1etric
1g. A
>d on
obust
y the
abase
ed by
1 and
irgely
some
ation,
metric
uction
nages
/stem,
ase of
his is
is ofa
nages.
From a
; task.
'erator
orrect
ise.
: been
1g the
dare &
not be
;e not
which
y. The
of the
cation,
t been
atellite
rion ef
dopted
image-
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
o-image registration technique looking for corresponding
points through a gerarchical approach. In the published papers,
experiences of registering different kinds of satellite data have
been proposed (Landsat TM). Moreover, the procedure has
been applied also to data fusion of high resolution satellite
images (QuickBird) to a digital photo-plane. In all presented
tests, results based on transforming a set of independent check
points (IC Ps) have shown that the accuracy of image
registration is enough good for upgrading maps at scale
compatible with the resolution of the used imagery.
Recently the implementation of GEOREF algorithms in an
operational software has been completed and its operational
application to upgrade spatial databases of satellite images has
become possible. GEOREF runs under Microsoft Windows
environment and is composed of a main window (Figure 1)
divided in three different areas: a workspace on the left side,
showing the project structure, a viewing window on the right
side and a message window on the bottom.
The main tasks which are performed will be described in next
sub-paragraphs.
ze ELDER
aln ï
i Otia:
5 fo imagen tt
E gr deut
D tesagez nl
i Inegri raptis
RG Encrefersecna date
LL des | it
+
i
t]
jl
XC y ARE kho 210 Rate (VB, 1A ZA D) Es DAD Minti G00 = ETT meen (OD P
Figure 1. The GEOREF software’s main window.
2.1 Data input
Basically the input of GEOREF is made up of the pair of
images to be co-registered; the user is let to make the selection
about which of them plays as master and slave.
An external pre-processing stage is needed, consisting in the
extraction of both images from larger datasets. In case of multi-
spectral imagery, combinations of more than one channel could
be used, as done in experimental tests reported in the above-
mentioned papers by the authors. The goal of this task is to
render both images as similar to each other as possible
concerning radiometric aspects; moreover, application of image
enhancement techniques to improve contrast (e.g. a linear
stretching) is wellcome.
GEOREF accepts images at a radiometric resolution of 8-bit per
pixel, being this enough for matching algorithms. All common
image formats can be directly read by the software.
In case master is already geocoded to a cartographic reference
frame, this can be provided to GEOREF by means of an ESRI
“world” file, containing 6 coefficients of the affine
transformation from pixel-to-map.
LI
Us
2.2 Extraction of homologous points
The procedure to automatically extract CPs is based on a
standard image registration approach derived from digital
photogrammetry (see Heipke, 1997) and developing the method
proposed by Alparone et al. (1995).
Here we do not focus on the implemented algorithms, because
these have been already addressed in previous works, but we
would like to show some operational issues of GEOREF and the
way it can be applied to co-register images.
The registration process, once images have been imported into
the project, is based on the following items:
© setup of control parameters;
e extraction of interest points (IPs);
e image matching;
© robust estimation of geocoding transformation;
e visualization of resampled slave image to overlap the
master.
2.2.1 Setup of control parameters: Algorithm parameters
are available to the user and can be set from the “Project
configuration parameters” window (Figure 2). All parameters
are grouped into five groups:
e interest operator parameters;
e starting affine transform coefficients;
e matching parameters;
e outlier rejection parameters;
e georeferencing parameters.
On the left side of the configuration window are the interest
operator parameters and the start affine transform coefficients.
The well-known interest operator from Forstner (1986) is
applied to extract from both images a set of points which are
candidate to become CPs. To work out enough points, a set of
parameters must be setup, two of them being crucial:
e the interest window size; usually, the smaller it is, the
higher is the number of extracted IPs; on the other hand, if
the window is selected too small, poor contrasted features
could be found (default window size is 5x5 pixels);
e the minimun distance between two IPs (usually two times
the interest window size);
e the minimum point density; this parameter allows to check
the extraction of a sufficient number of IPs, depending on
the image kind, quality and content. By the way,
experience of the user is fundamental to properly select
minimum point density. If the fixed value would not be
reached, an adaptive procedure will restart the application
of Fórstner operator by reducing the interest window size.
In the *GEOREF start affine transform coefficients" frame, a
set of rough initial parameters can be entered (if known).
Otherwise, these can be computed by interactively
measurement of at least 3 CPs in both images.
On the right side of configuration window we find matching,
outlier rejection and georeferencing parameters.
The matching process is performed by Least Squares matching
technique (Griin, 1985); implementation details can be found in
Scaioni (1999). The success of this algorithm will depend on
the selection of following parameters:
e the size of searching window (default size is 9x9 pixels);
e the size of matching window (default size is 7x7 pixels);
