IEEE Geoscience and Remote Sensing Magazine - March 2019 - 21

march 2019

92.68

0.9181
0.9056
0.9033
na
0.8985
0.891
0.903
0.8592
76.46
0.834

92.45

-

-

overall accuracy (%)

Kappa coefficient K
(no unit)

0.7563

91.3
91.75
92.45
91.11
89.93
91.28
86.98
78.35
84.69

100

-
average accuracy (%)

77.47

91.95

99.15
89.64

91.82
na

98.14
98.73

90.65
91.02

100
99.15

91.24
88.54

99.78
52.64

77.19
86.34

97.46
187/473
running track

80.34

181/247
tennis court

97.67

74.74

100
100
99.19
98.58
99.6
100
100
100
83
100

100

184/285
Parking lot 2

99.6

76.49
87.02
89.92
76.84
80.35
65.26
76.14
75.09
70.53

98.86

192/1,041
Parking lot 1

61.4

91.07
92.7
90.74
92.32
98.08
99.14
88.56
60.23
70.99

68.53

181/1,054
railway

49.38

90.7
86.34
90.56
98.58
90.23
94.46
90.89
85.01
83.87

97.07

191/1,036
Highway

76.09

63.42
81.18
94.66
86.2
68.82
60.91
67.08
55.21
83.2

93.54

193/1,059
road

57.63

98.21
89.71
86.87
91.78
87.72
90.46
93.95
72.11
82.53

86.01

191/1,053
commercial

69.78

95.73

90.39
87.87

95.25
92.51

91.11
87.59

84.14
86.61

78.54
90.95

90.98
85.28

73.41
86.19

65.81
45.87

89.09
196/1,072
residential

40.65

182/143
Water

82.09

99.43

95.8
98.6
95.8
100
91.61
95.8
95.1
95.80
79.02
95.1

96.02

186/1,056
soil

97.2

99.72
99.62
93.15
99.05
100
100
98.76
97.25
98.11

100

188/1,056
tree

96.4

96.12
99.34
98.04
95.83
97.92
98.96
95.45
94.98
98.77

99.62

192/505
Grass, synthetic

97.54

98.61
100
100
100
100
100
100
70.69
99.8

80.63

Grass, stressed

98.02

99.44

81.58
78.73

94.92
98.59

98.36
82.91

81.48
99.91

81.39
82.53

98.68
84.96

78.06
82.24

98.31
96.43

83.48

190/1,064

Grass, healthy

98.4

198/1,053

cLaSS NamE

83.38

OTVca
(%) [138]
SLrca (%)
[143]
cNNGBFF
(%) [142]
aLWmJ-KSrc
(%) [141]

mULTISENSOr FUSION

mLrsub
(%) [140]
FFcK
(%) [139]
GBFF
(%) [136]
EPhSI+LIDar
(%)
cNNhSI
(%)
SVmhSI
(%)
rFhSI
(%)

SPEcTraL

ieee Geoscience and remote sensing magazine

TraINING/
TEST SamPLES

LOW-RANK MODELS
To avoid the curse of dimensionality and also
increase the efficiency of the analysis compared
to filtering approaches, low-rank models were

TaBLE 2. ThE cLaSSIFIcaTION accUracY VaLUES FOr ThE hOUSTON DaTa SET achIEVED BY DIFFErENT STaTE-OF-ThE-arT aPPrOachES.

FILTERING
Filtering approaches have been used intensively in the literature to effectively extract
contextual and spatial features by attenuating redundant spatial details (based on a criterion) and preserving the geometrical characteristics of the other regions. Among those
techniques, one can refer to 1) morphological
profiles (MPs) [144], which can be produced
by the sequential implementation of opening
and closing operators through reconstruction by considering a structuring element
of increasing size; 2) attribute profiles (APs)
[145], which can obtain a multilevel characterization of the input image by considering
the repeated implementation of morphological attribute filters; and 3) extinction profiles
(EPs) [146], which can obtain a multilevel
characterization of the input image by considering the repeated implementation of a
morphological extinction filter.
These approaches have been investigated frequently for the fusion of lidar and
HSI because they are fast and conceptually
simple and able to provide accurate classification results. For instance, in [147] and
[148], the spatial features of HSI and lidar
were extracted using APs. Then, they were
concatenated and fed to a classifier, leading
to quick, precise results in terms of classification accuracy. In [142], EPs were used to
automatically extract the spatial and elevation features of HSI and lidar data. The extracted features were stacked and then classified using an RF classifier. (The results
obtained by that approach can be found in
Table 2 as EPHSI +lidar . h
Filtering approaches like MPs, APs, and
EPs suffer from two shortcomings: the curse
of dimensionality and intensive processing
time for the subsequent classification steps
because they usually increase the number
of dimensions by stacking spectral, spatial, and elevation features extracted from
HSI and lidar, while the number of training samples remains the same. To address
these deficiencies, composite kernel- and
low-rank-based methods, which will be discussed in the following subsections, have
been suggested in the literature to effectively fuse HSI and lidar.

21
IEEE Geoscience and Remote Sensing Magazine - March 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2019
