Incorporating Spectral Similarity Into Markov Chain Geostatistical Cosimulation for Reducing Smoothing Effect in Land Cover Postclassification

Abstract

Spatial statistics provides useful methods for incorporating spatial dependence into land cover classification. However, the geometric features of land cover classes are difficult to be captured by geostatistical models due to smoothing effect. The objective of this study is to incorporate spectral similarity into the Markov chain random field (MCRF) cosimulation (coMCRF) model, that is, to propose a spectral similarity-enhanced MCRF cosimulation (SS-coMCRF) model, for land cover postclassification so that postclassification will cause less geometric loss. Two mutually complementary spectral similarity measures, Jaccard index and the spectral correlation measure, were employed as a constraining factor in SS-coMCRF. One medium spatial resolution scene with a complex landscape and one very high spatial resolution scene with an urban landscape were selected for case studies. Neural network classifier and support vector machine classifier were used to conduct land cover preclassifications. Both coMCRF and SS-coMCRF were used to postprocess preclassified images based on expert-interpreted sample datasets from multiple data sources. Compared with preclassified results that depend on only spectral information of pixels, postclassifications by both models achieved similar significant improvements in overall accuracy. However, compared with coMCRF, the SS-coMCRF model apparently improved postclassified land cover patterns by effectively capturing some geometric features (e.g., boundaries and linear stripes) and more details of land cover classes. In general, incorporating spectral similarity into land cover postclassification through SS-coMCRF may contribute significantly to the “shape” or geometric accuracy of classified land cover classes.