Escolar Documentos
Profissional Documentos
Cultura Documentos
Robert C. Thomson
The Robert Gordon University, School of Computing,
St. Andrew St., Aberdeen, UK; email: rcthomson@yahoo.com
Abstract
Strokes are relatively simple linear elements readily perceived in a network. Apart from their role as graphical elements, strokes reflect lines of
flow or movement within the network itself and so constitute natural functional units. Since the functional importance of a stroke is reflected in its
perceived salience this makes strokes a suitable basis for network generalization, through the preferential preservation of salient strokes during data
reduction. In this paper an exploration of the dual functional-graphical nature of strokes is approached via a look at perceptual grouping in generalization. The identification and use of strokes are then described. The
strengths and limitations of stroke-based generalization are discussed; how
the technique may be developed is also considered. Finally, the functional
role of strokes in networks is highlighted by a look at recent developments
in space syntax and related studies.
Key words: network, generalization, perceptual grouping, continuity,
space syntax
1 Introduction
The automatic generalization of geographic networks can be viewed, at its
simplest, as the progressive removal of segments in a principled fashion.
For the generalization to be effective there are two goals to be achieved.
682
Robert C. Thomson
First, the reduction should preserve as far as possible the important features of the network where the definition of importance will be context
dependent, and may involve linkages with other networks and various nonnetwork contextual data. Second, the visual character of the network map
should be preserved where possible.
In some situations the two goals may be essentially the same. This paper
argues that when no semantic information about a network map is available then the perceived importance of elements will correlate strongly with
their perceptual salience. It follows that network features that combine
both functional importance within the network and perceptual salience in
the map representation should be preferentially retained in generalization.
Strokes are relatively simple network elements of this type. Strokes are
computationally simple to derive from network data, and generalization
based on strokes has been found to effective in several contexts. This paper
reviews the use of strokes in generalization with the broad aim of clarifying their dual functional/graphical nature and how these properties support
their role in effective generalization.
Aims and Structure. First perceptual grouping and its role in generalization are discussed. The procedures implemented at the Atlas of Canada for
extracting strokes and using them in generalization will then be briefly described, noting how it is sometimes possible to adjust their functional or
graphical roles in a generalization. Stroke-based network generalization
and other analyses are then described. The strengths and the limitations of
the stroke-based approach are discussed, and how the latter are being addressed. A final section considers the relationship between space syntax
and network generalization. It will be shown that the functional role of
strokes in networks is highlighted by recent developments in space syntax,
and that space syntactic measures can be used to support stroke-based generalization.
The material presented here deepens, extends and updates the discussion
of strokes, which formed one part of a recent, broader review of geographic network generalization (Thomson and Brooks 2006). In particular,
the discussion of perceptual grouping has been revised to clarify the dual
nature of the stroke and to place the use of strokes in a wider context of the
search for patterns in networks that should be preserved in generalization;
review material has been updated, and extended to consider applications of
strokes outside generalization; a new section has been added examining
links between strokes, generalization and space syntax analysis of road and
street networks.
683
2 Perceptual Grouping
When looking at maps of road or river networks, natural linear elements
will be seen which extend through junctions. These elements were termed
strokes (Thomson and Richardson 1999), prompted by the idea of a curvilinear segment that can be drawn in one smooth movement and without a
dramatic change in line style. Strokes are paths of good continuation: they
move through the network with no abrupt change in direction or character
at junctions. Any geographic network can be completely decomposed into
strokes. According to map type, the longer strokes present could be expected to represent the main courses of rivers or major routes; shorter
strokes could be expected to represent tributary streams or minor roads.
Implications about fundamental relations among these elements follow
from their perceived saliences. A longer and smoother stroke will naturally
appear more important than one shorter and more meandering. Also, one
stroke may terminate against another with implications of occlusion, tribute and less importance. The resulting perceived salience is a useful indication of the relative functional importance in the network of the features
represented by these elements. Simply put, roads or rivers that look important in a network map usually are important. And saying that a stroke
looks important is in effect saying it is a salient perceptual group.
Perceptual grouping is fundamental to human vision. Even with no high
level or semantic knowledge available, the human visual system spontaneously organizes elements of the visual field, resulting in the perception of
groups: Some arrangements of picture elements will tend to be seen as belonging together in natural groups, which often appear to stand out from
the surrounding elements, i.e. as figures against grounds. Many perceptual grouping principles have been identified, such as proximity, similarity,
symmetry, uniform density, closure, parallelism, collinearity, cotermination and continuity (Wertheimer 1938).
