bul 2004
ng roads
tion from
ing in the
ed in the
order to
area is a
jn with a
) value in
/egetation
ge in the
sult from
bstacle, is
available
bjects are
| identical.
described
, but also
re is a 2D
ction (e.g.
icrete of”
egion with
ark region
1terest in a
lationships
icles, thus,
re 1): One
earby, and
r example,
of a field,
it solely on
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
the field boundaries. Accordingly, the geometries of these
different objects are identical or at least parallel with a short
distance in between in the case of location alongside roads,
rivers or railways.
3. STRATEGY TO EXTRACT FIELD BOUNDARIES
AND WIND EROSION OBSTACLES
3.1 General Strategy
The general strategy for the extraction of field boundaries and
wind erosion obstacles is derived from the modelled
characteristics of the objects taking into account the realization
of an automatic process flow (cf. Figure 2). CIR-images, GIS-
data and a Digital Surface Model (DSM) are the input data to
initialize the flowchart: Firstly, field boundaries and wind
erosion obstacles are extracted separately with two different
algorithms. At a later date, a combined evaluation of the
preliminary results due to the modelled geometrical and
DSM
, reduction
c of search domain 9
2 definition of 2
= ® D
= a 9
o C =
c o
c 3
= D
m s
regions of
interest
c
regions of
interest
introduction of further knowledge
search areas
field areas
introduction of boundaries
line extraction
introduction of hedges and tree rows
extraction using image and DSM
preliminary
field boundaries
verification using image and DSM
field boundaries wind erosion obstacles
combined evaluation
Jt.
refined extraction process
refined field boundaries
refined wind erosion obstacles
Figure 2. Flowchart of the strategy
thematic similarities of the objects is essential getting a refined
and integrated solution.
The strategy extracting the field boundaries separately (cf.
Figure 2 on the left side) starts with the derivation of the open
landscape from the GIS-data. In addition, within the open
landscape, regions of interest are selected using the road
network, rivers, railways, tree rows and hedges as borderlines in
order to handle the large datasets in an appropriate manner (cf.
Figure 4 for an example of a selected region of interest).
Consequently, the borderlines of the regions of interest are field
boundaries, which are already fixed, and the image analysis
methods are focused to the field boundaries within the regions
of interest. The homogeneity of the vegetation of each field
enables a segmentation of field arcas, processed in a coarse
scale to ignore small disturbing structures. Identical vegetation
of neighbouring fields leads to missing field boundaries, which
can be derived by a line extraction in a finer scale. Further
knowledge will be introduced at this time to exploit GIS-data,
which gives evidence of field boundaries within the selected
regions of interest. The derived field boundariés are in some
parts inaccurate and a snake algorithm is initialized to refine
their geometric accuracy.
The strategy extracting the wird erosion obstacles separately
starts with a differentiated exploitation of the GIS-data (cf.
Figure 2 on the right side). Focussing on the open landscape,
search buffers are defined using the prior knowledge from the
GIS-data: Search areas for potential wind erosion obstacles are
located alongside roads, rivers or railways as described in the
semantic model (cf. section 2) and have to be verified using
imagery and DSM-data. In contrary, there is no prior
information about the location of all other wind erosion
obstacles available. In addition to the extraction of high NDVI-
values and higher DSM-values than the surrounding area,
characteristics of the wind erosion obstacles as straightness, a
minimum length, width and height have to be considered.
The aim of the combined evaluation of the preliminary results is
to detect discrepancies between field boundaries and wind
erosion obstacles due to their modelled geometrical and
thematic similarities (cf. section 2). For example, it has to be
solved, how far the extracted geometric position of two
different objects has to be matched, if they are located nearby
each other. Similarly, extracted wind erosion obstacles without
a corresponding extracted field boundary have to be checked,
whether there is a field boundary, too, or whether the extraction
of the wind erosion obstacle was wrong. Consequently, the
combined evaluation and refined extraction process leads to a
consistent and integrative final result.
3.2 Extraction of Field Boundaries
3.2.1 Segmentation of Potential Field Areas: The extraction
of field boundaries begins with a segmentation in each region of
interest to exploit the modelled similar characteristics and
homogeneity of each field. As data source the red channel of
the CIR-images is used, which according to our experience
fulfils the homogeneity criterion best. To utilize the changing
vegetation from one field to the next, the absolute values of the
gradient are computed. The topography of the grey values is
used to accomplish a watershed segmentation (Soille 1999).
The resulting basins are marked with their corresponding mean
grey value in the red channel. Potential field areas are derived
grouping basins, if they lie next to each other and have a low
1067
