4-CSPD | Snow | Matrix (Mathematics)

This art icle was downloaded by: [ Niigat a Daigaku] , [ Gulab Singh

]
On: 14 December 2011, At : 17: 44
Publisher: Taylor & Francis
I nforma Lt d Regist ered in England and Wales Regist ered Number: 1072954 Regist ered
office: Mort imer House, 37- 41 Mort imer St reet , London W1T 3JH, UK
Geocart o Int ernat i onal
Publ i cat i on det ai l s, i ncl udi ng i nst r uct i ons f or aut hor s and
subscr i pt i on i nf or mat i on:
ht t p: / / www. t andf onl i ne. com/ l oi / t gei 20
Ut i l i zat i on of f our-component
scat t eri ng power decomposi t i on
met hod f or gl aci at ed t errai n
cl assi f i cat i on
Gul ab Si ngh
a
, Y. Yamaguchi
a
& S. - E. Par k
a
a
Gr aduat e School of Sci ence and Technol ogy, Ni i gat a Uni ver si t y,
Ni i gat a, Japan
Avai l abl e onl i ne: 24 May 2011
To ci t e t hi s art i cl e: Gul ab Si ngh, Y. Yamaguchi & S. - E. Par k ( 2011) : Ut i l i zat i on of f our - component
scat t er i ng power decomposi t i on met hod f or gl aci at ed t er r ai n cl assi f i cat i on, Geocar t o Int er nat i onal ,
26: 5, 377- 389
To l i nk t o t hi s art i cl e: ht t p: / / dx. doi . or g/ 10. 1080/ 10106049. 2011. 584978
PLEASE SCROLL DOWN FOR ARTI CLE
Full t erms and condit ions of use: ht t p: / / www. t andfonline. com/ page/ t erms- and-
condit ions
This art icle may be used for research, t eaching, and privat e st udy purposes. Any
subst ant ial or syst emat ic reproduct ion, redist ribut ion, reselling, loan, sub- licensing,
syst emat ic supply, or dist ribut ion in any form t o anyone is expressly forbidden.
The publisher does not give any warrant y express or implied or make any represent at ion
t hat t he cont ent s will be complet e or accurat e or up t o dat e. The accuracy of any
inst ruct ions, formulae, and drug doses should be independent ly verified wit h primary
sources. The publisher shall not be liable for any loss, act ions, claims, proceedings,
demand, or cost s or damages what soever or howsoever caused arising direct ly or
indirect ly in connect ion wit h or arising out of t he use of t his mat erial.
Utilization of four-component scattering power decomposition method
for glaciated terrain classification
Gulab Singh*, Y. Yamaguchi and S.-E. Park
Graduate School of Science and Technology, Niigata University, Niigata, Japan
(Received 13 January 2011; final version received 27 April 2011)
Glaciated terrain classification is important for hydrological and climate
change modelling. For this purpose, fully polarimetric Advanced Land
Observation Satellite-Phase Array L-band Synthetic Aperture Radar (ALOS-
PALSAR) data has been used over Indian Himalayan glaciated region.
PALSAR data has been analyzed based on the three and four component
scattering decomposition methods for glaciated terrain classification. These
methods have been applied on multi-looked 3 6 3 coherency matrix of ALOS-
PALSAR data. The analysis of these methods shows that the Freeman and
Durden three-component scattering power decomposition (3-CSPD) method has
over estimation problem in volume backscattering component as compared to
the Yamaguchi four-component scattering power decomposition (4-CSPD)
method. After finding suitability of 4-CSPD method over Himalayan glaciated
terrain, it has been combined with complex Wishart distribution for supervised
classification of ALOS-PALSAR image. By this way, an overall accuracy has
been found to be 93.38%. Even this procedure shows very high accuracy but
discrimination between vegetation and glacier snow/ice classes was not
properly done. To overcome this ambiguity, the probability difference between
surface backscattering and volume backscattering has been introduced as
further steps in classification procedure.
Keywords: decomposition; PALSAR; Himalayan; glaciated terrain
1. Introduction
Satellite remote sensing has great potential in the study of dynamically changing
environments related to the high altitude cold regions mainly because of its repetitive
capability and synoptic coverage. The land-covered features have unique reflectance
characteristics in different spectral bands of optical sensors, which may provide
information on physical properties as well as the areal extent of glaciated terrain
features under cloud-free conditions. However, they have some difficulty in rugged
high mountainous area: (1) optical images are often affected by clouds in
mountainous glacier areas. The Himalayan region is strongly affected by monsoons
and cloud cover is quite common, especially during the summer months. The
glaciated terrain classification is even hindered in Himalayan region due to a lack of
cloud-free optical images; (2) the mountain shadow makes it difficult to discriminate
between glacier areas and non-glacier areas and (3) an ambiguity between snow and
*Corresponding author. Email: g.singh@wave.ie.niigata-u.ac.jp
Geocarto International
Vol. 26, No. 5, August 2011, 377–389
ISSN 1010-6049 print/ISSN 1752-0762 online
Ó 2011 Taylor & Francis
DOI: 10.1080/10106049.2011.584978
http://www.informaworld.com
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

ice exists because both have similar optical properties in glacial areas (Racoviteanu
et al. 2009). Due to the strong spatial and time dependent dynamics of glaciated
terrain, regular and frequent mapping is necessary to monitor glaciated terrain, and
requires sensors that are time and weather independent. Synthetic Aperture Radar
(SAR) remote sensing with its all-weather capability, cloud penetration and
independence of sun illumination can add considerable robustness to classify the
glaciated terrain (Rott 1994, Singh and Venkataraman 2009). In case of monostatic
fully polarimetric SAR data, point targets are characterized by five parameters (three
amplitude and two relative phases). Therefore, SAR full polarimetry techniques can
lead to a significant improvement in the quality of classification and segmentation
results in comparison to conventional single-channel SAR. Fully polarimetric SAR
also allows a discrimination of different types of scattering mechanisms. This
becomes possible because the received power depends strongly on the actual
backscattering process. Received backscattering power can be divided into a sum of
various backscattering contributions by using polarimetric target decomposition
methods (Cloude and Pottier 1996). Thus, polarimetric decomposition method can
be utilized for extracting the corresponding target type in fully polarimetric
Advanced Land Observation Satellite-Phase Array L-band Synthetic Aperture
Radar (ALOS-PALSAR) images over glaciated terrain.
In the literatures (Cloude and Pottier 1996, Yamaguchi 2007, Cloude 2009, Lee
and Pottier 2009), polarimetric target decomposition methods are categorized into
two types: the first type is coherent decomposition methods that are directly
performed on the scattering matrix. The second type is incoherent decomposition
methods based on the second order statistics of polarimetric information, e.g. on
the coherency matrix. The incoherent decomposition methods (Freeman and
Durden 1998, Yamaguchi et al. 2006, Yajima et al. 2008) decompose the coherency
matrix as the incoherent sum of scattering power of a distributed target. Since
most of the targets are distributed in natural earth surface, this type of target can
only be characterized statistically. Singh (2010) proved that incoherent decom-
position provides sufficient information for classification in glaciated terrain
features such as debris-covered ice, snow, barren rock, etc., using fully polarimetric
data.
In this work, the Yamaguchi four-component scattering power decomposition
(4-CSPD) method (Yamaguchi et al. 2006, Yajima et al. 2008) is applied to identify
glaciated terrain features in the part of Indian Himalaya. The 4-CSPD method
decomposes polarimetric radar power into surface, double bounce (DB), volume and
helix power scattering. This method is an extension of the Freeman and Durden
three-component scattering power decomposition (3-CSPD) method (Freeman and
Durden 1998) to general scattering case with non-reflection symmetry condition.
However, the 4-CSPD method has following advantages: (1) straightforward
implementation; (2) scattering power calculations are easy; (3) the decomposed
powers correspond to physical scattering mechanisms, i.e. surface scattering, DB
scattering, volume scattering, helix (circular polarization) scattering; (4) output
colour-coded images are directly recognizable and easy to understand. Moreover,
this paper presents the 4-CSPD method (Yamaguchi et al. 2006, Yajima et al. 2008)
suitability as compared to the 3-CSPD method (Freeman and Durden 1998) for
glaciated terrain features identification. A new methodology has been discussed by
combining the complex Wishart distribution (CWD) (Wishart 1928) and 4-CSPD
method for glaciated terrain classification.
378 G. Singh et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

