277
AUTOMATIC OBJECT EXTRACTION FOR CHANGE DETECTION AND GIS UPDATE
Stefan Hinz
Remote Sensing Technology, Technische Universitaet Muenchen, 80 333 Muenchen, Germany
Stefan.Hinz@bv.tu-muenchen.de
KEY WORDS: Object Extraction, Image Understanding, Evaluation, Disaster Management
ABSTRACT:
In this paper we will revise the recent developments in the field of automatic object extraction for change detection and GIS update
under different aspects: (1) level of automation; (2) available sensors and image characteristics; (3) effectiveness and processing
time; and (4) verification/editing and quality of results. For most applications, none of these aspects can be considered isolated from
the others. We will exemplify the importance of these aspects by the recent efforts made, in order to support the management and
prevention of natural hazards and civil crises, e.g. caused by floodings or earthquakes.
1. INTRODUCTION
Aerial and satellite imagery plays a major role for the
population and update of GeoDatabases, and many approaches
have been developed to automate the involved processes to a
large extent. Especially, semi- and fully automatic extraction of
roads and buildings has gained much attention over the last
decades. We will first give a short overview of recent
developments in the field of automatic object extraction with
focus on roads and buildings, before we put special emphasis on
object extraction in the context of change detection and disaster
management. We discuss the issues of object extraction under
different aspects, in particular:
• Level of automation
• Available sensors and image characteristics
• Effectiveness and processing time
• Verification/editing and quality of results
For most applications, none of these aspects can be considered
isolated from the others. We will exemplify interaction and
differing importance of these aspects by the recent efforts made
to support the management and prevention of natural hazards
and civil crises, e.g. caused by floodings or earthquakes. This
application is in particular interesting since the importance of
the above points varies over time before, during, and after such
an event.
2. AUTOMATIC OBJECT EXTRACTION -
ISSUES AND DEVELOPMENTS
The automatic extraction of objects like roads and buildings in
natural environments is one of the challenging issues in photo-
grammetry and computer vision. Natural environments are often
characterized by a high complexity of the imaged scene. Fac
tors having great influence are, for instance, the number of dif
ferent objects, the amount of their interrelations, and the vari
ability of both. Moreover, each factor - and thus the scene com
plexity - is related to a particular scale. To accommodate for
such factors, techniques like detailed modeling, (automatic)
contextual reasoning, and internal evaluation of intermediate
results have proven to be of great importance over the past
years. It is clear that these techniques must be integral parts of
an extraction system to attain reasonably good results over a
variety of scenes.
For the following discussion of recent work, we selected four
important trends seen in object extraction over the last years and
grouped the approaches accordingly: (1) Integration of multiple
views; (2) Integration of functional and temporal characteristics;
and (3) Integration of scale-space behavior.
2.1 Integration of multiple cues
Much recent work deals with integration and simultaneous
processing of multiple cues for object extraction. Typical multi
cue approaches rely on combination of:
• different filters responses (textures, topographic sketch,
• different spectral characteristics
• different sensors (images, laser data, SAR, ...)
• image information and external data (GIS)
A major challenge for processing multi-cue data is the handling
of uncertainties during object extraction. Cues derived from
different sources should support and supplement each other. In
practice, however, numerous conflicting cues appear in case of
reasonably complex scenes. The derivation and utilization of
confidence measures is thus a key issue to provide a basis for
deciding which hypothesis may overrate another and which
should be eliminated.
An interesting earlier approach regarding the role of multiple-
cue processing is given in (Tupin et al., 1999). They deal with
finding consistent interpretations of SAR scenes (Synthetic
Aperture RADAR). Different low level operators are applied to
generate cues for one or more object classes (e.g. road and
river), represented by a confidence value characterizing the
match with the designated object classes. All confidence values
are combined in an evidence-theoretical framework to assign
unique semantics to each primitive attached with a certain
probability. Finally, a feature adjacency graph is constructed in
which global knowledge about interrelations between objects
(road segments form a network, industrial areas are close to
cities,...) is introduced in form of object adjacency probabilities.
Based on the probabilities of objects and their relations the final
scene interpretation is formulated as a graph labeling problem
that is solved by energy minimization.
For building extraction, cues are typically derived from
overlapping images, high resolution surface models (DSMs) and
GIS data. The DSM may stem from aerial laser scanner data, but
also from matching aerial images with high overlap percentage.
A straightforward approach for combining surface and cadastral
data is given in (Durupt & Taillandier, 2006). Although
supporting or conflicting hypotheses are not analyzed in very