UC Berkeley
UC Berkeley Previously Published Works
Title
A multi-scale analysis of 27,000 urban street networks: Every US city,
town, urbanized area, and Zillow neighborhood
Permalink
https://escholarship.org/uc/item/80n7572n
Journal
Environment and Planning B: Urban Analytics and City Science, 47(4)
ISSN
2399-8083 2399-8091
Author
Boeing, Geoff
Publication Date
2018-08-08
DOI
10.1177/2399808318784595
Data Availability
The data associated with this publication are available at:
https://dataverse.harvard.edu/dataverse/osmnx-street-networks
Peer reviewed
eScholarship.org Powered by the California Digital Library
University of California
Article
Urban Analytics and
City Science
A multi-scale analysis of
27,000 urban street
networks: Every US city,
town, urbanized area, and
Zillow neighborhood
Geoff Boeing
University of California, USA
Abstract
OpenStreetMap offers a valuable source of worldwide geospatial data useful to urban researchers. This
study uses the OSMnx software to automatically download and analyze 27,000 US street networks
from OpenStreetMap at metropolitan, municipal, and neighborhood scales—namely, every US city and
town, census urbanized area, and Zillow-defined neighborhood. It presents empirical findings on US
urban form and street network characteristics, emphasizing measures r eleva nt to graph theory,
transportation, urban design, and morphology such as structure, connectedness, density, centrality,
and resilience. In the past, street network data acquisition and processing have been challenging and ad
hoc. This study illustrates the use of OSMnx and OpenStr eetMap to consistently conduct street
network analysi s with extremely large sample sizes, with clearly defin ed network definiti ons and
extents for repr oducibility, and using nonplanar, directed graphs. These street networks and
measures data have been shared in a public repository for other resear chers to use.
Keywords
GIS, network analysis, OpenStreetMap, street networks, urban form, urban morphology
Introduction
On 20 May 1862, Abraham Lincoln signed the Homestead Act into law, making land across
the United States’ Midwest and Great Plains available for free to applicants (Porterfield,
2005). Under its auspices over the next 70 years, the federal government distributed 10% of
the entire US landmass to private owners in the form of 1.6 million homesteads (Lee, 1979;
Sherraden, 2005). New towns with gridiron street networks sprang up rapidly across the
Great Plains and Midwest, due to both the prevailing urban design paradigm of the day and
the standardized rectilinear town plats used repeatedly to lay out instant new cities
(Southworth and Ben-Joseph, 1997). Through path dependence, the spatial signatures of
Corresponding author:
Geoff Boeing, School of Public Policy and Urban Affairs, Northeastern University, 360 Huntington Ave, 310 Renaissance
Park, Boston, MA 02115, USA.
Email: g.boeing@northeastern.edu
EPB: Urban Analytics and City Science
2020, Vol. 47(4) 590–608
! The Author(s) 2018
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/2399808318784595
journals.sagepub.com/home/epb
these land use laws, design paradigms, and planning instruments can still be seen today in
these cities’ urban forms and street networks. Cross-sectional analysis of American urban
form can reveal these artifacts and histories through street networks at metropolitan,
municipal, and neighborhood scales.
Network analysis is a natural approach to the study of cities as complex systems (Masucci
et al., 2009). The empirical literature on street networks is growing ever richer, but suffers
from some limitations—discussed in detail in Boeing (2017) and summarized here. First,
sample sizes tend to be fairly small due to data availability, gathering, and processing
constraints: most studies in this literature that conduct topological or metric analyses tend
to have sample sizes ranging around 10 to 50 networks (Barthelemy and Flammini, 2008;
Buhl et al., 2006; Cardillo et al., 2006; Strano et al., 2013), which may limit the
generalizability and interpretability of findings. Second, reproducibility has been difficult
when the dozens of decisions that go into analysis—such as spatial extents, topological
simplification and correction, definitions of nodes and edges, etc.—are ad hoc or only
partly reported (e.g. Porta et al., 2006; Strano et al., 2013). Third, and related to the first
two, studies frequently oversimplify to planar or undirected primal graphs for tractability
(e.g. Barthelemy and Flammini, 2008; Buhl et al., 2006; Cardillo et al., 2006; Masucci et al.,
2009), or use dual graphs despite the loss of geographic, metric information (Batty, 2005;
Crucitti et al., 2006a, 2006b; Jiang and Claramunt, 2002; Ratti, 2004).
