Qualitative Representations in Large Spatial Databases

Baher A. El-Geresy
School of Computing, University of Glamorgan,
Treforest, Mid Glamorgan, CF37 IDL, Wales, UK
bageresy @glam.ac.uk
Alia I. Abdelmoty
Department of Computer Science, Cardiff University
Queens Buildings, Newport road, Cardiff, Wales UK
A.I.Abdelmoty @cs.cf.ac.uk

Abstract number of sources.

Qualitative representation of spatial information is im-
portant in large spatial databases. This type of informa-
tion can be used in manipulation and analysis when no pre-
cise geometric information exist or are needed. Qualitative
manipulation is particularly useful when integrating differ-
ent geographic data sets, which may have been drawn from
different sources at different times, which may have differ-
ent accuracy measures, and which may vary in the scales
of presentation resulting in possible multiple representa-
tions of geographic entities. In this paper a representa-
tion scheme, denoted the Ring structure is proposed where
objects are grouped and ordered in layers based on adja-
cency and orientation. The method facilitates qualitative
manipulation of the data and the efJicient derivation of im-
plicit injrormation using spatial reasoning techniques. The
method is shown to facilitate the integration of data sets by
enhancing the process of spatial matching across these sets.
The method is simple and can be used in many applffations
of large spatial databases, such as GIS, CAD and image
databases
1. Introduction
One of the main functions of spatial information systems
such as GISs is the unification and integration of different
data sets and making them available for coherent manipula-
tion and analysis by different applications. Integrating data
in spatial information systems involves the integration of di-
verse types of information drawn from a variety of sources
requiring effective matching of similar entities in these data
sets and information consistency across data sets. Typically
spatial information can be provided in different forms by a
A pre-requisite for the effective use and manipulation
of several diverse spatial data sets is the understanding of
the contents of the data sets and how they compare to each
other. In this paper, a qualitative approach is proposed to
spatial data representation where the spatial content of the
data sets is encoded in two forms. First the spatial objects
in the data sets are grouped together, in a hierarchical fash-
ion, in collections denoted Rings, based on their relative ad-
jacency. Adjacency and orientation relationships are then
explicitly encoded between some specified sets of object
in those rings and represented in a structure denoted the
adjacency-orientation matrix.
The issues in this paper are discussed using geographic
information in GISs as example, but are also applicable to
other types of spatial information systems and applications.
The work in this paper is done in the context of an ongoing
research project which aims at the development of methods
for the modeling and manipulation of hybrid data sets in a
GIS [ l 11.
The rest of this paper is structured as follows. Related
work is presented in section 2. In section 3 the proposed
approach is described. The implementation of the method
is described in section 4. This is followed in section 5 by
the application of the method to the problem of integrating
different spatial data sets. Conclusions are given in section
6.
2. Related Work
Approaches to the qualitative representation of images or
maps can be classified into two categories. In the first cate-
gory, spatial relations are studied and defined between pairs
of objects, e.g. defining relationships between two simple
regions or two linear objects, etc. In the second category,

