially carry the smallest scale information through the intermediate scales up to the macroscale. Yet, existence of one or multiple intermediate scales can significantly decrease the accuracy of multiscale formulations. We show that the accuracy of multiscale methods based on sequential upscaling approaches is strongly influenced by a combination of geometric and dynamical scale separation conditions. In the second part of the talk, we propose an algorithm of downscaling followed by segmentation to reconstruct the unresolved pore space from XCT images of natural geological porous media. The method allows to generate a high-resolution binary image of the porous medium from poorly resolved grey-scale images while preserving available sub-resolution information. We validate the method on synthetic unresolved images and compare their known pore-space distribution with the extracted one. The comparison shows that the method better represents the initial distribution than threshold-based segmentation. We apply the method to extract the pore space distribution from unresolved XCT images of two natural sediment columns, and use it to parametrize a capillary bundle model. The latter is then used to estimate the hydraulic conductivity by matching breakthrough behavior of passive solute transport.