Michael F. Goodchild

Max J. Egenhofer

Karen K. Kemp

David M. Mark

Eric Sheppard

This paper introduces a special issue of the journal on the subject of Project Varenius, a three-year effort funded by the US National Science Foundation to advance geographic information science. Geographic information is first defined as an abstraction of primitive tuples linking geographic locations to general descriptors. Geographic concepts originate in the human mind, and are instantiated in geographic information. Geographic information technologies apply digital methods to geographic information. Finally, geographic information science is defined as the set of basic research issues arising from these technologies. Three motivations are presented for research in this area: scientific, technological, and societal. Within the project, geographic information science is structured by a three-part framework that includes cognitive, computational, and societal issues. The paper ends with an introduction to these three parts, which define the infrastructure of the project and are discussed at length by the subsequent three papers.

The US National Science Foundation, the nation's largest sponsor of basic research, is currently funding a substantial effort, known as Project Varenius, to advance the science of geographic information. Support for the project began in February 1997 and will extend into February 2000; it is being provided through a cooperative agreement with the site of the National Center for Geographic Information and Analysis at the University of California, Santa Barbara. This introductory paper to this special issue describes the background to the project, its intellectual and scientific objectives, its motivation, and its structure. Three major papers follow on each of the strategic areas of Varenius research.

Although this journal changed part of its name in 1996 from Systems to Science, and although the term geographic information science (GIScience) seems to be catching on in various ways, as evidenced by the growth of the US-based University Consortium for Geographic Information Science (http://www.ucgis.org), nevertheless there are lingering uncertainties about exactly what geographic information science means, about its relationship to geographic information technologies, and about how to move it forward. The Varenius project proposal developed a narrow and possibly provocative view of the field, and the first part of this introduction describes this view, and the reasons that led to its adoption. It is followed by a discussion of the possible components of the field, and the three strategic areas of the project are introduced. The final section briefly describes the history and structure of the project, and the reasons for naming it in honor of a 17^th Century Dutch geographer; and leads into the three major papers that form the largest part of this special issue.

Geographic information (GI) can be defined as information about the features and phenomena located in the vicinity of the surface of the Earth. What distinguishes this particular type of information from other types is of course the presence of a reference to some geographic location, and all GI can be reduced eventually to a simple statement that at some location there exists an instance of some more generally recognized thing, where thing might be a class, a feature, a concept, a measurement of some variable, an activity, an organism, or any of a myriad possibilities. By geographic location we mean to include up to three spatial variables, depending on whether distance above or below the surface is important, and also time, if the thing is time-dependent. More formally, the fundamental primitive element of geographic information is the tuple <x,y,z,t,U> where U represents some thing present at some location (x,y,z,t) in space–time.

Many authors have reviewed the nature of the geographic space–time in which geographic information is embedded (Gatrell 1983, Couclelis 1998). The dimensions of the Earth and the accuracy of surveying technology limit the range of reasonable resolutions of geographic space, from perhaps 10⁷ m at the high end, or the linear dimension of the planet as a whole, to 1 m, which is the finest resolution associated with commonly-available Earth imagery and mapping, or perhaps as small as 1 mm, the limit of most practical surveying. Within these limits there are no complications from relativistic or quantum effects, and no reason not to treat geographic space–time as a simple, rigid Newtonian frame. Within that frame the geographic surface is a two-dimensional curved object, and it is generally convenient to redefine the spatial coordinates as latitude, longitude, and elevation (f ,l ,h), based on some agreed mathematical approximation to the true form of the surface; unfortunately, this leaves our primary method of defining location dependent on which of many vertical and horizontal datums is chosen for the purpose. Moreover, it was also convenient in the era when most GI had to be recorded on flat sheets of paper to transform latitude and longitude to a two-dimensional frame (u,v) using one of a large number of transformation functions or map projections P(f ,l ) (Bugayevskiy and Snyder 1995).

We conceive of the geographic frame as continuous, so a complete characterization of some variable such as temperature at all points in space, or in space–time, leads to an infinite number of tuples. GI is manageable therefore only under certain conditions which render the number of recorded tuples finite (in what follows the symbol x refers to some location; the number of space–time dimensions defining x will be determined by the context):

• Correlation. This allows aspects of what is present at x to be guessed from other aspects of what is present at x. We humans have become adept at identifying only a minimal number of things at a location, and inferring other things from them.

• Spatio–temporal autocorrelation. This allows aspects of what is present at x to be guessed, predicted, or inferred from what is present at some other location x+e. In general the ability to predict declines in both space and time with the magnitude of e, since spatial autocorrelation is usually stronger over short distances (Tobler's first law of geography; Tobler 1970), and many geographic phenomena also vary slowly if at all through time.
• Things of finite spatial extent. If a thing is recorded as present over a region (a connected set of locations) rather than at a single location, then a single tuple of GI can connect the thing with the region. In this way a large number of tuples of the form <x,y,z,t,U> can be replaced by a single tuple <R,U> where (x,y,z,t)Î R. This condition can be generalized to include things that form a recognizable pattern over the region, where the pattern is not necessarily the simplest one of uniformity. For example, the function f(x) = a + bx + cy; {xÎ R} implies a pattern of linear spatial variation over region R.

• Named regions. A set of locations R may be sufficiently well-known and stable to be worth naming. Provided everyone can agree on the tuple <N,R>, which identifies a place-name N with a region R, then tuples of the form <N,U>, describing a thing present in the region named U, are an economical way of recording GI. Unfortunately there is often lack of agreement over the exact geographic extent named N. Gazetteers, which are a major source of relations of tuples <N,R>, all too often approximate R by a single central point (f ,l ).

• Generalization. The number of tuples can be made manageable if the thing recorded at a location is allowed to deviate from the thing that is actually observed. If most things within a region are equal, it may be acceptable to record a single thing for the entire region. If the thing has numerical value, it may be acceptable to record only a mean for a region, rather than values at every location within the region. If the spatial variation of a numerical thing over the region is closely matched by a mathematical function, it may be acceptable to record only the function. A vast range of possible generalization rules are in common use in the domain of GI, both to record information and also to derive simpler approximations from information that is already recorded. The differences between what is recorded and what is observed (or believed to be true) are the basis for the study of uncertainty in GI (see, for example, UCGIS 1996: 125).

• Abstraction. If we can agree on a standard vocabulary of things, then the larger the vocabulary, the greater the prospects for information reduction. For example, the abstracted tuple <R,lake> can be expanded first into a large number of tuples <x,lake> for all xÎ R, and expanded further based on our understanding of the nature of the thing lake (e.g., uniformly fresh water, uniform elevation). The expansion of <R,hill> is more problematic because hill is a much less specific indication of the things present at each of the locations in R.

The set of possible abstractions is vast, although limited in some domains by attempts to establish authorities or thesauri of standard geographic terms (see, for example, the work of the U.S. Federal Geographic Data Committee, http://www.fgdc.gov, and Rugg et al. 1997), and discourage the use of others. It includes abstractions that connect things in one location at one time to things in other locations at other times, through conceptualizations of dynamic processes that affect geographic landscapes. It includes geographic prepositions, which define instances of relationships between more primitive things, as in the road crosses the park, or the house is next to the road (e.g., Mark and Egenhofer 1994).

If information is to be shared between people, it is essential that its meaning be understood by all of the parties to the sharing. In scientific practice, information is expected to satisfy the requirement of replicability, that any two observers of the same phenomenon would record the same information, to within the accuracy of their measuring instruments. It follows that the same two observers would interpret the information in the same way. In the scientific world the bases of measurement are subject to international agreement and standardization, and are free of ambiguity.

In the case of geographic information, the system of spatio–temporal referencing defined by (f ,l ,h,t) is defined according to international standards that have been agreed since the late 19^th Century (when the Greenwich Meridian was accepted as the global standard of longitude), although problems exist because of the multiplicity of datum standards in use. However some ambiguity still persists because of imperfect measuring instruments, which prevent two observers agreeing exactly on the spatio–temporal location of a given point. Two observers using simple GPS receivers to record the location of a point might disagree by as much as 100m. But much larger semantic problems exist because of lack of agreement on the meaning of things, the other part of the primitive geographic tuple (for recent discussions of semantic issues in GI see Kuhn 1997, Harvey 1998). Some things, notably physical variables, can be defined sufficiently well to meet the scientific standard. Others are fuzzy (e.g., Burrough and Frank 1996), meaning that there is sufficient ambiguity in their definition to lead to greater variation among observers than is attributable to inaccuracies in measuring instruments. Others are subject to regional variations in dialect, or variations in terminology between disciplines or social classes. Some, such as hill, are so far away from the norms of scientific description as to be virtually useless except in the loosest description.

Despite this range, however, we state as a basic principle that GIScience must embrace the study of GI that fails to meet the scientific criterion, although GIScience itself must be based on rigorous scientific principles. Although much of the information that is collected and processed by GI technologies is scientific, it is also true that GI technologies are widely used in other contexts, and that much of our knowledge of the world fails to meet rigid scientific standards of observation. It is also true that ambiguity is not always bad, and that precision is not always better, especially when it is inconsistent with accuracy. GI technologies may be asked to answer questions that are ambiguous but nevertheless important to the originator, such as 'Is Santa Barbara north of Los Angeles?' They may be used to process data, from sources such as soil maps, that include information that is inherently fuzzy, but nevertheless is believed to be useful in certain contexts.

Perhaps the most powerful argument for this basic principle rests on expectations about the nature and interests of general GIS users. People habitually work with GI in ways that fall short of scientific standards of rigor, just as they use other precise technologies like the telephone to communicate imprecise information. They give each other directions, for example, that are full of poorly-defined and ambiguous terms, such as near, not far, about. A GIS that hopes to be useful in this environment can be designed according to one of two principles. First, it can be designed to work only along rigorous scientific lines, and require that users adjust their normal behaviors to match, insisting, for example, that they not use terms like near in interacting with the system, unless they have first given near rigorous definition. Alternatively, a GIS can be designed to support normal discourse (for an example of GIS interaction using the term near see Robinson et al. 1986). Of course one might argue that the second option is impossible as long as the computer is a machine with perfectly mechanistic behaviors; or that it is impossible in principle to design a precise machine that can interface smoothly with a less-than-perfect human intelligence. Nevertheless, recent experience in computing technologies does seem to support the notion that computing applications can be designed so as not to intrude significantly on the normal cognitive processes of their users. At worst, the possibility is an important research question.

In the previous section we argued that GI is constructed from primitive tuples <x,U> through processes of abstraction, use of common vocabularies, and generalization. We use the term geographic concepts to describe all of the generic components of GI, including:

• the concepts that provide the basis for GI itself (e.g., the geodetic system, projections, metrics of distance);
• elements of geographic thesauri, the standard authorities that define a common vocabulary of feature types (e.g., lake, reservoir, river, city, building);
• the contents of gazetteers that establish the positions of named places;
• the terms and phrases that define relationships between geographic point sets (e.g., near to, north of, crosses, intersects with); and
• the generic classes of things that define the phenomena present at geographic locations (e.g., variables such as elevation, temperature; land cover classes; land ownership; zoning regulations).

Some of these concepts meet scientific standards of replicability, but others do not. Some are simple, such as metrics of distance, or systems of Earth coordinates, while others are far more complex, including tropical storm, neighborhood and esker. Some have meaning that is shared by virtually the entire human population, such as longitude, while others have meanings that are confined to disciplines, language groups, regions or other information communities.

All of the previous discussion is independent in principle of whether digital computers are used in any stage of the handling of GI. In general, it is possible to represent geographic variation through models of two general classes: analog and digital. By an analog model, we mean that the model is represented in some space as a scaled replica of reality; the space is usually physical (e.g., a paper map), but may also be electrical (e.g., transmission of seismic records as analog electrical signals). By a digital model, we mean that the recorded properties are coded into some discrete alphabet, using an agreed set of rules. To all intents and purposes that alphabet is the binary alphabet, and digital implies the use of modern digital computer technology. Devising rules for recording important properties of the real world in the binary alphabet of the digital computer is one of the most challenging tasks of GIScience.

The use of digital technology conveys enormous advantages over analog modeling, and these have driven the rapid development and adoption of digital GI technologies over the past four decades. The relevant technologies include GPS, remote sensing, image processing, soft photogrammetry, the surveyor's total station, scanners, virtual environments, and plotters, all of which are to some degree specific to GI. Digital technology allows for easy editing, since there is no need to interact with a physical model; calculation and manipulation of data through the use of arithmetic and logical operations; reliable storage and handling, since it is much easier to protect digital systems from additional error and unwanted noise; and sharing, since digital information can be transmitted at the speed of light and at very low cost.

We use the term geographic information system (GIS) to describe the most generic and powerful of these technologies, embracing all forms of digital analysis, manipulation, querying, communication, retrieval, and output. Early versions of GIS were software applications on large mainframes; later, mainframes were replaced by desktop workstations. Today, however, GIS must be understood as a complex system of distributed data and processing resources, designed to support manipulation, analysis, modeling, and decision-making based on digital GI.

Digital information technologies have become so pervasive in today's world that it is difficult to find examples of information that is not digital at some stage in its life. Telephone voice communication has shifted almost transparently from analog to predominantly digital, and is moving rapidly to packet-switching technologies that handle conversations as small, independent packets of bits. The contents of paper sheets are now commonly transmitted using the digital protocols of FAX. Paper maps can be scanned, transmitted by FAX, and digitized using one of a large number of accepted rules. The mere existence of digital coding at some point in the life of the data is not very significant, however; much more significant are the methods used to code the data, and the constraints that this coding imposes. Conversion to digital form imposes constraints if its inverse fails to restore the full content of the information (by full content we mean all of the elements of the data that are known to be useful). For example, scanning the contents of a map with a resolution S (followed by its inverse, in this case plotting onto paper) deletes all of the variation at resolutions less than S; digitizing a smooth curve as a polyline (a sequence of points connected by straight lines) deletes any information not captured by the polyline.

There is nothing inherent in the information loss that results from coding in digital form, since with sufficient care it is possible to express virtually any information in digital form without loss. Communication is mediated by the senses, and we have abundant ways of representing visual and acoustic signals in digital form; only in the case of tactile and olfactory communication is there any suggestion that digital encoding may fail to capture signals. However, the constraints imposed by the coding schemes in common use for digital data almost always result in information loss, because they are normally chosen to achieve a balance between loss on the one hand, and volume of data and ease of manipulation on the other.

We are now in a position to define GIScience. Information science generally can be defined as the systematic study according to scientific principles of the nature and properties of information. From this position it is easy to define GIScience as the subset of information science that is about GI.

This definition is straightforward, but it fails to address a key question: why does GI form a subset that requires specialized study, or what is special about GI? Several authors have addressed this question, normally under the rubric 'what is special about spatial?' (unfortunately there seems to be no acceptable equivalent for 'geographical'). Anselin (1989) takes a statistical approach, arguing that GI differs from the mainstream of statistical information in the almost ubiquitous presence of spatial dependence and spatial heterogeneity. Each of these properties deviates from the normal assumptions of statistical methods; specifically, spatial dependence causes problems for the independence assumption that underlies many tests, and spatial heterogeneity causes problems for assumptions of homoscedasticity and the stationarity of other statistical properties.

Rhind (1996) takes a broader perspective that includes the institutional and societal context of GI, arguing that economic, legal, and public policy issues define the special nature of GIS. We believe these arguments and those of Anselin are convincing, and fully justify GIScience (and see also Goodchild, 1992) as a separate and significant subset of information science.

An independent series of arguments leads to a justification for GIScience based on the needs of the GI technologies for basic research (Goodchild 1992; Wright et al., 1997). GI technologies are so important and influential, it is argued, and raise such interesting basic questions, that a substantial investment is needed in the research issues that underlie the technology and determine its long-term development. Although large investments have been made in these technologies over the past three decades, there remain many impediments to greater efficiency, more effective analysis and modeling, and greater use; and these impediments may yield to sustained research in the appropriate disciplines.

Thus we have two distinct but convergent bases for defining GIScience, one deriving from information science generally, and the other from the GI technologies. Other terms may be equally acceptable, including geomatics, which is in common use in some parts of the research community; spatial information science, geoinformatics, and perhaps others.

GIScience re-examines some of the most fundamental themes in traditional spatially-oriented fields such as geography, cartography, and geodesy, in the context provided by the emerging digital age and the society in which it is embedded and which it influences. It incorporates recent developments in cognitive and information science, together with more specialized research in established disciplines such as computer science, statistics, mathematics, and psychology. It also incorporates developments in our understanding of the nature of society, and the forces that structure and influence it. It is motivated in part by questions of why certain geographic problems cannot easily be addressed with current GIS software. Similarly, the difficulty that some people have in using state-of-the-art GISs, or in moving from one GIS to another, raises basic questions about human spatial cognition, about how to capture and represent geographic knowledge in an information system, and about the continuing interference by immature technology in the performance of substantive tasks.

On a more technical level, GIScience needs to resolve a number of questions such as:

• How do we conceptualize geographic worlds?
• How are geographic worlds constructed by society, and how does society affect our understanding of geographic space?
• How do we capture (measure) geographic concepts, recognize them in the field or in remotely-sensed information, and identify their accuracy and quality?
• How do we represent geographic concepts, especially in digital environments, with incomplete information, with alternative data models in the same systems, and perhaps with different representations of the same phenomena (multiple representation)?
• How do we store, access, and transform geographic concepts with as little loss of information as possible during data sharing, file transfer, and data archiving?
• How do we explain geographic phenomena through the application of appropriate methods of analysis, and models of physical and human processes.
• How can geographic information technologies help to reveal and constrain understandings and interpretations, particularly in the human sciences?
• How do we visualize geographic concepts, as two-dimensional static maps on printed media or electronic displays, or as animated displays or hypertext documents?
• How do we use geographic concepts to think about geographic phenomena, to make decisions about places, and to seek explanations for geographic patterns and phenomena?

Traditionally, such questions have been addressed by researchers working within existing disciplines, and much progress has been achieved. However, the work has been spread across many research fields, and often has been conducted within very different research traditions. No systematic conceptual framework has emerged from these relatively isolated efforts. Commonalties among the questions and their solutions may be missed in fragmented research environments. Furthermore, cross-disciplinary work is often risky in academia, especially to early-career researchers, because promotion standards often given highest priority to scientific outlets that are defined by the boundaries of traditional disciplines, and tend to assign greatest prestige to their centers. We believe that by addressing these questions within the framework provided by the emerging field of GIScience, we can help to reduce institutional impediments to progress in these research areas, and to encourage the exploration of issues in ways that go beyond the solution of immediate problems.

In addition to pure intellectual curiosity, we see research in GIScience as motivated from three distinct directions: scientific, technological, and societal.

Research in GIScience serves the needs of science and scientists in two ways. First, it addresses areas where our understanding of key geographic notions and their appropriate representations is currently incomplete. We see such basic research as especially important at this time in areas where human conceptualizations and digital implementations of concepts interact and conflict. Second, GIScience contributes to the conceptualizations, methods, and tools with which scientists approach geographically-distributed phenomena. Thus it contributes to the infrastructure of science, particularly for those disciplines whose subject matter is distributed over the Earth's surface, and for which a geographic perspective is likely to prove useful.

Historically, the development of the GI technologies has been influenced only in a limited way by the needs of science. GISs have their roots in government agency data-gathering and decision-making, the design disciplines like landscape architecture and planning, and the mapping sciences of cartography and surveying. GPS was a military development. Only in remote sensing has there been a longstanding link to scientific applications. But this situation has changed rapidly in the past decade, particularly in disciplines like anthropology, hydrology, and terrestrial ecology, where the broadly-based functionality of GIS is able to provide a comprehensive software environment for a terrestrially-based science. Not surprisingly, much interest in GIS has originated in geography, with its comprehensive interest in phenomena on the surface of the Earth. In other fields, however, such as oceanography (Wright and Goodchild, 1997) and atmospheric science, the role of GIS is currently limited to preprocessing of boundary conditions and visualization, since the two-dimensional, static map metaphor used in current GISs has proven too restrictive for sciences concerned with the transient behavior of fluids in three dimensions. On the other hand certain generic issues of GIScience, such as error modeling and propagation in spatial data, are eminently relevant to these sciences also.

