flight height
age scale of
his so-callei
available for
or the Swis
1
er
'art of DSM
image of the
>construction
ips 1:25°000
s for building
be added. Al
1 data for the
DSM (differ
l') would als
value is 0 (Fig. 2).
Markus Niederöst
allow simpler techniques like a thresholding in the normalized DSM. On the other hand the procedure described below is
also applicable in projects where only the DSM is available and thus more general.
3.2.1 Calculation of height bins: The height information in
the format of a regular grid of z-coordinates is cut into binary
height bins, each of them corresponding to one height interval.
The chosen height difference between two bins is 1 m. Genera-
tion of height bins with a thickness of 1.5 m results in an over-
lap between adjacent bins to avoid problems in the subsequent
steps caused by inaccuracy of the DSM or the normalized DSM
respectively. Each rasterpoint in a bin equals 1 if its z-coordi-
nate lies in the height interval of the current bin, otherwise the
3.2.2 Find elements in bins: The bright objects consisting of a
group of adjacent white pixels bounded by black ones are
detected and for each element an enclosing rectangle (Fig. 2) as
well as the centre of this rectangle are determined, providing
one list of elements ordered from the highest to the lowest bin.
*ER Nu | idi Oe VE.
Fig. 2: Four consecutive slices for a part of the test
area (heights h1>h2>h3>h4) and enclosing
3.2.3 Tracking along z: In order to detect the blobs, the list of boxes for one building
elements is processed from the highest bin down to the lowest
bin. If through a sequence of several height bins a chain of elements can be found at the same location (distance of
centres of gravity < 7 m in object space), a possible blob is detected and described by the second highest enclosing
rectangle of the chain. The size of this rectangle approximatively corresponds to the building dimensions.
Main problem of the blob detection is that buildings cannot be separated from vegetation. Furthermore buildings
standing close to each other or trees standing too close to buildings are detected as one blob.
The blob detection results in one list of northwards oriented rectangles as input vector data for the house reconstruction.
3.3 Multichannel classification
The two color channels R (red) and G (green) from the RGB color space, the channel a* (redness-greenness) and L*
(brightness) of the CIELAB color space and the channel S (saturation) of the HSI color space have been used to calculate
several channels for the classification procedure. The channels of CIELAB and HSI can be derived directly from the
RGB color space (Pratt, 1991). Adding one channel containing the normalized DSM, the following channels were used:
* Channel containing shadows, derived from the S channel of the HSI color space by thresholding
* Channel a* from the CIELAB color space
* Channel containing texture. Texture means the number of edge pixels in a circle around the centre pixel (diameter
8 m) as proposed in Braun (1999),
* Channel containing the degree of artificiality (DoA). The result of (G-R)(G-R) was first inverted and then a
histogram equalization was done. The brighter pixels belong to man-made objects (high grade of artificiality),
while darker pixels are part of vegetation (see Fig. 5).
* Channel containing the normalized DSM (nDSM)
All channels were stretched to the interval [0, 255]. The classification which is an improvement of the procedure
described in (Braun, 1999) consists of the following steps:
* Removal of shadows (using channel S). Regions with shadows are excluded from further processes.
Unsupervised K-means classification (2 classes) to separate above-ground regions from ground regions (input
channels a* and nDSM). Regions classified as ground regions are excluded from further processes.
Unsupervised K-means classification (2 classes) of above ground regions to separate man-made objects from
vegetation regions (input channels texture and degree of artificiality). Regions classified as vegetation regions are
excluded from further processes.
Postprocessing to remove small objects and to fill small gaps
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 637