International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
apparent resemblance in shape and position. The discrepancies
between the data sets based on the different ways of
acquisition, modelling and updating have been described at the
beginning. But due to the diversity in digitizing the analogue
geoscientific source maps and the data modelling of ATKIS,
objects representing the same real-world objects differ in the
number and geometry of segments (see Fig. 2)
Thus, investigating corresponding partners between the ATKIS
and the geoscientific data sets, would lead not only to
unsatisfying results but to relation errors. Therefore the
investigation for corresponding objects has to be performed
based on the aggregation of segments.
Using an overlapping test and by evaluating the overlap-area
composed to the area of the segments to be tested, selection sets
will be build, these selection sets will be stored as aggregated
groups (with 1 to n elements). In order to find valid
correspondences, all possible pairs of combination of neighbour
objects will be checked against each other in the search process
(see Fig. 3). Alternatively, we can use a breadth search
procedure for finding the object clusters.
Fig. 2 : Segmented objects from the reference data set ATKIS
(left image), and from the geological map (right
image).
In order to define the neighborhood, either a buffer with a fixed
distance or a triangulation can be used. A parameter free
approach to identify clusters is based on an hierarchy of
neighborhood graphs (Anders 2003).
5.3.2 Geometry based matching
The matching of the selection sets (e.g. the aggregated
segments) will be checked individually using different
measures.
In the current prototype the following measures for determining
object similarity are used:
eHausdorff distance: The length of the greatest local
deviation between the two shapes. The lower the
deviation, the higher the score.
eSymmetric difference: The areas found in one shape
only. The more the two shapes overlap, the lower the
symmetric difference, and the higher the score.
eCompactness difference: The difference between each
shape's compactness, which is the area-to-perimeter
ratio. The more similar the compactness of the two
shapes, the higher the score.
eAngle Histogramm: The difference between each
shape's angle histogram, which is a histogram of the
angles that the segments make with the positive x-
axis, weighted by segment length. The more similar
the histograms for the two shapes, the higher the
score.
For each geometric criterion a result between 0 and 1 is
calculated and the mean value for each correspondence is
evaluated. Different combinations of segments from the
selection set of one data set are tested with the corresponding
selection set (e.g the combinations of segments) from another
data set . The highest result between to segment combinations
will be kept as link. This process will be repeated until no more
appropriate links can be established.
| data set A |
(geometry based matching )
uiu
E rm
GNE
| data set B |
Fig. 3 : Selection set for geometry based matching between
objects from two different data sets.
Once the correspondences between the selection sets have been
found in the matching step, it has to be decided, whether the
objects correspond exactly or if they differ due to update
processes, which have been applied to one data set, but not to
the other one. The automatic investigated links will be
visualized to the operator, but before the next step — the change
detection — will be performed, a manual correction of the links
will be possible. Depending on geometric descrepancies,
different types of change can be identified (see section 6.1).
6. CHANGE DETECTION
Objects which have been selected through geometric integration
and have been considered as a matching pair could be
investigated for change detection. A simple intersection of
corresponding objects is used for the change detection. Yet, the
mentioned differences may cause even more problems which
are visible as discrepancies in position, scale and shape. These
discrepancies will lead to unsatisfying results and make the
evaluation of the resulting elements almost impossible (Fig. 4).
Therefore firstly, a local transformation will be applied, leading
to a better geometric correspondence of the objects. To this end,
the iterative closest point algorithm (ICP) developed by (Besl &
1252
Internati
McKay,
between
elements
In our fi
reference
objects fi
the ATKI
Fig. 4 : Re
At the end
the given
corresponc
The ICP €
geometric
geoscienti
adaptation
6.1 Inter:
Following
detection '
simple inte
into differc
be conduct
and soil sci
At this tin
identified:
to their res;
9. I
n
e. .T
a
rt
c
