NJ
en into
; of the
s based
Points
ry. For
that the
| and in
optical
dinates
dinates
whose
zery. In
> use of
form of
nerated
"RPCS,
even if
rs) or a
0 quite
ntation,
ed on a
terrain
ntation
. and is
olution
> image
several
d The
srid are
already
ith an a
ed in a
| points
in our
nd QR
of the
RPC is
tion; in
red and
1 RPCs
respi et
ated by
sary in
e a re-
model,
ion.
several
ns: the
for all
igorous
or; the
1e ones
cimum,
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
quite few GCPs, if additional refinement transformations are
used.
For these reasons, the use of RPCs could be conveniently
extended also to high resolution SAR imagery, which presently
are not supplied with RPCs, with the exception of Radarsat-2.
Moreover, the RPFs model is an useful tool, in place of the
rigorous one, in the DSMs generation process, since it
establishes a straightforward correspondence between image
and ground coordinates, enabling a significant reduction of the
computational time.
3. IMAGE DENOISING
SAR imagery are affected by a high level of noise (speckle) due
to the inherent nature of radar backscatter. The source of this
noise is due to random interference between the coherent
returns issued from the numerous scatterers present on the
imaged surface, at the scale of the wavelength of the incident
radar signal.
Speckle noise gives the SAR image a grainy appearance and
prevents an efficient targets recognition and texture analysis,
crucial issue for the image matching .
Therefore, an image pre-processing step to reduce speckle noise
is required before starting the matching procedure.
In order to determine which adaptive image filters allow to get
the best results in terms of DSMs accuracy, a series of tests
were performed, varying filter type and window size.
In details, the applied filters were: Lee, Gamma Map, Frost,
Median (Shi and Fung, 1994); the correlation window size was
changed from a (5 x 5) to a (11 x 11) size, considering odd sizes
only. Also, the same filter has been passed several times (up to
three) over the images. Tests performed on several high
resolution SAR SpotLight imagery collected both by
TerraSAR-X and COSMO-SkyMed showed that Gamma Map
and Lee filters give approximately the same results in term of
final DSM accuracy (RMSE values), and the best ones were
obtained with a (7 x 7) correlation window.
4. IMAGE MATCHING
As already mentioned, the DSM accuracy is strictly related to
the matching process. In order to obtain good stereo geometry,
the optimum configuration for the radargrammetric application
is when the target is observed in opposite-side view; however
this causes large geometric and radiometric disparities,
hindering image matching. A good compromise is to use a
same-side configuration stereo pair with a base to height ratio
ranging from 0.25 to 2 (Meric et al. 2009) in order to increase
the efficiency in the matching procedure.
Many different approaches for image matching have been
developed in recent years. The main step of image matching
process is to define the matching entity, that is a primitive (in
the master image) to be compared with a portion of other
(slave) images, in order to identify correspondences among
them.
According to the kind of matching primitives, we can
distinguish two basic techniques, the already mentioned ABM
and FBM. In ABM method, a small image window composed
of grey values represents the matching primitive and the
principal methods to assess similarity are cross correlation and
Least Squares Matching (LSM); on the other hand, FBM uses,
as main class of matching, basic features that are typically the
easily distinguishable primitives in the input images, like
corners, edges, lines, etc.
37
These strategies, if separately applied, do not appear suited to
manage the strong geometrical deformation (like foreshortening
and layover) and the complex and noisy radiometry (speckle) of
SAR imagery. Therefore, an original matching procedure has
been developed.
Our matching method is based on a coarse-to-fine hierarchical
solution with an effective combination of geometrical
constrains and an ABM algorithm, following some ideas of
(Zhang and Gruen, 2006) but with a complete original
procedure and implementation.
After image pre-processing, the two images are resampled
reducing at each pyramid level the original resolution. The
correspondences between points in the two resampled images
are computed by correlation. In this way the surface model is
successively refined step by step, until the last step
(corresponding to the original image resolution), when the final
dense DSM is reconstructed. The advantage of this technique is
that at lower resolution it is possible to detect larger structures,
whereas at higher resolutions small details are progressively
added to the already obtained coarser surface. It has to be
underlined that, differently from other matching algorithm
recently developed (Perko et al., 2011), no a-priori known DSM
(for example the well known SRTM DEM) is necessary to
guide the DSM generation.
5. RESULTS
5.1 Data set
For the radargrammetric tests, imagery with a proper geometric
configuration, suitable for radargrammetric application, have
been acquired on Trento test field (Northern Italy). The
TerraSAR-X imagery have been provided by DLR in the
framework of the international project “Evaluation of DEM
derived from TerraSAR-X data” organized by the ISPRS
(International Society for Photogrammetry and Remote
Sensing) Working Group VII/2 “SAR Interferometry”, chaired
by Prof. Uwe Soergel — Leibniz University Hannover.
Acquisition |Coverage| Incidence ;
Area tte (Km?) Angie (deg) Orbit B/H
19/01/2011 | 5x 10 24.10 Desc
0.35
14/01/2011 | 5x 10 38.90 Desc
Trento
22/01/2011 | 5x10 31-15 Asc
0.35
16/01/2011 | 5x 10 44.19 Asc
Table 1. Features of TerraSAR-X SpotLight imagery
Two same-side stereo pairs have been considered for the DSMs
generation tests, acquired on ascending and descending orbits
respectively. The basic features of the images are listed in Table
1. A LiDAR airborne DSM supplied for free by the Provincia
Autonoma di Trento was available as reference for the accuracy
assessment.
5.0 Radargrammetric DSM analysis
A descending and an ascending stereo pairs were available, so
that the DSM could be reconstructed using two different views.
The two stereo pairs have been processed separately and the
corresponding point clouds have been assessed.