(a)
(b)
Fig. 1. A simple network with 8 arcs and 9 nodes resolves (a) into 4 strokes (b)
684
Robert C. Thomson
Perceptual grouping is a fundamental component of perceptual organization the ability of a visual system to spontaneously organize detected
features in images, even in the absence of high level or semantic knowledge (Palmer 1983; Witkin and Tenenbaum 1983; Lowe 1987). These
principles are recognized as the basis for parsing the visual world into surfaces and objects according to relatively simple visual characteristics, in a
process that operates independently of the domain being represented. Their
importance in map interpretation (MacEachren 1995) and generalization
(DeLucia and Black 1987) has also long been recognized.
The assumption that line-drawing recognition is a learned or cultural
phenomenon is not supported by the evidence (Lowe 1987). The mechanisms being used for line drawing or map understanding have presumably
developed from their use in recognizing three-dimensional scenes. The
role of perceptual organization in vision is to detect those image groupings
(such as those of Fig. 2) that are unlikely to have arisen by accident of
viewpoint or position. These image groupings therefore indicate probable
685
3 Stroke-based Generalization
The methods of stroke-based generalization used at the Atlas of Canada
have previously been described and illustrated (Thomson and Brooks
2000, 2002; Brooks 2003). The key steps are stroke building and stroke
ordering.
686
Robert C. Thomson
687
that there are no degenerate cases that cause the algorithm to fail, but to
achieve meaningful results it is essential that the data have correct connectivity and, in the case of hydrologic networks, correct flow direction. (According to the context for generalization, one-way restrictions on traffic
flow could be ignored or taken into account by additional rules on arc linking at junctions.)
The resulting strokes are concatenations of arcs each representing a path
through the network from one terminal node to another. There are no restrictions imposed on the overall shape of the stroke that emerges. No consideration is taken of the total curvature of the stroke [c.f. Chaudhry and
Mackaness (2005) implementation], and the resulting stroke could in certain cases self-intersect or form a complete loop thus allowing some orbital routes or ring roads, say, to be represented by single strokes.
3.2 Stroke Ordering
Once the strokes are constructed for a network they must be ordered appropriately, effectively assigning a ranking value or weight to each. The
generalization can then proceed through the removal of some suitable proportion of strokes, in what can be viewed as a thresholding process. The
data can be represented to the cartographer in an interactive display with
slider control of the threshold percentage. Percentage reductions based on
criteria such as Tpfers radical law (Tpfer and Pillewizer 1966) could be
applied, but the interactive method was preferred.
The success of the generalization will depend both on the suitability of
the strokes found and the method used for their ordering. It is to be expected that the ordering method will depend on factors such as the criteria
used in stroke building, the availability of arc attribute data, and the purpose of the generalization.
Stroke attributes on which to base the ordering will be derived from the
attributes of their constituent arcs and nodes. Two important stroke attributes are ratio measures that do not depend on thematic attributes of constituent arcs: stroke length, which is readily found, and a measure of connectivity that can be derived easily from the degrees of the nodes within
the stroke. Other stroke attributes depend on the available arc attribute
data. Road class/category is an important road attribute with values that are
normally nominal, e.g. motorway or single track. However workable
ordinal values usually follow from consideration of the relative road qualities. It may even be possible to derive workable ratio values to represent
road class attributes. The problem is similar to that of deriving 'friction
values' for roads, i.e. dimensionless multiplicative factors, which estimate
688
Robert C. Thomson
the effective length of a road segment given its attributes (Richardson and
Thomson 1996).
It may be noted that this method of ordering strokes does not use the
implications about relative stroke importance that may be drawn from observing how one stroke terminates at a junction when another passes
through it.
Network Connectivity. In the implementation at the Atlas of Canada
strokes are ordered for generalization in two stages. In the first stage a
stroke salience order is produced through sorting first by road category/class data (if available), and then by stroke length. There is no consideration of strokes role in connectivity and hence no guarantee that the
network would not become disconnected if generalization were applied directly to these strokes. Nevertheless, these strokes have been found to be
useful for several applications (Section 3.3) and good generalization results can often be produced without further consideration of connectivity.
