3. SHACKLETON RIM DTM
This section presents the process of generating a high resolution
DTM for the candidate landing site Shackleton Rim 1 (SR1) by
means of photogrammetric stereo processing of line scanner
imagery of LROC NAC. The process is divided into the three
steps image ingestion, bundle adjustment and DTM matching.
3.1 Image Ingestion
The first step is searching for stereo pairs overlapping the target
region. In support of this task a GIS database of LROC NAC
imagery was set up by using index information available from
PDS. A spatial query for the SRI site at -89.781°, -155.848°
yielded about 1.300 hits at the time of writing (April 2012). For
stereo coverage the LRO spacecraft has to carry out off-nadir
slews of about 20° which is possible up to three times per day
(Robinson et al. 2010). The incidence and azimuth angles
should be as similar as possible in order to support the DTM
matching. These constraints reduce the number of stereo pairs
to four which come into consideration for the processing of SR1
site. Fig. 2 displays the selected stereo pairs on top of a
colorized LOLA shaded relief dataset.
Figure 2. LROC NAC stereo pairs at SR1 site
As can be seen the images cover the SRI site on the rim of
Shackleton crater near the south pole of the Moon. It is striking
that large regions within the LROC NAC strips are very dark.
As stated in the introduction the Moon's low inclination results
in very high incidence angles close to 90? at the poles so that
dependent on topography and azimuth of the sun large areas are
temporarily or permanently in shadow. Thus, generating DTMs
at the poles of the Moon is very cumbersome. Small areas of
LROC NAC strips have to be processed separately and
mosaicked in order to yield a continuous DTM. After the
download of the images from PDS ISIS routines are used for
import, radiometric calibration and SPICE kernel initialization.
3.2 Bundle Adjustment
In order to improve camera pointing a bundle adjustment is
carried out. Using ISIS tools tie points are automatically
determined by image matching (cross correlation and least
squares matching). Because of the large shadowed areas an
equal distribution of tie points over the entire block was not
obtained. However, this is desirable for line scanner sensors
(Schmidt et al. 2008) in order to achieve fine scale
improvements of the camera pointing. Unfortunately, only 337
tie points have been determined so that we could not include
spacecraft velocity, spacecraft acceleration, camera angular
velocity and camera angular acceleration (i.e. offset and drift)
as unknowns into the bundle adjustment. Spacecraft position
was not included because no GCPs were measured. Only the
camera pointing was improved (cf. Lee et al. 2012) by fitting a
polynomial of degree 2 to the camera angles.
The solution converged after seven iterations resulting in a
standard deviation for the image coordinates of about 1.5 pixels.
As the images have an average ground sampling distance
(GSD) of about 1 m this value corresponds to an accuracy of
about 1.5 m. Sub-pixel accuracy is desirable for an accurate co-
registration of all input images but due to the unfavorable
illumination conditions this aim was not feasible. Nevertheless,
a significant improvement over the a priori accuracy of the
LRO orbits (Mazarico et al. 2012a) could be achieved.
3.3 DTM Matching
The derivation of the high resolution DTM is carried out by
digital image matching using the Ames Stereo Pipeline (ASP).
As pre-processing steps the input images are low-pass filtered
in order to reduce image noise. Using the LOLA dataset and the
improved camera pointing from bundle adjustment a pre-
rectification is carried out which crops the area of interest and
generates quasi-epipolar images of the input pair. This step
eliminates geometric differences between the two images with
respect to scale, rotation and line exposure times. Additionally,
this step also reduces perspective differences by taking into
account the LOLA DTM which results in a considerable
reduction of the search space for the correlation algorithm. A
LOLA dataset with a resolution of 120m was used, dated
March 2011; the available higher resolution versions were
found to contain gross errors which result in severe artifacts in
the rectified images. However, it is expected that soon a dataset
with higher quality will be available (Mazarico et al. 2012b).
In Fig. 3 the pre-rectified left input image cropped to the area of
interest (SRI site) is displayed. As can be seen the area is only
partially illuminated. To the right a 5 km long part of the crater
rim and its permanently shadowed interior can be seen. The
illuminated part has a width of about 1.5 km and to the left a
temporarily shadowed area is depicted. Furthermore, it can be
seen that due to the high incidence angle of 88.56? small
topography features cast long shadows which disturb the
correlation process.
The DTM matching procedure is divided into four steps. At first
integer disparity estimates are computed for each pixel using
the normalized cross correlation as similarity measure. In order
476
