EXPERIENCE WITH AUTOMATIC DEM GENERATION 
F. Ackermann and W. Schneider 
University of Stuttgart 
Germany 
Commission IV 
ABSTRACT: 
The development of methods for the automatic generation of digital elevation models (DEM) by digital image processing has made 
great progress recently, and computer programs are becoming available. The paper refers to the MATCH-T system which claims 
particularly high performance. Several DEMs, derived automatically from digitized aerial photographs, are analysed with regard to 
accuracy and speed of operation. The results confirm that the conventional standards of DEM generation are surpassed with regard 
to accuracy and economy. 
KEY WORDS: DEM, Digital Image Processing 
1. The MATCH-T System 
One of the recently developed systems for the automatic DEM 
generation by digital image processing is the MATCH-T 
system. It is based on original research at the University of 
Stuttgart (Forstner, Hahn 1988; Hahn 1989) and has been 
developed as a program at INPHO GmbH, Stuttgart 
(Ackermann, Krzystek 1991; Krzystek, 1991; Krzystek, Wild 
1992). This paper is only concerned with empirical results. 
Therefore the system features are only recalled very briefly. 
The MATCH-T system operates with pairs of digital or 
digitized aerial photographs. The image data are digitally 
oriented and normalized in order to have the benefit of epipolar 
geometry. The method is based on extraction of feature points 
in both images, by using the Forstner interest operator. The 
feature points are matched to pairs of points and analytically 
processed to terrain points. The DEM is derived patch-wise by 
robust fitting of finite elements to the terrain points. The 
system is designed for high performance. Therefore, the 
processing goes through pyramidal data structures (image 
pyramid, feature pyramid, 7 or 8 levels), in order to be fast and 
autonomous, i.e. practically independent of pre-knowledge 
about the terrain. The number of extracted feature points is 
very high, leading usually to about 500 000 terrain points per 
stereo-pair. The derived DEM may consist of 50 000 to 80 
000 grid points. Thus, a finite element grid unit may contain 
perhaps 10 terrain points. This redundancy allows outliers (i.e. 
points on trees or houses, and remaining mismatches) to be 
automatically recognized and eliminated. The redundancy is 
also the basis for high DEM accuracy, and it allows the 
derivation of an internal accuracy indicator. It is based on the 
empirical residuals and theoretical propagation of errors 6, = 
Co Q_)- 
ZZ 
986 
2. Controlled Empirical Results 
2.1 The test material 
In this paper, without further theoretical discussion, some 
empirical investigations and results are presented concerning 
the accuracy performance and the time performance of the 
MATCH-T system. All results refer to the batch version of the 
program and to the processing on the Silicon Graphics 4D35 
UNIX workstation. The system is currently being transferred 
to the Intergraph image workstation. The photographs were 
digitized with the Zeiss PhotoScan PS1, at 15 pm pixel size. 
The tests with five different stereo-pairs are summarized 
hereafter. A very detailed description is not possible nor 
necessary. In table 1 the essential features of the five projects 
are collected. 
The photographs were digitized for the areas of stereo-overlap. 
The DEM computation covered the model areas in principle 
completely. Only some small built-up areas and forest areas 
were defined by border lines and excluded from the accuracy 
evaluation, although they were included in the DEM 
computation. The terrain areas are reasonably suited for 
automatic DEM generation, as they are mostly open, although 
covered with scattered vegetation and houses which the system 
must be able to handle automatically. The first three projects 
relate to rather smooth terrain and large-scale photography. 
Projects 4 and 5 apply medium-scale photography and concern 
hilly and finally very rough and rocky terrain. 
The figure 1 gives an impression of the types of terrain and 
surface coverage. 
Pi 
image s« 
  
Pr 
image sc 
  
edges of total 
Figure 1: Ir