Figure 1. Location map of study area.
2. Study area
The location map is shown in Figure 1. The Alaknanda river catchment,
Uttarakhand, India has many glaciers. Satopanth and Bhagirath Kharak glaciers
are the major glaciers among them in this catchment. The Satopanth and Bhagirath
Kharak glaciers are approximately 13 and 18.5 km long with an average width of
750–850 m, covering an area of 21.17 and 31.17 km
2
, respectively. The upper
Alaknanda river catchment covers an area of 1544.08 km
2
, out of which 70.70 and
107.22 km
2
are covered by the Satopanth and Bhagirath Kharak sub-watersheds,
respectively (Nainwal et al. 2008).
The elevation ranges between 2000 meters above sea level (m.a.s.l) and 7000
m.a.s.l. The Alaknanda river, which is the main tributary of Ganga river, originates
at the snout of the Satopanth glacier. The area falls between latitude 30840
0
N and
30850
0
N and longitude between 79815
0
E and 79828
0
E. Satopanth and Bhagirath
Kharak glaciers are shown in Figure 2(a)–(d). This glaciated region includes snow,
debris-covered glacier and barren rock targets.
3. Data used
In this study, we acquired the PALSAR fully polarimetric, single look complex, level
1.1 data of 12 May 2007 with 21.58 incident angle and nominal pixel spacing
(azimuth 6 range) 3.54 (m) 6 9.36 (m). PALSAR, on-board the ALOS, was
launched on 24 January 2006 by the Japan Aerospace Exploration Agency (JAXA).
It operates in the L-band frequency of 1.27 GHz (23.6 cm wavelength). It is well
known that L-band microwave signals penetrate through dry snowpack with
negligible volume backscatter from snow. However, if snowpack is wet, the situation
becomes different. If the moisture exceeds 1% in highly accumulated snowpack, L-
band frequency suffers from attenuation in the snowpack while reflection or
backscatter from the snowpack comes out (Abe et al. 1990). In general, snow cover
area becomes wet in May (early summer) over Himalayan snow bound area with
Geocarto International 379
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

significant melting. The magnitude of backscatter depends on the snow density and
water content, and the depth of snowpack. Since the snowpack on 6 May 2007 is
rather wet due to the beginning of snow melting, the snowpack contains water and is
not so transparent for L-band frequency. Snowpack is also heterogeneous with snow
grain particle compressed during winter season and has rather high density due to
snow accumulation and melting cycle. The co-polarization backscatters (HH and
VV) increases with snow volume, while the cross-polarization backscatter (HV)
remains low as compared to co-polarization. Since the HV components contribute
only to volume scattering, the main polarimetric response from snowpack becomes
surface scattering in the L-band (Abe et al. 1990).
Figure 2(b) shows the Pauli colour composite image (12 May 2007) which gives
the clear information about single scattering (snow cover area over glacier and non-
Figure 2. (a) Cloud free ALOS-AVNIR-2 image of 24 May 2010 (b) Pauli RGB of PALSAR
of 12 May 2007 (c) Photo of Satopanth (ST) and Bhagirath Kharak (BK) entrance (d) Photo
view of Satopanth glacier.
380 G. Singh et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

glacier), DB (dihedral features) and volume scattering (debris-covered glacier) in the
study area.
4. Method and technique
4.1. Decomposition method
Once scattering matrix S is acquired with fully polarimetric radar, we can define the
scattering vector k as
k ¼
1
ffiffiffi
2
p
S
HH
þ S
VV
S
HH
À S
VV
2S
HV
_
_
_
_
ð1Þ
where S
HH
, S
VV
, S
HV
are elements of scattering matrix S assuming the reciprocal
condition of S
HV
¼ S
VH
The coherency matrix is given as
T ½ Š ¼ k:k
ÃT
¸ _
¼
T
11
T
12
T
13
T
21
T
22
T
23
T
31
T
32
T
33
_
_
_
_
ð2Þ
where *
T
denotes complex conjugation and transposition, and h.i denotes ensemble
average in an imaging window.
The 4-CSPD method divides the measured coherency matrix into four sub-
matrices representing physical scattering mechanisms (Yamaguchi et al. 2006,
Yajima et al. 2008)
T ½ Š ¼ f
s
T
s
½ Š þ f
d
T
d
½ Š þ f
v
T
v
½ Š þ f
c
T
c
½ Š ð3Þ
where f
s
, f
d
, f
v
and f
c
are coefficients to be determined. [T
s
], [T
d
], [T
v
] and [T
c
] are
expansion coherency matrices corresponding to surface, DB, volume and helix
scattering, respectively.
The single-bounce scattering model is represented by surface scattering
phenomena from slightly rough surface in which the cross-polarized component
is negligible. The expansion coherency matrix for surface scattering is as
follows:
½T
s
Š ¼
1 b
Ã
0
b b j j
2
0
0 0 0
_
_
_
_
; b j j ¼
S
HH
À S
VV
S
HHþS
VV
¸
¸
¸
¸
¸
¸
¸
¸
< 1: ð4Þ
The DB scattering model is based on the hypothesis of double reflections from
right angle structures. DB structure includes road surface–building wall, ground
trees, and man-made targets, etc. The expansion coherency matrix for DB scattering
is as follows:
T
d
½ Š ¼
a j j
2
a 0
a
Ã
1 0
0 0 0
_
_
_
_
; a j j ¼
S
HH
þ S
VV
S
HH
À S
VV
¸
¸
¸
¸
¸
¸
¸
¸
< 1: ð5Þ
Geocarto International 381
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

