2004 
  
small 
long 
‘such 
ed on 
bject- 
rpose 
5, and 
les Or 
r and 
is on 
phical 
Y. the 
lefine 
“ina 
eature 
| have 
nation 
S. Lee 
riteria 
ftware 
esults 
nation 
ere Is 
these 
aimed 
¢ line 
ers as 
blem. 
rocess 
uild a 
2000) 
prove 
multi- 
rk on 
ational 
nercial 
me of 
in an 
pher's 
.reGIS 
upport 
id map 
d ina 
1 line 
cus of 
r only 
liti and 
nt role 
.2001). 
Master 
major 
f map 
etween 
:essing 
ouring 
outines 
;eyond 
hat use 
youring 
line or 
to an 
t of an 
(1973) 
as the 
S tools 
hat the 
comes 
  
  
International Archives of the Photogrammetry, Remote Sensing 
from Attneave's (1954; cited by Visvalingham, 1999) theory that 
curvature conveys informative points on lines. Many other pieces of 
research have subsequently enhanced Douglas and Peucker's 
algorithm (e.g. Wang and Muller, 1993 ànd 1998: Visvalingham and 
Whyatt, 1993; Ruas and Plazanet, 1996) in the area of curvature 
approximation applying various thresholds. Oosterom (1995) 
criticized these types of algorithms as time-consuming, so he 
introduced the reactive-tree data structure for line simplification that 
is applicable to seamless and scale-less geographic databases. There 
is still, however, a need for the cartographer's interaction in 
generalizing lines/curves to make them “fit-for-use”. 
A majority of map features are represented as lines or polygons that 
are bounded by lines. Skopeliti and Tsoulos (2001) developed a 
methodology for the parametric description of line shapes and the 
segmentation of lines into homogeneous parts, along with measures 
for the quantification of shape change due to generalization. They 
stated that measures for describing a positional accuracy are 
computed for manually generalized data or cartographically 
acceptable generalization results. Muller et al, (1995) imply that 
ongoing research into line generalization is not being managed 
properly. Most of the research in generalization has focused on single 
cartographic line generalization instead of working on data modelling 
in an object-oriented environment to satisfy database generalization 
requirements. In contrast, other researchers (e.g. Visvalingham and 
Whyatt, 1993) have highlighted a need to evaluate and validate 
existing generalization tools rather than developing new 
generalization algorithms and systems. So far standard GIS software 
applications do not fully support automatic generalization of line 
features. This research focuses on integration and utilization of 
generalization operators using the ArcGIS 8.2 Generalize tool in 
order to generalize a road network database from GEODATA TOPO- 
250K Series 2 to produce smaller scale maps at 1:500,000 and 
1:1000,000. 
40 GENERALIZATION FRAMEWORKS 
An excellent classification of generalization assessment tools based 
on measures, conditions and the interpretation of generalization result 
is provided by Skopeliti and Tsoulos (2001). Peter and Weibel (1999) 
presented a general framework for generalization of vector and raster 
data to achieve more effective translation generalization constraints 
into assessment tools to carry out the necessary generalization 
transformation. Peter (2001) developed a comprehensive set of 
measures that describe geometric and semantic properties of map 
objects. These are the core parts of a generalization workflow from 
initial assessment of the data and basic structural analysis, to 
identification of conflicts and guiding the transformation process via 
the generalization operators, and then qualitative and quantitative 
evaluation of the results. The following discussion provides a critical 
review of the relevant generalization research based on measures, 
constraints or limitations, and integration of measures into the 
generalization process. 
In connection with generalization constraints, Peter (2001) 
categorized constraints based on their function (graphical, 
topological, structural and Gestalt) and spatial application scope 
(object level micro, class level macro, and group of 
objects/region/partition of the database level — meso). The constraints 
relevant to the micro level (object) include minimum distance and 
size (graphical), self-coalescence (graphical), separatability 
(graphical), separation (topological), islands (topological), self- 
intersection (topological), amalgamation (structural), collapsibility 
(Structural), and shape (structural). To assess generalization quality 
for linear features, constraints have been employed (Peter and Weibel, 
1999: Yaolin et al. 2001). Constraints for the micro level (object 
classes) include size ratio (structural), shape (structural), size 
distribution (structural) and aliment/pattern (Gestalt). Finally, Peter 
(2001) divided meso level (objects groups) constraints into 
neighbourhood relationships (topological), spatial context (structural), 
aggregability (structural), auxiliary data (structural), 
a ignment/pattern (Gestalt), and equal treatment (Gestalt). For a 
detailed description of the above constraints readers are referred to 
Peter and Weibel (1999); Skopeliti and Tsoulos (2001); Peter (2001); 
and Jiang and Claramunt (2002). 
