GIS AND TRANSPORTATION: STATUS AND CHALLENGES
Michael F. Goodchild1
The circumstances of the first use of the term geographic information system are lost in the sands of time (Marble, 1989; Tomlinson, 1989), although several excellent histories of the field now exist (Coppock and Rhind, 1991; Foresman, 1998). But it is clear that much of the initial impetus came from the group of graduate students in quantitative geography at the University of Washington in the late 1950s; and that one of these was Duane Marble, who later moved to Northwestern University and developed a rudimentary form of GIS in support of transportation studies in the Chicago area.
ransportation research requires a wide range of models and forms of analysis, but makes use of a comparatively small number of types of data, the vast majority of which have some form of geographic reference. Later, the Census Bureau’s interest in managing data collection led to the development of the DIME structure, and later TIGER, as methods of representing street networks with a high level of internal consistency. Although GIS-T is the subject of an annual conference in the U.S., an expanding literature, and numerous specialized software applications, there is as yet no book devoted solely to the topic. In a recent review, Waters (1999) concluded that “it is possible to state unequivocally that GIS-T has ‘arrived’ and now represents one of the most important application areas of GIS technology.” In this paper I attempt to address
why that should be, to identify some of the more important research issues in the field, and to point to some interesting and challenging trends. The discussion is organized according to three distinct views: the map, navigation, and behavior.
THREE PERSPECTIVES
The map view
Seen from above, the Earth’s surface is an enormously complex jumble, with very little in the way of obvious order. One of the types of features most easily identified by the human eye are the linear transportation corridors: rivers, canals, railroads, and roads. Even here, however, the simple linear model breaks down. Rivers flow into lakes and seas, and people and vehicles can escape the linear road system in parking lots and other trafficable areas. Nevertheless, the linear system is relatively stable, since rivers rarely
move, and roads are expected to have substantial lifetimes, and it provides a very convenient way of organizing and referencing much human activity. Homes and businesses are mostly located within short distances of the public road network, and so can be readily identified by specifying position on that network. Vehicles are mostly confined to the road network also, and trains are strictly confined to the railroad network. Moreover, as a linear system it is possible to specify location in a one-dimensional space, with a single number. One-dimensional references are inherently simpler than twodimensional references, since only one measurement is needed, and the distance between two such linear measurements can be much easier to compute or estimate intuitively than the distance between two points in two- or three-dimensional referencing systems, especially if the points are close to each other. Thus human societies generally have chosen to identify home location by street address, and to deliver mail using that system, in preference to two-dimensional systems such as latitude and longitude, or national grids.
Efforts to digitize the U.S. street network began in the 1960s, driven by the needs of the Bureau of the Census for an easy way of allocating individuals to reporting zones. If street addresses could be converted to coordinates, then simple routines could be used to identify any of the complex hierarchy of reporting zones: blocks, census tracts, counties, cities, or states. The same function—conversion of one-dimensional street address to two-dimensional coordinates—found abundant applications in market research and led in part to the growth of the field we now know as geodemographics, since it could be used to convert lists of customer addresses to areal counts. But the structure chosen for the Bureau’s DIME files for the 1970 census reflects what in retrospect seems like an odd choice: to see the street network as a collection of nodes and links. The traditional view of a street is reflected in our naming system, which assigns the same name to a roughly linear feature running through multiple intersections. But DIME chose to break thisf eature into a series of individual links. The advantage lies in the ability to institute checks of logical consistency, because all of the links surrounding a block must form a closed figure. But it has a number of significant disadvantages: • dependence on the precise definition of an intersection (what to do about traffic circles, or intersections not at grade, or intersections with laneways that may not count as streets); • sensitivity to creation of new intersections, which may require extensive modifications to the database; • conflict with traditional ways of thinking about street networks; and • redundancy since street names must be repeated for each link. Despite these problems, links and nodes have remained the prevailing view of networks, driven in part by the almost mythical significance given to topology in the traditions of GIS.
Another major difficulty with the link/node view concerns events or features which occur at points within links, or over stretches of links that do not match their endpoints. In the simple DIME model it is possible to attach an attribute to a link or a node, but not to arbitrarily defined parts of links. Dynamic segmentation extends the basic link/node model by making it possible to refer to points in the linear referencing system that are not at nodes (Goodchild, 1998). Route and milepost schemes model the network consistently with traditional practice, by creating a single record for every street, and a separate but appropriately linked record for every intersection. The map view is inherently limited by the need to represent real features as onedimensional spaces or centerlines. In principle positional accuracy is limited to one half of the street width transverse to the street, and is typically similar along the street; the locations of intersections are similarly subject to uncertainty. Features such as businesses and houses may have much larger offsets from the linear system, and again such information is typically lost in the map view. Important information about the side of the street (e.g., whether a house is on the north or south side) must be conveyed topologically, as a binary attribute, or inferred from the numbering system, since it cannot be obtained from the geometric information in such representations.
