IEEE Geoscience and Remote Sensing Magazine - June 2017 - 24

Table 5. The unIVeRSITY OF hOuSTOn liDaR DaTa: ClaSSIFICaTIOn aCCuRaCIeS ObTaIneD bY
ORIGInal aPs anD aPPRs (The ValueS WITh The hIGheST aCCuRaCY aRe In bOlD).
RaW

MPns

MPRs

MPPRs

ORIGInal aPs

aPPRs

aPa

aPs

aPi

aPs

aPPRa

aPPR s

aPPR i

aPPR s

number of features

1

21

21

21

21

21

21

63

21

21

21

63

oa

31.34

65.86

57.69

68.31

55.08

51.73

53.73

63.65

69.97

59.26

57.85

72.66

aa

37.57

68.76

63.92

70.18

58.37

55.65

56.89

65.33

72.67

63.28

62.05

74.81

K

26.53

63.11

54.24

65.63

51.47

47.96

49.88

60.57

67.41

55.94

54.32

70.31

Grass healthy

35.90

37.51

44.16

48.05

60.02

61.63

63.44

61.63

60.97

70.47

62.58

76.16

Grass stressed

2.73

50.19

45.02

66.54

42.39

16.64

24.91

49.81

51.88

18.80

39.00

62.12

Grass synthetis

87.92

97.82

94.26

94.85

86.34

86.34

87.52

86.34

88.12

85.35

91.09

88.71

tree

40.34

78.88

65.34

70.74

55.40

68.28

90.72

71.40

68.56

75.47

60.32

73.20

soil

14.49

73.30

70.46

79.36

28.41

35.61

33.05

36.08

79.83

78.50

80.30

89.77

Water

70.63

63.64

86.71

78.32

68.53

64.34

78.32

70.63

78.32

76.92

81.82

88.11

residential

57.37

68.38

65.21

70.80

66.14

60.63

35.07

74.35

77.71

54.85

58.40

69.50

commercial

14.62

96.39

73.31

90.31

70.09

52.90

37.42

65.81

62.77

54.23

42.55

71.51

road

22.0

50.71

40.89

60.06

49.95

37.58

47.21

53.45

57.44

53.26

34.37

47.78

Highway

5.50

38.32

34.75

44.21

27.61

37.45

54.05

64.09

58.98

36.78

47.20

65.83

railway

67.65

81.31

43.55

79.41

89.37

96.77

90.82

98.86

89.75

76.85

88.24

98.48

Parking lot 1

12.10

55.24

49.38

54.47

44.67

27.76

47.17

56.96

74.35

51.30

38.81

57.44

Parking lot 2

11.93

58.60

63.51

69.12

61.75

62.46

39.65

65.61

62.81

67.37

61.05

63.51

tennis court

93.12

100.0

98.38

68.83

97.17

97.17

96.36

97.17

99.60

100.0

89.07

89.88

running track

27.27

81.18

83.72

77.59

27.70

29.18

27.70

27.70

81.61

49.05

56.03

80.13

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

figure 17. The classification maps for the University of Houston LiDAR data. Thematic maps using (a) a raw hyperspectral image, (b) MPNs,
(c) MPRs, (d) MPPRs, (e) APm, (f) APPR m, (g) APs, and (h) APPRs.

soil, and road and parking lot 1). Moreover, these objects
are connected in typical remote-sensing scenes. That is
why geodesic reconstruction (even original APs) performs
much worse than those with no reconstruction and with
partial reconstruction, with OA decreases that are almost
2-10%. Many classes are misclassified, grass stressed and
soil are confused, and road and parking lot 1 are confused. Working as connected filters, geodesic reconstruction and conventional AFs treat connected objects (even
24

in different classes) as a single object. Without reconstruction, the borders of rectangular objects are obviously degraded to round ones. With partial reconstruction, the
classification performances are greatly improved. APPRs
produce the best OA, AA, and K, with 7% OA improvements over MPNs and 9% OA improvements over APs.
With partial reconstruction, the OA improvements over
original APs are 14, 8, and 4% for the attribute of area, Std,
and MI, respectively.
ieee Geoscience and remote sensing magazine

june 2017
Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2017
