International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
os database b training areas
:
pixel-based
classification
object-based
classification
Figure 1. Workflow
objects in a GIS database. Both classification steps are based on
a supervised maximum likelihood classification. The training
areas are derived automatically from a GIS database in order to
avoid the time consuming task of manual acquisition. In a pixel-
based classification, the grayscale values of each pixel in
different input channels and possibly some other pre-processed
texture channels are used as input. For the classification of
objects (groups of pixels), we have to define new measures that
can be very simple (for example the mean gray value of all
pixels of an object in a specific channel) but also very complex,
like measures that describe for example the texture, the 3D
shape, the homogeneity or the pixel-based classification result
of an object.
3.2 Preprocessing of the input channels
The input channels for the pixel-based and object-based
classification are multispectral bands, LIDAR data and a texture
channel. In the following, the pre-processing of this data is
described.
3.2.1 Pre-processing of the LIDAR data: In Haala and
Walter (1999) it was shown that the result of a pixel-based
classification can be improved significantly by the combined
use of multispectral and LIDAR data because they have a
complementary “behavior”. There are two different kinds of
information in LIDAR data: (a) the height of the terrain and (b)
the local height of the objects on the terrain. For the
classification only the local height is used. Therefore we use a
Digital Terrain Model and “subtract” it from the LIDAR data.
The result is a normalized DHM that contains only the local
height of the objects (see Figure 2).
LIDAR DTM
normalized DHM
Figure 2. Pre-processing of LIDAR data
724
3.2.2 Texture channel: Ditferent landuse classes cannot be
distinguished only by their spectral behavior but also because of
different textures. In our approach, we use a texture operator
based on a co-occurrence matrix that measures the contrast in a
5 * 5 pixel window. Figure 3 shows the used texture operator in
an example. The input image is shown in Figure 3 a, the texture
(calculated from the blue band) in Figure 3 b.
a) b)
Figure 3. Calculation of the texture
3.3 Pixel-based classification
The result of the pixel-based classification is used as an
additional channel for the object-based classification. The
training areas are derived from an already existing GIS database
in order to avoid the time consuming task of manual
acquisition. This can be done, if it is assumed that the number
of changes in the real world is very small compared with the
number of all GIS objects in the database. That assumption can
be seen às true because we want to realise update cycles in the
range of several months.
In the pixel-based classification all pixels have to be classified
into the landuse classes: houses, streets, greenland. trees and
water. The GIS data (ATKIS; see section 4) that is used in this
project is captured in the scale 1:25.000 and does not contain
the object class houses. Therefore a heuristic is needed that
derive automatically training areas for this class.
The training areas for houses are derived from the object classes
of the GIS database that are representing settlements. In ATKIS
there exist four settlement object classes: residential areas.
industrial areas, areas of mixed use and areas of special use.
The following three conditions are used to select pixels that are
representing houses: (1) the pixels must be located in one of the
four object classes for settlements, (2) the pixels must have a
local height above the terrain and (3) the pixels must not
represent vegetation.
Figure 4 shows the process chain to calculate the training areas
for houses. First, all pixels are selected which are representing
settlement objects. From these pixels all pixels are selected
which are above the ground. A vegetation index is used in order
to eliminate all pixels that are representing vegetation. The
result contains only pixels that are likely to represent roofs of
houses. Figure 5 shows the training areas for houses on an
example. The input image is shown in Figure 5 a and the
calculated training areas in Figure 5 b.
Internati
settle
obi:
pix
norms
DH
Figur
34 Ob
The bas
single pi
objects
dimensic
based or
n-dimen:
appearar
automati
3.4.1
between
first to
classes.
to decid
industria
German:
"- ft^ Ss
Not all
only thre