As in many other instances, the development and adoption of GIS tools in the scientific community raises questions about the influence of tools on the conduct of science, and whether such tools can alter the ways of doing science as the microscope and the telescope did in the past (Abler, 1987). Should the scientist insist on knowing exactly what operations are performed by the tools, or is this principle bound to be weakened as science becomes more complex, more collaborative, and more interdisciplinary? Do the choices that the use of a geographic database imposes on its users constrain the science that can be performed, in ways that are often out of the immediate control of the scientist? Under what circumstances is the scientist willing to trust data that he or she did not collect, and will the increased technological ability to share scientific data over the Internet and using the World-Wide Web (WWW) change them (Onsrud and Rushton, 1995)? Such questions about tools often have their roots in theoretical questions about appropriate representations, operations, and concepts.

Our second motivation derives from the technology push that we are currently experiencing. The proliferation of faster computing hardware and the emergence of an information infrastructure are quickly changing the ways people think and work, creating digital worlds.

New information technologies have significant influence on the advancement of GIScience through the design and use of GISs. They enable geoscientists to collaborate in new ways, sharing large spatial data collections or performing tasks together without the need to be present at the same location at the same time. Digital worlds embed new paradigms. They move bits rather than atoms, offering access to data and information without any need to relocate physical media. They offer everyone the chance to make information publicly available. The telephone has been the medium for the exchange of voice, but the new information highways allow users to collaborate by exchanging digitally-coded data that stand for text, voice, images, and more complex structures such as GI.

Digital worlds form a new culture of computing, in which the user is paramount.

Ease of use is critical, and only the user's tasks should matter; systems should reflect the needs of users, without requiring them to be concerned with technicalities. The current limitations of GIS are such that it is clear that the technology will have to be reinvented repeatedly, and we doubt whether the GIS of ten years from now will be recognizable to its current practitioners. Of particular importance are visual representations, which replace numeric representations, and visual thinking. While visual representation boosts communication ('A picture is worth a thousand words'), it hides internal representations. This perception causes a dilemma for scientists who in general desire to understand fully what the system they are using does with their data. One might argue that the cultures of science and digital worlds diverge on this issue, counter to the commonly-held belief that computing provides an ideally supportive environment for science.

While many commentators present a uniformly rosy picture of a technology-based future, the academic sector has a useful role to play in dispassionate assessment. Geographic information science should also be about the flaws in bullish predictions, including questions of equity, narrowness of representation, and many other issues. These are addressed at length in this volume by Sheppard et al.

Bullish predictions of a glowing future for digital technology and GIS aside, we also believe it is important that such a fast-growing and groundbreaking technology be subjected to the kinds of dispassionate reflection at which the academic sector excels. No other group is likely to take the kinds of long, hard looks at GIS that are needed if its benefits to society are to be maximized, and its potential abuses avoided or controlled.

Geographic perspectives are fundamental to an understanding of the interplay between local and global environments; the couplings between physical processes in the terrestrial subsurface, atmosphere and oceans, and their interactions with the human world; and the integration of processes and policies over geographically varying boundary conditions. As such, they may offer approaches to the solution of many of society's most pressing problems. For example, the US National Research Council report Science Priorities for the Human Dimensions of Global Change concluded that:

While we can claim only the most indirect of linkages between GIScience and the solution of problems of global hunger, unemployment, or crime, each of these issues provides a context in which GI technologies play an important part, and where it is important that these tools be as carefully thought out and as effective as possible.

At a more immediate level, GIScience research addresses such issues as the implications of GI technologies for systems of democratic representation; the potential for popular empowerment through concepts of electronic democracy; the legal liability associated with GI and GI technologies; the potential for invasion of privacy, surveillance, and control; and the implications of GI technologies for the organization of human activities in geographic space. Many similar issues arise in connection with all aspects of digital technology; an important question for GIScience in each case is whether the geographic context makes the problem in any sense unique.

Based on these definitions and motivations, we can now begin to examine GIScience in detail, and to discuss various ways of partitioning the field. Our guiding principle in doing so is the need for research progress: how best can we move the field forward?

Goodchild (1992) reviewed the various disciplines that might have a role in contributing to progress in GIScience. These included the disciplines that have traditionally focused on GI, and on one or more of the associated technologies. Cartography largely predates the digital era, as do surveying, geodesy, and photogrammetry, though all four have been enormously stimulated by it. The GIScience disciplines also should include those that have substantial contributions to make, but for which GI has not been seen as an important focus in the past: in this category we include statistics (spatial statistics), economics (information economics); cognitive science (spatial cognition), psychology (environmental and developmental psychology, and social psychology), and mathematics (geometry).Geography and computer science clearly have contributions to make to GIScience, though in both cases the significance of the field within the broader objectives of the discipline is open to debate (in the case of geography, see Wright et al. 1997). We believe, however, that any partitioning of the field of GIScience based on traditional disciplines will work against the needs of the field, which are surely best addressed through multidisciplinary collaboration.

The US-based University Consortium for Geographic Information Science (UCGIS) has taken a consensus-building approach, asking each of its institutional members and their delegates to identify key research areas, and distilling a structure from the response. At its 1996 Assembly the UCGIS arrived at ten topics (Table 1), and these were further refined at the 1998 Assembly. Papers on each of the topics are available at http://www.ucgis.org, and a summary paper has been published (UCGIS 1996). The ten topics are roughly equal, however, only in the level of support they received from the membership; no other coherent basis has been suggested for dividing the field in this way.

Table 1: The ten research topics of the US-based University Consortium for Geographic Information Science (source: UCGIS 1996)

The Varenius project uses a method of structuring the field that differs from both of these previous options in having a coherent intellectual basis. We propose that the domain of GIScience addresses three distinct arenas:

• the individual, as user of technology, observer of geographic phenomena, source of conceptualizations, and maker of decisions;
• the system, defined as the entire complex of digital GI technologies and its supporting hardware, software, and networking foundations; and
• society, including its institutions, customs, communities, norms, and standards.

GIScience must address specific questions in all three arenas, including:

Individual: how do people conceptualize the world around them, and reason about it using those conceptualizations?

System: how can we design GIS to achieve maximum performance and functionality, with minimum information loss or other constraint?

Society: what processes determine the adoption of GIS in society, and its use by institutions, and how does the adoption of GIS change the way society constructs space?

But more significant perhaps are the questions that arise at the boundaries between the three arenas, such as:

Individual–System: to what degree does the use of digital coding constrain the ability of individuals to record and communicate knowledge of the geographic world?

System–Society: what will be the impact on societal issues, such as privacy, as a result of rapid growth in the use of GIS, and how is GIS a construction of society?

Society–Individual: how will GI technologies change the individual's access to information, and the ability of governments to monitor society's members?

This tripartite division also allows us to identify the areas of GIScience most in need of attention by specific disciplines, to encourage participation by members of those disciplines in research, and to develop research ideas within a coherent framework. Below we identify the disciplines we believe are most likely to contribute to each of the three research areas:

Individual: cognitive science, environmental psychology, linguistics;

System: computer science, information science;

Society: economics, sociology, social psychology, geography, political science.

More specifically, the following titles seem to us to summarize the important issues of geographic information science:

Cognitive Models of Geographic Space (issues of the individual);

Computational Methods for Representing Geographic Concepts (issues of the system);

Geographies of the Information Society (issues of society).

We have struggled with an appropriate title for the societal area, and the title given is far from satisfactory, but we are unable to find a better one. Geographies of the Information Society suggests an undue emphasis on future patterns of human activity, under the influence of such technologies as telecommuting, and thus misses all of the other issues that arise because of the complex interactions between GI technologies and society. However, we continue to use it as the title of this area (note that Sheppard et al. also refer to this area as the more anonymous 'Apex').

The project is designed around this tripartite structure. Each of the three strategic areas is overseen by a panel of international experts, with the responsibility to monitor progress and manage project activities in the area. The three panels are chaired by David Mark (Cognitive Models of Geographic Space), Max Egenhofer (Computational Methods for Representing Geographic Concepts), and Eric Sheppard (Geographies of the Information Society). During the lifetime of the project they will each sponsor workshops on significant and promising research topics in the area, and undertake related activities to foster progress. The three papers that form the largest part of this special issue describe the three research areas and their activities in detail.

The project is named for Bernardus Varenius, who is perhaps best known as the author of the Geographia Generalis (1650), a work that ranks among the most influential in the history of the discipline of geography. The influence of Varenius on his contemporaries, particularly Sir Isaac Newton, has been documented by Warntz (1981, 1989) and will not be reviewed here. Rather, we cite two simple reasons for honoring him in this way:

• The division of the discipline of geography into general and special branches, rather than the more conventional human and physical. This division avoids the problems that arise when separate human and physical understanding must be integrated. More importantly, it identifies general geography as including the principles of Earth measurement, and the dynamic processes that create the landscape, while special geography documents the unique characteristics of places, and the boundary conditions of processes. This cleavage dominated methodological discussion in the discipline of geography for much of this century (Johnston 1991); but it finds a compelling resolution in GISs, which combine the special (the data) with the general (the algorithms, principles, concepts, and models) into a functioning whole. We make this argument advisedly, however, being aware of the limited ability of GIS to deal with dynamics, and thus with representation of process.
• The conviction, clearly apparent to Newton, that the geography of Varenius was founded on principles of scientific observation and reasoning.

The following three papers describe the three areas of the project in detail. Each presents a definition of its domain, a review of our current level of understanding, and a prospectus for the future. The three papers vary somewhat in approach, as appropriate to the nature of each topic.

The primary mechanism of the Varenius project is the specialist meeting, which brings together an international group of experts in a workshop setting to review progress in a given area, identify researchable topics that will move the area forward, and define a research agenda. Within the project, mechanisms exist for pursuing topics after the specialist meeting through seed grants for proposal preparation, and funds to support visits to other institutions to develop collaborations. Each workshop also results in a report, which is published in the NCGIA Varenius series and made available electronically. Further details on the Varenius project can be found at the NCGIA web site, http://www.ncgia.org. They include information on past and future meetings, meeting reports, and additional information on the project structure and administration.

The project includes resources to support three specialist meetings for each of the three strategic research areas. The topics for these were chosen by the respective panels, as representing areas of high significance to geographic information science where rapid progress might be anticipated. Further details of each, and the basis on which they were chosen, are given in the three papers. The papers also include results of the meetings and subsequent research where these were available at the time of writing.

The authors gratefully acknowledge the support of the U.S. National Science Foundation, which provides core support for the Varenius project under Cooperative Agreement SBR 9600465 to the University of California, Santa Barbara.

David M. Mark

Christian Freksa

Stephen C. Hirtle

Robert Lloyd

Barbara Tversky

This paper reviews research in geographic cognition that provides part of the theoretical foundation of geographic information science. Free-standing research streams in cognitive science, behavioral geography, and cartography converged in the last decade or so with work on theoretical foundations for geographic information systems to produce a coherent research community that advances geographic information science, geographic information systems, and the contributing fields and disciplines. Then, we review three high-priority research areas that are the topics for research initiatives within the NCGIA's Project Varenius. Other topics consider but ranked less important at this time are also reviewed.

As geographic information becomes ubiquitous in a variety of domains and field applications, computational models of geographic cognition become increasingly important to the growth of a science of geographic information. This paper reviews the history of research in cognition of geographic space, summarizes the current state of the field, and suggests several important open issues regarding the cognitive component of geographic information science.

It is important to recognize the distinction between geographic space and space at other scales or sizes. Palm-top and table-top spaces are small enough to be seen from a single point, and typically are populated with manipulable objects, many of which are made by humans. In contrast, geographic or large-scale spaces are generally too large to be perceived all at once, but can best be thought of as being transperceptual (Downs and Stea 1977: 197), experienced only by integration of perceptual experiences over space and time through memory and reasoning, or through the use of small-scale models such as maps. Some of our discussions of geographic cognition might now apply to spatial cognition at other scales.

Our review of cognitive models of geographic space is made in the context of providing a sound theoretical foundation for geographic information systems (GIS). The First International Advanced Study Symposium on Topological Data Structures for Geographic Information Systems, organized and hosted by Harvard University in October 1977, included many papers that raised theoretical issues about the nature of geographic information and of computational systems to deal with it. Among the more important papers in this vein are those by Chrisman (1979), Kuipers (1979), and Sinton (1979). Kuipers' paper in particular established a cognitive component to these theoretical foundations. Also in 1979, formal topological foundations for GIS were published Corbett (1979). Boyle et al. (1983) stated that progress in GIS was impeded by a lack of theory. In that same year, Jerome E. Dobson published his now famous 'Automated Geography' paper in The Professional Geographer, raising the idea that widespread adoption of computational models was changing the discipline (Dobson 1983). Around 1987 there was a flurry of publication activity (Abler 1987, Frank 1987, Peuquet 1988, NCGIA 1989).

The idea of a science of geographic information, a scientific or intellectual field behind and around geographic information systems, emerged rather suddenly in the late 1980s, perhaps as part of the maturation of theory and intellectual content in GIS. Michael Goodchild's July 1990 keynote address at the Spatial Data Handling international symposium in Zurich was entitled 'Spatial Information Science'. The word spatial was changed to geographical as Goodchild's keynote became an article (Goodchild 1992), and in a few short years, a new field of study, a new science, had emerged. For a further review of geographic information science (GIScience), see the introduction by Goodchild et al. in this issue.

The idea that a science of geographic information should have a cognitive foundation emerged with the development of GIScience itself in the late 1980s, when it was included in the successful proposal for the National Center for Geographic Information and Analysis (NCGIA 1989). The National Science Foundation's solicitation for proposals for an NCGIA included five main topics (bullets) characterizing key areas for research by the Center. One of these topics was to research and develop a 'general theory of spatial relations and database structures'. The successful bidders for the NCGIA argued that such a theory must necessarily include a component that linked cognition and computation, and outlined a research initiative called 'Languages of Spatial Relations' (Mark 1988, 1989, Mark et al. 1989, Mark and Frank 1990, 1991), the second research initiative undertaken by the NCGIA.

A direct attention to cognitive issues in the GIS discourse appears to have originated in the mid-1980s, when GIS researchers saw the potential of cognitive science to provide insights on how to develop a richer theoretical basis that could be provided by Euclidean geometry and graph theory. The first papers clearly identifiable with this theme were V. B. Robinson's work of fuzzy logic models of the meanings of spatial relations (Robinson et al. 1985, 1986a, b). Smith et al. (1987) described a knowledge-based GIS that included many cognitive concepts. Shortly thereafter, papers on cognitive aspects of geographic information science were presented at an international meeting in Crystal City, Virginia (Frank 1987, Mark et al. 1987).

Cognitive geography and geographic cognition is a very broad topic, and some aspects of geographic cognition are of only marginal relevance to geographic information science. Some of those more peripheral topics are excluded from this review, although they are important topics in their own right, and may eventually prove critical to a complete cognitive foundation for geographic information science. One set of such topics are those that deal with neurophysiology and neuropsychology. Although the physiological architectures of human cognitive systems may eventually provide explanations for observable aspects of human spatial cognition and behavior, we have decided to exclude them from the study domain because at present their relevance is peripheral. Furthermore, current methods of instrumentation for brain observation do not generally provide sufficiently detailed measures of localized activity in the brain to provide insights at the level of current questions in geographic information science. Secondly, we exclude cognition of spatial relations and positions at astronomical scales—whereas they are also of potential interest, such spaces are so unlike terrestrial spaces that their inclusion here would be unlikely to provide insights for geographic cognition. And third, we exclude reasoning about purely geometric figures and patterns, again because of marginal relevance and space limitations, although, given the close ties between geographic reasoning and geometric reasoning, including the newer field of diagrammatic reasoning, we have included selected papers in cases where such papers provide useful insights into map cognition or geographic cognition.

In the remainder of this paper, we first provide a history of the field, including the several independent research threads that converged in the 1980s to produce a cognitive research theme in geographic information science. Next, we discuss the major research themes in cognitive geography and geographic cognition today, under the headings of acquisition of geographic knowledge, mental representations of geographic knowledge, geographic knowledge use, and communication of geographic information. Then, in the last main section of this paper, we review the major topics considered for investigation under the Varenius project (Kemp et al. 1997), emphasizing the three researchable topics identified under the project, and mentioning other topics considered.

Maps have been used to provide external representations of geographic information for thousands of years, and Thrower (1972) provides an excellent review of the history of maps. Also, as revealed by its etymology, geometry itself is said to have had its origins in land surveying after annual floods of the Nile river. Connections between geospatial cognition and human activity and survival are even more ancient. Although many aspects of this may be universal, researchers have found wide cross-cultural variation in how geographic space is conceptualized for tasks such as navigation (Gladwin 1970, Lewis 1972), and research on cognition and behavior at geographic scales developed in several fields and disciplines. In the remainder of this section, we review some of the research streams that came together in the 1980s to form a cognitive foundation for geographic information science.

The origins of psychology as a scientific discipline are variously dated to the founding of Wundt's laboratory in Leipzig in 1879 or to the publication of James' Principles of Psychology in 1890 or to the founding of the American Psychological Association in 1892 (Hilgard 1987). Although cognitive psychology can be traced back as far as 19th Century Germany, to the psychophysics studies of Weber, Fechner, and Wundt, to the mental operations researchstudies of Donders, and to thememory experiments on memory of Ebbinghaus, its continuity was interrupted by the hegemony of behaviorism. The modern origins of cognitive psychology date from the 1940s and 1950s. Two parallel but unrelated strands united in the development of theories of information processing. One strand was practical, growing out of the human factors work that was part of the war effort. The other strand was theoretical, primarily borrowed from information theory and formal linguistics. These strands were woven together by George Miller (e.g., Miller 1956) and his colleagues and students in the United States, and Donald Broadbent (e.g., Broadbent 1958) and his colleagues and students in England, among many others. If an official date is needed, the publication of Neisser's book Cognitive Psychology (1967) serves as well as any for the beginning of cognitive psychology as a separate field.

The pursuit of spatial cognition was delayed even after the restrictions of behaviorism were overcome, due to a bias among many leading researchers based on the idea that the underlying language of thought was like language, and that the visual and spatial world could be reduced to language processing. Research on imagery provided a persuasive case that, at best, such reductionism ignored the truly fascinating issues (Kosslyn 1980, Shepard and Podgorny 1978). In the background, developmental psychologists were rediscovering Piaget, including his work on children's concepts of space. At the same time, geographers were investigating how people perceived and remembered the geographic world, and neuropsychologists were recording activation in rats' brains as they learned mazes (O'Keefe and Nadel 1978). More recently, Landau and Jackendoff (1993) provided a useful review of this topic.

Trowbridge's (1913) early paper aside, many people would date the modern period of cognitive studies of geographic environments to the work of E. C. Tolman. His classic 1948 paper 'Cognitive maps in rats and men' introduced the term 'cognitive map', and made an explicit link between experimental behavior of laboratory animals on one hand, and wayfinding and navigation abilities of people on the other (Tolman 1948). Two other early benchmark works of great influence were Piaget and Inhelder's 1956 book L'Espace Chez L'Enfant (translated as The Child's Conception of Space; Piaget and Inhelder 1956), and Kevin Lynch's 1960 book The Image of the City (Lynch 1960). Piaget and Inhelder were psychologists, but Lynch was an urban planner and landscape architect.

The period 1978–1985 saw serious empirical work on geographic cognition being conducted and published by psychologists. Most of these efforts were aimed at revealing how environments are mentally represented, by focusing on distortions in judgments about the environment. Stevens and Coupe (1978) were among the first to provide empirical evidence of hierarchical spatial reasoning, and showed how this powerful heuristic can distort judgments and memory of spatial relations. Hirtle and Jonides (1985) showed that hierarchical organization can be based on function as well as boundaries, and that it affects distance as well as direction judgments. Evans and Pezdek (1980) used reaction times to study distance judgments, finding evidence for mental rotation effects for environments learned from maps but not for environments learned from experience. Tversky (1981) presented evidence that perceptual organizing principles can distort judgments of spatial relations. For example, Americans typically think that South America is aligned directly south of North America, when in fact most of South America is much farther east. Holyoak and Mah (1982) showed that the perspective taken on an environment distorts distance judgments, so that near distances are judged to be relatively larger than far ones. Thorndyke and Hayes-Roth (1982) studied acquisition of environments from maps or exploration. They found superior direction judgments in the group that learned from exploration, and superior straight-line distance estimates from those who learned from maps. Kozlowski and Bryant (1977) found that people's estimates of their senses of direction predicted acquisition of a computer maze. There are several reviews of this foundational literature, including those by Evans (1980), Golledge and Stimson (1997), Downs and Stea (1977), and Tversky (in press, a and b).

