1л: Paparoditis N., Pierrot-Deseilligny M.. Mallet C... Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France, September 1-3. 2010 
In order to obtain the DTM from classified ground points the 
interpolation techniques applied often consider the properties of 
the terrain, e.g. in the variogrant of kriging or in the 
neighbourhood size of local interpolation methods. For high 
demands break lines and peak heights may be inserted at those 
locations where the global assumptions on the terrain are 
violated, e.g. the assumptions on smoothness. For surface 
models the situation is more complex as different objects, 
including terrain, roofs, vegetation canopy, etc., have to be 
considered. The surfaces of these objects have different 
properties, which have effect on the appropriate interpolation 
method. Examples include the treatment of height jumps, 
methods for filling data gaps, or what constitutes an outlier. 
Additionally, the sampling interval may become relatively low 
for some objects, e.g. individual trees, which suggests that no or 
only minimal filtering of random measurement errors shall be 
performed. Thus, applying one of the DSM computation 
methods listed above with only one set of parameters, can 
hardly fulfil the different interpolation demands arising from 
different object surfaces. 
The following two approaches for calculating DSMs shall 
demonstrate this. 
• For areas with surfaces discontinuities e.g. along 
breaklines, step edges or forests the DSM calculation 
based on the highest point within a defined raster cell 
(DSM max ) leads to a good approximation of the real 
surface. As shown in Fig. la the DSM max values are 
near to the tree tops and near to the terrain within 
forest gaps. Furthermore, the DSM inax represents the 
surface in a proper form along building edges and 
roof ridges (Fig. lb). On the other hand, for inclined 
smooth surfaces (Fig. lb) the derived DSM lliax leads 
to an artificial surface roughness and to an 
overestimation of the modelled surface. 
• In contrast to the DSM niax , the surface height 
calculation based on moving least squares 
interpolation (DSM mls ) smoothes surface 
discontinuities as illustrated in Fig. la (1-3). On the 
other hand the DSM m i s represents a good 
approximation of inclined smooth surfaces (Fig. lb). 
Obviously, a mixture of both models would be the best choice 
in the above examples, with the DSM mls used for smooth 
surfaces (e.g. street and roof) and the DSM max for rough 
surfaces (e.g. the high vegetation areas). However, due to the 
missing semantic information of the acquired data, the land 
cover (class) of the recorded echoes is unknown in advance. 
The contribution of this article is a conceptual framework, its 
methods, and the description of a specific implementation for 
the computation of improved DSMs. Based on (i) land cover 
proxies and (ii) properties of the data coverage (e.g. gaps), 
different algorithms for the DSM computation are applied. The 
algorithms used are not new, but it is pointed out that the 
synthesis of methods is the key idea presented in this paper. The 
overall aim is to generate DSMs, which allow improving visual 
analysis and algorithms working on the DSM and nDSM. 
In section 2 the conceptual framework and its specialization 
addressing the above examples are presented. This section also 
includes the implementation within OPALS (2010). 
O = First echoes ■■■ = z-Value DSM,«. 
C'i - Intermediate + last echoes = z-Value DSM™, 
Figure 1. Illustration of DSM calculations based on the highest 
point within a defined raster cell (DSM max ) and moving least 
squares interpolation (DSM mls ) for a forested area (a) and for a 
building roof (b). In (a) the DSM mls underestimates the heights 
of tree tops and overestimates the heights of the terrain for 
forest gaps. In (b) the DSM max overestimates the inclined roof 
plane and introduce an artificial roughness. 
Depending on the surface roughness either the DSM niax or the 
DSM m | S is used for the determination of the DSM height. The 
supposed work flow is applied for different ALS data (e.g. 
discrete, full-waveform) with different point densities for 
Austrian test sites located in forests, agricultural land and urban 
areas as introduced in section 3. In section 4 the examples are 
presented and compared to traditional DSMs. 
2. METHODS 
The basic principle of the suggested approach is to first apply 
data analysis on the original point cloud and - depending on its 
outcome - chose different algorithms for DSM computation 
(see also Fig. 2). Each analysis step results in a layer (/ b ...) 
with qualitative or quantitative information, all together nl 
layers. A rule-based analysis of the different layers at each grid 
position selects the appropriate interpolation method (iri\, ...), 
from all together run methods. This can also be seen as selecting 
the elevation of a specific (primary) surface model computed for 
the entire area. In contrast to the layer data, the surface model 
values must be metric. 
The rules for selecting the appropriate elevation are provided by 
an expert. Such a rule may be to choose a specific 
method/model, but also the combination of different methods 
(e.g. the average) might be feasible.