You are on page 1of 3

1

Improved Spectral-Spatial Classification of


Remotely Sensed Images using Kernel Feature
Space Representation
Sergio Bernabe, Student Member, IEEE, Prashanth Reddy Marpu, Antonio Plaza, Senior Member, IEEE,
Mauro Dalla Mura, Member, IEEE, and Jon Atli Benediktsson, Fellow, IEEE

AbstractOver the last few years, several new strategies


have been proposed for spectral-spatial classification of remotely
sensed image data. However, the availability of high spectral
resolution has been generally considered as a requirement in
many remote sensing applications. Unfortunately, in many cases
such detailed spectral resolution is not available. In multispectral
imaging, for instance, only a few spectral bands are generally
available. Moreover, many remote sensing instruments with
high spatial resolution can only provide panchromatic bands.
In this paper, we explore the possibility to perform advanced
spectral-spatial classification of remotely sensed images when
the available spectral resolution is very limited. Specifically,
we propose a new strategy which first expands the spectral
dimensionality of multispectral/panchromatic images by using
the kernel principal component analysis (KPCA). Then, we use
extended morphological attribute profiles (EMAPs) to characterize spatial information, prior to spectral-spatial classification
using support vector machines (SVMs). This strategy allows us to
increase the dimensionality and further improve the classification
results that can be obtained for remotely sensed imagery with
very limited spectral resolution. Our experimental assessment
is based on comparing the classification results obtained with
hyperspectral images with those obtained after applying the proposed methodology to the corresponding panchromatic versions
of the hyperspectral data. Interestingly, our results demonstrate
that the proposed methodology can provide similar or even
superior classification results for the panchromatic data than
those obtained for the original hyperspectral image. This opens
many innovative perspectives from the viewpoint of applying
hyperspectral imaging algorithms to remotely sensed data sets
with limited spectral resolution.
Index Terms

I. I NTRODUCTION
II. C ONCLUSIONS AND FUTURE RESEARCH
R EFERENCES

Sergio Bernabe and Antonio Plaza are with the Hyperspectral Computing
Laboratory, University of Extremadura, E-10003 Caceres, Spain. Funding
from CEOS-SPAIN project by the Spanish Ministry of Science and Innovation,
reference AYA2011-29334-C02-02, is acknowledged.
Prashanth Reddy Marpu is with the Masdar Institute of Science and
Technology, P.O. Box 54224 - Abu Dhabi, UAE.
Jon Atli Benediktsson is with the Faculty of Electrical and Computer
Engineering, University of Iceland, 101 Reykjavik, Iceland.

TABLE I
OVERALL ACCURACY ( IN PERCENTAGE )
Overall Accuracy

AND STANDARD DESVIATION OBTAINED AFTER APLYING THE

ROSIS PAVIA UNIVERSITY

SVM C LASSIFIER FOR

AVIRIS INDIAN PINES

DIFFERENT SCENES .

GOOGLE MAPS - INDIAN PINES

Features

Standard

50 pixels

5% Training

5% Training

10% Training

15% Training

5% Training

10% Training

Original Hyperspectral

80.64% 0.00

84.09% 2.06

93.38% 0.21

75.60% 1.15

81.95% 0.54

84.46% 0.45

15% Training

RGB Multispectral

65.81% 0.00

62.67% 5.62

78.87% 0.40

48.33% 0.71

49.63% 0.29

49.86% 0.38

52.52% 1.41

54.94% 0.65

55.55% 0.51

EMAP(RGB)

76.85% 0.00

91.87% 0.63

96.89% 0.23

79.63% 0.73

83.71% 0.47

85.89% 0.44

76.64% 0.79

81.69% 0.77

84.18% 0.32

KPCA n1p0(RGB)

66.13% 0.00

65.60% 4.26

78.92% 0.26

47.97% 1.20

49.50% 0.46

49.89% 0.17

53.91% 0.84

55.99% 0.62

57.16% 0.46

KPCA n1p5(RGB)

66.01% 0.00

63.93% 4.61

78.62% 0.38

48.83% 0.38

49.59% 0.38

84.46% 0.45

52.78% 0.71

55.36% 0.46

55.96% 0.31

KPCA n2p0(RGB)

67.94% 0.00

62.86% 3.71

78.61% 0.32

48.25% 0.62

49.25% 0.46

49.70% 0.30

52.96% 0.62

55.33% 0.48

55.92% 0.53

EMAP(KPCA n1p0(RGB))

