International Journal of Geographical Information Systems,

1992, Vol. 6, No. 1, 31–4

Geographical Information Science*

Michael F. Goodchild

Abstract. Research papers at conferences such as the European Geographical Infor-

mation Systems (EGIS) and the International Symposia on Spatial Data Handling

address a set of intellectual and scientific questions which go well beyond the limited

technical   capabilities   of   current   technology   in   geographical   information   systems.

This paper reviews the topics which might be included in a science of geographical

information. Research on these fundamental issues is a better prospect for long-term

survival and acceptance in the academy than the development of technical capabil-

ities. This paper reviews the current state of research in a series of key areas and

speculates on why progress has been so uneven. The final section of the paper looks

to the future and to new areas of significant potential in this area of research.

1      Introduction

The geographical information system (GIS) community has come a long way in the

past   decade.   Major   research   and   training   programmes   have   been   established   in   a

number of countries, new applications have been found, new products have appeared

from an industry which continues to expand at a spectacular rate, dramatic improve-

ment continues in the capabilities of platforms, and new significant data sets have

become available. It is tempting to say that GIS research, and the meetings at which

this   research   is   featured,   are   simply   a   part   of   this   much   larger   enthusiasm   and

excitement, but there ought to be more to it than that.

What, after all, is the purpose of all of this activity? Expressions such as ‘spatial

data handling’ may describe what we do, but give no sense of why we do it. This

was one of the themes behind Tomlinson’s keynote address at the First International

Symposium on Spatial Data Handling in Zürich in 1984 (Tomlinson 1984). The title

of the conference suggests that spatial data are somehow difficult to handle, but will

that always be so? It suggests a level of detachment from the data themselves, as if

the U.S. Geological Survey were to send out tapes labeled with the generic warning

‘handle with difficulty’. It is reminiscent of the name of the former Commission on

Geographical Data Sensing and Processing of the International Geographical Union.

A quick review of the titles of the papers at that or subsequent meetings should be

*  Based   on   keynote   addresses   by   the   author   at   the   Fourth   International   Symposium   on   Spatial   Data

Handling, Zürich, July 1990 (Goodchild 1990), and EGIS 91, Brussels, April 1991 (Goodchild 1991).

enough to assure anyone that their authors are concerned with much more than the

mere handling and processing of data — from a U.S. perspective, that the community

is more than the United Parcel Service of GIS.

Geographical information systems are sometimes accused of being technology

driven, a technology in search of applications. That seems to be more true of some

periods of the 25-year history of GIS than of others. For example, it is difficult to

suggest that Tomlinson and the developers of the Canadian Geographical Information

System   (CGIS)   (Tomlinson et   al.  1976)   were   driven   by   the   appallingly   primitive

hardware capabilities of 1965. On the other hand the prospect of a menu driven, full

colour,   pull-down   menu   raster   GIS   in   the   386-based   personal   computer   on   one’s

desk has clearly sold many systems in the past few years. Technological development

comes in distinct bursts, and so does the technology drive behind GIS. It may be

the motivation behind the desire to handle spatial data, but it fails to explain many

of the diverse research efforts being reported at meetings and in the literature.

There   have   also   been   phases   when   applications   have   driven   GIS.   CGIS   itself

was an application in search of a technology, and the drive was sufficiently strong

to lead to the prototype of the first map scanner, and to numerous other technological

developments   (Tomlinson  et   al.  1976).   McHarg   had   worked   out   the   principles   of

the map overlay technique (McHarg 1969) long before Berry and others automated

them in MAP and its derivatives (Berry 1987); school bus routing software has been

around much longer than the problem’s implementation in a standard GIS. But again,

much of the subject matter of GIS research lies well beyond any reasonably fore-

seeable application.

There   have   also   been   phases   when   applications   have   driven   GIS.   CGIS   itself

was the widespread distribution of Landsat and SPOT imagery, and the availability

of digital elevation models and street files in many countries have certainly led to

applications well beyond those used to justify the data’s compilation. TIGER, for

example,   appears   to   be   spawning   its   own   industry   of   updaters,   repackagers   and

application developers, although it exists in principle only to serve the needs of the

1990 U.S. Census.

However, although the driving seat of GIS is undoubtedly crowded, I would like

to deal in this paper largely with the fourth driver located apparently irrelevantly in

the back seat, the ‘S’ word. It seems to me that there is a pressing need to recognize

and develop the role of science in GIS. This is meant in two senses. The first has

to do with the extent to which GIS as a field contains a legitimate set of scientific

questions, the extent to which these can be expressed and the extent to which they

are   generic,   rather   than   specific   to   particular   fields   of   application   or   particular

contexts.   To   what   extent   is   the   GIS   research   community   driven   by   intellectual

curiosity about the nature of GIS technology and the questions that it raises? And

if GIS can be motivated by science, then what are its subfields, what are its questions,

and what is its agenda? The second sense has to do with the role of GIS as a toolbox

in science generally — with GIS for science rather than the science of GIS. What

do we need to do to ensure that GIS, and spatial data handling technology, play their

legitimate role in supporting those sciences for which geography is a significant key,

or a significant source of insight, explanation and understanding?

To do this we must first establish that spatial, or rather, geographical data are

unique,   and   that   their   problems   cannot   therefore   be   subsumed   under   some   larger

field.   We   must   also   establish   that   there   are   problems   which   are   generic   to   all

geographical data, or at least establish that it is possible to distinguish those that are

from those that are not. For example, the accuracy of attributes on a choropleth map

of crime statistics would seem to be very little informed by knowledge of attribute

accuracy   for   geographical   data   generally,   but   to   require   instead   a   level   of   under-

standing   of   the   specific   problems   of   crime   statistics.   However,   the   accuracy   of

population estimates for an arbitrarily defined polygon may well be known from,

or at least informed by, the general properties of the modified areal unit problem

(Openshaw 1977).

