Multiple-point geostatistical simulation for post-processing a remotely sensed land cover classification

Abstract

A post-processing method for increasing the accuracy of a remote sensing classification was developed and tested based on the theory of multiple-point geostatistics. Training images are used to characterise the joint variability and joint continuity of a target spatial pattern, overcoming the limitations of two-point statistical models. Conditional multiple-point simulation (MPS) was applied to a land cover classification derived from a remotely sensed image. Training data were provided in the form of “hard” (land cover labels), and “soft” constraints (class probability surfaces estimated using soft classification). The MPS post-processing method was compared to two alternatives: traditional spatial filtering (also a post-processing method) and the contextual Markov random field (MRF) classifier. The MPS approach increased the accuracy of classification relative to these alternatives, primarily as a result of increasing the accuracy of classification for curvilinear classes. Key advantages of the MPS approach are that, unlike spatial filtering and the MRF classifier, (i) it incorporates a rich model of spatial correlation in the process of smoothing the spectral classification and (ii) it has the advantage of capturing and utilising class-specific spatial training patterns, for example, classes with curvilinear distributions.