Volume scattering can be observed if SAR beam penetrates into a medium.
Scattering by trees or branches, subsurface or snow/ice layers, etc. are examples of
volume scattering. For the volume scattering model, we choose one of the following
matrices according to the magnitude balance of jS
HH
j
2
and jS
VV
j
2
(Yamaguchi et al.
2006).
When 10log (hjS
VV
j
2
i/hjS
HH
j
2
i) ! 2 dB
½T
v
Š ¼
1
30
15 À 5 0
À5 7 0
0 0 8
_
_
_
_
: ð6Þ
When j10log (hjS
VV
j
2
i/ hjS
HH
j
2
i) j 52 dB
½T
v
Š ¼
1
4
2 0 0
0 1 0
0 0 1
_
_
_
_
: ð7Þ
When 10log (hjS
VV
j
2
i/ hjS
HH
j
2
i) 72 dB
½T
v
Š ¼
1
30
15 5 0
5 7 0
0 0 8
_
_
_
_
: ð8Þ
Helix scattering power is equivalent to circular polarization power. This term
appears in urban and mountainous area for L-band data. The helix scattering
expansion matrix, which takes into account of non-reflection symmetry condition, is
as follows:
½T
c
Š ¼
0 0 0
0 1 Æ j
0 Ç j 1
_
_
_
_
: ð9Þ
The corresponding scattering powers (the surface scattering power P
s
, the DB
scattering P
d
, the volume scattering power P
v
and the helix scattering power P
c
) are
directly obtained from the expansion coefficients when we apply decomposition. The
decomposition takes into account an imbalance of the co-polarized channel power.
For the case of j10log (hjS
VV
j
2
i/ hjS
HH
j
2
i) j 52 dB, the decomposed power
expression becomes as (Yamaguchi 2007):
P
c
¼ f
c
¼ 2jImfT
23
gj ð10:1Þ
P
v
¼ f
v
¼ 4T
33
À2P
c
ð10:2Þ
P
s
¼ f
s
ð1 þ jbj
2
Þ ð10:3Þ
P
d
¼ f
d
ð1 þ jaj
2
Þ: ð10:4Þ
382 G. Singh et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

3.2. Complex Wishart classifier
The CWD is expressed as (Wishart 1928)
P T ½ Š=T
m
ð Þ ¼
n
nq
T ½ Š j j
nÀq
e
ÀnTr T
m
½ Š
À1
T ½ Š ð Þ
K½T
m
Š
n
; K ¼ p
qðq À 1Þ
2
ÀðnÞ:::Àðn À q þ 1Þ ð11Þ
where n is the number of looks and q is the polarimetric dimension and [T
m
] = E[T].
Using the complex Wishart distribution of the coherency matrix [T], an appropriate
distance measure, d
m
, can then be calculated according to Bayes maximum
likelihood classification as (Lee and Pottier 2009)
d
m
ð T ½ Š T
m
½ Š j Þ ¼ Tr T
m
½ Š
À1
T ½ Š
_ _
þ ln T
m
½ Š ð Þ ð12Þ
thus leading to a minimum distance classification independent of the number
of looks used to form the multi-looked coherency matrix [T] (Lee and Pottier
2009):
T ½ Š 2 T
m
½ Š if d
m
T ½ Š ð Þ < d
j
T ½ Š ð Þ 8j 6¼ m: ð13Þ
3.3. Decomposed power probability
The 4-CSPD method decomposes total backscattering power into surface scattering
(P
s
), DB backscattering (P
d
), volume backscattering (P
v
) and helix backscattering
(P
c
). Total backscattering power (TP) can be defined as
TP ¼ P
s
þ P
d
þ P
v
þ P
c
: ð14Þ
With the help of Equation (7), we can define the probability of surface
backscattering and volume backscattering decomposed components as follows:
Probability of surface backscattering; P
surface
¼
P
s
TP
ð15Þ
Probability of volume backscattering P
volume
¼
P
v
TP
: ð16Þ
Therefore from Equations (8) and (9)
À 1 DP
sÀv
¼ P
surface
À P
volume
1 ð17Þ
Equation (17) helps us to determine the dominated scattering component from
surface scattering and volume scattering in the 4-CSPD method’s decomposed
image. If DP
s7v
is positive, we determine that surface scattering is the dominant
contribution. On the other hand, if DP
s7v
is negative, we determine that volume
scattering is the dominant contribution. The threshold does not begin from zero
because we want to take into account the noise variation in case that both the
probabilities are close to zero.
Geocarto International 383
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

Figure 3. Flow chart of PALSAR image classification.
3.4. Classification procedure
Using 4-CSPD method and CWD, a supervised classification methodology has been
presented for fully polarimetric SAR images classification. The flow chart of
developed methodology is shown in Figure 3.
The sequence of this procedure is followed as given below:
(1) First of all, a multi-looked (six times in azimuth direction and one time in
range direction) coherency matrix has been generated.
384 G. Singh et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

(2) For reduction, the speckle noise from the polarimetric SAR data, the
polarimetric refined Lee filter (Lee et al. 1999) with window size 7 6 7 has
been applied on coherency matrix.
(3) 4-CSPD method has been applied on de-speckled coherency matrix and 4-
CSPD method false colour composite (FCC) image has been generated. In
4-CSPD method FCC image, red colour is assigned to DB scattering, green
colour is assigned to volume scattering and blue colour is assigned to surface
scattering. Training samples have been allotted on the basis of visually
comparing four component colour composite image with AVNIR-2 snow
cover image and field information.
(4) CWD has been applied on de-speckled coherency matrix and computed the
averaged coherency matrices from the assigned classes. These computed
mean matrices have been used as the class centres. All pixels are classified
based on their Wishart distance measure and criteria (Equations (12) and
(13)) from class centres.
(5) Finally, based on conditional approach, the probability difference (DP
s7v
)
has been used for resolving volume scattering ambiguity from vegetation and
glacier snow/ice. The conditional approach is defined as
EITHER glacier snow/ice class IF (DP
s7v
! 0.05 and vegetation class) OR
classified classes OTHERWISE, i.e. an additional class (glacier snow/ice class) has
been added to supervised classified image, where in, the probability difference
(DP
s7v
) image has ! 0.05 value but supervised classified image shows vegetation
class. Otherwise, supervised classified pixels remain same in final classified image.
4. Results and discussion
The 3-CSPD and 4-CSPD methods have been applied on L-band PALSAR data
over the Satopanth and Bhagirath Kharak glaciated region. 3-CSPD and 4-CSPD
methods FCC images are shown in Figure 4. In 3-CSPD and 4-CSPD methods, FCC
images of PALSAR data over Satopanth glacier region (Figure 4), blue to deep blue
colour represents surface scattering from snow cover over glacier area (accumulation
zone) and permanent snow cover at mountains peaks. Red colour represents DB or
dihedral scattering mechanism. Debris-covered glaciers (ablation area) are shown in
green colour. Glacier moraine dammed lakes appear in deep blue colour.
The comparison of visual interpretation has been done for both decomposed
FCCs with the Pauli RGB (Figure 2(b)) and AVNIR-2 (Figure 4) image. The
differences are clearly seen between 3-CSPD and 4-CSPD methods FCCs in Figure 5
(enlarged part of red colour rectangular area in Figure 4). Most of the differences can
be visibly identified in DB and surface scattering component, and these components
are clearly exposed (Figure 5(b)) in 4-CSPD method FCC. Some of these differences
that are indicated by yellow circles no. 1, 2 and 3 in Figure 5(b) (4-CSPD FCC)
correspond to yellow colour circles no. 1, 2 and 3 in Figure 5(a) (3-CSPD FCC),
respectively. Furthermore, 4-CSPD method gives very sharp information about
dihedral features in the study area as compared to 3-CSPD method and Pauli RGB
images. Therefore, the fourth component of four component scattering power
decomposition (4-CSPD) method represents the helical scattering phenomenon,
which occurs due to slope surface of target. It has been seen in Figure 6 that the helix
(Pc) scattering component shows high value (greater than 710 dB) at steep slope
Geocarto International 385
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