68
1098-8068/01$10.000 2001 IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.
 approaches attempt to describe continuous spaces by de-
so
scribing sets of objects and relationships in these spaces.
In the first category, several methods were proposed. Ex-
amples are the work of of the Cohn et a1 [3] and the work
of Egenhofer et a1 [6] and Jen and Boursier [IO]. The set of
topological relations between two spatial objects, e.g. con-
vex regions, are first defined. Then those are used to de-
fine relationships between more complex objects such as
regions with holes. In [7], eight topological relations be-
tween simple regions were used to represent composite and
non-composite fields using a method similar to consistency
networks.
In the second category, the main approaches proposed
define spatial scenes using symbolic arrays and minimum
bounding rectangles [2, 91. However, it is recognised that
approximating objects by their minimum bounding rectan-
gles may produce misleading results. In a different ap-
proach, Lundell [ 141 used graphical illustrations to repre-
sent the adjacency between composite and non-composite
physical fields. Composite fields are represented by draw-
ing connected lines between the different representations
of data layers or themes. The representation of change is
depicted through a sequence of diagrams. A computation
model for this method is however not directly envisaged.
Glasgow and Papadias [9] showed how a symbolic array
can represent whole map scenes schematically. Methods for
I O3 I A' ;
checking consistency in spatial databases have been limited
to checking topological consistency of pairs of spatial ob-
jects and not to whole map scenes [12, 41. In [7] consis-
tency networks were used to check the consistency of a spa-
Figure 1. (a) Sample map scene. (b) Corre-
tial scene containing regions with holes. In [ 171consistency
sponding Rings. (c) Forming rings in maps
of topological relations between multiple representations of with disconnected objects.
objects, specifically between parts and aggregate represen-
tations, is given.
be propagated to deduce orientation relations between non-
3. The Approach adjacent objects. The representation levels are given in the
rest of this section.
3.1. Representation Levels I. The Ring Structure The semi-infinite region
around the map edges is the reference ring and is bounded
The basic idea of the proposed approach is simple. The by the four map edges, north ( N ) ,east ( E ) ,south, (S)and
map or spatial scene is divided up into concentric layers west ( W ) .Adjacent objects from the edge of the map form
of adjacent objects in a form resembling the annual growth the first ring RI.Adjacent objects to the first ring form the
rings on a tree trunk. The map edges constitute the first ring second ring RZ and so on. This process ends when all the
and serves as a frame of reference for the rest of the scene. objects are related to a ring. The last ring formed is the core
Objects are then grouped in a hierarchical division of the of the spatial scene. An example of this process is shown in
map into successive rings and related by explicitly defining figure 1 where successive rings are shown with no orienta-
adjacency and orientation relationships between objects in tion or adjacency relationships.
every ring. Adjacency and orientation relationships are then In some cases the map may have more than one core,
used to relate objects across consecutive rings. It is shown where rings may enclose more than one disjoint ring. In
how the inter and intra-ring relationships can be represented this case the core of the tree is made up of more than one
using one structure. ring.
Relationshipsare explicitly represented between a subset In other cases, the map may contain a set of disconnected
of objects on the map. It is shown how this information can objects. Virtual rings can then be created around groups of

