The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1298
are available (stereo imaging direction). A threshold of 0.5
pixel is used throughout this report for acceptance of forward
intersection results.
2.4 Blunder reduction
Blunder reduction is done during and after image matching and
during forward intersection. Most of the methods for image
matching are mentioned in section 3.1. Additionally, some
blunders can be detected via quasi-epipolar reprojection of the
stereo pair. For this best points from hierarchical matching are
used to estimate an affine transformation from fore to aft
CARTOSAT image. The standard deviation a for the column
residuals of this affine transformation is very small (a = 0.2-0.3
pixel). The fore coordinate pairs of the mass stereo tie points
are transformed with this affine transformation. All points are
rejected for which the absolute value of the difference of the
column coordinates of the aft tie points and the transformed
fore tie points is larger than the threshold 3a. This blunder
check is not independent of the blunder check via the residuals
in image space during forward intersection.
2.5 DSM interpolation and orthoimage generation
A regular DSM is generated from the point clouds produced by
forward intersection via triangulation and interpolation (Hoja et
al. 2005). DSM editing via cloud- and water-masks etc. is under
investigation, as well as DSM fusion for separately produced
CARTOSAT-1 DSM and existing DSM from other sources.
After derivation of the DSM orthoimages with user defined
datum and projection can be generated using the affine
corrected RPC and the DSM. Accuracy evaluations based on
orthoimage comparison of aft and fore sensor are given in
(Lehner et al. 2007) with more details.
3. RESULTS FOR THE DIFFERENT TEST SITES
3.1 Catalonia
A reference DTM with a GSD of 15 m (height accuracy 1.1m)
and 10 orthoimages with a resolution of 0.5 m are provided by
the Institut Cartografic de Catalunya (ICC). 70 GCP have been
measured in the orthoimages and the stereo partner Cat-A with
sub-pixel precision. These measurements have been
automatically transformed into Cat-A/F tie points via least
squares matching using mass points from hierarchical matching
for initial guesses of Cat-F coordinates and least squares
matching (LSM) with several window sizes. Thus, 68 GCP for
Cat-F could be derived - well fitting to the Cat-A GCP in terms
of stereo tie points. 6 window sizes from 17 to 27 have been
used in LSM in order to get statistical values for the accuracy.
The mean standard deviations in rows and columns for the 68
GCP and 6 window sizes were below 0.1 pixel. Consistent
stereo tie points are derived by this procedure.
The standard deviations of the residuals in RPC correction are
given in table 3 and a plot of the residual vectors for the fore
image is provided in figure 2. The shift parts of the affine
transformations for RPC correction are given in table 4 for all
stereo pairs, normally (besides M1A/F) corresponding well with
the expected orbit/attitude accuracy of a few hundred meters
(better than 160 m along-track and 100 m across-track).
After RPC correction the DSM is derived through image
matching and forward intersection. A few numbers for
illustrating the results are given in table 5. For each stereo pair
the number of matches passing through forward intersection
(using the threshold of 0.5 pixel for all 4 residuals in image
space) is given for the best tie points from hierarchical
matching (HM) and for all points after region growing (RG).
Additionally, the mean height difference and the standard
deviation of the height differences to the reference DSM/DTM
are provided. There is no distinction on land use. Thus, for the
small Taching area with a high percentage of forests with tall
trees this gives a wrong picture which will be put right to some
extent in the section 3.3 on Bavaria.
Image
Number
of
GCP
Shift part of affine
transformation (pixel)
row
column
Cat-A
70
-32.66
-16.95
Cat-F
68
-48.42
1.47
MIA
31
-2231.59
-685.14
M1F
30
-2335.86
-546.46
M2A
9
-40.08
-6.01
M2F
9
-50.83
-0.56
Bav-AT
14
52.38
-8.60
Bav-FT
14
62.07
38.18
Table 4: Shift parts of the affine transformations for RPC
correction for visualization of absolute positional accuracy of
original RPC
Stereo pair /
matching
step
Number of
accepted tie
points
(million)
Height difference:
reference DTM/DSM minus
Cartosat-DSM (m)
mean
a
Cat / HM
0.06
-0.6
1.83
Cat / RG
7.08
-1.0
3.05
Ml / HM
0.01
-1.9
2.24
Ml / RG
4.82
-1.4
3.80
M2 / HM
0.03
-1.5
2.21
M2 / RG
6.14
-1.1
3.53
Bav / HG-T
0.006
-3.6
4.18
Bav / RG-T
1.09
-3.6
7.29
Table 5: Number of tie points accepted during forward
intersection for 2 matching steps (HM: excellent points from
hierarchical matching / RG: all points from region growing) and
mean and standard deviation of height differences to the
reference DSM (Ml/2) or DTM (Cat and Bav, T: Taching area
only)
Orthoimages Cat-A-ortho and Cat-F-ortho are computed based
on corrected RPC and the DSM produced via triangulation and
interpolation from the accepted matches from region growing.
The matching between Cat-A/F-ortho gives shift vectors with a
mean of (0.10, 0.06) and a standard deviation of (0.16, 0.24) for
rows and columns, respectively (in pixel, about 45000 tie points
excluding the coast area are used). Thus, in regions of
acceptable density of matches, i.e. in areas where the DSM
interpolation has enough support, the fit between the
orthoimages is very satisfactory.