2       What is unique about spatial data?

In many facilities management systems, the role of the GIS is to provide an alter-

native key to data, a method of access based on geographical location. In essence,

a spatial database has dual keys, allowing records to be accessed either by attributes

or by locations. However, dual keys are not unusual. The spatial key is distinct, as

it allows operations to be defined which are not included in standard query languages.

For   example,   it   is   possible   to   retrieve   all   point   records   lying   within   an   arbitrary,

user-defined polygon, an operation which is not defined in standard query languages

such as SQL. In essence, the spatial key is multidimensional, but again multidimen-

sional keys are known from other areas, and analogues of point in polygon retrieval

can be defined for non-spatial dimensions.

What distinguishes spatial data is the fact that the spatial key is based on two

continuous   dimensions.   It   is   possible   to   visit   any   location   (x, y)   in   the   real,   geo-

graphical world, defined in principle with unlimited precision, and return a value

for a variable, for example, topographic elevation z. Terrain is thus characterized by

an infinite number of tuples <x, y, z>. In network applications z is defined only for

locations   on   the   network,   but   the   number   of   tuples   is   still   infinite   if   variation   is

continuous along this one-dimensional structure of links and nodes. Time series also

have continuous keys, but are rarely conceived, measured or represented as contin-

uous,   and   there   appears   to   be   little   commonality   of   interest   in   the   problems   of

temporal data handling. By contrast, there is ample evidence of commonality in the

spatial data handling disciplines.

Many of our data models, particularly polygon networks and triangulated irreg-

ular   networks   (TINs),   reflect   an   underlying   view   of   space   as   continuous   and   the

need   to   accommodate   the   user   who   wishes   to   determine  z  at   some   arbitrary   and

precise (x, y). One implication of this is that there exists a multiplicity of possible

conceptual   data   models   for   spatial   data,   and   that   the   choice   between   them   for   a

given phenomenon is one of the more fundamental issues of spatial data handling.

Another distinctive feature of spatial data is what Anselin (1989) refers to as

spatial dependence, the propensity for nearby locations to influence each other and

to possess similar attributes. Without spatial dependence, there would be no reason-

able   prospect   of   creating   even   approximate   views   of   continuous   spatial   variation

within a discrete, finite machine. It is not uncommon for tuples which have similar

values of a key to have similar values of other attributes, but the structure of spatial

dependence is unusual, relying as it does on both dimensions of the (x, y) key, with

similarity determined by a metric.

Finally, geographical data are distributed over the curved surface of the earth, a

fact which is often forgotten in the limited study areas of many GIS projects. We

have worried for centuries about how to portray the earth’s surface on a flat sheet

of paper, and have developed an extensive technology of map projections. However,

as a result we have few methods for analyzing data on the sphere or spheroid, and

know little about how to model processes on its curved surface. Moreover, we tend

to have treated GIS displays as if they were virtual sheets of paper, and insisted on

viewing   geographical   data   as   if   they   were   projected   to   a   flat   surface,   instead   of

exploiting the potential of electronic display to create views of the globe itself. We

need to develop the appropriate techniques for working with the globe, and making

use of solid modeling rather than conventional two-dimensional graphics, if we are

to understand geographical processes at the global scale and contribute effectively

to   global   science.   We   must   rescue   the   orthographic   projection   from   its   present

obscurity.

3      The content of geographical information science

Having established that geographical information has unique properties and prob-

lems,   we   can   now   review   the   set   of   generic   questions   which   might   make   up   a

geographical information science. This can be done in a largely linear fashion, from

data   collection   to   analysis,   although   some   themes   tend   to   cut   across   this   simple

arrangement. However, it seems appropriate to begin this review with a disclaimer.

What I present in this paper is in many ways my own view, and I would expect it

to be challenged. I think my own biases will become clear in what follows. Because

of the field’s diversity and dynamism it is difficult, if not impossible, for any one

individual to attempt a general overview. What follows is therefore almost inevitably

incomplete and uneven.

Research   is   often   identified   as   either   pure   or   applied   —   driven   by   basic   and

innocent human curiosity or by the practical everyday needs of human society. Many

GIS are a response to human needs for information management and analysis, and

in that sense one might expect GIS research to be more applied than pure. However,

one view of pure research is that it is research that has not yet found application;

pure   research   is   a   long-term   investment   just   as   applied   research   is   a   short-term

investment. From an academic perspective, pure research is often associated with

higher prestige, but applied research with greater funding. I have tried to cover the

full range from pure to applied, feeling that both are important to GIS. At the same

time ‘basic research’ is the primary purpose of the U.S. National Center for Geo-

graphical Information and Analysis, and the center is very fortunate in being funded

to do research the applications of which may lie years or even decades into the future.

During the design phase of the CGIS in the 1960s, it became clear that the only

practical way to input the large number of maps needed would be by some form of

scanning   device   (Tomlinson  et   al.  1976). At   that   time   no   scanner   for   map-sized

documents existed, and it was necessary to invent one. A prototype drum scanner

was   built   by   IBM   Canada   and   successfully   tested,   at   what   by   modern   standards

would be regarded as vast expense. Other parts of the CGIS design team were busy

inventing   other,   equally   fundamental   and   now   familiar   solutions   to   technical   GIS

problems, such as the Morton order.

In   the   almost   three   decades   of   development   of   GIS   that   are   now   behind   us,

similar   ‘how   to   do   it’  research   has   produced   a   large   number   of   algorithms,   data

structures, spatial indexing schemes and other technological solutions. Some of these

are unique to GIS, but many have been reinvented in several related disciplines. The

Morton order, for example, occurs in the literature of several spatial data handling

fields under different names (Samet 1989), and descriptions of algorithms for finding

Thiessen polygons are spread over a wide range of journals. At the same time there

is   a   growing   sense   in   GIS   research   that   our   emphasis   has   changed,   as   more   and

more of the underlying technical problems of GIS are solved. Attention has moved

from primitive algorithms and data structures to the much more complex problems

of database design, and the issues surrounding the use of GIS technology in real

applications. The following sections identify some of these key issues.