69

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.
 so
05 01
I
RI 01 02 03
R2 09 06
E
07 08 09
N
R3
Figure 2. Representing Inter-Ring relation-
ships. (a) Sample map. (b) Orientation rela-
tionships are encoded only between adjacent
objects.
objects as shown in figure l(c). The figure shows a distribu-
tion of disjoint regions in a map where no natural division
of rings is obvious.
Note that the virtual rings divides the containing objects
into regions, namely Ao, A I ,A2 and As, which are used to
encode the adjacency and orientation relationships. Only
objects with outer edges adjacent to the ring edges are re-
lated to it. In this case, the virtual rings act as frames of ref-
erence for their enclosing objects in exactly the same way
the map edges are for the map scene.
11. Representing Inter-Ring Relationships
In this step, orientation relationships are represented be-
tween adjacent objects in different rings. For every object
in every ring in the tree, define its relative orientation with
adjacent objects. An example of encoding this type of rela-
tionships for the map in figure 2(a) is given in 2(b).
111. Representing Intra-Ring Relationships
Here, orientation relationships are represented between ad-
jacent objects in the same ring. This is achieved in two
steps:
1. Firstly, a matrix is used to encode the adjacency rela-
tionships between objects in the ring. In figure 3(a) a
matrix is used to hold adjacency relationships between
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.
E 07
0 E 081
s 0 s 1 0 9 1
05
3 Figure 3. (a) Adjacency matrix for objects in
R2 of figure 2. (b) Half the symmetric adja-
cency matrix is sufficient to capture the scene
representation. (c) Adjacency-orientation matrix
is formed by modifying the adjacency matrix
to hold orientation relations between adja-
cent objects in the ring. (d) The combined
inter- and intra- ring adjacency-orientation
matrix.
objects in R2. The fact that two entities are adjacent
is represented by a value (1) in the matrix and by a
value (0) otherwise. For example, 0 6 is adjacent to
both objects 07 and 09, object 07 is adjacent to object
os and object 08 is adjacent to object 09. Since ad-
jacency is a symmetric relation, the resulting matrix
will be symmetric around the diagonal. Hence, only
half the matrix is sufficient for the representation of
the space topology and the matrix can be collapsed to
the structure in figure 3(b).
2. Secondly, the values of (1)in the matrix is replaced by
the relative orientation relationship between the corre-
sponding objects as shown in figure 3(c). This struc-
ture is then denoted Adjacency-Orientation Matrix.
Note that So represents the infinite embedding space of the
map.
IV. Representing the Combined Inter- and
Intra-Ring Relationships Representation levels I1
and I11 above can be combined by representing the com-
plete set of objects in the scene in the adjacency-orientarion
matrix. Hence, a uniform structure can be used to capture
the adjacency and orientation relationships between all the
70
 map objects. The combined matrix for the map in figure 2
is shown in figure 2(d).
The matrix can be kept compact by exploiting the tran-
sitive property of the orientation relations. Relations be-
tween non-adjacent entities can be deduced using qualita-
tive reasoning. The convention of orientation relations is
R(column,row).
4. Implementation of the Approach using Ori-
entation Rings
In the previous section, the ring structure proposed is
constructed using adjacency relationships. Orientation re-
lationships are stored between adjacent objects and derived
for others. In this section, the algorithm for implementing
the method is described.
The structure for representing objects in orientation rings
is first described followed by detailed discussion of the
derivation of possible orientation relationships. It is as-
