We designed the following experimental procedure to vali-
date the proposed method on road extraction. An impervious
surface was first extracted from
HSR
images by first-level super-
vised classification using region spectral features. Then, three
road extraction schema, i.e., using common
LW
measure (the
LW method),
LBLW
measure (the
LBLW
method), and
LBLW
mea-
sure combined with depth searching (the
LBLW
&DS method),
were applied on the impervious surface to select road regions
with
LW
s or
LBLW
s that exceeded a threshold (
T
). For all three
methods, a high threshold
T
excluded many false roads (i.e.,
Precision
measure increases) than a low
T
, whereas the chance
of omitting real roads increases (i.e.,
Recall
measure decreases).
Thus, threshold
T
was adjusted from low to high, and the
pre-
cision-recall
and
F-measure
curves of the three methods were
drawn to evaluate their robustness during threshold changes.
In
HBC-SEG
method, regions are generally merged to their ex-
treme sizes (the edge position) under small scales. The small-
est default scale of 10 was used for all image segmentations in
our experiments. Regions were then classified using an
SVM
classifier with Gaussian radial basis function (RBF) kernels.
The Gaussian RBF kernels had a penalty factor
C
of 25 and a
kernel width
σ
of 40. The region spectral mean values of all
image bands were used in classification. A minimum amount
of training samples were selected for
SVM
classification if the
impervious surfaces were extracted totally. The classification
results were not edited and possible errors were left. Second-
level road extraction was conducted directly on the extracted
impervious surfaces for an objective method evaluation.
The inputs for second-level road extraction were as fol-
lows: threshold
T
ranging from 2.5 to 4.5, with a step of 0.5.
For the
LBLW
method, a region was extracted as roads if its
LW
or
LBLW
was larger than
T
. Thus, the result obtained by the
LBLW
method was a superset of the results obtained by the LW
method. The results of the
LBLW
&DS method, which inte-
grated a depth-searching step into the
LBLW
method, formed a
superset of the
LBLW
results. The road width threshold
T
2
was
16 pixels. All inputs were fixed in all experiments.
For each of the experimental areas, training samples were
selected thrice, the corresponding first- and second-level clas-
sifications were performed, average accuracy measures were
obtained, and accuracy curves were drawn. The regions ob-
tained in image segmentation could have different sizes, and
thus, method accuracy was evaluated in two ways, namely,
by counting the road region number and by summing up the
road region areas, for a comprehensive comparison.
Experimental Analyses
The quantitative results of road extraction are prsesented in
Table 2 and Figure 5. Accuracy in the two experimental areas
was evaluated by region number and area, and the results
exhibited slight differences, which led to the same conclusion.
Given a threshold
T
, the recall ratios of the three methods
were
LBLW
&DS >
LBLW
> LW, whereas their precision measures
were reversed as LW >
LBLW
>
LBLW
&DS. These results were
reasonable because the region sets obtained by the three meth-
ods were LW
Í
LBLW
Í
LBLW
&DS. For the
LBLW
and
LBLW
&DS
methods, the risk of including false road regions increased as
the number of included true road regions rose. Errors resulted
because some false roads might have large
LBLW
s. In addition,
the depth-searching step of the
LBLW
&DS method was mainly
based on road ductility along several straight lines, which un-
avoidably included some false road regions with small
LBLW
s
that were attached to these straight lines. These errors were
more apparent in urban areas than in rural areas. However, the
LBLW
and
LBLW
&DS methods obviously performed better than
the LW method on the
F-measure,
given the same threshold
T.
The compensation on
recall ratio
far exceeded that on
preci-
sion
loss when the
LBLW
and
LBLW
&DS methods were com-
pared with the LW method. In addition, the
LBLW
&DS method
had flat
F-measure
curves in all the experiments. Along with
the increase in
T
, an increasing number of road regions were
abandoned by the
LBLW
and LW methods, which caused the
F-
measure
curves to decline. However, the
LBLW
method restored
a significant number of regions through the depth-searching
step, and thus, maintained high
F-measure
accuracy. In this
scenario, the depth-searching process reduced algorithm pa-
rameter dependency and increased method robustness.
The
precision-recall
curves also verified the aforemen-
tioned conclusions. Given the same threshold
T,
the
LBLW
&DS
method exhibited the lowest
precision
and the highest
recall
T
able
2. A
ccuracy
of
R
oad
E
xtraction
Areas Evaluation
schemes
T
LW Method
LBLW Method
LBLW&DS Method
Precision Recall
F-Measure Precision Recall
F-Measure Precision Recall
F-Measure
Area 1
Region
number
2.5 0.587
0.442
0.504
0.576
0.624
0.599
0.542
0.696
0.609
3.0 0.665
0.367
0.473
0.655
0.555
0.601
0.599
0.648
0.623
3.5 0.739
0.304
0.431
0.718
0.472
0.569
0.643
0.591
0.616
4.0 0.768
0.218
0.340
0.760
0.388
0.514
0.662
0.555
0.604
4.5 0.831
0.161
0.270
0.833
0.313
0.456
0.690
0.519
0.593
Region
area
2.5 0.618
0.416
0.497
0.605
0.629
0.617
0.569
0.699
0.627
3.0 0.692
0.350
0.465
0.676
0.565
0.616
0.625
0.652
0.638
3.5 0.752
0.295
0.424
0.736
0.490
0.588
0.668
0.600
0.632
4.0 0.763
0.222
0.344
0.765
0.417
0.540
0.684
0.571
0.622
4.5 0.821
0.173
0.286
0.835
0.348
0.492
0.712
0.534
0.610
Area 2
Region
number
2.5 0.536
0.425
0.474
0.510
0.542
0.525
0.495
0.671
0.569
3.0 0.597
0.329
0.424
0.562
0.425
0.484
0.539
0.614
0.574
3.5 0.643
0.255
0.365
0.617
0.342
0.440
0.565
0.566
0.565
4.0 0.653
0.183
0.286
0.647
0.284
0.395
0.581
0.549
0.565
4.5 0.694
0.134
0.224
0.674
0.214
0.325
0.594
0.512
0.550
Region
area
2.5 0.528
0.426
0.472
0.497
0.561
0.527
0.477
0.681
0.561
3.0 0.589
0.337
0.429
0.553
0.450
0.496
0.521
0.623
0.567
3.5 0.637
0.264
0.374
0.605
0.374
0.462
0.549
0.577
0.563
4.0 0.659
0.198
0.305
0.638
0.319
0.425
0.569
0.564
0.566
4.5 0.697
0.151
0.248
0.670
0.247
0.362
0.584
0.529
0.555
154
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
