Land Cover Classification of Remotely Sensed Imagery UsingMultiple-point Geostatistics

Objective A post-processing method is proposed based on the theory of multiple-point geostatistics.The method extracts prior spatial structures from a training image,and infers the pattern distributionand correlation of classes.A spatial correlation model can be established from training image,which ispreferable to the traditional two-point-based variogram model.An experiment was performed on aLandsat TM image,wetlands with a complicated distribution were extracted.The method was com-pared to the spatial filtering and the contextual Markov random field(MRF)classifier.This approachincreases overall classification accuracy,and has advantages when dealing with classes that have curvi-linear distributions.