4, 9-11 Nov. 1999 
lower points (which are in 
' prediction is used for the 
stribution of laser scanner 
und surface is no longer a 
ribution with a strong bias 
1e points near the ground 
ributed whereas the vege- 
iduals with reference to the 
ep a rough surface approx- 
whether ground points or 
fluence. Thus, the surface 
between the ground points 
ified weight function from 
ute weights from residuals. 
irst interpolation step and 
. Each measurement (i.e. 
a weight according to its 
nsidered in the next inter- 
/eights attract the surface, 
influence. Therefore, the 
to the ground, disregard- 
vegetation points obtain a 
vious step. This process is 
1. We use it to re-compute 
measurements into ground 
e. vegetation points in the 
ication is done on the basis 
als. For a detailed descrip- 
and [Pfeifer et al., 1999]. 
of the methods mentioned 
tive approach, filtering of 
and the classification and 
aneously. Of course it can 
urces with an asymmetric 
rithm for laser data 
ots of experience with this 
antages and deficiencies of 
; are not only valid for our 
lata processing in general. 
ome general laser scanner 
the classification are per- 
°p terrain this is an advan- 
always performed relative 
ie ground surface may be 
on steps, nevertheless the 
vays be captured. This is 
es which consider only the 
ork on original data or it 
ve pre-classified data. An 
ation perform by the com- 
ner data results in a higher 
model derived from these 
the points are given in an 
. 
occurences 
  
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999 
  
  
  
  
  
  
  
  
  
  
-4.0 -2.0 0.0 2.0 
4.0 6.0 8.0 10.0 
Figure 2: Residual distribution after the first interpolation step. The ground points are clustered at around —3m, whereas the 
vegetation points have residuals up to 10m. The weight function p(r)which is used to determine a weight for an observation 
(a z-measurement) is superimposed. Note, that the origin g of the weight function is negative and that the left branch of the 
weight function is identical to 1. Thus, ground points obtain higher weights than vegetation points. 
3. High degree of automation. Mainly the initial setting 
of parameters and the end-inspection are left to the 
user. 
. There is the possibility to eliminate negative blunders 
as well. By an appropriate setting of the weight func- 
tion negative errors can be given a lower weight and 
less influence, too. For this end, the weight function 
shown in figure 2 would decrease also for the left branch 
of the function. Of course this decrease need not be 
symmetrical to the right side. 
However, in order to maintain structures like break lines 
as good as possible, a soft filtering of negative errors is 
required. On the other hand this prevents the detection 
of negative blunders. Small edges are always blurred, 
whereas the general structure can be preserved. 
. The ground model has a very high quality. This is 
due to the interpolation process of linear prediction. 
On the other hand it is necessary to solve an equation 
system which has a dimension equal to the number of 
points. Therefore, this algorithm can only be applied 
patch-wise. 
Deficiencies: 
1. There is still interactive post processing required: 
Dense bush groups of larger size with very low penetra- 
tion rate cannot be detected. Thus, a manual inspec- 
tion with the help of digital ortho photos and/or other 
data sources is still necessary. Depending on the time 
of flying and the type of tree, the penetration rate can 
be as low as 0%. Dense deciduous trees during sum- 
mer time or young densely planted conifer trees can re- 
flect all laser rays in the tree tops [Rieger et al., 1999b] 
and [Rieger et al., 19992]. If this occurs for larger ar- 
eas, laser scanning is not an applicable method. 
Usually very large buildings are not eliminated. The 
situation corresponds to the point just mentioned. In 
such a case another data source like a cadastral map 
(or again the digital ortho photo) are necessary to de- 
tect and/or eliminate such artefacts. On the other 
hand, smaller buildings but also bridges are eliminated. 
It depends on the purpose of the project, if this is an 
advantage or not. 
2. Negative errors occur, too. By this we mean laser 
points which are "measured" below the terrain. Be- 
cause the algorithm puts more emphasise on the lower 
points, these are usually classified as ground points. 
This leads to a (topsyturvy) cone like pattern in the 
surface model, where the peak is at the negative er- 
ror and the basis of the cone is on the actual ground 
surface. One source for these errors can be multi-path 
reflections. We observed such blunders in water ar- 
eas as well as in urban areas. In water areas (fig. 1) 
we observed a number of neighbouring points (ca. 4) 
in a scan below the surface. They are between 0.5m 
and 2m under "ground" (i.e. the water surface). As 
all approaches stress the lower points, this behaviour 
is common to all approaches. 
Generally, structures above the local mean surface (e.g. 
embankments) appear smaller, structures below the 
mean surface (e.g. ditches) are enlarged. This can lead 
to a shift of terrain features. 
3. There is no consideration of break lines in the terrain. 
Thus the edges of an embankment (or similar terrain 
features) are usually blurred. 
4. For our algorithm the setting of the parameters is 
rather sophisticated, depending on the parameter set- 
ting negative blunders may be stressed 
5. The computation times are rather long. Compared to 
simpler algorithms this approach requires considerably 
more computation time. However, as the process runs 
automatically, once the parameters are set, this is less 
of a problem. Furthermore, the increase in the quality 
of the DTM justifies this additional effort. 
The first three deficiencies are general problems of laser scan- 
ner data evaluation. They also apply to grid generation mech- 
anisms, especially if they favour lower points. Solutions need 
to be found in these areas in order to speed up the processing 
of laser scanner data. For a high quality of the final DTM 
either break lines are necessary, or the point density has to 
be very high (1 point per m?). 
4 Examples 
In the meantime a considerable amount of experience has 
been gained in the processing of laser scanner data and the