You are on page 1of 23

Transactions in GIS, 1999, 3(2): 137±159

Research Article

Use of a Three-Domain Representation to


Enhance GIS Support for Complex
Spatiotemporal Queries

May Yuan
Department of Geography
University of Oklahoma

Abstract
As the development of Geographic Information Systems (GIS) proceeds to advanced
scientific and societal applications, there is an emerging need to enhance GIS support
for complex spatiotemporal queries. Dynamic GIS representations (as opposed to
static, map-based representations) that can integrate proper data elements in the
production of geographic information are required. This paper demonstrates the use
of a three-domain representation that facilitates compilation of higher-level
information (such as frequency and rate) from preliminary data records (such as
time and location) stored in a database. The three-domain representation is
compared with snapshot, space-time composite, and spatiotemporal object models
using a sample data set for forest transitions. While the three-domain representation
is a normalization of these data models, it offers a conceptual alternative that enables
GIS to represent spatiotemporal behaviors of geographic entities, in addition to
entities as well as histories at locations as emphasized in most GIS data models. The
comparison shows that the three-domain representation has combined the strengths
of the space-time composite and spatiotemporal object models. Moreover, it enables
aggregations of analytical use along with dynamic mappings between geographic
concepts and locations, a distinct capability that takes GIS query processing beyond
the level of information support offered by static map-based data models.

1 Introduction

The development of Geographic Information Systems (GIS) has advanced from


automatic mapping tools to a facility enabling a wide range of spatiotemporal queries

Address for correspondence: May Yuan, Department of Geography, University of Oklahoma,


100 E. Boyd Street, Sarkey Energy Center #684, Norman, OK 73019, USA. Email:
myuan@ou.edu

ß 1999 Blackwell Publishers, 108 Cowley Road, Oxford OX4 1JF, UK and
350 Main Street, Malden, MA 02148, USA.
138 May Yuan

and information computing. The foundation for a GIS to provide an adequate query
support lies in the design of data models and algorithms (Worboys et al 1990). A data
model is the conceptual core of an information system; it defines data object types,
object relationships, operations, and rules to maintain database integrity (Codd 1980).
In particular, a scheme used to define data object types and object relationships is `a
representation,' which delineates information that is describable and deductible
through the subsequent design of GIS algorithms. The advantages and disadvantages of
a representation have profound influences on what information can be represented and
computed in an information system. Moreover, the design of GIS representations will
affect the difficulty of handling geographic complexity, scale differences,
generalization, and accuracy (Burrough and Frank 1995). Improvements in
representations can significantly enhance the usefulness of GIS to interactive queries
and modeling support of a GIS (Goodchild 1992).
Support for geographic information queries is one of the primary GIS functions.
The degree to which a GIS can support spatiotemporal queries depends upon its
embedded representations. While simple spatiotemporal queries only search stored
data records, complex spatiotemporal queries need a representation to handle the
distributions of objects and events in space and time and thereby discover higher-level
information buried in GIS data records. Most GIS data models support simple
spatiotemporal queries, but their support for complex queries is limited. The goal of
this paper is to demonstrate a representation that can enhance GIS support for complex
spatiotemporal queries. In Section 2, several types of spatiotemporal queries are
discussed with reference to a sample data set to characterize the basic query types that a
GIS representation needs to support for information discovery. In Section 3, a three-
domain representation (Yuan 1996) is applied to unveil spatiotemporal information
embedded in the sample data set. In Section 4, the benefits of using a three-domain
model to represent forest transitions are illustrated through a comparison with other
data models. Key arguments and ideas are summarized in Section 5.

2 Types of Spatiotemporal Queries

Geographic information queries ask what, where, and when questions with reference
to geographic properties, spatiality, temporality, and the interactions between
geographic entities (cf Langran 1991, Peuquet 1994). These queries can be categorized
using an example of forest transitions with three geographic entities: old growth, clear-
cut, and fire-burn patches (Figure 1). From 1600 to 1960, the boundaries of the
geographic entities varied due to clear-cutting, burns, and the addition of new old-
growth patches. Areas of clear-cuts expanded significantly from the south, while
several fires occurred inside the old growth. On many occasions, these geographic
entities appeared at more than one location; they were aggregations of spatial patches.
While the data set is small, the amount of spatiotemporal information embedded in the
data is rich. In particular, the sample data include information about movements of
boundaries and aggregations of spatial objects. The dynamic nature of the embedded
information challenges conventional static map-based representations to support
complex spatiotemporal queries that require extracting higher-level information from
preliminary GIS data records. This study categorizes the following four types of
spatiotemporal queries to provide a systematic means to acquire information implicit

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 139

Figure 1 Sample forest transitions in space and time. The example includes
spatiotemporal constructs about change (e.g. change of land-cover types), aggregation
(e.g. incremental expansion of clear-cut area by adding new cuts), and movement (e.g.
movement of clear-cut boundaries).

in a GIS database: (1) queries about attributes of entities, (2) queries about locations,
spatial properties, and spatial relationships, (3) queries about time, temporal
properties, and temporal relationships, and (4) queries about spatiotemporal behaviors
and relationships.

2.1 Queries about Attributes of Entities


Attribute queries seek information regarding attributes of entities and relationships,
which may be independent of space and time. If geographic locations or time of
existence are critical to the identification of entities, these criteria should be included in
the definitions of these entities. Examples of attribute queries regarding the sample
forest transitions include: `Up to 1960, how many clear-cuts resulted from logging?' and
`How many clear-cuts were caused by fires?' These questions have no spatial
components, so they are manageable in non-spatial database management systems
(DBMS). However, changes of properties at a location may alter the identification of
geographic entities, such as ``reincarnation'' of the old-growth from clear-cuts or burns
(Figure 1e). This suggests two basic semantic relationships between the old-growth and
clear-cuts (as well as burns): (1) change of the old-growth to clear-cuts, and (2) change
of clear-cuts back to old-growth after a long period of time. A GIS needs to support
queries about these semantic relationships.
While most GIS studies refer to semantic relationships as attribute relationships,
the term semantics is more appropriate here to stress the essential geographic concepts
that a GIS database attempts to represent (the meaning of geographic data). `Semantics'
has been broadly used in the fields of database modeling and knowledge representation
to refer to data abstractions with direct correlation to how data arise (Hull and King

ß Blackwell Publishers Ltd. 1999


140 May Yuan

1987). In this paper, semantic information has two aspects: (1) contextual information
about components of the world, and (2) attribute information of these components.
Contextual information is domain dependent; in that discernable components
(including features, entities, events, or processes) and their attributes can vary with
user views. For example, farmers and engineers apply different criteria to characterize
the usefulness of soils. Information scientists have developed methods to capture a
diverse range of semantics and still preserve the integrity of a database. Commonly
used methods include entity-relationship analysis (Chen 1976), conceptual structuring
(Sowa 1984), semantic modeling (Hull and King 1987, Morehouse 1990), schema and
view integration (Nyerges 1991), and object-oriented analysis (Coleman et al 1994).
These methods have been applied to model semantic relationships, resolve user views,
and facilitate data sharing in building databases. Query about semantic information
has been the primary utility in DBMS via development of relational and object-oriented
data models with correspondent query languages (cf Date 1995).