This study addresses these limitations by conducting a morphological analysis of urban street
networks at multiple scales, with large sample sizes, with clearly defined network definitions and
extents for reproducibility, and using nonplanar, directed graphs. In particular, it examines
27,000 urban street networks—represented as primal, nonplanar, weighted multidigraphs
with possible self-loops—at multiple overlapping scales across the US, focusing on structure,
connectedness, centrality, and resilience. It examines the street networks of every incorporated
city and town, census urbanized area, and Zillow-defined neighborhood in the US. To do so, it
uses OSMnx
1
—a new street network research toolkit (Boeing, 2017)—to download, model, and
analyze these street networks at metropolitan, municipal, and neighborhood scales. These street
networks and measures data sets have been compiled and shared in a public repository at the
Harvard Dataverse
2
for other researchers to use.
The purpose of this paper is threefold. First, it describes and demonstrates a new
methodology for easily and consistently acquiring, constructing, and analyzing large samples
of street networks as nonplanar directed graphs. Second, it presents empirical findings of
descriptive urban morphology for the street networks of every US city, urbanized area, and
Zillow neighborhood. Third, it investigates with large sample sizes some previous smaller-
sample findings in the research literature. This paper is organized as follows. In the next
section, it discusses the data sources, tools, and methods used to collect, model, and analyze
these street networks. Then, it presents findings of the analyses at metropolitan, municipal, and
neighborhood scales. Finally, it concludes with a discussion of these findings and their
implications for street network analysis, urban morphology, and city planning.
Methodology
A network (also called a graph in mathematics) comprises a set of nodes connected to one
another by a set of edges. Street networks can be conceptualized as primal, directed,
nonplanar graphs. A primal street network represents intersections as nodes and street
segments as edges. A directed network has directed edges: that is, edge uv points one-way
from node u to node v, but there need not exist a reciprocal edge vu.Aplanar network can be
represented in two dimensions with its edges intersecting only at nodes (O’Sullivan, 2014;
Boeing 591
Viana et al., 2013). Most street networks are nonplanar—due to grade-separated
expressways, overpasses, bridges, tunnels, etc.—but most quantitative studies of urban
street networks represent them as planar (e.g. Barthelemy and Flammini, 2008; Buhl
et al., 2006; Cardillo et al., 2006; Masucci et al., 2009; Strano et al., 2013) for tractability
because bridges and tunnels are uncommon in some cities. Planar graphs may reasonably
model the street networks of old European town centers, but poorly model the street
networks of modern autocentric cities like Los Angeles or Shanghai with many grade-
separated expressways, bridges, and underpasses (Boeing, 2018b).
Study sites and data acquisition
This study uses OSMnx to download, construct, correct, analyze, and visualize street
network graphs at metropolitan, municipal, and neighborhood scales. OSMnx is a
Python-based research tool that easily downloads OpenStreetMap data for any place
name, address, or polygon in the world, then constructs it into a spatially-embedded
graph-theoretic object for analysis and visualization (Boeing, 2017). OpenStreetMap is a
collaborative worldwide mapping project that makes its spatial data available via various
APIs (Corcoran et al., 2013; Jokar Arsanjani et al., 2015). These data are of high quality and
compare favorably to CIA World Factbook estimates and US Census TIGER/Line data
(Frizzelle et al., 2009; Haklay, 2010; Maron, 2015; Over et al., 2010; Wu et al., 2005; Zielstra
and Hochmair, 2011). In 2007, OpenStreetMap imported the TIGER/Line roads (2005
vintage) and since then, many community-led corrections and improvements have been
made (Willis, 2008). Many of these additions go beyond TIGER/Line’s scope, including
passageways between buildings, footpaths through parks, bike routes, and detailed feature
attributes such as finer-grained street classifiers, speed limits, etc.
To define the study sites and their spatial boundaries, we use three sets of geometries. The
first defines the metropolitan-scale study sites using the 2016 TIGER/Line shapefile of US
Census Bureau urban areas. Each census-defined urban area comprises a set of tracts that
meet a minimum density threshold (US Census Bureau, 2010). We retain only the urbanized
areas subset of these data (i.e. areas with greater than 50,000 population), discarding the
small urban clusters subset. The second set of geometries defines our municipal-scale study
sites using 51 separate TIGER/Line shapefiles (again, 2016) of US Census Bureau places
within all 50 states plus DC. We discard the subset of census-designated places (i.e. small
unincorporated communities) in these data, while retaining every US city and town. The
third set of geometries defines the neighborhood-scale study sites using 42 separate shapefiles
from Zillow, a real estate database company. These shapefiles contain neighborhood
boundaries for major cities in 41 states plus DC. This fairly new data set comprises nearly
7000 neighborhoods, but as Schernthanner et al. (2016) point out, Zillow does not publish
the methodology used to construct these boundaries. However, despite its newness it already
has a track record in the academic literature: Besbris et al. (2015) use Zillow boundaries to
examine neighborhood stigma and Albrecht and Abramovitz (2014) use them to study
neighborhood-level poverty in New York.
