Modelling semantic uncertainty of land classification system

Abstract

Land use and land cover information are fundamental for study of earth's ecosystem, global carbon cycling, global climate, atmospheric composition, energy and water balance, biodiversity, ecologically mediated diseases, and other concerns. Several global, regional and national land inventory projects have been carried out to acquire land classification data, e.g., AFRICOVER, CORINE, NLCD, etc. Correspondingly, a vast number of land classification systems have been developed using different philosophical principles to satisfy different purposes. Semantic problems, e.g., semantic overlap, have been identified by few researchers. However, classification systems are usually applied without considering their semantic problems that will cause confusion and could be misleading. Therefore, to address this research gap, this thesis focuses on the semantic uncertainties of classes in classification systems. First, classes are formalized for quantitative calculation. Based on the characteristics of land class definitions, classes are divided into two types: concept and category. A concept can be rebuilt by applying product operations and union operations, whereas a category can be formalized using an equation set of concepts. A class can always be formalized by applying product operations and union operations. Second, a reference system is established to uniquely represent all classes. The reference system is set up based on the contrast among classes using a bottom-up method through addition of classes step by step. A reference system is composed of reference concepts, which contain contrast components, not-contrast components, and complement components. All classes can be optimally and economically represented by a combination of reference concepts. Finally, different models for measuring semantic uncertainties are proposed based on the reference system. These models are divided into three groups: (1) uncertainties between classes, including semantic overlap and semantic similarity; (2) uncertainties between hierarchical levels, including semantic gap and semantic overflow; and (3) semantic interoperability between different classification systems. Characteristics of these models are also analysed in this thesis. Throughout the thesis, the National Land Cover Database Classification Systems (NLCD CS) of the United States are used for demonstration. The results reveal that the proposed theories and models are feasible and that semantic uncertainties are widespread in the NLCD CSs.