A second stage revision of stroke order eliminates the possible problem
of roads becoming disconnected during stroke removal that has sometimes
been reported (Zhang 2004b; Elias 2002). The reordering procedure used
at the Atlas of Canada has been published (Thomson and Brooks 2000),
albeit with a typing error. The algorithm ensures that if a stroke is removed
then all pairs of its stroke neighbors remain connected by paths through the
strokes that remain. The search for alternative paths is not exhaustive, but
this did not adversely affect the results.
This reordering of strokes is effectively subordinating the perceptual salience of strokes to their structural role in maintaining network connectivity. Thus generalization on the basis of the revised order can lead to
relatively salient strokes being removed while less salient strokes survive
because of their greater structural importance.
3.3 Using Strokes
Use in Generalization. The most successful application of stroke-based
generalization has been to hydrographic networks. The system developed
at the Atlas of Canada was used to generalize the hydrology for a published 1:4M scale map of the three northern territories from source material in the GeoBase Level 0 hydrology dataset. The source data had been
cleaned, attributes had been added, connectivity corrected and directionality computed (Brooks 2000, 2003). The results were comparable in quality
with previous methods, and the new approach brought additional advantages in production (see below).
689
690
Robert C. Thomson
connections. Heinzle et al. (2005) used strokes in detecting certain important patterns in road networks.
The use of strokes and its related concepts in space syntactic network
analysis is considered below (Section 4.2).
3.4 Further Development of Stroke-Based Generalization
Although stroke-based generalization preferentially retains the more salient strokes in a network and this helps to preserve the network character,
this may not be sufficient when higher level structures are perceptible in
the network i.e. patterns formed from groupings of arcs or strokes. The
network may lose aspects of its character during generalization if certain
patterns are disrupted or lost by stroke removal. The problem is most likely
to occur in dealing with urban road networks, where such patterns are
commonly found, and unlikely to arise with river networks.
As noted (Section 2), characteristic higher-level network patterns were
identified by Zhang (2004a) and Heinzle et al. (2005). All were instances
of one or more perceptual grouping principles. Ideally, network generalization should preserve these patterns where possible, perhaps necessarily in
attenuated form, in order to preserve the network character. Stroke based
generalization as described above does not attempt to retain any higherlevel patterns.
The use of strokes can be viewed as the first, important step in preserving perceptual groupings in network generalization. Strokes can comprise
several network arcs and in that sense are intermediate-level structures, but
they are relatively simple perceptual groups and often serve as a basis for
higher-level groupings. Preserving the more salient strokes helps preserve
one facet of the network character during generalization. The surviving
strokes may then preserve some higher-level perceptual groupings that are
based on them, but this cannot be guaranteed without more relatively sophisticated processing. The work of Edwardes and Mackaness (2000) was
an important advance in this direction.
Using a Networks Areal Dual. Edwardes and Mackaness recognized that
the stroke-based approach to generalization is concerned only with the set
of linear road or street objects and so may not provide sufficient consideration of the networks areal properties. Their solution to this problem (Edwardes and Mackaness 2000) goes a significant way in addressing the
shortcomings of the stroke-based approach described above.
Strokes are used as one tool for characterizing the network in order to
provide global information to the generalization process, but their method
691
adopts the areal dual of the network as a second structure for characterization, using the two structures simultaneously to perform the generalization.
The urban spaces are partitioned into city blocks using minimum cycles of
streets, and generalization proceeds by sequential fusion of adjacent blocks
when a block area is below a scale dependent threshold. A partition always
aggregates with its neighbor across the weakest boundary, hence the weakest strokes are removed from the network and the effect on the overall
good continuation of the network is minimized. The area size threshold
can be varied using a function relating it to district density, which helps
preserve network density patterns. Network connectivity is handled implicitly, since block aggregation cannot disconnect the network, although some
special cases need additional processing. The algorithm produces good results, identifying and retaining the essential areal, linear, semantic and
density patterns of the network and its constituent roads.
692
Robert C. Thomson
693
694
Robert C. Thomson
(ICN) was adopted. ICN is identical with stroke construction in all important respects. The vertices of the resulting dual graph thus represent
strokes, the arcs indicate stroke intersections, and syntactic analyses are as
for conventional graphs derived from axial maps.
