nen
SE
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istánbul 2004
ridge on the pixel is calculated. Then the second derivative of a
pixel that lies on the extended direction of the ridge is
examined. If the absolute second derivative on the pixel is more
than lth, the pixel is considered as a part of the ridge. This
process will continue until there is no candidate pixel for seed
point.
2.3 Line Segment Classification Stage
2.3.1 Straight Line Extraction
Road segments extracted in the previous stage are not straight
line but curvature or combination of lines (poly lines) because
the employed method only consider local angle relationship of
cach ridge pixels and does not consider global shape of
segments. As the method follow pixels as long as possible,
resulting segments go through two different roads or even
bridge true roads and false linear object. As these mingled
segments make climination of false segments and construction
of road topology difficult, curved segments must be
decomposed to array of straight lines.
Another method is so called Douglas-Peucker (DP)
approximation algorithm. This algorithm divides a segment at a
vertex where distance from the segment becomes farthest. This
simple process is recursively repeated to resulting segments
until distance of all vertexes is less than a tolerance.
DP method is rather simple graph based approach. Initial
approximation of a curve is the straight line that connects two
end points of the curve. Then a furthest vertex from the straight
line is picked up. If distance of the vertex from the straight line
is more than a tolerance, this approximated poly-line is
subdivided at the vertex into two shorter poly-lines. This
process is repeated until every vertex on the curve is within a
certain distance from resulting poly-lines.
2.3.2 False Line Elimination
Result of line detection is a blend of true road segments and
false road segments such as buildings or elongated vegetated
area. As these false segments often prevent correct road
network construction, Classification of line segments is a
necessary step.
where r is a positive weighting constant, — k is the number of
class, Dis(k) is the discrepancy between after-threshold and
original images. Then k and threshold value is determined to
minimize C(k).
For multi-spectral image, both a supervised and a unsupervised
method are considered. A supervised method gives a better
result than unsupervised ones if proper training data set is
provided. An Unsupervised method does not need training data
set but an operator must give the appropriate class number and
meaningful category to a classified result. A supervised method
is adequate when objects included in the image are known. An
unsupervised method, to the contrary, is adequate when.
included objects are not known or property of each object is
dispersed.
In this study, classified image is extracted using an
unsupervised classification method offered by a remote sensing
software, ERDAS.
2.4 Line Segment Grouping Stage
As most road segments extracted by the centre line detector are
not directly connected and include some false road segments,
grouping and linking of them is also needed. 1 have only found
few articles discussing this issue. I have uses two methods
proposed by Crevier (1999).
In the method, each pair is first tested by rule based screening.
There are three tests for geometric relation and one for
photometric relation. The angular difference between each line
and the line that bridges the gap is the first inspection. The
second one is the transverse gap. It measures lateral offset
between lines. The third one is the longitudinal gap. It
represents “net gap” without lateral offset. If any of these
values exceed predetermined thresholds, this pair is rejected.
The last one is the contrast difference. It checks contrast
between segments and backgrounds. If a line is whiter than
background whereas another line is darker than background,
this pair is rejected.
Each pair that passes rule based screening test is connected and
Operation of ALOS is not clear but PRISM and AVNIR- c Seed Link Num. Correct- | Complete-
2 may not work simultaneously due to bandwidth threshold | threshold | of Line ness (96) ness (96)
limitation. If both images cannot be obtained 0.60 10.0 1.0 1922 10.9 69.1
simultaneously, DEM must be created from images to 0.80 10.0 1.0 1450 10.6 54.7
project both image onto proper geographic position. 1.00 10.0 1.0 1174 TES 582
However, there are several difficulties in DEM creation. 120 10.0 1.0 998 11.8 473
It is time-consuming work and DEM accuracy is 1.40 8.0 0.8 952 12.8 56.1
insufficient in urban area that causes improper 1.60 8.0 08 1058 12.6 58.4
registration of imagery. For that rcason, grey value image 1.80 5.0 0.8 984 12.9 58.6
classification is worth Lo consider. : 2.00 5.0 0.5 825 139 52.9
Multilevel threshold is expected to filter out high 25 50 0.5 848 142 54.9
brightness objects such as buildings & vehicles and low 2.40 5.0 0.5 636 TE 484
brightness objects such as trees and shadows. Removal of 2.60 50 05 711 51 525
these objects will reduce false extraction of road T : = : T = 13
segments and improve correctness. This process employs 2 3.0 0.5 333 143 ES
automatic multilevel threshold method (ATC) proposed 3.0 2.0 0.5 620 15.2 A6 3 sd
by Yen et al (1995). Most feature of the method is that
the number of class can be determined automatically
because the cost function include the number of class
whereas user must specify optimal number from experience.
In the method, Cost function of this method C(k) is defined as
C(k)- ptDis())* (log, (E)) ()
Table 1. Result of centre line detector.
chains of line segments are obtained. Then the validity of each
chain should be examined. The measure of junction's deviation
(d) of each line pair is calculated:
Internati
He
3.1 Lin
Accordit
width is
Table 1
Seeding
results s
increase:
suggests
and resu
Complet
It has a
Gaussiar
width of
pixels (=
scene ar:
result in
extractio
Table 2
Gaussiar
threshok
slightly
threshok
reliable |
32 Fal
Table 3
Gaussiat
threshol
ATC se
cach cla
30 and 8
The coli
containe
GIS soft
means tl
The col
many “(
by the r
buffer is
Th
Thr
a. €
BJ
se
€. Coi
d. Coi
road
Co
Con
Table 3.
