| The movement of people, goods, capital and information over space, generally known as "spatial interaction", has been a long-standing concern in geography due to the complex nature of its underlying processes. When studying this type of data, we are often faced with a large amount of data with multiple dimensions, which can easily overwhelm traditional spatial analytical methods, developed for conditions when little information is available, and thus many assumptions need to be made. Furthermore, aside from the conventional entropy and microeconomic foundations of spatial interaction models, we have relatively few insights into the processes through which different types of spatial interaction patterns are formed among different geographic regions. Spatial structure and spatial interdependencies in particular is still very much handled in these models in an ad hoc fashion.; In this dissertation, I have adopted self-organizing maps (SOM) as a technique to reduce the complexity of spatial interaction data, both in terms of data amount and dimensionality, at once. SOM is basically used as a Spatial Exploratory Data Analysis (ESDA) tool to suggest new hypotheses for explaining the movement across space. Using a large domestic air travel dataset as case study, this dissertation studies how the characteristics of the air transport system interact with the spatial interaction system to create relationships among sets of trip origins and sets of trip destinations, as well as sets of O-D pairs. In order to fully explore the patterns in SOM feature maps and geographic maps of flows and movements, an interactive visual data mining (VDM) environment is developed in which different visualization forms are linked together. Findings from the case studies support the contention that SOM is well capable of picking out structures in a large spatial interaction database and can be used to examine the properties constitutive of clusters. By integrating SOM feature maps with other visualization forms, especially geographic maps, we are able to explore how the structures in the attribute space are distributed in the actual geographic space. This capability is proven to be vital during any geographic knowledge extraction process. |
