Professional Documents
Culture Documents
Bono 2011
Bono 2011
Bono 2011
a r t i c l e i n f o a b s t r a c t
Keywords: The magnitude 7.0 earthquake that struck Haiti in January 2010 led to an unprecedented effort in collect-
Road network ing and providing geographical information in support of humanitarian aid. Although most of the com-
Accessibility piled datasets and generated maps were able to provide specific and detailed information regarding
Graph theory the location of damaged buildings and road interruptions, none or little information was available to
Damage
describe the accessibility—or otherwise—of the urban space. Here we try to offer an alternative method
Earthquake
Spatial analysis
to define the urban accessibility landscape in the aftermath of earthquake damage, by combining simple
graph theory concepts and GIS-based spatial analysis to assess how the urban space accessibility
decreases when the road network is damaged.
Ó 2011 Elsevier Ltd. All rights reserved.
0966-6923/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jtrangeo.2011.08.002
1444 F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455
Here we first show how the street vector data are converted work measures, however simple, combined with GIS-based road
into a graph network and how this is analyzed to identify roads network data sets is becoming more common: for example, in
isolated by the disruptions on the adjacent streets. the moving-segment approach used by Steenberghen et al.
Secondly, through the use of raster-based techniques, we (2010) an influence distance from a given location was defined in
compute the accessibility of the urban space for the two cases of terms of topological distances—rather than geographic—to uncover
damaged and undamaged road networks. hidden trends in the distribution of traffic black spots in the Brus-
Finally, we combine the accessibility measures of the two net- sels city area; in the study by Jenelius and Mattsson (accepted for
works—pre and post the hazard, in order to identify urban areas publication) the Dijkstra shortest path measure is used for dy-
that can pose barriers to the emergency operations. namic re-calculation of the vulnerability of disrupted networks
The case study focuses on the damaged Port Au Prince and Car- for generic natural hazards specific to Sweden.
refour urban road network in order to analyze the effects of the In this paper we try and go some way to continue with this phi-
disruptions of the 2010 seismic event on urban space accessibility. losophy in order to generate a visualization of a damaged road net-
work that is, in a way, a fusion of GIS and basic graph theory
concepts. Our specific approach to assessing accessibility for dam-
2. Research issues and literature overview aged road networks relies on the basic concepts of graph cluster
analysis.
Graph theory has been successfully applied to many different The term accessibility has been widely used in many fields and
systems (Albert and Barabási, 2002): from biological processes with different meanings to express different situations and pur-
(Jeong et al., 2000) to infrastructures (Crucitti et al., 2006; Jenelius poses. In general, it refers to the ease of reaching a service, place
et al., 2006; Albert et al., 2004), to communications and social rela- or topographical item. As a consequence, different measures and
tionships (Barabasi et al., 2002); thus, an object, an ensemble of approaches for the computation of accessibility have been defined
items, a process, relational matrices, etc, can all be represented (Makri and Folkesson, 1999 and Geurs and van Wee, 2004).
in terms of nodes and links called graphs or networks. Depending on the required type of analysis, accessibility in GIS
Network analysis was introduced into GIS systems to solve can be performed to study the degree and type of disaggregation,
vehicle routing problems, to find shortest paths, or to perform ori- the definition of origins and destinations, the measurement of tra-
gin destination and optimum route analysis (Curtin, 2007). Going vel impedance and the measurement of attractiveness.