and low values (less than 720 dB) are found over snow-covered area. Himalayan
topography has gentle to steep slope, which behaves like oriented target from the
direction of radar and oriented target does not hold reflection symmetry condition
where the 3-CSPD method works, which causes the over estimation of volume
scattering in 3-CSPD method. But in 4-CSPD method, volume scattering is reduced
by the fourth component where volume scattering is more than helix scattering
component. The main reason for the reduction of the volume component for the 4-
CSPDmethod is due to the reflection symmetric space where the 4-CSPDmethod works.
In other words, it corrects for rotation along the line of sight (LOS) while it decomposes.
Therefore, a 4-CSPD method is suitable for fully polarimetric PALSAR data
decomposition over Himalayan glaciated terrain as compared to the 3-CSPD method.
Moreover, using 4-CSPD method decomposed image and CWD, ALOS
PALSAR data was classified into six major classes (e.g. snow, non-snow and
unidentified/layover, Figure 7). Classification technique was applied on coherency
matrix, and training samples were taken from the four component colour composite
image with the help of visual interpretation of 4-CSPD method FCC and AVNIR-2
image (which are shown in Figure 4).
Figure 4. ALOS-AVNIR-2 image of 6 May 2007 (upper), and 3-CSPD model FCC of 12
May 2007 PALSAR data (middle) and 4-CSPD model FCC of 12 May 2007 PALSAR data
(bottom) (Red colour rectangular on 4-CSPD model FCCs enlarged view are shown in
Figure 5).
386 G. Singh et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

Figure 5. (a) 3-CSPD Model FCC of 12 May 2007 PALSAR data (b) 4-CSPD Model FCC
of 12 May 2007 PALSAR data.
Figure 6. Helix scattering (P
c
) component of 4-CSPD for PALSAR image of 12 May 2007.
Figure 7. PALSAR classified image of 12 May 2007.
Geocarto International 387
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

The most common tool used for assessing the classification accuracy is the
confusion (or error) matrix (Table 1). The columns in a confusion matrix (Table 1)
represent test data that have been collected via field observation and interpretation of
4-CSPD method FCC and ALOS AVNIR-2 image, while rows represent the labels
assigned by the classifier. The main diagonal entries of the Table 1 represent the
number of pixels that are correctly classified. By this way, overall classification
accuracy has been found to be 93.38%. Both user’s accuracy and producer’s accuracy
of vegetation class are more than 80%, but vegetation class could not be
discriminated from glacier snow/ice by using alone complex Wishart classifier with
defined training samples. For further improvement of classified image (Figure 7),
a probability difference image (DP
s7v
) has been used for separating these classes. The
probability difference image shows lowvalue over vegetation area and high value over
glacier snow/ice area. By using conditional approach, it is possible to resolve the
ambiguity between vegetation class (Figure 7) and glacier snow/ice (Figure 8).
Table 1. A confusion matrix composed of six glaciated terrain classes.
Snow Rock DCG
Double
bounce
(DB)/
settlement
Vegetation/
ice Layover
Row
sum
Snow 19885 67 73 0 87 65 20177
Rock 106 1594 0 29 282 0 2011
DCG 0 0 3629 0 94 469 4192
DB/settlement 0 72 0 742 11 0 825
Vegetation/ice 0 477 224 0 3072 13 3786
Layover 0 0 32 0 0 735 767
Column sum 19991 2210 3958 771 3546 1282 31758
Producer’s
accuracy (%)
99.46 72.12 91.68 96.24 86.63 57.33
User’s
accuracy (%)
98.55 79.25 86.58 89.96 89.96 95.86 Overall
accuracy¼
93.38%
Figure 8. Vegetation and ice separating in PALSAR classified image of 12 May 2007 by
using probability difference between surface scattering and volume scattering probabilities.
388 G. Singh et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

5. Summary and conclusion
In this work, full polarimetric PALSAR data of high altitude glaciated terrain in
Himalayan region has been analysed based on 3-CSPD and 4-CSPD methods and
information of various terrain features have been extracted. It has been found that
the 4-CSPD method discriminates better terrain features such as snow cover,
dihedral (DB) and glacier features as compared to 3-CSPD method.
The supervised classification procedure shows overall accuracy of 93.38%, but
the ambiguity of separating the vegetation from glacier snow/ice has also been found
in classified image. Therefore, the probability difference (DP
s7v
) has been combined
with supervised classification procedure to resolve the ambiguity between vegetation
and glacier snow/ice. In future work, this methodology will be assessed with more
time series data to check the resolving capability of the ambiguity between vegetation
and glacier snow/ice.
References
Abe, T., Yamaguchi, Y., and Sengoku, M., 1990. Experimental study of microwave
transmission in snowpack. IEEE Transactions on Geoscience and Remote Sensing, 28 (5),
915–921.
Cloude, S.R., 2009. Polarisation: applications in remote sensing. New York: Oxford University
Press.
Cloude, S.R. and Pottier, E., 1996. A review of target decomposition theorems in radar
polarimetry. IEEE Transactions on Geoscience and Remote Sensing, 34, 498–518.
Freeman, A. and Durden, S.L., 1998. A three-component scattering model for polarimetric
SAR data. IEEE Transaction on Geoscience and Remote Sensing, 36 (3), 936–973.
Lee, J.S., Grues, M.R., and de Grandi, G., 1999. Polarimetric SAR speckle filtering and its
implication for classification. IEEE Transactions on Geoscience and Remote Sensing, 37,
2363–2373.
Lee, J.S. and Pottier, E., 2009. Polarimetric radar imaging: from basics to application. Boca
Raton: CRC Press, Taylor & Francis Group.
Nainwal, H.C., et al., 2008. Temporal changes in rate of recession: evidences from Satopanth
and Bhagirath Kharak glaciers, Uttarakhand, using total station survey. Current Science,
94 (5), 653–660.
Racoviteanu, A.E., et al., 2009. Challenges and recommendations in mapping of glacier
parameters from space: results of the 2008 Global Land Ice Measurements from Space
(GLIMS) workshop, Boulder, Colorado, USA. Annals of Glaciology, 50 (53), 53–69.
Rott, H., 1994. Thematic studies in alpine areas by means of polarimetric SAR and optical
imagery. Advances in Space Research, 14 (3), 217–226.
Singh, G., 2010. SAR polarimetry techniques for snow parameters estimation. India: Thesis
(PhD). Indian Institute of Technology Bombay, India.
Singh, G. and Venkataraman, G., 2009. LOS PALSAR data Analysis of snow cover area in
Himalayan region using four component scattering decomposition model. In: 2008
International conference of recent advances in microwave theory and applications
(Microwave-08), 21–24 November 2008, Jaipur, India. New York: Institute of Electrical
and Electronics Engineers (IEEE), 772–774.
Wishart, J., 1928. The generalised product moment distribution in samples from a normal
multivariate population. Biometrika, 20A (1–2), 32–52.
Yajima, Y., et al., 2008. POLSAR image analysis of wetlands using a modified four-
component scattering power decomposition. IEEE Transaction on Geoscience and Remote
Sensing, 46 (6), 1667–1673.
Yamaguchi, Y., 2007. Radar polarimetry from basics to applications: radar remote sensing using
polarimetric information (in Japanese). Japan: IEICE. ISBN 978-4-88552-227-7.
Yamaguchi, Y., Yajima, Y., and Yamada, H., 2006. A four-component decomposition of
POLSAR images based on the coherency matrix. IEEE Geoscience and Remote Sensing
Letter, 3 (3), 292–296.
Geocarto International 389
D
o
w
n
l
o
a
d
e
d

