environmental elements make major contributions to this
phenomenon (Portigal
et al.
1997; Zhang
et al.
2006). They
cause significant confusion in image classifications. Efforts
such as weighted
SMA
and spectral transformation are made
by researchers to minimize within-class variability and
maximize between-classes variability in order to acquire more
accurate results.
Although many scholars have applied several spectral
transformed schemes in their applications, there is not a
comprehensive comparison between them. This study com-
pared 26 spectral transformations in three different study
areas. We repeated tests 100 times with different end-member
spectra to reveal the reliability of each scheme. Janesville,
Asheville, and Columbus are far from each other. Residential
areas, commercial areas, soil, trees, and grass are the major
land cover types in each of these areas. These regions can be
viewed as typical urban and suburban environments in the
United States. Therefore, it is meaningful to compare these
three locations with the same end-member model, to test the
reliability of each transformed scheme.
Spectral variability is widely present in remotely sensed
imagery. Thus, research with different backgrounds may
select different end members from images. Moreover,
end-member selection is a key step for successful
SMA
. Dif-
ferent end members may give different results. Therefore,
evaluating a scheme’s performance based on one-time end-
member selection may not be appropriate. Thus, the signifi-
cant-difference test was based on 100 repeated tests, which to
a certain degree includes many potential spectra. In this case,
the general reliability of each transformed scheme can be il-
lustrated from these 100 repeated tests. Moreover, three study
areas were tested in order to further prove each transformed
scheme’s reliability. Thus, the results of the significant-differ-
ence test can be reliable in this study.
Is There a Significant Difference After Applying a Transformed Scheme?
Results from paired-sample
t
tests demonstrated that a signifi-
cant difference of
MAE
existed in many transformed schemes.
On the one hand, some significant differences were positive,
meaning that the
MAEs
were reduced. On the other hand, some
significant differences were negative, meaning that the
MAEs
increased. Only
NSMA
had consistent positive differences in
all three study areas, as the
MAEs
were reduced significantly.
Other transformed schemes, such as
DA1–3
,
ICA
, and
MNF
, il-
lustrated a positive significant difference in one or two study
areas but a negative one in the remaining area(s). The rest of
Table 2. Results of paired-sample
t
test and improved statistic.
Schemes
Paired-sample t test
Number of improved tests
Average improvement (%)
Mean difference
Sig. (two-tailed)
Jane.
Ashe.
Col.
Jane.
Ashe.
Col.
Jane.
Ashe.
Col.
Jane.
Ashe.
Col.
DA1
−0.008
0.015 0.012
0.054
0.011 0.003
38
58
66
28.0
47.3
23.8
DA2
−0.008
0.020
0.002
0.043
0.001
0.618
36
67
53
29.3
54.9
27.2
DA3
−0.009 0.012 −0.003 0.059 0.058 0.429
40
59
46
31.0
58.5
23.1
PCA
0.001
−0.001 −0.006 0.894 0.864 0.208
49
47
41
29.2
53.0
32.2
ICA −0.009 −0.015
0.014
0.031 0.024
0.002
37
37
59
28.0
55.0
29.6
MNF
−0.009
0.007
0.030
0.069 0.219
0.000
45
52
79
28.4
43.2
35.3
TC
−0.001 −0.002 −0.001 0.908 0.763 0.890
52
49
49
28.6
46.5
21.5
BN
−0.028 −0.014 −0.016 0.000 0.050 0.002
25
44
40
26.0
44.6
22.8
CR
−0.087 −0.036
0.030
0.000 0.000
0.000
11
32
87
26.5
43.6
24.1
GHP
−0.069 −0.102 −0.113 0.000 0.000 0.000
11
6
2
21.7
15.6
48.6
GLP
−0.001
0.001 0.002
0.848 0.912 0.753
55
52
55
26.0
21.2
14.5
HP
−0.073 −0.103 −0.112 0.000 0.000 0.000
10
5
1
19.8
11.1
11.7
LP
−0.013 −0.013
0.005
0.061 0.100 0.318
46
46
59
25.6
24.8
13.5
NSMA 0.015 0.022 0.014 0.001 0.000 0.001
67
69
62
32.8
38.9
29.2
Tie1
−0.014
−0.009 0.003
0.081
37
47
23.6
21.6
Tie2
−0.012
0.002
−0.016 0.009 0.774 0.003
35
54
41
27.3
59.3
6.2
Tie3
−0.008
0.005 0.001
0.051 0.381 0.870
37
52
55
25.3
18.5
18.4
Tie4
0.003
0.016
−0.005 0.544
0.002
0.281
49
58
45
30.4
48.5
22.7
Tie5
−0.002
0.011 0.010
0.623 0.063 0.024
46
57
61
33.1
51.8
25.1
Tie6
−0.009
0.012
−0.008 0.054
0.049
0.093
39
54
46
27.6
57.5
24.9
Tie7
−0.009
0.008
−0.010 0.071 0.152 0.023
44
54
38
28.7
51.1
23.2
DWT1 −0.013 0.000
0.002
0.030 0.977 0.682
43
55
56
27.8
46.4
25.2
DWT2 −0.014 −0.010 −0.013 0.024 0.073 0.018
45
43
35
25.9
37.2
30.7
DWT3 −0.014 −0.007 −0.004 0.027 0.272 0.424
43
45
48
29.0
38.5
25.1
DWT4 −0.007 −0.007 −0.004 0.094 0.294 0.527
40
52
50
25.6
37.0
29.1
DWT5 −0.009 −0.008
0.002
0.133 0.232 0.711
43
44
57
30.3
40.0
23.2
Ashe.: Asheville; BN: band normalization; Col.: Columbus; CR: continuum removal; DA: derivative analysis; DWT: discrete
wavelet transform; GHP: Gaussian high-pass;
; ICA: independent component analysis;
Jane.: Janesville; LP: low-pass; MNF: minimu
ectral mixture analysis; PCA: principal
components analysis; Sig.: significance; TC: ta
tion.
Tie spectral transformation in Asheville (La
Tie2 in order to match the Tie2 of
Janesville and Columbus (Landsat 8 Operational Land Imager). Italics indicate nonsignificant better performance compared to
the untransformed scheme; bold indicates significantly better performance.
526
July 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