Useful features of the ICN (stroke-building) model were that it allows
complex chains like loops and tailed loops to be recognized, and captures
most of the continuity of paths throughout urban networks. Also, being
based on a pure spatial principle of continuity, it avoids problems of social
interpretation. From the combined analyses of the dual graph and a primal
graph (derived directly from the original network map) new and useful
measures of network structure are being developed (Porta et al. 2005).
5 Conclusions
Strokes are simple elements of a network, readily perceived in its map,
whose visual salience broadly reflects functional importance within the
network. The perceived paths of good continuation in the map indicate
natural lines of flow in the physical network. Because of the good general
correlation between these two aspects of strokes they form a suitable basis
for network generalization, through the preferential preservation of the
more salient strokes in data reduction.
This paper approached the functional-graphical nature of strokes via a
wider consideration of perceptual grouping. Perceptual groups in a representation of the world such as an image or map generally reflect important
features in the world being represented. Retaining the more salient perceptual groups in a network during generalization should therefore help to preserve both its functionally important features and the visual character of
the map.
Hence the use of strokes can be viewed as a first, important step in preserving perceptual groupings in generalization. Strokes are relatively simple groups and often serve as a basis for higher-level groupings. Preserving
the more salient strokes helps preserve one facet of the network character
during generalization; the surviving strokes may preserve some higherlevel perceptual groupings that are based on them, but this cannot be guaranteed without more relatively sophisticated processing. For networks
without such patterns, such as river networks, strokes provide a basis for
effective and efficient generalization as demonstrated in the production
of commercial maps.
Methods for implementing network generalization on the basis of
strokes were described. Recent applications of strokes in generalization
695
and other analyses were reviewed. The limitations of stroke-based generalization were also discussed, and further development of the technique
was considered.
Finally, some links between network generalization, space syntax analysis, and strokes were highlighted. Here strong support for the view that
strokes represent important structural/functional units of networks was
found in examples of space syntactic network analysis incorporating the
use of strokes or closely similar elements.
Acknowledgments
The discussions with Rupert Brooks and William Mackaness concerning
the topics of this paper are greatly appreciated.
References
Brooks R (2000) National Atlas of Canada producing first map using automated
generalization of framework data. Cartouche 39
Brooks R (2003) Atlas of Canada open-source generalization tools. Online document available at http://www.cim.mcgill.ca/~rbrook/atlas_gen/
Chaudhry O, Mackaness M (2005) Rural and urban road network generalization
deriving 1:250000 from OS MasterMap. In: Proc 22nd Int Cartographic Conf,
La Corua, Spain
Dalton N (2001) Fractional configurational analysis and a solution to the Manhattan problem. In: Peponis J (ed) Proc 3rd Int Space Syntax Symp, Atlanta,
Georgia, pp 26.126.13
Dalton N, Peponis J, Conroy Dalton R (2003) To tame a TIGER one has to know
its nature: extending weighted angular integration analysis to the description
of GIS road-center line data for large scale urban analysis. In: Hanson J (ed)
Proc 4th Int Space Syntax Symp, London, pp 65.165.10
DeLucia A, Black TA (1987) Comprehensive approach to automatic feature generalization. In: Proc 13th Int Cartographic Conf, pp 168191
Downs TC, Mackaness WA (2002) Automating the generalization of geological
maps: the need for an integrated approach. The Cartographic J 39(2):137152
Edwardes AJ, Mackaness WA (2000) Intelligent generalization of urban road networks. In: Proc GIS Research UK 2000 Conf, York, pp 8185
Elias B (2002) Automatic derivation of location maps. IAPRS 34(4), Geospatial
Theory, Processing and Applications
Figueiredo L. (2004) Linhas de Continuidade no Sistema Axial. Unpublished MSc
Dissertation, Federal University of Pernambuco Recife
696
Robert C. Thomson
