Markus Niederöst
Fig. 3: Orthophoto of the test area Fig. 4: Classification result
Application of the multichannel classification procedure to the test area (Fig. 3) produced a result which shows a clear
separation of buildings from other image content (Fig. 4). Although trees close to buildings could not be completely
eliminated and shapes of buildings are often not accurate, the result is sufficient for being used in further steps.
In order to provide data which can be used as approximate data for the reconstruction, the classification result is
processed to get a list of bounding boxes, including one bounding box for each classified building. Those northward
oriented rectancles serve as input vector data for the building reconstruction.
4 BUILDING RECONSTRUCTION
The building reconstruction process is done for
each initial building seperately. Input for the
reconstruction can either be the approximate
vector data provided by L+T (VECTOR25),
the result from the blob detection or a bound-
ing box derived from the classification result.
The building reconstruction steps described
below are done using image data based on the
calculated color orthophoto and height infor-
mation. Those are the degree of artificiality
(DoA) described in '3.3 Multichannel classifi-
cation’, the absolute height values from the
DSM, the normalized DSM and the L* chan-
nel (brightness) which is used to calculate ori- :
entation and magnitude of the grey value Fig. 5:
Used data for building reconstruction:
gradients using the Sobel operator (Fig. 5).
a) L* channel (brightness)
: b) magnitude of grey value gradients (from L*)
This new approach profits of the fact that a C) orientation of grey value gradients (from L*)
multiplication of the normalized DSM with the dyrormalized DSM
degree of artificiality results in a separation of
artificial objects above ground from all other
objects.
e) degree of artificiality (from R and G)
f) multiplication of the normalized DSM with the
degree of artificiality
4.1 Translation
First step of the building reconstruction is the improvement of the planar position. A binary mask (Streilein, 1999)
calculated from the approximate vector data is moved on a regular raster (rasterwidth 2.5 m), going from -10 m to 10m
in x- and y-direction. On each of those 81 positions the following score (Eq. 1) is calculated:
638 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
SC
m =
wy
dx,
Do.
The wei
weight
decreast
reduce t
the initi:
rect. Thi
the neig
the dist:
caused :
data is n
42 Rec
Usually
approxii
detectio
For the
vectors i
and the
essentia.
discrete
sum giv
Fig. 7:
43 Sc
The resi
building
position
shape) i:
is neces:
of the ve
further :
similar
artificial