International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
The used classifier uses four features that are calculated for each
segment:
e A measurement for the shadiness,
e a measurement for the uniformity,
e a measurement for the contour angularity and
e a measurement for straight contour lines.
The shadiness is calculated by use of the grey value histogram
see in figure 5. The threshold ¢ in equation | is defined as the
first local minimum in the grey value histogram H(i), where ?
is the grey value. The measurement is the ratio of shade pixels
to all pixels in the image. Buildings cause a characteristic high
shadiness in inhabited areas regions.
t ;
t HG)
shadiness = = (1)
X HO
The uniformity is calculated by a mean a of the local variances
c2 of the image grey values. In figure 5 a contour image is shown
that is calculated from the local variance matrix with a threshold
decision.
uniformity = Mma (02) (2)
The calculation of a measurement of angularity is based on a lin-
ear Hough transformation for straight lines ( Duda and Hart, 1972)
of the contour image (cp. Hough space in figure 5). A straight
line forms a local maximum in the Hough space, orthogonal lines
form maxima in the Hough space. The values of the Hough space
are totalized over the angle and illustrated as a histogram in fig-
ure 5. The discrete Fourier transformation (DFT) is calculated
in the next step and the DFT (k = 2) is taken as a measurement
for the angularity, because vertical lines have an angle distance of
90? in the Hough space.
DFT(k — 2)
imagesize?
(3)
angularity —
The calculation of a measurement of straight contour lines
M SCL is done by searching local maxima in the Hough space
H S (x, y) because straight lines shape a local maximum in the
Hough space. The results of the region based classifier are shown
in figure 7 and discussed in chapter 4. The results for the years
1988 and 2001 are achieved in two ways with and without use of
a previous classification.
S acis HS(z, y) (4)
(size(H S(x, y))?
MSCL =
parallel sun beams \
. \ \
south
—
/ zenith angle sz
cast,
A
pad azimut angle sa ut :
7 >
/
WES
nadow area
/
e
*
north
Figure 6: Illumination Model for Buildings
3.1.2 Linear Regression Classification The features of each
region are combined in a feature vector that is basis for the linear
regression classification algorithm method described in (Meyer-
Brötz and Schürmann, 1970). The method was implemented by
using the scilab (Gomez, 1998) environment. The a priori prob-
abilities are taken from the state transition diagram considering
the previous classification. They are empirical set, because the
random sample was limited. Result of the classifier is a vector
with probability for each differentiated class. GEOAIDA decides
for the most probable class for each segment. Here both segmen-
tations are taken into account.
32 Structural Classifier
The structural building extraction operator models buildings as
complex structures consisting of different parts (cp. (Müller et
al., 2003)). It assumes an illumination model shown in figure 6.
The angles « and A are calculated from the exact date and time
of the image capture and the sun angle. Hypotheses for shades
and roofs are generated using two different image segmentation
operators. To get the buildings, the roofs are grouped with one or
more shades. The neighborhood relations regard the illumination
model presented in figure 6.
Shades of buildings are derived with a threshold decision in the
image. The threshold can be calculated automatically from the
histogram, so that images with different illumination can be pro-
cessed. Since shades are generally not visible in a green color
channel, the green color has been masked during shade detection.
Pixel taken as shade have to fulfill the condition grey value «€
threshold and hue < 90° V hue > 150°.
Roofs are generated in a more complex procedure. Here the so-
called color structure code (Priese et al., 1994), (Rehrmann and
Priese, 1998) is used to segment the entire image. Additionally
greenish areas are masked and roofs are accepted only in the other
parts. An additional size criterion restricts acceptable roof hy-
potheses considering the size.
Shades generated by buildings have a limited area, so for exam-
ple shade near a forest can be excluded. The compactness and
orthogonality of roof labels is additionally measured to validate
buildings. The grouping of shades and roof labels leads to val-
idated buildings. The neighboring position of a shade to a roof
has to fulfill the illumination model. Sometimes it is not possible
to differentiate between the roof of a building and for example an
adjacent parking area. In that case the expected size for a build-
ing is exceeded, does not fulfill the model, and the grouping !5
rejected during the analysis.
1246
Internati
Numbe
U3 nv —
UA 4
To evalua
for a GIS
region-ba
tected but
the regior
figure 7).
The evalu
input ima
in the inp
a detectio
made:
The two 1
of the m:
where TP
son and G
GEOAID,
ing detect
considere
detected: :
of the buil
The devel
be used ir
knowledge
ties for sta
regression
The appro
acquired ii
entiated: |