Although DIME and its successor TIGER were created by the public sector, there is sufficient business interest in street centerline databases and their applications to support a substantial private sector in most industrialized countries. This is especially true in those countries where printed street maps are not widely available or in common use, or where addressing systems are not as simple as in the U.S. and Europe (e.g., in Japan, where houses are commonly not numbered sequentially). Increasingly, then, more than one database is available for an area. High quality databases can now be created by renting a vehicle suitably equipped with kinematic GPS. Since such databases inherit the imperfections in the processes used to create them, any two databases can be expected to differ by amounts comparable to their positional inaccuracies.
Church et al. (1998) describe some of the problems that are emerging in GIS-T and ITS applications of street centerline databases as a consequence of such differences, and Noronha’s paper at this conference explores the issues in greater detail. Whereas interoperability is normally understood in GIS to refer to the syntactic and semantic issues of definition and content that produce different databases from the same information (Goodchild et al., 1999), the case here adds a new dimension of accuracy, which may turn out to be the most problematic dimension of all, since it casts doubt on the basic ability of street centerline databases to support reliable conversion between geographic coordinates and linear referencing systems.
Moreover, Noronha has shown that interoperability is also impeded by high rates of failure to match features, such as occurs if a street is present in one database but not in another, or if its name is missing or in conflict. Linear addressing systems also tend to break down in rural areas, where a patchwork of efforts has attempted to replicate the simplicity of the urban system with varied success.
The navigation view
Although unimportant to the Bureau of the Census, the link/node view of a street network has a significant advantage in supporting navigation, since a path through a network is readily expressed as a series of decisions at nodes. The algorithms to find optimum paths through networks, such as the shortest path, or the path of least expected travel time, also are based on link/node structures (e.g., Dijkstra, 1959), as are more complex and harderto- solve routing problems.
Routing problems require certain classes of attributes that may not be present in databases created from the perspective of the map view. It is important to identify oneway streets as attributes of links, and barriers to hrough traffic. But entirely new structures are needed to support information about turns between links, or conditions of access from one link to another (Goodchild, 1998). ESRI’s ARC/INFO, for example, extends the basic link/node structure with a turntable to record information about turns between links, and can use it to distinguish between intersections at grade andoverpasses. More generally, use of street centerline databases to support navigation requires a more complex view of topology than that of the DIME and TIGER files, which used it simply to establish consistent representations of planar features. Support for naviga tion is an indispensible part of intelligent transportation systems (ITS), where it is needed to assist drivers in designing routes, and in modifying routes in response to new information. But such effective navigation requires a massive extension of the attributes provided in the map view, to include dynamic attributes such as levels of congestion and travel speeds, temporary obstructions, and temporary turn restrictions. Few if any of these attributes are visible from above, so expensive ground-based collection is almost always required. Routes found on databases representing the map view, such as those provided by an increasing number of WWW-based services, are sometimes inaccurate or impossible because these essential attributes, and the structures needed to support them, are typically not present.
he one-dimensional perspective embedded in the map view presents numerous problems when applied to navigation. Streets may or may not have median dividers, and U-turns may or may not be legal at gaps in dividers. The advent of multi-lane freeways and turn lanes has also led to a fundamental change in the way drivers navigate. In the 19th Century gridiron city it was sufficient to tell a driver to “turn left at First Street”.
Today, a driver must anticipate left turns much earlier, in order to move into the correct lane where one is present, and must transition between intersecting freeways by taking the correct ramp. In short, it is no longer sufficient to provide navigational instructions based on a link/node view of the street network. Instead, the configuration of the street as a collection of lanes must be represented in the database, and vehicles must be tracked at the lane level of detail. Geometric representation of lanes is much more expensive than simple representation of street centerlines, and requires a substantially higher level of positional accuracy. Goodchild (1998) shows how a representation can be built that has the same level of geometric accuracy as the street centerline, but includes a topological representation of the relative positions of lanes and their connectivity. This compromise representation would be sufficient for lane-level navigation, and much less expensive than full lane geometry, and much of it could be built from aerial observation. Systems to support navigation must somehow deal with two additional complications. First, people and vehicles are not necessarily confined to the linear network, and will at times depart from it, in parking lots, unrecorded or unrecognized roads, or on private property. Systems that match vehicle tracks to networks must deal with this problem (e.g., White, 1991), and it is a significant problem for routing also. Second, navigation often requires the use of more than one mode, or comparative evaluation of modes. For example, a commuter may combine road and rail travel to work. Few efforts have been made to create databases that combine modes, by representing both road and rail networks and their interconnections, for example, but these would be essential for multimodal routing.