sumed that the adjacency of all objects in the map is first
explicitly stored in an adjacency matrix.
1. Creating Orientation Rings
The example in figure 4 is used to illustrate the proce-
dure.
The first orientation ring is formed from objects which
are intersecting or are adjacent to the map edges.
Objects are collected in ordered sets starting from the
north-east (upper-right) corner of the map and moving
in a clockwise direction around the edges. Ring 0 is
defined by the set of all objects on the four edges,
namely, E(1,2,3,4,5,6,),S(6,7,8,9,1O),W(lO)
and N(10,11,12,13),as shown in figure 4(b).
The second, and successive, rings are constructed us-
ing the following steps.
Ignore all objects already referenced in previous rings
from further consideration.
Define the minimum bounding rectangle (mbr) for all
remaining objects in the map, as shown in 4(c).
Identify the orientation edges of the next ring using the
following procedure.
0 Let S be the set of all mbrs of objects not yet
assigned to a ring.
0 The eastern edge is defined such that it collects
all outermost eastern objects as follows. Let
z l ( E ) = max(z,i,(S)). For all Se E S,
such that, z,,,(Se) 2 q ( E ) ,then Q ( E ) =
min(zmaZ(Se)). The eastern edge of the ring
lies between z l ( E ) and x z ( E ) . The process is
shown in 4(d), where the eastern edge limits are
71
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.
defined by the shaded vertical area in the figure.
Further illustration of the process is shown in fig-
ure 5 where A, B , C, D , E , F, G are the mini-
mum bounding rectangles for the objects in set
S. In this case, q ( E ) = xmin(D) and Se =
{ A , B , C } , hence Q ( E ) = z,,,(B) and the
eastern edge is between z,in (D) and x,,, (23).
0 Similarly, to define the western edge of the ring,
let z l ( W ) = min(x,,,(S)). For all S, E S,
such that, x,in(Sw) 5 x l ( W ) ,then z2(W) =
max(x,i,(S,)). The western edge of the ring
lies between 2 1 ( W )and z 2 (W ) .
Use the same methodology to define the north,
west and south edges. This results in the identi-
fication of the orientation frame of reference for
the first ring.
0 An ordered list of objects intersected by the four
edges in clockwise order from east to north de-
fines objects of the first ring.
5. Repeat the process from step two above to identify suc-
cessive rings and stop when all objects in the map are
related to a ring. The map in figure 4 is defined by three
rings, Ring 0, Ring 1and Ring 2, as shown in 4(f). The
procedure for creating successive rings is shown in fig-
ure 6.
It should be noted that the mbrs are only used for defin-
ing the orientation reference frames. The rings store ordered
intersections with the actual objects and not their mbrs, and
thus provide a very accurate measure of orientation rela-
tions. This fact distinguishes our method from all others
based on the manipulation of mbrs [ 131.
The method is expected to perform well in cases where
objects are not intertwined, e.g. intertwined spirals. Such
objects are not common in geographic data sets.
2. Deriving Orientation Relations
Orientation relations between objects are derived from
the orientation ring structure as follows:
1. Orientation relationships between consecutive objects
on every ring edge are determined by their order. For
example, east and west edges in every ring determine
the north and south relationships, e.g. in figure 4(b),
object 1 is north of object 2 which is north of 3, etc.
Similarly, the north and south edges determine west
and east relationships.
2. Transitivity of order relations are used to determine
orientation relations between objects on similar ori-
entation edges in consecutive rings. If A is on the
east edge of ring 0 and B and C are on the east
. edges of rings 1and 2 respectively,then, east(A,B ) A
east(€!,C ) -+ east(A, C ) .
 I I I ! I