(1999a).

One can find some evidence that individual geographers were interested in behavioral research before 1960, with work on imaginary worlds (Wright 1947, Kirk 1951) and the perception of hazardous environments (White 1945). However, behavioral research became an important part of the discipline of geography during the 1960s and 1970s. Research in the 1960s related to images of the city (Lynch 1960), environmental images (Lowenthal 1961), decision-making processes (Wolpert 1964), and mental maps (Gould 1966) inspired a generation of geographers to consider behavioral issues. They were soon applying new methods to old ideas (Rushton 1969) and relating concepts developed in behavioral geography to more-mainstream geographic problems (Horton and Reynolds 1971). Blaut and Stea (1971) published an important article that was an early presentation of cognitive foundations for geographic learning. A number of books appeared that formed the foundation of this new behavioral interest. Cox and Golledge (1969) considered the behavioral problems in geography. Geographers and psychologists began to collaborate in the 1970s, and some of these collaborations gave rise to books that considered their common interests in learning about the geographic environment, cognitive maps, and spatial behavior (Downs and Stea 1973, 1977, Moore and Golledge 1976).

Although early behavioral geography has had both its critics (Bunting and Guelke 1979) and defenders (Downs 1979, Rushton 1979, Saarinen 1979), behavioral geography has continued to grow and make progress (Aitken 1991, 1992, Kitchin 1996). Behavioral topic considered by geographers expanded into a variety of interests in the 1980s, including earthquake hazard information (Palm 1981), the use of imagery to store geographic information (Lloyd 1982), the spatial abilities of the sexes (Gilmartin and Patton 1984), consumers' cognition of distance (Coshall 1985), and the influence of anchor points in the environment (Couclelis et al. 1987).

The current decade has seen more collaboration between geographers and psychologistsgeographers, psychologists, computer scientists, linguists, and others with a focus on spatial cognition. A special issue of GeoForum in 1992 presented a collection of papers related to cognitive issues. A book edited by Gärling and Golledge (1993) presented both geographical and psychological approaches to studying behavior and environment. Portugali (1996) edited a book on the construction of cognitive maps, with chapters by both geographers and psychologists. Other recently published books providing a geographic perspective to cognitive issues are by Golledge and Stimson (1997), Lloyd (1997),and Golledge (1998), and Kitchin and Freundschuh (in press). For a review of behavioral geography within American academic geography, see Golledge and Aitken (1991).

Although studies on the perception of orientation (Gulliver 1908) and imaginary maps (Trowbridge 1913) showed an early interest in the cognitive processes used in map reading, cartographers did not demonstrate much additional interest in cognition until the 1970s. An important exception from the 1950s was a landmark study of distortions in perceived sizes of cartographic symbols. Flannery (1956, 1971) examined graduated circles, a cartographic symbol technique which originally scaled circles so that their total areas were in linear relation to some quantity being symbolized, such as city population. Flannery noted a tendency to underestimate the relative sizes of larger circles, and he calibrated this perceptual bias using psychophysical laboratory methods. The result of his work was to add to conventional cartographic practice a rule whereby radii of graduated circles were scaled in proportion to 0.57 times the logarithm of the quantities they were to represent (Robinson 1960). This correction introduces a systematic geometric exaggeration of larger circles, in principle to compensate in advance for the underestimation evidently found in normal perception. In a communication model of cartography, if the decoder (human perception during map reading) is known to introduce systematic distortions, then the inverse of the distortion should be used when constructing the map, so that the perceived map is unbiased.

Maps must provide accurate information to be useful, but they also must have an understandable message and be aesthetically pleasing. When cartographers began to study the nature of maps to understand symbolization and design principles (Robinson 1952, Robinson and Petchenik 1976), this resulted in an appreciation of maps as communication tools (Board 1967, Kolány 1969) and the discovery of a need to understand the cognitive processes used by map readers. To fulfill this need, some cartographers embraced an experimental paradigm and studied the interaction between the map and map reader. Sheppard and Adams (1971) studied drivers' use road maps for route finding. Other important early issues were the organization of information (Dent 1972, Lloyd and Yehl 1979), the perceptual response to cartographic symbols (Cox 1976, Gilmartin 1981, Kimerling 1985), and the visual comparison of maps (Lloyd and Steinke 1977, Muehrcke 1973).

In additional to being familiar with computer graphics technology, the current generation of cartographers must be part artist and part cognitive scientist as they try to construct better maps (MacEachren 1995). Some of the cognitive topics recently considered by cartographers include use of color on maps (Brewer et al. 1997, Brewer and Olson 1997, Mersey 1990), visual search processes used in map reading (Lloyd 1997, Nelson 1995), and learning processes used with maps and graphics (Lloyd 1994, Lloyd and Carbone 1995). The idea of using visualization to discover patterns has only recently been discovered by a number of disciplines, but is an old and familiar concept for cartographers (MacEachren and Taylor 1994, Monmonier 1990). The design and use of interactive maps (Patton and Cammack 1996, Slocum and Egbert 1993) and map animations (DiBiase et al. 1992, Peterson 1995) have been of particular interest to cartographers.

As the field of artificial intelligence matured, the knowledge representation and processing methods developed in the laboratory were applied to real world problems in the 1970s. Several applications to geographic information became visible. In 1978, Benjamin Kuipers published an elaborate model for representing human knowledge of large-scale space in an artificial intelligence framework (Kuipers 1978). Kuipers' model built largely on the intuitive and descriptive work of Kevin Lynch (1960). In a follow-up paper, Kuipers discusses the 'map in the head' metaphor frequently employed to account for people's ability to find their way in space (Kuipers 1982). The author uses computational models to refine the too simplistic metaphor and to discuss its implications in detail. In a subsequent paper, Kuipers discusses cognitive maps, their structure, and potential alternatives by employing a thought experiment in robotics (Kuipers 1983).

Davis (1983, 1986) looked at a cognitive map as a knowledge base; he developed a theory of representation, retrieval, and assimilation of geographic knowledge and implemented his theory in the MERCATOR system. MERCATOR is conceived for the use by a robot whose task is to build up a coherent representation of his visually perceived environment. Yeap (1988) also developed a computational theory of cognitive maps. Yeap's work emphasizes the cooperation of different loosely coupled modules representing different levels of information. The approach is motivated by Marr's (1982) investigations into the human representation and processing of visual information. Finally, Munro and Hirtle (1989) and Wender (1989) have developed connectionist models of cognitive maps, in contrast to the symbolic approaches listed above.

Critical evidence from linguistics entered the picture in 1983 with Leonard Talmy's seminal work on how space is structured in language (Talmy 1983). This paper started the important cognitive–linguistic research thread in cognitive geographic research. The artificial intelligence system CITYTOUR (André et al. 1987) was designed to answer natural language questions about the spatial relationships between objects in a city. Approaches from artificial intelligence were also applied to classic problems in cartography, especially automated name placement and other aspects of map design (Buttenfield and Mark 1990, Freeman and Ahn 1984). Couclelis (1986) reviewed artificial intelligence in geography during its early stage of widespread impact in the discipline.

Wayfinding, defined as the mental processes involved in determining a route between two points and then following that route, has long been an important site for studying spatial cognition. Kevin Lynch's 1960 book Image of the City paid particular attention to making cities more navigable. In a dissertation aimed at modeling common-sense reasoning in general, Benjamin Kuipers chose learning the geography of a place as a case study. The main results were published in a journal article already mentioned above, that was to have considerable influence on the field (Kuipers 1978). Shortly after the publication of that work, Riesbeck (1980) described a related problem and implemented a system to judge the clarity of driving directions, given no knowledge of the actual geographic layout.

In the 1980s, the development of microcomputers made it possible to consider designing navigation aid systems for private automobiles, that would keep track of the location of the vehicle, relate that position to an on-board digital street map, and provide navigation assistance to the driver. An obvious way to communicate with the driver would be to display maps, and this was the design of early implemented systems such as Etak's Navigator (Zavoli et al. 1985). A parallel line of work developed systems to provide verbal descriptions of routes, mainly to give to someone renting an automobile and requiring directions to some attraction. Elliott and Lesk (1982), Streeter et al. (1985), and Streeter and Vitello (1986) all studied the nature and content of driving directions, and related these characteristics to principles of knowledge representation in artificial intelligence. This line of work was picked up by people in the GIS community (Mark 1985, Mark and McGranaghan 1986, Mark et al. 1987, McGranaghan et al. 1987), and by others working on wayfinding and navigation (Golledge et al. 1993, Gopal et al. 1989). By the late 1980s, this thread was being related to other aspects of cognitive studies of geographic space and process.

The development of a series of conferences with refereed proceedings, under the name 'Conference on Spatial Information Theory' (COSIT), was an important factor in the development of a community and field of study in cognitive foundations of geographic information science, and the maturation of the field. The COSIT meetings grew out of a series of workshops, NATO Advanced Study Institutes (Mark and Frank 1991), and NSF-sponsored specialist meetings concerned with cognitive and applied aspects of representing large-scale space, particularly geographic space. In these meetings, the need for a well-founded theory on spatial information processing was identified. The conference series was established in 1993 as an interdisciplinary biannual European conference on the representation and processing of information about large scale (geographic) space after a successful international conference on the topic had been organized by Andrew Frank and others in Pisa in 1992 (Frank et al. 1992). The 1992 Pisa meeting has subsequently been informally referred to as 'COSIT zero'. After two successful European COSIT conferences (COSIT'93, Elba, Italy, Frank and Campari 1993; and COSIT'95, Semmering, Austria, Frank and Kuhn 1995), the conference became a truly international enterprise when COSIT'97 was held in the United States (Hirtle and Frank 1997). COSIT'99 is scheduled to take place in Germany. COSIT brings together researchers and methodologies in the area of spatial information theory from different disciplines, in particular: Geography, Geodesy, and Geoinformation Science; Computer Science and Artificial Intelligence; Cognitive Science; Cognitive and Environmental Psychology; Architecture and Environmental Design; Cognitive Anthropology and Psycholinguistics; and Philosophy of Mind. COSIT covers theoretical implications of empirical investigations, formal models, applications, and spatial information technology, and the community of like-minded scholars that participates in the COSIT meetings is evidence of a coherent field of study at the interface between cognitive science and geographic information science.

Much of the research reviewed in this section was not conducted in the framework of geographic information systems or geographic information science, or even in the context of computation. However, these research themes and communities are critical input to computational theories of geographic cognition, and to formal models of geographic space, phenomena, and features that contribute to the foundations of geographic information science.

Clearly, a multidisciplinary effort is required to develop and validate cognitive models for geographic space that are compatible with computation and that can form part of the theoretical grounding of GIS. The fields of geography and cartography provide concepts and relations for geographic space. Cognitive and environmental psychology provide empirical investigations and models of human factors. Artificial intelligence provides formalizations and computational models, as well as ontologies and structures for the development of cognitive models. Linguistics provides a link to the construction of descriptive spatial phrases and the use of spatial metaphors. Philosophy provides a theoretical foundation for spatial concepts.

In this section we review major research themes in cognitive geography and geographic cognition, not by disciplinary perspective or research methods, but by stages in a hypothetical information flow model for spatial and geographic cognition.

For humans, knowledge of space is acquired in many different ways. Although the prototypic experience may be actual exploration, by the time children are talking, telling them where to find the cookies may be all that is needed. Still older children can and do use maps for finding their way. Actual exploration itself is complex. There is good evidence that visual information about space has different qualities from kinesthetic and vestibular information, and that both differ from acoustic or tactile information (e.g., Berthoz et al. 1995, Loomis et al. 1993). For example, updating orientation changes is more accurate following sightless real movement than imagined movement, although no parallel differences exist for translation. Characterizing the kind of spatial information imparted by each modality and describing how they are integrated in actual behavior are topics ripe for investigation. Despite their differences, all of these modalities provide valid and often substitutable information about space, athough the embodied information is more important for local guided navigation, and the cognitive for judgements in larger-scale space.

The process of extracting geographic knowledge from locomotion through a space requires a series of complex interactions. As Montello (1997) argues, the conversion of sensorimotor information into geographic knowledge is an indirect process, in which environmental features are used to generate spatial characteristics, such as distance information. Such environmental features include not only physical characteristics, such as turns, landmarks, intersections, and barriers, but also travel time, travel effort, and aesthetic qualities of the space. Such characteristics are often visually acquired, but might also be acquired through other modalities. Virtual reality (VR) provides an alternative spatio–temporal experience for locomotion, which can mimic some of the complexities of movement through space (Berendt and Jansen-Osmann 1997), but may not provide a full sensory experience to the traveler. Geographic information systems themselves also provide ways to learn unfamiliar spaces and reason about geographic phenomena. Subtle differences provided by perceptual texture gradients or kinesthetic variations in surface qualities are often lacking in VR simulations, and current GISs certainly do not provide direct transperceptual experiences, but often are closer to an interactive version of map use.

Alternative media for acquisition often focus on the schematization of geographic knowledge for improving communication efficiency. This might include the use of maps (Head 1991), descriptions in natural language (Taylor and Tversky 1992), or spatial abstractions, such as data charts or other visualization techniques (MacEachren 1995; Tversky, in press c). The ability to construct mental models from text is an essential component to the understanding of narrative stories (e.g. Morrow et al. 1987). Furthermore, spatial mental models constructed from text are similar in content to those constructed by studying maps of same scene in terms of generic spatial knowledge (e.g., Federico and Franklin 1997, Taylor and Tversky 1992). However, geographic information acquired from pictorial input appears to be retained longer than similar information obtained from text (Federico and Franklin 1997).

For over two decades, the acquisition of spatial knowledge has been modeled by the continuum of landmark, route, and survey knowledge (Siegel and White 1975). This trichotomy, while once assumed to be acquired in a strict ordinal fashion, is now believed to be acquired, at least partially, in parallel in many situations (Hirtle and Hudson 1991, Thorndyke and Hayes-Roth 1982). Route information is gleaned in parallel with identification of landmarks (Presson and Montello 1988), and survey information is constructed in parallel with building routes when possible (Moar and Carleton 1982).

It is important not to minimize the linkage between the environment and the representational schemes that are used to navigate through it. As Hutchins (1995) argued, the environment provides a context for learning with constant feedback and adjustment. Learning, according to Hutchins, is the adaptive reorganization in a complex system which includes the environment and communication among actors in that environment. Edwards (1997) argues for a combination of cognitive and geometric approaches, in which two representational structures, views and trajectories, provide the basic building blocks of spatial cognition. Hiscognition; his framework, which he calls geocognostics, provides a model for understanding the biases that are inherent in the learning and use of GIS, among other applications. This is but one of several models that combine views and trajectories.

Any behavioral expression of spatial knowledge requires both knowledge representation and knowledge retrieval. Separating the contributions of representation and retrieval is difficult, if not impossible. Converging evidence from different retrieval tasks strengthens the case that effects are due to knowledge representation. Even so, because different knowledge is retrieved for different tasks, knowledge representations of space are probably not best conceived of as coherent, unchanging wholes, but rather as conglomerations of information drawn from different sources and modalities and pulled together for a particular purpose.

With those provisos in mind, can spatial knowledge be characterized in any general ways? Several metaphors have been proposed for representation and processing of geographic knowledge: the cognitive map, the cognitive atlas, and the cognitive collage. To most psychologists at least, the term 'cognitive map' has connotations of metric properties, like a drafted, cartographic map. This conception comes mostly from Kosslyn's work on imagery, which has argued that images are like internalized perceptions and quite true to what is seen (Kosslyn 1980). 'Cognitive atlas' was introduced by Kuipers (1982) as a term to refer to a collection of cognitive maps, perhaps of different scales, and with gaps. Tversky (1993) introduced the term 'cognitive collage' to emphasize the fact that mental representations driving judgments and wayfinding are fragmented, partial, constructed, and multi-media.

There seems to be ample evidence that spatial knowledge does not have the metric qualities that maps do. As noted earlier, geographic space, though locally flat, is organized hierarchically, both in cognition and in administrative practice (e.g., Allen and Kirasic 1985, Chase 1983, Hirtle and Jonides 1985, Maki 1981, McNamara 1986, Stevens and Coupe 1978). Spatial information is organized by geographic boundaries, by economic categories, and by functional groupings of all kinds. The consequences of imposing hierarchical structure are that distances within categories are judged to be smaller than distances between categories (Hirtle and Jonides 1985), that direction judgments between categories are faster than those within (e.g., Maki 1981, Wilton 1979), and that directions of elements within a unit are distorted to the directions of the encompassing unit (Stevens and Coupe 1978, Tversky 1981).

Geographic elements, then, are organized as parts of larger geographic units. That hierarchical organization affects distance and direction judgments. Geographic elements are also organized one to another. When one of the elements is a better landmark, then that organization is asymmetric. Ordinary elements are judged to be closer to landmarks than landmarks to ordinary elements, violating usual metric conventions (e.g., McNamara and Diwadkar 1997, Sadalla et al. 1980). When the geographic elements are more or less comparable, such as North and South America, then they are organized together and remembered as more aligned geographically than they actually are (Tversky 1981).

In retrieving the relevant geographic information to make a judgment, people may take a particular perspective on the set of information. The perspective, too, can alter judgments. Holyoak and Mah (1982) found that people judged distances between pairs of near cities to be larger relative to pairs of distant cities, where near and distant were determined by an imagined east- or west-coast perspective.

The arrangement of the physical environment as experienced is also known to affect distance judgments, specifically, the amount of clutter or the number of intersections and nodes or the presence of barriers. On the whole, these increase distance estimates, but often reversals are obtained (e.g., Newcombe and Liben 1982, Sadalla and Magel 1980, Sadalla and Staplin 1980a, b, Thorndyke 1981). Irregular environments are remembered as more regular; for example, streets and rivers as straighter (Chase and Chi 1981, Milgram and Jodelet 1976) or more parallel or perpendicular than they actually are (Byrne 1979, Golledge and Spector 1978, Moar and Bower 1983, Tversky 1981).

These are only some of the ways that people's knowledge of the geographic world differs systematically from the actual geographic world. Together, these findings suggest that mental representations of the geographic world are not stable, map-like entities that can be consulted as maps can be viewed. Rather, they seem to be constructed for a particular goal, drawing from the multiple sources of scattered information available those bits of information that seem relevant. Mental representations of geographic information seem to be constructed from elements, such as roads, landmarks, cities, land masses, the spatial relations among them, and the spatial relations of them to the larger units encompassing them. This schematization of the geographic world provides a framework for integrating information from different sources, modalities, and occasions. Like all schematizations, it also simplifies the complex and categorizes the continuous, allowing distortions as well as integration (Tversky 1992, Tversky and Lee 1998).

Not only are the sources of knowledge of space diverse, but the uses to which that knowledge is put are similarly varied. The prototypical use is finding one's way, but spatial knowledge is also used to make geographic judgments, such as estimates of distance and direction. Indeed, wayfinding seems to require implicit if not explicit judgments of distance and direction.

McDermott and Davis (1984) developed an artificial intelligence model for planning routes through uncertain territory. In their system, topological and imprecise metric information are represented and used for selecting a promising path towards a given goal. Route planning is modeled as a process of finding the overall direction and topology of the path, then filling in the details by deciding how to go around barriers.

Towards the end of the 1980s researchers from different disciplines independently developed formal and computational approaches for representing and processing knowledge about large-scale space. While Egenhofer's (1989) and Frank's (1991) work was directly driven by issues in geographic reasoning, the approaches of Freksa (1991a, b, 1992, Freksa and Zimmermann 1992, Zimmermann and Freksa 1993, 1996, Freksa and Barkowsky 1996), Hernandez (1994), Mukerjee and Joe (1990), and Schlieder (1993) were largely motivated by cognitive considerations independent of specific applications. The work of Cohn and coworkers (e.g. Cohn et al. 1993), Faltings (1995), Guesgen (1989), and Ligozat (1994) has its roots in formal (logic-based, geometric, topologic) concepts while Jungert's work on navigation (e.g. Jungert 1988) grew out of database-oriented research concerned with encoding pictorial information.

