The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
146
Figure 1. A classical sequence for registration algorithms
Registration algorithms rely on corresponding “control
elements” to calculate the parameters of the geometrical
transformation that sets the correspondence between all the
points of both images. These “control elements” are either
landmarks or locations characterized by their neighborhood.
Landmarks can be points, contours, line intersections, or regions
that have to be extracted from images before: approaches using
such landmarks are called “feature-based” ones. Otherwise,
when using pixel neighborhood to perform such a matching
process, we talk about “area-based” approaches.
“Feature-based” methods try to identify corresponding
landmarks in both images. In fact, this identification is not
performed directly on landmarks but on shape descriptor values
(descriptions) that represent them. We only consider those
shape descriptors that remain invariant through any possible
transformation from an image to another. They must also have
two important properties that are uniqueness and stability: two
different landmarks must be represented by two different
descriptions (uniqueness property) but slight changes on a
landmark (because of noise, for example) must not change its
description (stability property). Finally, landmark descriptions
of both images are compared through a similarity measure that
helps in finding how to pair landmarks.
“Area-based” methods use a similarity measure to identify two
areas that are considered as neighborhoods of corresponding
pixels. Then, there is to define a criterion that depends on this
similarity measure and whose optimization provides the
registration transformation.
Whatever approach we decide to use, we need to make choices
(about landmarks, similarity measure,) that take into account the
way SAR images are built but also the noise that is a relevant
feature of these images. Anyway, the transformation model has
to be guided by three considerations that are the geometrical
deformation during the acquisition process, the required
precision of the registration, and the use of the expected result.
Global deformations, local deformations, or both can
characterize such transformations.
b) Specificity of Spaceborne platform Radar systems
2.2.1 Finding the position when acquisition is performed
from a spatial platform: We need to have a rough
approximation of the two images geographic parameters (i.e.
location and orientation of the corresponding areas) in order to
efficiently initialize the registration process. This information
could be derived from sensor position parameters but the orbit
and orientation of the platforms that convey these sensors can
be modified by several external actions (Arbinger and D’Amico,
2004), as, for example:
Earth gravity field irregularities, and sun/moon
interactions
Atmospheric friction in the case of satellites whose
height is between 300 and 800km
Photonic pressure
Orbit and orientation parameters are not captured in a
continuous way but obtained through estimation from key
positions, and this estimation requires quite a long time to be
processed: it can take a day to several weeks depending on the
precision (Wessel et al., 2007). In addition, even if we had all
the metadata for computing such parameters, we would not
have enough information on the sensor functioning to exploit
efficiently these metadata (Eastman et al., 2007). Finally, we
can say that only knowing these external parameters is not
sufficient for solving the registration problem between two
Radar images. From now, we will suppose that the two images
globally represent the same area but are not precisely registered.
2.2.2 Geometrical deformations: A main drawback of SAR
systems is to provide image geometrical deformations
(Lillesand et at, 2004) that result from the way of sorting points
by using their distance to the antenna. The ground position of a
point can be slightly (or more) wrong because it has been
“seen” as shifted toward the antenna. This kind of error is more
significant when the ground height is not uniform: depending on
height variations, the target density (a point on the ground is a
target for the Radar) can increase and generate artificially clear
areas in the image (“highlighting” effect), or it can decrease and
create abnormally dark areas in the image (“lowlighting” effect).
When height variations are very large, even the target sorting
may be wrong (“layover” effect) and some of the targets can
disappear because they are masked (“shadowing” effect).
Other deformations have to be taken into account (Richards,
2006): they are related to the earth curvature and to the width of
the viewing angle. In such cases, the scanned areas that are far
in the nadir direction are widely opened: this situation results in
an important non-homogeneity in the pixel distribution, in
particular along image boundaries.
2.2.3 Radar image radiometry: Interpreting images from
Radar systems is very difficult because of the complexity of the
processes involved in their generation (Rees, 2001). Signal
intensity for each point - or target - is encoded as a grey level
in the resulting image and depends on the way the Radar wave
interact with the target. This interaction relies on both sensor
features and target features. Sensor features are its wavelength,
its polarization and its viewing angle. Target features are its
roughness, its dielectric characteristic (it has a high value for
metallic elements, and it is correlated to its moisture content),
its shape and orientation.
All these parameters mutually interact and thus, it is very
difficult to exactly know which are their individual
contributions to the returning signal. For example, the target
roughness parameter depends on the target itself but also on the
incident angle and on the wavelength that has been used for
illuminating the target. Several mechanisms related to the
reflection and backscattering processes are involved in the
image generation: specular reflection, diffusion, comer
reflection, and volume diffusion.