Abstract

Geographic Information Systems (GIS) data for a region is of different types and collected from different sources, such as aerial digitized color imagery, elevation data consisting of terrain height at different points in that region, and feature data consisting of geometric information and properties about entities above/below the ground in that region. Merging GIS data and understanding the real world information present explicitly or implicitly in that data is a challenging task. This is often done manually by domain experts because of their superior capability to efficiently recognize patterns, combine, reason, and relate information. When a detailed digital representation of the region is to be created, domain experts are required to make best-guess decisions about each object. For example, a human would create representations of entities by collectively looking at the data layers, noting even elements that are not visible, like a covered overpass or underwater tunnel of a certain width and length. Such detailed representations are needed for use by processes like visualization or 3D modeling in applications used by military, simulation, earth sciences and gaming communities. Many of these applications are increasingly using digitally synthesized visuals and require detailed digital 3D representations to be generated quickly after acquiring the necessary initial data.

Our main thesis, and a significant research contribution of this work, is that this task of creating detailed representations can be automated to a very large extent using a methodology which first fuses all Geographic Information System (GIS) data sources available into knowledge base (KB) assertions (instances) representing real world objects using a subprocess called GIS2KB. Then using reasoning, implicit information is inferred to define detailed 3D entity representations using a geometry definition engine called KB2Scene. Semantic Web is used as the semantic inferencing system and is extended with a data extraction framework. This framework enables the extraction of implicit property information using data and image analysis techniques. The data extraction framework supports extraction of spatial relationship values and attribution of uncertainties to inferred details. Uncertainty is recorded per property and used under Zadeh fuzzy semantics to compute a resulting uncertainty for inferred assertional axioms. This is achieved by another major contribution of our research, a unique extension of the KB ABox Realization service using KB explanation services.

Previous semantics based research in this domain has concentrated more on improving represented details through the addition of artifacts like lights, signage, crosswalks, etc. Previous attempts regarding uncertainty in assertions use a modified reasoner expressivity and calculus. Our work differs in that separating formal knowledge from data processing allows fusion of different heterogeneous data sources which share the same context. Imprecision is modeled through uncertainty on assertions without defining a new expressivity as long as KB explanation services are available for the used expressivity. We also believe that in our use case, this simplifies uncertainty calculations. The uncertainties are then available for user-decision at output. We show that the process of creating 3D visuals from GIS data sources can be more automated, modular, verifiable, and the knowledge base instances available for other applications to use as part of a common knowledge base. We define our method’s components, discuss advantages and limitations, and show sample results for the transportation domain.