2.2 Queries about Location, Spatial Properties, or Spatial Relationships


Spatial query support is one of the primary GIS functions for answering questions
about where and what (Peuquet 1994). Questions about `where' ask about spatial
properties of given geographic entities, such as `Where were clear-cuts in 1950?' In
contrast, questions about `what' search for geographic entities with certain spatial
properties, such as `What was the type of the forest patch at location X in 1800?'
Spatial queries may proceed on one or multiple GIS layers. A GIS layer contains spatial
objects of a single object type (cell, point, line, or polygon) and correspondent sets of
attribute data at locations defined by these spatial objects. The allocation of attribute
data to spatial objects simplifies GIS support for spatial queries on one layer. Queries
that need to access geographic information from multiple layers require overlay
operations (spatial joins) to unveil spatial relationships among geographic themes
(Tomlin 1990). Use of spatial overlays to produce new geographic information from
coded geospatial data through spatial analysis is one of the most salient features that
distinguish GIS from computer aided design (CAD), non-spatial database management
systems (DBMS), and computer mapping packages (Cowen 1988). With the
attachments of attribute data to locations, a GIS can establish spatial relationships
through overlay operations, compute the area and perimeter for each of the forest
patches, retrieve adjacent forest patches of a given patch, delineate buffer zones, and
identify forest patches within a specified area. Likewise, porting these capabilities to a
temporal dimension is important to support queries about temporal information in a
GIS.

2.3 Queries about Time, Temporal Properties, and Temporal Relationships


Temporal queries ask for information that evolves over time. Both GIS and
information scientists have directed considerable research effort toward temporal
databases and temporal query languages in the last decade (Tansel et al 1993).
Information scientists have proposed schemes to incorporate time into relational
databases by time-stamping a relation (a table, Figure 2a, Gadia and Vaishnav 1985),
individual tuples (ungrouped relations, Figure 2b, Snodgrass and Ahn 1985), or
individual cells (grouped relations, Figure 2c, Gadia and Yeung 1988). A detailed

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 141

Figure 2 Examples of representations of temporal information in a relational data model.

review of temporal data models is provided by Ozsoyoglu and Snodgrass (1995).


Because information scientists usually work with non-spatial data, their models
represent the evolution of semantic information about what and when changes occur to
identified entities. GIS researchers, however, have to work with spatial data. Data
modeling is a much greater challenge when changes occur in both space and time.
Temporal queries seek geographic information at a particular time, during a
period, or about temporal relationships among geographic entities. Langran (1992b)
characterized temporal queries as simple temporal queries (e.g. what is the state of a
feature at time t?) and temporal range queries (e.g. what happens to a feature over a
given period?). Examples of these queries include `What was the state of old growth in
1950?' and `How did the distribution of old growth change from 1800 to 1960?' In
addition to Langran's two types of temporal queries, temporal relationship queries are

ß Blackwell Publishers Ltd. 1999


142 May Yuan

Table 1 An example table for temporal information support and computing

Land cover type Area (hectares) Year

Old growth 2,234,500 1600


Old growth 2,145,200 1700
Clear-cut 337,000 1700
Burn 352,300 1700
Old growth 2,113,000 1800
Clear-cut 469,200 1800
Burn 352,300 1800
... ... ...
... ... ...

added here to include questions like `In which year did we have the largest area of
clear-cuts converting to old-growth cover from 1600 to 1950?' A simple search and
arithmetic computation can be performed on the data records in the table that includes
areas of old-growth forests for given years to reveal the changes to old growth during
the period of interest (Table 1). Challenges arise when area measurements are
unavailable and, hence, computations are required to resolve interactions among
geographic entities in both space and time. Capabilities to manage interactions between
spatial and temporal information components are pre-requisites for GIS support of
spatiotemporal queries.

2.4 Queries about Spatiotemporal Behaviors and Relationships

Support for spatiotemporal queries is at the heart of developing a temporal GIS that
will facilitate understanding and modeling of the dynamic geographic world. Langran
(1992b) characterized spatiotemporal queries as `simple spatiotemporal queries' (e.g.
what is the state of a region at time t?) and `spatiotemporal range queries' (e.g. what
happens to a region over a period?). A simple spatiotemporal query seeks information
about a state at a given time, such as `Where were the old-growth patches in 1800?'
Answers to this type of query are snapshots of an area. Comparatively, a
spatiotemporal range query asks questions about the history of a region over time,
such as `How did the land-cover composition change in the area from 1600 to 1800?
(Figures 1a±c)'. Responses can be a loop of snapshots with tables describing changes at
locations in the area.
Complementary to Langran's two types of spatiotemporal queries that only look
for static information of histories at locations, this study also adds `spatiotemporal
behavior queries' to stress the need for dynamic information describing behaviors of
events and processes. Despite the importance of handling events and processes in a
temporal GIS (Peuquet 1994), current GISs lack methods to support queries about
dynamic characteristics of geographic entities, such as frequency of fire occurrences,
expansion of clear-cuts, and rate of decrease in old growth coverage. This type of query
requires tracing the evolution of a given entity, i.e. to follow an entity in space and time
and aggregate summaries of correspondent GIS data spatially and temporally. This is
relatively simple in aspatial databases since every entity is represented by the same data

