CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
104
which is essential in crisis scenarios. Depending on the type and
complexity of the input data, the system can be run in a fully
automatic or semi-automatic mode, where a human operator can
edit intermediate results to ensure the required quality of the final
results.
Section 2 describes the generic near-realtime classification sys
tem with the objective to classify and evaluate objects using re
mote sensing and other available data. In Section 3 the system is
applied to road objects in case of natural disasters. Two test sce
narios of flooded areas are used to verify the system. By means
of manually generated reference data, the applicability and effi
ciency of the system is evaluated in Section 4. Finally, further
investigations in future work are pointed out.
2 CLASSIFICATION SYSTEM
The fusion of multi-sensor images is an important issue, because
the corregistration between optical and radar images is still a cur
rent research topic (Pohl and Van Genderen, 1998). Methods
such as mutual information can be applied for the system (Inglada
and Giros, 2004). The system has to deal with multi-temporal
images having the possibility to derive important information on
time. This leads to an even more complex corregistration pro
cess. Change detection algorithms can provide information about
the variation of assessed objects. In this article the temporal fac
tor is neglected, but will be an essential part in future research.
The main core of the system represents the classification. The
goal is to classify each object into a different state Si. For each
object probabilities are derived belonging to a certain state. The
methods estimating the probabilities depends on the data: typ
ical examples are multispectral classification or fuzzy member
ship functions (Figure 2).
The goal of the developed classification system is the assessment
of GIS-objects using up-to-date remote sensing data. The system
is designed in a general and modular way to provide the opportu
nity to label GIS-objects into different states. Typical states de
scribe the functionality of infrastructure objects as roads or build
ings. The generic system embeds different kinds of image data:
multi-sensor as well as multi-temporal data. Additionally, any
kinds of available data sources and spatial knowledge, which con
tributes information for the assessment, can be embedded. Typi
cal examples are digital elevation models (DEM) and further G1S
information, e.g. land cover or waterways. The minimum re
quirement of the system are the objects to be assessed and one
up-to-date image which provides the information for the assess
ment.
Time
Point t-|
Optical
Imagery 1
SAR
Imagery 1
DEM ►
GIS-
object
1
Classification
System
GIS ►
I
Optical
Imagery 2
Time SAR
Imagery 2 Classification
System
GIS-
Object
I
data 1
► method 1 -
—►
PlIlSi) Pil:S;>
••• > Pd,S,
data 2
► method 2 -
—►
Pd.Sp Pd:S;r
■■■ t Pd;S!
...
— - -
—►
■■
data n
► method n
—►
PaSh PJ„S:>
••• ) Pd&
Figure 2:
Derivation of probabilities
from data
using various
methods
Beside the derivation of the individual probabilities from each
data source the combination plays a decisive role:
Psi =Pd 1 ,s 1 ®Pd 2 ,s 1 <S> • • • <8>Pd n ,s x
Ps 2 = Pd 1 ,s 2 <S> Pd 2 ,s 2 ® ■ ■ ■ ® Pd n ,s 2
: (1)
PSi = Pd 1 .Si ® Pd 2 .Si ® • • • <8> Pd n ,Si■
The variable Pd n ,Si denotes the probability that the state Si oc
curs given data d n ■ The indices i and n describe the number of
available states and data, respectively. The result pSi shows the
probability that a GIS-object belongs to the state S',-. For each
type of data weights w n can be introduced in order to cope with
the different influence of infonnation content. Hence, Equation 1
for one state i leads to:
i
up-to-date
map
Figure 1 : Classification system
PSi = m • Pdi,Si ® • • • <8> w n • Pd n ,Si■ (2)
Finally, the object is assigned to the state S, with the largest prob
ability pSi. A basic characteristic of the whole system is the
combination at the probability level in order to remain flexible
concerning the available data.
The classification system depicted in Figure 1 can be subdivided
into different components. Starting point are the GIS-objects to
be assessed. Secondly, the input data as imagery or digital eleva
tion models which contribute the information for the assessment.
In the following this information is called data. Thirdly, the clas
sification system by itself and, finally, a resulting up-to-date map.
3 MODEL FOR ROAD OBJECTS
After describing the generic system, a model is shown which as
sesses linear objects as roads after flooding. However, this model
is transferable to other linear objects like railways and further
