A descriptor-based method combined with a partition approach is proposed to reconstruct three-dimensional (3D) microstructures based on a set of two-dimensional (2D) scanning electron microscopy (SEM) images. The features in the SEM images are identified and partitioned into small features using the watershed algorithm. The watershed algorithm first finds the local gray-level maxima, and partitions the features through the gray-level local minima. The 3D size distribution and radial distribution of the small spherical elements are inferred, respectively, based on the 2D size distribution and radial distribution using stereological analysis. The 3D microstructures are reconstructed by matching the inferred size distribution and radial distribution through a simulated annealing-based procedure. Combining with the proposed partition approach, the descriptor-based method can be applied to complex microstructures and the computational efficiency of the reconstruction can be largely improved. A case study is presented using a set of 2D SEM images with nanoscale pore structure from the low-density CSH (calcium silicate hydrate) phase of a hardened cement paste. Cross sections were randomly selected from the reconstructed 3D microstructure and compared with the original SEM images using the pore descriptors and the two-point correlation function with satisfactory agreement. Using the 3D reconstructed model, the properties of the sample material can be investigated on such a small scale as demonstrated in this paper on quantifying the absolute permeability.