false alarms are acceptable for hazard detection, while omit-
ted landslides may trigger unexpected loss.
The semi-automatic method omits numerous landslides
in Figure 8a2 through 8n2. One of the main reasons is that
the semi-automatic method is quite sensitive to the lighting
condition or the atmospheric environment when images are
taken, as the cases shown in Figure 4c1, 4d1, and 4f1. The
other possible reason is that the landslide occurred in our
study is more complicated than the experimental data (Li
et
al.
, 2016). The landslides in Li et al. (2016) are much clearer
and bigger, and the background is simpler, rather than the
cases in our study with various rocks, bare soil, roads and
clouds being background.
An objective comparison in terms of landslide detection
accuracy is conducted to further analyze the performance of
our method. Given the ground truth images, we calculated
the Producer’s Accuracy and User’s Accuracy in each study
region for both methods and listed in Table 1.
Generally, our method gets higher accuracy than semi-
automatic methods in terms of both most User’s Accuracy
and Producer’s Accuracy. The semi-automatic method gets
extreme low Producer’s Accuracy and User’s Accuracy in
several cases for research area One, such as study regions A,
C, D, and F (marked as green). It indicates that semi-automatic
method omits and commits many landslides in these cases.
Compared with semi-automatic method, the proposed method
gets more reasonable Producer’s Accuracy and User’s Ac-
curacy, but for the case of study region D and E, where rocks
and roads make up massive background objects, a majority of
rocks and roads are miss-recognized as landslides. Concern-
ing the second research area, i.e. study regions I, J, M, and
N, the Producer’s Accuracy of the proposed method is still
much higher than the semi-automatic method, but the User’s
(i)
(j)
(m)
(n)
Figure 7.
Continued from last page.
Table 1. Comparison of Accuracy Performances (%) of Our
Method and Semi-Automatic Method (Li
et al
., 2016)
Image
Method
Producer’s Accuracy User’s Accuracy
A
Our method
70.77
63.75
Semi-automatic
9.43
29.31
B
Our method
35.96
64.89
Semi-automatic
41.22
51.87
C
Our method
73.06
53.89
Semi-automatic
3.40
3.91
D
Our method
79.93
20.97
Semi-automatic
1.69
0.94
E
Our method
56.71
36.04
Semi-automatic
35.89
27.84
F
Our method
58.32
61.88
Semi-automatic
10.03
9.23
I
Our method
80.794
41.315
Semi-automatic
49.361
59.195
J
Our method
80.775
51.193
Semi-automatic
39.151
63.343
M
Our method
81.960
69.394
Semi-automatic
50.717
68.634
N
Our method
71.773
83.926
Semi-automatic
69.885
61.781
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2017
359