Figueiredo L, Amorim L. (2004) Continuity lines: aggregating axial lines to predict vehicular movement patterns. In: Proc 3rd Great Asian Streets Symp,
Singapore
Figueiredo L, Amorim L (2005) Continuity lines in the axial system. In: Van Nes
A (ed) Proc 5th Int Space Syntax Symp, Delft
Heinzle F, Anders K-H, Sester M (2005) Graph based approaches for recognition
of patterns and implicit information in road networks. In: Proc 22nd Int Cartographic Conf, La Corua, Spain
Hillier B, Hanson J (1984) The Social Logic of Space. Cambridge University
Press, Cambridge
Jiang B, Claramunt CA (2004) A structural approach to model generalization of
an urban street network. GeoInformatica 8(2):157171
Klarqvist B (1993) A space syntax glossary. Nordisk Arkitekturforskning 2:1112
Lowe DG (1987) Three-dimensional object recognition from single twodimensional images. Artificial Intelligence 31(3):355395
MacEachren AM (1995) How Maps Work: Representation, Visualization, and Design. Guilford Press, New York
Mackaness WA (1995) Analysis of urban road networks to support cartographic
generalization. Cartography and Geographic Information Systems 22:306316
Mackaness WA, Mackechnie GA (1999) Automating the detection and simplification of junctions in road networks. GeoInformatica 3(2):185200
Misund G, Johnsen KE, Lindh M (2003) Hierarchical GML modeling of transportation networks. Presented at the 2nd Annual GML Developers Conf [http://
www.ia.hiof.no /~gunnarmi/omd /gmldev_03 (accessed 24.02.06)]
Palmer S (1983) The psychology of perceptual organisation: a transformational
approach. In: Beck J (ed) Human and Machine Vision. Academic Press, New
York, pp 269339
Porta S, Crucitti P, Latora V (2004) The network analysis of urban streets: a dual
approach. arxivorg preprint cond-mat/0411241
Porta S, Crucitti P, Latora V (2005) The network analysis of urban streets: a primal approach. arxivorg preprint cond-mat/0506009
Richardson DE, Thomson RC (1996) Integrating thematic, geometric and topological information in the generalization of road networks. Cartographica
33:7583
Thom S (2005) A strategy for collapsing OS Integrated Transport Network dual
carriageways. In: 8th ICA Workshop on Generalization and Multiple Representation, La Corua, Spain
Thomson RC (2003) Bending the axial line: smoothly continuous road centre-line
segments as a basis for road network analysis. In: Hanson J (ed) Proc 4th Int
Space Syntax Symp, London, pp 50.150.10
Thomson RC, Brooks R (2000) Efficient generalization and abstraction of network data using perceptual grouping. In: Proc 5th Int Conf on GeoComputation, Greenwich UK
Thomson RC, Brooks R (2002) Exploiting perceptual grouping for map analysis,
understanding and generalization: the case of road and river networks. In:
697
Blostein D, Kwon YB (eds) Graphics Recognition: Algorithms and Applications (= LNCS 2390). Springer, Berlin, pp 141150
Thomson RC, Brooks R (2006) Generalization of geographical networks. In:
Mackaness WA, Ruas A, Sarjakoski T (eds) The Generalization of Geographic Information: Models and Applications. Elsevier, Amsterdam
Thomson RC, Claridge E (1989) A 'computer vision' approach to the analysis of
crystal profiles in rock sections. In: Pietikainen M (ed) Proc 6th Scandinavian
Conf on Image Analysis, Oulu, Finland, pp 12081215
Thomson RC, Richardson DE (1999) The 'good continuation' principle of perceptual organization applied to the generalization of road networks. In: Proc 19th
Int Cartographic Conf, pp 12151223
Tpfer F, Pillewizer W (1966) The principles of selection: a means of cartographic generalization. The Cartographic J 3(1):1016
Turner A (2001) Angular analysis. In: Peponis J (ed) Proc 3rd Int Space Syntax
Symp, Atlanta, Georgia, pp 30.130.11
Turner A (2005) Could a road-centre line be an axial line in disguise? In: Van Nes
A (ed) Proc 5th Int Space Syntax Symp, Delft
Wertheimer M (1938) Laws of organization in perceptual forms. In: Ellis W (ed)
A Source Book of Gestalt Psychology. Harcourt Brace, New York, pp 7188
Witkin AP Tenenbaum JM (1983) On the role of structure in vision In: Beck J,
Hope B, Rosenfeld A (eds) Human and Machine Vision. Academic Press,
New York, pp 481583
Zhang Q (2004a) Modeling structure and patterns in road network generalization.
ICA Workshop on Generalization and Multiple Representation, Leicester, UK
Zhang Q (2004b) Road network generalization based on connection analysis. In:
Proc 11th Int Symp on Spatial Data Handling, Leicester, UK