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 143

object with a persistent identifier throughout the course of the entity's existence. In a
GIS, however, changes can occur to both semantic and spatial characteristics. As a
result, identities of geographic meanings and spatial representations are likely to vary
throughout the evolution of a GIS database. A space may continue being fragmented
when change overlaps in space and time. Subsequently, spatial object identifiers will
also change via the fragmentation. Over time, a GIS database has a different set of
spatial objects with new identifiers to represent the entity. Although abilities to
maintain persistent spatial object identifiers in representing an entity are one of the
primary issues in the development of a temporal GIS (Langran 1992a, Peuquet and
Duan 1995, Raper and Livingstone 1995), many GIS data models fail to provide this
important function. Hence, a GIS with data of area burnt by multiple fire events cannot
easily answer queries like `When and where did the fire events reach their maximal rate
of spread?' Later, it will be shown that a three-domain representation can maintain
persistent identifiers for spatial objects (Section 3) and incorporate computational
procedures to solve a sample query of this type through computation of a change
matrix (Sections 4.2 and 4.3).
Numerous spatiotemporal representations or indexing schemes have been
proposed in the GIS literature (see Abraham and Roddick 1996 for details), though
all have a certain degree of difficulty in supporting complex spatiotemporal queries.
Parallel to the aforementioned three time-stamping relational database approaches
(Table 1 and Section 2.3), temporal information has been incorporated into GIS data
models by time-stamping layers (the snapshot models, Figure 3a, Armstrong 1988),
attributes (space-time composites, Figure 3b, Langran and Chrisman 1988), and spatial
objects (spatiotemporal objects, Figure 3c, Worboys 1992). Armstrong's (1988)
snapshot model has problems supporting questions about fire frequency when more
than 50 snapshots need to be overlaid. Langran and Chrisman's (1988) space-time
composite model can support the computation of frequency; however, it encounters
two major problems with information computing. First, the model provides
unidirectional mappings from space-time composites to time and semantics (i.e.
geographic meanings of these space-time composites) because information has to be
associated with spatial representations of space-time composites. As long as locations
are disjoint, the space-time composite model represents them with two distinct records,
even if the two locations are footprints of the same process. This result causes
difficulties when supporting complex semantics- and time-based queries, such as `How
did the rate of spread of fire A change from t1 to t2?' The second major problem of the
space-time composite model is the need for re-structuring space-time composites when
new spatiotemporal information overlaps with the existing space-time composites
(Langran and Chrisman 1988, Langran 1992a), which results in spatial objects without
persistent identifiers.
The spatiotemporal object model proposed by Worboys (1992) is a three-
dimensional version of the space-time composite model, but it is able to maintain
persistent object identifiers by stacking additional spatiotemporal objects on the top of
existing ones instead of overlaying them to resolve new sets of space-time composites
(cf Figure 3c). Yet, the spatiotemporal model sill encounters problems of unidirectional
mappings from spatiotemporal objects to semantic information in that information has
to be directly or indirectly related to spatiotemporal atoms. Hence, information cannot
be recorded unless it is associated with both space and time. However, many kinds of
information are non-spatial, such as vegetation taxonomies or relationships between

ß Blackwell Publishers Ltd. 1999


144 May Yuan

Figure 3 Examples of representations of spatiotemporal information in a GIS


environment.

fire behavior and fuel contents. In addition, unidirectional mappings from space (or
space and time) to semantics also constrain support for queries about spatiotemporal
behaviors of events or processes because information has to be tied to locations and
therefore is unable to represent motion and processes (see Sections 4.2 and 4.3 for
further discussion of these issues).
Peuquet and Duan's (1995) event-based spatiotemporal data model (ESTDM) and
Raper and Livingstone's (1995) OOgeomorph support the mappings from semantic to
spatial objects, but these models are specifically designed for cell- and point-based
representations, respectively. While the raster-based model, ESTDM, outperforms the

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 145

snapshot model in reducing data volume and increasing computational efficiency, it


requires significant revisions to be adaptable to a vector framework (Peuquet and Duan
1995). Similarly, OOgeomorph is a point-based system designed for modeling change
of coastlines. The system is inapplicable in modeling other spatiotemporal information
of higher dimensions as needed in the forest-transition example.
Spatiotemporal information extends beyond simple snapshots or homogeneous
changes at a locality (space-time composites or spatiotemporal objects) and should
allow line- and polygon-based representations (rather than being limited to a particular
representation like ESTDM and OOgeomorph). GIS representations should be
dynamic and allow for bi-directional mappings among semantics and space through
time to portray location- or entity-based information. Location-based information
includes changes in semantics (geographic meanings) of spatial objects (spatial
properties of location, shape, geometry remain constant), while entity-based
information emphasizes changes of an entity or a process in space and time. A
dynamic GIS representation needs to account for semantic, spatial, and temporal
heterogeneity throughout the existence of an entity.

3 A Three-domain Representation to Support Spatiotemporal Query in GIS

Yuan (1996) proposed the conceptual framework of a dynamic three-domain


representation that departs from the static-map metaphor. Figure 4 shows the
conceptual structure of the three-domain representation consisting of semantic objects,
temporal objects, spatial objects, and domain links. Temporal and spatial objects
represent the temporal and spatial properties of geographic semantics (including
themes, entities, events, and processes) that a GIS database is designed to convey.
Domain links may be one of three types: (1) pointers that associate objects of different
domains to represent spatiotemporal facts, (2) mathematical functions that indicate
spatiotemporal behaviors, and (3) physical models that predict trends or processes in
space and time. This section illustrates the first implementation of the three-domain
framework. While the three-domain representation can be implemented in either geo-
relational structures or, probably better, in object-oriented frameworks, the geo-
relational approach is used here for its structural simplicity.

3.1 The Semantic Domain


Table 2a lists the semantic objects throughout the forest transitions. Semantic objects
represent entity classes (i.e. old-growth, clear-cut, and fire burn) with unique identifiers
throughout the transitions. Identification of semantic objects is subject-dependent, and
separate tables are necessary for objects of different relations to maintain semantic
integrity. However, the forest example contains only one semantic table because these
semantic objects share a common set of attributes. The table is extendable to multiple
tables when more information regarding these entities are available.

3.2 The Temporal Domain


Temporal objects included in the transitions are five time points from T1 to T5 (Table
2b). Generally, time is recorded in a database as an attribute of data objects. Instead,

ß Blackwell Publishers Ltd. 1999


146 May Yuan

Figure 4 A conceptual framework for a three-domain model.

the three-domain representation handles time as identifiable objects with attributes


(e.g. measurement units) and relationships (e.g. transformation rules). The
representation assumes that no sudden changes occur from Tx to Ty unless the
semantic or spatial object has a link to another temporal object between Tx and Ty. For
gradual and continuous changes, semantic objects can associate with proper
interpolation functions to estimate variations between Tx and Ty.. Using interpolation
functions, the model can infer any state of the geographic entities by projecting these
entities to space at a time of interest. Temporal objects include points (instants as
defined in Freska 1992) and lines (intervals as defined in Allen 1983). Temporal
structures and relationships elaborated in Snodgrass and Ahn (1986), Freska (1992),
and Worboys (1990) are easily handled in the temporal domain.

