The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B7. Beiji j 2 QOS
Figure 7. Causal RX (a), ACAD (b),
Figure (5) shows the visual results of global weighted RX-base local CFT (c), global CFT (d)
methods.
Mob*I l*X ‘■>‘>0* W**<B*4 «Wft?
Figure 5. Weighted RX (a), weighted RX-UTD (b),
Weighted modified RX (c), weighted normalized RX (d)
Figure (6) shows the visual results of local RX-base algorithms.
Figure 6. Local RX (a), local modified RX (b),
local normalized RX (c), local RX-UTD (d)
For comparing all implemented algorithms together, their area
under ROC curves, number of correct detections (from 70 target
pixels) and false alarm detections in 99% confidence level
displayed in table (1).
Method
AUC
Correct
Detected
False Alarm
1
Global NRX
0.6429
11
201
2
Global UTD
0.8237
29
189
3
Local NRX
0.8981
47
143
4
Causal RX
0.9248
39
141
5
Global RX-UTD
0.9413
46
133
6
ACAD
0.9464
54
33
7
Local DWEST
0.9569
59
106
8
Local MRX
0.9625
55
99
9
Global MRX
0.9641
49
132
10
Local CFT
0.9666
58
52
11
Global RX
0.9686
53
108
12
Local UTD
0.9698
59
113
13
Global DWEST-RX
0.9706
60
14
14
Global DWEST
0.9719
63
110
15
NSWTD
0.9731
64
89
16
Local RX
0.9764
59
67
17
Local RX-UTD
0.9801
64
86
18
Global CFT
0.9823
60
41
19
Global Weighted MRX
0.9861
60
89
20
Global Weighted UTD
0.9934
60
81
21
Global Weighted RX
0.9944
61
71
22
Global Weighted RX UTD
0.9946
61
68
Table 1. Comparative result of anomaly detection algorithms
Moreover, all of these algorithms are compared with
computational complexity point of view and they are tested by
simulated hyperspectral data with various additive Gaussian
noises (20:1, 10:1 and 5:1 signal-to-noise ratio) to investigate
noise sensitivity of them. For example figure (8) shows the
307