A Novel Method for Information Prediction

Many interpolation methods were proposed to predict or reconstruct unknown information. However, when the conditional data are quite few or even there are no conditional data, predicted results are often poor. Originally, a method called multiple-point geostatistics (MPS) originated from geostatistical fields and it allows extracting multiple-point structures from training images, after that MPS can copy these structures to the regions to be simulated. However, original MPS can only predict discretized variables. To overcome the disadvantage, a method using continuous MPS based on filters is proposed to predict the unknown information composed of continuous variables. Filters are used to realize dimension reduction, and a filter score space can be created using filters. All similar training patterns fall into a cell in the filter score space to create a prototype. During prediction, a training pattern from a cell is randomly drawn, and then is pasted back onto the simulation grid. Experimental results show that our method can effectively predict the unknown information of a region.