The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
1057 
3. IMPLEMENTATION 
The proposed method was implemented on the intersection 
point cloud produced from HRSC hlOll orbital imagery 
(Figure 2). The images cover part of eastern Ares Vallis at 
approximately 1.12° to 11.74°N and 335.32° to 336.42°E. 
(a) (b) (c) (d) 
Figure 2. Left to right: (a) MOLA points, (b) HRSC 
intersection points, (c) detected noises (green) over HRSC 
intersection points, and (d) re-fed non-noise points (magenta) 
over noises (using hlOl 1 ortho-image as background). 
3.1 Noise Reduction 
First of all, the MOLA points over the same extent were 
extracted. As introduced in Section 2.3 they were then 
compared with a gridded MOLA DTM and a number of 
obvious errors were removed. Meanwhile the hi Oil orbital 
imagery was processed through the VICAR software and a 3D 
intersection point cloud was derived. A median filter was then 
applied to both point clouds and in total 11,884 MOLA points 
(Figure 2 (a)) and 28,098 HRSC 3D intersection points (Figure 
2 (b)) were available for surface matching. 
Prior to commencing the matching, it was critical to decide the 
threshold value for determining noise in the HRSC point cloud. 
Ebner et al. (2004) studied the height difference between HRSC 
intersection points and the MOLA DTM, in which three HRSC 
images (h0018, h0022 and h0068) were investigated. Their 
results demonstrated the average difference in height between 
MOLA DTM and HRSC object points was on the order of 70 m. 
Hence an empirical value of 70 metres was adopted as the 
threshold value in this paper so these points were flagged as 
noise during surface matching and removed from the HRSC 
point cloud afterwards. 
Surface matching was finished after 11 iterative computations. 
As a result, a total of 7,452 points (26.5%) were treated as noise 
and removed. In the top diagram of Figure 3, large errors 
occurring in the original 3D intersection point cloud are visible 
from the side view. However, after surface matching, most of 
these errors are successfully removed (Figure 3, bottom). 
Moreover, the spatial distribution of the removed noises is 
shown in Figure 2 (c) and it can be observed that most of the 
errors mainly occurred in two areas. First of all, many errors 
appear to exist at the edge of the orbital strip. In addition, when 
compared with MOLA points (Figure 2 (a)), it was observed 
that the other set of errors were located in the area where there 
was a shortage of MOLA points occurring. As noise reduction 
was implemented based on the MOLA surface, the HRSC 
intersection points might be removed during matching if not 
enough information was supplied by the MOLA TIN surface. 
This issue was especially obvious in the area of large height 
variation, such as the edges of craters and channels. Therefore it 
appears that some points being treated as errors might not be 
true errors. Further noise inspection was required. 
Figure 3. Original 3D point cloud (top) and point cloud after 
noise reduction using surface matching (bottom). 
3.2 Inspection of Removed Noise 
As introduced in Section 2.2, further inspection of removed 
noises was carried out in the hi011 ortho-image. The 
corresponding pixels of the removed noise point were marked 
in the ortho-image and the general topography (e.g. rough or 
flat area) at the located pixel was determined. To achieve this, 
U. S. Geological Survey (2007) suggested that at least three 
pixels were needed to represent the shape of a feature. However, 
this paper has adopted a 5 by 5 pixel window due to the 
relatively lower image resolution. The window was then