Unsupervised Delineation of Urban Structure Types Using High Resolution RGB Imagery

Abstract

GIScience 2016 Short Paper Proceedings Unsupervised Delineation of Urban Structure Types Using High Resolution RGB Imagery J. Niesterowicz 1 , T. F. Stepinski 1 , J. Jasiewicz 1,2 Space Infromatics Lab, University of Cincinnati, 401 Braunstein Hall, Cincinnati, OH 45221-0131, US Email: {niestejk; stepintz}@mail.uc.edu Instititute of Geoecology and Geoinformation, Adam Mickiewicz University, Dziegielowa 27, Poznan, Poland Email: jarekj@amu.edu.pl Abstract We present a method for delineating Urban Structure Types (USTs) using only high resolution RGB images. As the method is unsupervised, it does not require training; the interpretations of delineated USTs are assigned a posteriori. The method utilizes freely available software and performs delineation in a short time even for very large images. A 1-meter resolution image of the entirety of Los Angeles is delineated as an example. We have found seven distinct USTs which were given interpretations based on examination of their patterns. These interpretations are validated by population statistics. The method aims at broaden the usage of USTs delineations for applications in urban and social studies. 1. Introduction Urban Structure Type (UST) is a distinct spatial pattern of the urban structure at the neighborhood scale, which can be interpreted in terms of the type of activity or of residential pattern. Classification of a city into USTs complements standard land cover/land use classification by working at a scale that is significantly coarser than an individual pixel. Fairly extensive literature exists on how to delineate USTs from remotely sensed data (for example, see Heiden et al. 2012), but, because these works focus on supporting effective urban planning, they use multisource data and supervised learning. This means that they are restricted to very few places where this data exists and where the significant cost of supervised analysis is justified by the need. There also exists extensive literature on using single-source data (RGB or multispectral images) but only in the context of separating two specific types of USTs – formal from informal settlements (slums); for example see Graesser et al. (2012). Algorithms proposed there are restricted to this single purpose; they also are predominantly based on supervised learning. In this paper we present an approach to delineation of USTs that uses only RGB images (many of which are freely available online) as input, delineates an exhaustive set of USTs, is based on training-free, data-driven unsupervised principles, and can process very large input data in a reasonable time. In addition, our method relies only on existing public domain software. Our motivation is to make the delineation of USTs more broadly accessible to analysts from different disciplines. The methodology is described and applied to a ~2 billion pixel 1 m-resolution image covering the greater Los Angeles area. 2. Methodology Our method is based on the concept of Complex Object-Based Image Analysis (COBIA) (Vatsavai 2013; Stepinski et al. 2015). In COBIA a raster (not necessarily restricted to an image) is divided arbitrarily into a grid of local blocks of cells. We refer to these blocks as motifels –