93.41% 0.00

93.22% 0.71

97.18% 0.64

62.75% 3.72

74.45% 0.63

77.32% 2.90

75.88% 1.14

83.35% 1.11

86.84% 0.62

EMAP(KPCA n1p5(RGB))

94.09% 0.00

90.61% 2.38

97.18% 0.62

63.86% 2.74

73.91% 2.16

77.34% 3.14

72.25% 1.98

80.25% 1.08

83.65% 1.40

EMAP(KPCA n2p0(RGB))

91.87% 0.00

91.61% 1.36

97.25% 0.63

61.40% 3.14

71.62% 3.68

77.86% 2.30

70.61% 1.54

80.05% 1.12

84.58% 0.21

2*[60pt]
(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

Fig. 1.
Classification results for ROSIS PAVIA UNIVERSITY (using SVM classifier in the top and RF classifier in the bottom, trained with
5% of the available samples per class). (a) Ground-truth. (b) Original hyperspectral (93.38%). (c) RGB multispectral (78.87%). (d) Emap(RGB)
(96.89%). (e) Emap(KPCA n2p0(RGB)) (97.25%). (f) Original hyperspectral (87.66%). (g) RGB multispectral (77.08%). (h) Emap(RGB) (96.76%). (i)
Emap(KPCA n1p5(RGB)) (98.81%). (j) Class accuracies between SVM and RF with the two better classifications.

TABLE II
OVERALL ACCURACY ( IN PERCENTAGE )
Overall Accuracy

AND STANDARD DESVIATION OBTAINED AFTER APLYING THE

ROSIS PAVIA UNIVERSITY

RF C LASSIFIER FOR

AVIRIS INDIAN PINES

DIFFERENT SCENES .

GOOGLE MAPS - INDIAN PINES

Features

Standard

50 pixels

5% Training

5% Training

10% Training

15% Training

5% Training

10% Training

Original Hyperspectral

71.42% 0.13

72.63% 2.66

87.66% 0.28

69.84% 1.30

75.38% 0.56

77.70% 0.51

15% Training

RGB Multispectral

64.87% 0.11

61.68% 2.62

77.08% 0.27

44.27% 0.54

46.17% 0.43

47.05% 0.28

46.74% 0.62

49.41% 0.24

50.97% 0.42

EMAP(RGB)

80.81% 0.30

88.61% 1.38

96.76% 0.23

80.93% 1.07

84.78% 0.49

86.46% 0.37

76.00% 0.63

81.10% 0.69

83.40% 0.41

KPCA n1p0(RGB)

65.26% 0.11

65.32% 2.83

78.06% 0.22

44.62% 0.62

45.95% 0.52

46.30% 0.37

51.78% 0.81

53.64% 0.51

53.88% 0.32

KPCA n1p5(RGB)

65.44% 0.18

65.64% 2.39

78.48% 0.26

44.36% 0.62

45.83% 0.47

46.24% 0.53

52.48% 0.71

54.27% 0.35

54.47% 0.33

KPCA n2p0(RGB)

65.97% 0.14

65.50% 2.59

78.46% 0.24

44.95% 0.73

46.45% 0.47

46.81% 0.36

52.91% 0.85

54.66% 0.36

54.82% 0.46

EMAP(KPCA n1p0(RGB))

92.45% 0.15

94.72% 0.88

98.61% 0.13

88.65% 0.63

92.67% 0.59

94.25% 0.42

82.26% 0.81

87.63% 0.42

90.13% 0.53

EMAP(KPCA n1p5(RGB))

94.23% 0.35

94.92% 1.15

98.81% 0.14

88.74% 0.64

92.65% 0.40

93.96% 0.28

81.95% 0.65

87.65% 0.39

90.02% 0.38

EMAP(KPCA n2p0(RGB))

93.96% 0.57

95.05% 1.09

98.67% 0.14

88.25% 0.53

92.26% 0.61

93.90% 0.33

81.46% 0.46

87.65% 0.57

90.04% 0.48

2*[60pt]
(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

Fig. 2. Classification results for AVIRIS Indian Pines (using SVM classifier in the top and RF classifier in the bottom, trained with 5% of the available samples
per class). (a) Ground-truth. (b) Original hyperspectral (75.60%). (c) RGB multispectral (48.33%). (d) Emap(RGB) (79.63%). (e) Emap(KPCA n1p5(RGB))
(63.86%). (f) Original hyperspectral (69.84%). (g) RGB multispectral (44.27%). (h) Emap(RGB) (80.93%). (i) Emap(KPCA n1p5(RGB)) (88.74%). (j) Class
accuracies between SVM and RF with the two better classifications.

You might also like