Despite these differences of concern, the approaches converge remarkably with respect to some basic issues. The approaches have in common that they employ qualitative rather than quantitative information. Most of them are strongly related to Allen's work on qualitative temporal reasoning (Allen 1983, Freksa 1992). While temporal reasoning is concerned with one dimension only (the time axis), spatial reasoning deals with orientation information in addition to distance information. As it became evident that these independently developed approaches have much in common (Freksa and Röhrig 1993) and are suited to complement one another (Röhrig 1998), the European research network Spacenet was formed to promote interaction and exchange between the different perspectives and approaches. A recent overview by Cohn (1997) covers qualitative spatial representation and reasoning techniques.

Reasoning about large-scale space is relevant not only to immediate applications in geography, but also is particularly relevant to robot navigation (e.g., Kuipers and Levitt 1988) and to human navigation in virtual environments (May et al. 1995). The comparison between human navigation performance in real and in virtual environments can provide important insights into the underlying cognitive mechanisms (Mallot et al. 1998)

Orientation and navigation performance can be effectively enhanced through geographic maps. The cognitive processes involved in representing knowledge in maps and in map reading therefore are of particular interest for studying the use of geographic knowledge. Cartographic research about the use of maps as media for representing spatial knowledge (McEachren 1995) is supplemented by formal studies describing cognitive processes involved in extracting and combining knowledge from maps to draw inferences useful for identifying geographic landmarks and routes (Barkowsky and Freksa 1997, Berendt et al. 1998). It is still a topic of investigation to determine under which circumstances spatial, functional, or featural aspects are most heavily used in solving particular tasks. The German Science Foundation (Deutsche Forschungsgemeinschaft) supports a 6-year priority program for interdisciplinary basic research on spatial cognition (Freksa et al. 1998).

Maps can be viewed as special diagrams specifically designed for representing geographic knowledge about spatial regions. Diagrammatic reasoning research therefore provides important foundations for the analysis and synthesis of spatial reasoning in geographic contexts. Diagrammatic reasoning is one of the early research areas in artificial intelligence; Glasgow et al. (1995) published an excellent collection of important contributions to the field.

Besides viewing maps as vehicles for performing spatial inferences, they can be studied as a means of communicating spatial knowledge. Thus the topic of generating sketch maps for the purpose of describing places or routes or for complementing verbal descriptions of space is of interest (Tappe and Habel 1998). One ancient and reliable means of conveying geographic information is a map. Useful maps, like other useful graphics, are not simply reductions in size of actual worlds; rather, useful maps extract the essential information and eliminate the inessential (e.g., Tversky 1999b).in press c). Of course, what is essential and what is inessential depends on the goals of the user. The schematization of graphics often parallels the schematization of the mind.

Yet another way to convey spatial information is through language, also a venerable way to communicate about space.space Useful analyses of the connections between language and space appear in many of the papers in Bloom et al. (1996) as well as a paper by Landau and Jackendoff (1993). Spatial expressions typically describe a target or figural object in relation to a background object (e.g., Talmy 1983), as in 'the church is west of the town hall'. Languages typically have several different reference systems for describing spatial relations. 'West of' uses an extrinsic (also called geocentric or environmental) reference system. An expression like 'the church is left of the town hall' is ambiguous in English, using either a relative or intrinsic reference system. According to Levinson (1996), a relative reference system is centered on a viewer and uses a three-term relation: the church is in front of the town hall from the viewer's perspective. An intrinsic reference system is a two-term relation projected from the intrinsic sides of an object (or person). In this case, the church is in front of the front of the town hall. Neuropsychological evidence indicates that locations are perceived according to all three (and maybe more) reference systems (Behrmann and Tipper 1998).

What perspective is used for description depends on a number of variables. There are some languages that do not use the relative reference frame, using an extrinsic reference frame instead (Levinson 1996). For languages that use all three reference systems, the relative frame seems to be used primarily for environments that can be viewed from a single point (Taylor and Tversky 1996, Ullmer-Ehrich 1982). For larger environments that cannot be seen from a single viewpoint, people's descriptions use both extrinsic reference frames as in survey descriptions and intrinsic reference frames with a person as the central reference object as in route descriptions (Taylor and Tversky 1996). Quite frequently, people switch perspective, usually without signaling. When environments have features on several size scales and contain many alternative routes, there is a shift in preference toward more survey descriptions.

An additional issue arises in two-person interactions. Whose perspective do speakers adopt, their own or that of their addressees? Schober (1993) found that in the majority of cases, speakers take the perspectives of their addressees, and that this tendency was even more pronounced in cases where the addressees were unknown. In a variation of Schober's task, Mainwaring et al. (unpublished) also found that both American and Japanese speakers preferred their addressees' perspectives. They attributed this to considerations of relative cognitive load rather than politeness as variations in cognitive load led to variations in perspective.

In face-to-face communication, gesture as well as words convey spatial location. Emmorey et al. (in preparation) found that the oral component alone of videotaped descriptions was not sufficient for conveying the environment; the gestures in many cases supplemented and disambiguated the verbal information. Many of the accompanying gestures were iconic; for example, gestures indicating turns and intersections. Gestures accompanying spatial descriptions are in many cases language-specific, following language-specific schematizations of space (Kita et al. in press).

Route directions are of particular interest as an aid to wayfinding. In his analysis of a large corpus of route descriptions, Denis (1997) found two types of statements: references to landmarks and prescriptions of actions. Route directions can be segmented by actions that use landmarks as referents. In another set of studies, judges rated spontaneous directions for quality. Those rated as good followed the structure proposed by Denis (1997). Travelers using the highly rated directions were more likely to find their way in Venice than travelers using poorly rated directions (Denis et al. 1998). Sketch maps are also a good tool for wayfinding, and, in fact, schematize routes in much the same way as verbal directions (Tversky and Lee 1998).

The Varenius project (Kemp et al. 1997) is organized around three main topical areas in GI Science, one of which is cognitive models of geographic space. Each topic area is directed by a panel of five individuals, and the authors of this paper make up the cognitive panel. Each panel has identified several important topics for further research, and has ranked them in order to select three topics that became Varenius research initiatives and the subjects for workshops. In this section, we review the three researchable topics that were identified by the Varenius project as the cognitive topics of highest priority. We also review other identified topics that, while worthwhile, were considered lower in priority. Research initiatives begin with Specialist Meetings, which are workshops that bring together 30–40 researchers from a wide range of disciplinary perspectives to provide detailed priorities for researching the topic. Such meetings were conducted for each of the following three topics between May 1998 and February 1999.

Degree of geographic detail is one of the most poorly understood and most confusing of the fundamental geographic concepts that underlie our cognition of geographic objects, spaces, and phenomena (Montello 1993). Scale, although often used ambiguously and poorly defined, is nevertheless an important component of naive geography, which asserts that geographic scales are different in fundamental yet unspecified ways from things at other scales (Egenhofer and Mark 1995b). If this is so, do certain spatial processes suddenly come into existence at some specific scale? Are they not present at micro-levels (larger scales)? Do all spatial processes have an emergence threshold, or are spatial and geographic processes really scale invariant but ignored as larger and larger scales (i.e., smaller and smaller areas) are examined? How do we rationalize the different uses of the term (e.g., in cartography, compared to in environmental modeling)? Is scale more important in the physical domain than in the human? Are spatial cognitive processes scale-dependent? What scales of representation lend themselves most to visualization, and to other forms of representation? How does a scale change influence granularity or clarity of data? Does a change in scale involve loss of original structure and emergence of artificial structure or patterns? To what extent is scale at the crux of traditional geographic arguments of form versus process? Does scale change involve changing models? How do visual representations of spatial phenomena need to change with a change in scale? While GIS and other automated cartographic systems allow rapid changes of display scale, they do not take into account the perceptual requirements of the viewer. Do the perceptual requirements of the viewer change along with a change in scale, or are they constant across many scales? What types of data are most amenable to display at multiple scales? How should those data be manipulated when scale changes, through cartographic generalization? All of these questions are important to this research initiativel.

As society makes the transition to digital worlds, associated metaphors for geographic detail are likely to change also. Metric scale or representative fraction, the measure of geographic detail dominant in the cartographic world, has no well-defined meaning in a digital world of seamless perspectives on geography in which the user is free to zoom and pan at will. Other metaphors, such as the view from space, may replace metric scale with less familiar dimensions such as the distance of the viewpoint from Earth, as they do in Microsoft's Encarta Atlas. This suggests two fundamental objectives for this initiative, in addition to those identified earlier: 1) can we identify the fundamental, invariant dimensions of the concept of geographic detail that survive the transition from analog to digital, and 2) can we identify the mapping between these dimensions and the terms and metaphors commonly used in naive geography? The initiative on 'Formal Concepts of Geographic Detail' is led by Daniel Montello and Reginald Golledge, both of the University of California, Santa Barbara, and its Specialist Meeting was held in Santa Barbara in May of 1998.

The ability to manipulate, interpret, and store information about changing environments is a critical skill human survival, and also is very important for geographic information science. Models of the cognition aspects of dynamic spatial representations are necessary for understanding temporal and spatial changes in spaces or maps, for the manipulation of temporal geographic data, and for navigation through changing spaces. Furthermore, the use of representational information may be dependent on the context of the problem, with different entity types resulting in the adoption of different spatial metaphors for reasoning and understanding. For example, land use changes might be viewed as changes in attributes of a fixed location, whereas an advancing forest fire is thought of as a moving entity of change shape and size. And, at a different temporal scale, the former process, involving no real motion, might be talked and reasoned about as the spread or sprawl of development. Some other examples of dynamic geographic processes include navigation through changed environments, diffusion of diseases, and much slower processes such as glaciations, mountain-building episodes, or continental drift and plate tectonics. At a database level, we are concerned with issues such as forming discrete representations of continuous phenomena or continuous representations of discrete phenomena. Cartographically, the emphasis is on animation, but many methods have been used to show temporal phenomena on static maps. The use of dynamic and manipulable interfaces also must be investigated within the same conceptual framework used for observing dynamic phenomena in the real world.

The Varenius research initiative on 'Cognition of Dynamic Phenomena and their Representations' is led by Stephen Hirtle of the University of Pittsburgh and Alan MacEachren of the Pennsylvania State University. The initiative takes a dual and parallel look at dynamic phenomena in geographic space itself, and at their representation in dynamic displays of geographic information. If research finds that there are systematic differences in human cognitive responses to various kinds of change and motion in geographic space, then different representations may be appropriate for the different situations. If different kinds of computer displays also trigger different kinds of human memory, reasoning, or decision-making, then the match between cognitive models for the phenomenon being represented and those for the display methods will influence how intuitive and usable the display will be. This research initiative held its Specialist Meeting in Pennsylvania in October of 1998.

Space can be experienced directly, through vision, hearing, touch, and other modalities, as well as indirectly, primarily through language. Space can be viewed from many different perspectives, and conceived of from perspectives that have not or cannot be viewed. How do people interact with multiple modalities and multiple frames of reference? How do they integrate and reconcile the varied information, if and when they do? What are the relative advantages and disadvantages of each kind or source of spatial information? These are issues that have arisen in linguistics, philosophy, computer science, anthropology, and psychology, as well as in geography, in theoretical as well as applied contexts. However, there are many open questions, especially with respect to human behavior and learning in natural situations. Understanding how people combine or juggle information from a variety of sources in a variety of forms is important to geographic information science and GIS in at least two ways. First, it is important in deciding how to provide additional information to system users, dependent in part upon what they already know. Second, the ways in which people represent and combine geographic information may help in the design of computerized systems to do the same thing.

Some specific topics serve as examples: relative, intrinsic, and absolute reference frames for describing locations; heads-up and north-up maps in navigation systems; mixing gaze, route, and survey perspectives in descriptions; tactile, auditory, visual localization; orientation-free vs. orientation-specific representations; expressing differing modalities or frames through language; and cross-cultural differences in the use of reference frames. The Varenius initiative on 'Multiple Modes and Multiple Frames of Reference for Spatial Knowledge' is led by Scott M. Freundschuh of the University of Minnesota, Duluth, and by Holly Taylor of Tufts University, and the Specialist Meeting was held in Santa Barbara in February of 1999.

The Varenius project panel on cognitive models of geographic space identified several other cognitive research topics that, while not considered quite as important as the three detailed above, were still thought to be worthy of research attention. In this section, we give brief overviews of several of these topics.

Ontology of geographic entities. Categories are central to human cognition. Psychologists and other cognitive scientists have developed a well-established model of the nature of categories. The model, grounded in the work of Eleanor Rosch (1973, 1978) and reviewed in depth by Lakoff (1987), has shown that, while cognitive categories are often like mathematical sets, they frequently depart from classical set theory in important ways, having consensus best examples, internal structure, and indistinct category boundaries. Mark (1993) discussed how this model of categories might provide a theoretical foundation for definition of entity types in geospatial information transfer standards such as the US Spatial Data Transfer Standard (Fegeas et al. 1992). The above theory of cognitive categories is well-established and widely accepted. However, it is based almost entirely on studies of categorizations of biological entities, artifacts, and other manipulable objects. Recently, Smith and Mark (1998) provided evidence that geographic objects are different in fundamental (ontological) ways from the sorts of objects studied by Rosch and her colleagues and followers, and suggested that category formation might be different here also. A research initiative on this topic would focus on the nature of basic-level categories of geographic entities, and their role in geographic cognition. The committee recognized this as a very important topic in its foundational role in geographic information science. They also felt, however, that the researchable questions, as well as the methods to be used in investigating it, were well established in psychology. Thus this topic is a good candidate for research, but probably would not benefit as much from a workshop as would some of the other topics discussed and described above, where the approach to doing the research, and the priority subtopics, are less well defined.

Mental maps. The entire area of geographic knowledge at an individual level, or mental maps and related topics, is important. Research questions include the ways in which knowledge of geographic space is represented in the brain, how it is recalled, how people reason to derive new knowledge, and what are the relative roles of different sources of geographic information. Are there important differences due to scale? Is geographic information represented differently depending on its source, when it is learned from maps versus learned from text versus learned from real-world experience? Are there important cultural or sex-related differences in the answers to the above questions? Although these are critical topics with more research needed, there already is a considerable literature on the topic, and an active research committee. The probable marginal effect of a Varenius project initiative on mental maps on progress in this area was not judged to be as great as the probable effect on research progress of the three highest priority topics described above. Lastly, it is a very broad topic, and the topics addressed by the initiatives on detail and reference frames and modalities will contribute to understanding of mental maps.

Formalizing spatial relations. Spatial relations are one of the most distinctive aspects of spatial or geographic information, and thus a better understanding of the cognitive aspects of spatial relations, and their formalization in computational models, is critical to the advancement of geographic information science. As with some of the other topics discussed above, these is an active research community and research literature on spatial relations which has momentum independent of the Varenius project (for example, see Mark and Egenhofer 1994a, b, 1995, Egenhofer and Mark 1995a, Mark et al. 1995, Shariff et al. 1998), and thus the probable impact of a Varenius initiative on progress in formalizing spatial relations was not thought to be as high as for some of the other topics.

Additional topics. Six other topics were identified by the panel on Cognitive Models of Geographic Space as being of some potential interest and importance. These are less well-defined than the topics outlined earlier in this section, and are listed here for completeness and to alert readers to topics in need of attention by the research community.

One of these is the issue of place—what are the cognitive models of place and neighborhood, and can these be implemented in computational environments? What would a place-based, rather than coordinate-based GIS look like, and what could it do, and not do? Another topic would examine navigation in virtual spaces: how similar is this to navigation in real spaces, and how can the look and feel of virtual spaces be designed to maximize navigability? Another topic would address issues in the design of graphic displays and diagrammatic reasoning. This topic may be too far away from geographic information science, since it deals with diagrams as diagrams rather than as representations of geographic spaces. Some aspects of this topic are covered under the Varenius initiative on Cognition of Dynamic Phenomena and their Representations. The panel also felt that the role of experience in the ability to use displays was worthy of some attention from researchers, but arguments against giving it high priority are similar to those just presented for diagrammatic reasoning. Another topic would be to study the design and implementation of cognitive agents for GIS. This is somewhat related to the knowledge discovery initiative under the computational models panel of the Varenius project. Lastly, the very broad topic of the semantics and structure of geographic space was raised as an important goal.

In this paper, we have reviewed research in geographic cognition that provides part of the theoretical foundation of geographic information science. Free-standing research streams in cognitive science, behavioral geography, and cartography converged in the last decade or so with work on theoretical foundations for geographic information systems to produce a coherent research community that advances geographic information science, GIS, and the contributing fields and disciplines. The Internet is now delivering albeit simple GIS functions to the general public, and systems for use by untrained people provide new challenges for systems designers. Many of those challenges relate to the cognitive models of geographic space and phenomena that are held by members of the public (Egenhofer and Mark 1995b, Mark and Egenhofer 1996). Cognition by spatially-aware professionals and other experts must not be ignored.

Emerging geographic technologies such as Global Positioning System (GPS) receivers and wireless information systems also provide cognitive research challenges for GIScience. Can virtual scenes be surperimposed on the real world in such a way as to augment the geographic information available to people in the field? What are the cognitive implications of such systems, and how can knowledge of principles of human cognition of geographic spaces inform the design of augmented reality systems and other forms of geographic field computing?

There are also basic research challenges that lie in the nature of geographic cognition itself. How exactly does the relative size of objects or spaces influence how they are cognized, if it does at all? If geographic cognition is different from spatial cognition at other scales, is the difference somehow indexed to the size and physical capabilities of the human body? How many of the differences between CAD (computer-assisted design) and GIS software result from the differences in how their application domains are dealt with in human cognition, and can formalized knowledge of the exact ways that geographic and non-geographic spatial cognition differ be used to make better and more easily used software? Studies of geographic cognition, and of computational models based on findings of such studies, will continue to be an important basis for geographic information science.

Max J. Egenhofer

Janice Glasgow

Oliver Günther

John R. Herring

Donna J. Peuquet

Over the past ten years, a subfield of GIScience has been recognized that addresses the linkage between human thought regarding geographic space, and the mechanisms for implementing these concepts in computational models. This research area has developed an identity through a series of successful international conferences and the establishment of a journal. It has also been complemented through community activities such as international standardization efforts and GIS interoperability. Historically, much of the advancement in computational methods has occurred at—or close to—the implementation level, as exemplified by attention to the development of spatial access methods. Significant progress has been made at the levels of spatial data models and spatial query languages, although we note the lack of a comprehensive theoretical framework comparable to the relational data model in databases management systems. The difficult problems that need future research efforts are at the highly abstract level of capturing semantics of geographic information. A cognitive motivation is most promising as it shapes the focus on the users' needs and points of view, rather than on efficiency as in the case of a bottom-up system design. We also identify the need for new research in fields, models of qualitative spatial information, temporal aspects, knowledge discovery, and the integration of GIS with database management systems.

In the past, much research in the computational domain of geographic information systems (GIS) concerned the development of fast and efficient implementations of traditional cartographic concepts for data storage, retrieval, and analysis. Increased functionality had also characteristically been accompanied by increased conceptual complexity, as improvements were most often motivated by short-term needs, resulting in ad hoc solutions. With the increasing availability of GISs, there is an increasing need to provide users—from scientists to average citizens—with tools that allow them to solve their problems better in a more intuitive and user-friendly manner.

In light of these observations, a subfield within Geographic Information Science (GIScience) has developed over the last ten years that addresses the linkage between human thought about geographic space and the mechanisms of computational models. It is particularly concerned with the interface between the real world as perceived and computational geographic worlds. This aspect is important from a scientific perspective, because through such a cognivitely-motivated approach, geographic concepts that have always been intuitive but never formalized can be developed into a more formal framework. Such formalizations will enable space-time analyses of large-scale geographic processes that are impossible to perform without the aid of computers.

As part of this latter process, interactions with other fields, including mathematics, computer science, and statistics, need to be tailored. Through formalization, geographic concepts become unambiguously defined such that the danger of miscommunication and misuse is reduced. There is a strong tradition in some parts of GIS that draws upon fundamental mathematical principles, such as the use of algebraic topology , point-set theory , and partially ordered sets . Other approaches have relied on models developed in computer science, probably best shown in GIS debates over the use of relational and object-oriented or logic-based data models. Formalizations of geographic processes also lead to spatial languages—not natural language, but formal languages understandable to people as well as machines. These considerations make computational methods an indispensable part of GIScience.