and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
n2 
Ww 
In relation to application of measures for the evaluation of 
generalization results, there are several measures to assess 
performance. These can be classified as being either qualitative and 
quantitative methods. To date most of the generalization 
transformation results have been evaluated qualitatively based on 
aesthetic measures. Recently Skopeliti and Tsoulos (2001) developed 
a methodology to assess linear feature integrity by employing 
quantitative. measures that determine if specific constraints are 
satisfied. Researchers began to develop formal approaches that 
integrated generalization constraints and measures for development 
of coherent frameworks and workflows (e.g. Peter and Weibel, 1999; 
Yaolin et al. 2001). In this regard, Skopeliti and Tsoulos (2001) 
incorporated positional accuracy measures to quantitatively describe 
horizontal position and shape, then to assess the positional deviation 
between the original and the generalized line, and to relate this to line 
length after and before the generalization. A technique such as cluster 
analysis (qualitative assessment) was used for the line shape change 
and the averaged Euclidean distance (quantitatively assessment). 
Also, McMaster (2001) discussed two basic measures for 
generalization that include procedural measures and quality 
assessment measures. These measures involve a selection of a 
simplification algorithm, selection of an optimal tolerance value for a 
feature as complexity changes, density of features when performing 
aggregation and typification operations, determining transformation 
of a feature from one scale to another such as polygon to line, and 
computation of the curvature of a line segment to invoke a smoothing 
operation. 
It should be noted that quality assessment measures evaluate both 
individual operations, e.g. the impact of simplification, and the 
overall quality of generalization (i.e. poor, average, excellent). 
Despite all these efforts there is no comprehensive, universal and 
concrete process for generalization measurement techniques. 
However, Ibid (2003) provided a review of existing measurement 
methods for automatic generalization in order to design a new 
conceptual framework that manages the measures of intrinsic 
capability, in order to design and implement a generalization 
measurement library. To apply quantitative measures, Kazemi (2003) 
used two methods of the Radical Law (Pfer and Pillewizer, 1966; 
Muller, 1995) and an interactive accuracy evaluation method to 
assess map derivation. The Radical Law determines the retained 
number of objects for a given scale change and the number of objects 
of the source map (Nakos, 1999). 
While the majority of developed frameworks for the generalization of 
cartographic data, such as those by Lee (1993), Brassel and Weible 
(1998) and Ruas and Plazanet (1996), deliver generic procedural 
information (Peter and Weibel, 1999), the one briefly discussed in 
this paper is designed more specifically for the derivation of multiple 
scale maps from a master road network database (see Kazemi, 2003). 
Large portions of Kazemi's proposed framework may be considered 
generic (e.g. conditions/parameters/constraints definition). However, 
most parts deal specifically with road generalization. Generalization 
operators in the ArcGIS software are tested to generalize roads above 
the conceptual generalization framework for derivative mapping. The 
method is empirically tested with a reference dataset consisting of 
several roads, which were generalized to produce outputs at 
1:500,000 and 1:1,000,000 scales (Ibid. 2003). According to visual 
interpretation, the results show that the derived maps have high 
correlations with the existing small-scale road maps such as the 
Global Map at 1:1,000,000 scale. As the methodology is only tested 
on roads, it is worthwhile to extend it to various other complex 
cartographic datasets such as drainage networks, power lines, and 
sewerage networks, in order to determine the suitability of the 
methodology proposed here. Additionally, various kinds of linear. 
areal and point cartographic entities (e.g. coastlines, rivers, 
vegetation boundaries, administration boundaries, land cover, 
localities, towers, and so on) should also be studied. 
There is no universal semi-automatic cartographic generalization 
process (Costello et a/., 2001; Lee, 2002), because off-the-shelf tools 
do not provide an aesthetically robust and pleasing cartographic 
solution. The current ArcGIS map production tools are significantly