DETERMINATION OF TERRAIN MODELS BY DIGITAL IMAGE MATCHING
METHODS
Christoph Bauerhansl ^ *, Franz Rottensteiner ® Christian Briese *
Institute of Photogrammetry and Remote Sensing, Vienna University of Technology, Gufhausstrafe 27-29,
A-1040 Vienna, Austria — {cba,cb} @ipf.tuwien.ac.at
"School of Surveying and Spatial Information Systems, The University of New South Wales,
UNSW Sydney NSW 2052, Australia — f.rottensteiner@unsw.edu.au
Commission IV, WG IV/6
KEY WORDS: DEM/DTM, Matching, Hierarchical, Modelling, Photogrammetry
ABSTRACT:
Today, digital terrain models (DTMs) are used in many fields of science and practice. When modelling the earth's surface it is
necessary to make a clear distinction between terrain models, i.c. models representing the terrain in the sense of the ‘bare soil’, and
surface models, i.e. models that also include artificial buildings and vegetation. A DTM should not be influenced by off-terrain
points such as points on vegetation and on buildings. Hierarchical robust filtering, a method for eliminating the influence of the off-
terrain-points in DTM generation, has been shown to give good results for airborne laser-scanner-data. In this paper, we want to
show that this method can also be applied successfully to improve the quality of DTMs created by image matching techniques. Those
techniques deliver a digital surface model containing disturbances such as houses and forests, even if filtering methods are an
integral part of the matching process. Hierarchical robust filtering, implemented in the program package SCOP++, can be used in
order to eliminate these errors in the DTM. The results presented in this paper show the improvement of DTMs created by matching
methods that can be achieved by this method, using test data from different areas of interest.
1. INTRODUCTION
Digital terrain models (DTMs) are important components in
Geographic Information Systems, and they are used in many
fields of science and practice. There are different ways of
representing a DTM in the computer. Often the terrain is
represented by heights in a regular grid. For a high-quality
description of the terrain, a hybrid raster can be used,
containing not only the grid heights, but also geomorphologic
elements such as break lines or spot heights. The elevations of
the grid points are not measured directly, but they have to be
determined from irregularly distributed points and the
geomorphologic elements, e.g. by linear prediction, or by
interpolation based on finite elements (Kraus, 2000). The
original points can be acquired in different ways. Traditionally,
they were measured manually in stereoscopic images. Image
matching methods have been successfully applied to automate
DTM generation from digital aerial images (Gülch, 1994;
Krzystek, 1995), which has resulted in operational software
modules such as MATCH-T by INPHO GmbH (INPHO, 2003)
that are widely-used today. in addition to photogrammetric
techniques, the original data for DTM generation can also be
acquired by airborne laser scanning (ALS) (Kraus, 2000).
[t is common to both image matching techniques and ALS that
the original point cloud represents the earth's surface as it is
seen from the sensor's vantage point. The original point cloud
does not only consist of points located on the terrain, but it also
contains off-terrain points on houses, trees, or other objects.
Thus, a model interpolated from that point cloud is a digital
surface model (DSM) rather than a DTM. For applications such
as orthophoto production, a DSM might be sufficient. For other
applications it is essential to eliminate the off-terrain points to
* Corresponding author.
obtain a model that really represents the terrain. In image
matching, robust interpolation techniques are used to eliminate
these off-terrain points (Krzystek, 1995), but problems arise in
densely built-up regions and in forests, and manual intervention
is often required to remove remaining errors.
With respect to ALS data, hierarchical robust linear prediction
has been shown to give excellent results in densely built-up and
forested areas (Kraus and Pfeifer, 1998; Briese et al., 2002). It
is the goal of this paper to show how this method can be applied
to improve DTMs derived by image matching. We start with a
description of the characteristics of DSMs derived from image
matching and with an outline of the filter algorithm. After that,
we show how the filter algorithm is adapted to the specific
characteristics of point clouds derived by image matching.
Finally, we will present results achieved for various types of
terrain and land cover.
2. DTM GENERATION USING IMAGE MATCHING
In this work, we used the program MATCH-T from INPHO
GmbH (INPHO, 2003) for the generation of a DSM from aerial
images. MATCH-T applies feature based matching to generate a
dense point cloud. From this point cloud, an elevation grid is
interpolated by the finite element method, applying robust
estimation to eliminate false matches (Krzystek, 1995). The
major goal of this work was to create a DTM without buildings
and vegetation. MATCH-T has various parameters which
control the point density in the matching process and the degree
of smoothing during the grid interpolation. By these parameters,
the user can control the degree to which the resulting elevation
grid represents the terrain (Summit Evolution, 2001):
414
Internati
e The gr
paranx
applic:
limits
matchi
effect
smoot]
The re
selecte
matchi
The €
param
a DTN
smoot!
close ¢
The d
sparse
point |
dense :
MATC
additic
measu
deviati
on th
interpc
MATCF
required
width is
terrain p
trees an
inv the r;
chosen s
results, «
shows a
1.23 m.
Fionre 1
As in ge
the prop
width fo
improve
predictic
that tec
possible
density I
3. HI
We use
interpol:
