The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
1052 
Figure 4 shows an evaluation of the crater DTMs at different 
iteration stages. 
: i : 
3*00 F Jf 
V 
(a) First iteration (matching 
window 10) 
•**’¿id« ■ 
, **°i 
r, v-*" ^ 
•056 ^ 
(b) Second iteration (matching 
window 8) 
j 
r. 
(c) Third iteration (matching 
window size 6) 
|fe , 
a**; 
d) Final iteration (matching 
window size 4) 
Figure 4. Crater DTM model fitting evolution at each iteration 
stage (1 pixel =20m) 
3. RESULT AND ASSESSMENT 
Two regions were selected for detailed study, Elysium Planitia 
(9-11°N,148-158°E) and Iani Vallis (0-7°N, -14-19°E). They 
were chosen for their geological interest and different 
topographies, Elysium being a flat frozen sea area (Murray et 
al., 2005) and lani being a chaotic region with deep valleys and 
massive flood drainage (Gupta et al., 2008). 
Figure 5. Overview of Crater processing system for 3D and 3D 
crater database generation 
To process such large area data sets, the processing chain shown 
in Figure 5 was constructed. The photogrammetric processing 
such as 3D intersection and ortho image generation uses the 
DLR VICAR software which is available to HRSC team 
members and their associates (Scholten et al., 2005). At first, a 
wide area HRSC DTM up to 50m resolution and 40m ortho 
image sets was created and then a 2D crater GIS was processed 
by the processing chain, described in section 2.1. Then the 2D 
detection result and global DTM are fed forward into the 3D 
processing line which can automatically generate DTMs and 
ortho image “chips” of individual craters. The target range of 
crater processing is R> 400m for optical images. 3D processing 
was implemented in small areas such as HRSC image h2099 
over Elysium and h9023 in Iani. However, the comprehensive 
3D processing for both entire areas is still ongoing and will be 
shown in the final presentation. 
3.1 Generation and assessment of 2D crater Database 
A performance evaluation of the 2D craters was made using 
Building Detection Metrics (Shufelt, 1999). These were devised 
in order to provide an objective measure of performance for 
automated building detection algorithms, a situation analogous 
to automated crater detection: 
Detection Percent = (100 * TP) / (TP + FN) 
Branching Factor = FP / TP 
Quality Percent = (100 * TP) / (TP + FN + FP) 
where True Positive (TP) means the pixel is correctly identified 
as part of a building/crater object; False Positive (FP) means 
that a pixel is incorrectly identified as part of a building/crater 
object but is actually background; False Negative (FN) means 
that a pixel is incorrectly identified as part of the background 
when it is actually part of a building/crater object. 
However, there are problems with using these metrics for 
automated crater detection because of the lack of reliable 
ground truth at the relevant resolution. 
Therefore two processing stages are undertaken: 
Stage 1. Edit and Assessment 
• read each detection result from a text file & create a 
polygon shapefile containing each crater result 
• overlay each shapefile on its base image for manual 
verification & digitisation 
• tag TP, FP and delete FNs 
• establish the crater diameter lower limit cut-off in 
pixels for assessing the algorithm performance 
• compile the data and calculate statistics quantifying 
the performance for each image and detection result 
Stage 2. Merge and create 2D GIS file 
• merge the crater data sets from Stage 1. 
• resolve duplicates where orbital images overlap 
• produce a merged, georeferenced 2D crater shapefile 
In order to establish a detection cut-off minimum diameter in 
pixels, the detection metrics are calculated as a function of 
diameter. Plots of detection percentage against diameter in 
pixels are constructed and the cut-off minimum crater diameter 
in pixels for the chosen level of detection is determined. The 
assessment metrics are then calculated by only counting craters 
with diameters greater than this cut-off value. A cut-off limit of 
8 pixels was selected as representing the algorithm’s current 
limits for acceptable detection and quality rates, representing 
diameters of 320 metres and 200 metres for Elysium and Iani 
respectively as shown in Figure 6. Assessment results obtained 
for the Elysium and Iani images are shown in Table 2. 
(a) 2,543 Craters for the image strips in Elysium Planitia