3.1        Data collection and measurement

If spatial reality is continuous and subject to complex structures of spatial depen-

dence,   then   how   should   it   be   compiled   and   measured?   More   generally,   how   do

people perceive the real world of geographical variation, structure it and learn about

it? Although   many   of   these   questions   are   part   of   the   research   agendas   of   remote

sensing, photogrammetry, geodesy and cognitive psychology, the lines of demarca-

tion are far from distinct. Should GIS or remote sensing concern the problems of

transferring   information   from   one   technology   to   the   other,   and   more   importantly

making   good   sense   of   it?   Is   it   GIS   or   remote   sensing   if   ancillary   geographical

information is used to improve the accuracy of classification or if an image is used

to update a GIS layer? Ultimately it matters little to which of the many pigeonholes

we assign each topic. There are undoubtedly substantial scientific questions here,

which require a depth of understanding of the nature of spatial variation, and one

person’s remote sensing may well be another’s geographical information science.

The   process   of   discretization,   with   its   implied   generalization,   abstraction   and

approximation, takes place as data are collected, interpreted or compiled, and choices

are   made   at   this   stage   that   affect   the   ultimate   uses   of   the   data. When   those   uses

change, as they have been doing with the widespread use of GIS, it may be necessary

or   beneficial   to   rethink   the   process   of   data   collection.   For   example,   with   digital

management and delivery of census data, is it still appropriate to conduct a census

on   a   decennial   basis?   Is   the   traditional   approach   to   geological   field   mapping   the

most appropriate if the eventual objective is a digital three-dimensional representa-

tion of the subsurface? How will topographic mapping change now that it is cost-

effective to survey new features using the Global Positioning System? Geographical

data collection is often the domain of specialists in well established disciplines, so

it may be many years before these kinds of questions are investigated or answered.

To date the introduction of GIS seems to have had very little effect on the process

of data collection.

3.2       Data capture

Enormous strides have been made in the technology for capturing digital geograph-

ical data in the past decade, and the systems now on the market are capable of a

high   level   of   intelligence   in   interpreting   scanned   map   documents.   The   problem

remains the poor quality of the documents, and the ambiguities that are caused by

aspects of map design. As a result, manual digitizing remains a widely used approach,

despite its high cost, tedium, and failure to show significant improvements in effi-

ciency. Two   trends   may   change   this   situation   substantially   in   the   next   few   years.

One is the increasing avoidance of the map document as a step in the data compilation

and input process. Surveying and photogrammetry are moving away from compila-

tion using paper maps, and the more interpretive fields such as land use, vegetation

or   soil   mapping   are   likely   to   follow   suit.   The   digital   total   station   is   likely   to   be

followed by the digital plane table and perhaps even the digital field geology note-

book. The other is the long recognized possibility that comparatively minor changes

in a map’s design can make it vastly easier to scan and interpret (Shiryaev 1987).

3.3       Spatial statistics

As spatial data are always an approximation or generalization of reality, they are

full of uncertainty and inaccuracy. A change of data model or scale can introduce a

loss   of   information,   as   can   digitizing   or   scanning.   Processing   in   a   finite   machine

also inserts its own form of uncertainty, although this is often insignificant in relation

to the errors inherent in the data themselves. Many human geographical constructs

are   implicitly   uncertain,   including   spatial   objects   (‘Indian   Ocean’,   ‘Europe’)   and

their relationships (‘in’, ‘across’). Whether we think of uncertainty in set theoretical

terms   through   notions   of   fuzziness   or   in   statistical   terms   through   the   calculus   of

probabilities, the study of spatial data uncertainty, its measurement and modeling,

and the analysis of its propagation through the processes of spatial data handling

are undoubtedly part of geographical information science. How should one compile

an accurate representation of geographical variation for input to a database? How

should one represent the uncertainty or inaccuracy present in a digital representation?

How can uncertainty be propagated from database to GIS products?

Geographical data bring their own special set of problems to spatial statistics.

Whereas in medical imaging the problem may be to determine the true location of

objects from ‘dirty’ pictures (Besag 1986), in geographical images there is often no

clear concept of truth, as objects are often the products of interpretation or gener-

alization. We need much better methods of measuring and describing uncertainty,

particularly in the complex spatial objects common in GIS. We need better methods

for dealing with the world as a set of overlapping continua, instead of forcing the

world   into   the   mould   of   rigidly   bounded   objects.   Most   of   the   answers   to   these

questions   will   have   to   come   from   spatial   statistics,   but   geographical   information

specialists   must   provide   the   motivation   and   the   examples,   and   define   the   overall

objectives and constraints.

Although all geographical data are uncertain to some degree, all of the current

generation of GIS follow the common practice in cartography and represent geo-

graphical   objects   as   if   their   positions   and   attributes   were   perfectly   known;   data

quality may or may not be addressed in a separate statement. The consequences of

uncertainty   for   GIS   products   are   never   estimated.   Recent   research   has   followed

several different and productive lines in attempting to address the problem of data

quality.   One   is   to   match   precision   to   accuracy.   In   a   locational   sense,   this   means

using limited precision in data representation and processing, most often through

the   use   of   a   raster   whose   size   is   determined   by   data   accuracy. Various   forms   of

quadtree structure have also bee used to fit locational precision to known levels of

accuracy. There have been several recent papers on finite resolution processing in

GIS (e.g. Franklin 1984, Dutton 1989) and finite resolution geometry is an active

research area in mathematics.

Another productive approach has been to incorporate techniques from geosta-

tistics, notably kriging, as the statistical basis of these techniques makes uncertainty

explicit. We now have several useful models of digitizing error, and its consequences

for estimated measures such as area (e.g., Chrisman and Yandell 1988, Keefer et al.