-
I .
I

XI@)I x2(E)

range for East edge

Figure 5. The process of creating orientation

rings.

1 S = set of all objects I

Sc is first ring
(collect objects
intersecting map

(e>
N

Define E, W, N, S
edges of next ring
w E

Sc is followingring
(collect objects
intersectingwith

(0
. Figure 4. Process of creating orientation
rings. (a) Example Map. (b) Ring 0 defined
u s k g the map edges as the first orientation
frame of reference. (c) MBRs-of the all ob- Figure 6. Process of creating successive
jects except those intersected by Ring 0. (d) rings.
Finding the east edge of Ring 1. (e) Orienta- .
tion edges of Ring 1. (f) All three orientation
rings for the map.

72

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.
 , 3. Projections of ring edges determine further refined ori-
entation relationships, e.g. Switzerland is in the north-
- east of Spain.
4. Further distinction between orientation relations can
be made using the order of the ring, e.g. Germany is
further away in the northeast direction of Spain than
Switzerland. Note, that objects intersecting more than
one ring is recorded in both.
5. Application: Integration of Different Data
Sets
Typically spatial information can be provided in different
forms by a number of sources. For example, data sources
in GIS include maps, field surveys, photogrammetry and re-
mote sensing. Data sets can be collected at different scales
or resolutions in different times. The problem of integration
is further complicated as the data may be collected in in-
compatible ways and may vary in reliability. Some details
may be missing or undefined. Incompatibilities between
different data sets can include incompatibilities between the
spatial entities for which data are recorded, including dif-
ferences in dimension, shape and positional accuracy. Two
possible approaches can be identified for handling the inte-
gration problem:
1 . Convert all the data sets to a single format and store the
most detailed representation of the data (at one scale)
[151.
2. Permit inconsistent data to co-exist with explicit infor-
mation being provided on the inconsistencies between
the data sets [16, 11.
The first approach is limiting since, it is not feasible for ap-
plications requiring immediate response to user queries to
wait for the derivation of the required representation from
existing or stored ones, and digital map coverages are of-
ten variable in quality from one place to another. A single
large scale representation may not be available for an entire
region of interest.
On the other hand, the second approach is more flexi-
ble. To realise this approach, first the relationship between
the data sets is identified which can be used to assess the
inconsistency between the sets.
An important assumption here, is that in integrating spa-
tial data sets qualitative relationships, namely, topological
and orientation have to be “mostly” preserved.
Identifying Spatial Relationships between
Different Data Sets
A pre-requisite for comparing data sets in space is the
knowledge of their relative spatial extension, i.e. the
amount of overlap between the data sets. An illustration
73
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.
Data set A Data set B
(b)
Figure 7. a) Two different spatial data sets A
and B. R1, and Rlb are the first rings in both
sets respectively. b) Some possible relation-
ships between the data sets.
of the possible relationships between data sets in space is
shown in figure 7. Spatial relationships between data sets
can be derived by comparing the relative location of their
constituent objects. This comparison can be achieved di-
rectly by comparing their Ring structures and the associated
adjacency-orientation matrices.
The intersection method developed initially by Egen-
hofer and Herring [5] and later modified and extended in
[8],is used to formally define qualitative relationships be-
tween two spatial objects. This method can be used to
define relationships between full spatial data sets. In [ 5 ] ,
point-set topology is used for the definition of the object
components as interior ( A ” ) ,boundary (6A) and exterior
( A - ) . Spatial relationships between the objects considered
are the result of the exhaustive combinatorial intersection of
their components represented in a matrix which is used as
a model for representation. Eight topological relationships
can exist between two simple regions. The relationships and
their corresponding intersection matrices are shown in fig-
ure 8(a). If orientation and topology are concurrently con-
sidered, then the set is extended to 169 different relation-
ships as identified by [13]. A subset of the 169 relations is
shown in 8(b).
The above method is applied for whole data sets as fol-
lows.
1. The outer rings ( R I )for the data sets are considered to
be their boundaries and the rest of the objects (in the
rest of the rings) make their interior.
 I \ I " ' 1

lAo;
{A01 All A2 1 ~ 0 All
1 A2 !A01 All A2
B 1 1 - 1 8 1 1 1 8 1 1 1 8 1 0 0
8 1 0 0 8 1 1 0 B l l l B O 1 0
8 1 0 0
A2 B l O O B l l l B O O 1

Figure 9. (a) Two spatial data sets. (b) Their

relative spatial relation defined by comparing
their ring and adjacency-orientation matrices.

2. Objects in the boundary and interior in one set are then

Figure 8. (a) Eight possible topological re-
lationships between two simple regions and
their intersection matrices. (b) 169 possible
topological-orientation relationships. (only
part of the set is shown).
compared with the objects in the boundary and interior
in the other set. I.e. the point-set topology is replaced
by an object-based topology comparison. Figure 9 il-
lustrates the derivation of an example overlap relation-
ship between two data sets using this method.
More than eight possible relationships can be derived
using this method as the specific adjacency and orien-
tation of the objects are considered, and hence differ-
ent varieties of overlap, touch and containment can be
distinguished.
Spatial reasoning techniques can be used to propagate and
derive new implicit spatial relationships. In particular in the
case where the details of the objects are not complete in
one data set (e.g. in the case of complex objects formed by
aggregating several parts) but can be derived by integrating
several sets.

6. Conclusions

In this paper, a qualitative approach is proposed to spa-

tial data representation where the spatial content of the data
sets is encoded in two structures using the adjacency and