Computational implementations are complementary to other areas of GIScience (e.g., cognitive models and geographies of the information society), and they play a central part in making advances in GIS theory accessible to a large audience, the audience of the users. While computational implementations undoubtedly require at one point engineers and programmers to cut code, the steps that fuel these implementations are complicated and cannot be accomplished without considering how people think about geographic space and time, how to translate human conceptualizations into formalisms that allow these processes to be repetitively consistent, and how to make people interact more naturally with information systems. These three concerns show how computational methods span across GIScience: Spatial thinking extends to the cognitive area; use and interaction about geographic information reaches out to the societal implications; and formalizations of geographic concepts are the core of the research agenda in computational methods.

This article assesses the progress and status of research in the area of computational methods for representing geographic information. After an assessment of this field's community with its major activities, conferences, and publication outlets (the next section), we review some of the results of 15 years of research in computational methods for representing geographic information. In the final section, we look ahead with a discussion of the needs for future research.

Researchers and practitioners in the geosciences have long been working on computer solutions to match their specific needs. Commercial GISs are among the most important outcomes of these efforts. Traditionally, computer scientists have been involved only marginally in the development of such systems. This started to change in the 1980s, and the increasing number of contacts is now bearing fruit. The design and implementation of data management tools for spatial applications is pursued by an interdisciplinary community of people from academia, government, and industry. A critical foundation lies in the concept of spatial database systems as an enabling technology for a variety of application software, such as CAD systems or GIS. There have been several interdisciplinary projects of high visibility, including the U.S. National Center for Geographic Information and Analysis (http://www.ncgia.org), the Sequoia 2000 project , and the Alexandria Digital Library .

It was also in the 1980s that a number of conference series were launched that provided a forum for the exchange of research results. In 1984, GIS researchers initiated the International Symposium on Spatial Data Handling (SDH), which quickly became a forum for communication between geographers, engineers, and computer scientists interested in analyzing and manipulating geographic data with the help of computers. Five years later, an NCGIA research initiative led to the First Symposium on Large Spatial Databases (SSD) . Compared to SDH, SSD has a stronger emphasis on computer technology and draws its audience primarily from researchers who focus on database systems. In 1993 a new direction was established with the First International Conference on Spatial Information Theory (COSIT), which aims at bringing together an interdisciplinary group of researchers who span an even wider range of disciplines, from cognitive science to geography and computer science. All these conference series continue to be held biennially, and their proceedings typically appear as books with major publishers, such as SSD , SDH , and COSIT . These activities focus on the GIScience community and are complemented by publications in the traditional computer science domain. Such conferences as ACM SIGMOD and Very Large Data Bases (VLDB) typically each year have a session on spatial data management. In addition, an annual ACM Workshop on GIS has been held since 1993, catering primarily to computer scientists who work on models for spatial data and their efficient implementations.

Progress in computational methods for representing geographic information can also be measured in terms of archival outlets that regularly publish articles on this topic. For many years, the International Journal of Geographic Information Systems (now the International Journal of Geographic Information Science) has been the umbrella for most GIS journal papers. Occasional papers on computational issues in GIS have appeared in several computer science journals, such as IEEE Transactions on Knowledge and Data Engineering, ACM Transactions on Database Systems, the Journal of Visual Languages and Computing, and the VLDB Journal, which published a special issue on Spatial Databases . With the formation of the journal Geoinformatica in 1996, the field passed another important milestone, resulting in a journal that is dedicated to research results at the interface between GIS and computer science. Overview articles on the status of the field are infrequent. The field moves quickly and several issues that were seen as critical in the early 1990s have lost their appeal. Two more recent reviews provide complementary material to the present assessment of the status in the field.

Although the book market in GIS has been growing quickly, there have been few books focusing on computational aspects of GIS. Laurini and Thompson exposed technical issues in GIS to a larger audience. Hanan Samet's books provide an in-depth coverage of spatial data structures and geo-algorithms. More recently, Worboys gave an up-to-date computing perspective on GIS.

Another measure of success is how results from research have found their way into widely available products and standards that may have a broad impact on designers and users. In both aspects, strong linkages to research results are very visible. GISs have become significantly easier to use through the adoption of a Windows-Icons-Menus-Pointers (WIMP) design. GISs also have started to embrace database management systems, an approach that has been suggested and requested since the early 1980s . It is now being enabled by industry's support of specialized spatial versions of database systems (e.g., Spatial Data Blades and the Spatial Data Option) or middleware (e.g., the Spatial Data Engine). Today's implementations show significant similarities with such research results as R-trees , quadtrees , models for topological relations , database architecture , and object-oriented models . Other impact, particularly at the level of geo-algorithms, is more difficult to assess since in most cases industry has abstained from publishing technical details about their approaches, and user documentations reveal no view behind the scenes.

At the organizational level, an important event happened in 1994, when an international group of GIS users and vendors founded the Open GIS Consortium (OGC, http://www.opengis.org). OGC has quickly become a powerful interest group to advance open systems approaches to geoprocessing and promote an Open Geodata Interoperability Specification. Such a computing framework and set of software specifications would support interoperability in the distributed management of geographic data (Kottman 1999). OGC seeks to make geographic data and geoprocessing an integral part of enterprise information systems. Together with other standardization efforts, such as the International Organization for Standardization's Technical Committee on Geographic Information/Geomatics (ISO TC/211), often diverse computational methods are being consolidated, thereby making GIS technologies more mature.

We have organized the large variety of computational methods into parts that follow the traditional approach used in computer science for modeling and implementing application domains. At the highest level, one is concerned with spatial data models and spatial query languages, as these parts of a GIS are visible and accessible to users. Since the late 1980s query languages have migrated from an end-user facility to an entry for application programmers, facilitating standardized storage, updating, and retrieval of information. Discussions about higher-level aspects of access have found new fields in the area of user interface design and human-computer interaction. Considerations about user interface design are truly at the interface between the cognitive and computational domains of GIScience and successful approaches typically require multidisciplinary participation, including designers and users. Current GIS user interface techniques have moved from the more archaic command lines through WIMP (Window-Icon-Menu-Pointer) designs to direct manipulation , an interaction mechanism by which users see and manipulate objects to affect commands .

To implement geographic concepts in a high-level, compact, and reusable way, spatial data types have become the standard method. Although early attempts were made to rely exclusively on those data types that are offered by standard programming languages and database systems , it has become common practice to identify spatial data types and to link them with their related operations. Object-oriented methods, with encapsulation and hiding of implementations, have favored this approach, and it also has made its way into relational database systems in the form of extended relational or object-relational database systems.

The third level of our review is concerned with implementation aspects, primarily the need for fast access to the necessary spatial, multidimensional data elements from linear storage devices . This topic has a long research tradition, and accounts probably for the most frequently researched topic related to GIScience. It includes some of the most frequently cited papers in the database domain—Guttman's R-tree and the R+-tree by Sellis et al. . We focus on implementations that are typically tailored to vector data models, since this has been the traditional viewpoint in much of the computational GIS domain. Complementary areas, such as image processing, remote sensing, and vision, have relied on different data models and representations, which start with a representation of the underlying space, often in the form of regular spatial subdivisions such as pixels, and aim at extracting higher-level spatial concepts from this structure.

In this section, we review some aspects that have been identified as stable foundations for modeling and implementing geographic concepts. It is by no means exhaustive.

The data management requirements of spatial applications differ substantially from those of traditional business applications. Business applications tend to have simply structured data records. There is only a small number of relationships between data items, and transactions—converting a database from one consistent state into another —are comparatively short. Relational database systems meet these requirements extremely well. Their data model is table-oriented, therefore providing a natural fit to business requirements. By means of the transaction concept, one can check integrity constraints and reject inconsistencies.

For spatial applications, however, conventional concepts from database management systems are often inadequate. Spatial databases contain multidimensional data with explicit knowledge about objects, their extents, and their locations in space. The objects are usually represented in the cartographic tradition of some vector-based format, and their relative position may be explicit or implicit. They often have a complex structure: a spatial data object may be composed of a single point or several thousands of polygons or various collections of polygons, lines, and points often with complicated consistency constraints. These objects rarely follow regular shapes or patterns of distribution across space. It is usually impossible to store collections of such objects in a single relational table with a fixed tuple size. Moreover, spatial data are dynamic: insertions and deletions are interleaved with updates, and data structures have to support this dynamic behavior without deteriorating over time. Spatial databases tend to be large, typically occupying several hundred gigabytes of storage. The seamless integration of secondary and tertiary memory is, therefore, essential for efficient processing .

Recent database research has helped to solve many related problems. These include both extensions to the relational data model and the development of flexible object-oriented approaches for spatial information .

Any serious attempt to manage spatial data in a relational database framework requires some significant extensions at the logical and the physical level. These kinds of extension need to be supported at the query language level as well. Besides an ability to deal with spatial data types and operators, this involves concepts to support the interactive working mode that is typical for many GIS applications. Pointing to objects or drawing on the screen with the mouse are typical examples of these dynamic interactions. Further extensions at the user interface level include the graphical display of query results, including legends and labels; the display of unrequested context to improve readability; and the possibility of stepwise refinement of the display (logical zooming).

For many years, the database market has been dominated by the Structured Query Language SQL. There has been a long discussion in the literature as to whether SQL is suitable for querying spatial databases. It was recognized early on that relational algebra and SQL alone are not able to provide this kind of support . "Why not SQL!" gives numerous examples of SQL's lack of expressive power and limitations of the relational model in general when applied to spatial data. At the user interface level, one encounters difficulties when trying to combine retrieval and display aspects in a single SQL query. Besides requiring specialized operators, this kind of combination usually leads to long and complex queries. The integration of selection by pointing (to the screen) is also problematic. There is no support in SQL for the stepwise refinement of queries, which is particularly important in a spatial database context where users often ask questions iteratively. The underlying problem is that SQL does not provide a notion of state maintenance that allows users to interrupt their dialogue at a given point and resume their work later.

In some sense, however, with the success of SQL the discussion about its appropriateness has become a moot point. The question is not whether SQL should be used—SQL is and in the foreseeable future will be used to query spatial databases. The question is rather which kind of extensions are desirable to optimize user friendliness and performance of the resulting spatial data management system.

Various extensions to SQL have been proposed to deal with spatial data , including PSQL , Spatial SQL , GEOQL , and the SQL-based GIS query languages for KGIS , and TIGRIS . Current efforts under the umbrella of the ANSI Committee on SQL3 are developing an integrated version of such spatial extensions, called SQL/Multimedia (SQL/MM), which is a suite of standards that specify type libraries using SQL's object-oriented facilities.

An essential weakness in traditional commercial databases is that they do not provide any spatial data types . Following their orientation towards classical business applications, they may sometimes offer non-standard types such as date and time in addition to the classical data types integer, real, character, and string. Spatial data types, however, are not included in any of today's standard commercial DBMS. On the other hand, such data types are a crucial requirement when it comes to processing geographic data.

For vector data, there have been several proposals on how to define a coherent and efficient spatial algebra . It is generally assumed that the data objects are embedded in d-dimensional Euclidean space Ed or a suitable subspace. Any point object stored in a spatial database has a unique location in the universe, defined by its d coordinates. Any point in space can be occupied by several point objects stored in the database. A (convex) d-dimensional polytope P in Ed is defined as the intersection of some finite number of closed halfspaces in Ed, such that the dimension of the smallest affine subspace containing P is d. A hyperplane H supports a polytope P if >Ø and P is completely contained in one of the halfspaces defined by H. If H is any hyperplane supporting P then is a face of P. The faces of dimension 1 are called edges; those of dimension 0 vertices. By forming the union of some finite number of polytopes Q1, …, Qn, one obtains a (d-dimensional) polyhedron Q in Ed that is not necessarily convex. Following the intuitive understanding of polyhedra, one usually requires that the Qi (i = 1, …, n) have to be connected. This also allows for polyhedra with holes.

One often uses the terms line and polyline to denote a one-dimensional polyhedron, and the terms polygon and region to denote a two-dimensional polyhedron. If, for each k (0 £ k £ d), one views the set of k-dimensional polyhedra as a data type, one obtains the common collection of spatial data types (i.e., point, line, and polygon). Combined types sometimes also occur. Curved objects can be obtained by extending these definitions.

Since there is neither a standard spatial algebra nor a standard spatial query language, there is also no consensus on a canonical set of spatial operators. Different applications use different operators, although some operators (such as intersection) are more common than others. Spatial operators can be classified in several different ways, reflecting fundamentally different perspectives and objectives. A common distinction is based on different geometric properties of spatial relations, leading to groups of topological, directional, and metric relations. Topological relations are invariant under topological transformations, such as translation, rotation, and scaling . Examples are overlap, disjoint, and inside. Direction relations refer to the location of two spatial objects with respect to a reference frame , yielding quantitative values (e.g., 44º 34') or qualitative values (e.g., north and sourthwest; or left and right). Metric relations capture distances, either quantitatively (e.g., 24.5 km) or qualitatively (e.g., near and far). For the evaluation of such spatial predicates in a database context, the spatial join operator has been introduced . A spatial join takes two sets of spatial objects as input and produces a set of pairs of spatial objects as output, such that each pair fulfills the given spatial predicate. Examples include, "Find all houses that are less than 10 km from a lake" or "Find all buildings that are located within a wetland."

A different query perspective is given if operators are classified according to their signatures , that is, the input and output behavior of each operation. In order to be considered a spatial operator, at least one of the operators has to be of a spatial data type. The input behavior refers to whether it is a unary or binary operator, as well as to the type of its operands. Operators over more than two operands are typically broken down into a sequence of binary operations. The output behavior refers to the type of result. This categorization distinguishes unary and binary spatial operators with Boolean, scalar, or spatial results. Of particular interest are set operators that compute the union, difference, or intersection of two spatial objects or sets of objects, for which Tomlin's Map Algebra provides a framework. Map overlays are an important application of set operators and a series of efficient operators have been proposed . Notorious problems regarding these operators, however, include the lack of a closure property and the handling of boundary phenomena. The efficient computation of set operators has received a lot of attention in the computational geometry literature .

The efficient computation of spatial operators requires special implementations of spatial data types. Over the years, the shortcomings of some more primitive representations have been recognized and semantically more powerful methods have been developed. The early representations in terms of vertex lists have been typically replaced by representations that better capture topological properties. A vertex list is a list of a polygon's vertices. It is sufficient for basic graphic output and well suited to support certain similarity operators, such as translation, because it corresponds to the addition of the translation vector to each of the coordinates. Problems with this particular representation, however, arise when comparing polygons because the list is not unique. For example, the same triangle could be described by the lists [(1,1), (5,1), (4,4)], [(5,1), (4,4), (1,1)], [(4,4), (1,1), (5,1)], or [(2.5,2.5), (1,1), (5,1), (4,4)]. Furthermore, there are no invariants with respect to set operations; therefore, translations, rotations, or scalings change each element of the representation. This also means that it is difficult to determine whether two vertex lists represent congruent or similar polygons. Most critical is the introduction of redundancy if two or more lines or polygons coincide in one or more points. The coincidence is only captured through the common (identical) coordinate values, which provides significant problems during updates. For instance, any consistent update of the boundary of two neighboring land parcels becomes a task of finding all vertex lists that contain a particular coordinate pair, and at the outset it is unclear how many such occurrences there are. Another data structure with similar deficiencies is the representation of a line as a 4-dimensional point (and a polygon as a list of such lines). While access methods and indices can be designed to cluster and retrieve such stored elements efficiently, the loss in the semantics is significant. The reliance on coordinates to determine identity has been found to be a particular fallacy as geometric transformations over the finite number systems of computers often violate basic assumptions about geometry .

To store the geometric data structures, most commercial (relational) databases provide long fields (also called binary large objects or memo fields) that serve as simple containers. One of the columns in the relation is declared to have variable length. The geometric representation is then stored in such a long field in a way that only the application programs can interpret, while the database system itself usually cannot decode the representation. It is, therefore, impossible to formulate or process SQL queries against that column. This long field approach complies with the OpenGIS Simple Features Specifications proposed by the Open GIS Consortium.

Abstract data types (ADTs) provide a more robust way to integrate complex types into a database system. The basic idea is to encapsulate the implementation of a data type in such a way that one can communicate with instances of the data type only through a set of well-defined operators. The internal implementation of the data type and its operators are hidden to any external users, who have no way to review or modify those interior features. Object-oriented and object-relational databases systems use the concept of the abstract data type for defining the structure of object classes. A class is a collection of objects of the same abstract data type. They thus all have the same structure and behavior as they share the same operations. Classes support two basic concepts underlying abstract data types: abstraction and encapsulation. An object can only be accessed through the operators defined on its class, that is, it is only characterized through its behavior. The user is prevented from applying unsuitable operators to the object, and its internal representation is hidden. Operators (methods) and attributes are attached to a class, which means that they are valid for all objects that belong to it. Classes may form an inheritance hierarchy. This means that all attributes and methods of a class apply to its subclasses as well, unless they are explicitly overwritten. Object-oriented concepts can easily be adapted to the implementation of spatial data types and operators .

Retrieval queries on a spatial database often require the fast execution of geometric search operations, such as point or range queries or spatial joins. Of particular concern is here the often massive spatial data sets that need to be searched. Early proposals for multidimensional data structures, such as the K-D-tree or the quadtree , focused on memory-resident data and, therefore, did not take secondary storage management explicitly into account. Despite the growing size of available RAM, GIS applications are typically disk resident as the size of the datasets is still too large to be stored entirely in RAM. Sequential search is unacceptably slow for most spatial databases; therefore, spatial search operators need special support at the physical level to guarantee good performance for spatial query processing, particularly as the size of a database grows. Traditional databases, however, lack explicit support for searching spatially. To support efficient spatial search, one needs special multidimensional access methods.

The access methods designed with secondary storage management in mind allow their operations to be closely coordinated with the operating system to ensure that overall performance is optimized. Of importance in the design of spatial access methods is the physical organization of storage devices and the goal to minimize the number of operations to secondary storage. A common assumption is that most spatial searches are I/O-bound rather than CPU-bound. Since CPU performance continues to increase at a rate much faster than disk access time, it is likely that in the future spatial access methods will depend even more on I/O. Applications with objects of complex shapes, however, may incur major CPU costs for the refinement steps necessary to filter data retrieved, thereby changing the balance with I/O .

Today's secondary storage devices are linearly structured. The main problem for the design of spatial access methods is that there exists no total ordering among multi-dimensional spatial objects that would preserve their spatial proximity. Most spatial queries and interesting geographic configurations are related to the neighborhood of a specific phenomenon; therefore, it is detrimental that there exists no mapping from a two- or higher-dimensional space onto a one-dimensional space such that any two objects that are spatially close in the higher-dimensional space are also close to each other in the one-dimensional sorted sequence. This makes the design of efficient access methods in the spatial domain much more difficult than in traditional databases, where a broad range of efficient and well-understood access methods is available.

One-dimensional access methods, such as linear hashing , extendible hashing , and the B-tree , are an important foundation for multidimensional access methods. A natural approach to handle multidimensional search queries consists in the consecutive application of such single key structures, one per dimension. This approach can be inefficient , however, since each index is traversed independently of the others without exploiting the possibly high selectivity in one dimension for narrowing down the search in the remaining dimensions. Another interesting approach is to extend hashing by using a hash function that takes a d-dimensional vector as argument. A structure based on this idea is the grid file . Unfortunately, this approach suffers from possibly superlinear directory growth.

There is a great variety of requirements that multidimensional access methods should meet, based on the properties of spatial data and their applications :

• Dynamics: As data objects are inserted and deleted from the database in any given order, access methods should continuously keep track of the changes.
• Secondary/tertiary storage management: Despite growing main memories, it is often not possible to hold the complete database in main memory. Access methods therefore need to integrate secondary and tertiary storage in a seamless manner.
• Broad range of supported operations: Access methods should not support just one particular type of operation (such as retrieval) at the expense of other tasks (such as deletion).
• Independence of the input data: Access methods should maintain their efficiency even when the input data are highly skewed. This point is especially important for data that are distributed differently along the various dimensions.
• Simplicity: Intricate access methods with special cases are often error-prone to implement and thus not sufficiently robust to be used in large-scale applications.
• Scalability: Access methods should adapt well to growth in the underlying database.
• Time efficiency: Spatial searches should be fast.
• Space efficiency: An index should be small in size compared to the size of the data set.
• Concurrency and recovery: In modern databases where multiple users concurrently update, retrieve, and insert data, access methods should provide robust techniques for transaction management without significant performance penalties.
• Minimum impact: The integration of an access method into a database system should have minimum impact on existing parts of the system.