beyond the most common interpretations of networks within GIS Here we adopt the travel impedance approach as it is measured
systems, graph theory analysis has also been implemented in order in units of distance (or time) and best expresses the consequences
to capture new information that is usually not manageable solely of disruptions on the network. Although more complex topological
using a GIS. Our approach here is similar in spirit to prior investi- measures are available, distance measures and network partitions
gations where a number of topological analyses were run on GIS are the simplest and most suitable for our analysis. Accessibility
platforms in order to perform analyses on electricity, gas, and road measures are commonly used when performing accessibility anal-
networks (Carvalho et al., 2009; Bono et al., 2010; Poljanšek et al., ysis to locate services in towns (e.g., stores, schools or emergency
2010), to investigate these networks’ vulnerability or resilience to units) in order to fulfill specific requirements of proximity to the
attacks, failures and natural hazards. In particular, the analysis of affected population. Given that the study of the accessibility in ur-
urban streets carried out by Bono et al. (2010) where the roads ban and regional areas is closely related and dependent on the
were converted into directed or undirected dual mode graphs (Por- transportation networks (railways, streets and public transporta-
ta et al., 2006a, 2006b), allows us to convert the practical aspects of tions systems), recent studies have addressed the complex issue
real road networks (one-way roads, road class, length, etc), into of this mutual dependency when, for example, investigating the
graphs which can then be used to investigate their statistical and impact that new investments or modifications to the original
topological properties. transport systems may have on the population or the economy
The vulnerability of transportation networks—more specifically (Gutierrez et al., 2010), when investigating accessibility and trans-
road networks—to various natural hazards has recently begun to port demand (Kamruzzaman et al., 2011) or when assessing the
be studied by combining GIS-derived data and network concepts. spatial accessibility to primary care services (McGrail and Humph-
So, for example, Sohn (2006) has studied vulnerability to floods reys, 2009).
of road networks from an accessibility point of view. Jenelius In our study we address such issues from the perspective of
et al. (2006) and Jenelius and Mattsson (accepted for publication) how accessibility is reduced (or breaks down) following a disaster.
have examined (using an approach close in spirit to our own here- Because we want to analyze how urban blocks are affected and, in
in) the vulnerability of the Swedish road network, in the first in- particular, are isolated as a consequence of road disruptions, here
stance by examining the effect on the whole network when we consider the urban streets as the starting point to access urban
closing one link at a time, and in the second instance by examining blocks (i.e., the built up areas as detected on the basis of satellite
the impact on the whole network for area-covering disruptions imagery). Therefore, we measure the accessibility in terms of
when a multiplicity of interconnected nodes are affected. shortest time/distance from the road network to all the built-up
In other GIS-based network studies, accessibility indices have urban areas. Based on this definition, the main transportation sys-
been used as urban planning tools (as is the case of Vandenbulcke tem (i.e., the road network) is to be considered as the origin where
et al., 2009) to gauge the travel times within the highly urbanized emergency service teams may depart from, and the built-up urban
Belgium transport system, which, although seemingly in contrast areas as the destination.
to the problems faced by emergency services in disaster-affected
areas, does have many commonalities as, essentially, the goal is
to traverse a network as efficiently as possible. However, a com- 3. Methods and data preparation
bined approach that performs both network and GIS-spatial analy-
sis is less common (Juliao, 1999; Schintler et al., 2007). Moreover, A graph or network is a collection of nodes or vertices V(G) and
when dealing with networks, GIS is more generally adopted only in edges E(G) that make up the graph G (V|E). Graphs can be undi-
the data pre-processing and to convert source spatial data into net- rected, if the links can be used in both directions, or directed
works; i.e., it is simply used to prepare the data in such a form so (i.e., one-way streets in urban streets networks). Directed links
that it can be parsed into a graph. However, the use of graph net- are commonly referred to as arcs.
F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455 1445
In this analysis we choose to generate an undirected road net- work, if links are missing (i.e., blocked or damaged roads), the con-
work for two reasons: firstly because the number of one-way tiguous nodes in the system may become unreachable and, as a
streets in our study area (Port Au Prince and Carrefour in Haiti) result, the graph may be decomposed into several different compo-
is a small percentage (less than 3%, according to the OSM data) nents (see Fig. 1). Connectivity analysis through strong components
of the total of the streets, and secondly because in an emergency identifies the isolated elements of the road network.
