nbul 2004
nd 1:5000
xpect that
coincide'.
functions
jus result
ar spatial
enced by
avoid this
oes more
ited, their
s of the
; checked
a set that
mportant.
>d out.
| that has
f interest
ction’ is
rlapping)
object is
and the
ion if the
ken using
nore than
lered the
using the
functions
rst step),
oof the
area).
S
always
gregation
become
1g only
ICES), no
hich can
ata set at
> of two
iderlying
sets (in
have the
different
propriate
duce the
nding on
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
what types of data are available, different algorithms can be
developed. In our approach a low-resolution data set can be
generated from a high-resolution data set by using a low-
resolution data set containing a different theme.
For the scope of this paper we concentrate on a particular case
(see also Section 5), i.e. two data sets in two different scales
(high resolution and low resolution) and with slightly different
themes. The goal is to obtain a new data set that has the
resolution (scale) of the first data set but the theme (objects) of
the second data set. We assume that the themes are similar, i.e.
at certain classification level, the objects can be matched. The
interest here was primarily on the applicability of spatial
operations. We have established the cross-references between
the themes of the two data sets manually, which is the most
common way to link objects at the thematic level. The
algorithm for deriving a new representation can be specified as
follows:
e For each object with a particular theme, establish link
with all the objects from the second data sets that
interact (following the algorithm described in Section
3) AND have a similar theme. For example ‘river’ and
‘bottom of river’
e Create an ‘aggregation’ of the objects of the second
data set that fall in one object (from data set 1).
* Assign to the objects of data set 1, the theme
properties of data set 2. Only matched objects have to
be considered.
The algorithm was implemented in Oracle Spatial 9i (using
PL/SQL) and the spatial functions SDO AGGR UNION (for
the aggregation of objects).
5 CASE STUDY
The two algorithms were tested on three data sets named here
DSI (scale 1:1000), DS2 (scale 1:1000) and DS3 (scale 1:50000
with the following characteristics:
e DSI and DS2 have different objects and different
geometries, but the scale is the same. DS1 has been
created on the basis of topographic boundaries (grass,
river, forest, etc.), while the DS2 on the basis of
maintenance characteristics (road, facilities, etc.). In
principle, it is possible to have boundaries in DS2 that
do not follow topographic boundaries.
* DS2 and DS3 are both designed for maintenance
purposes. Consequently the theme of the objects is the
same (given with a unique code) however the scale is
different (see Figure 1). The polygons of the DS3 are
defined with much less points compared to the
polygons in DS2.
More details on the data sets and the tests can be found in
Zlatanova et al 2003 and Binkhorst and Zlatanova, 2003.
The three datasets are produced and maintained by three
different organisations within the Ministry of Transport and
Public Works on different platforms and in different systems.
This case study is part of a larger project on usability of Oracle
Spatial for the support of infrastructure work-processes within
the Ministry. The three data sets were imported in Oracle
Spatial 9i in three different tables (using SDO GEOMETRY
data types), indexed (R-tree spatial index) and validated (as
specified in Section 2).
We tested the fist algorithm for all the objects belonging to
DSI&DS2 and DS2&DS3. Since the objects of the maintenance
225
maps (DS2 and DS3) have the same theme, linking the objects
was tested also considering thematic codes.
The results of the tests can be summarised as follows:
DS2&DS3: These two maps supposed to have objects that can
be classified as having relations 1:1 (ie. Group 1, Section 3).
The statistics show that using the geometrical procedure about
85% of the objects of the two representations can be
automatically linked, i.e. one object from the DS2 is matched
with only one object from DS3. The tests have clearly shown
that the threshold for overlapping areas can be very relaxed, all
the objects with overlapping area larger that 60% can be
considered as one object. In only few cases the overlapping area
was smaller than 60% although the objects were the same (e.g.
long, narrow objects). How many objects from the DS3 are not
matched with any object was difficult to check, since the area
covered by the DS3 was much larger. There were several
reasons for not matching the remaining 15%, but they can be
mostly considered as errors. For example, one of the objects in
DS3 covers two objects in DS2, or missing objects (Figure 2).
The results of this algorithm were better compared to the results
from comparing theme codes (due to typing errors in the text
string).
P: m SA
A j NY a
Pd %
PA LA uen
3 Ee pms ^s =
p nup Nen
A PA 4 ^» us €
xe E P T emt
REA L 2 e. fie Se
PA : ts rT rod
2 d
Pd es e e
pa : SL rat Te et
se
Lie oe
her hs 2 gems
mem
aem
ittm nid
Me Rep
eco a
Figure 2: Missing object in one of the data sets (the thick
polygon)
DS1&DS2: In this case, the thematic definition of objects was
different and therefore there were more cases of the last three
groups defined in Section 3. For example an object “river side”
from DS2 may overlap with “grass” and “trees” from the DS].
Clearly, when a DS1 object has 100% overlap with DS2 object
then the object is completely ‘inside’ the DS2 object (i.e. group
2) and can be linked to it. It has to be a multi-step process in
which first objects that have a 100% match can be linked, then
link ‘obvious’ objects (those which are e.g. 60% or 70% or
more inside a matching object). The remaining objects will
require some additional ‘rules’ (e.g. assign the object to the
matching object with the largest overlap if this overlap is more
than 50% of the input object). We expect some objects will
remain which cannot be matched. This may require a match
between the objects defined in a different way. The last step of
the algorithm was not tested. We have matched all the objects
that have an overlap of 100% (58%), larger than 80% (70%) and
larger than 60% (74%).
Similarly to the second algorithm, a simplification of the objects
match between DS1 and DS2 will be achieved applying an
object aggregation with respect to theme classification. For
example, the river in DS2 is subdivided into several additional
parts with respect to the usage. All these parts (which interact)
are first united (the thick polygon in Figure 3, up). DSI does not
have theme subdivision of rives, but it has subdivision with
respect to data collection procedure (aerial stereo-pairs). Figure
3 (down) illustrates the resulting new object.