A common approach to meet these requirements consists of a two-step process: (1) choosing an approximation (e.g., a simpler shape, such as a bounding rectangle) that can be indexed and serves as a fast filter and (2) using the original geometry to assert the retrieval condition only for the initially retrieved objects to eliminate false hits. An index may only administer the MBR (minimum bounding rectangle) of each object, together with a pointer to the description of the object's database entry (the object ID). With this design, the index only produces a set of candidate solutions. This step is therefore termed the filter step. For each element of that candidate set we have to decide whether the MBR is sufficient to decide that the actual object must indeed satisfy the search predicate. In those cases, the object can be added directly to the query result. However, there are often cases where the MBR does not prove to be sufficient. In a refinement step we then have to retrieve the exact shape information from secondary memory and test it against the predicate. If the predicate evaluates to true, the object is added to the query result as well, otherwise we have a false drop.

Spatial access methods have been among the most extensively investigated research areas in computational implementations for GIS. Details about the large variety of methods, and the often subtle differences, are given by Samet and Gaede and Günther . Research in this area has been theoretical as well as experimental, typically with a straightforward hypothesis that the new access method requires less disk accesses or simply runs faster than a subset of previously developed methods. While the differences can be measured, even for experts it has become increasingly difficult to recognize the pros and cons of each access structure, because every new method seems to claim better theoretical or empirical performance than at least one other access method that has been published previously. There is no lack of experimental and theoretical studies that analyze and compare the performance of many of the access methods; however, at present no access method has proven itself to be superior to all its competitors in whatever sense. Even if one benchmark declares one structure to be the clear winner, another benchmark may prove the same structure to be inferior. A key question is how generalizable the results are. More complexly structured data often lead to significantly different performance figures. Often also variations in distribution or density affect how suitable a particular method is. Both time and space efficiency of an access method strongly depend on the data to be processed and the queries to be answered. An access method that performs reasonably well for rectangles with the same orientation may fail for arbitrarily oriented lines. Strongly correlated data may render an otherwise fast access method irrelevant for any practical application. An index that has been optimized for point queries may be highly inefficient for arbitrary region queries. Large numbers of insertions and deletions may degrade a structure that is efficient in a more static environment. Initiatives to set up standardized testbeds for benchmarking and comparing access methods under different conditions are important steps in the right direction .

Commercial products have resorted to access methods that are easy to understand and implement. Typical examples are quadtrees in Oracle 8, SICAD, and Smallworld GIS; R-trees in the relational database system Informix; and z-ordering , which was adapted and integrated under the term HHCODE into Release 7.3 of Oracle. Performance seems to be of secondary importance for the selection, which comes as no surprise given the relatively small differences among methods in virtually all published analyses. Simple and robust methods are preferred, which can be tuned and tightly integrated with other system components.

As with any forecasting, it is difficult and risky to predict where future computational methods for representing geographic concepts will lead. We can only observe trends that are either underway or have recently started. A complementary approach is a discussion of what aspects would benefit from further research because they have not yet been developed sufficiently.

Since the early 1990s, we have seen a growing influence of cognitive considerations on the next generation of computational methods . In the past, often hardware aspects were driving the development of computational methods. In the future, the semantics of spatial information need to be addressed. A second motivation for new research in computational methods is the continuing technological push with new information technologies that need to be integrated with traditional GIS functionalities. Miniature GPS receivers, cellular phones, and other wireless communication devices will contribute to making GIS a mobile technology with the potential for the development of novel Spatial Information Appliances . They will contribute to larger spatial data collections, which will get offered through the Web, some commercially, some for free.

The need for research in computational methods for representing geographic phenomena has been recently emphasized as part of several workshops at the U.S. National Science Foundation. The reports on Critical Research Issues in Geographic Information Science and the NSF Digital Government Initiative include complementary, and at times overlapping, discussions with this viewpoint.

Most data models of today's GISs follow the cartographic tradition of organizing spatial data as points, lines, and polygons. Since this approach facilitates the creation of good quality graphics in the form of maps, it has almost become a synonym for GIS. Significant amounts of geographic information, however, do not match easily with this model and complementary alternatives need to be investigated.

Fields in GIS are characterized by values at locations, have a continuous distribution of values of a domain, an inability to be completely measured, and a need for approximation (through interpolation or functions). Since they expose distinct ontological foundations and properties , they lead to different data structures and implementations . The current imbalance between models for discrete and continuous geographic phenomena needs to be overcome by truly interoperating GISs, allowing users to perform analyses beyond the limitations of a single spatial conceptualization. While object representations have reached a level of maturity, the lack of a similar level of formalization, compatibility, and general acceptance is a major impediment. A tight integration with database systems is also necessary, and provides new challenges through the need to retrieve derived (i.e., interpolated) values rather than stored values. The research needs can be broadly structured into four domains :

• ontological perspectives of fields,
• cognitive aspects of fields,
• definitions of operations on fields, and
• formalization of fields.

'Ontology of Fields' was selected as the topic of one of the three Specialist Meetings organized as part of the Varenius project. The initiative was led by Donna Peuquet and Barry Smith, a philosopher from the University at Buffalo, and held in Bar Harbor, Maine, in June 1998.

Today's GISs capture geographic information at the geometric level in quantitative terms. One needs to know about an object's location, extent, and shape in order to record it in a GIS. It has been recognized, however, that a purely quantitative approach does not match human cognition and spatial reasoning . For example, mapping biologists' narrative descriptions of geographic locations into a Cartesian coordinate space is a struggle. The properties of such a setting are different from the traditional approach—small sets of symbols on an ordinal and nominal scale in a discrete space versus quantitative calculations in an infinitely precise, continuous space .

Representations of qualitative spatial information are needed to deal with partial information, which is particularly important for spatial applications when only incomplete data sets are available. Natural language descriptions and discourse are typical examples. Neither Cartesian coordinates nor pictorial representations are adequate. Foundations for the representation of qualitative spatial information have been developed in Artificial Intelligence with the primary focus on qualitative spatial relations and their inferences :

• inferences across different types of qualitative spatial information through interoperability across different models,
• integration of qualitative and quantitative spatial reasoning methods, and
• linkage between the semantics of natural-language terminology and models for qualitative spatial information.

Most of today's computational methods in GIScience treat geographic phenomena as static. A variety of conceptual models for time in GIS have been studied , but to date little impact has been made on commercially available tools. The linkage between space and time requires, among others, the modeling of different types of times , the incorporation of processes , and the most fundamental aspects of change . Temporal databases and temporal reasoning have a long tradition in computer science, but the integration with spatial databases (http://www.dbnet.ece.ntua.gr/~choros/) and spatial reasoning is only in its infancy. With the organization of several workshops, however, this research area has developed significant momentum.

A promising approach is the focus on a set of particular set of spatial phenomena that share the same space-time behavior. An ESP workshop focused on Life and Motion of Socio-Economic Units and recently others have placed emphasis on continuously moving, point-like objects .

Massive amounts of spatial data are being collected, either now or in the near future. New technologies will lead to ever increasing sizes of data sets, at greater levels of spatial and temporal detail. Already in place are plans for EOS, which will generate several terabytes a day of remotely-sensed imagery. In addition, the integration of GPS receivers into a large variety of spatial appliances will lead to massive records about movement of people and objects at high levels of temporal resolution. In a similar way, airborne or terrestrial video will soon become a vast data source, providing near-continuous coverage of selected activities. Such new, highly detailed, and massive data sources have the potential of enabling the scientists to perform novel types of analyses.

Such large datasets that are beyond the comprehension of a single person and computational methods are indispensable to discover new knowledge. Such knowledge may be about recurring spatial and spatio-temporal patterns, clusters, associations, outliers, or anomalies that characterize interesting situations . Effective spatial data mining methods need to be coupled with efficient algorithms that schedule the processing of very large, possibly distributed data sources .

'Discovering Geographic Knowledge in Data-Rich Environments' was selected as the topic for the third Specialist Meeting organized under the Varenius project's Panel on Computational Methods for Representing Geographic Concepts. It was organized under the leadership of Harvey Miller (University of Utah), Jiawei Han (Simon Fraser University), and John Herring (Oracle Corporation) and held at Microsoft Research Labs in Redmond, Washington in March, 1999.

Modern database technology is essential for the efficient handling of geographic data. For the necessary integration of GIS and modern database technology, there are essentially four options:

• Extension of an existing GIS with database functionalities. Most GISs in the 1980s and early 1990s were representative of this approach.
• Coupling of a GIS with a commercial DBMS. Such a coupling has been common for some time for storing the non-spatial data in a commercial relational database. Since the mid-1990s, many vendors started to store spatial data in a relational DBMS as well. Usually, the relations just serve as containers that manage the geometries as unstructured long fields.
• Extension of an existing DBMS with spatial functionalities . Recent commercial products include Illustra's DataBlades and Oracle's Cartridges, which package domain-specific extensions in modules.
• Open toolbox approaches that see a GIS just as a collection of specialized services in an electronic marketplace. While there is currently no commercial system that strictly follows this architecture, many vendors are starting to integrate similar ideas into their products.

Due to strong customer pressure, the trend towards such open GIS (fourth option) continues to increase significantly. Commercial database systems can be integrated into open architectures in a relatively simple manner. A GIS can thus gain directly from the traditional strengths of a modern database system, including an SQL query facility, persistence, transaction management, and distribution. Most importantly, however, more openness would facilitate the integration of GIS with mainstream business software. While data analysis has always been an important part of GIS, the breadth and depth of related work has increased considerably since the early 1990s. By extending their functional spectrum beyond the traditional domains of data capture, storage, and visualization, GISs are gradually moving into the mainstream of computing. Rather than providing support just for the geosciences, GIS vendors are trying to position their products as spatial data management components that should be a part of just about any information system architecture—simply because just about any information has a spatial aspect. Interfaces to business software such as Microsoft Office or SAP's R/3, and the development of spatial decision support systems are among the most visible signs of this trend. Open systems were also the focus of the first Specialist Meeting organized under the Varenius project by the Panel on Computational Methods for Representing Geographic Concepts. The meeting on 'Interoperating Geographic Information Systems' was held in Santa Barbara in December 1997, and organized by an international steering committee led by Michael Goodchild, Max Egenhofer, and Robin Fegeas (U.S. Geological Survey). The meeting immediately followed an international conference on the same topic; selected papers from the conference appear in an edited volume (Goodchild et al. 1999).

In order to achieve these ambitious goals, GIS vendors have to provide data analysis capabilities that go far beyond simple map overlays. Moreover, they have to package their modules in a manner that allows the easy integration of selected functionalities into a given business package. The new eXtended Markup Language XML may play an important role in this integration process. The resulting shift from GISs to GIServices is one of the great challenges for the next decade.

There is a need to continue to improve the foundations of computational methods, advancing them to the next level of sophistication. For example, new multi-media data types are becoming available, and we need computational methods to extract geographic content. Fresh approaches, particularly those based on cognitive considerations, should be pursued rather than making small increments to established algorithms and data structures. The most dramatic effect of such approaches will be on the user interfaces of geographic information systems. In lieu of training people, GIS user interfaces need to be made more intuitive, providing also better integration into the problem solving process. We recognize that this subfield of GIScience relies on researchers from diverse backgrounds, and close interactions are needed to make significant progress. Such interactions need to span academia, industry, and government to address the users' needs, account for technological advancements, and enable technology transfer. An important economic factor is the high demand in industry for people with knowledge in computational GIS methods.

The scope for GIS applications is broad for the future. Only through the design, development, and evaluation of complex real-world systems will we realize the full potential of computation for GIS. Some of the more important domains that should be considered for future research are spatial navigation, transportation, environmental modeling, and sales and marketing. GIS research can also benefit from and have impact on other areas of computational science. For example, advanced vision systems contribute to the understanding of satellite data. Intelligent robotics requires the storage and manipulation of geographic information. What we learn from the implementation of GISs can also be transferred to other domains, such as the analysis of molecular spatial databases.

In addition to the Varenius project funding from the National Science Foundation, Max Egenhofer's work is further supported by NSF grants IRI-9613646 and BDI-9723873, by grants from the National Imagery and Mapping Agency under grant number NMA202-97-1-1023; the Air Force Research Laboratory under grant number F30602-95-1-0042; the National Aeronautics and Space Administration under grant number COE/97-0015; Bangor Hydro-Electric Co.; by a research contract from GE Research and Development Center; and by a Massive Digital Data Systems contract sponsored by the Advanced Research and Development Committee of the Community Management Staff.

Thanks to Volker Gaede (Oracle, Germany) for his contributions to section 3.3.

Eric Sheppard

Helen Couclelis

Stephen Graham

J. W. Harrington

Harlan Onsrud

This article presents the Varenius perspective on the societal dimensions of geographic information technologies and the geographical dimensions of information technologies in general, and puts them in the context of the research literature of the last ten years. The central themes examined are: theoretical perspectives on the societal implications of geographic information technologies; the changing significance of key geographic concepts in the information age; and societal aspects of the practical application of geographic information technologies. The relationships between these themes and three NCGIA Varenius research initiatives on geographies of the information society are summarized, and some directions for future research in this broad area are outlined.

This review article introduces the theme of the third component of the Varenius project, known as the Apex, relates it to the other two major Varenius themes, and presents its constituent research areas as reflected in the three research initiatives organized by the panel. By exploring recent writings and calls for papers in this area, the article also attempts to elucidate more generally the state of research and thinking on the geographies of the information society as of 1998, as viewed from the perspective of geographic information research. Its purpose is thus threefold. First, we introduce the three Apex research initiatives held in October and November of 1998, in the context of a more general assessment of the field. In chronological order these are: (1) 'Place and Identity in an Age of Technologically Regulated Movement', (2) 'Empowerment, Marginalization, and Public Participation GIS', and (3) 'Measuring and Representing Accessibility in the Information Age'. Second, the paper will serve as a benchmark by which to assess, a few years from now, the specific contributions of the Varenius project to that increasingly vital research area. Third, we hope that this effort by a small group of researchers associated with the Varenius project will elicit a broader discussion and alternative interpretations of the issues involved, leading to further concerted research programs in the area.

The title 'geographies of the information society' must be qualified for the purposes of this essay, as it can mean many different things. It can mean the actual geographies that evolve on the surface of the earth in the information age: the changes in and among places resulting from the increased ability to store, transmit, and manipulate vast amounts of information, and the new patterns of geographic differentiation, privilege and disadvantage that these changes are bringing about (the geographic consequences of 'informationalism': see Castells 1989, 1996, Graham and Marvin 1996). Or it can mean the virtual geographies that are directly the product of the information and electronic communication technologies: the geographic study of the invisible but almost ubiquitous information networks, with their nodes, links, connectivities, and flows, along with the social, cultural, economic, and professional networks that coalesce around the electronic ones. 'Cybergeography', or the 'geography of cyberspace', currently the domain of a small though rapidly growing research community, is likely to become a mainstream area of research in the next few years (Adams and Warf 1998, Kitchin 1998, Mitchell 1995). 'Geographies of the information society' can also designate the conceptual geographies gradually constructed within individual and social consciousness through the representations of the Earth conveyed by digital geographic information technologies: this would in a sense be closer to the original meaning of geography as 'writing about the Earth'. How might we write about the Earth, in its infinite variety and interacting dimensions, in a context where the possible forms of representation are constrained by and filtered through the stringent logical and technical requirements of digital systems? What are the epistemological implications of tailoring our understanding of the Earth's complexity to that which can be recorded on electronic media, and accessed only with the requisite hardware, software, network connections, and technical know-how? The kind of knowledge that emerges within the fairly well-circumscribed universe of geographic information systems is the product of specific technological, institutional, corporate, and intellectual trajectories, and the geographies that we are able to write in that context are necessarily constrained by the limits of that particular worldview (see Sheppard 1995).

Clearly these different meanings of 'geographies of the information society', though distinct, are interconnected. They all are about places and relations among places and the individual and social lives that are an integral part of these, and they necessarily also include the circular connections between how people understand, create, and use these places and relations in the information society. This is why, rather than opt for one or the other interpretation, the Varenius Apex has taken an eclectic view, synthesizing questions about actual, virtual, and conceptual geographies. The choice of questions has been guided by two kinds of considerations. First, they must relate to geographic information science and technology, which is the focus of the Varenius project, and more specifically to the two other major Varenius themes: Cognitive Models of Geographic Space, and Computational Methods for Representing Geographic Concepts. The second criterion is more opportunistic. It considers the state of the art in geographic thinking and writing about the information society as of early 1998, the year when these questions were formulated, and seeks to identify issues that have been widely recognized as being of major import but have not yet been the subject of substantial research efforts elsewhere.

The paper is organized as follows. We first outline briefly the Varenius project and the three Apex research initiatives within it. We then review three major research areas within the wider theme of the geographies of the information society, each of which has a special connection with one of the three Apex initiatives: (a) theoretical perspectives on the societal implications of geographic information technologies; (b) changes in the meanings of key geographic concepts induced by the information age, and the empirical correlates of such changes; and (c) the relationship between the practical application of geographic information technologies and the societal context within which this is occurring. Our review is necessarily eclectic, focusing on aspects of these broader themes that can be directly linked to the three Varenius initiatives. We close by sketching a tentative road map for future research in the geographies of the information society.

The question of how the geographic world is understood by humans and represented in machines is at the core of the Varenius research agenda. The thematic area entitled Cognitive Models of Geographic Space investigates how people understand geographic entities and relations, how they reason and talk about these, in what respects these cognitive models are similar or dissimilar with the formal representations of these entities and relations in geographic information technologies, and how can the latter be better in tune with the former. The Computational Methods for Representing Geographic Concepts area views similar questions of geographic concept definition and adequacy from the perspective of computer languages, data models, visualizations, and interoperating systems. The Apex thematic area is in many ways the societal counterpart to the individual cognition stressed in Cognitive Models of Geographic Space. Here it is the social origin and evolution of geographic concepts that is highlighted, along with the societal issues arising from the changing patterns of opportunity, privilege, power, disadvantage, discrimination, equity, or liberty that these changing concepts reflect. The three Apex initiatives bring the geographic information science perspective to bear on these issues, investigating the role of geographic information science and technologies in helping both to construct and to study these evolving information-age geographies. Moreover, an increasing number of social theory scholars suggest that, before the combined commercial, government, and academic enterprise pursues the next generation of geographic information technologies, we might well try to identify the social bases and limitations of the current generation (Curry 1997a, Goss 1995a, Pickles 1995b, Sheppard 1995). How have the current models of geographic space in our systems been developed? What cognitive and social understandings of geographic space may have been left out? How have these lacunae affected the utility of the systems for different groups, or for different purposes? How might they be filled? The three Apex initiatives are designed to address selected aspects of these broad and profound questions, along with some more technical but equally important ones dealing with how we can understand, measure and represent the evolving new geographies.

The purposes and foci of the three Initiatives are best seen in the original calls for papers (see www.ncgia.ucsb.edu/varenius/initiatives/ncgia.html). The rationale for each is summarized in the following excerpts from the calls.

This initiative investigates geographical aspects of the interplay between information technologies and society, in particular the dissolution of traditional territorial identities and the formation of new ones around new kinds of communities and places held together largely by electronic connections:

This initiative investigates the effects of information and communication technologies on accessibility, the resulting changing meanings of that fundamental geographic concept, the societal and geographic implications of these changes, and the role of geographic information technologies in both bringing about and helping study these changes:

This initiative approaches the GIS and Society theme from a more applied perspective, focusing on the potential role of GIS as a tool (or, on occasion, liability) in the democratic process of grassroots community self-determination. This initiative was developed to attract significant numbers of community-level GIS practitioners as well as academics:

Each of these initiatives fits within a wider research area reaching far beyond the Varenius effort, namely: (a) theoretical perspectives on the societal context and implications of geographic information technologies, and on the geographic implications of information technologies in general; (b) the study of changes in the meanings of key geographic concepts induced by the information age, and the empirical implications of these changes; and (c) issues of democracy and individual rights arising from the practical application of geographic information technologies. In the following three sections we view each of the three Apex initiatives against the background of past research in these three broad areas.

