of reasons 
s and poor 
| on editing 
ed by using 
data fusion 
the original 
depend on 
on is less 
ssessing the 
1l affect the 
of the data 
he gridded 
he resultant 
the terrain 
] hence no 
ons. 
nistic and 
tly on the 
formulae 
based on 
ich is the 
bunting for 
iso called 
1 to reduce 
odel of the 
ethods are 
erpolation, 
a dimpled 
(Maune et 
hbourhood 
hbour. The 
points are 
that use a 
ses overall 
hat passes 
for gently 
uitable for 
helpful in 
ilable data 
e distance 
veights for 
e points is 
e sampled 
1eous data 
iuse of the 
the model 
11 trend of 
S better if 
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India, 2002 
4.0 POTENTIAL METHODS OF IMPROVING THE 
QUALITY OF DEMs 
4.1 General 
From the conclusions in the previous two sections, three 
methods stand out as being candidates for improving the 
quality of DEMs: 
1. Improved stereomatching, which may include the use 
of new techniques such as wavelet transformations, 
the use of breaklines and the use of new strategies that 
recognise terrain slope and land cover. 
2. Improved representation of the terrain through 
selection of grid size and method of interpolation. 
3. Data fusion techniques that exploit synergy and 
remove blunders. 
4.2 Improving stereomatching 
The obvious way of improving the stereomatching is to 
improve the algorithm used. Algorithms have been 
developed over many years and comprise highly complex 
code, there is therefore clearly an inertia to be overcome. 
New techniques such as the wavelet transform has been 
applied to image matching, (for example, He-Ping Pan, 
1996, Tsay, 1998) and this could be a promising line of 
investigation. Before pursuing this though it is necessary to 
establish an a priori case that improvements for terrain 
analysis would be forthcoming. 
New algorithms have been developed for specific purposes 
such as building extraction; a study has been done at UCL, 
(Sohn and Dowman, 2001) and concludes that the 
introduction of 3D breaklines improves the DEM that can be 
generated. Paparoditis et al (1998) and Cord and Declercq 
(1999) also discuss modified matching algorithms for 
building extraction. These studies concentrate on high 
resolution data of urban areas and are therefore not 
immediately applicable to terrain analysis, however the use 
of breaklines might be transferable. 
Breaklines can be used in most software packages. The 
disincentive in using these however is that the only ways to 
obtain ridge lines at present are from existing DEMs or 
maps, or by manual following. Since the object of terrain 
analysis is automatic extraction of such features, this is not a 
satisfactory approach, although it is one that may be 
explored further in the section of data fusion. 
It is possible to infer breaklines from the TIN. The two 
forms of breaklines are soft breakline and hard breakline. 
The soft breakline ensures that known elevation values 
along a linear feature are maintained and ensures that linear 
features and polygon edges are maintained in a TIN surface 
model by enforcing the breakline as TIN edges. They are 
generally synonymous with 3-D breaklines because they are 
they are depicted with series of x/y/z co-ordinates. For 
example, rounded ridges or trough of a drain may be 
collected using soft breaklines. A hard breakline defines 
interruptions in surface smoothness and abrupt changes in 
surfaces, such as cliffs, ridges, building footprints, streams 
and shorelines. Although some hard breaklines are 3-D 
breaklines, they are often depicted as 2-D breaklines 
because features such as shoreline and building footprints 
are normally depicted with series of x/y co-ordinates only 
and include no elevation data. 
Another approach would be an improved strategy for image 
matching that might include more than one data set (see data fusion 
below), or such techniques as an adaptive approach depending on 
the steepness of the terrain, the presence of breaklines and the 
texture. Leica Geosystems SOCET Set software gives the use a 
choice on non-adaptive, or adaptive strategies for matching. The 
former allows the user to specify the type of terrain and the 
characteristics of the output DEM, and uses fixed parameters in the 
matching. Tests have shown that in general the adaptive approach, 
which adapts parameters during the matching process, gives better 
results. (Drummond, et al 1997, Caner, 2001), Although details of 
the SOCET Set algorithm are not available, this results suggests 
that automatic adaptive strategies can be effective. Fox and Gooch 
(2001) suggest a *Failure Warning Model' that might be used with 
adaptive matching. This identifies low accuracy areas of 
automatically generated DEMs, based in comparison of two DEMs 
generated using slightly different strategy parameter settings. 
Rosenholm (1987, 1988), Rauhala (1987) and Li (1989) focused 
on multipoint matching algorithms that are image-based and use 
simultaneous computation of parallaxes in grid points, which are 
connected with bilinear finite elements describing the parallax 
differences. Thus, the object model used is a continuous surface 
with continuous first derivatives. They also use additional 
fictitious, weighted, continuity constraints on parallaxes to 
strengthen the connections between the grid points. Rosenholm 
uses multipoint matching to bridge areas with poor signal content. 
The possible constraints on the parallaxes are: 
l. The second derivatives of the parallaxes are zero 
(minimization of the curvature). The results of Rosenholm 
are based on the use of this constraint. 
2. The first derivatives of the parallaxes are zero (minimization 
- of the slope). 
This method can be of importance at the border of the grid. 
Rauhala uses more global solutions (matching of even whole 
images) with the view to increase the accuracy and especially the 
reliability of matching without any success. 
4.3 Improved representation of the terrain 
The size of feature that can be shown is ultimately limited by the 
pixel size of the imagery. However the size of the window used 
for matching, which can be changed, is also important, and 
parameters can be changed, as discussed. above. The method of 
interpolation, however, is often fixed. The methods used by 
software packages within digital photogrammetric workstations do 
not tend to be well explained, and in any case, cannot be modified. 
Improvements may therefore be dependent on exporting data for 
use in new packages. 
Image matching only generates the elevations of the surface visible 
to the sensor. Manual editing can be used to remove surface 
features, and there are techniques for generating a bare earth 
model, or digital terrain model (DEM), but this not advocated for 
terrain evolution work at fine scales, as the surface will inevitably 
have been generalised. and possibly distorted. It is however 
necessary to remove blunders and this can be done within surface 
fitting routines. 
Scientists from the Earth sciences have made their own 
investigation into methods of surface fitting and have proposed the 
use of techniques such as dynamic modelling, geostatistics and 
fuzzy classification, (Wilson et al, 2000) Use of these tools could 
lead to techniques for handling uncertainty, identification of scale 
dependent filters and handling of sub-grid scale variability. 
ANUDEM, a package to generate gridded DEMs from point data, 
making use of additional information such as stream networks has 
619