Overall.Acc
=
+
+ + +
PT NT
PT PF NT NF
,
(9)
mean IOU
=
+ +
+
+ +
(
)
/
,
PT
PT PF NF
NT
PF NT NF
2 (10)
where
PT
represents a positive true representing the number
of samples whose prediction and ground truth are both built-
up areas and
PF
,
NF
, and
NT
have similar meanings as
PT
according to their prediction and ground truth. User.Acc and
Prod.Acc focus on the built-up area corresponding to the false
alarm rate and the missing alarm rate, respectively. The other
two concern the overall segmentation precision. Overall,
accuracy represents the segmentation precision of the entire
image. The mean
IOU
means the
IOU
averaged across the built-
up area and non–built-up area. These indexes are the most
commonly used in image segmentation tasks, and the first
three are often used in classification tasks as well.
Overall Performance of Our Proposed Built-Up Area Detection Algorithm
We validate the
LMB-CNN
on the testing sample set, which
contains 22 656 samples of built-up area and 109 803 samples
of non–built-up area. As we can see in Table 2, the
LMB-CNN
achieves 99.36% overall accuracy in the test block set. At
the same time, it processes 4588 samples per second on the
forwarding inference. The experimental results show that the
excellent performance of
LMB-CNN
meets the requirements of
speed and accuracy of built-up area extraction.
Effect of the FCN and Postprocessing Steps
We first compare the proposed
LMB-CNN
+
FCN
algorithm with
the direct
LMB-CNN
classification to show the difference be-
tween classification and segmentation. The experiments are
performed on the test images of 10 240×10 240 pixels. One
of the results is shown in Figure 10, in which there are many
false alarms of the classification result, and some built-up
areas have obvious missing alarms within the area. For the
proposed method, there are only a few differences compared
to ground truth.
In order to explain the effect of the postprocessing, we
show the
FCN
segmentation and postprocessing results of the
local area. As Figure 11 demonstrates, wit
some isolated small blocks are removed, a
are also converted to be smooth so that the
more natural. But postprocessing leads to
on accuracy; that is, it improves Overall.Acc from 0.9665 to
0.9690 on a pixel-labeled 2048×2048 data set. The reason is
that the area of each built-up region is generally large, and the
coarse edge will not drop accuracy significantly.
The experimental results of all 10 test images are listed in
Table 3. As shown in the table, the proposed algorithm has a
significant improvement in User.Acc, although the Prod.Acc
has a slight drop, and the Overall.Acc increased to 98.75%.
Overall, the proposed algorithm has an obvious significance.
Effect of the Voting Procedure
There are eight channel outputs from our
FCN
model, and the
final segmentation result is obtained by voting. Our intention
is to achieve a balance between the false alarm and missing
alarm by adjusting the threshold, so we conduct an xperiment
and show the results in Table 4. The single output means there
is only one channel outputting from the
FCN
model, and T = 4
represents the voting threshold for eight channel outputs. The
Table 2. Performance of
LMB-CNN.
Models
Built-Up Area Accuracy Non–Built-Up Area Accuracy Overall Accuracy Speed (fps)
Layers
Model Size (MB)
LMB-CNN
0.9917
0.994
0.9936
4588
9
3.18
Figure 10. Experiment result on the image located in
Shandong. (a) Panchromatic image. (b) Ground truth. (c)
Result of
LMB
-
CNN
classification. (d) Result of the proposed
LMB
-
CNN
+
FCN
.
Figure 11. The effect of the post-processing. (a) Local area
results before postprocessing. (b) Isolated small blocks
removed after postprocessing. (c) Built-up areas with jagged
edge before postprocessing. (d) Boundaries of the built-up
area more smooth after postprocessing.
744
October 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