Research examining the two-way relationship between geographic information technologies and the societal context within which they are being developed and applied has undergone an explosion in this decade, triggered by a debate between social theorists and GIS specialists in the early 1990s. Initially this debate was highly antagonistic. An acrimonious exchange between Taylor (with Overton) and Openshaw pitted a social theoretic perspective critical of the positivist and reductionist nature of GIS against a visionary perspective of GIS as liberating and unifying geography (in chronological order: Taylor 1990, Taylor 1991, Openshaw 1991, Taylor and Overton 1991, Openshaw 1992, Taylor and Overton 1992, Openshaw 1993, Dobson 1993). This abstract discussion of philosophical and methodological limitations and possibilities was followed by more empirical arguments, highlighting a number of negative social impacts associated with the spread of GIS applications (Lake 1993, Pickles 1991, Smith 1992). As a consequence, by 1993 there was little constructive communication between GIS researchers and practitioners on the one hand, and social theoretic critics of the GIS research agenda on the other. Two 'cultures of indifference' focusing on the same technology had emerged (Pickles 1998).

This began to change at a workshop on Geographic Information and Society organized by the NCGIA in November 1993 in Friday Harbor, WA, shortly before the publication of Ground Truth, a landmark book on that general theme edited by John Pickles (1995a). The meeting brought together representatives of both viewpoints and evolved into a constructive engagement between the two. Publication of a special issue of the journal Cartography and GIS (Poiker and Sheppard 1995) enriched the discourse on these questions, giving proponents from both sides a chance to air their different perspectives constructively in print. It is interesting to compare that issue with the Pickles volume for a sense of how debates shifted as a result of the meeting. A further major outcome was the NCGIA Initiative 19, entitled 'GIS and Society: The Societal Implications of How People, Space, and Environment are Represented in GIS' that was launched by a group of Friday Harbor meeting participants with the purpose of laying out an agenda for research in the area of GIS and society.

The specialist meeting for Initiative 19 took place in Minnesota in March 1995 and drew equally and in a very constructive atmosphere on the ideas and perspectives of both social critics and GIS researchers (Harris and Weiner 1996, Pickles 1998). Participants at that meeting identified the following themes as worthy of further research: limits of representation in GIS; the societal impact of GIS use; a critical history of GIS; ethics, privacy and GIS; alternative GIS, or GIS2; the use of GIS in debates about global change; and gender and GIS. Some of these themes have since been pursued systematically by Initiative 19 participants and others.

By 1998 Michael Curry could assert that it was time to move beyond the GIS and society debate (Curry 1998). This does not mean that a common perspective has emerged, nor that it should emerge; tension between research perspectives is healthy for the development of any research program. Some continue to claim that there is still too little reflection within the GIS community on societal context and the social implications of the technology, whereas others argue that these problems have been over-rated or can be addressed within the domain of ongoing GIS research. Yet things have changed a great deal since the debates of the early 1990s. Both sides now recognize that a number of GIS researchers have a track record of concern for the social implications of geographic information technologies, and that such concerns are becoming more widely shared within the GIS community. It is equally clear that critics often work from a sophisticated understanding of geographic information technologies, and that the debate no longer is between those who only understand GIS and those who only understand social theory. New bi-partisan networks of collaboration have resulted just as we witness the emergence of new debates taking shape along new, less well-defined boundaries. These are as much within as between the two groups that originally constructed themselves as being in polar opposition to one another (Chrisman 1987, Couclelis in press, Curry 1995, Edney 1991, Flowerdew 1998, Harris and Weiner 1998, Miller 1995, Obermeyer 1995, 1998, Pickles 1997, Sui 1994, Wright et al. 1997).

Theoretical perspectives on geographic information technologies inform all three Apex initiatives, and remain an active area of research outside the themes focused on by Varenius. For example, Nyerges and Jankowski (1997) examine how collaborative decision-making with GIS can be conceptualized through the lens of structuration theory, and Harvey and Chrisman (1998) are actively exploring how contemporary theorizations of the practice of science provide insight into the two-way relationship between society and geographic information technologies. There is an extensive literature outside geography addressing these issues. Yet the initiative most directly related to the theoretical issues underlying the Initiative 19 research tradition is that entitled 'Place and Identity in an Age of Technologically Regulated Movement'. This initiative seeks to unravel the societal implications of both the vastly increased personal mobility and the increasing use of information technologies, and in particular geographic information technologies, to monitor and record the movements of people from place to place and at any time. It addresses profound questions regarding the ongoing detachment of community, identity and territory from contiguous spatially bounded places. At the same time, questions of privacy, surveillance, and democracy loom large. The initiative addresses these issues in the context of the following more specific questions, spelled out in the call for papers:

• What have been the traditional means for the regulation of borders? In what ways have they been successful in promoting territorially based identities?
• How has the development of modern communications and especially geographical technologies altered the regulation of flows of people, goods, and information?
• To what extent has the regulation of borders at various scales—from neighborhood to nation state and beyond—moved away from geographical borders, and been replaced by ubiquitous forms of control?
• How are these various regulatory regimes related to personal and group identity?
• How have alternative, non-place-based identities been promoted and maintained? How have they been controlled, and how successful have these controls been?
• What lessons relevant to the world of the Internet can be learned from these experiences?
• What future is there for borders and boundaries in a world where 'there is no there'?

The meanings of geographic concepts, as with all concepts, keep evolving under the combined effects of a changing empirical world and changing societal modes of thinking about and acting in that world. Geographic information technologies contribute to the ongoing redefinition of geographic meanings on both these fronts, by affecting both the tangible urban and regional structures that we study as well as our ways of dealing—conceptually, politically, or practically—with these new geographies. Moreover, geographic information technologies have the peculiar distinction of both contributing to changes in the meanings of geographic concepts, and of attempting to provide suitable and sufficiently robust representations for these meanings. This section considers that double aspect of geographic concept change, and outlines how the Varenius project and its immediate precursors have been addressing that issue.

The implications of the diffusion of geographic information technologies for how the geographic world is represented and understood were a major theme in Initiative 19. Geographic information systems are the most successful technologies in history for representing, manipulating, analyzing, and storing knowledge about the geographic world. At the same time, their power to privilege certain kinds of representations and to generate their own variety of geographic meanings cannot be overestimated. Considerable critique has focused on the ontological and epistemological assumptions underlying GIS, in particular the Euclidean, Cartesian, and positivist conceptions of geographic space on which it is built (Pickles 1995b, Sheppard 1995). According to that critique, GISs tend to embed a powerful and particular ontology of space deep into the practices that surround their application. As a consequence, many key geographic concepts are being implicitly redefined by both GIS developers and users to fit the constraints of that positivist, Cartesian ontology. Take for example the fundamental concept of place. Critics point out that the ontology of GIS sees the geographic world as being subdivided into spatial units that can be represented as sharply defined, contiguous, non-overlapping polygons. Captured or processed data become the reduced, sole signifier of such units—towns, neighborhoods, census tracts—and at the same time their representation and simulation. The signs become coded as the signified, and place is reduced to a geometric object (Shields 1995). But place is not some unitary Euclidean phenomenon—some well-circumscribed spatial unit that can be represented unproblematically as a shape on a computer screen. Rather, as critical theorists point out, it is a dynamic, relational phenomenon made of multiple, superimposed space–times, which jostle and compete through the subjectivities, discourses, and representations of the social world. As Massey (1993: 66) puts it, places need to be defined in dynamic and relational rather than static and geometric terms; as 'articulated moments in networks of social relations' rather than as 'areas with boundaries around'. Stressing the constant 'becoming' of places, Thrift (1996: 1485) writes: 'There is [...] no big picture of the modern city, but only a set of constantly evolving sketches'. Nothing could be further from the ontology of the Cartesian grid. The above-mentioned Varenius initiative on place and identity addresses some of these issues.

According to the critics, similar problems hold for other aspects of geographic space that geographic information technologies also reduce to geometric configurations (primarily, points, polygons, networks, and bounded regions), as only these can satisfactorily be represented through the (digital) Cartesian gaze (Virilio 1991). These artifacts provide rigid categorizations through which meanings about the geographic world are constructed and communicated: one space becomes 'urban' and 'housing', the next 'green belt' or 'environmentally-sensitive area', the other an 'area of social deprivation', and so on. These definitions then go on to influence social action, spatial practices, and, in turn, the on-going production of space. Ontology thus shapes representation and linguistic construction in what Bibby and Campbell (1998) call a process of 'representational stabilization'—the production of taken-for-granted boundaries and differentiated categorizations about spaces, characteristics, and boundaries across the Euclidean plane that go on to shape the application and impacts of GISs in practice. For example, the use of geodemographic databases to support GIS-based informal redlining by banks, insurance firms, or retailers, may aggravate spatial discrimination which then further undermines the socio-economic fortunes of an area, leading recursively to further redlining and an accelerating slide down the spiral of decay. As Bibby and Campbell (1998: 9) suggest, the ways in which socio-spatial entities are configured or represented in terms of geometrical boundaries and polygons allows them to be reduced to essential, 'ordinary things'. This closes meaning, 'reifies spatial definitions', and generates a spurious integrity for categorizations and generalizations about the social world (see Shields 1995). Moreover, as Pickles (1995b) and Curry (1994) have argued, GIS researchers have all too often understood the dynamics of this process of representation and meaning creation as a purely technical one, expecting that the honing of algorithms and the availability of faster computers will eventually lead to the perfect GIS-based representational tool. This is what Pickles (1991: 80) has called a 'mythos of liberating technology'.

These are indeed serious issues that have not gone unnoticed by the GIS community. Since the days of the Friday Harbor meeting and Initiative 19 an increasing number of GIS researchers and practitioners have critically reflected upon and analyzed the fundamental epistemological principles and geographic conceptualizations that provide the essential underpinning to their work. Crisp polygons are no longer seen as unproblematic representations of places and phenomena (Burrough and Frank 1995). Socio-economic units of data collection or administration are no longer understood as inert geometrical configurations. The multiple levels of meaning of the phrase 'GIS and society', from the most mundane and practical to the ontological, are being examined (Couclelis in press, Nyerges 1991). Finally, as mentioned earlier, a large part of the purpose of the Varenius theme of Cognitive Models of Geographic Space is to take a close critical look at the discrepancies between human and mechanical representations of geographic concepts.

However, changes in the meanings of geographic concepts do not only—or even primarily—occur through societal re-interpretations and re-constructions. Geographical concepts also change because the corresponding entities and structures change in the empirical world around us. For example, the meaning of distance is profoundly affected by the socio-economic, institutional, and technological developments of the information age (Cairncross 1997, Couclelis 1996a, Virilio 1995). These not only make physical travel faster, easier, and more of an integral part of everyday life than at any time before; they also, in a sense, annihilate distance by rendering instantly accessible, from any suitably equipped location, information originating at places that may be arbitrarily remote in geographic space. The shifting emphasis from distance to connection as the key explanatory variable of socio-economic structures in geographic space reflects contemporary trends towards a 'network society' (Castells 1996). The term implies that the relational interweaving of connections and disconnections within and between places, based largely on information technologies (IT), telecommunications, and fast transport grids, is reaching unprecedented depth and intensity (Amin and Graham 1997, 1999, Castells 1996). Physical proximity no longer signifies meaningful connection, as is usually implied by the 'first law of geography'—namely, that nearby things tend to be more connected. Far-distant nodes, spaces, and places (airports, sea ports, elite spaces, corporate locations, back offices, financial trading areas, Internet sites, 'cyberspaces', and media flows) can be drawn together into intimate exchange with each other across the planet, while being relationally severed from physically adjacent barrios, ghettos, backcountries, and other marginalized areas (Boyer 1996, Graham 1998). Instantaneous, trans-global switchings of billions of dollars between financial centers, as well as the international airline flight paths, pass over, through, within and between places where mobility and access of every kind are highly limited and circumscribed.

The Varenius Initiative on Measuring and Representing Accessibility in the Information Age is designed explicitly to address issues arising from the changing meanings of distance, connection, and access in contemporary society, as spelled out in the call for papers. In contrast to the place and identity initiative, which takes a broad theoretical perspective on the changing significance of that other fundamental geographic concept, place, the accessibility initiative focuses on the more technical and measurable aspects of such changes. As in the case of place and identity and Initiative 19, there was a precursor NCGIA-sponsored meeting to the accessibility initiative that was held in Baltimore, MD in September 1996. The meeting, entitled 'Spatial Technologies, Geographic Information, and the City', identified accessibility as a pivotal concept in understanding the far-reaching changes in urban organization and society brought about by the information age (see Couclelis 1996b). Some of the questions addressed by the accessibility initaitve are as follows:

• What are the information age counterparts to the accessibility and potential surfaces developed for interaction in physical space?
• What space–time topologies need to be developed to accommodate both the physical and virtual worlds?
• How do emerging conceptions of virtual space map onto traditional conceptions of geographic space and how do we handle their interfaces analytically?
• How can interactions and accessibility gradients within these new hybrid spaces (and space–times) be represented and visualized within GIS?
• How useful are traditional spatial interaction and urban computable general equilibrium models for the analysis of the new forms of accessibility? How should they be altered?
• What are the technical and societal impediments to network access in different social domains, particularly for geographic information?
• What representations can highlight patterns of lack of access independently of the lack of interaction?

Next to the more theoretical investigations, an extensive body of research has examined the adoption and utilization of geographic information technologies. Questions examined include the human, organizational, and institutional aspects of GIS implementation (Azad 1993, Azad and Wiggins 1993, Budic 1993, Budic and Godschalk 1994, Croswell 1991, Pinto and Azad 1994); the assessment of the institutional and societal value of geographic information and geographic information technologies (Dickinson 1989, Dickinson and Calkins 1988, Donelan 1993, Epstein and Duchesneau 1990, Gillespie 1991, 1992, Lopez 1996a, b, Smith and Tomlinson 1992, Steeger 1991); the evaluation of GIS adoption and management by organizations (Budic 1994, Campbell 1994, Obermeyer 1990, Obermeyer and Pinto 1994, Onsrud and Pinto 1993, Pinto 1994); the diffusion of geographic information innovations (Grimshaw 1994, Masser and Onsrud 1993, Onsrud and Pinto 1991); the sharing of geographic information among organizations (Onsrud and Rushton 1995); and the design of new kinds of geographic information systems based on social and cultural goals (Chrisman 1987). Much less has been documented about the extensive growth in the development and use diffusion of GIS by military and espionage institutions (but see Cullis 1995).

While this body of research is documenting the diffusion of current GIS technology and its utility for agencies and institutions, it has paid only a limited degree of attention to examining the societal context influencing the nature of GIS practices, or to the broader societal implications. Nicholas Chrisman, together with Francis Harvey, has explored the question of the nature of the practices that result from GIS adoption (Harvey 1997, Harvey and Chrisman 1998). Rather than focusing on GIS as a fixed set of technical tools with presumably beneficial impacts from appropriate adoption, they argue that the practice of GIS depends not only on the technology used but also on the cultural and institutional context within which it is implemented. Drawing on comparative case study analysis of GIS use in different national cultures, and in different institutional contexts in the US, and on the writings of Bruno Latour (1987) about the practice of science, they argue that the technology (both GIS and geographic databases) and the context co-evolve. Thus, rather than conceiving of GIS as a fixed entity diffusing through social institutions, it is conceived of as an evolutionary practice; an emergent property of the interdependence between technology and societal context.

Other researchers have focused on the societal implications stemming from the kinds of representations of the world that current geographic information technologies privilege (Sheppard 1995). They observe that whereas GIS may be beneficial for public and private agencies and institutions, it remains vital to interrogate the impact of GIS practices on the remainder of society. In this view, even democratic political processes do not mean that state agencies are accountable to, or act in the interest of, all social groups; and neither do market processes guarantee that profit-seeking firms promote general economic and social welfare. The databases and analytical capacities that state agencies derive from using GIS have two kinds of potentially negative consequences. On the one hand, GIS enhances the capacity of state agencies to increase their surveillance of society, with negative consequences for those social groups whose interests conflict with those of state agencies and their representatives (Clark 1994, Clarke 1998, Gandy 1989, Pickles 1991, Smith 1992). On the other hand, in conflicts between state agencies and civic society over land use and other locational issues, privileged access to GIS and spatial databases can enhance the power of those agencies and thus reduce the effectiveness of democratic processes to influence state policy (Aitkin and Michel 1995, Archer and Croswell 1989, Edney 1991, Lake 1993, Miller 1995, Yapa 1991). Similarly, analysis of the use of GIS by private firms for geodemographic marketing has shown that sub-populations and neighborhoods are characterized by consumer profiles that homogenize the treatment of those groups and places, in terms of both the products marketed to them and the putative impact of targeted marketing on consumer habits and neighborhood characteristics (Curry 1997a, b, Goss 1995a, b). Provocative hypotheses about the concrete impact of such processes on the homogeneity of and differences between neighborhoods and their residents, stemming from these observations, remain to be investigated.

The tendency privileging access to GIS in public and private institutions, relative to their availability to the general public and to civic and grassroots organizations, is opposed by a counter-tendency towards smaller and cheaper systems that can be installed and used by the general public. This counter-tendency has the capacity to enhance the power of the organizations and movements of civic society in democratic processes, and thereby increase their potential influence over public agencies and private institutions (Barndt and Craig 1994, Craig and Elwood 1998, Dangermond 1988, Edney 1991, Elwood and Leitner 1998, Ferber 1992, Schmitt 1997, Shiffer 1998, Van Der Meuelen and Lively 1994, Yapa 1989). In this context it has been noted, for example, that GIS is being used more frequently by Indian tribes (Bird 1995, Brown et al. 1995, Jarvis and Spearman 1995, Kemp and Brooke 1995, Marozas 1993, Nietschmann 1995). In addition, the US Department of Housing and Urban Development (HUD) is subsidizing and promoting the adoption of GIS by neighborhood groups (Kingsley et al. 1997, US HUD 1997). The adoption of GIS by such organizations raises further questions about social consequences, however. There are numerous institutional and political barriers to the successful use of GIS by community and grassroots groups (Barndt 1998, Clark 1998, Harris and Weiner 1998, Hutchinson and Toledano 1993, Yapa 1991). Elwood and Leitner (1998) show that community organizations have very different capacities to take advantage of GIS technology, depending on their awareness of technology; their financial resources; the networks of which they are a part; the availability of appropriate expertise; and their ability to integrate GIS use into organizational practices. In addition, community and grassroots organizations are also not an automatic panacea for deleterious state and private sector actions. Elwood and Leitner (1998) find that neighborhood organizations seem willing to employ GIS for surveillance purposes, recreating at a local scale some of the negative consequences of GIS use. Community and neighborhood organizations are engaged in their own power struggles, within the organization, with other such organizations, and with more formal agencies and institutions, and the outcomes of those struggles do not necessarily benefit all those who such organizations purport to represent.

Negative social consequences may also emerge from inequalities in access to spatial data. Data are the raw material to which intellect may be applied in order to create and to learn, from which answers may be sought, and from which new intellectual works may arise. Without access to data, geographic information technologies cannot be applied to the exploration of real-world physical or social phenomena. In the past, government has maintained a heavy role in collecting geographic digital data and characteristics of locations. It also has organized massive amounts of other government-collected information through geographic indicators. The dominance of government in the collection of geographic data gives rise to the concern that the state might unintentionally or purposely control the ability of citizens to ask questions. If data do not exist or access to collected data is denied, unlimited access to hardware, software, and expertise will be of little benefit to a citizen inquirer. The potential trend towards diminishment or loss of the current US public commons in geographic information is an important aspect of these issues. Government agencies at all levels in the US are moving towards contracting with private firms for the collection and maintenance of geographic databases whenever possible rather than using government personnel. As a result, in many instances, private ownership interests are now attached to many of the geographic data sets and products used by government agencies in their decision-making. Such databases become far less available for general use by citizens. Corporate firms with vested interests in large accumulations of factual information are pressing for new laws that would grant them ownership interests in compilations of information that are far more extensive than provided under traditional copyright law concepts (e.g., H.R. 2652 1998). As businesses and other societal institutions move to offering data and information primarily in digital formats, the concept of fair use is dwindling and may be eliminated altogether.

It would be a mistake to rely primarily on the marketplace to define the relationships between public access and private rights in geographic data sets in future digital library environments, and yet the prevailing response of the research community to date when confronted with these issues has been to claim that 'the marketplace will take care of it'. The research and scientific community is only just beginning to recognize the potential ramifications that loss of the public commons in geographic data might have and has yet to begin concerted research to explore this arena. 'Without equitable access to GIS data and the technology, small users, local governments, nonprofit community agencies, and non-mainstream groups are significantly disadvantaged in their capacity to engage in the decision-making process' (Harris et al. 1995: 203).

