IS systems 
key to the 
tive of this 
as: (a) The 
elongation, 
This paper 
scussed are 
| by using 
s from the 
1998], the 
y using the 
studies are 
roving the 
detecting 
lating GIS 
y was the 
dating the 
combining 
ently and 
unly, in the 
tive, a rule 
system was 
Taken into 
metric and 
ers of the 
gation, etc; 
ects of the 
well. 
carried out 
" approach. 
stection was 
sions which 
holds of the 
of change to 
ss) for each 
lentification.” 
'ameters: (a) 
uch as area, 
(c) Textural 
y; and (d) 
he objects. 
defined for 
eveloped to 
a building 1s 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
described by: specific spectral parameters, small area, compact 
shape, access way, etc. On the other hand, a street is described 
by: specific spectral parameters, elongated shape, connected to 
other streets, moderate slope, etc. 
The algorithm for change identification was developed as a 
hierarchical decision tree, to integrate the different rules. The 
algorithm is implemented in two steps. The first step is to mark 
objects and clusters of pixels which were "changed". The 
detection was implemented in "object-wise" manner rather than 
single pixel treatment, as used in traditional methods. The 
labeling process is done by using the region growing 
segmentation method, which takes into consideration. four 
different data sources: change intensity index, edge gradient, 
radiometric data and the objects from the existing spatial 
database. The second step is to identify the type of "change" 
using the set of rules. 
3. DATA SOURCES 
In this research, three data sources were used: National GIS 
spatial data layers, including the hypsographic data; Orthophoto 
generated from color photographs, for two different epochs; 
Multi-spectral IKONOS imagery. All of these data were fused 
and integrated into one spatial database, which was developed 
for the research experiments. The types of objects to be treated 
and tested were determined during the integration process. In 
addition, a quality control process [Peled, Gilichinski, 2004] 
was implemented to test the GIS spatial database. The results of 
the quality control process indicated serious problems especially 
in the values of the TYPE-CODE attribute for general objects 
and of the WIDTH attribute for road links. The detection of 
errors, their correction and removal, were essential to the 
system learning process and the specific spectral determination 
for each type-code processing. 
4. EXPERIMENTS 
4.1 Change Detection 
The change detection process was implemented by comparing 
color orthophotos from different epochs. Different factors of 
detecting the "regions of changes" were tested. These factors 
are: (a) Change detection method category, whether it is a-pre- 
classification or post-classification method; (b) Comparison 
implementation method for single pixels or for a small window 
around the single pixel; (c) Comparison between the RGB 
components or between other color components, such as HLS, 
HSI, L*a*b*, etc.; (d) Radiometric normalization; (e) Noise 
removal; (f) Methods for defining the "regions of changes" and 
the determination of thresholds for defining "significant" 
changes. The experiments included many combinations of these 
factors. Figure 1, illustrates some of the results of these 
experiments. According to these procedures two conclusions 
were made. The first was to use the Euclidean spectral distance 
in the L*a*b* color space to define the change magnitude. The 
second conclusion was that the optimal threshold to define a 
significant change is close to the value of one standard deviation 
of the change magnitude for all pixels in the research area. 
499 
4.2 Rule-Based System 
The rule-based system which was developed includes sets of 
rules which supply a unique description for each type of objects. 
These sets of rules integrate radiometric, geometric, textural and 
topological parameters. The radiometric parameters include the 
distribution and other statistical parameters of the grey level 
values of each band for the pixels within each object (see figure 
2). These parameters were computed only for objects within 
"no-change" regions. The geometric parameters include 
descriptors which define the geometrical characteristics of the 
object, such as area, perimeter, elongation, compactness, 
moments of inertia, etc. The textural parameters describe the 
textural template of grey level values for each band, such as: 
contrast and homogeneity. The topological parameters include 
topological and spatial relations between the objects from 
different types. These relations take into consideration instances 
such as if the object is within urban, rural, industrial, flat or 
mountainous zones. 
4.3 Segmentation 
In the segmentation algorithm distinguished are two groups of 
changed objects: (1) Objects, in the spatial database, that were 
changed totally; and (2) Clusters of pixels that were changed 
and are only part of objects within the ‘old’ database. 
4.3.1 Whole Objects 
Two methods were implemented to detect whole objects which 
were changed. In the first method, a change index was defined 
for each object in the existing spatial database. This index is 
calculated by averaging the change magnitude for each pixel 
within the object. If the value of this index exceeds specified 
criteria then the related object will be tagged as a changed 
object. Figure 3 shows some examples of objects which were 
tagged as "totally changed". In the second method, a quality 
control process is performed. The quality control process 
detects objects which have irregular radiometric parameters. 
While building the rule-based system, for each type of objects 
were calculated the average and the standard deviation values of 
the histograms of grey level values in each radiometric band. 
For each grey level a range of normal population were 
determined by the average + 20 (0 = standard deviation). The 
quality control process counts, for each object, the number of 
grey level ranges which have population (percent of pixels) 
outside this range of “normal population”. If the counting 
results exceeded a specified threshold then the treated object 
was marked as "incorrect" type or "changed" object. 
4.3.2 Clusters of Single Pixels 
After the process of detecting whole objects which where 
changed, single pixels were tagged if the average of the change 
magnitude of the neighborhood pixels exceeded a specified 
threshold. The neighborhood pixels are defined by an operator 
of 9X9 pixels in size. This process is implemented for each 
pixel which falls within a specified region or any large area 
objects. The objective of defining this type of pixels is to detect 
whether a small part of large-area-objects was changed without 
affecting the change index (defined for the whole object). The 
segmentation process is implemented for the segmentation of 
clusters of these single pixels and their partitioning into separate 
objects. The newly developed segmentation algorithm is based 
on the region growing method. Seed pixels are selected