ınbul 2004
t-up of an
d satellite
ping aerial
amera and
ing project
X2 des.
orted. As
D data are
VPF to the
ction 4.1).
s complex
inJECT as
erface
p1 (VPF) |
D (VPF) |
)
PF)
date
scribed in
or modified
data can be
iis standard
ed to DXF
) extraction
owser (e.g.
ce between
| GeoMedia
1L2 format
With this
chemes can
eatures can
'ometry and
bject.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
For the import of GML data into GeoMedia, the existing GML
data server can be used, whereas the export of GML data from
GeoMedia can be carried out with Intergraph's GML export
module.
For the data exchange between inJECT and Dynamo two new
conversion tools were developed by the Universität der
Bundeswehr, Munich (Shi et al. 2001). The tool for the import
of GML into Dynamo generates a Dynamo data dictionary from
the XML schema and a Dynamo object space from the GML
data. The tool for the export of GML data from Dynamo creates
a XML schema and a GML file derived from the Dynamo data
dictionary and object space.
We have implemented this GIS interface for the following vector
data types:
* VMapl (Vector smart map Level 1, corresponding to
1:250.000 maps, 2D features)
e FFD (Foundation feature data, corresponding to 1:50.000 —
1:250.000 maps, 3D features)
* German military basic vector database, corresponding to
1:10.000 — 1:25.000 maps, 2D features)
4.2 Parcel Extraction
We have implemented an algorithm to extract in a semi-
automatic way regions from aerial and high-resolution satellite
images. The algorithm is able to extract:
* near-homogeneous parcels, whose statistical parameters
are different from the ones corresponding to their
neighbours.
* parcels with quite well defined boundaries i.e. other
near-homogeneous parcels, or linear elements.
There are some other kinds of parcels that can be only identified
by contextual reasoning, where the knowledge interpretation
drives to identify them. Those parcels, however, can not be
extracted by our approach.
Our approach is a combination of deformable models and region
growing techniques in a statistical framework under the
Minimum Description Cost environment. This approach is
called region competition.
In general the automatic feature extraction from images estab-
lishes a correspondence between some image characteristics and
models. In our particular case, deformable models are used,
which describe general geometric and radiometric characteristics:
degree of continuity and kind of radiometry to be recovered.
Snakes is a special case of a general technique that matches a set
of deformable models with an image by minimizing a cost
function, that represent in our case an energy function,
composed by a weighted combination of internal (model) and
external energies (image).
Lage = Wyegion EB coton + Dodge Beige ( 1)
E,4.. 1s based on the image contrast and will attract boundaries
to contours of high image gradients, whereas £,,,;,, pushes the
snake to enclose quite homogeneous areas.
473
With / being the image, we define
E dy zm TIN Fr. y? (2)
For the first term of (1) we define the region homogeneity as
follows:
A region R is considered homogeneous when their intensity
values are consistent with the ones that will be obtained in case
of being generated by a family of pre-established probability
distributions P(I:o), where & are the distribution parameters.
Our objective is to have a representation of the image composed
by quite homogeneous areas, so we represent the image
decomposed into region entities:
M
E, oni, fo, M s fs | ds — il log P(It y) : o;)dxdy 4 »)
is 203 Jo WR enm (3)
The first term in (3) describes the curve length of 8R; that is the
boundary of Rand ju is the code length codification. The
second term is the addition of the cost of codifying each pixel
intensity inside the region R; with probability P(1:01) . Y stands
for the codification length to describe R,.
Since we solve the total snakes energy equation by the steepest
descendent method, we end up with formula (4) that describes
the movement of the curve v which represents the boundary
between two regions that compete for having each of the
boundary pixels.
v ; 5
my 1- roii + log Polos) ) E AVIVIVI|(4)
kcQ()
Q(v) 7 fk | v belongs to D) and K;(v) are the curvature at I.
The threshold to decide where to stop the movement is defined
by the last term of (4). The movement is completed in the case
of reaching an edge, or when the region growing process stops.
So equation (4) weights two possible non-excluding situations:
the contours move to high gradients or enclose pixels with
similar statistical parameters. In the case of near homogeneous
regions the statistical model chosen is the Gaussian distribution,
so the statistical parameters for each parcel are the mean and the
deviation. In our case these parameters are user-defined because
the operator delivers a starting polygon composed of at least 3
vertexes to the system i.e. a rough approximation of the parcel
to be extracted. From this darting polygon the system will
initialise the statistical parameters and the initial model, that will
be deformed until the result is reached. Figure 3 shows an
example of parcel extraction.
In a final processing step the parcel contours are smoothed using
the smoothing tool as described in subsection 4.3. The
smoothing tool simplifies the contours and eliminates their
oscillations (Figure 4).
