be considered for first-level classification. However, classifica-
tion accuracy was not increased significantly by these features
in our experimental areas. In addition, the proposed scheme
was mainly designed for validating
RLPAF
on road extraction,
and classification accuracy on impervious surfaces was not
the main focus. Thus, the most simple classification scheme
was designed for first-level impervious surface extraction.
Second-level rule-based classification extracts high
LBLW
segments within an impervious surface as roads. The proposed
method utilizes spectra and shape features comprehensively
through the two-level classification scheme and exhibits better
performance than
OBIA
that uses a common
LW
feature. How-
ever, the aforementioned procedure cannot extract road regions
with low
LBLW
s. Road borders can be extracted or fitted piece-
wise as straight lines because roads are generally man-made
linear features with even widths and directions. Moreover,
these lines possibly span several regions (called IPS-neighbors
in this study). Straight lines can serve as direction guides for
searching road regions with low
LBLW
s. Thus, a straight line-
guided depth-searching step was designed to extend initial,
broken road regions into a road network. This searching step
involves several region-to-line, line-to-line, and line-to-region
conversions, which are mainly based on
IPSL
-neighborhood re-
lationships. Road regions with low
LBLW
s and sufficient
LBW
s
are appended and formed the final road network. The pseudo-
codes of the road searching step are presented in Algorithm 1.
Experiments
Experimental Data
The proposed technical framework and algorithm was imple-
mented with Visual C++ 2010. The operating system used was
Windows-7 with an Intel(R) Xeon(R) E5620 2.40
GHz
CPU
and
3.48 GB RAM. The methods were applied in different
HSR
im-
ages for validation. Two experimental areas were selected for
method illustration. Table 1 shows the image type, resolution,
size, imaging date, and location of the two test areas.
Experimental Procedure
Commonly used measures, including
Recall
,
Precision
, and
F-
measure
, were employed to evaluate road-extraction accuracy.
Recall
TP
TP FN
Precision
TP
TP FP
F measure
Precision Recal
=
+
=
+
−
= ×
×
2
l
Precision Recall
+
,
(12)
where
TP
is the number of true positives;
FP
is the number
of false positives;
FN
is the number of false negatives; and
F-measure
is a combination of precision and recall, which is
a harmonic mean of the two measures. An ideal road extrac-
tion method should have high precision and recall ratios. In
practice, however,
Precision
and
Recall
measures are gener-
ally in conflict with each other. Thus,
F-measure
was used as
the comprehensive index to evaluate method performance in
our experiments. After segmentation, we first screened out
“real” road segments by visual interpretation and compared
the machine-extracted roads with their visual counterpart
to evaluate method accuracy. In visual interpretation, if a
segment has over 50 percent road pixels, then the segment
is marked as a road. This scheme excludes the influence of
segmentation on evaluating road extraction accuracy.
After image segmentation and straight line extraction,
segment-and-line relationships were built, which involved
three inputs.
T
a
is equal to 15, which indicates that straight
lines shorter than 15 pixels are not considered.
T
b
is equal to
3.0, which indicates that the projected lengths of segments
should be less than thrice their contained and intersected
straight lines.
T
c
is equal to 0.90 for the unilateral and tangent
relationships of segments. These inputs were specified by
sample testing and were used as uniform default inputs in all
experimental analyses.
Figure 4. Impervious surface classification and road extraction.
A
lgorithm
1. P
seudo
-
codes
of
the
D
epth
S
earching
R
oad
N
etwork
A
lgorithm
.
Input: Thresholds of
LBLW
(
T
1
),
LBW
(
T
2
), and straight line
length (
T
3
)
For each region,
P
1
in the impervious surface with an
LBLW
that is larger than
T
1
(with respect to straight line
L
1
) and an
LBW
that is smaller than
T
2
Mark
P
1
as road;
Set all straight lines parallel to
L
1
(including
L
1
) as {
L
},
which are tangent to
P
1
and are longer than
T
3
;
//region-to-
line conversion; use these lines to search neighboring road
regions.
For each straight line
L
2
in {
L
}
Within the impervious surface, get the
IPSL
–neighbors
of
P
1
with respect to
L
2
with
LBW
s less than
T
2
and form
set {
P
};
// line-to-region conversion; search possible road
regions with small
LBLW
s but suitable LBWs.
For each
P
2
in {
P
}
Mark
P
2
as road;
Get all straight lines parallel to
L
2
that are tangent to
P
2
and
longer than
T
3
, and add these lines to {
L
};
//line-to-line con-
version; depth searching along L
1
, L
2
, and succeeding lines
can continue…
End For
End For
End For
T
able
1. E
xperimental
data
Experimental
area
Image
type
Image
resolution
Image size
(pixels)
Data
description
Area 1
ALOS multi-
spectral
image with
four bands
10 m
576
×
493
Collected in
February 2007
in Jiangning,
Nanjing, China
Area 2
China’s GF-1
multi-spec-
tral image
with four
bands
8 m
1013
×
1032
Collected in
2014 in Jiaji-
ang, Nanjing,
China
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2016
153
