results show that false alarm is twice the rate of missing alarm
when
FCN
has a single output. When the voting threshold is
set to 4, false alarm and missing alarm are basically the same.
As the voting threshold continues to increase, the false alarm
will decrease further. In addition, the time consumption of
this operation is negligible because the computational con-
sumption of voting takes only 0.2% of the whole process.
Comparison with Existing Methods
Comparison with Existing Unsupervised Algorith
We compare the proposed algorithm w
pervised built-up area extraction algori
CGEO
(Li
et al.
2015), Harris corner-bas
Szirányi 2013), anisotropic rotation-invariant textural measure
(PanTex) (Pesaresi
et al.
2009), and
MBI
(Huang and Zhang
2012). The
MBI
is used primarily to extract buildings, so we
add a Gaussian smooth and use it to extract built-up areas,
where the scale parameter
s
= 2 +
i
× 5(
i
= 0–5). The PanTex,
MHEC
, and
MBI
are threshold-based algorithms, and
CGEO
is
graph based, which utilizes superpixels to reduce complex-
ity. In addition, the results of all the algorithms are removed
in small areas below 2500 pixels according to our actual test
remote sensing image resolution. Furthermore, we also dem-
onstrate the results of our model with single output, meaning
only one channel outputting from the
FCN
model. This model
is denoted as Our-single in Figure 12 and Tables 5 and 6.
We test the five approaches in test images with a size of
2048×2048 considering the time consumption. As shown
in Figure 12, the graph-based
CGEO
and proposed algorithm
have the best performance, but
CGEO
obviously has many false
alarms and does not work on some images. The other algo-
rithms basically do not work.
We calculate the evaluation indexes and show them in
Table 5. In addition, the elements of our statistics are pix-
els, although the essence of our algorithm is based on image
blocks. Considering the User.Acc, our algorithm performs best
in half the images; the next is
MHEC
. However,
MHEC
does not
work in some images, while ours is stable at approximately
80%. On the other hand, the proposed algorithm behaves best
on Prod.Acc, and
CGEO
is second. That means that
MHEC
and
CGEO
can do well in only one way, but the proposed algorithm
does well in both. In addition, the poor performance of our
algorithm on User.Acc is due to the lack of subtle margin re-
fine procedures. For Overall.Acc and Mean
IOU
, our algorithm
achieves the best performance on most images. Furthermore,
our two models (i.e., the model with single output and the
trategy) have very similar performance on
pplication of built-up area extraction,
algorithm speed is one of the most important aspects. We
calculated the average processing time of each algorithm on
an image with 2048×2048 pixels and show the times in Table
6. As can be seen, even the fastest unsupervised contrast al-
gorithms require more than 60 times as much processing time
as ours. The proposed algorithm has important practical value
considering the speed and precision.
Comparison with Existing Supervised Algorithms
In previous experiments, our algorithm was compared with
several unsupervised methods. However, our models are
obtained through supervised learning. Therefore, we compare
the proposed algorithm with several supervised built-up ex-
traction algorithms, including
MKL
+
GC
(Tao
et al.
2012),
MMM
(Li
et al.
2017a) ,
VW
(Weizman and Goldberger 2009), Deep-
lab v3+(Chen
et al.
2018), and
LMB-CNN
+
GC
(Tan
et al.
2019).
Here the feature extractor of Deeplab v3+ is
ResNet
-101. To
highlight effectiveness of our architecture, we implement the
feature extractor of our method as
ResNet
-101 which is same as
Deeplab v3+, and this entire model is denoted as
ResNet
+
FCN
.
All six compared models are trained by the same sample
data set as our
LMB-CNN
+
FCN
model. We test them on the test
data set just as before. The binary results are illustrated in
Figure 13/
LMB-CNN
+
GC
,
ResNet
+
FCN
, and our
LMB-CNN
+
FCN
are
Table 3. Results of
LMB-CNN
classification and proposed algorithm.
No.
Overall.Acc
User.Acc
Prod.Acc
LMB-CNN
Proposed
LMB-CNN
Proposed
LMB-CNN
Proposed
1
0.9729
0.9843
0.8035
0.9245
0.9311
0.9012
2
0.9736
0.9856
0.8158
0.9345
0.8949
0.8954
3
0.9914
0.9959
0.7984
0.9235
0.9794
0.9515
4
0.9693
0.9881
0.7525
0.9072
0.9808
0.9660
5
0.9827
0.9910
0.9258
0.9829
0.9855
0.9684
6
0.9868
0.9923
0.9338
0.9676
0.943
0.9606
7
0.9776
0.9795
0.9457
0.9454
0.9485
0.9586
8
0.9845
0.9900
0.7866
0.8949
0.9201
0.8927
9
0.9489
0.9796
0.7899
0.9288
0.9575
0.9546
10
0.9771
0.9883
0.8662
0.9508
0.9435
0.9459
All
0.9765
0.9875
0.8569
0.9431
0.9517
0.9468
Table 4. Results of different voting thresholds.
No.
False Alarm
Missing Alarm
Time(s)
Single Output
T = 4
T = 7
Single Output
T = 4
T = 7
Single Output
T = 4
T = 7
1
0.1224
0.0755
0.0643
0.0706
0.0988
0.1100
8.78
8.78
8.80
2
0.1097
0.0655
0.0558
0.0878
0.1046
0.1228
8.75
8.77
8.70
3
0.0887
0.0765
0.0635
0.0291
0.0485
0.0533
8.71
8.71
8.74
4
0.1450
0.0928
0.0768
0.0270
0.0340
0.0427
8.71
8.79
8.76
5
0.0264
0.0171
0.0140
0.0219
0.0316
0.0358
8.73
8.73
8.78
6
0.0653
0.0324
0.0246
0.0329
0.0394
0.0449
8.71
8.75
8.81
7
0.0696
0.0546
0.0463
0.0345
0.0414
0.0489
8.78
8.80
8.76
8
0.1509
0.1051
0.0918
0.0824
0.1073
0.1190
8.77
8.74
8.73
9
0.1064
0.0712
0.0608
0.0302
0.0454
0.0520
8.73
8.77
8.80
10
0.0700
0.0492
0.0425
0.0482
0.0541
0.0617
8.77
8.71
8.80
All
0.0860
0.0569
0.0479
0.0412
0.0532
0.0612
87.44
87.55
87.68
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2019
745
