International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
removing trees and buildings from DSM. This process is called 
"filtering." The manual refinement is time consuming. The strategy 
for automatic filtering is based on a combination of geometric 
conditions together with Linear Prediction. The filtering was 
applied to photogrammetric acquired data via automatic image 
matching. 
    
Figure 2: DSM - Yellow points belonging to surface of objects. 
   
8B 8 e 8 vB 
Figure 3: DHM - Only red points — belonging to bare earth 
surface. 
BREAKLINES 
Digitized lines that define critical changes (natural or 
manmade) in topographical shape. The most important reason 
why break lines should be located in global object reconstruction is 
the demand to keep the number of unknown geometric, DEM 
parameters low so that large image areas could be processed 
simultaneously and as fast as possible. As in the least squares 
adjustment the size and the structure of the normal equation matrix 
depend on the unknown quantities, and as the object surface 
elements can be eliminated in the matching, the amount of 
unknowns depends mostly on the number of DEM grid points. 
Beside the largest errors of the matching occur due to break lines 
when a continuous model is used in object surface reconstruction. 
Thus, if a continuous model is used without considering break lines, 
the result of the image matching is not reliable especially at break 
line locations. These reasons imply that in global object 
reconstruction break lines should be detected at first so that break 
line areas could be better modelled. 
Thus, break lines can be detected by computing one ortho-image 
per aerial input image and the difference (or deviation image in case 
of more than two input images) ortho-image at the same mage 
pyramid level, and interpreting large differences between the ortho- 
images as defects in the geometric model originating, e.g., from not 
modelled break lines. 
For the interpolation we use the *moving plane" method. The plane 
is stopping at the break line, and will not eliminate the points 
beyond it. (see figure below). 
The laser scanner data are a point cloud without structural 
information. A qualified terrain model, on the other hand, exists in 
the inclusion of structure lines, especially break lines. A terrain 
model, integrating a close dawn grid (raster data) and structure 
lines (vector data) , is called hybrid DHM. The raster data as well 
as the vector data is smoothed, the hybrid DHM shows 
discontinuities in the first derivation at the break lines. 
The solution found, combines filtering and interpolation of terrain. 
Is called a robust interpolation or robust linear prediction. The 
algorithm is embedded in a hierarchical approach, however, the 
filtering of the laser scanner data on one level will be described 
first. In this Algorithm a rough approximation of the surface is 
computed first. Next, the residuals, i.e. the oriented distances from 
the surface to the measured points are computed. Each (z) 
measurement is given a weight according to its distance value, 
which is the parameter of a weight function. The surface is then 
recomputed under the consideration of the weights. A point with a 
height weight will attract the surface, resulting in a small residual at 
this point, whereas a point that has been assigned a low weight will 
have little influence to the figure of the surface. If an oriented 
distance is above certain value, the point is classified as off-terrain 
point and eliminated completely from this surface interpolation. This 
process of weight iteration is repeated until all gross are eliminated 
or a maximum number of iterations is reached. Program 
RASCOR we maximum 2 iterations. The reason for this limitation 
is that more itterations will result in a very smooth earth surface due 
to the risk of elimination of too many points. 
For the interpolation we use linear prediction. In this method the 
classification and DHM generation are performed in one step, 
there is no assumption that the terrain is horizontal. It is applied 
patch wise to the data, which result is an adaptive setting of the 
shift origin of the weight function. Furthermore, the base are 
determined for each patch separately, too. The process yields a 
smooth surface, that means the accidental (random) measurement 
errors have also been filtered. 
However, the algorithm relies on a “good mixture” of ground and 
off-terrain (vegetation) points, which can also be seen as a high 
frequency of change from change from ground to vegetation points. 
This is necessary for a reliable determination of the shift value for 
the origin of the weight function. In this high frequency is not given, 
we need to provide the nput data in a suitable form. This can be 
achieved by inserting the robust linear prediction in a hierarchic 
environment. 
On the right is the DHM derived by the linear prediction without 
any manual intervention. 
5 
  
Figure 4 Surface interpolated using the original data. DHM 
derived by the filtered data. 
PROGRAM RASCOR AND FILTERING 
RASCOR is a program developed for automatic 
improvement of digital surface models to digital elevation models 
(raster data set). 
Program RASCOR can analyze, improve, smooth and interpolate 
a digital elevation model (DEM) which may be created by 
automatic image matching or laser scanning (LIDAR) in an equal 
Interna 
spacing 
solid g 
building 
by ma» 
sudden 
interpol 
height « 
and a | 
tilted pl 
(smooth 
surface 
The rec 
analysis 
as hom 
RASCC 
line wil 
inclinati 
Progran 
the filte 
accepte 
RASCC 
This me 
mountai 
differen. 
neighbo 
errors. ' 
larger el 
Even la 
elevatio 
change 
building 
optimal 
the builc 
Other la 
by a m 
longer [ 
profiles 
longer p 
against t 
linear re 
be used. 
not ever 
All thes 
which I 
analysed 
polynomr 
program 
with diff 
is usuall 
the desc 
In the 
neighboi 
tolerance 
weight f 
neighboi 
cutting ii 
do not I 
but this 
automati