Visualizing Spatial Uncertainty of Multinomial Classes

Area-class maps are conventionally delineated by human hands based on limited observed (or high-quality) data and expert knowledge. It is recognized that areaclass maps contain spatial uncertainty because the classification of unobserved locations is not completely certain. Spatial uncertainty associated with an area-class map may be quantified by spatial statistical approaches through estimating the occurrence probability of a class at each unobserved location. In this study, we use both indicator kriging and the recently proposed Markov chain geostatistics (MCG) to simulate spatial distribution of multinomial classes and assess spatial uncertainties associated with simulated results. The spatial uncertainties are represented by occurrence probability vectors, which are further visualized as occurrence probability maps. Results show that given the same observed dataset the spatial uncertainty assessed by MCG is apparently lower than that assessed by indicator kriging. In simulated realizations, MCG generates apparently higher PCCs (percentages of correctly classified locations) than indicator kriging. It is concluded that the spatial uncertainty assessed by MCG should be closer to the real spatial uncertainty associated with an area-class map.