Characterizing Land Cover Structure with Semantic Variograms

Abstract

This paper introduces the semantic variogram, which is a measure of spatial variation based upon semantic similarity metrics calculated for nominal land cover class definitions. Traditional approaches for measuring spatial autocorrelation for nominal geographical data compare classes between pairs of observations to determine a simple binary measure of similarity (identical/different). These binary values are summarized over many sample pairs separated by various distances to characterize some spatial metric of correlation, or variation. The use of binary similarity measures ignores potentially substantial ranges in similarity between different classes. Through the development of category representations capable of producing quantifiable measures of pair wise class similarity, descriptive spatial statistics that operate upon ratio data may be employed. These measures, including the semantic variogram proposed in this work, may characterize spatial variability of categorical maps more sensitively than traditional measures. We apply the semantic variogram to National Land Cover Data (NLCD) for three different study sites, and compare results to those from a multiple class indicator semivariogram. We demonstrate that substantial differences exist in observed short-range variability for the two metrics in all sites. The semantic variograms detect much lower short-range variability due to the tendency of semantically similar classes to be closer together.