Abstract:
This dissertation addresses the development of metrics for scene similarity assessment using abstract spatial relations among objects comprising of a scene. This is done in the context of an ontological approach as it can support scene matching and ontology classification. We consider that abstract spatial relations are more important than absolute and/or quantitative spatial relationships in deriving semantic features for an ontology driven scene similarity. The motivation of this study arises from the semantic gap that currently exists in the majority of scene modeling methods and the insufficiency on describing higher-level knowledge representation that applies in a specific spatial context. Therefore there is currently a need to develop novel metrics to successfully describe the abstract spatial context of complex features and better communicate this information in the context of ontologies. Our approach is based on the creation of spatial signatures for complex features as they express the spatial relations of their components. These relationships, which can be articulated through an extension of fuzzy Allen relationship, can then be quantified through the Histogram of Forces approach. Once such expressions are derived two or more scenes can be compared through the correlation analysis. Borrowing from data mining principles, the process can become computationally more effective by analyzing variations in the multiple relations that exist among various components. The proposed framework is applied in two representative test cases that express established and emerging geospatial analysis challenges. The first test case addresses structured built environments in satellite imagery, using airport compounds. The second test case addresses unstructured environments in crowdsourced data, using social media contribution patterns in the aftermath of natural disasters.