b
y

[
N
i
i
g
a
t
a

D
a
i
g
a
k
u
]
,

[
G
u
l
a
b

S
i
n
g
h
]

a
t

1
7
:
4
4

1
4

D
e
c
e
m
b
e
r

2
0
1
1

38%. The glaciated terrain classification is even hindered in Himalayan region due to a lack of cloud-free optical images. glaciated terrain 1. PALSAR data has been analyzed based on the three and four component scattering decomposition methods for glaciated terrain classification. Introduction Satellite remote sensing has great potential in the study of dynamically changing environments related to the high altitude cold regions mainly because of its repetitive capability and synoptic coverage. the probability difference between surface backscattering and volume backscattering has been introduced as further steps in classification procedure. The land-covered features have unique reflectance characteristics in different spectral bands of optical sensors. 26. After finding suitability of 4-CSPD method over Himalayan glaciated terrain.jp ISSN 1010-6049 print/ISSN 1752-0762 online Ó 2011 Taylor & Francis DOI: 10. 5. (2) the mountain shadow makes it difficult to discriminate between glacier areas and non-glacier areas and (3) an ambiguity between snow and *Corresponding author. For this purpose. it has been combined with complex Wishart distribution for supervised classification of ALOS-PALSAR image. However.Geocarto International Vol.informaworld. Niigata University. Yamaguchi and S. To overcome this ambiguity. Keywords: decomposition. 377–389 Utilization of four-component scattering power decomposition method for glaciated terrain classification Gulab Singh*. Even this procedure shows very high accuracy but discrimination between vegetation and glacier snow/ice classes was not properly done. Email: g.niigata-u.584978 http://www. fully polarimetric Advanced Land Observation Satellite-Phase Array L-band Synthetic Aperture Radar (ALOSPALSAR) data has been used over Indian Himalayan glaciated region. Park Graduate School of Science and Technology. August 2011. Niigata. final version received 27 April 2011) Glaciated terrain classification is important for hydrological and climate change modelling.-E. they have some difficulty in rugged high mountainous area: (1) optical images are often affected by clouds in mountainous glacier areas.com .2011. PALSAR.ac.1080/10106049. Himalayan. The analysis of these methods shows that the Freeman and Durden three-component scattering power decomposition (3-CSPD) method has over estimation problem in volume backscattering component as compared to the Yamaguchi four-component scattering power decomposition (4-CSPD) method. No. Japan Downloaded by [Niigata Daigaku]. These methods have been applied on multi-looked 3 6 3 coherency matrix of ALOSPALSAR data. an overall accuracy has been found to be 93. especially during the summer months.singh@wave. The Himalayan region is strongly affected by monsoons and cloud cover is quite common. Y. By this way. [Gulab Singh] at 17:44 14 December 2011 (Received 13 January 2011. which may provide information on physical properties as well as the areal extent of glaciated terrain features under cloud-free conditions.ie.

Singh (2010) proved that incoherent decomposition provides sufficient information for classification in glaciated terrain features such as debris-covered ice.e. polarimetric decomposition method can be utilized for extracting the corresponding target type in fully polarimetric Advanced Land Observation Satellite-Phase Array L-band Synthetic Aperture Radar (ALOS-PALSAR) images over glaciated terrain. surface scattering. 2006. (4) output colour-coded images are directly recognizable and easy to understand. this paper presents the 4-CSPD method (Yamaguchi et al. 2008) is applied to identify glaciated terrain features in the part of Indian Himalaya. The incoherent decomposition methods (Freeman and Durden 1998. This method is an extension of the Freeman and Durden three-component scattering power decomposition (3-CSPD) method (Freeman and Durden 1998) to general scattering case with non-reflection symmetry condition. SAR full polarimetry techniques can lead to a significant improvement in the quality of classification and segmentation results in comparison to conventional single-channel SAR. Fully polarimetric SAR also allows a discrimination of different types of scattering mechanisms. point targets are characterized by five parameters (three amplitude and two relative phases). Yajima et al. the 4-CSPD method has following advantages: (1) straightforward implementation. The 4-CSPD method decomposes polarimetric radar power into surface. In the literatures (Cloude and Pottier 1996. Received backscattering power can be divided into a sum of various backscattering contributions by using polarimetric target decomposition methods (Cloude and Pottier 1996). barren rock. Singh et al.g. In this work. cloud penetration and independence of sun illumination can add considerable robustness to classify the glaciated terrain (Rott 1994. using fully polarimetric data. Yajima et al. on the coherency matrix. volume scattering. 2006. volume and helix power scattering. Cloude 2009. i. 2006. (2) scattering power calculations are easy. Synthetic Aperture Radar (SAR) remote sensing with its all-weather capability. ice exists because both have similar optical properties in glacial areas (Racoviteanu et al. [Gulab Singh] at 17:44 14 December 2011 . Thus.378 G. Yamaguchi 2007. Moreover. However. regular and frequent mapping is necessary to monitor glaciated terrain. double bounce (DB). 2009). this type of target can only be characterized statistically. A new methodology has been discussed by combining the complex Wishart distribution (CWD) (Wishart 1928) and 4-CSPD method for glaciated terrain classification. In case of monostatic fully polarimetric SAR data. helix (circular polarization) scattering. polarimetric target decomposition methods are categorized into two types: the first type is coherent decomposition methods that are directly performed on the scattering matrix. Yamaguchi et al. This becomes possible because the received power depends strongly on the actual backscattering process. The second type is incoherent decomposition methods based on the second order statistics of polarimetric information. 2008) suitability as compared to the 3-CSPD method (Freeman and Durden 1998) for glaciated terrain features identification. etc. 2008) decompose the coherency matrix as the incoherent sum of scattering power of a distributed target. DB scattering. Lee and Pottier 2009). Downloaded by [Niigata Daigaku]. Singh and Venkataraman 2009). snow. Therefore. e. Due to the strong spatial and time dependent dynamics of glaciated terrain. the Yamaguchi four-component scattering power decomposition (4-CSPD) method (Yamaguchi et al. Since most of the targets are distributed in natural earth surface.. and requires sensors that are time and weather independent. (3) the decomposed powers correspond to physical scattering mechanisms. Yajima et al.

