mantic Model,
of the (up to
‘eature base to
lge in semantic
eated in many
mage contents.
and the image
ile non-verified
image analysis
restellt werden,
natisierung des
len Merkmalen
1 (SN) sind die
dern aufgebaut
AnalyseprozeB
topographische
ersten Schritt
n einem allge-
schritt zu einer
ihren.
re
| and size, e.g.
it object struc-
get a specific
ing alignment,
atural' objects
.g. settlement-
essed in rules,
ecisions. How-
r uncertain in-
unt during the
these features
à human oper-
ated processing
of external knowledge and general rules in structured form is
necessary. Digital topographic databases offer such an exter-
nal knowledge about the location, size, shape and semantic
meaning of objects, which have to be extracted from satellite
image information. In many countries — especially Europe and
North-America — digital databases in this sense are created
or will be finished in the next few years. For example the
surveying administration in Germany digitize specific topo-
graphic map series and aerial photographs to build up differ-
ent digital databases, where ATKIS (Amtlich Topographisch-
Kartographisches Informationssystem) is the most important
one for an integrated processing with satellite image infor-
mation. The relevant part of ATKIS for image analysis
are mainly 3 digital landscape models DLM ( — Digitales
Landschaftsmodell) related to different scales (1 : 25 000,
1: 200 000, 1 : 1 000 000). This work is based on DLM200
(1 : 200 000) because the improved image analysis should be
used for updating the digital database iteratively. Therefore,
the image information should include more details than the
database to be updated. A second reason may be the better
availability of information in this scale all over the world.
In a first step this work concentrates on areal objects of
DLM200, which are main classes of permanent usage, like
settlements, forests, water areas and special agricultural ar-
eas (meadows, vineyards etc.). For these objects we have
a-priori a complete geometrical and semantical description
for the whole image area. Therefore, by this symbolic formu-
lated knowledge the integration of a model-driven top-down
approach is possible, where in conventional satellite image
analysis data-driven bottom-up methods are common.
In contrast to this conventional methods, automated feature
extraction and determination of semantic meaning has to take
quite new aspects into account. Up to now an human opera-
tor chooses interactively only few representative training areas
during a supervised classification. This happens by visual in-
terpretation and judgement of the operator based on his/her
experience. 'Clean’ training areas are chosen, i.e. without dis-
turbance or corruption, e.g. by mixture of classes (parks/lakes
inside settlements), geometrical errors (displacements) at the
edge of an object or digitizing errors. With DLM200 informa-
tion all objects inside the image can be used as training areas
to extract the actual values of the features out of the image,
but the above mentioned disturbances have to be excluded.
Therefore, a robust estimation of the features must be car-
ried out. On the base of the extreme large samples an error
tolerant method can be created (with the assumption that a
majority of the whole concerned pixels in the DLM-objects
are valid for the related class).
Because of the object-oriented nature of the above mentioned
features the basic concept of this new approach is to build up
image segments (by means of DLM-Information), which are
candidates for semantic objects. Inside these segments the
features have to be determined ('learned’) as basic parameters
for the decision process (chapter 4) by means of integrated
knowledge processing. After this analysis a change detection
can be carried out, which leads to an update of the digital
database.
3 FEATURE EXTRACTION
3.1 Segmentation
The results of common segmentation methods such as region
growing or edge based methods are not satisfying for inho-
753
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
mogenous areas. Especially settlement areas show a great
variability in reflectance values because of numerous mixed
classes and mixed pixels, i.e. settlements usually contain
also other classes like meadow, forest, water or agriculture
((VOEGTLE, SCHILLING 1995)). Therefore, methods were
developed which concentrate only to the significant charac-
teristic of the object. With the assumption that this charac-
teristic appears more often inside the concerning object than
other ones, a robust segmentation can be achieved by mod-
elling of the interrelations in the image information. So the
segment boundaries can differ between image and database
information, if the actual status (image) has changed (e.g.
increasing of settlement areas, stubed areas in forest etc.).
Example 'Settlement': Settlements are characterized by
a clustering of artificial, man-made objects like buildings,
streets etc., which contain no vegetation. Because of a (nor-
mally) planned development, these objects have specific in-
terrelations (e.g. distances) and build up specific structures.
In the image information spatial clusters of pixels with a poor
density of vegetation can be found in settlement areas. Be-
cause of the great amount of other objects which can appear
inside these areas, an enormous inhomogenity in reflectance
values is usual. This phenomena (semantic class settlement
don't coincide with reflectance class) results in dissatisfying
segmentations.
As preprocessing step of a new segmentation approach the
pixels with a poor density of vegetation are extracted by the
method of Normalized Difference Vegetation Index NDVI:
(IR — R)
NDVI = ————-
(IR + R)
where IR — reflectance value in near infrared domain
R — reflectance value in visible red domain
With a threshold operation based on a statistical analysis
of the NDVI-histogramm the vegetationfree pixels can be
marked directly or these pixels are used to build up a spec-
tral signature for a consecutive classification. In both cases
we get a more or less spatial clustered class of pixels (fig. 1
and 2, Landsat TM, 26.04.1993). A first approach for mod-
elling neighbourhood characteristic was tried by iterative ap-
plication of mathematical morphology operators. The basic
idesa was to close the (smaller) gaps by dilation, afterwards
the original size should be achieved by the same amount of
erosion. Investigations have shown the principle applicability
of this method, but if the number of iterations is not high
enough, the contourlines and the shape resp. are disturbed
(fig. 3). If the number of iterations is too high, the shape
gets totally smoothed.
To overcome the problems of morphology operators another
approach which is used normally in determination of Dig-
ital Terrain Models (DTM), was investigated. Neighbour-
hood relations can also be modelled by triangulation net-
works, e.g. Delaunay triangulation, which connects adjacent
elements with shortest distances (fig. 4). A statistic of tri-
angle perimeters inside the training areas can be determined.
This will be used to select thoses triangles belonging to a
spatial cluster. After fusion of valid adjacent triangles we get
segments congruent to the spatial clusters (fig. 5).
This method has some important advantages compared with
morphology operators: with the distances between adjacent