1988). Finally, there have been several successful efforts to model geographical data

sets as random fields, or derivatives of random fields, and to use this approach to

model   uncertainty   in   GIS   objects   (e.g.,   Goodchild   1989).   Between   all   of   these

methods, we probably now have an adequate set of models of accuracy from which

to build an error-tracking GIS. However, spatial statistics is not an easy field, and

many of these techniques go well beyond elementary statistics in their conceptual

sophistication.

3.4       Data modeling and theories of spatial data

Data   models   are   the   logical   frameworks   which   we   use   to   represent   geographical

variation in digital databases. As each must be an approximation, the choice between

alternative models constrains not only the functions available, but also the accuracy

of products. Of all the developments in GIS in the past decade, perhaps the most

exciting   has   been   the   proliferation   of   data   models,   and   the   growing   literature   on

their relative merits. The debate over raster and vector goes back to the earliest days,

but has now been joined by debates over objects, layers, the philosophy of object

orientation, hierarchical models of complex objects, and the entire range of possi-

bilities inherent in time dependence and three dimensions. Despite the interest, we

still do not have a complete and rigorous framework for geographical data modeling,

even in the static two-dimensional case, and without one it is difficult to see how

GIS can escape the constraints imposed by specific system implementations. How

much capability is being lost by forcing contemporary applications into the multi-

layer raster model used by many systems, or the point–line–area coverage model

used by many others? This is both a pure and an applied research problem. On the

one   hand,   we   must   develop   a   comprehensive   framework   for   geographical   data

modeling, with an associated terminology, to provide the basis for standards and an

ideal against which specific systems can be measured. On the other hand, an abstract

framework is of little value if it does not influence practice, through implementation

in the products of the vendors. Here the real issue is whether it is possible to enlarge

or ‘retrofit’ the data model underlying an existing product, or whether any attempt

to do so is doomed to cause inconsistency and incoherence.

These issues are precipitating lively discussion over the entire question of the

degree   to   which   we   view,   analyse,   represent   and   model   the   world   as   discrete   or

continuous,   as   a   collection   of   objects   or   a   set   of   fields.   Do   we   think   in   terms   of

variables   with   defined   values   everywhere   in   space,   or   of   an   empty   space   littered

with possibly overlapping objects? In essence, these issues have brought the GIS

debate from the comparative obscurity of internal data structures to the much more

general issues of how we understand geographical variation. Everyday human expe-

rience sees a world of objects, but the science of natural processes deals more with

continuous variation (Frank and Mark 1991). Thus the object oriented debate threat-

ens to pit the New Agers against the embattled remnants of the Enlightenment, and

what could be more stimulating than that?

3.5       Data structures, algorithms and processes

Many of the results of basic research which have accumulated over the past 25 years

in this field of research concern internal representations of data, and the algorithms

which   operate   on   them.   The   quadtree   (Samet   1989),   band   sweep   algorithms   for

overlay (White 1977), analysis of computational complexity (Preparata and Shamos

1988) and the arc-node data structure (Peucker and Chrisman 1975) are all intellec-

tual breakthroughs of lasting significance. Many challenging problems remain, for

example in the design of efficient algorithms to minimize overposting and in other

areas of cartographic design, or in developing better methods for converting between

various terrain data models. Many systems now handle data through database man-

agement   systems,   and   data   structure   issues   have   moved   more   and   more   into   the

realm of computer science. We seem, however, to have reached a point where all of

the simpler, more generic problems have been solved, and where what remains is a

set of difficult, context-specific problems. It seems clear, for example, that further

advances   in   the   conversion   of   terrain   data   models   (for   example,   from   contour   to

TIN) will require a much better understanding of the nature of terrain (Mark 1979),

and will perhaps have to be specific to terrain type (e.g. fluvial versus glacial). There

will   also   continue   to   be   a   need   for   research   on   efficient   methods   of   storage   and

access to deal with the enormous volumes of data likely to become available in the

coming decade.

3.6       Display

Geographical   information   systems   have   often   been   criticized   for   failing   to   give

adequate attention to principles of cartographic design (Buttenfield and Mackaness

1991), or for regarding the map as a simple store of information rather than a tool

for communication. If we think of the database as the truth, then a map is no more

than a store, as there is often a simple correspondence between objects in the database

and objects on the map. However, if the database is seen merely as an approximation

of   the   geographical   truth,   then   the   design   of   output   displays   is   critical,   as   it   can

affect the user’s view of the world. Such simple things as the choice of background

colour, or the contrast between adjacent polygons (McGranaghan 1991) can have a

significant effect.

The capabilities of electronic display go far beyond those of conventional car-

tography. We need research on the design of animated displays, three dimensional

display, the use of icons and metaphors in user interfaces, continuous gradation of

colour   and   tone,   zoom   and   browse,   multiple   media   including   voice   and   pointing

devices, multiple windows which allow simultaneous access to spatial and temporal

series of multivariate data. We need to use the electronic medium to think far beyond

improvements to the design of choropleth maps. All of these are fundamental prob-

lems to a science of geographical information.

3.7       Analytical tools

A   GIS   is   a   tool   for   supporting   a   wide   range   of   techniques   of   spatial   analysis,

including processes to create new classes of spatial objects, to analyse the locations

and   attributes   of   objects,   and   to   model   using   multiple   classes   of   objects   and   the

relationships between them. It includes primitive geometric operations such as cal-

culating the centroids of polygons, or building buffers around lines, as well as more

complex operations such as determining the shortest path through a network. The

functionality of leading products continues to grow, with no obvious end in sight.