3.3 The Spatial Domain


The spatial domain consists of spatial objects representing composite spatial
configurations of the study area over time, similar to Langran and Chrisman's
(1988) Space-Time Composite Model. However, in order to maintain persistent
identifiers for spatial objects, the spatial domain includes a spatial graph to record
transitions of spatial objects so that we can trace spatial object 5 as an aggregation of
spatial objects 8, 9, 13, 14, and 15 (Figure 5e). The shaded nodes in the spatial graph
represent the most current spatial configuration in the area; these are the simple spatial
objects included in the space table (Table 2c). Only spatial objects in the latest spatial
configuration are stored in the domain with coordinates, geometry, and topology.
Geometrical and topological information is manageable using layer-table structures
(i.e. a layer with an attribute table as used in ARC/INFO) or layer-object frameworks

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 147

Table 2 Use of the three-domain model to represent spatiotemporal information about


the sample forest transitions. The space table contains the current configuration and the
spatial graph in Figure 5e is used to complete the three-domain representation.

a. Semantics Table b. Time Table

Sem. ID Landcover Management Address Time ID Time Operator ID

1 Old Growth USFS 12 Forest Rd. 1 1600 2439


2 Clear-cut A. Log Co. 3 Clear Dr. 2 1700 2439
3 Burn USFS 12 Forest Rd. 3 1800 7473
4 Clear-cut B. Log Co. 45 Pine Ave. 4 1950 1029
5 1960 1029

d. Domain Link Table (Links among temporal,


c. Space Table semantic, and spatial objects)

Space ID Area Perimeters Sem. ID Time ID Space ID List

4 A1 P1 1 1 1
6 A2 P2 1 2 2
8 A3 P3 2 2 3
9 A4 P4 3 2 4
10 A5 P5 1 3 5
11 A6 P6 2 3 3, 6
12 A7 P7 1 4 7, 10
13 A8 P8 4 4 8, 9
14 A9 P9 1 5 10, 11, 13
15 A10 P10 2 5 6, 12
3 5 4, 14, 15

The table lists semantic associations among


spatial objects. Every data record represents a
combination of spatial objects associated with a
particular semantic object. A new data record
will be added when the association has been
changed for any semantic object. For example,
spatial object 4 is associated with semantic
object 3 at T2, and the semantic association
held until T5 when the semantic object 3
should relate to spatial objects 4, 14, and 15.
Further, retrieving and aggregating all spatial
The spatial graph records objects associated with Time ID  4 can derive
parenthood transitions the state of the forest at T4.
among spatial objects in
Space Table.

ß Blackwell Publishers Ltd. 1999


148 May Yuan

(i.e. a layer with a set of geographic objects as used in SmallWorld). Aggregated spatial
objects are excluded from the space table, but they can be derived by tracing the spatial
graph. For example, the spatial graph shows that spatial object 3 is an aggregation of
spatial objects 11 and 12. The object's geometrical and topological properties can be
computed based on its component objects 11 and 12. The structure of aggregate and
simple spatial objects enables representations of disjoint as well as overlapping entities
in a GIS. The aggregation is similar to the concepts of Dynamic Segmentation and
Regions in ARC/INFO GIS, but the spatial graph's ability to: (1) map locations to
semantic objects with both persistent spatial object identifiers, and (2) maintain
transitions among spatial objects distinguishes the spatial domain from these other two
spatial data models.

3.4 Domain Links


A domain-link table (Table 2d) illustrates the links among semantic, temporal, and
spatial objects, and it establishes semantic associations among spatial objects. Semantic
associations connect spatial objects with comparable geographic meanings. For
instance, all spatial objects of old-growth patches are linked to the same semantic
object (Old-growth). This tabular structure is in fact inefficient in handling one-to-
many relationships among semantic and spatial objects. An alternative for organizing
links among the three types of objects is an object-oriented framework. Table 2d is
used here because of its structural simplicity.
The semantic domain (Table 2a), temporal domain (Table 2b), spatial domain
(Table 2c and Figure 5e), and domain links (Table 2d) provide the basic references to
infer spatiotemporal information about forest transitions. Table 2d incorporates bi-
directional mappings (from space, time to semantics or vice versa) among the three
domains to represent both location- and entity-based information. For example, it
represents a location (spatial object 3) that was clear-cut by logging (semantic object 2)
in 1700 (temporal object 2). It also indicates that logging resulted in four patches
(spatial objects 3, 6, 8, and 9) up to 1950. In addition, the table suggests that there was
only one burn (semantic object 3) in 1700 (temporal object 2), thereafter no fires
occurred until 1960 (temporal object 5), and two burns were recorded (at spatial
objects 14 and 15) in 1960. From a technical perspective, the three-domain
representation is a normalization of current spatiotemporal GIS data models because
it eliminates repetitive data records in spatial attribute tables for snapshots, space-time
composites, and spatiotemporal objects by only using a single attribute record for
separated locations with common properties. The order of normalization is a function
of spatiotemporal heterogeneity. When every spatial object has unique attributes and
time limits, the three-domain representation will be at the same order of normalization
as the spatiotemporal object model. In addition to the technical aspects, the three-
domain representation provides a conceptual alternative that enables a GIS to represent
and analyze geographic information for motion and other spatiotemporal behaviors via
mappings from entities and processes to locations at different times. Moreover, the
three-domain representation is readily extendable to 3D space and 1D time because it
handles space and time separately. This extension can provide an important framework
for spatiotemporal queries about processes, such as smoke dispersion and geochemical
diffusion.

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 149

Figure 5 Transitions in a spatial domain, and a spatial graph. Only the latest spatial
configuration at T5 is stored in the spatial domain. The spatial graph is used to represent
transitions among spatial objects from T1 to T5. Shaded leaves represent spatial objects in
the latest spatial configuration and are included in its correspondent space table (Table
2c). The transitional configurations (T1 to T4) can be deduced from relationships of spatial
objects encoded in the spatial graph.

4 Spatiotemporal Query Support: Comparison of Spatiotemporal


Representations in GIS

All five prominent spatiotemporal representations discussed (Section 2.4) have distinct
strengths in spatiotemporal information modeling, but they, like the three-domain
representation, also have limitations. The following comparison aims to evaluate their
support for complex queries on higher-level information because all of them can handle
simple queries. In contrast to simple queries that retrieve data records stored in a
database, complex queries require computation (including aggregation and
decomposition) based on stored data records to produce higher-level information.
The comparison excludes ESTDM and OOgeomorph because of their constraints on
specific data structures (Section 2.4). Hence, the following sections compare the three-
domain representation with snapshots, space-time composites, and spatiotemporal
objects using the sample forest transitions.

