International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
ADS 40 camera (Leica Geosystems) or the ALTM-3033/3070 
(Optech) with the Emerge DSS camera. Schiewe (2004) gives 
a detailed overview of these systems. 
2.2 FALCON sensor system 
In 2002 the company TopoSys (Germany) released its airborne 
FALCON system which not only delivers laser scanning 
elevations but also acquires multi-spectral imagery 
simultaneously (Lohr, 2003). In the following a scene covering 
the City Memmingen (located in Southern Germany) and its 
neighbourhood will be processed. 
The imaging sensor consisting of a line array scanner that 
acquires data in four bands in the visible and near infrared 
spectrum with a ground pixel size of 0.5 m and at a radiometric 
resolution of 11 bit. 
The laser scanner operates with a glass fibre array producing 
a parallel acquisition pattern. The resulting elevation data with 
a height accuracy of about + 0.2 m are delivered with a raster 
grid size of (0.5 m)” or (1.0 m)”, respectively, having a swath 
width of 250 m for a flying altitude of 1000 m. 
A special feature which is of major interest for the following 
study is the ability to capture multiple reflections of the laser 
beam (Lóffler, 2003): From each laser pulse of the FALCON 
system the respective lowest and highest values (first echo and 
last echo, respectively) are recorded (other systems might 
record even more reflections). It is important to note that this 
separation is only possible if the two echoes are at least one 
pulse length (i.e. 5 ns, corresponding to 1.5 m, taking the 
speed of light into account) apart from each other. That means, 
that rather low objects like bushes cannot be separated from 
the ground by subtracting first and last echo. On average one 
  
  
  
  
  
  
  
  
  
  
raster 
cells E | | | J 
FE-high : ® € e { e 
LIÉ 1/9 |. 
-— — — 
FE-high ® . 
- LE-low : 
= —@— 
Figure 1. Schematic profile representation of a building 
through different reflections 
grid cell of (1.0 m)? of the derived Digital Elevation Model is 
“hit” by a couple of laser beams. Hence, we have the 
possibility of distinguishing the respective highest and lowest 
values of the first and last echoes, respectively (called FE-high, 
FE-low, LE-high, LE-low). As a consequence, the respective 
elevation models differ in their representation of the scanned 
surface. For example (see figures 1 and 2), the first echo- 
highest values (FE-high) are able to represent narrow, 
outstanding objects (like buildings and vegetation) while 
narrow gaps in between disappear. Outstanding objects are 
extended in size. On the other hand, the last echo-lowest 
values (LE-low) show the opposite properties. Consequently, 
the difference (FE-high minus LE-low) leads to the 
representation of vegetation and (rather broad) building edges. 
  
Internat 
With th 
previou: 
increase 
eventua 
levels b 
been lai 
on featu 
several 
describe 
demons: 
level fu 
applicat 
edges. 
top- 
down 
  
Figure 2. Different representation of buildings through 
different laser pulse reflections: image data (upper 
left, for comparison purposes), FE-high (upper 
right, buildings too large), LE-low (lower left, 
buildings too small) and difference FE-high minus 
LE-low (lower right, broad building edges). 
3. GENERAL REMARKS ON PERCEPTUAL 
ORGANISATION AND FUSION 
The process of (automatically) interpreting a remotely sensed 
scene is strongly correlated to the process of cognitive 
perception (see also figure 3). The latter does not only include 
the acquisition and representation of various stimuli by a 
human being, but also their organisation and interpretation. 
Cognitive perception can be seen as a hybrid process, i.c. it 
contains procedures in both directions, top-down (or model 
driven) and bottom-up (or data driven) at the same time. 
Furthermore, it is not simply a linear process, but it also 
includes feedback mechanisms at various stages. 
However, actual implementations of scene interpretation 
algorithms do not consider all these principles. In particular, 
we have still a big, application dependent gap between the 
extracted features and the related object characteristics as 
described in the knowledge (or memory) representation. The 
necessary bridging process of perceptual organisation (like 
grouping of lines and areas into meaningful structures) has 
been neglected too much in the past and is mainly responsible 
for unsatisfying classification results (Schenk, 2003). 
606 
Figure 3. 
4.1 Def 
For the d 
those cui 
maxima 
or the b 
gradients 
type, whi 
as a sul 
surface « 
objects (« 
of terrai 
DTM, lil 
the comr 
because t 
generatec 
Surface e 
or they b 
and uppe 
embankn