Study area The location map is shown in Figure 1. This glaciated region includes snow.a. Satopanth and Bhagirath Kharak glaciers are the major glaciers among them in this catchment. covering an area of 21. was launched on 24 January 2006 by the Japan Aerospace Exploration Agency (JAXA). The Alaknanda river catchment. In general. It is well known that L-band microwave signals penetrate through dry snowpack with negligible volume backscatter from snow. we acquired the PALSAR fully polarimetric. on-board the ALOS. the situation becomes different.58 incident angle and nominal pixel spacing (azimuth 6 range) 3. If the moisture exceeds 1% in highly accumulated snowpack. out of which 70. .1 data of 12 May 2007 with 21. The area falls between latitude 308400 N and 308500 N and longitude between 798150 E and 798280 E. Data used In this study. India has many glaciers. originates at the snout of the Satopanth glacier. level 1.5 km long with an average width of 750–850 m. The elevation ranges between 2000 meters above sea level (m.17 km2. snow cover area becomes wet in May (early summer) over Himalayan snow bound area with Figure 1.s.08 km2. Location map of study area. 1990). Lband frequency suffers from attenuation in the snowpack while reflection or backscatter from the snowpack comes out (Abe et al. The Alaknanda river.l.17 and 31.36 (m).54 (m) 6 9. However.a. debris-covered glacier and barren rock targets. if snowpack is wet. Satopanth and Bhagirath Kharak glaciers are shown in Figure 2(a)–(d).22 km2 are covered by the Satopanth and Bhagirath Kharak sub-watersheds.s. 2008). The upper Alaknanda river catchment covers an area of 1544. respectively. It operates in the L-band frequency of 1.Geocarto International 379 Downloaded by [Niigata Daigaku]. Uttarakhand.l) and 7000 m. 3. which is the main tributary of Ganga river.27 GHz (23.6 cm wavelength). [Gulab Singh] at 17:44 14 December 2011 2. PALSAR. single look complex.70 and 107. The Satopanth and Bhagirath Kharak glaciers are approximately 13 and 18. respectively (Nainwal et al.

(a) Cloud free ALOS-AVNIR-2 image of 24 May 2010 (b) Pauli RGB of PALSAR of 12 May 2007 (c) Photo of Satopanth (ST) and Bhagirath Kharak (BK) entrance (d) Photo view of Satopanth glacier. Since the HV components contribute only to volume scattering. the snowpack contains water and is not so transparent for L-band frequency. Snowpack is also heterogeneous with snow grain particle compressed during winter season and has rather high density due to snow accumulation and melting cycle. Since the snowpack on 6 May 2007 is rather wet due to the beginning of snow melting. Downloaded by [Niigata Daigaku]. while the cross-polarization backscatter (HV) remains low as compared to co-polarization. Singh et al.380 G. Figure 2(b) shows the Pauli colour composite image (12 May 2007) which gives the clear information about single scattering (snow cover area over glacier and non- . The magnitude of backscatter depends on the snow density and water content. [Gulab Singh] at 17:44 14 December 2011 Figure 2. The co-polarization backscatters (HH and VV) increases with snow volume. the main polarimetric response from snowpack becomes surface scattering in the L-band (Abe et al. 1990). significant melting. and the depth of snowpack.

DB (dihedral features) and volume scattering (debris-covered glacier) in the study area. SHV are elements of scattering matrix S assuming the reciprocal condition of SHV ¼ SVH The coherency matrix is given as 2 3 T11 T12 T13 ÃT ¼ 4 T21 T22 T23 5 ð2Þ ½TŠ ¼ k:k T31 T32 T33 where *T denotes complex conjugation and transposition. 4. fd. fv and fc are coefficients to be determined.1. The 4-CSPD method divides the measured coherency matrix into four submatrices representing physical scattering mechanisms (Yamaguchi et al. Method and technique Downloaded by [Niigata Daigaku].i denotes ensemble average in an imaging window. [Td]. we can define the scattering vector k as 2 3 SHH þ SVV 1 k ¼ pffiffiffi 4 SHH À SVV 5 ð1Þ 2 2SHV where SHH. 2006. Yajima et al. DB. respectively.Geocarto International 381 glacier). Decomposition method Once scattering matrix S is acquired with fully polarimetric radar. [Tv] and [Tc] are expansion coherency matrices corresponding to surface. SVV. [Ts]. 2008) ½TŠ ¼ fs ½Ts Š þ fd ½Td Š þ fv ½Tv Š þ fc ½Tc Š ð3Þ where fs. The expansion coherency matrix for surface scattering is as follows: 2 3 . [Gulab Singh] at 17:44 14 December 2011 4. and h. volume and helix scattering. The single-bounce scattering model is represented by surface scattering phenomena from slightly rough surface in which the cross-polarized component is negligible.

.

1 bà 0 .

.

4 b j bj2 0 5. j bj ¼ .

SHH À SVV .

< 1: ½Ts Š ¼ ð4Þ .

S .

and man-made targets. etc. ground trees. The expansion coherency matrix for DB scattering is as follows: 2 2 3 . HHþSVV 0 0 0 The DB scattering model is based on the hypothesis of double reflections from right angle structures. DB structure includes road surface–building wall.

.

j aj a 0 .

.

4 aà 1 0 5. jaj ¼ .

SHH þ SVV .

< 1: ð5Þ ½Td Š ¼ .

S À S .

HH VV 0 0 0 .

etc. is as follows: 2 3 0 0 0 ½Tc Š ¼ 4 0 1 Æ j 5: ð9Þ 0 Çj 1 The corresponding scattering powers (the surface scattering power Ps. For the volume scattering model. we choose one of the following matrices according to the magnitude balance of jSHHj2 and jSVVj2 (Yamaguchi et al. Singh et al. which takes into account of non-reflection symmetry condition. 2006). The helix scattering expansion matrix. the volume scattering power Pv and the helix scattering power Pc) are directly obtained from the expansion coefficients when we apply decomposition. are examples of volume scattering. [Gulab Singh] at 17:44 14 December 2011 When j10log (hjSVVj2i/ hjSHHj2i) j 52 dB 2 2 0 1 1 ½Tv Š ¼ 4 0 4 0 0 When 10log (hjSVVj2i/ hjSHHj2i) 72 dB 3 0 0 5: 1 ð7Þ 2 15 5 1 4 5 7 ½Tv Š ¼ 30 0 0 3 0 0 5: 8 ð8Þ Helix scattering power is equivalent to circular polarization power. For the case of j10log (hjSVVj2i/ hjSHHj2i) j 52 dB. subsurface or snow/ice layers. This term appears in urban and mountainous area for L-band data. The decomposition takes into account an imbalance of the co-polarized channel power. Scattering by trees or branches. When 10log (hjSVVj2i/hjSHHj2i) ! 2 dB 2 15 1 4 À5 ½Tv Š ¼ 30 0 À5 7 0 0 3 ð6Þ 0 5: 8 Downloaded by [Niigata Daigaku]. the DB scattering Pd. Volume scattering can be observed if SAR beam penetrates into a medium. the decomposed power expression becomes as (Yamaguchi 2007): Pc ¼ fc ¼ 2jImfT23 gj Pv ¼ fv ¼ 4T33 À 2Pc Ps ¼ fs ð1 þ jbj2 Þ Pd ¼ fd ð1 þ jaj2 Þ: ð10:1Þ ð10:2Þ ð10:3Þ ð10:4Þ .382 G.