Despite widespread recognition that analysis is central to the purpose of a GIS,

the lack of integration of GIS and spatial analysis, and the comparative simplicity

of the analytical functionality of many systems continues to be a major concern. In

the early days of the statistical package SAS, there was a very rapid increase in the

range of tests and techniques implemented in the system. Unfortunately, the same

has not been true of GIS, and remarkably little progress has been made in incorpo-

rating the range of known techniques of spatial analysis into current products.

There are many reasons for this. One obvious reason is the heavy emphasis in

the GIS marketplace on information management rather than analysis. The lucrative

markets for GIS technology have comparatively unsophisticated needs, emphasizing

simple queries and tabulations. Another is the relative obscurity of spatial analysis,

a set of techniques developed in a variety of disciplines, without any clear system

of codification or strong conceptual or theoretical framework. Even now it is difficult

to identify more than a handful of texts (e.g., Haining 1990, Upton and Fingleton

1985). Although one might expect that GIS could provide the basis for a system of

codification for spatial analysis, the poor level of current understanding of geograph-

ical data models is a major difficulty. Tomlin (1990) has made one of the few attempts

to add some sort of structure or framework to the proliferation of GIS functions,

which in the case of ARC/INFO is already around 103. We badly need a taxonomy

of spatial analysis, developed perhaps from an enumerated set of data models, but

going well beyond the primitive geometrical operations.

At this stage, integration of GIS and spatial analysis is proceeding slowly, in at

least three different modes. Some analytical capabilities are being added directly to

GIS,   for   example   in   the   recent   expansion   of   functionality   in   several   modules   for

network analysis. Some progress is being made in loosely coupled analysis, where

an   independent   analysis   module   relies   on   a   GIS   for   its   input   data,   and   for   such

functions as display. However, still missing is an effective form of tight coupling,

in which data could be passed between a GIS and a spatial analysis module without

loss of higher structures, such as topology, object identity, metadata, or various kinds

of relationships. At present this is impossible, to a large extent because of a lack of

standards for data models. Instead, coupling has to occur at a lower level, and higher

structures have to be rebuilt on an arbitrary basis.

Integration   between   GIS   and   spatial   analysis   might   also   take   the   form   of   a

language, whose primitive elements would represent the fundamental operations of

spatial   analysis.   The   beginnings   of   such   a   language   already   exist   in   the   macro

languages   of   many   of   the   current   generation   of   GIS,   and   in   various   attempts   to

extend SQL to spatial operations. However, all of these are specific to, and heavily

dependent on limited data models, and there is remarkably little similarity between

them at this time. At Santa Barbara we have been attempting to define a common

language from an analysis of the languages used by a variety of current GIS, but a

more satisfactory solution would begin with the conceptual framework provided by

a comprehensive data model.

Another   problem   in   integrating   GIS   and   spatial   analysis   is   that   in   the   former

discretization of space is explicit, whereas in many forms of spatial analysis it is

often either implicit, or unspecified. Many forms of spatial analysis are written on

continuous fields, and fail to deal with the uncertainties introduced by the inevitable

process of discretization. For example, in GIS there can be no measure of slope that

is independent of discretization, and similarly the length of an area object’s boundary

is   dependent   on   its   digital   representation.   However,   slope   and   length   commonly

appear as unqualified parameters in spatial models. In this sense, the integration of

GIS and spatial analysis is a two-way process, in which the inadequacies of both GIS

and spatial analysis must be addressed.

Most   of   the   current   generation   of   GIS   provide   some   sort   of   macro   or   script

facility, allowing the user to define products from complex sequences of operations,

but to invoke them with a single instruction. Although these often include the ability

to   construct   customized   environments   and   interfaces,   they   do   not   as   yet   provide

tools which are specific to the needs of spatial analysis. One limited exception is

Prime/Wild’s ATB, a set of tools constructed on top of System/9 which allows the

user to work with complex analyses, visualize their sequences and manage interme-

diate results. Tools like this will be needed increasingly if GIS are to move into an

era of more sophisticated analysis and decision support, because it is not uncommon

for relatively simple GIS products to involve processing tens of layers through similar

numbers of primitive steps. We need to research methods for keeping track of data

lineage   and   error   propagation,   backtracking   to   recover   intermediate   results,   and

preventing   the   user   from   combining   operations   in   incorrect   or   meaningless   ways

(Lanter 1990). We also need research on ways of incorporating this sort of analysis

into the GIS acquisition and planning process.

This emphasis on complex multistage analysis and the generation of products

from a multilayered database seems very different from research on knowledge based

systems, spatial reasoning and spatial query. One of the attractions of the GIS field

is its breadth of applications, and the correspondingly extreme variety of environ-

ments for the design of user interfaces. In data modeling, the important question is

not whether extended relational or object oriented models are better for geographical

data,   but   what   types   of   geographical   data   are   best   modeled   by   each   approach.

Similarly, the important research issue in the design of user interfaces is to determine

the   optimal   environment   for   each   of   the   many   types   of   GIS   application. What   is

best for a vehicle navigation system may be entirely different from what is best for

a forest resource manager with a deeply seated fear of keyboards and VDUs, either

colour or monochrome.

3.8       Institutional, managerial and ethical issues

Research   is   just   beginning   to   appear   on   the   issues   involved   in   implementing   and

managing GIS, especially in large institutions. This is difficult research, and general-

izations are not discovered easily. However, the success of several large projects in

the   U.S.A.,   and   the   discussions   surrounding   several   large   acquisitions   by   federal

agencies, have created the opportunity for a number of useful case studies. Many more

are   needed,   particularly   given   the   importance   of   such   research   for   improving   the

institutional   environment   in   the   future.   We   need   a   much   better   understanding   of

the processes of adoption of GIS technology and its effects on organizations; of the

value of geographical information and the benefits of GIS; and of processes for utilizing

geographical information in decision making. Theoretical frameworks for addressing

many of these issues already exist in the relevant social science disciplines, and we

need to make much more effective use of them in tackling the specific issues of GIS.

