2.6 Thomas Löcherbach
This work presents a method which adjusts prior GIS-
information to given image data. In an adjustment process,
the GIS-data is fit to the edges extracted from the images.
The aim is the reconstruction of the geometry of land-use
parcels and their classification.
Data Source: Multispectral image and GIS-information
(field boundaries)
Object Model: The object model contains 3D polygons,
representing the geometry of the land-use units. The radio-
metric part consists of one feature vector per object, e.g. the
field mean, within-field variance, or a field histogram.
. The geometry of the image model is represented by a 2D
polygon network. The radiometric part contains the assump-
tion of homogeneous features within a given object and a fea-
ture edge model describing the transition between two neigh-
boring fields along the land-use boundary.
The observations of the adjustment process are the intensity
values of the images and the map coordinates as prior in-
formation. The correct object coordinates, the transformation
between image and object space, and the feature vectors per
fields are derived in an estimation procedure.
Prior Knowledge: The GIS-data is used as prior informa-
tion.
Strategy: Aim of the procedure is the estimation of the ge-
ometry of the land-use parcels. A field is assumed to be a
homogeneous area. To set up the observation equations the
transition of the feature of one field to the feature of the neigh-
bouring field is modelled. Therefore the image is partitioned
into regions along each boundary. From the pixels within one
region along a boundary the position of the boundary and
the features of the areas on both sides of the edge may be
estimated. If fields are large compared to the pixel size, the
pixels in the center of the field may be used to estimate the
radiometry, not the geometry of the boundaries. The proce-
dure is an iterative process, where each iteration step results
in a new boundary position.
Differences between map and image may have several rea-
sons: they may be shifted to some degree, there might be
additional boundaries in the map which do not exist (or are
not visible) in the image, or there are boundaries which are
not contained in the map. The procedure aims at a recon-
struction of the elements in the map, thus the 3rd case is
not treated here. Detailed information on the method can be
found in [Lócherbach 1994]
Results: The experiments reveal that the shifts between
the data sets can be adapted very well (conf. Figure 5, up-
per). If, however, large displacements occur, further mod-
elling would be necessary, e.g. to impose constraints con-
cerning the parallelism of paths (conf. Figure 5, lower).
Figure 5: Initial maps (left) and corresponding estimated
maps (right) superimposed on original image
3 Evaluation
Since most of the participants concentrated on the data set
flat only these results will be presented and compared in de-
tail. The evaluation bases on a reference data set which was
measured manually on a digital stereo workstation. The ac-
curacy of the manual measurement can be expected to be in
the range of 0.20 to 0.30 m due to measurement and defini-
tion uncertainty. Thus the results of the participants can be
compared against this reference. The buildings were repre-
sented simply by the 3D-coordinates of their roofs, i.e. each
building is described by 6 coordinates. The comparison is
done based on the differences in x,y, and z-coordinates be-
tween the manual measurement and the individual reports of
the participants.
The following table gives the mean value of the differences
(RMS) in the coordinates of the roofs of the buildings!:
05 |m Oy|M om
el
a
e
DEM-Image-Fusion)
(DEM)
The maximal RMS of all participants lie in the range of 0.4
to 1.5 m. These figures are higher than the expected val-
ues, which is probably due to the definition uncertainty of the
building. The positional accuracy (in x and y) of methods us-
ing image data is higher than those using solely the DEM.
The following general conclusions can be drawn from the
test:
> The problem of detection and recognition can be solved
for the building objects from the range data alone - pro-
vided that the data is dense enough compared to the
object size (cf. Weidner and Fayek). Range data is very
suitable for localization, especially when objects dis-
tinctively emerge from background. Even if not, generic
models help to increase the reliability of object recogni-
tion (e.g. the model that buildings consist of nearly pla-
nar surfaces, or the fit of a parametric building model to
the data set).
! The results of Fayek are not included, since his aim was the detection and not the reconstruction of the buildings.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
772
The i
not k
The
Lóch:
princi
of pa
datak
For tl
in cor
(see :
In sul
data :
searc
speci
varie:
able i
With
ners)
data
resoli
inforr
Thei
out a
[Frits
man-
Ref:
Collir
/
|
Fayel
L
(
Fiere
PM AY
