road network graph instead of multi-leg intersections (like the
one shown by dotted circle in the cut-out of Figure 12c).
The
RMS
E values in the Table 5 are approximately 1-meter
in all the experiments which are acceptable for the images
used. However, the values of junction
RMS
errors are far from
the desirable values. Since finding the corresponding position
of the junction point in both the reference and the vectorized
result is not expected; the junction geometrical accuracy is
less accurate than that of roads.
Comparison with Other State-of-the-Art Approaches
In this section, the evaluation of the proposed vectorization
approach compared with two different approaches for road
network vectorization is presented. The first approach uses a
Self-Organizing Road Map (
SORM
) algorithm which combines
a K-medians spatial clustering approach with a post-conver-
gence node linking
MST
algorithm (Doucette
et al
., 2001). The
second one uses a combination of a novel Increasing Ellipse
Clustering (
IEC
) methodology and the fuzzy ellipse-shaped
clustering (Mokhtarzade
et al
., 2010).These two approaches
employ the idea of image space clustering for road vectoriza-
tion similar to the proposed methodology.
In the following a pan-sharpened Ikonos image, an ur-
ban aerial image, and a pan-sharpened QuickBird image are
depicted in Figure 14a through 14c, respectively. These data
sets contain roads with different shapes and widths and also
numerous three and four-leg intersections. The outputs of
different road detection methodologies including an object-
oriented classification methodology (Nghi and Mai, 2008), an
object-oriented classification by means of eCognition
®
software
(Mokhtarzade
et al.
, 2010), and an artificial neural networks
(Mokhtarzade
et al
., 2007) were used for Figure 14a through
14c, respectively. The variety of different road detection
methodologies were applied to evaluate the capability of the
proposed algorithm to deal with different road binary images.
The results of applying the proposed vectorization
methodology, the
SORM
algorithm, and the
IEC
and fuzzy
(a)
(b)
(c)
(d)
Figure 12. Results of non-urban Hobart: (a) original image, (b) detected road image, (c) results of the proposed methodology, and (d)
results of MST: Correctly extracted roads are given in dark solid lines, incorrectly extracted roads in light solid lines, and missing roads in
dotted lines.
116
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