The use of GIS requires heavy dependence on secondary data sources. Means and methods for maintaining and extending access to secondary sources is a major challenge for the research community. Similarly, exploration of approaches and methods for collecting and merging local data and local knowledge with other forms of knowledge within geographic information processing and analysis environments remains a substantial research challenge.

Several areas of law are substantially influencing the development of geographic information systems and, in turn, controversies over geographic information and technologies are helping to form information policies and laws at local to national levels. In 1994, the National Center for Geographic Information and Analysis began a research initiative on Law, Information Policy, and Spatial Databases. The overall goals of this initiative were to advance scientific understanding of the law and information policy within spatial database environments; raise the quality and content of the debate about law and GIS by identifying issues in concrete terms with a high degree of specificity; report observations of the law in action in order to explore the impact of spatial databases on public information policy and law, and, conversely, report observations of the law's acceptance of GIS uses and practices; identify emerging problems at the interface of law, information policy, and spatial databases in order to address those problems prospectively, with particular focus on legal issues relevant to the National Spatial Data Infrastructure; and divulge, test, and contribute knowledge useful in the improvement of public policy and formulation of law with respect to the use of spatial databases and related technologies.

Progress has been made by researchers in both the US and Europe in pursuit of all of these goals over the past several years. However, the exploration of any one issue and its resolution in a particular context typically gives rise to numerous additional knowledge gaps and unanswered questions. For instance, even if a workable legal and institutional model is developed and implemented for balancing personal privacy protection and public access interests in spatial data in a single state or jurisdiction, differences in laws and social conditions may make the model inappropriate for application in other jurisdictions. Advancements in technology may also make a workable model today inappropriate tomorrow. Although the challenges are daunting, the academic research community fulfills an extremely important societal role by continually questioning the logic and validity of arguments presented by the various parties in information policy debates. It often falls to the research community to document through observations the truth or falsity of claims and to collect evidence on the actual ramifications of following one information policy or legal approach over another. The research community is particularly suited to this task since the work of individual researchers is subjected to peer review, and full disclosure of research methods and data is the expected norm. In addition, academic researchers often have the ability to identify affected interests not represented in social policy debates and find ways of exposing, articulating, and interjecting the interests of those who may be disfranchised.

Focus areas for investigation of the law and information policy in action have included intellectual property rights in spatial databases (Lopez 1993, Onsrud and Lopez 1997), access rights of citizens to publicly held information (Lawrence 1990, 1993, Lopez 1994, Onsrud 1998b, Onsrud et al. 1996, Perritt 1995a, Rhind 1992), privacy rights and principles (Onsrud et al. 1994), liability in the use and distribution of geographic information system data and products (Anderson and Stewart 1995, Johnson and Danby 1995, Onsrud 1998a, Perritt 1995b), and ethical issues in the use of geographic information (Onsrud 1995).

The breadth and extent of the legal, ethical, and policy issues that geographic information technologies and data sets give rise to is enormous. The challenge for the research community is to continually revisit the issues in order to critique existing social and institutional models, and develop new models that may better satisfy and benefit diverse stakeholders in society.

A further group of researchers have begun to raise the question of whether the GIS systems and databases developed are appropriate for neighborhood and community organizations. In this view, their development as useful tools for public agencies and private firms does not guarantee that they are the appropriate technology for other potential users. Sheppard (1995) has argued that different social groups employ different 'ways of knowing' (different ways of reasoning about and making sense of the world) which need not be consistent with those underlying conventional databases and GIS software. He suggests that an important area of research is to determine the degree to which this is the case, and its implications. Dan Weiner, Trevor Harris, and co-researchers note, for example, that much of the local knowledge about the history of land ownership in South Africa is not recorded in standard property files, but resides in the oral histories of local peasants who have been progressively denied access to land that they once had a right to. Through development of the Kiepersol GIS, they have innovatively combined conventional GIS mapping routines with overlays containing this local knowledge. This knowledge is in a form quite different from that of conventional databases: it is qualitative, taking the form of narratives rather than statistics, and is partial in its coverage and even contested by different informants. Yet they have demonstrated that it can be combined visually with conventional data, providing a different perspective on land ownership rights than would otherwise be available (Harris et al. 1995, Weiner et al. 1995).

In a similar vein, Yapa (1998) has argued that use of a conventional layer-based GIS to analyze poverty results in a particular spatial representation, a mapping of poverty that treats poverty as located in the actions of poor people, rather than in broader societal causes that marginalize certain social groups. He advocates object-oriented GIS as an alternative that generates a more systemic representation of the causes of poverty. A more skeptical assessment of the possibilities of extending the kinds of representation possible with GIS is provided by Robert Rundstrom's analysis of American Indian conceptions of space and place, and their compatibility with GIS (Rundstrom 1991, 1995). He argues that Indian understandings of space are deeply rooted in local context, and cannot be abstracted into the kinds of general principles about topology and geographical data that lies at the root of GIS software and databases. This incompatibility raises significant questions, he suggests, about the ability of even Indian organizations to employ GIS in ways that are consistent with the views and understandings of Indian cultures.

The difficulties associated with incorporating local knowledge into GIS have been highlighted in research using GIS for the assessment of environmental justice, and risk assessment more generally. On the one hand, careful application of GIS to publicly available Toxic Release Inventory (TRI) and census data has been able to detail the closeness of different social groups to places releasing toxic chemicals. Work carried out at a variety of scales and resolutions, using different measures of proximity, and different definitions of the risk of exposure, have shown the complexity of the relationship between the location of noxious facilities and that of low income, minority, and dependent populations. In some cases clear inequities to the disadvantage of such groups are documented; in other cases, higher income populations are closer to TRI sites (Bowen et al. 1995, Burke 1993, Chakraborty and Armstrong 1997, Cutter et al. 1996, Glickman 1994, McMaster 1990, McMaster et al. 1997, Nyerges et al. 1997b, Scott et al. 1997, Scott and Cutter 1997, Sui and Giardino 1995). On the other hand, the causal processes behind such patterns cannot be inferred from GIS analysis. Patterns do not reveal process, and understanding process requires detailed and intensive examination of the local historical and geographical context (Bryant 1995, Pulido 1996, Pulido et al. 1996, Sheppard et al. 1998).

Making such GIS capacity available to local neighborhood or environmental groups concerned with state proposals to locate a garbage transfer site or an incinerator in their neighborhood, with developing a good neighbor agreement with a local firm, or with generally documenting and improving the local physical environment, exposes limitations to conventional use of GIS in novel contexts. First, many of the environmental problems of local concern are not documented in the standard databases used for nation-, state-, or city-wide GIS analysis; local-scale data collection creates the possibility of a much more comprehensive analysis of local public health risks in the environment than is possible in larger-scale analyses. Second, local knowledge about environmental problems may be anecdotal and in narrative form, requiring careful further investigation and novel methods of analysis that may or may not be compatible with those available in GIS. Third, local priorities for environmental improvement may again challenge the capacity of GIS as the appropriate technology for addressing them. Finally, provision of GIS to neighborhood organizations is fraught with the kinds of difficulties of implementation, and unforeseen local social impacts, described above. In response to these problems, neighborhood involvement in the development and completion of environmental inventories, in the implementation and design of databases, and in discussions about the appropriateness of GIS-based analyses for addressing neighborhood concerns is important. More generally, this addresses the question of developing public participation GIS.

During the specialist meeting for NCGIA's Research Initiative 19 on GIS and Society the question of 'what could GIS be?' was continually raised in the context of GIS that would be more responsive to the needs of broader segments of society and in different ways. At the meeting a small working group developed an initial set of criteria for what was then called GIS2. The criteria developed by this group included the following:

1. A GIS2 would increase emphasis on the role of participants in creation and evaluation of data.
2. A GIS2 would accommodate an equitable representation of diverse views, preserving contradictions, inconsistencies, and disputes against premature resolution.
3. System outputs would be redefined to reflect the standards and goals of the participants.
4. A GIS2 would be capable of managing and integrating all data components and participant contributions from one interface. Components would include e-mail, the Web, access to data archives, presentation of parallel texts and counter texts in diverse media, real-time data analysis, standard base maps and data sets, sketch map capabilities, and field note capabilities.
5. A GIS2 would preserve and represent the history of its own development and be more capable of handing time components than existing GIS.

This initial formulation of GIS2 thus envisioned new technological capabilities that could have a much greater ability than current ones in enabling the process of communication. Technologies and processes to be embedded in his new geographic information systems environment would allow diverse members of communities to explore and interact with each other in manners allowing them to improve their own community conditions and relationships. This new information systems environment would allow enhanced participation by all groups or individuals with interests in the outcomes of disputes.

Because the term GIS2 appeared to imply a straightforward extension of existing current geographic information systems (i.e., GIS1), a term that would describe a systems environment more focused on communication, process, and participation among many interested parties was sought to describe this new domain of research and development. No shorthand title is sufficient to describe the complete body of concepts that individuals wanted to embed in this new way of looking at geographic information technologies and their relationships to individuals and communities, but the ensuing discussions in search of an appropriate term raised a series of interesting issues.

In preparation for a follow-up workshop at the University of Maine to further develop and explore the concept of GIS2 (NCGIA 1996), one of the many terms suggested to better describe the set of concepts being pursued was 'Public Participation GIS'. Xavier Lopez suggested the term because a well-developed literature already existed relating to public participation in decision making in the planning community, and it appeared that a merging of the concepts embedded within that literature stream with the technical and social concepts being explored in the geographic information science community came close to encompassing the intended research domain.

One critique of the term Public Participation GIS (PPGIS) was that public participation seems to imply the development of concepts and systems directed primarily at meeting the needs of and enabling grass roots, community, and marginalized groups. An alternative suggested was to pursue concepts and capabilities that would serve the interests of all of those interested in the outcomes of publicly debated decisions, including not only grassroots and community groups but also government and commercial sector groups. In response to this concern, the term 'Public Forum GIS' was proposed. Although this term might suggest openness and opportunity for all interests to be heard, it seems to imply that final decision-making is primarily and foremost in the control of government decision-makers and thus the term suggested a less active role for others. PPGIS is a more active term. In addition, public participation as conventionally understood within the planning community does indeed imply identification of all interested groups, and involves invitations to all such groups to participate in consensus building processes. If the term PPGIS implies to some people a greater focus on meeting the specific needs of marginalized groups this meaning was acceptable as well, because the academic community is in a far better position than the marketplace or government to focus on developing tools, techniques, and processes for these groups. Thus the ambiguity in the term PPGIS was deemed desirable.

A focus on communities and the integration of diverse interests in communities seems a more accurate description to some people of the concepts to be encompassed by the research domain and therefore the term 'Community-Integrated GIS', or 'Community-Integrative GIS' was preferred by these people. For others, the term 'Participatory GIS' was more appropriate. Regardless, the title for the domain of interest is far less important than the concepts and ideas within it, to which a large number of people with diverse perspectives and backgrounds are being drawn.

In summary, PPGIS research seeks an adaptable suite of analytic tools, communication technologies, and participatory processes that will expand the roles of communities in defining questions in which location or geography have a bearing on the issues addressed; increase public participation in data creation and evaluation; increase opportunities for community self-exploration and self-improvement; increase the breadth and depth of participation in decisions of broad public interest; and enable wider public acceptance of the results of decision making in which place or space play a significant role (Schroeder 1997). An extensive literature on PPGIS is now beginning to emerge. Recent examples include articles in the Proceedings of the 1997 ACSM/ASPRS Annual Convention, Volume 5 (e.g., Neyrges et al. 1997a); the Proceedings of Auto-Carto 13; the Proceedings of the 1997 UCGIS Annual Assembly and Summer Retreat, http://www.spatial.maine.edu/ucgis/assembly_schedule.html; and a special issue of Cartography and Geographic Information Systems on Public Participation GIS, Volume 25, No. 2 (Barndt 1998, Obermeyer 1998).

The Varenius initiative on Empowerment, Marginalization, and Public Participation GIS addresses a broad range of issues related to the relevance of GIS representations for community organizations and public participation GIS, as spelled out in the call for papers:

• The implications of map-based representations of information for community groups;
• The nature of GIS knowledge distortion from grassroots perspectives;
• The value and impact of increasingly sophisticated analyses for understanding key issues and marginalizing certain groups;
• The ways in which socially differentiated communities and their local knowledge are or might be represented within GIS, and the impacts on communities of differential access to hardware, software, data, and expertise in GIS production and use;
• The educational, social, political, and economic reasons for lack of access, and exemplary ways communities have used to overcome these barriers;
• The implications of conflicting knowledge and multiple realities for spatial decision making;
• GIS as local surveillance;
• Current attempts to use GIS to empower communities for spatial decision making;
• Changes in local politics and power relationships resulting from the use of GIS in spatial decision making;
• Successful implementations of a public participatory GIS;
• What community groups need in the way of information, and the role GIS plays or could play in meeting this need;
• The economic impact of government pricing policies on small and large businesses.

Looking forward, a variety of potential avenues for future research on the geographies of the information society can be discerned. The questions raised in the context of the individual Apex initiatives, and the recommendations for research agendas arising from those meetings, will help assess the influence of these initiatives over the broader research scene for the next three to five years. A parallel agenda-setting effort is reflected in a white paper on 'GIS and society' by the University Consortium for Geographic Information Science (UCGIS 1998), with specific recommendations about research which are worth noting here briefly. The UCGIS recommends that attention given to the impact of GIS on society be counter-balanced by attention to the impact of society on the evolution of geographic information technologies; that attention to the determinants and consequences of GIS in public agencies and institutions be balanced by the study of their use by private firms and by community and grassroots groups; and that attention to empirical questions regarding the societal determinants and consequences of GIS be counterbalanced by attention to ethical and legal implications (for more details, see http://www.ucgis.org).

The Varenius roadmap largely intersects with the UCGIS recommendations but is sketched here more specifically as the logical continuation of the three Apex initiatives. Thus the theoretical perspective on geographic information, information technology, and society, best represented by the place and identity initiative, must eventually lead to the development of case studies investigating the relevance of the theoretical views within concrete places and situations. Second, the attempt to conceptualize accessibility in the information age should be followed by similar investigations of how to measure and represent socially induced changes in other fundamental geographic concepts such as proximity, community, region, and so on. Third, the concerns raised by the current practice of GIS should lead to the study of new tools, new data schemes, and new institutional frameworks for facilitating and safeguarding democratic values at all levels.

The evolving relationships between society and GIS can take many forms depending on local context and circumstances, and it seems that at this stage we know less about the actual than we do about the potential consequences. We believe that further progress is best pursued through a series of carefully selected case studies of particular organizational and geographic contexts. Since less is known about GIS in the private sector and in community organizations, case studies of these contexts will be particularly useful both to further develop and also to challenge and improve our understanding of theoretical scenarios.

We need case studies and eventual generalization regarding the widespread and increasingly sophisticated use of geo-demographic analysis. Under what corporate (large or small firms, local or global marketing and production strategies, industry setting), localized or social-context circumstances (entrepreneurship, capital availability, data availability) and under what geographic characteristics (degree of heterogeneity of regions and neighborhoods) does geodemographic modeling lead to redlining or other forms of geographic exclusion or, on the other hand, more efficient niche marketing and better, specialized location and investment decisions? What sorts of effects has the widespread use of neighborhood or block-group-targeted marketing campaigns had on heterogeneity of tastes and households within and across neighborhoods? What are the implications of widespread use of spatial-interaction and location-allocation models in corporate geography for competition, for public welfare, and for corporate welfare (e.g., how effectively and appropriately have these tools been used)? What public-policy measures (e.g., formal regulation of the dissemination and use of private, household information) would be effective in reducing the negative impacts of geo-demographic analysis and marketing?

Second, we need further case studies of the use of GIS software by neighborhood organizations in low-income and minority communities, seeking to improve the social and physical environment available to community residents; its effect on the ability of these organizations to make or negotiate improvements; and its effect on the internal coherence of these organizations and their ability to represent the diversity of views of local residents. Case studies of such organizations beginning to transform the nature of GIS practices through their actions do not yet exist, but as they emerge they will begin to give a better sense of the possibilities and limits of current software, compared to new technologies (see below).

Third, there is still a need for case study analysis of the use of GIS in public agencies. Many such studies are documenting the influence of GIS on the actions of government agencies, and on the capacity of the general public to assert democratic influence over those agencies, and important questions in this area still remain. Of particular interest, however, are studies now beginning to emerge which are showing how the practices of GIS use in such agencies are themselves actively constructing GIS conventions and norms.

Finally, there is a great need for case studies of controversial applications of GIS, paying attention to what can be learned about appropriate ethical principles and legal regulatory mechanisms.

In concert with and drawing on such individual case studies, comparative analysis across case studies will be important to tease out which kinds of contextual conditions affect which kinds of outcomes. This will be as important for the study of how social practices influence the evolution of GIS technologies as it is for the study of the social implications of GIS. Such analysis should compare both case studies of similar organizational contexts in different places, and case studies of different organizational contexts in similar places. A successful outcome of such comparisons would be the development of mid-range generalizations about the relationships between GIS and society, and about ethical and legal principles, which may be capable of further examination through a combination of extensive empirical analyses and new targeted case studies. Progress on these questions will depend crucially on fostering collaborative research networks.

This research direction would attempt the synthesis of the societal view of geographic concepts taken under the Apex with other Varenius research in the area entitled 'Cognitive Models of Geographic Space', which approaches similar questions from a cognitive perspective. For reasons of research efficiency Varenius took a disjunctive approach to the study of basic geographic concepts, treating them either as cognitive constructs, as in the case of scale, or as evolving socio-empirical phenomena, as in the case of place and accessibility. The dialectic between the cognitive and the societal must now be played out in a new research arena where these concepts can be defined and represented with the richness appropriate to their dual role: compatible with the cognitive structures of individual GIS users, while at the same time reflective of the contextual and evolving meanings that the information society bestows on them. That fusion is needed for the development of a new generation of geographic information technologies that are advanced and sophisticated enough to play the roles proponents have always hoped for and skeptics have always doubted. The more applied and practical aspects of that ambitious program would be picked up by the third stream of research, continuing the work of the Public Participation GIS initiative.

It is important to develop a parallel area of research into new types of GIS technologies, perhaps more reflective of the flexibility and communicative logic of Java and the Web than the complex logic of expert programs over which users have little influence. To be effective in designing geographical information systems that are appropriate for all areas of society, such developments should combine the practical experiences of new users struggling with currently dominant GISs; the expertise of programmers, graphic artists and communications specialists; and the experience and expertise of individuals skilled in the study of GIS and society. Focused research in this area will increase the possibility of lateral development of new approaches to GIS which can qualitatively enhance their relevance for an equitable and democratic society.

One important aspect, where research is already making advances, is in our ability for (and our software for) simultaneously and interactively showing, manipulating, and commenting on computer-based information. How can these technologies be combined to allow group interaction and analysis via GIS? What are the relevant differences in human perception of geographic relationships and human visualization of computer-generated models, that affect the ability of multiple individuals to interact in this way? Assuming that the technological and dissemination issues are resolved, what are the social, communication, perception, and cognitive barriers to the interactive use of spatial-display and spatial-analytic tools? This research is moving GIS and related technologies away from the norm of being tools for the individual analyst to store and manipulate spatial information, but there are many other aspects of software development, ranging from interface design to Web-based technologies of communication, that are necessary before the GIS2 vision can be approximated.

Finally, relevant institutions (UCGIS, software developers, and specialty groups of researchers whose main focus is not GIS) should develop demonstration projects and compilations of the use of GIS in human-geographic research: migration, labor analysis, or central-place modeling. Specifically, demonstrations are needed of innovative uses of data and new data sources, and of new data consortia. These same institutions, along with funding agencies such as NSF and private foundations, need to catalyze the development of GIS software that employs more rigorous approaches to spatial statistics, regionalization schemes, and more probabilistic approaches to estimates of spatial interaction; new analytic methods; and new, automated modeling tools.

It is to be hoped that this sketch of a roadmap will stimulate the kind of criticism and further debate needed to articulate a sustained, rigorous, and reflective agenda for research into geographies of the information society by both practitioners of geographic information science and their critics.