A Density-Based Spatial Flow Cluster Detection Method

Abstract

GIScience 2016 Short Paper Proceedings A Density-Based Spatial Flow Cluster Detection Method Ran Tao 1 , Jean-Claude Thill 1 Dept. of Geography & Earth Sciences, University of North Carolina at Charlotte, 9201 University City Blvd, Charlotte, NC Email:{ rtao2; jfthill}@uncc.edu Abstract Understanding the patterns and dynamics of spatial origin-destination flow data has been a long- standing goal of spatial scientists. In this paper we introduce a density-based cluster detection method tailored for disaggregated spatial flow data. The basic idea is to first measure flow density considering both endpoint coordinates and flow lengths, and combine it with state-of-art density-based clustering methods. We experiment with a carefully designed synthetic dataset. The results prove that our method can effectively extract flow clusters from various situations encompassing varied flow densities, lengths, hierarchies and, at the same time, avoid issues of Modifiable Areal Unit Problem (MAUP) of flows endpoints, loss of spatial information, and false positive errors on short flows. 1. Introduction Spatial flows, also known as spatial interactions (SI) between georeferenced places, have been an enduring study object in a wide range of research fields. With the widespread adoption of location-aware technologies and the global diffusion of geographic information systems (GIS), spatial interaction data have been enriched in several respects including volume, type, availability, ubiquity, and spatiotemporal granularity (Yan and Thill 2009; Guo et al. 2012). While it brings unprecedented opportunities to improve our understanding of SI processes and thus enriching SI theories, it also brings the analytical challenge of developing more data-drive approaches tailored for SI data (Yan and Thill 2009). As a common data mining technique, cluster detection has proved useful in exploratory analysis of large sets of spatial flows. One approach measures the spatial relationships among origins and destinations, respectively, before combining them, as the basis for clustering flows. Here, spatial relationships can be contiguity or proximity of origin or destination regions (Guo 2009; Zhu and Guo 2014). However these methods are sensitive to uneven distribution and ad hoc zoning definition of flow endpoints; besides they are prone to false positive errors on short- distance interactions. Another type of methods use flow geometry to bundle nearby ones (Cui et al. 2008). While the results usually have desirable visual clarity, these methods compromise through loss of valuable spatial information. In this paper we introduce a new method that not only can extract spatial flow clusters from various situations including varying flow densities, lengths, hierarchies, but also avoids problems like MAUP, false positive errors, and loss of information. 2. Methodology Of various clustering methods, we choose to design our flow clustering method in the density- based tradition because of its capability to discover clusters of arbitrary shape and to filter out noise. Moreover, density-based methods like OPTICS (Ankerst et al. 1999) can effectively reveal hierarchical structures in the data since its byproduct, the reachability plot, is convertible to a dendrogram (Sander et al. 2003; Campello et al. 2013). Hereafter, we first introduce the proximity metric tailored to spatial flows; then we explain the clustering method step by step.