Istanbul 2004
yentually, the
m scale space
the Deutsche
n von Straßen
rtationen, No.
System zur
rmationen aus
des TNT der
uttgart, 162 p.
dge Detection.
and Machine
to derive a
. International
., and Tonjes,
vledge based
‘borne Remote
13-320:
nputer Vision.
23 p.
Detection with
| of Computer
mert, F., 1990.
understanding.
md Machine
en System zur
undungsdaten.
pver, Band 3
retation von
undungsdaten.
, 104 p.
er und multi-
itationsschrift,
Schwerpunkt
f Curvilinear
Analysis and
ion und 3D-
Luftbildern.
Berichte VDI,
.
and, A. (Ed.):
poration, New
AUTOMATIC EXTRACTION OF ROAD NETWORKS IN URBAN AREAS
FROM IKONOS IMAGERY BASED ON SPATIAL REASONING
J. Gao and L. Wu
School of Geography and Environmental Science, University of Auckland, Auckland, New Zealand
jg.gao@auckland.ac.nz
Commission III, Working Group 111/4
KEY WORDS: Remote Sensing; Mapping; Extraction; IKONOS; Automation; Urban
ABSTRACT:
In this study we developed a spatial reasoning-based method of automatically extracting roads in a densely populated suburb of
Auckland, New Zealand from IKONOS data. First, all of the four multispectral bands were grouped into 20 clusters in an
unsupervised classification, two of which corresponded to road networks. This intermediate result was then converted into a binary
image of road and non-road pixels. This binary image was then further processed with spatial reasoning in two ways. First, all
isolated or small clusters of pixels were examined spatially to determine if there were other isolated pixels in their immediate vicinity.
If no neighbouring pixels were found, they were considered as noise and removed from the image. If neighbouring pixels were
found, their position in relation to the pixel under consideration was further analyzed. If they were aligned with existing pixels along
a certain orientation, then they were regarded as a portion of a disjoined road and retained in the output image. Seconds, these
disjoined road segments were later joined together to form a road network. The extracted road network was unified to a constant
width because trees planted along both sides of a road caused its width to vary in different sections. The detected results using a
threshold of six pixels show that most roads can be extracted at a reasonable accuracy level.
1. INTRODUCTION
Roads in dynamic cities tend to change very frequently even
within a short period of time. Road maps of these areas have to be
updated periodically, preferably from current satellite images to
meet the urgent need of urban planners. With the advances in
remote sensing, more and more high quality and fine spatial
resolution satellite images have become available from different
platforms. For instance, the recently emerged IKONOS satellite
imagery has a spatial resolution of 4 m in the multispectral mode
and of 1 m in the panchromatic mode. These images enable the
extraction of even minor streets in urban areas. They have raised a
renewed possibility of timely and efficiently updating changed
road networks in urban areas.
Extraction of road networks from remote sensing images can be
accomplished either manually or automatically. Manual extraction
is subject to the analyst’s experience and skills. Roads can be
recognized reasonably well even from noisy images that contain
incomplete information about roads if s/he is familiar with the
study area. However, this manual method is expensive and time-
consuming. By comparison, automatic extraction of road network
information involves significantly less time and expense, even
though it is more complex methodologically.
Automatic extraction of roads from satellite images faces several
challenges because the image appearance of roads depends upon
the spatial resolution of the satellite images. In addition, the
extraction is hampered by noise on satellite images. Ground
objects such as trees along a street can obstruct the image of roads.
Vehicles on the road may cover certain parts of a road and make it
difficult to detect on the image.
So far various automatic methods have been developed to extract
roads from satellite images. These methods fall into five broad
categories: ridge finding, heuristic reasoning, dynamic
programming (DP), statistical tracking, and map matching (Xiong,
331
2001). Ridge finding is a classic method in which an input
image is edge-filtered to obtain the magnitude and direction
of linear features, including roads (Nevatia and Babu, 1980).
Wang et al. (1992) developed a way of detecting ridges from
SPOT data. In this gradient direction profile analysis method
four predefined directions for each pixel are calculated first
and the gradient direction for a pixel is the direction of the
maximum slope among the four defined directions around the
pixel. The road segments have the same ridge direction and it
is perpendicular to the gradient directions of the pixels with
the bridge. Analysis of the gradient profile will generate the
ridge pixels. The road network or segment can be obtained by
linking all the ridge points. Steger (1996) introduced
differential geometry to ridge finding. This method uses
curve or surface fitting techniques to locate ridges on remote
sensing imagery. If the image intensity surface is represented
by a mathematical equation, the first and second derivatives
of the equation can be analyzed to locate edges.
In DP, roads are modelled as a set of mathematical equations.
The derivatives of the grey values perpendicular to the
direction normal to the road tend to be maximized, while
derivatives along the road direction are minimized. Roads
appear to be straight lines or smooth curves. Their local
curvature has an upper bound. DP is advantageous in finding
curves in noisy pictures, for it can bridge weakly connected
feature elements automatically while the program searches
for optimal solutions (Gruen and Li, 1995).
Statistical inference models are particularly suitable for
detecting roads with complexity and uncertainty (e.g. bridges,
road width variation, vehicles and shadows on the roads and
image noises, etc). Barzohar and Cooper (1996) explored the
method further and developed a stochastic approach that can
be applied to automatic extraction of highly sophisticated
roads. A geometric-stochastic model formulates road width,
direction, grey level intensity and background intensity as a
