ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision‘, Graz, 2002 
  
  
" N : n de ë mw. qe - 
Figure 10: Shading of the DTM with gross errors in the MOLA 
dataset 
these gross errors can be recognized. Of course, the generation 
of a DTM from the whole Mars surface by manual correction is 
not practicable and therefore an automatic procedure for error 
elimination is required. 
In the first test we tried to use hierarchical (due to the very 
inhomogeneous point density) robust interpolation to eliminate 
these errors. The elimination of the scan errors was possible 
with this technique, but due to the roughness of the Mars 
surface we also eliminated points in regions without scan errors. 
The rough surface did not fit to our functional model of linear 
prediction, which is able to generate very smooth surfaces. 
Better results were obtained by analysing scan line segments 
instead of the residuals of each individual point. It showed up 
that the average filter value of a segment (i.e. RMS of the 
residuals of the points belonging to this segment) could be used 
to eliminate those segments with gross errors. Due to the fact 
that correct scan line segments next to a segment with gross 
errors also get a higher RMS it was necessary to apply this gross 
error elimination in an iterative manner. Because not all points 
along a scan line segment are affected by gross errors we 
analysed the discrepancies of all eliminated points in respect to 
a DTM computed without these gross error segments and 
accepted all previous eliminated points within a certain user 
defined threshold value. 
This iterative method proceeds like the following: 
1. Compute a DTM with all points. 
2. Compute the RMS per scan line segment and eliminate lines 
with a high RMS. 
3. Compute a DTM with the accepted scan line segments. 
4. Former eliminated points are accepted if they are within a 
certain tolerance to the DTM . 
5. Compute a new DTM. 
A p=p(f) 
  
20m | 75m 200m 
  
  
  
  
  
  
— 
f 
Figure 12: Symmetric box weight function with decreasing 
extend in each iteration step for the elimination of scan line 
segments 
0.0m 
    
  
Wh 1 4 4 E * 
Figure 11: Shading of the DTM after automatic gross error 
elimination 
The steps 2 to 5 are repeated with iteratively decreasing 
tolerance values until the tolerance of the RMS per line reaches 
a user defined threshold value. The results of this process can be 
seen in fig. 11. 
This method corresponds to the robust interpolation with a box 
weight function with decreasing extend in each iteration step 
(fig. 12). In contrast to the previous examples, the weight 
function is not applied to the filter values of single points but 
for complete scan line segments. 
5. CONCLUSIONS 
We have presented a very general concept of gross error 
elimination for DTM generation (surface computation) and 
achieved good results for the presented datasets. What can be 
seen in the example section is that the general concept for gross 
error elimination is the same for all projects. Only a few 
parameters must be adapted to the characteristics of the specific 
datasets (weight function, number of hierarchical levels). This 
adaptation is performed manually. We have derived standard 
parameters, which are adapted to the characteristics of each 
project. This is performed in a trial and error basis with 2 or 
maximal 3 repetitions of the computation. Of course only the 
interpretation of the intermediate results requires human 
resources, the computation itself is performed totally 
automatically. In our experience most of the adaptations are 
necessary in the coarse levels. The functional model for surface 
computation is in the presented examples linear prediction, but 
any model can be used, which is capable of considering 
individual weights for the given points. 
Additionally two methods of data densification have been 
presented (sec. 4.1 and 4.3). 
In the future we will have to consider new measurement 
systems, which will provide further information of the sensed 
objects like laser scanner systems, which allow now the 
registration of multiple echoes and the intensities of the 
returned signal and in future even full waveform capture. An 
other important topic will be sensor combination. Nowadays 
sensor systems, which combine laser scanner sensors with 
digital line cameras, already do exist. The big task for the future 
will be to integrate all these information sources into one 
modelling process in order to achieve better results.