Geocarto International 3. Total backscattering power (TP) can be defined as TP ¼ Ps þ Pd þ Pv þ Pc : ð14Þ With the help of Equation (7). K¼p 2 ð11Þ Downloaded by [Niigata Daigaku]. DB backscattering (Pd). we determine that volume scattering is the dominant contribution. an appropriate distance measure. Decomposed power probability The 4-CSPD method decomposes total backscattering power into surface scattering (Ps). if DPs7v is negative. If DPs7v is positive. we can define the probability of surface backscattering and volume backscattering decomposed components as follows: Probability of surface backscattering. can then be calculated according to Bayes maximum likelihood classification as (Lee and Pottier 2009)   dm ð½TŠj½Tm ŠÞ ¼ Tr ½Tm ŠÀ1 ½TŠ þ lnð½Tm ŠÞ ð12Þ thus leading to a minimum distance classification independent of the number of looks used to form the multi-looked coherency matrix [T] (Lee and Pottier 2009): ½TŠ 2 ½Tm Š if dm ð½TŠÞ < dj ð½TŠÞ 8j 6¼ m: ð13Þ 3. Psurface ¼ Ps TP Pv : TP ð15Þ Probability of volume backscattering Pvolume ¼ Therefore from Equations (8) and (9) À1 DPsÀv ¼ Psurface À Pvolume 1 ð16Þ ð17Þ Equation (17) helps us to determine the dominated scattering component from surface scattering and volume scattering in the 4-CSPD method’s decomposed image. On the other hand.3.2. Complex Wishart classifier 383 The CWD is expressed as (Wishart 1928) nnq j½TŠjnÀq eÀnTrð½Tm Š Pð½TŠ=Tm Þ ¼ K½Tm Šn À1 ½TŠÞ qðq À 1Þ ÀðnÞ:::Àðn À q þ 1Þ . we determine that surface scattering is the dominant contribution. . [Gulab Singh] at 17:44 14 December 2011 where n is the number of looks and q is the polarimetric dimension and [Tm] = E[T]. The threshold does not begin from zero because we want to take into account the noise variation in case that both the probabilities are close to zero. volume backscattering (Pv) and helix backscattering (Pc). dm. Using the complex Wishart distribution of the coherency matrix [T].

The sequence of this procedure is followed as given below: (1) First of all. Flow chart of PALSAR image classification. a supervised classification methodology has been presented for fully polarimetric SAR images classification.384 3. Classification procedure G. The flow chart of developed methodology is shown in Figure 3. a multi-looked (six times in azimuth direction and one time in range direction) coherency matrix has been generated.4. . Singh et al. Downloaded by [Niigata Daigaku]. [Gulab Singh] at 17:44 14 December 2011 Figure 3. Using 4-CSPD method and CWD.

Glacier moraine dammed lakes appear in deep blue colour. i. In 3-CSPD and 4-CSPD methods. Otherwise. Most of the differences can be visibly identified in DB and surface scattering component. In 4-CSPD method FCC image. where in. blue to deep blue colour represents surface scattering from snow cover over glacier area (accumulation zone) and permanent snow cover at mountains peaks. Therefore. Training samples have been allotted on the basis of visually comparing four component colour composite image with AVNIR-2 snow cover image and field information. green colour is assigned to volume scattering and blue colour is assigned to surface scattering. 4. It has been seen in Figure 6 that the helix (Pc) scattering component shows high value (greater than 710 dB) at steep slope . The conditional approach is defined as EITHER glacier snow/ice class IF (DPs7v ! 0.e. an additional class (glacier snow/ice class) has been added to supervised classified image. All pixels are classified based on their Wishart distance measure and criteria (Equations (12) and (13)) from class centres. 4-CSPD method gives very sharp information about dihedral features in the study area as compared to 3-CSPD method and Pauli RGB images. The comparison of visual interpretation has been done for both decomposed FCCs with the Pauli RGB (Figure 2(b)) and AVNIR-2 (Figure 4) image. Some of these differences that are indicated by yellow circles no. 2 and 3 in Figure 5(a) (3-CSPD FCC). the probability difference (DPs7v) image has ! 0. the probability difference (DPs7v) has been used for resolving volume scattering ambiguity from vegetation and glacier snow/ice. Furthermore. 1999) with window size 7 6 7 has been applied on coherency matrix. the polarimetric refined Lee filter (Lee et al. (5) Finally. which occurs due to slope surface of target. respectively. based on conditional approach. 1. The differences are clearly seen between 3-CSPD and 4-CSPD methods FCCs in Figure 5 (enlarged part of red colour rectangular area in Figure 4). (4) CWD has been applied on de-speckled coherency matrix and computed the averaged coherency matrices from the assigned classes. the speckle noise from the polarimetric SAR data. Results and discussion The 3-CSPD and 4-CSPD methods have been applied on L-band PALSAR data over the Satopanth and Bhagirath Kharak glaciated region. 1. 3-CSPD and 4-CSPD methods FCC images are shown in Figure 4.05 value but supervised classified image shows vegetation class. Red colour represents DB or dihedral scattering mechanism. FCC images of PALSAR data over Satopanth glacier region (Figure 4).Geocarto International 385 Downloaded by [Niigata Daigaku]. red colour is assigned to DB scattering. the fourth component of four component scattering power decomposition (4-CSPD) method represents the helical scattering phenomenon. These computed mean matrices have been used as the class centres. (3) 4-CSPD method has been applied on de-speckled coherency matrix and 4CSPD method false colour composite (FCC) image has been generated. Debris-covered glaciers (ablation area) are shown in green colour. and these components are clearly exposed (Figure 5(b)) in 4-CSPD method FCC.05 and vegetation class) OR classified classes OTHERWISE. 2 and 3 in Figure 5(b) (4-CSPD FCC) correspond to yellow colour circles no. [Gulab Singh] at 17:44 14 December 2011 (2) For reduction. supervised classified pixels remain same in final classified image.

non-snow and unidentified/layover. But in 4-CSPD method. Classification technique was applied on coherency matrix.g. snow. In other words. which behaves like oriented target from the direction of radar and oriented target does not hold reflection symmetry condition where the 3-CSPD method works. and 3-CSPD model FCC of 12 May 2007 PALSAR data (middle) and 4-CSPD model FCC of 12 May 2007 PALSAR data (bottom) (Red colour rectangular on 4-CSPD model FCCs enlarged view are shown in Figure 5). it corrects for rotation along the line of sight (LOS) while it decomposes. Downloaded by [Niigata Daigaku]. a 4-CSPD method is suitable for fully polarimetric PALSAR data decomposition over Himalayan glaciated terrain as compared to the 3-CSPD method. ALOS PALSAR data was classified into six major classes (e. which causes the over estimation of volume scattering in 3-CSPD method.386 G. Therefore. using 4-CSPD method decomposed image and CWD. Moreover. Himalayan topography has gentle to steep slope. volume scattering is reduced by the fourth component where volume scattering is more than helix scattering component. Figure 7). . [Gulab Singh] at 17:44 14 December 2011 Figure 4. ALOS-AVNIR-2 image of 6 May 2007 (upper). The main reason for the reduction of the volume component for the 4CSPD method is due to the reflection symmetric space where the 4-CSPD method works. and training samples were taken from the four component colour composite image with the help of visual interpretation of 4-CSPD method FCC and AVNIR-2 image (which are shown in Figure 4). Singh et al. and low values (less than 720 dB) are found over snow-covered area.