4.1 The Snapshot and Three-domain Representations


The snapshot model (Figure 3a) is the simplest way to represent spatiotemporal
information, but its capability to support complex spatiotemporal query is the most

ß Blackwell Publishers Ltd. 1999


150 May Yuan

limited. A snapshot represents a state at a given time. Similar to time-stamped


relational tables (cf Figure 2a), there is no temporal relation built among spatial states.
Therefore, the snapshot model has limited support for queries that require accessing
data from multiple snapshots. For example, it is difficult for the snapshot model to
resolve `Where is the largest burn that occurred from 1600 to 1960?' Computation
based on the snapshot model requires retrieval of all burns from every snapshot and
comparisons of the areal coverage of these burns. Such operations will become
unmanageable when the number of snapshots is large (say, more than 50 years of forest
fire layers). Even with automatic scripts, the operations still require significant
computer time and memory to hold retrieved objects from multiple layers when the
system is opening another snapshot for search. Essentially, the snapshot approach
inevitably needs to scan all spatial objects on all snapshot layers to ensure an
exhaustive search for fire burns. Because snapshots duplicate spatial objects that
remain unchanged from one snapshot to another, an introduction of a new snapshot
will at least double the number of spatial objects from its previous snapshot. The
search time increases in proportion to the total number of spatial objects on all
snapshots.
Using the three-domain representation to answer the query about the largest burn
recorded in the database is straightforward. As shown in Figure 6, the three-domain
model reduces the total number of spatial objects in two ways. First, it normalizes data
tables to eliminate duplicated semantic information by storing data about change
instead of the entire snapshot. Second, it enables retrieval of the semantic object of `fire
burn' from the semantic table and use of the selected semantic ID to eliminate spatial
objects that need to be retrieved from the space table. Hence, the search in the three-
domain model proceeds only on those spatial objects associated with `fire burns.'

Procedures Results

(1) Select Landcover = `Burn' Sem ID Landcover Management Address


from SEMANTICS TABLE 3 Burn USFS 12 Forest Rd.

(2) Select Sem ID = 3 Sem ID Time ID Space ID List


from DOMAIN LINK TABLE 3 2 4
3 5 4, 14, 15

(3) Select Space ID = 4, 14, or 15 Space ID Area Perimeter


from SPACE TABLE 4 (value) (value)
14 (value) (value)
15 (value) (value)

(4) Calculate Max (Area) Space ID Area Perimeter


4 (value) (value)

Figure 6 Procedures to compute the largest burn in the database based on the three
domain representation of the sample forest transitions (Table 2).

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 151

Because the number of spatial objects with fire burns is much less than the total
number of spatial objects across all snapshots, an increase in temporal data (years of
forest transitions) has much less impact on the search time.

4.2 The Space-Time Composite Model and Three Domain Model


Langran and Chrisman's (1988) space-time composite model decomposes a space over
time into increasingly smaller fragments to identify the maximum number of common
spatiotemporal homogeneous units, each with a distinct attribute history (Figure 3b).
The basis of the space-time composite model is a categorical map designed for an
attribute that exhausts the space available with a single classification. Hence, a space-
time composite model cannot express multiple attributes at a location, such as
population and average household income at T1. It needs look-up tables or indexing
schemes for additional semantic or temporal information, which is essentially the idea
of the three-domain model. Furthermore, the space-time composite model, inherited
from categorical maps, is static and spatially exhaustive. As a result, the model cannot
represent processes and overlapping entities.
Another disadvantage of the space-time composite model is semantic information
duplication that provides good support for location-based query but hampers query
support for entity-based information. Two disconnected clear-cut patches, for
example, will require two separate data records to store their attributes even if they
have identical histories. The three-domain model, in contrast, applies semantic
associations to eliminate redundancy by linking spatial objects with the same
geographic meanings to one single semantic description. To query histories at
locations, the three-domain representation provides links from spatial objects
(locations) and temporal objects to semantic objects to resolve locational histories.
The mappings are essentially equivalent to space-time composites.
The space-time composite model can support Langran's (1992b) six basic query
types in space and time (Section 2). However, it still has difficulty in facilitating queries
about spatiotemporal behaviors and relationships proposed in this paper, such as
`What is the rate of areal increase in clear-cuts from 1600 to 1950?' Figure 7 illustrates
an example using space-time composites to represent the sample forest transitions. The
model represents times (1600, 1700, etc) as data fields (attribute names) used by the
standard query language (SQL) to specify the attribute of interest in a selection, such as
AREA in `select AREA > 500.' As a result, the space-time composite model requires
separated Boolean operations to examine if the time is within the search period of
interest (ie from 1600 to 1950). GIS users have to examine proper data fields and repeat
selections individually on these fields (select 1600 = `Clear-cut', select 1700 = `Clear-
cut', select 1800 = `Clear-cut', etc) until the value of a data field is beyond the
temporal range of interest (>1950). The examination is difficult, if not impossible, in
an automatic fashion using SQL because data fields are distinct variable names and
cannot be used as values for arithmetic comparison with 1950. When the number of
times stored in a GIS is large (say 50 different times), the operation becomes impractical
to manually check all 50 data fields in the space-time composites before calculation of
rate of change in spatial coverage can proceed. Use of temporal query languages cannot
ease this search on space-time composites because temporal query languages also
require time to be stored as data values (as shown in Figure 2b and c and not as data
fields (cf Tansel et al 1993).

ß Blackwell Publishers Ltd. 1999


152 May Yuan

Polygon ID Area 1600 1700 1800 1950 1960

1 A4 Old-growth Burn Burn Burn Burn


2 A6 Old-growth Old-growth Clear-cut Clear-cut Clear-cut
3 A8 Old-growth Old-growth Old-growth Clear-cut Clear-cut
4 A9 Old-growth Old-growth Old-growth Clear-cut Clear-cut
5 A10 ± ± ± Old-growth Old-growth
6 A11 Old-growth Clear-cut Clear-cut Clear-cut Old-growth
7 A12 Old-growth Clear-cut Clear-cut Clear-cut Clear-cut
8 A13 Old-growth Old-growth Old-growth Old-growth Old-growth
9 A14 Old-growth Old-growth Old-growth Old-growth Burn
10 A15 Old-growth Old-growth Old-growth Old-growth Burn

Figure 7 Use of Space-Time Composite Model (Langran and Chrisman 1988) to represent
spatiotemporal information about the sample forest transitions. The polygon IDs in
parentheses are identical to the space IDs in Figure 5e and Table 2c&d for comparison
with the three-domain model.