Despite the problems involved in adopting any new technology, GIS have been

widely adopted in local government, utilities and resource management agencies.

In   fact,   the   introduction   of   GIS   has   had   a   major   effect   on   the   management   of

geographical information in society. At the same time there is increasing concern

over the power of GIS for surveillance and invasion of privacy. The research com-

munity has a responsibility to monitor and study the more substantive aspects of the

GIS phenomenon, including its significance to society as a whole. What will GIS

mean to the balance of power in society? Will it be a technology available only to

the empowered, or will it somehow serve to even the distribution of power? Thus

far there have been remarkably few studies of the ethics of GIS.

4      Tests of commonality

The   preceding   sections   have   looked   at   various   candidate   areas   for   inclusion   in   a

geographical information science. In each case there are clearly challenging scientific

questions to be posed and researched. There is no reason to believe that the list is

complete, or that there are not additional and substantive questions in other related

areas. In each case the spatial context appears to be distinctive, although clearly it

is more so in some than others. For example, we might debate whether the spatial

context was distinctive in the area of decision theory, but the issue seems clear-cut

for data modelling.

In   the   NCGIA   research   plan   (NCGIA   1989),   we   argued   that   the   absence   of

solutions to issues such as these constituted impediments to the effective applications

of   GIS   technology.   Other   discussions   of   the   GIS   research   agenda   have   come   to

similar conclusions, although with different emphases (Craig 1989, Maguire 1990,

Masser 1990). Many are old issues, recognized long before the advent of GIS in

fields such as cartography, geodesy and geography. Some may not be unique to GIS.

For example, it is not immediately obvious that GIS technology diffuses in a fun-

damentally different fashion, or shows fundamentally different patterns of adoption

from other technologies. Is the measurement of GIS benefits a unique problem, or

an example of the more general problem of measuring the benefits of information

technology? Of course these questions are in themselves research issues.

At the same time it is very important to identify those areas where GIS have

created new and unique issues that are not common to other fields. In the early days

of GIS, it was possible to argue that the technology was filling an existing gap, and

making possible tasks that had been previously identified, but that were not easy to

carry   out   manually.   The   use   of   GIS   or   suitability   analysis,   by   overlaying   layers

(Tomlin   1990),   mirrors   the   manual   technique   popularized   by   McHarg,   although

admittedly   adding   some   interesting   new   capabilities.   CGIS   was   justified   on   the

grounds that the computer was a cost-effective alternative to hand measurement of

overlaid   areas.   But   GIS   make   it   possible   to   do   things   with   data   that   the   data’s

gatherers may never have envisioned. GIS technology is producing radical changes

in   the   way   geographical   data   are   collected,   handled   and   analysed,   and   it   will   be

many years before the impact of existing technology is felt, let alone the impacts

of future developments.

Here   are   some   of   the   issues   that   seem   unique   to   GIS:   how   to   model   time-

dependent geographical data; how to capture, store and process three-dimensional

geographical   data;   how   to   model   data   for   geographical   distributions   draped   over

surfaces embedded in three dimensions; how to explore such data, for example, what

exploratory metaphors are useful; and how to evaluate the geographical perspective

on information and processes relative to more conventional perspectives?

These   are   important   issues   for   GIS,   and   the   GIS   community   needs   a   strong

commitment to research if it is going to make significant progress on them. As issues

that arise within the context of GIS, they are not of major concern in other disciplines.

However, at the same time the GIS community can benefit enormously from inter-

disciplinary research. Statisticians can make a very valuable contribution to solving

the error problem in GIS, and research in cognitive psychology may be helpful in

designing the cognitive aspects of user interfaces in GIS.

This argument leads naturally to a proposed definition of GIS research: research

on the generic issues that surround the use of GIS technology, impede its successful

implementation,   or   emerge   from   an   understanding   of   its   potential   capabilities.   Is

this ‘research about GIS’ or ‘research with GIS’? In a sense it is both, because these

are issues that are both fundamental to the technology of GIS, and also issues that

must be solved before the technology can be successfully applied. If the problems

of doing research with GIS are generic, then they are best tackled as part of the GIS

research agenda. However, problems that are specific to the application of GIS in a

particular field clearly need to be addressed in the context of that field, and with the

benefit of its expertise. Accuracy issues provide a useful example. There are aspects

of the accuracy problem that span a wide range of types of geographical data, and

need to be solved using generic models of uncertainty, analogous to the role played

by the Gaussian distribution in the theory of measurement error. However as noted

earlier, an analysis of crime data using a GIS will also raise problems of accuracy

that   are   specific   to   that   particular   application,   and   need   an   understanding   of   the

processes operating in criminology and in the collection of crime data if they are to

be understood fully.

However, mere existence of scientific questions is far from an adequate basis

for   a   science.   Is   there   a   commonality   of   interest   here?   Can   these   subfields   find

sufficient basis for interaction that they will develop the lasting accoutrements of a

science,   such   as   journals,   societies,   books   and   philosophers?   Will   researchers   in

these subfields behave as a group of scholars? Is there a valid analogy between the

systems and science of geographical information on the one hand (tools supporting

researchers) and statistical packages and statistics on the other? Statistics is a highly

formalized discipline, but more technologically oriented groups can be found in such

areas   as   exploratory   data   analysis,   statistical   visualization   and   applied   statistics.

Certainly the relationship between science and tools is stormy at times, but never-

theless vital to the success of both. The ongoing debate over the value of statistical

software in teaching statistics has interesting implications for the same issue in GIS.

It may be useful to look briefly at the arguments for a commonality of interest

in geographical information science, first in principle and then in practice. The field

is small — rhetoric about growth in the industry aside, no one would suggest that

the field of GIS is a major discipline. It is distinct, with its own reasonably unique

set   of   questions.   And   it   is   certainly   challenging   and   innately   appealing.   On   the

negative   side,   it   is   multidisciplinary,   competing   with   longstanding   cleavages   and

rivalries.   It   lacks   a   core   discipline,   unlike   the   statistical   analogy,   where   there   has

been a steady growth in the number and size of academic departments for the past

few decades. One of the claimants to the core, geography, has traditionally been a

non-technical   field,   and   in   some   areas   of   social   geography   there   is   a   strong   and

fundamental antipathy to technological approaches.

