"TON
Air
ing the pixel-
ut the image
iowledge and
TA algorithm
en supervised
likelihood. On
tation, several
ome from the
r's and user's
1 the detailed
ENTED
S
classification
els in an imag
ıltispectral data
ed, the spectral
is used as the
ifferent feature
based on their
pperties. In thi
ic in character.
tof radiance
ngth bands fo
o the family of
pixel-by-pixd
ited land cover
issical plus one
been employd
nimum-distance,
led information
in the following
) software based
n the eCognition
ects. This meam
Jould ideally X
shape combined
ties can be used
ig the general
ed within à cles
er-class and ds
-classes or to lé
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
subclass (es). With, respect to the multi-scale behaviour of the
objects to detect, a number of small objects can be aggregated
to form larger objects constructing a semantic hierarchy.
Likewise, a large object can be split into a number of smaller
objects which basically leads to two main approaches of image
analysis: A top-down and a bottom-up approach (see
eCognition User Guide, 2003 and Benz, U., et al., 2003). In
eCognition both approaches can be realised performing the
following steps:
o Creating a hierarchical network of image objects
using the multi-resolution segmentation. The upper-
level image segments represent small-scale objects
while the lower-level segments represent large-scale
objects.
o Classifying the derived objects by their physical
properties. This also means that the class names and
the class hierarchy are representative with respect to
two aspects: the mapped real-world and the image
objects’ physically measurable attributes. Using
inheritance mechanisms accelerates the classification
task while making it more transparent at the same
time.
* Describing the (semantic) relationships of the
network's objects in terms of neighbourhood
relationships or being a sub- or super-object. This
usually leads to an improvement of the physical
classification res. the class hierarchy.
* Aggregating the classified objects to semantic groups
which can be used further for a so called
‘classification-based’ segmentation. The derived
contiguous segments then can be exported and used in
GIS. The semantic groups can also be used for further
neighbourhood analyses.
These steps describe the usual proceeding when working with
eCognition. While the first two steps are a mandatory, the latter
two steps may be advisable according to the user's objectives
and content of the image. :
In the segmentation phase, following parameters should be
assigned as accurate as possible, of course, suiting with the
reality.
Scale parameter: this parameter indirectly influences the
average object size. In fact this parameter determines the
maximal allowed heterogeneity of the objects. The larger the
scale parameter the larger the objects become.
Color/Shape: with these parameters the influence of color vs.
shape homogeneity on the object generation can be adjusted.
The higher the shape criterion the less spectral homogeneity
influences the object generation.
Smoothness/Compactness: when the shape criterion is larger
than 0 the user can determine whether the objects shall become
more compact (fringed) or more smooth.
Segmentation phase is folloed by the classification of images.
¢Coginition software offers two basic classifiers: a nearest
neighbour classifier and fuzzy membership functions. Both act
as class descriptors. While the nearest neighbour classifier
describes the classes to detect by sample objects for each class
Which the user has to determine, fuzzy membership functions
describe intervals of feature characteristics wherein the objects
do belong to a certain class or not by a certain degree. Thereby
Sach feature offered by eCognition can be used either to
describe fuzzy membership functions or to determine the
feature space for the nearest neighbour classifier. A class then is
described by combining one or more class descriptors by means
of fuzzy-logic operators or by means of inheritance or a
combination of both (see Figure 1). As the class hierarchy
should reflect the image content with respect to scale the
creation of level classes is very useful. These classes represent
the generated levels derived from the image segmentation and
are simply described by formulating their belonging to a certain
level. Classes which only occur within these levels inherit this
property from the level classes. This technique usually helps to
clearly structure the class hierarchy.
P
/
/
>» i M \
Figure 1. Hierarchical network of image
6
3. STUDY AREA AND IMAGE DATA
The test site is Zonguldak and its close vicinity, located in
Western Black Sea region of Turkey. It is famous with being
one of the main coal mining area in the world. Although losing
economical interest, there are several coal mines still active in
the region. Testfield has a rolling topography, in some parts,
with steep and rugged terrain. While urbanized part is located
alongside the sea coast, there are agricultural and forested areas
inner regions. The elevation ranges roughly up to 1800m.
Figure 2. Landsat ETM+ image (band 3,2,1) of the study area
For the analysis, Landsat-7 ETM image (see Figure 2) covering
the test site and taken on 04.07.2000 has been utilised. At the
processing phase, all spectral channels expect thermal one were
used and their properties are given in Table 1. Reference
datasets employed during the classification procedures of
1119