Multiple-point statistics (MPS) is a useful approach to reconstruct three-dimensional models in the macroscopic or microscopic field. Extracting spatial features for three-dimensional re-construction from two-dimensional training images (TIs), and characterizing non-stationary features with directional ductility are two key issues in MPS simulation. This study presents a step-wise MPS-based three-dimensional structures reconstruction algorithm with the sequential process and hi-erarchical strategy based on two-dimensional images. An extension method is proposed to construct three-dimensional TIs. With a sequential simulation process, an initial guess at the coarsest scale is simulated, in which hierarchical strategy is used according to the characteristics of TIs. To obtain a more refined realization, an expectation-maximization like iterative process with global optimization is implemented. A concrete example of chondrite micro-structure simulation, in which one scanning electron microscopy (SEM) image of the Heyetang meteorite is used as TI, shows that the presented algorithm can simulate complex non-stationary structures.