The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
1058 
scanned across each pixel regarded as noise and the standard 
deviation of the DN values covered in the window was 
computed. If a low standard deviation was observed, the pixel 
was deemed as being located in a flat area, and furthermore, 
this point was confirmed as noise. As a result 2,418 out of 7,452 
points remained as noise while the others were re-fed into the 
non-noise 3D point cloud. The overall distribution of the re-fed 
points is shown in Figure 2(d). It is noted most of the re-fed 
points occurred at the area of large height variation while noise 
points located at the image edges remained as noise. 
3.3 Iterative Noise Reduction 
Since a local shortage of MOLA points would impact the result 
of surface matching, it was proposed to replace MOLA points 
with non-noise HRSC intersection points produced in 
Section 3.2 (Figure 4 (a)) as reference surface. Surface 
matching was then performed again with the original 
intersection point cloud (Figure 2 (b)). The matching was 
finished after 5 iterative computations and in total 1,906 points 
was treated as errors. The distribution of the removed errors is 
shown in Figure 4 (b). Compared with the result from the first 
matching, the number of detected noisy points was significantly 
reduced. Moreover, compared with Figure 2 (c), it is observed 
that the updated reference surface effectively improved the 
determination of possible errors. 
4. QUALITY ASSESSMENT 
In order to assess whether there was any improvement after 
noise reduction, a DTM quality assessment was carried out. To 
accomplish this, height differences between the MOLA points 
and the HRSC 3D intersection points were computed. The 
derived height difference was considered as an indicator of 
DTM quality. As the spot size of MOLA footprint on Martian 
surface is about 160 m (Smith et al., 2001), a buffer distance of 
160 m was applied to each MOLA point. The HRSC 
intersection points located in the buffer zone of MOLA points 
were selected and their height differences were computed. 
Four sets of point clouds were compared against MOLA points, 
including the original HRSC intersection points (Set 1), the 
point cloud after the first noise reduction (Set 2), the point 
cloud after noise inspection (Set 3) and the point cloud after the 
second noise reduction (Set 4). The statistics of the height 
difference are listed in Table 1. 
Set 1 
Set 2 
Set 3 
Set 4 
Total points 
28,098 
20,646 
25,680 
26,138 
Points 
examined 
848 
737 
812 
815 
Max 
6223.1 
140.5 
228.2 
140.4 
Min 
-6413.3 
-176.4 
-211.5 
-176.6 
Mean 
-16.835 
-3.178 
-3.079 
-3.317 
Stdev 
606.870 
23.463 
34.731 
33.365 
Skewness 
-2.992 
-0.221 
0.094 
-0.112 
Kurtosis 
71.710 
7.326 
7.082 
4.648 
Table 1. Statistics of height differences for four datasets. 
a slightly larger maximum, minimum and standard deviation 
values were derived. Finally, when the point cloud in Set 3 was 
used as a reference surface, the quality of the DTM after noise 
reduction (Set 4) achieved the same level of Sets 2 and 3, in 
which an even better kurtosis value was derived. 
(a) (b) 
Figure 4. Points for reference surface (a), and detected noise 
(green) over HRSC intersection points (red) (b). 
5. EXAMINATION OF DTM MOSAICS 
It was expected that a large difference statistics would be 
obtained for Set 1 as many errors can be observed in the 
original HRSC point cloud in Figure 3 (top). While positive 
results were derived after the first stage of noise reduction 
(Set 2), in which the height differences were close to a normal 
distribution. Furthermore, it is also revealed that the quality of 
DTM after noise inspection (Set 3) was close to Set 2, although 
From the results derived in Section 3, it was revealed that 
numerous errors were distributed at the edge of the DTM. This 
issue might affect the performance of DTM mosaicking. This 
section therefore aims to examine the impact of noise reduction 
on adjacent DTMs. To achieve this, four HRSC orbital images 
(hi000, hi011, hi022 and h0103), which cover a section of the 
large outflow channel in Ares Vallis (Figure 5) were employed.