MULTITEMPORAL INTERPRETATION OF REMOTE SENSING DATA
Sonke Miiller®*, Guilherme Lucio Abelha Mota? and Claus-Eberhard Liedtke“
^ Institut für Theoretische Nachrichtentechnik und Informationsverarbeitung, University of Hannover, Germany -
{mueller, liedtke } @tnt.uni-hannover de, “corresponding author
" Department of Electrical Engineering, Pontifical Catholic University of Rio de Janeiro, Brazil -
guimota@ele.puc-rio.br
KEY WORDS: Multitemporal, Land Use, Classification, Urban, Aerial, Knowledge Base.
ABSTRACT
The automated interpretation of aerial image data is a task with increasing significance for several applications, e.g. quality control and
automatic updating of GIS data, automatic land use change detection, measurement of sealed areas for public authority uses, monitoring
of land erosion etc. The use of additional sensors could improve the performance of the automated classification; however, because
of additional costs or simple unavailability of data, this approach should be avoided. One possibility to stabilize an automatic image
analysis is using remote sensing data of the same region of different dates that is often existing. This paper presents a method how
a monotemporal knowledge representation can be expanded by a temporal component to take advantage of previous classifications
of the same scene and knowledge about the time dependency of the object classes. The present approach proposes the combination
of a semantic network, representing the generic description of the scene, and a state transition diagram, modeling the possible state
transitions for each one of the classes of interest. The probabilities of the state transition diagram are introduced as a priori knowledge
in a statistical classification procedure. Experimental results from
a series of three aerial images from 1983 up to 2001 of a suburban
region near Hannover are shown in order to illustrate the potential of the proposed multitemporal approach.
1 INTRODUCTION
The quality of geodata refers to the geometric and semantic cor-
rectness as well as to its up-to-dateness. Among these features,
most part of the effort is devoted to keep the consistence between
the geodata and the respective area. In order to achieve this aim,
it is desirable to develop an image processing system that is able
to generate automatically up-to-date geodata.
The quality of the outcome of an automatic image analysis de-
pends on the used input data and on the knowledge about the
investigated scene. In the present approach, actual aerial or satel-
lite images are the standard input for the scene interpretation. In
many cases, the increment of the cost or unavailability hinder the
utilization of additional sensor data. Moreover, the fact that aerial
and satellite images are produced in standardized intervals and
quality permits the system to work on data acquired in different
time instances.
In this paper we restrict on aerial images with a resolution of
0.31 25-7 and differentiate the following object classes :
e Inhabited area
e Forest
e Agriculture
Figure | presents example input data. These images were ac-
quired in 1983, 1988 and 2001. While the images of 1983 and
1988 are gray scale, the image of 2001 is originally colored. In
order to standardize these data, the image of 2001 was converted
to gray scale.
In this paper, the knowledge based system GEOAIDA is em-
ployed. The system was developed to interpret a scene consid-
ering aerial photographs or other raster data. The original con-
ception of GEOAIDA (Biickner et al., 2002) aims at interpreting
remote sensing data by exploiting an exclusively hierarchical de-
scription of the problem given by a semantic network. The sys-
tem is being extended actually in order to incorporate features to
manipulate temporal data.
Inside GEOAIDA, the interpretation of remote sensing data from
a given scene aims at finding out its structural and pictorial de-
scriptions. The structural description has the same structure of
the semantic network and is bound to the pictorial description.
This approach allows simultaneous access to information about
the object type, the geocoordinates and all other attributes calcu-
lated during the analysis.
The temporal approach proposed is based on (Pakzad, 2002)
which employs a transition graph to describe the temporal de-
pendencies between the classes of interest. Such strategy enables
the user to formulate temporal a priori knowledge and to use it
during the automatic analysis in connection to an older classi-
fication of the scene. Thus, in the present paper, besides the
structural knowledge, knowledge about temporal dependencies
is exploited to refine decisions in the interpretation process. The
temporal knowledge is used in a statistical classification process
and directly influence the interpretation result. In a previous work
(Müller et al., 2003) the usage of temporal knowledge was limited
to generate hypothesis for possible new states of region, without
the usage of real probabilities.
The remaining part of this paper is organized as follows. Section
2 describes briefly the approach. Section 3 presents the classifi-
cation strategy. Section 4 presents the experimental results and
section 5 the conclusions.
2 PROPOSED APPROACH
The proposed approach is based on the knowledge based image
interpretation system GEOAIDA (Bückner et al., 2002) devel-
oped at the Institut für Theoretische Nachrichtentechnik und In-
formationsverarbeitung, University of Hannover. In GEOAIDA,
1244
Interne
p=0.65
1
x
Figure 2
State Trz
| input
2 input
3 5 result
+ Coat
5 [result
the struct)
mantic ne
used state
indicate a
two image
placed ne:
ited area,
is very uni
are possib
from fores
sible with
In figure 3
1983 is pre
knowledge
used as a p
aerial imag
of 2001. T
edge, the ii
temporal ki