74

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.
 orientation relationships between objects. The representa- the 6th intemational Symposium on Spatial Data Handling,
tion schemes are directly derivable from the actual object volume 1 , pages 73-89, 1994.
representations, thus guaranteeing consistency of the infor- [ 111 C. Jones, D. Kidner, and J. Ware. Equivalence Tests for
mation content in these structures. This is an advantage over Integration of Multi-Source Spatial Data. In Proceedings
the constraint network representation of the spatial data sets EuroCarto X l l l , Ispra, Italy, pages 13-22. Joint Research
Centre, European Commission, 1996.
[71. [12] B. Kuijpers, J. Paredaens, and J. den Bussche. On topologi-
A method for implementing the ring structure was de- cal elementary equivalence of spatial databases. In F. Afrati
scribed. An algorithm for systematically defining rings of and P. Kolaitis, editors, to appear in the Proceedings of6th
orientation relations is given with which orientation rela- International Conference on Database Theory (ICDT '97),
tions between adjacent and non-adjacent objects could be pages 4 3 2 4 6 . Springer Verlag LNCS-1186, 1997.
accurately defined using a minimal set of pre-stored and cal- [13] S.-Y. Lee and F.-J. Hsu. Spatial Reasoning and Similarity
culated information. The method also has the advantage of Retrieval of Images using 2D C-String Knowledge Repre-
minimising the uncertainty associated with chain reasoning sentation. Pattern Recognition, 25(3):305-318, 1992.
when applied in this domain. [ 141 M. Lundell. A Qualitative Model of Physical Fields. In Pro-
ceeding of the Fourteenth National Conference on Artijicial
The work presented here is a step towards defining qual-
Intelligence, AAAI-92, pages 1016-1021. AAAI Press/ The
itative spatial data models and structures which allows for
MIT Press, 1996.
the incorporation of spatial reasoning engines in large spa- [I51 T.Nyerges. Schema Integration Analysis for the Develop-
tial databases. ment of CIS Databases. Int. J. Geographic Information Sys-
tems, 3(2):153-183, 1989.
[16] I. Shepherd. Information Integration and CIS. In
References D. Maguire, M. Goodchild, and D. Rhind, editors, Ge-
ographic Information Systems, volume l , pages 337-60.
[ 13 D. Briggs and H. Mounsey. Integrating Land Resource Data Longman Scientific & Technical, 1991.
into a Eurpean Geographicsal Information Systems. Applied [I71 N. Tryfona and M. Egenhofer. Consistency among Parts and
Geography, 9( 1):5-20, 1989. Aggregates: A Computational Model. Transactions in CIS,
[2] S. Chang, Q. Shi, and W. Yan. Iconic Indexing by 2-D l(3): 189-206, 1997.
Strings. IEEE Trans. Pattern Analysis and Machine Intel-
ligence, PAMI9(3):413-428, 1987.
[3] Z. Cui, A. Cohn, and D. Randell. Qualitative and Topolog-
ical Relationships in Spatial Databases. In Design and Im-
plementation of Large Spatial Databases, third Symposium
SSD '93, LNCS 692, pages 396-315. Springer Verlag, June
1993.
141 M. Egenhofer, E. Clementini, and P. Di Felice. Evalu-
ating Inconsistencies among Multiple Representations. In
Proceedings of the 5th International Symposium on Spa-
tial Data Handling, volume 2, pages 901-918, Charleston,
1994. IGU Commission of GIs.
[5] M. Egenhofer and J. Herring. A Mathematical Framework
for the Definition of Topological Relationships. In Proceed-
ings of the 4th international Symposium on Spatial Data
Handling, volume 2, pages 803-13, 1990.
[6] M. Egenhofer and J. Sharma. Topological Relations between
Regions in R2 and Z2.In Design and Implementation of
Large Spatial Databases, third Symposium SSD '93,LNCS
692. Springer Verlag, June 1993.
[7] M. Egenhofer and J. Sharma. Assessing the consistency
of complete and incomplete topological information. Ge-
ographical Systems, l( 1):47-68, 1997.
[8] B. El-Geresy and A. Abdelmoty. Order in Space: A Gen-
eral Formalism for Spatial Reasoning. Int. J. on Artijicial
Intelligence Tools, 6(4):423-450, 1997.
[9] J. Glasgow and D. Papadias. Computation Imagery. In
B. Chandrasekaran and J. Glasgow, editors, Diagrammatic
Reasoning, pages 435-480. AAAI Press, 1995.
[lo] T. Jen and P. Boursier. A Model for Handling Topologi-
cal Relationships in a 2D Environment. In Proceedings of

75

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on October 21, 2009 at 14:13 from IEEE Xplore. Restrictions apply.