For each of these geometries, we use OSMnx to download the (drivable, public) street
network within it, a process described in detail in Boeing (2017) and summarized here. First
OSMnx buffers each geometry by 0.5 km, then downloads the OpenStreetMap ‘‘nodes’’ and
‘‘ways’’ within this buffer. Next, it constructs a street network from these data, corrects the
topology, calculates street counts per node, then truncates the network to the original,
desired polygon. OSMnx saves each of these networks to disk as GraphML and
shapefiles. Finally, it calculates metric and topological measures for each network,
592 EPB: Urban Analytics and City Science 47(4)
summarized below. Such measures extend the toolkit commonly used in urban form studies
(Ewing and Cervero, 2010; Talen, 2003).
Street network measures
Brief descriptions of these OSMnx-calculated measures are discussed here, but extended
technical definitions and algorithms can be found in e.g. (Albert and Baraba
´
si, 2002;
Barthelemy, 2011; Brandes and Erlebach, 2005; Costa et al., 2007; Cranmer et al., 2017;
Dorogovtsev and Mendes, 2002; Newman, 2003, 2010; Trudeau, 1994). The average street
segment length is a linear proxy for block size and specifies the network’s grain. Node density
divides the node count by the network’s area, while intersection density excludes dead-ends to
represent the density of street junctions. Edge density divides the total directed network
length by area, while street density does the same for an undirected representation of the
graph (to not double-count bidirectional streets). Average circuity measures the ratio of edge
lengths to the great-circle distances between the nodes these edges connect, indicating the
street pattern’s curvilinearity (cf. Boeing, 2018a).
The network’s average node degree quantifies connectedness in terms of the average number
of edges incident to its nodes. The average streets per node adapts this for physical form rather
than directed circulation. It measures the average number of physical streets that emanate from
each node (i.e. intersection or dead-end). The distribution and proportion of streets per node
characterize the type, pervasiveness, and spatial dispersal of network connectedness and dead-
ends. Connectivity represents the fewest number of nodes or edges that will disconnect the
network if they are removed and is thus an indicator of resilience. A network’s average node
connectivity (ANC)—the mean number of internally node-disjoint paths between each pair of
nodes—more usefully represents how many nodes must be removed on average to disconnect a
randomly selected pair of nodes (Beineke et al., 2002; Dankelmann and Oellermann, 2003).
Brittle points of vulnerability characterize networks with low average connectivity.
A node’s clustering coefficient represents the ratio between its neighbors’ links and the
maximum number of links that could exist between them (Jiang and Claramunt, 2004;
Opsahl and Panzarasa, 2009). The weighted clustering coefficient weights this by edge
length and the average clustering coefficient is the mean of the clustering coefficients of all
the nodes. Betweenness centrality evaluates how many of the network’s shortest paths pass
through some node (or edge) to indicate its importance (Barthelemy, 2004; Huang et al.,
2016; Zhong et al., 2017). A network’s maximum betweenness centrality (MBC) measures the
share of shortest paths that pass through the network’s most important node: higher
maximum betweenness centralities suggest networks more prone to inefficiency if this
important choke point should fail. Finally, PageRank ranks nodes based on the structure
of incoming links and the rank of the source node (Agryzkov et al., 2012; Brin and Page,
1998; Chin and Wen, 2015; Gleich, 2015; Jiang, 2009).
In total, this study cross-sectionally analyzes 27,009 networks: 497 urbanized areas’ street
networks, 19,655 cities’ and towns’ street networks, and 6857 neighborhoods’ street networks.
These sample sizes are larger than those of any previous similar study. The following section
presents the findings of these analyses at metropolitan, municipal, and neighborhood scales.
Results
Metropolitan-scale street networks
Table 1 presents summary statistics for the entire data set of 497 urbanized areas. These
urbanized areas span a wide range of sizes, from the Delano, CA Urbanized Area’s 26 km
2
Boeing 593