Comparatively, the three-domain model supports queries about the rate of change
in a spatial coverage of clear-cuts from 1600 to 1950 (Figure 8). Based on this database,
there was no clear-cut until 1700 in this area. After 1700, R1 and R2 are computed for
the rates of areal changes in clear-cuts from 1700 to 1800 (T2 to T3) and 1800 to 1950
(T3 to T4), respectively. The sample query illustrates the fundamental difference
between the three-domain model and the space-time composite model beyond simple
normalization. The difference in query support is due to the separation of semantic,
spatial, and temporal information in the three-domain model. The separation enables
information access by specifying criteria of semantics, space, or time. While simple
spatial objects in the `latest spatial configuration' (Figure 5e) are identical to the spatial
representation in the space-time composite model (Figure 7), these simple spatial
objects are not used to define histories as they are in the space-time model. Instead, the
three-domain model links spatial and temporal objects to semantic objects to represent
histories at locations. Alternatively, the model links semantic and temporal objects to
spatial objects to describe spatiotemporal behaviors of a geographic process. The
spatial graph enables the domain link table to provide direct and persistent mappings
between semantics and space at all times; for example, the record of a clear-cut at
location 3 in 1700 remains unchanged regardless of how the space may be further
fragmented in a later time. Therefore, the model is able to represent overlapping
entities by having two entities linked to aggregate spatial objects that share common
simple spatial objects. The model can also represent a spatially disjoint entity by

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 153

Procedures Results

(1) Select Landcover = `Clear-cut' Sem ID Landcover Management Address


from SEMANTICS TABLE 2 Clear-cut A. Log Co. 3 Clear Dr.
4 Clear-cut B. Log Co. 45 Pine Ave.

(2) Select 1600  Time  1950 Time ID Time Operator ID


from TIME TABLE 1 1600 2439
2 1700 2439
3 1800 7473
4 1950 1029

(3) Select (SemID = 2 or 4) Sem ID Time ID Space ID List


AND (TimeID < 5) from 2 2 3
DOMAIN-LINK TABLE 2 3 3, 6
4 4 8, 9

(4) Search SpaceID = 3 in SPATIAL GRAPH

(5) Select SpaceID = 6, 8, 9, 11, or 12 Space ID Area Perimeter


from SPACE TABLE 6 A6 P6
8 A8 P8
9 A9 P9
11 A11 P11

(6) Calculate Sem ID Time ID Area


Area (SemID=2, TimeID=2)=A3=Sum (A11, A12) 2 2 Sum (A11, A12)
Area (SemID=2, TimeID=3)=Sum (A3, A6) 2 3 Sum (A3, A6)
Area (SemID=4, TimeID=4)=Sum (A8, A9) 4 4 Sum (A8, A9)

(7) Calculate
Rate (Area(Sum(A11, A12), Sum(A3, A6)) Time (Time(2), Time(3)) = R1
Rate (Area(Sum(A3, A6), Sum(A8, A9)) Time (Time(3), Time(4)) = R2

Functions
Area (SemID, TimeID): function that aggregates area of spatial objects associated with the
specified SemID and TimeID in a domain-link table.
Sum (X1, X2, . . .): function that calculates addition of enclosed values (X1, X2, etc.).
Time (x, y): function that calculates the difference between time x and time y based on
records in a time table, including necessary transformation of time units.
Area (x, y): function that calculates the difference between time x and time y based on
records in a space table, including necessary transformation of areal units.
Rate (x, y): function that calculates the rate of x over y.

Figure 8 Procedures to compute the rate of change in clear-cuts from 1600 to 1950 in
the database based on the three-domain representation of the sample forest transitions
(Table 2).

ß Blackwell Publishers Ltd. 1999


154 May Yuan

linking the entity to two separate spatial objects. Likewise, the mappings allow
retrieval of all burns occurring at a location to support queries about histories at
locations as supported by the space-time composite model.

4.3 The Spatiotemporal Object Model and Three-domain Model


The spatiotemporal object model incorporates a temporal dimension orthogonal to a
2D space to model the world by spatiotemporal objects (ST-objects) and
spatiotemporal atoms (ST-atoms). Similar to space-time composites, ST-atoms are
the largest homogeneous unit in which properties hold in both space and time. While
spatiotemporal composites have a common set of time steps, every ST-atom has a
distinct temporal range independent of the others. ST-atoms form ST-objects to
represent complex geographic entities or phenomena, similar to the way in which
dynamic segmentation is used to represent a highway with sections of different
authority or maintenance schedules. Such an atom-object structure enables the ST-
object model to represent variations in attributes over space and time (Worboys 1992).
Therefore, it is a more flexible representation than the space-time composite model. In
addition, the atom-object structure maintains persistent identifiers for both ST-objects
and ST-atoms by stacking new ST-atoms upon existing ones.
Figure 3c shows an example of using the ST-object model to represent regional
change from t1 to t3. In the initial state, t1, there were two landuse types: agriculture
and urban. At t2, a portion of agricultural land is converted to an urban area, and at t3,
part of the urban area became an industrial zone. Instead of overlaying all three
snapshots at t1, t2, and t3 to form a space-time composite layer, the ST-object model
has ST-atoms to represent spatially and temporally homogeneous landuse blocks of
agriculture (S1 and S2), urban (T1, T2, and T3) and industry (U). These ST-atoms form
three ST-objects: S, T, and U, to represent change of a landuse type in space and time.
For example, ST-object S consists of ST-atoms S1 and S2 for agricultural land that
decreased from S1 to S2 during the period from t1 to t2. Hence, a ST-atom is defined by
a location (eg S1) and a time interval (eg t1 to t2). Semantic associations among ST-
atoms with related properties (ie the same landuse type) form ST-objects, which are
aggregations of ST-atoms across space over time. The atom-object structure in the ST-
object model is comparable to semantic associations among spatial objects encoded in
the domain link table (Table 2d). With semantic associations, both the three-domain
and spatiotemporal object models are able to represent separate or overlapping areas
by associating a semantic object or ST-object (eg a fire) with multiple isolated spatial
objects or ST-atoms.
The spatiotemporal object model has two drawbacks. First, it duplicates
coordinates of adjacent atoms, such as the common boundaries among ST-atoms 3,
11, and 12. Because space and time have distinct measurements, integrated topological
relationships for space and time are unavailable to support encoding common
boundaries and topology among ST-atoms. Consequently, the model cannot apply the
space-filling principle to maintain spatial data integrity, which may cause problems in
error checking of spatial data (Langran 1992b). In contrast, the three-domain model,
like the space-time composite model, is a topological model in which adjacent spatial
objects share common boundaries. With a spatial graph, it also maintains transitions of
spatial objects. Lack of such transitions among spatial objects encoded in the database
is in fact the second drawback of the spatiotemporal object model.

