USING HEURISTIC METHODS AND MACHINE LEARNING TO ENHANCE THE POSITIONAL ACCURACY OF HISTORICAL GEOSPATIAL DATASETS

The recent availability of multiple geospatial datasets can be attributed to advancements in location-based technologies. The merging of various datasets, commonly known as conflation, is essential to strengthening the content of existing datasets by integrating information from various sources. However, complexities arise when these merged datasets contain distinct representations of the same real-world entities with differing accuracy, projections, data structure, and consideration of details. Although conflation has been an interest to researchers for a long time, the existing methods do not cater to the specific needs of some datasets. For example, historical datasets have lower resolution with richer attribute information, while newer datasets are more positionally accurate but lack detailed or relevant attribute information. In addition, contemporary conflation algorithms usually demand manual review and verification for these datasets as the conflation process still lacks an efficient automated technique.

Through the work in this dissertation, both historical and modern datasets are investigated and explored using optimization and learning-based tools to aid in identifying corresponding and unifying elements in these datasets. The elements identified through these algorithms are validated against verified spatial information known as the ground truth dataset. Research results can inform practitioners with various tool options and notably reduce the time required to merge datasets compared to manual methods.