situation the circulation of emergency vehicles is not necessarily The damaged road network is then composed of all the roads
limited by the traffic rules. Therefore we have transformed all that belong to the major component only; i.e., the streets that
the streets into undirected edges (E) and the street junctions into can still be driven through after the earthquake. All the damaged
nodes (V). roads, as well as the edges that belong to minor components, are
The importance of the network analysis in the case of a crisis not part of the active network and are therefore discarded.; the
scenario (where not all the urban streets are open to traffic) is process is described as follows:
strictly related to identifying which parts of the road network are Having imported all the data onto a GIS database, a Depth First
still functional in order to reach earthquake-affected urban areas. Search (DFS) algorithm is implemented as an SQL stored procedure
For example, as a result of the disruptions, it is possible that some to loop across all the edges’ adjacent vertices and so extract the
urban areas happen to be inaccessible because some, or all, of the connected components (see the pseudocode in Table 1). Starting
connecting streets or paths to reach these areas are damaged. In from one road segment, defined by its univocally defined end-
network analysis terms, this corresponds to identifying the so- nodes (FromID and ToID), the procedure initiates the Partition
called strong components of a network, which we define below. counter and starts searching for all the streets that share a node
with that segment. The connected streets are then marked as vis-
3.1. Definition and identification of the network’s components ited and assigned with the same Partition number. The procedure
loops recursively until all the connected streets are visited. If more
A graph is fully connected if there is a path from every vertex to streets are left unvisited, the Partition counter is incremented by 1
every other vertex in the network; hence all the nodes are mutu- and the successive street is selected. The procedure ends when the
ally reachable. If the graph is not fully connected, it is fragmented entire network is visited.
into several independent connected components. In a road net-
Fig. 1. Road network with located road blocks (a), correspondent graph network (b), and three components (green, gray and red vertices) resulting from the splitting of road
segments (X marks) in correspondence of debris’ location (c). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this
article.)
1446 F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455
Fig. 2. Disruption points (a) are snapped to the nearest road lines (b), a debris buffer zone is created along the contiguous road segments (c), and then split at the intersections
(d).
F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455 1447
be dependent on the order of importance of all the raster datasets road network vector feature set (Fig. 4a) from which the accessibil-
(grids) or from a map algebra operation. ity is computed. The friction layer is defined by the cost surface of
When dealing with areas that cannot be included in the cost the urban built-up area considering a travelling speed of 6 km/h
evaluation (i.e., barriers that prevent free movement), a null value (walking speed), assuming that this land use type can be travelled
is used. at a minimum human speed only.
The evaluation of the cost distance from a cell to an adjacent The cost surface is then merged with a river raster dataset with
one is performed on the basis of the network representation of a fictitious constant cell cost; the merged raster is then reclassified
the raster grid (see Fig. 3), and where each cell’s centroid is a node (Fig. 4b) so that the cells in correspondence with waterways are
(ESRI, 2004); edges are defined by the links connecting the adja- defined as No Data (i.e., these empty cells will not be accounted
cent eight nodes (four perpendicular to the cell’s perimeter and in the computation of the cost distance).
four along the diagonals). The calculation of the least cumulative cost from each
Having defined the composite cost of an area, the cost distance cell to specified source locations over a cost raster is performed
is then computed as the distance from each single cell previously in the cases of undamaged (Fig. 5a) and damaged networks
defined as a source, to all the others on the basis of the defined (Fig. 5b).
impedance of each link. The impedance is defined by the costs Depending on the method used to build the damaged network
(from the cost surface) associated with the origin and destination (e.g., removing all the street lines or splitting roads in correspon-
cells of each path considered, and the direction of movement dence of detected damaged areas), a different source raster set is
through the cells. generated; thus affecting the computation of the cost distance.