In practice, commonality of interest is evident in the proliferation of GIS meet-

ings,   and   we   are   beginning   to   see   a   supply   of   books   and   journals.   However,   the

scientific   track   at   GIS   meetings   is   often   small.   People   who   attend   GIS   meetings

need a constant supply of novelty, whether in scientific research or vendor products,

and will soon desert if the supply dries up.

5       Options for the future

Looking back over nearly three decades of GIS research, it is clear that the greatest

progress has been made on the best defined and easiest problems, where solutions

lay   in   advances   in   the   technology   itself.   Rapid   progress   was   made   on   algorithms

and data structures in the 1970s and 1980s, but many of the difficult problems of

data modeling, error modeling, integration of spatial analysis and institutional and

managerial issues remain. Some of these may be unsolvable: for example, there may

simply   be   no   generalities   to   be   discovered   in   the   process   of   adoption   of   GIS   by

government agencies, however easy it may be to pose the research question.

Other issues have already been solved in a pure research sense, but implemen-

tation   remains   a   major   question   of   applied   research.   In   accuracy,   for   example,   a

substantial set of techniques has been defined, but the problem of moving them into

actual application remains. The academic research environment is set up to pursue

significant areas of research, but is generally poor at providing the means of imple-

mentation. For that we need a software industry that is tightly coupled to the research

community,   but   able   to   find   the   resources   to   motivate   development.   More   impor-

tantly, we need an education system that responds rapidly to new research and is

able to build new concepts quickly into its programmes. Unfortunately, the higher

education sector is too often characterized by conservatism, and it may take many

years for new ideas to work themselves into the curriculum.

Research in GIS is like geographical data — the more closely one looks, the

more interesting issues appear. GIS research has only begun to tackle the important

issues in the research agenda. We are in an enviable position, working in a field with

such   strong   motivation   and   such   a   strong   underlying   industry,   and   with   such   an

interesting set of problems spanning so many disciplines and fields. I hope I have

shown in this paper that the handling of spatial information with GIS technology

presents   a   range   of   intellectual   and   scientific   challenges   of   much   greater   breadth

than the phrase ‘spatial data handling’ implies — in effect, a geographical informa-

tion science. The term ‘geographical’ seems essential — much of what GIS research

is about concerns the geographical world and our relationships with it, and the term

is much richer than ‘spatial’. The change in meaning of the ‘S’ word — from systems

to science — seems to be going well, as evidenced by the success of the spatial data

handling series of conferences, the move of the AutoCarto series to fully refereed

papers,   the   new   texts,   subscriptions   to   the International   Journal   of   Geographical

Information Systems, and submissions of GIS papers to such established journals as

Geographical Analysis, Computers and Geosciences, Computer Vision, Graphics and

Image Processing and publications of the Regional Science Association and the IEEE.

I   hope   I   have   also   shown   that   a   strong   scientific   programme   serves   not   only

itself, but also the needs of industry and GIS users. GIS needs a strong scientific

and intellectual component if it is to be any more than a commercial phenomenon,

a short-lived flash in the technological pan. It is too easy to see current GIS as a

hardware and software technology in search of applications, and to see the field of

GIS as defined by the functional limits of its major vendor products. We need to

move from system to science, to establish GIS as the intersection between a group

of disciplines with common interests, supported by a toolbox of technology, and in

turn supporting the technology through its basic research. As currently perceived,

GIS sometimes seem about as close to a science as FORTRAN is to algebra.

In recent years we have seen a growing cleavage in GIS between two traditions,

that of spatial information on the one hand and that of spatial analysis on the other.

The spatial information tradition stresses large inventory databases, and gives geog-

raphy the role of an access mechanism. The spatial analysis tradition stresses rich

functionality and a range of data models, and gives geography a fundamental role

in   analysis   and   modelling.   The   two   traditions   share   common   data   structures   and

algorithms, and rely on the same sources of data and hardware. However, this is not

enough to convince the academy of the existence of a scientific field. To claim this

we need to take a broader view, and to include data modelling accuracy, cognition,

reasoning,   human–computer   interfaces   (HCI)   and   visualization,   and   to   show   how

these are integral parts of both traditions.

Without such arguments, the GIS field will fragment, and the GIS storm will

blow itself out. Associations as fundamentally disjoint as the Association of Amer-

ican Geographers and AM/FM will find it impossible to justify joint sponsorship of

conferences.   Vendors   will   specialize   in   data   input   workstations,   spatial   analysis

workstations or facility management systems, with little potential for interaction or

integration. This would be tragic.

How can we ensure a lasting future for both geographical information systems

and science? Disciplines are like tribes, with their own totems, symbols and mem-

bership rules, languages and social networks. The GIS tribe is currently very cohe-

sive; it is well funded, the field is exciting and much useful research is being done.

However, in the longer term the field has not done well at behaving as a science,

and the academy is still doubtful about whether it needs to be taken seriously. Science

is hard and places heavy obligations on its practitioners. We have been too busy,

and   technology   has   been   moving   too   quickly.   Too   much   of   our   literature   is   in

conference proceedings, which bring fast exposure but only to limited audiences,

and   lack   sufficient   quality   control.   Few   people   have   had   the   time   to   write   the

textbooks or to identify the intellectual core, or to publish the good examples.

