The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
Status 
Before BA 
After BA 
Mean (pixel) 
1.9 
0.0 
Maximum (pixel) 
9.2 
2.6 
Standard deviation (pixel) 
4.4 
0.85 
Number of points 
500 check points 
Table 3. Statistics on back-projection residuals at Columbia 
Hills 
4. TOPOGRAPHIC MAPPING AT MER LANDING 
SITES 
4.1 DEM Generation 
'•Ì 
♦ 
# 
*'■- -i 
Legend 
Vertical Difference 
MB 
I-*--» 
Figure 6. 3-D surface map of the area of Spirit traverse 
We used 644,609 interest points and 1,556,677 grid points to 
perform 3-D stereo intersection of ground points based on the 
bundle-adjusted EO parameters. Ordinary Kriging with a 
spherical semi-variogram model was used to generate a 1-m- 
resolution DEM of the area covering the entire traverse of Spirit. 
Figure 6 shows the resulting 3-D surface. 
4.2 DEM Comparison 
To validate the quality of our DEM, we performed a 
comparison with a DEM generated by USGS using the same 
data. The comparison was done both in horizontal and in 
vertical directions. The two DEMs were registered through a 2- 
D similarity transformation based on six manually identified 
corresponding points. The root mean square error of these 
points after the transformation is 0.9 m. After horizontal 
registration, a vertical registration was done by a shift and a 
rotation in such a way that the average of the vertical 
differences was zero. Grid-to-grid vertical differences were then 
calculated. The result indicates that most areas have differences 
less than 1 meter (Figure 7). The standard deviation of all grid 
differences between those DEMs is 0.63 meter which 
corresponds to 2 pixels on the image. 
Figure 7. Vertical difference between OSU DEM and USGS 
DEM 
There are a few features with large differences of a few meters. 
The most significant one is a small crater in the south west 
comer of the DEM. Further investigations will be performed to 
figure out the causes of the large differences. The strip artifact, 
though smaller than 1 meter in elevation, will also be 
investigated. Orthophoto will be generated using the validated 
DEM, the original HiRISE imagery and the bundle-adjusted EO 
parameters. 
5. CONCLUSIONS 
In this paper, we presented a rigorous photogrammetric 
processing approach for automatic DEM generation from 
HiRISE images. Our approach employs a coarse-to-fine 
hierarchical matching method that can provide dense and 
reliable tie points for both DEM generation and bundle 
adjustment. First, interest points were matched up to the level of 
original image scale. Then grid points were defined and 
matched, also moving from coarse to fine grid scale. For quality 
control, an automatic error detection algorithm was incorporated 
at each hierarchical level. We evaluated the performance of our 
matching results for a test area covering the entire Spirit 
traverse. At intermediate levels, the mean residual remained 
lower than 0.3 pixel, providing a TIN surface of parallax 
difference to provide estimation for dense grid points matching. 
At the final level, with 3-pixel grid spacing, the mean residual 
was less than 0.11 pixel. We performed a bundle adjustment to 
reduce the inconsistencies between HiRISE stereo images. In 
the bundle adjustment, second-order polynomials were used to 
model the change of EO parameters over time. We chose 500 tie 
points from the matched interest points for bundle adjustment. 
The performance was evaluated based on back-projection 
residuals of the independent check points. The mean residual 
was reduced from 1.9 pixels to zero. Also, the standard 
deviation of the residuals decreased from 4.4 pixels to 0.85 
pixel. We created a DEM of the area covering the Spirit rover 
traverse, which was compared with a DEM generated by the 
USGS. 
ACKNOWLEDGEMENTS 
Funding for this research by the NASA Mars Applied 
Information Systems Research Program and the NASA Mars 
Exploration Program and is acknowledged. 
991