For this reason the orthogonal paths among adjacent cells are In the examples of Fig. 5, cell color intensities are related to the
computed as the average of the two cells’ costs. For diagonal move- cost distance; dots are the cell centroids with size proportional to
mentsp the
ffiffiffi cost is computed as the average of the two cells’ costs the cell value.
times 2 so as to compensate for the longer distance. Damage to a road network affects the path travelled; in partic-
ular, damaged roads may result in the complete isolation of some
Orthogonal cost co ¼ ðcost1 þ cost2 Þ=2
areas. Considering the example of Fig. 5, it can be seen how the re-
pffiffiffi moval of the damaged street, red line in Fig. 4 (or dashed line in
Diagonal cost cd ¼ 2ðcost1 þ cost2 Þ=2 grayscale image), changes the original cost distance values of the
The cumulative cost between two cells ni and a ni+1 is then calculated undamaged case (Fig. 5a) by reducing the number of green (dark
as the cost to reach cell ni from ni+1 plus the average cost of the gray in grayscale image) cells (i.e., areas contiguous to street lines)
source and destination cells ni and ni+1. The cost distance can be then in the lower part of the raster area (Fig. 5b).
computed from any cell to a source cell as the cumulative cost of all The reduced accessibility (or the level of isolation) of the urban
the links along the least-cost path from each cell to all the sources. environment (Fig. 6) is the result of the difference of the two
cumulative costs (Fig. 5) computed for the damaged and undam-
3.3. Cost-distance analysis aged network. Cells unaffected by the disruptions are reached with
unchanged effort (white cells in Fig. 6), whereas the increased cost
Considering an urban area, we select one source layer and a fric- is due to the missing parts of the network affecting the distance
tion layer. The source layer is the raster dataset converted from a from the defined sources (undamaged roads).
Fig. 3. Cost function: how the cost distance is computed in a raster set (size and color of centroids according to the cost distance relative to the centre). (For interpretation of
the references to color in this figure legend, the reader is referred to the web version of this article.)
1448 F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455
Fig. 4. Vector feature sets and computed cost surface (with overlay of vectors): (a) road network (black lines), damaged road in red (dashed in grayscale version) and
waterways in cyan (gray in grayscale version); (b) cost surface (NoData in white); roads are the sources for the cumulative cost analysis. (For interpretation of the references
to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 5. Accessibility analysis example: (a) cost distance (undamaged network) and (b) cost distance (damaged network).
3.4. Data sources seismic event. In addition to road data, waterways are also avail-
able from OSM and were included in the analysis.
The GIS model was implemented using ArcGIS 9.2 software by OSM road vectors are presented in aggregated lines that belong
ESRI, including Spatial Analyst for raster analysis and ModelBuilder to each single named street with no discontinuities (Fig. 7a).
for structured model development. The data repository is a Micro- Although this format is efficient for map drawing applications, this
soft SQL-based geospatial database that offers the possibility to de- structuring of roads leads to the more simple straight conversion
velop complex data processing with stored procedures (see Table 2 into different formats, which, unfortunately makes compiling net-
for the list of datasets used in the analysis). works difficult or impossible because the interconnection informa-
For the analysis of the Port Au Prince and Carrefour road net- tion is missing.
works, Open Street Map (OSM) data were processed and converted For this reason, the original OSM data were processed in order
into a shapefile dataset. OSM contains the most complete and up- to define a proper structure suitable for the needs of network anal-
to-date data on the Haitian roads, thanks to the contribution of ysis (Fig. 7b) with graph nodes and edges corresponding to the
many individuals and organizations who volunteered after the physical intersections. As we were dealing with a large area, we
F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455 1449
Table 3
OSM road types and correspondent defined FRC.
(equivalent to the node id) in the <nd> tags. The tag k = ‘‘highway’’
defines the road elements, and its attribute, v, the type of road (pri-
mary, secondary, service, etc); we process this information to de-
fine the functional road classes FRC. All the other way elements
(e.g., waterways) are ignored.
The Open Street Map source .osm file was then processed by a
stored procedure in the MS SQL database and converted into a
.gml file. All the way elements are split into the constitutive seg-
ments; as the topology of the road network is defined by how
edges (ways) are connected, ways’ end-nodes coordinates are re-
trieved and stored for each single street. The generated .gml file
is converted into a shapefile suitable for the ArcGIS with the
FWtools utilities.
Having defined the end-nodes for each street element, we can
extract the network with the same approach used by Bono et al.