I   believe   we   ensure   the   future   of   GIS   by   thinking   about   science   rather   than

systems,   and   by   identifying   the   key   scientific   questions   of   the   field   and   realizing

their intellectual breadth. Geographical information systems are a tool for geograph-

ical information science, which will in turn lead to their eventual improvement. We

need to speak to the academy, both directly and through key articles and texts, on

the philosophy, methodology and foundations of the field, and by placing GIS papers

in strong journals. All three communities — users, vendors and researchers — have

vital and symbiotic roles to play, and we will serve all three best by playing ours

in the fullest possible sense.

Acknowledgments

The National Center for Geographic Information and Analysis is supported by the

National Science Foundation through grant SES 88-10917.

References

Anselin, L., 1989, What is special about spatial data? Alternative perspectives on spatial data

analysis. Technical Report 89-4 (Santa Barbara, CA: National Center for Geographic

Information and Analysis).

Berry, J. K., 1987, Fundamental operations in computer-assisted map analysis. International

Journal of Geographical Information Systems, 1, 119–136.

Besag, J., 1986, On the statistical analysis of dirty pictures. Journal of the Royal Statistical

Society, B48, 259–302.

Buttenfield, B. P., and Mackaness, W. A., 1991, Visualiztion. In  Geographical Information

Systems: Principle and Applications edited by D. J. Maguire, M. F. Goodchild and

D. W. Rhind (London: Longman).

Chrisman,   N.   R.,   and Yandell,   B.   1988, A   model   for   the   variance   in   area.  Surveying   and

Mapping, 48, 241–246.

Craig, W.   J.,   1989,   URISA’s   research   agenda   and   the   NCGIA. Journal   of   the   Urban   and

Regional Information Systems Association,  1, 7–16.

Dutton, G., 1989, Modeling locational uncertainty via hierarchical tessellation. In Accuracy

of Spatial Databases, edited by M. F. Goodchild and S. Gopal (London: Taylor &

Francis), pp. 125–140.

Frank, A. U., and Mark, D. M., 1991, Language issues for GIS. In Geographical Information

Systems: Principles and Applications, edited by D. J. Maguire, M. F. Goodchild and

D. W. Rhind (London: Longman).

Franklin,   W.   R.,   1984,   Cartographic   errors   symptomatic   of   underlying   algebra   problems.

Proceedings,     International    Symposium      on   Spatial   Data    Handling,    Zurich,  pp.

190–208.

Goodchild, M. F., 1989, Modeling error in objects and fields. InAccuracy of Spatial Databases,

edited by M. F. Goodchild and S. Gopal (London: Taylor & Francis), pp. 107–114.

Goodchild, M. F., 1990, keynote address: spatial information science. Proceedings, Fourth

International Symposium on Spatial Data Handling, Zurich,  1, 13–14.

Goodchild, M. F., 1991, Keynote address: progress on the GIS research agenda. Proceedings,

EGIS 91, Brussels, pp. 342–350.

Haining, R., 1990, Spatial Data Analysis in the Social and Environmental Sciences  (Cam-

bridge: Cambridge University Press).

Keefer, B. J., Smith, J. L., and Gregoire, T. G., 1988, Simulating manual digitizing error with

statistical models. Proceedings GIS/LIS 88 (Falls Church, VA: American Society of

Photogrammetry and Remote Sensing/American Congress on Surveying and Map-

ping), pp. 475–483.

Lanter, D. P., 1990, Lineage in GIS: the problem and a solution. Technical Paper 90-6 (Santa

Barbara, CA: National Center for Geographic Information and Analysis).

Maguire, D. J., 1990, A research plan for GIS in the 1990s. The Association for Geographic

Information Yearbook 1990, (London, Taylor & Francis), pp. 267–277.

Mark, D. M., 1979, Phenomenon-based structuring and digital terrain modeling. GeoProcess-

ing, 1, 27–36.

Masser,   I.,   1990, The   Regional   Research   Laboratory   initiative:   an   update.  The Association

for Geographic Information Yearbook 1990  (London: Taylor & Francis), pp. 259–263.

McGranaghan,   M.,   1991,   Modeling   simultaneous   contrast   on   choropleth   maps.  Technical

Papers,    1991,   ACSM/ASPRS/Auto-Carto           10   Annual   Convention,    Baltimore,    MD,

March 25–29, 1991, Vol. 2, pp. 231–240.

McHarg, I. L., 1969, Design with Nature (New York: Doubleday).

NCGIA (National Center for Geographic Information and Analysis), 1989, The research plan

of the National Center for Geographic Information and Analysis) International Jour-

nal of Geographical Information Systems, 3, 117–136.

Openshaw, S., 1977, A geographical solution to scale and aggregation problems in region-

building, partitioning and spatial modelling.  Transactions of the Institute of British

Geographers, 2 (NS), 459–472.

Peucker, T. K., and Chrisman, N. R., 1975, Cartographic data structures. American Cartog-

rapher, 2, 55–69.

Preparata, F. P., and Shamos, M. I., 1988,  Computational Geometry: An Introduction (New

York: Springer-Verlag).

Samet, H., 1989, The Design and Analysis of Spatial Data Structures (Reading, MA: Addison-

Wesley).

Shiryaev, E. E., 1987, Computers and the Representation of Geographical Data (New York:

Wiley).

Tomlin, C. D., 1990, Geographic Information Systems and Cartographic Modeling (Engle-

wood Cliffs, New Jersey: Prentice Hall).

Tomlinson, R. F., 1984, Keynote address: geographical information systems — a new frontier.

Proceedings, International Symposium on Spatial Data Handling, Zurich, 1, 2–3.

Tomlinson, R. F., Calkins, H. W., and Marble, D. F., 1976, Computer Handling of Geograph-

ical Data (Paris: UNESCO).

Upton, G. J., and Fingleton, B., 1985, Spatial Data Analysis by Example (2 volumes) (New

York: Wiley).

White, D., 1977, A new method of polygon overlay. Proceedings, Advanced Study Symposium

on Topological Data Structures for Geographic Information Systems, Harvard.