obviously superior to others. As can be seen, the
MMM
per-
formance is next only to them, and there are obviously many
false alarms in
MKL
+
GC
. VM does not work at all, and the
VW
treats pixels as the processing unit. Deeplab v3+, which is a
pixelwise segmentation algorithm, performs unstable in built-
up area extraction since there are many different materials. In
addition, it needs huge memory so that it cannot process the
images with 10 240×10 240 pixels in our computer platform.
ResNet
+
FCN
significantly outperforms DeepLab v3+ and is
slightly inferior to our
LMB-CNN
+
FCN
.
We calculate the four evaluation indexes and show them in
Table 7. Our algorithm performs stably in every testing image
on all metrics. The
MKL
+
GC
behaves best in Prod.Acc, but it
does worse in User.Acc, as can be seen in Figure 13. Although
the test accuracy of
LMB-CNN
-
GC
is similar to that of our
algorithm, the test accuracy of our algorithm will be improved
by increasing the training set, which may be better than that
of
LMB-CNN
-
GC
. In addition, our algorithm is faster than
LMB-
CNN
-
GC
, and the time is improved by 25%.
MMM
’s perfor-
mance is also acceptable, but the time consumption of
MMM
is
more than 200 times as much as ours, as shown in Table 8.
The superiority of
ResNet
+
FCN
relative to Deeplab v3+ dem-
onstrates that it is our two-step learning process and
FCN
ar-
chitecture rather than a better backbone that lead to accuracy
improvement. Although Deeplab v3+ achieves state of the art
in several existing common image data sets, it performs unsta-
bly in built-up area extraction of remote sensing images. The
reasons are twofold. First, it is hard to train Deeplab v3+ with
the whole remote sensing image because of its high memory
consumption, so we have to divide each training image into
Figure 12. Results of comparing unsupervised algorithms and ours, each row corresponding to a 2048×2048 pixel test image.
746
October 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