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 155

ST-atom ID Area Land cover type St-object ID

1 A! [1600±1700] Old-growth 1
2 A2 [1700±1800] Old-growth 1
3 A3 [1700±1950] Clear-cut 2
4 A4 [1700±1960] Burn 3
5 A5 [1800±1960] Old-growth 1
6 A6 [1950±1960] Clear-cut 4
7 A7 [1950±1960] Old-growth 1
8 A8 [1950±1960] Clear-cut 2
9 A9 [1950±1960] Clear-cut 4
10 A10 [1950±1960] Old-growth 1
11 A11 [1960±1960] Old-growth 1
12 A12 [1960±1960] Clear-cut 2
13 A13 [1960±1960] Old-growth 1
14 A14 [1960±1960] Burn 5
15 A15 [1960±1960] Burn 6

Figure 9 Use of spatiotemporal Object Model (Worboys 1992) to represent


spatiotemporal information about the sample forest transitions. ST-atoms hold constant
properties in space and time. Formation of ST-objects is based on spatiotemporal contiguity
of ST-atoms with the same land cover in the forest transitions (cf Figures 3c and 5).

Without the transitions, the spatiotemporal object model has difficulty in


supporting queries for semantic relationships at locations over time, such as `Based
on the database, how much clear-cut area returned to old-growth in the last 100 years?'
as asked in 1997. Figure 9 shows an example of using the ST-object model to represent
information about the sample forest transitions. Using the ST-object model, the
computation of the information requested by the above query requires search through
the life span of each ST atom for records with starting time and ending time that
include the year of 1897 (= 1997ÿ100). The result includes all ST objects with ID
greater or equal to 5 (Figure 9). Next, it is necessary to determine which ST atoms were
clear-cuts but later became old growth. The information about change from one land-
cover type to another is implicit because every ST object is of a single land-cover type,
and there is no transitional information about how two ST-objects relate over time.
Hence, the next step needs to retrieve clear-cut ST atoms (6, 8, 9, and 12) from the
selected set (ie ST-atoms with ID greater or equal to 5). The following selection
retrieves old-growth ST atoms from the same pool; it selects ST atoms 5, 7, 10, 11, and
13. The next step is to spatially overlay the selected old-growth ST atoms with the
selected ST-atoms of clear-cuts to reveal the overlaid area between the two groups. The
last step is then to compare the life span of the old-growth and clear-cuts in the
overlaid area. The answer of this query is the area recorded with a later life span of old-
growth. This is the most difficult step because the computation involves spatial overlay

ß Blackwell Publishers Ltd. 1999


156 May Yuan

Procedures Results

(1) Select Landcover = Sem ID Landcover Management Address


(`Old Growth' or `Clear-cut') 1 Old Growth USFS 12 Forest Rd.
from SEMANTICS TABLE 2 Clear-cut A. Log Co. 3 Clear Dr.
4 Clear-cut B. Log Co. 45 Pine Ave.

(2) Select Time  (1997±100) Time ID Time Operator ID


from TIME TABLE 4 1950 1029
5 1960 1029

(3) Select (SemID = 1 or 4) and (TimeID = 4 or 5) Sem ID Time ID Space ID List


from DOMAIN-LINK TABLE 1 4 7, 10
1 5 10, 11, 13

(4) Search Parent (SpaceID = 7, 10, 11, or 13)


in SPATIAL GRAPH

(5) Select (SpaceID =0, 3, 5, or 7) and Semi ID Time ID Space ID List


(Sem ID = 2 or 4) from DOMAIN-LINK TABLE 2 3 3, 6

(6) Search Child (SpaceID = 3)


in SPATIAL GRAPH

Functions
Parent (X1, X2, . . .): function that traces a spatial graph to retrieve parent objects of X1, X2, etc.
Child (X1, X2, . . .): function that traces a spatial graph to retrieve parent objects of X1, X2, etc.
Figure 10 Procedures to compute the area that has changed from clear-cut to old-
growth forest in the last 100 years as asked in 1997.

of selected objects, temporal overlay of their associated life span, and determination of
which area has attributes of clear-cuts before old growth.
Computation of the requested information is comparably straightforward using
the three-domain model (Figure 10). The search starts by identifying semantic objects
representing old-growth or clear-cut patches in the semantics table (Table 2a) and
temporal objects that represent a later time than 1897 (= 1997±100) in the time table
(Table 2b). The remaining question becomes `find all spatial objects that are: (1) linked
to selected temporal objects and semantic objects of old-growth in the domain link
table (Table 2d); and (2) linked to spatial objects associated with the semantic object of
old-growth in the spatial graph (Figure 5e).' As a result, the search resolves that spatial
object 11 was a clear-cut in 1800 (as part of spatial object 3 with temporal object 3), but

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 157

it became an old-growth area in 1960 (temporal object 5). According to the database,
spatial object 11 is the only location fulfilling the search criteria.

5 Conclusions

This study has categorized the basic types of spatiotemporal queries and applied a
three-domain framework to facilitate these queries. It has also compared the
framework with other prominent spatiotemporal representations in GIS. The three-
domain representation consists of both semantic associations and transitions among
spatial objects, so that it provides bi-directional mappings to enable both entity-based
and location-based query processing.
With the bi-directional mappings, the three-domain representation combines
advantages from the space-time composite and spatiotemporal object models. Both the
three-domain and spatiotemporal object models incorporate semantic associations to
aggregate spatial objects with the same geographic meaning (semantics). The three-
domain representation uses domain links to associate spatial objects with the same
properties in a snapshot regardless of their spatial adjacency. Similarly, the
spatiotemporal object model uses the atom-object structure to combine spatiotemporal
atoms to facilitate queries about geographic entities. Without semantic association, the
space-time composite model is unable to support entity-based queries, as discussed in
Section 4.2. In contrast, both the three-domain and space-time composite models
maintain spatial transitions of locational properties. While the three-domain model
holds the transitions in a spatial graph with domain links to semantic objects
describing geographic meanings at locations, the space-time composite model records
locational histories at the largest, homogeneous spatial units. Therefore, both models
support queries about changes at locations. Without transitions among spatial objects,
the spatiotemporal object model does not support queries about histories at locations,
as illustrated by the sample query in Section 4.3. Because the snapshot model lacks both
semantic associations and transitions among spatial objects, its capability to support
complex spatiotemporal query is the most limited. Moreover, the three-domain model
is able to represent spatially distributed features (eg discrete forest patches), processes
(eg storms), and movement (eg spread of a fire) using links among semantic, temporal,
and spatial objects. The dynamic nature of the three-domain representation
distinguishes the model from other spatiotemporal representations in GIS.
This study has demonstrated the ability of the three-domain model to improve GIS
query processing with conventional SQL structures. Such query support is critical for a
GIS to provide information for scientific analysis and modeling. Hitherto, query
processing and optimization for GIS databases has been relatively underdeveloped
(Samet and Aref 1995). The three-domain model provides an alternative scheme to
enhance GIS query support through not only technical improvements in data
normalization, but conceptual advancements in both location-based and entity-based
query processing in a dynamic framework.