(2010) for the network analysis. The advantage of transforming
spatial datasets into graph networks relies on the possibility to
Fig. 6. Example of reduced accessibility result.
easily identify the connectivity of the original dataset (identifica-
tion of disconnected streets islands within the network), and to as-
sess the isolated streets due to the seismic damages.
Table 2
From the road vectors (divided into the single elements com-
GIS source datasets. prised between junctions) a point vector dataset is created from
all the vertices of the streets and two fields FromID and ToID, cor-
Dataset Type Source
responding to the origin–destination (OD) pairs, which are created
Road network Polyline Open Street and updated with the unique identifiers of the end-nodes of each
Map
Road interruptions Point UNOSAT and
single street. Functional Road Class (FRC) indices are created for
JRC each street on the basis of the OSM description (see Table 3).
Administrative boundaries Polyline GeoFabrik Damage surveys were conducted by different groups in order to
JRC PANTEX built-up mask for Port-au-Prince from Polygon JRC release prompt evaluation of the post-event situation—mainly in
SPOT imagery
support of the humanitarian and emergency organizations. The
analyses were performed on the basis of the available satellite
images and the detection of damage was performed either manu-
pre-processed the .osm downloaded dataset by removing the ally or automatically.
unnecessary elements and attributes, but only where the road UNOSAT, the UN Institute for Training and Research Operational
types compatible with vehicles transit were included (i.e. the Satellite Applications Programme, compiled a point dataset report-
OSM original data composed of the data primitives’ ways, nodes ing road disruptions; this is converted into a shapefile from the ori-
and relations). ginal kml dataset available online. The points defining the observed
OSM data are encoded in XML where node elements are the ba- disruption or the location of the debris on the roads are first spa-
sic elements for the definition of ways and contain the latitude/lon- tially joined to the closest street segment; streets are then assigned
gitude coordinates. Each way element contains the set of nodes that with the correspondent UNOSAT damage state (0 – no damage,
it is composed of; these nodes are specified by the ref attribute 1 – partially blocked, 2 – blocked). For our analysis the UNOSAT
Fig. 7. (a) OSM extract plain road network nodes in the GIS and (b) network processed intersections (red nodes, black in grayscale image).
1450 F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455
Fig. 8. Examples of new findings in GeoEye satellite images during the evaluation of the UNOSAT road disruptions dataset.
dataset of the road disruptions was considered as the basic gener- Table 4
ating set/data. The dataset contains the interruption points with Damaged network’s detected components.
two severity levels (road partially interrupted or fully blocked) Number of components Number of street segments
but in the analysis only the interruptions fully blocking the streets 1 25,315
(damage state = 2) were considered as partially blocked streets still 2 8
allow the transit, even if with some difficulty. The disruption data- 1 6
set was integrated with new findings observed during the evalua- 2 5
3 4
tion of the UNOSAT dataset; new disruption points were created in
15 3
Google Earth and an updated kml dataset was generated (see 8 2
Fig. 8). 113 1
Fig. 9. The Port Au Prince and Carrefour damaged road network. Some roads (in red, white in grayscale image) are isolated from the network even if undamaged.
F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455 1451
Fig. 10. Cost surface in minutes for the undamaged (left) and damaged (right) urban road network (central PaP detail).
1452 F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455
reported, and two power plants were left un-operational due to work. After the components were identified, the road network
failures in the transmission lines and adjacent structures. was clipped to the desired extent.
From other international agency reports (Eberhard et al., 2010; The damaged road network in Fig. 9 is derived from the origi-
ARUP, 2010; Fierro and Perry, 2010), it appears that the infrastruc- nal—undamaged—one by splitting the street lines at their intersec-
ture network in Haiti was severely affected by the seismic event. tions with the disruption-points dataset (previously snapped to the
Many large landslides along Highway 204 were reported, as well closest lines, see Section 3.1).
as in the foothills of the La Selle Mountains between downtown The disruption points were snapped onto the closest street lines
Port Au Prince and Pétionville. (see Fig. 2) with a sufficiently precise snap tolerance to allow all
Because they are usually well-engineered designed structures, the interruption points to be snapped to the lines, whilst ensuring
most of the bridges resisted the seismic forces, though some of a one-to-one relation between an interruption and the closest
them suffered damage to the shear keys at the supports due to street; these were then divided at the intersection of the snapped
insufficient or inadequate steel reinforcements detailing. damage points, creating the new dataset of the damaged network.
