tion 
ction and 
nagement 
; like data 
ations are 
' solely a 
be related 
generate 
building’, 
two laser 
ts have to 
results in 
Bereichen 
erken um 
ellen. Im 
isen, z.B. 
igentliche 
itpunkten 
Anbauten 
bjekte mit 
nale wie 
egetation’ 
erung hin 
rden nun 
'eworden' 
Airborne 
1al system 
t provides 
ding slant 
nd spatial 
es of this 
/, fast data 
the height 
tions and 
efore, the 
yr weather 
r disaster 
1adows on 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
data acquisition and analysis like it can be observed for aerial 
images. In contrast to aerial images laser scanning provides 
dense height data with a specific characteristic of the acquired 
data. For instance buildings - and most vegetation objects - are 
characterised by a series of significant gradients along their 
border lines. Modern laser scanning sensors are able to record 
data in two different modes, first and last pulse response, which 
leads to somewhat different appearance of 3D objects in the 
derived laser scanning data (Steinle & Vógtle, 2000), what can 
be used for a better analysis of the information, e.g. to detect 
deciduous trees. 
In general, changes or modifications of 3D objects can be 
observed by examining changes of their geometry. Using laser 
scanning data, elevation differences inside or around a building 
area have to be detected and analysed for this purpose. 
In this approach, changes will be detected by comparison of 
digital surface models (DSMs) of two different dates (t, and t;). 
An analysis of solely a differential DSM (= DSM(t;) - DSM(t;)) 
would lead to ambiguities, e.g. attachments or modifications of 
buildings could not be related to the affected building. 
Additionally, changes of all other objects like trees, bushes, 
infrastructure devices etc. have to be excluded when intending 
to use this methodology in the context of a disaster management 
system where the task is to provide rescue measures for people 
being trapped in damaged buildings. Therefore, in a first step a 
segmentation based on a specific region growing algorithm is 
applied to generate separate 3D object areas. In a second step 
each segmented object has to be classified in order to exclude 
all non-building objects. After elimination of these objects, the 
correspondence between accordant building segments of the 
two data sets (at the dates t, and t;) has to be determined by 
inspecüng overlapping areas. Inside overlapping objects parts 
the DSMs of both dates are analyzed to classify the buildings 
into not-modified, heightened or decreased. Not overlapping 
parts are analysed in regard to new or teared-off structures. 
2. LASER SCANNING DATA 
The basis for the change detection procedure are laser scanning 
derived DSMs. Laser scanning is operational since 
approximately 10 years. It is a measurement technique using 
basically the reflections of an emitted laser light signal, 
produced by the aircraft-borne system itself, to determine the 
distance between the system and the measured point. On the 
aircraft, positioning and orientation sensors of high precision 
are installed in addition. Therefore, 3D coordinates of a 
measured point can be computed (more details see e.g. Wehr & 
Lohr, 1999). 
Here, data acquired by the pulsed laser scanning systems 
TopoSys I and TopoSys II were used (see TopoSys, 2004). The 
measurements were carried out in last pulse mode and the 
measured points interpolated onto a regular grid of Im x Im 
grid cell size, i.e. digital surface models (DSM) of Im 
resolution were produced. 
A first flight over the city of Karlsruhe (south-west Germany) 
was carried out in January of 1998 using TopoSys I. another 
one in spring 2002 using TopoSys II. Both flights were carried 
out with an airplane in a flying altitude of about 800m above 
ground. 
429 
3. SEGMENTATION AND CLASSIFICATION OF 3D 
OBJECTS 
For each of the two original data sets DSM (t,) and DSM (t) a 
segmentation of 3D object areas has to be performed. In order 
to get exclusively 3D objects on the surface of the Earth - i.e. to 
eliminate the influence of topography - a so-called normalized 
DSM (nDSM) was determined for each data set. Therefore, in a 
rough filtering process, points found to reliably lie on the 
ground were extracted (e.g. Weidner & Forstner, 1995; Sithole 
& Vosselman 2003; Vogtle & Steinle, 2003) which can be 
interpolated to a digital terrain model (DTM). The resulting 
nDSM for each date can be calculated easily by subtracting this 
DTM from the original DSM. 
A segmentation procedure is now applied to these nDSMs 
which is based on an adapted region growing algorithm. 
Because this application is focussed on buildings and their 
changes, the parameters of this region growing algorithm are 
chosen in such a manner that nearly all buildings are segmented 
but only few vegetation objects. Taking into account that a 
building has certain minimal dimensions, this process starts at a 
point (crystallisation point) in which neighbourhood n points 
(e.g. n=5 or n=9) exceed the minimal building height (e.g. 
2.5m). lteratively, new neighbouring points are added to this 
area if their height difference does not exceed a maximum 
acceptable value - defined for instance by the maximal slope of 
roof planes which may occur in urban areas, e.g. l.2m 
(homogeneity criterion). Therefore, region — growing 
automatically stops at object boundaries of buildings and higher 
vegetation objects. 
Using these parameter values, nearly all buildings could be 
segmented in the used test site “Karlsruhe”. A lot of vegetation 
objects with significant height texture were not segmented due 
to the restrictive definition of homogeneity criterion, a 
favourable effect for this application. 
An extraction of object-relevant features is performed inside 
each segment area, e.g. amount of significant border gradients, 
height texture, first/last pulse differences, shape parameters and 
laser pulse intensities. These features are used to classify all 
segments into the main object classes buildings, vegetation and 
terrain (Vogtle & Steinle, 2003). Different classification 
methods based on fuzzy logic and maximum likelihood have 
been tested for this kind of data, in the case of fuzzy logic also 
different inference schemes like Maximum, Minimum, Product 
or Weighted Sum were applied. Our investigations show that 
over all classification rates of about 8996 to 9596 - dependent on 
the characteristics of different test sites - may be obtained 
(Tóvári & Vógtle, 2004). 
Now, these classification results are used to exclude all non- 
building objects - in test site Karlsruhe mainly vegetation; due 
to a quite flat terrain surface no small terrain rise was included, 
e.g. rocks. The final result is an segmentation image where only 
building areas remain, each with an individual segment number, 
nearly all other objects are excluded; an example is shown in 
Figure 2. It is the result obtained for the scenery shown in 
Figure 1, where a part of a nDSM determined for the campus of 
Karlsruhe university is given in grey-value coded 
representation. 
If reference data (e.g. previously created and verified 3D city 
models) are available for the first date t; - as it is the normal 
case for applications in disaster management - the change 
detection process for buildings can be started directly without 
this segmentation and classification step. In this case a