6. Automatic
zonal features.
and Remote
etric research.
[, Schmidt D,
of automatic
150.
e-art of digital
Week 1993,
layr W, 1994.
orientation.
and Remote
ela A, 1995.
ration system.
with scene
me 2486:156-
M, Bryan J K,
at, SPOT, and
Programming
orted Pictures
ansactions on
(2).
ral analysis of
lew York. 199
n I, 1994. An
X forests and
International
ote Sensing,
and digital
Week » 1993,
otogrammetry.
irlsruhe:21 -32.
orientation of
al Archives of
174-181.
er J-P, 1988.
ntrol Features
PRS Congress,
iltiple Regions
EEE Trans. on
-7:107-109.
UNCERTAINTY IN GIS SUPPORTED ROAD EXTRACTION
George Vosselman
Faculty of Geodetic Engineering
Delft University of Technology
Thijsseweg 11, NL-2629 JA Delft
The Netherlands
e-mail: g.vosselman@geo.tudelft.nl
Commision Ill, Working Group 3
KEY WORDS: Uncertainty, GIS Updating, Change. Detection, Image Interpretation, Statistics, Modeling
ABSTRACT
Most image interpretation methods suffer from a lack of knowledge about the scene contents. In the case of updating road
databases by interpretation of aerial images, the outdated database is a váluable source of knowledge. Both the old database
and the generic, often heuristic, models of the scene objects contain uncertain data. This paper deals with different ways
of representing this uncertainty and combining data from different uncertain knowledge sources. It is shown how uncertainty
affects the quality of the road extraction and how modeling uncertainty can contribute to a better result.
1 INTRODUCTION
After one or two decades of digital mapping many countries
are completing or already have completed extensive digital
topographic databases, both in small and large scales. The
updating of these databases will therefore be one of the major
topics for many national mapping organisations. Although
the efforts for updating are expected to be significantly lower
compared to those of the initial database creation, there is a
widespread interest in developing tools for (semi-)automatic
mapping in digital images to further reduce the costs.
Contrary to initial expectations, the interpretation of aerial
images for the purpose of mapping roads or houses has shown
to be an extremely difficult task to automate. Simple schemes
of thresholding, edge detection and grouping are clearly in-
sufficient. One has come to realize that a human operator
exploits an enormous amount of knowledge to interpret im-
ages and that an extensive knowledge base will be necessary
to even partially solve tasks like understanding the complex
aerial images.
Several ways are being explored to incorporate more knowl-
edge into the image interpretation. A first one is to use the
knowledge base of a human operator. By using the com-
puter's speed for low level image processing and recognition
of simple patterns and leaving the more difficult interpreta-
tion tasks to the operator, interactive algorithms are a very
attractive way to speed up the mapping process. Further im-
provement should be possible if one can provide computer
algorithms with detailed specific and generic models of the
objects and their interrelationships that are encountered in
aerial images. How to represent this knowledge is one of the
major research issues. A third important source of knowledge
is, of course, the database with the objects that have been
mapped at a previous occasion and now need to be updated.
The outdated databases not only outline many objects that
only need to be verified instead of detected and measured,
but also supply the context within which new objects may be
found.
Since the modeling of the scene contents is an extremely dif-
ficult task, it is clear that a considerable degree of automa-
tion in mapping can only be achieved by combining most
of the knowledge available. As most of the future map-
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
ping projects will be concerned with updating, the outdated
databases should be considered a very important source of
knowledge.
All sources of knowledge will, however, contain errors. These
errors will influence the interpretation process. In order to
assess the quality of the image interpretation results we will
therefore need to describe the quality of the information for
each individual source. Furthermore it is necessary to describe
all steps in the interpretation process such that the quality of
the different sources can be propagated to the final result.
This paper is about all kinds of uncertainties that are en-
countered during the updating of a topographical database
by interpreting digital aerial images. In particular, we will
focus on the extraction of roads using the context of an out-
dated road database.
Representation of uncertainty in data and processing of un-
certain data are relevant to a wide range of disciplines. Espe-
cially in the field of artificial intelligence, many research efforts
have been devoted to these topics [Kanal et al., 1986]. More
recently, user requirements and standardisation efforts have
also lead to better descriptions for uncertainty in (or quality
of) geographical information [Guptill and Morrison, 1995].
After briefly describing the data and processing steps needed
for the updating of road maps (section 2), we therefore first
give an overview of representations of uncertainty as they
have been developed in different disciplines (section 3). In
section 4 some examples are given of how these uncertainty
descriptions relate to the data used for road map updating.
The problem of image interpretation is a reasoning problem
in which many sources of evidence need to be combined with
rules and heuristics in order to generate the most likely ex-
planation of the image. In section 5 methods for processing
these uncertain data are reviewed. This also includes the
propagation of uncertainty towards the final interpretation
result. Section 6 then describes how these methods are being
applied or could have been applied to the extraction of roads
from aerial imagery. The last section summarizes the findings
and outlines the work that remains to be done to assess the
uncertainty in updated road databases.
909