Several streets of the urban areas, in particular the Port Au This operation creates two new street nodes assigned with a un-
Prince downtown area and the Leogane and Carrefour districts, ique ID, thus ensuring that the network analysis will consider the
were blocked as a consequence of the collapse of buildings and two newly created nodes as independent and not connected.
by the presence of debris. The disruptions dataset is composed of a total of 379 locations,
which split the network causing the total number of edges to in-
crease from 25,153 to 25,533 in the fully connected main
4.1. Spatial analysis of the urban area of Port Au Prince component.
The nodes created by the line-splitting operation are numbered
In order to avoid border effects in the network components sequentially, starting from the higher node ID. Two new table fields
identification (where roads on the area’s limits may be discon- are created (IDfrom2 and IDto2) and updated with the nodes’ new
nected due to the cut-out operation), a 5 km buffer of the analysis IDs based on the geographical coordinates. The identification of the
selected area was initially considered, thus considerably reducing network’s connected components is now performed on the basis of
the dimensionality (especially in graph terms) of the whole net- the new connectivity relations.
Fig. 11. Damaged network with isolated roads (a) and reduced accessibility index (b).
F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455 1453
straightforward to discern the best path, whilst accounting at the the accessibility, as the two halves of the street can be accessed
same time for the degree of isolation of certain areas, due to the (even if with the disadvantage of travelling along, what is techni-
modified topological connectivity. Introducing the reduced accessi- cally, a dead end street).
bility index may help understand the town’s—further reduced—
transport capacity. 6. Conclusions
Thus, if the surface cost in both cases (damaged and undam-
aged) is compared, it is possible to quantify how road isolation af- The difficulty in deploying resources into an effective supply
fects the ease of getting to areas close to the removed components. chain was made abundantly clear to a near real-time watching
It can be seen from Fig. 11 b that the reduced accessibility of the world during the Haiti earthquake of 2010; however immediate,
urban area is not directly proportional to the damage to the built such images could only provide an overview of an extremely com-
infrastructure: it is not necessarily true that the distribution of plex supply chain running into difficulties due to the lack of basic
building damages will affect transit on the adjacent streets. More- infrastructure information. Indeed, one of the key points concerns
over, peripheral zones may result in over-ranked accessibility indi- the changing accessibility of urban areas brought about by the clo-
ces as they are served by a limited number of roads (i.e., peripheral sure of roads due to debris (de la Torre et al., in press). Here we
areas are even more easily isolated because they start off as being have presented a method that integrates spatial and network anal-
badly connected in the first place). ysis; this was implemented to assess the degree of isolation of ur-
The reduced accessibility (Figs. 11b, 12 and 13) shows how ban blocks as a consequence of major disruptions to the urban road
some areas became completely inaccessible to vehicles after the network due to collapsed buildings and debris. This method was
earthquake, with a higher concentration in the central part of the applied to the Port Au Prince urban area in Haiti by parsing the re-
town of Port Au Prince. Whereas most of the isolated areas are rel- ported disruptions of the 12 January 2010 earthquake into a before
atively small, disruptions on the road network also generated a few and after connectivity network.
large inaccessible areas. The analysis of the reduced accessibility/level-of-isolation in
It must be noted that the present reduced accessibility accounts Port Au Prince shows, what appears to be, a moderate level of iso-
only for the elimination of disconnected components. However, lation after the earthquake. Only 0.53% of the total length of the ur-
whenever a single road is blocked but still connected at both ends ban road network, or approximately 11.14 km of streets, was
to the major network component, the disruption does not affect completely inaccessible. However, as the street ‘‘islands’’ (or
F. Bono, E. Gutiérrez / Journal of Transport Geography 19 (2011) 1443–1455 1455