Pattern extraction from spatial data - Statistical and modeling approache

Exploratory and statistical spatial data analyses are commonly used in a wide range of research fields, such as epidemiology, disease surveillance and crime analysis. Spatial epidemiology, for example, needs to detect significant spatial clusters of disease incidents to help epidemiologists identify environmental factors and spreading patterns associated with certain diseases and then take action accordingly. Existing spatial analysis approaches mostly focus on the analysis of spatial lattice data, i.e., observations are organized by locations such as county or census tract. With the wide spread of location-aware technologies such as GPS and smart phones, increasing amount of spatial interaction data become available, e.g., information about human daily mobility, traveling and migration.;The goal of this dissertation is to develop new methodologies for the analysis of both spatial lattice data and spatial interactions data, with a focus on statistical and modeling approaches. The contribution of this dissertation includes three new methodologies for spatial scan statistic (Chapter 2), flow scan statistic (Chapter 3), and spatial interaction modeling (Chapter 4).;The first developed methodology is a new spatial scan statistic incorporating smoothing and regionalization techniques. The contribution is three-fold: 1) the new method can detect irregular shaped spatial clusters, which is more efficient and effective than existing methods; 2) the method can alleviate the multiple-testing problem by dramatically reducing the cluster search space with hierarchical regionalization; and 3) the integration of a smoothing strategy addresses the small-area problem, which significantly improves the accuracy of cluster detection. The new method is evaluated with a series of benchmark data that are widely used in related literature.;The flow scan statistic developed in this research is specifically designed for spatial interaction data to detect significant flow clusters. To my best knowledge, this is the first scan statistic approach for spatial interaction data and it can extract significant flow clusters from very large origin-destination (OD) data sets such as migration and taxi trips. The proposed flow statistic scans the OD data with a flow tube (which consists of a neighborhood at the origin and another neighborhood at the destination) to search significantly higher-than-expected flows between locations. A test statistic based on the Generalized Likelihood Ratio (GLR) is specifically designed, which works with both area-based and point-based spatial interaction data. The new method is demonstrated and evaluated with case studies of the county-to-county migration data in U.S. and a synthetic point-based OD flow data.;The third method presented in this dissertation is a spatial interaction modeling and analysis framework that consists of (1) a piece-wise spatial interaction model to understand the global patterns; (2) an extended spatial autocorrelation statistic based on Moran's I to examine the spatial distribution of model residuals; and (3) a new mapping approach to visualize local flow patterns (spatial clusters of model residuals) that cannot be explained by the configured model and global patterns. The developed model takes in account the distance, size and an accessibility measure for each flow and its origin/destination. The model outcome (i.e., coefficients) can reveal interesting global patterns. Moreover, followed with the statistical analysis and mapping of model residuals, one can further investigate the local deviations from the global trend and be able to gain a comprehensive understanding of the complex patterns in spatial interaction data. A case study is developed to analyze the migration among Metropolitan Statistical Areas (MSAs) of the United States. The major contribution of proposed framework includes an extended Local Moran's I statistic for analyzing flow residuals and a novel mapping method for visualizing the flow residual patterns.;Although the three methods are presented separately, they are related in several ways. The first and second methods both focus on scan statistics, with the first one improving the existing spatial scan statistics by detecting irregular-shaped clusters and the second extending the existing methods by analyzing spatial interaction data (i.e., location-to-location flows). The second and third methods are both the analysis for spatial interaction data, with the second method developing the exploratory statistics for actual flow data and the third method focusing on spatial interaction modeling and residual analysis.