Acknowledgements

Financial support for this project from the National Imagery and Mapping Agency
(NIMA) through the University Research Initiative Grant NMA202-97-1-1024 is

ß Blackwell Publishers Ltd. 1999


158 May Yuan

gratefully acknowledged. The author also thanks Dr John Wilson and three
anonymous referees whose comments have significantly improved this paper.

References
Abraham T and Roddick J F 1996 Survey of Spatiotemporal Databases. Adelaide, University of
South Australia, School of Computer and Information Science, Advanced Computing
Research Centre Technical Report CIS-96-011
Allen J F 1983 Maintaining knowledge about temporal intervals. Communications of the ACSM
26: 832±43
Armstrong M P 1988 Temporality in spatial databases. In Proceedings of GIS/LIS'88, 2.
Bethesda, MD, American Congress of Surveying and Mapping: 880±9
Burrough P A and Frank A U 1995 Concepts and paradigms in spatial information: Are current
geographical information systems truly generic? International Journal of Geographical
Information Systems 9: 101±16
Chen P P 1976 The entity-relationship model toward a unified view of data. ACM Transactions
on Database Systems 1: 9±38
Codd E F 1980 Data models in database management. In Brodie M L and Zilles S N (eds)
Proceedings of the Workshop on Data Abstraction, Databases, and Conceptual Modeling,
Pingree Park, Colorado, June 23±26 1980: 112±4
Coleman D, Arnold P Bodoff, S, Dollin C, Gilchrist H, Hayes F, and Jeremes P 1994 Object-
Oriented Development: The Fusion Method. Englewood Cliffs, NJ, Prentice Hall
Cowen D J 1988 GIS versus CAD versus DBMS: What are the differences? Photogrammetric
Engineering and Remote Sensing 54: 1551±5
Date C J 1995 An Introduction to Database Systems, 6th edition. Reading, MA, Addison-Wesley
Freska C 1992 Temporal reasoning based on semi-intervals. Artificial Intelligence 54: 199±227.
Gadia S K and Vaishnav J H 1985 A query language for a homogeneous temporal database. In
Proceedings of the ACM Symposium on Principles of Database Systems: 51±6
Gadia, S K and Yeung C S 1988 A generalized model for a relational temporal database. In
Proceedings of ACM SIGMOD International Conference on Management of Data: 251±9
Goodchild M F 1992 Geographical information science. International Journal of Geographical
Information Systems 6: 31±45
Hull R and King R 1987 Semantic database modeling: Survey, applications, and research issues.
ACM Computing Surveys 19: 201±60
Langran G 1991 Producing answers to spatial questions. In Auto-Carto 10: Technical Papers of
the 1991 ACSM-ASPRS Annual Convention. Bethesda, MD, American Congress of
Surveying and Mapping: 133±47
Langran G 1992a States, events, and evidence: The principle entities of a temporal GIS. In
Proceedings: GIS/LIS'92. Bethesda, MD, American Congress of Surveying and Mapping:
416±25
Langran G 1992b Time in Geographic Information Systems. London, Taylor and Francis
Langran G and Chrisman N R 1988 A framework for temporal geographic information.
Cartographica 25: 1±14
Morehouse S 1990 The role of semantics in geographic data modeling. Proceedings of the 4th
International Symposium on Spatial Data Handling, Zurich, Switzerland, 2: 689±98
Nyerges T L 1991 Geographic information abstractions: Conceptual clarity for geographic
modeling. Environment and Planning A, 23: 1483±99
Peuquet D J 1994 It's about time: A conceptual framework for the representation of temporal
dynamics in geographic information systems. Annals of the Association of American
Geographers 84: 441±62
Peuquet D J and Duan N 1995 An event-based spatiotemporal data model (ESTDM) for
temporal analysis of geographical data. International Journal of Geographical Information
Systems 9: 7±24
Ozsoyoglu G and Snodgrass R T 1995 Temporal and real-time databases: a survey. IEEE
Transactions on Knowledge and Data Engineering 7: 513±33

ß Blackwell Publishers Ltd. 1999


Complex Spatiotemporal Queries 159

Raper J and Livingstone D 1995 Development of a geomorphologic spatial model using object-
oriented design. International Journal of Geographical Information Systems 9: 359±84
Samet H and Aref W G 1995 Spatial data models and query processing. In Kim W (ed) Modern
Database Systems: The Object Model, Interoperability, and Beyond. New York, ACM
Press: 338±60
Snodgrass R and Ahn I 1986 Temporal databases. IEEE Computer, September, pp. 35±42
Sowa J F 1984 Conceptual Structures: Information Processing in Mind and Machine. Reading,
MA, Addison-Wesley
Tansel A U, Clifford J, Gadia S, Jajodia S, Segev A, and Snodgrass R (eds) 1993 Temporal
Databases: Theory, Design, and Implementation. Reading, MA, Benjamin/Cummings
Publishing Co
Tomlin C D 1990 Geographic Information Systems and Cartographic Modeling. Englewood
Cliffs, NJ, Prentice-Hall
Worboys M F 1990 Reasoning about GIS Using Temporal and Dynamic Logics. Santa Barbara,
CA, University of California, National Center for Geographic Informational Analysis
Center Technical Report 90±4
Worboys M F, Hearnshaw H M and Maguire D J 1990. Object-oriented data modeling for
spatial databases. International Journal of Geographical Information Systems 4: 369±83
Worboys M F 1992 A model for spatio-temporal information. In Bresnahan P, Corwin E, and
Cowen D (eds) Proceedings: The 5th International Symposium on Spatial Data Handling, 2.
San Jose, CA, American Congress of Surveying and Mapping: 602±11
Yuan M 1996 Modeling semantic, temporal, and spatial information in geographic information
systems. In Craglia M and Couclelis H (eds) Geographic Information Research: Bringing the
Atlantic. London, Taylor and Francis: 334±47

ß Blackwell Publishers Ltd. 1999

You might also like