Abstract: Conditional data influence the reconstructed results greatly in the reconstruction of spatial data. Reconstructed results often show a number of uncertainties when only sparse conditional data are available, so it is suitable to use indefinite interpolation to reconstruct spatial data. As one of the main indefinite interpolation methods, multiple-point statistics (MPS) can extract the intrinsic features of patterns from training images and copy them to the simulated regions. Because the traditional MPS methods using linear dimensionality reduction are not suitable to deal with nonlinear data, isometric mapping (ISOMAP) is combined with MPS to address the above issues. A method using MPS and ISOMAP for the reconstruction of spatial data is proposed for the accurate reconstruction of unknown spatial data by constructing pattern dataset, dimensionality reduction of patterns, classification of patterns and extraction of patterns, which has provided a new idea for dealing with nonlinear spatial data by MPS. The experimental results show that the structural characteristics of reconstructed results using this method are similar to those of training images.

Key words: training image, multiple-point statistics (MPS), pattern, isometric mapping (ISOMAP), dimensionality reduction