.Geocarto International 387 Downloaded by [Niigata Daigaku]. Helix scattering (Pc) component of 4-CSPD for PALSAR image of 12 May 2007. PALSAR classified image of 12 May 2007. Figure 6. (a) 3-CSPD Model FCC of 12 May 2007 PALSAR data (b) 4-CSPD Model FCC of 12 May 2007 PALSAR data. [Gulab Singh] at 17:44 14 December 2011 Figure 5. Figure 7.

55 Rock 67 1594 0 72 477 0 2210 72.25 DCG 73 0 3629 0 224 32 3958 91. a probability difference image (DPs7v) has been used for separating these classes.58 Vegetation/ ice 87 282 94 11 3072 0 3546 86. but vegetation class could not be discriminated from glacier snow/ice by using alone complex Wishart classifier with defined training samples. The main diagonal entries of the Table 1 represent the number of pixels that are correctly classified. Double bounce (DB)/ settlement 0 29 0 742 0 0 771 96.24 89. . overall classification accuracy has been found to be 93. while rows represent the labels assigned by the classifier. By this way.86 Row sum 20177 2011 4192 825 3786 767 31758 Overall accuracy¼ 93.33 95. The most common tool used for assessing the classification accuracy is the confusion (or error) matrix (Table 1). By using conditional approach. The probability difference image shows low value over vegetation area and high value over glacier snow/ice area. Singh et al.68 86. The columns in a confusion matrix (Table 1) represent test data that have been collected via field observation and interpretation of 4-CSPD method FCC and ALOS AVNIR-2 image.38%. For further improvement of classified image (Figure 7). Both user’s accuracy and producer’s accuracy of vegetation class are more than 80%.96 Layover 65 0 469 0 13 735 1282 57. [Gulab Singh] at 17:44 14 December 2011 Figure 8.12 79. A confusion matrix composed of six glaciated terrain classes. it is possible to resolve the ambiguity between vegetation class (Figure 7) and glacier snow/ice (Figure 8). G.38% Downloaded by [Niigata Daigaku].63 89.96 Snow Snow Rock DCG DB/settlement Vegetation/ice Layover Column sum Producer’s accuracy (%) User’s accuracy (%) 19885 106 0 0 0 0 19991 99.388 Table 1. Vegetation and ice separating in PALSAR classified image of 12 May 2007 by using probability difference between surface scattering and volume scattering probabilities.46 98.

New York: Oxford University Press. and Sengoku.Geocarto International 5. 2009. IEEE Geoscience and Remote Sensing Letter. 14 (3). Polarimetric radar imaging: from basics to application. 2009. Freeman. 1999. 32–52.. 20A (1–2). E. 34.. Y. Advances in Space Research. but the ambiguity of separating the vegetation from glacier snow/ice has also been found in classified image. Nainwal. India: Thesis (PhD). M.S. Lee. J. A three-component scattering model for polarimetric SAR data. 2008. et al.. Current Science. 1928. IEEE Transaction on Geoscience and Remote Sensing.R.. M. 915–921. In: 2008 International conference of recent advances in microwave theory and applications (Microwave-08). LOS PALSAR data Analysis of snow cover area in Himalayan region using four component scattering decomposition model. and Yamada. IEEE Transactions on Geoscience and Remote Sensing. Yajima. India. 217–226.. dihedral (DB) and glacier features as compared to 3-CSPD method.. Y. Yamaguchi. Taylor & Francis Group. ISBN 978-4-88552-227-7. Japan: IEICE. Polarisation: applications in remote sensing. A. J. E. Biometrika. Lee.. S. using total station survey. 1994. Grues. and Pottier. 1998. Jaipur. 2006.E. H. Boulder.. S. Yajima. In future work.. A review of target decomposition theorems in radar polarimetry.. 653–660. G. A. full polarimetric PALSAR data of high altitude glaciated terrain in Himalayan region has been analysed based on 3-CSPD and 4-CSPD methods and information of various terrain features have been extracted.L. Uttarakhand... 2363–2373.. IEEE Transactions on Geoscience and Remote Sensing. this methodology will be assessed with more time series data to check the resolving capability of the ambiguity between vegetation and glacier snow/ice.. Challenges and recommendations in mapping of glacier parameters from space: results of the 2008 Global Land Ice Measurements from Space (GLIMS) workshop.. 94 (5). 2009. S. POLSAR image analysis of wetlands using a modified fourcomponent scattering power decomposition. India. Therefore.. 2007. 37. G. 21–24 November 2008. Yamaguchi. et al. Indian Institute of Technology Bombay. Singh. Yamaguchi. and Pottier. 53–69. [Gulab Singh] at 17:44 14 December 2011 In this work. The generalised product moment distribution in samples from a normal multivariate population. the probability difference (DPs7v) has been combined with supervised classification procedure to resolve the ambiguity between vegetation and glacier snow/ice. and Durden.. T. 1996. Y. J. USA. Singh. Annals of Glaciology. IEEE Transaction on Geoscience and Remote Sensing. and Venkataraman.R.C. 2010.. Cloude. Racoviteanu. The supervised classification procedure shows overall accuracy of 93. It has been found that the 4-CSPD method discriminates better terrain features such as snow cover. Summary and conclusion 389 Downloaded by [Niigata Daigaku].. Polarimetric SAR speckle filtering and its implication for classification. 498–518. Cloude. Rott... 292–296. Thematic studies in alpine areas by means of polarimetric SAR and optical imagery. SAR polarimetry techniques for snow parameters estimation. Experimental study of microwave transmission in snowpack. References Abe.R. 36 (3). 772–774.. and de Grandi.S. IEEE Transactions on Geoscience and Remote Sensing. Y. A four-component decomposition of POLSAR images based on the coherency matrix. 1667–1673. G. 1990. Y. Colorado. Wishart. 3 (3). H. 28 (5). 2008. Radar polarimetry from basics to applications: radar remote sensing using polarimetric information (in Japanese). H. 2009.38%.. 50 (53). Boca Raton: CRC Press. . 46 (6). Temporal changes in rate of recession: evidences from Satopanth and Bhagirath Kharak glaciers.. 936–973. et al. New York: Institute of Electrical and Electronics Engineers (IEEE). G.

Sign up to vote on this title
UsefulNot useful