The behavioral view
The map view implies an essentially static perspective, and its success is in inverse proportion to the propensity of the network to change, by adding or moving links, adding intersections, or other modifications. The navigation view assumes that information of a dynamic nature must be represented on the static geometry of the network, but does not attempt to represent moving geometry. The third view, discussed in this section, deals explicitly with the behavior of discrete objects—vehicles, people, trains, or boats—on
and off the linear network. Hägerstrand (1970) was one of the first to examine the behavior of discrete objects moving in time with persistent identity. He introduced the notion of time as a thirddimension, with the rajectories of objects tracing paths in this three-dimensional space, constrained such that each object had exactly one intersection with any plane of constant time, in other words, that the locational coordinates of an object at time t could be expressed as single-valued functions of t. Mark and Egenhofer (1998) have termed such object trajectories geospatial lifelines.
In the past, such data have been very costly to collect, and consequently rare. A sample of individuals in Halifax, Nova Scotia, kept daily diaries in 1971 of their activities and locations, and these data were linked to individual records of socio-economic andm demographic characteristics. But although the data set has been subject to numerous analyses (e.g., Janelle et al., 1998), it remains one of a very small number of tensive samplings of individuals and their behavior. That situation is changing rapidly, however, as GPS transponders make it possible to track objects and sample locations intensively in time at reasonable cost. Other ITS sensing technologies, such as 'smart' loops in highway pavements, may make it possible to construct detailed travel diaries of individuals. Such data raise interesting issues of privacy that will have to be dealt with. They also create the need for effective methods of representation and visualization. We also lack a suitable suite of models and analytic techniques with which to test simple hypotheses on such data.
The Halifax data, for example, were recorded as a flat file of events, defined by a change of activity, and each of these events was geocoded. In a travel event, no information was available on the route followed, and analysis had to be confined to the beginning and ending locations of the travel event. In other situations, data might consist of samples of the locations of all objects at regular time intervals (snapshots). In tracking animals, battery weight is often a constraining factor since batteries can only be replenished through recapture, so it is important to minimize the number of locations transmitted. Hagerstrand's primary concern was with movement in continuous two-dimensional space. Similar issues arise if objects are confined to a linear network, and positions are represented in a linear system. A GIS to support the behavioral view, and thus the modeling of complex behavior, must also deal with other data types that result from aggregation of geospatial lifelines. These include flow matrices, defined as measures of the numbers of objects moving in a
given period of time between an origin area and a destination area. In such matrices all knowledge of each object's actual trajectory is lost. Numerous models exist of such aggregate origin/destination flows (Fotheringham and O'Kelly, 1989). Another data type represents the flow in each link of the network, as used for example in modeling modal splits and in traffic assignment models. In summary, the behavioral view requires a new series of representation methods, beyond those required by either the map or the navigation view. Many of these have been implemented in software, though it is unusual to find them all provided within the framework of a single, comprehensive GIS for transportation. Rather, GIS representations to date have tended to favor the map view, and to some extent the navigation view, reflecting the bias of the GIS software industry towards inventory and static representation rather than to analysis and dynamic modeling. In that sense, today's commercial GISs have made little progress at meeting the needs of comprehensive transportation planning or ITS.
References
Church, R.L., K.M. Curtin, P. Fohl, C. Funk, M.F. Goodchild, V.T. Noronha, and P. Kyriakidis (1998) Positional distortion in geographic data sets as a barrier to interoperation. Technical Papers, ACSM Annual Conference. Bethesda, MD: American Congress on Surveying and Mapping.
Coppock, J.T., and D.W. Rhind (1991) The history of GIS. In D.J. Maguire, M.F. Goodchild, and D.W. Rhind, editors, Geographical information systems: Principles and Applications. Harlow: Longman Scientific and Technical, Vol. 1, pp. 21–43.
Dijkstra, E.W. (1959) A note on two problems in connexion with graphs. Numerische Mathematik 1: 269–271.
Foresman, T.W., editor (1998) The history of geographic information systems: Perspectives from the pioneers. Englewood Cliffs, NJ: Prentice-Hall.
Fotheringham, A.S., and M.E. O'Kelly (1989) Spatial interaction models: formulations and applications. Boston: Kluwer.
Goodchild, M.F. (1998) Geographic information systems and disaggregate transportation modeling. Geographical Systems 5(1–2): 19–44.
Goodchild, M.F., M.J. Egenhofer, R. Fegeas, and C.A. Kottman, editors (1999) Interoperating geographic information systems. Norwell, MA: Kluwer.
Hägerstrand, T. (1970) What about people in regional science? Papers of the Regional Science Association 24: 1–21.
Janelle, D.J., B. Klinkenberg, and M.F. Goodchild (1998) The temporal ordering of urban space and daily activity patterns for population role groups. Geographical Systems 5(1–2): 117–138.