ph gy
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
stochastic process using Gibbs distributions. Roads are found by
maximum a posteriori probability estimation.
Geman and Jedynak (1996) developed another statistical model to
track roads through hypothesis testing. This approach uses the
testing rule that is computed from the empirical joint distributions
of tests (matched filters for short road segments) to determine
whether the hypothesis (road position) is true or not. The tests are
performed sequentially and an uncertainty or entropy
minimization procedure is devised to facilitate testing decisions so
that new tests can be analytically identified. Although this method
works best for coarse resolution images, it is also adaptable to
fine-resolution images (Xiong, 2001). Tupin et al. (1998)
proposed a two-step algorithm for almost unsupervised detection
of linear structures which are treated as road segment candidates.
During the first step, linear features which are treated as road-
segment candidates are extracted. In the second step, genuine
roads among the segment candidates are identified by defining a
Markov random field (MRF) on a set of segments, which
introduces contextual knowledge about the shape of road objects.
The map matching method is adaptable for road network
extraction in case of a large amount of data on road systems in
many parts of the world (Maillard and Cavayas, 1989). It
consisted of two major algorithms. The first algorithm focuses on
image-map matching to identify roads that can be found on the
map and the image. The second algorithm searches new roads
based on the assumption that these new roads are connected to the
old ones. This automatic approach is suitable for revising 1:50 000
scale planimetric data using panchromatic SPOT imagery. Stilla
(1995) developed a syntax-oriented method that uses map
knowledge as a supportive aid for image interpretation. Road
network structures are obtained first through map analysis. Then
image object models are defined and utilized to search for objects
that fulfil model expectations with a given tolerance. Assessment
on image objects with respect to its correspondence to the map
representations results in road object identification for a given
image scene.
Without pre-defined parameters or setting any threshold or to
describe statistically the classes to be extracted in pattern
recognition, a neural network is an alternative way in road
network extraction. The commonly used is back-propagation
network. Fiset and Cavayas (1997) described a map guided
procedure to automatically extract road network using this
network. Bhattacharya and Parui (1997) improved back-
propagation neural network for detection of linear feature from
IRS and SPOT images with a satisfactory result.
Differential shakes, also known as deformable contour models, is
a novel approach for the integration of object extraction and
image-based geospatial change detection (Agourls, 2001). In this
method, the extraction is carried out at three levels. At the low
level of analysis images are characterized and information
extracted on a pixel by pixel basis using only reflectance/emission
measurements (e.g., filtering, edge detection, segmentation, etc.).
At the intermediate level of analysis low level results are
symbolized and fused to form data structures for the high level
analysis. The high level analysis involves not only imagery but
also domain specific knowledge, ancillary sources, and
symbolized data from the intermediate level results to establish
reliable interpretation. The program first runs a template-matching
procedure that localizes potential road pixels, followed by the
optimization to identify potential road segments. To allow a more
inclusive search, a relaxed road model is utilized during this
search. After this search processing, all segments found will be
considered as a road candidate. Then the supervised ISODATA
332
classification procedure is applied to identify whether or not a
candidate is indeed a road.
Also known as rule- or knowledge-based method, the
heuristic method makes use of the human vision system.
Meisels and Mintz (1990) developed a three-stage reasoning
method for the extraction of simple man-made objects from
aerial photography. In the low level, image primitives are
considered as the building blocks of road and identified the
value by checking the neighbors’ value. During the
intermediate level analysis image primitives are combined
with line segments by using the reasoning mechanism. At the
high level of processing gaps are filled and segments grouped
by taking into consideration of distance, brightness and
uniformity among them. This method is flexible when the
problems concern linear feature alignment and fragmentation.
The above review indicates that a large number of studies
have been carried out to detect roads from remote sensing
images. Some of them have produced satisfactory results.
However, they are limited in that they are designed for the
extraction of roads from a particular type of imagery.
Whenever a different kind of data or geographic region is
involved, these methods are utterly incapable of the
extraction.
In this study we propose a new spatial reasoning-based
approach to extraction of road networks in urban areas from 4
m resolution IKONOS imagery. In this environment other
urban built-up areas can be easily mixed with roads. Besides,
urban vegetation can also obscure image properties of roads.
At such a fine resolution level, roads of various widths and
conditions have their unique spectral properties on the
imagery. All of these factors make their extraction
challenging. The challenge is overcome with spatial
reasoning which takes the spatial arrangements and spatial
relationship among road pixels into decision making. This
extraction has been implemented in an environment which is
not subject to the spatial resolution of the input imagery.
2. METHODOLOGY
2.1 Image properties of roads
As a kind of artificial entities, roads differ from other human-
made and natural objects in their surface material and
geometric configuration. Namely, they are usually paved with
asphalt or concrete, causing them to reflect incident solar
radiation strongly. A road is linear in shape and has a uniform
width which varies with its significance. Highways and
artillery roads are wider than inner city streets.
On remote sensing imagery roads are recognizable from their
distinctive tone, texture and shape. Strong reflection of the
incident radiation causes them to have a uniform light tone on
remote sensing imagery. Consisting of the same paving
material, roads have a smooth and identical texture. There is
a drastic change in tone and texture across road boundaries.
Geometrically, roads are elongated and spatially continuous.
They also have a uniform width with small curvatures. Their
physical appearance on an image, however, is affected by
image spatial resolution. On a course resolution image, a road
may not even register or appear to be a narrow line
represented by a string of pixels. On a fine resolution image,
however, the same road may appear as an array of cells with
two parallel boundaries. The spatial continuity of a road
Internai
network
the groi
road ne
2.2 Dat
The ori
covers
New 4
unsupel
groupe:
classifk
to road
image |
format,
image |
class!
This bii
to the «
along
unsupei
quite cc
isolated
elimina
segmen
disjoine
will be
the extr
3.-
The prc
reasoni
classifi
thinnin;
3.1 Noi
Roofs ¢
that of
road p
inspect;
spatiall
