e, distinction
between one-
of major roof
- L-, U-, and
s DSMs can
s and ortho-
in be used to
in the case in
ages become
and sufficient
vhich employ
| photogram-
mage or ob-
Iding bound-
le. To avoid
| should have
grid spacing
are close to
f discontinu-
ze should be
in be derived
bination with
jy the use of
s [Grün 1985,
rg.
yossible build-
»hological op-
uring element
ave problems
> buildings, or
raction of the
As, if they are
and accuracy.
ect low build-
5 outlines but
ctures with a
d borders, are
ing the DSM
a certain size.
igh the DSM.
gions that are
an be applied
ple and fast,
naximum and
known height
the estimated
- 4 m). The
s that detect
coarse detec-
uilding model
esults of this
efer to [Balt-
(A)
Figure 2: (A) ortho-image, (B-D) height bins of the Digital
Surface Model (DSM) with 1, 2, and 3 meter size (quan-
tization). By decreasing the bin size a better modeling of
the buildings is achieved. In addition, gabled roofs, T- and
L-shaped buildings, and buildings close to each other can be
distinguished.
4.3 Classification of 3-D Blobs
Objects other than buildings will often be detected as blobs,
for example trees, bridges/over-passes, transportation means,
and big poles. A first elimination of non-building blobs is
performed based on the area, height and minimum dimensions
of the detected blobs. A further separation can be achieved
by using the number and length of extracted straight lines as
well as the size and shape of compact homogeneous regions
within the projected blobs, the weighted histogram of the
local gradient orientation, spectral properties, and context.
Vegetation blobs, in particular trees, are the most prominent
non-building blobs that must be detected and eliminated.
Apart from using spectral information to separate trees from
buildings (see below), we propose a simple procedure which is
based on weighted local orientation histograms. A histogram
of the local orientations of all edge pixels within the pro-
jected blob region is computed. Each entry is weighted with
its magnitude. Assuming regularly shaped buildings, the his-
tograms of building blobs will often contain significant peaks
90° apart. Histograms of more complex buildings contain a
few additional peaks (usually one or two). On the contrary,
histograms of tree blobs are predominantly flat. For details
on the approach we refer to [Baltsavias et al. 1995].
4.4 Combining Color and False Color Infrared Images
with 3-D Blobs
In addition to the above rather simple procedures for blob
classification, we have also investigated into the use of color
and infrared images together with DSM blobs to separate
man-made (MMOs) from natural objects (NOs) [Sibiryakov
1996]. The RGB images are initially transformed into a more
suitable color space — the CIE (1976) L'a'b' color space
(abbr. CIELAB) [Wyszecki and Stiles 1982]. The CIELAB
color space separates the luminant and chromatic compo-
nents of color and is perceptually uniform. In uniform color
spaces, perceptual color differences are computed with Eu-
clidean distances.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
In the following analysis we use only the chromatic compo-
nents a* and b*. The lightness component L” was not used
in the classification, because different parts of a roof may
have different lightness, however, the same chromatic prop-
erties. The CIELAB color space allows us to describe colors
more similar to what is perceived by human beings, which is
very useful in handling images under non-uniform illumina-
tion conditions such as shade, highlight, and strong contrast.
A simple classification of the object classes (roads, buildings,
vegetation, cars etc.) is not possible using only color images,
because the different object classes overlap considerably as
can be seen in Fig. 4A. Especially objects with low chroma,
such as roads, shadows, trees, brown or grey roofs, and pa-
tios, overlap in their chromatic components.
A comparison between color and false color infrared images
(CIR) showed, as expected, that a separation between natural
and man-made objects is easier with CIR images. Figure 4B
shows the main clusters for the CIR image in the a" and b^
color components. CIR images have essentially three major
spectral classes: vegetation, man-made objects and bare soil,
and water. Roughly, the man-made objects form a high and
well-separated peak in the histogram of the a” channel, thus a
simple thresholding can be used for their detection. With such
a simple approach, bare soil mixes with man-made objects.
Color images have a larger spectral variability and are more
appropriate than CIR images when a larger number of classes
must be determined.
An unsupervised classification based on a simple k-mean clus-
tering is used for both the color and CIR images, where k is
the number of predefined classes. The clustering is based on
minimum distance (Euclidean distance was used). The clas-
sification is iterative in a binary tree fashion and it employs
1 - 3 classification steps, see Fig. 3.
a* and b* color components
of the original image
kz2
step 1 man-made objects natural objects
k=3
non-buildings
k=3
\
step 2 buildings
| k=3
step 3 | non-buildings : r Ï - | buildings buildings
T
I
Y Y |
end result through combination
of partial classification results
non-buildings |
and class region editing
Figure 3: An iterative, binary-tree classification scheme using
unsupervised k-mean minimum distance clustering, where k is
the number of predefined classes. When k=3, the third class
is the rejection class. The binary tree can be further densified
or reduced. The dashed lines show optional processing steps.
In the first classification step, the number of classes is 2 (NO
and MMO). In the second step the MMO image is selected
and k=3, i.e. buildings, non-buildings, and a third rejection
class. The aim of this step is to separate buildings from other
MMO, especially roads, or NO that correlate with MMO like
bare soil. This is successful to a large extent but some class
mixing does occur, e.g. some buildings are still included in
the non-building class. Thus, a third classification step is
323
