Design of a conceptual framework and approaches for geo-object data conflation

Issues of heterogeneity and incompatibility in geospatial data become increasingly important as data sources become more abundant. Scientific research and decision-making usually require geospatial data from a variety of producers, but it is not realistic to collect all data directly; therefore, it is important to utilize effectively data created by various agencies using different methodologies under different circumstances. The term conflation refers here to the problem of combining incompatible geospatial data. It has two major components: object matching and object transformation. This dissertation examines the characteristics of various conflation types and proposes a relational algebra framework for geo-object data conflation that contains two relational operators: probabilistic similarity join and consolidation. Uncertainty is introduced and evaluated in the conflation process as probability associated with alternative values of attributes derived from different sources. Lineage information is stored in a probabilistic database to connect alternative tuple values in the conflated dataset to tuples in source datasets from which they are derived. Storage of all alternative attributes enables a better estimate of attributes when more data are available without mistakenly discarding data of a seemingly lower quality.;New approaches have been developed to complete a conflation task. Object matching is first formulated as an assignment problem that takes into account all matched pairs simultaneously to achieve a global optimum in terms of percentage of correctly matched objects. This optimization method is compared to the widely used greedy method that matches objects one pair after another. As a result, a higher percentage of correctly matched pairs is achieved using the optimization method on test datasets. In addition, this optimized object-matching model is adjusted to address 1:m, m:1, 1:0, and 0:1 correspondences effectively for linear objects by taking advantage of the asymmetry of directed Hausdorff distance. Two approaches have been designed for datasets with independent and autocorrelated distortion. For datasets with autocorrelated distortion, a spatial transformation component is incorporated into the object-matching model as an affine transformation function to achieve a better matching result.