FULLY POLARIMETRIC ALOS PALSAR DATA APPLICATIONS FOR SNOW AND ICE STUDIES
G. Venkataraman1, Gulab Singh1, and Y. Yamaguchi2
1
Centre of Studies in Resources Engineering, Indian Institute of Technology Bombay, Mumbai, India
2
Department of Information Engg., Niigata University, Japan
Email-ID: [email protected] ; Phone.No.91-22-25767686
numerous small sized glaciers also and the
neighborhood falls between latitude 30o 30’ N and 31o
15’ N and longitude between 79o 15’ E and 79 o 30’ E.
The elevation ranges between 2000 m.a.s.l (Meters
Above Sea Level) to 7000 m.a.s.l. The Alaknanda
River, which is the main tributary of Ganga River,
originates at the snout of the Satopanth Bank glacier. In
this study, the fully polarimetric L-band ALOS
PALSAR data (acquisition date was May 12 and Nov.
27, 2007) have been used for snow classification over
Badrinath area, Uttarakhand in Indian Himalayan snow
covered region. ALOS-AVNIR data (acquisition date,
May 6, 2007) is used to interpret snow covered area and
non snow covered area and it helps in the selection of
the training sample of different features for supervised
classification.
Abstract: In this study, the capability assessment of fully
polarimetric L-band ALOS PALSAR data has been
carried out for snow discrimination from other targets.
Eigenvaluve based polarization fraction value has been
determined for assessing the capability of PALSAR data
for snow discrimination. Radar snow index has been
developed using polarization fraction and normalized
third eigenvalue of coherency matrix. It has been found
that radar snow index is more robust and simple to
implement that supervised classification.
I.INTRODUCTON
Mapping of snow and ice covered areas is important for
many applications such as prediction of floods,
snowmelt runoff modeling, water supply for irrigation
and hydropower stations, weather forecasts and
understanding climate change. Optical and nearInfrared (IR) remote sensing techniques are proved to
be promising for snow cover mapping. But, in the
presence of cloud cover and different weather
conditions optical and IR fail in acquiring snow cover
information. Microwave remote sensing has an
advantage over optical and IR techniques due to its all
weather capability, penetration through cloud and
independence of sun illumination. Various methods are
available for wet snow cover mapping using multitemporal SAR data. Due to high penetration capability
and low attenuation of dry snow cover at intermediate
frequency, dry snow cover behaves like transparent
media. Hence, discrimination of total (wet + dry) snow
cover using intermediate and low frequency SAR data,
still remains an active research field. The fully
polarimetric SAR data does contain more information
than the corresponding single or dual polarization SAR
data. Fully polarimetric data gives an optimization of
the polarimetric contrast and other polarimetric
parameters which may be very useful for accurate target
discrimination between snow and non-snow covered
areas.
III. METHODS AND TECHNIQUES
In this investigation, H/A/Alpha [1], three scattering
component [2] and four scattering component
decomposition model [3] have been applied on L-band
fully polarimetric ALOS-PALSAR data for extracting
the desired information. Using Wishart classifier,
ALOS-PALSAR data have been classified into major
distinct classes namely snow cover, vegetation, debris
covered glacier, rock and layover/unidentified areas.
Radar Snow Index (RSI) has been introduced for
accurate discrimination of snow from others targets.
IV. RSI DEVELOPMENT
The proposed RSI algorithm flow chart for
discriminating snow from other features is shown in
Fig.1. Eigenvalue based polarization fraction image
(Fig. 2(a)) shows that the snow cover area has high
polarization fraction (PF) value as compared to other
targets, which indicates that the return signal is
polarized or can be said that L- band ALOS PALSAR
data is capable of discriminating snow from other
targets. Therefore polarization fraction parameter can
be used for snow discrimination from other targets. Fig.
2(a) shows high PF value over lower central part of
image but this part of image is snow free and it is
covered by layover affected area due to low incident
angle of ALOS-PALSAR and high topography of
Himalaya. Hence there is need to explore supporting
parameters to discriminate snow cover properly. In this
investigation, eigenvalue images of coherency matrix
II. STUDY AREA AND DATA USED
The present study confines to Badrinath region in
Himalayas, and this test site comprises extensive snow
cover areas having various degrees of glaciations that
act as a huge fresh water reservoir. Snow cover in the
Himalayan region play an important role in the Earth’s
climate system. The test site covering Panpatia,
Satopanth, Bhagirath Kharak, Suraji Bank glacier and
978-1-4244-9566-5/10/$26.00 ©2010 IEEE
1776
IGARSS 2010
cover area and showing very high value for layover
affected area in the image. For selected portion of
image which is affected by topographic distortion, the
normalized Ȝ1 is varying in low range (0 to 0.56) as
compared to Ȝ2 range (0 to 0.80) and Ȝ3 (0 to 1) range,
and hence it is not able to discriminate snow from other
targets. Normalized Ȝ2 range is mixing with layover
affected area in the upper part of image, whereas
normalized Ȝ3 have wide range from 0 to 1 (Fig. 2(b)),
which is able to separate the affected area and other
features. Hence, Normalized Ȝ3 has been used to
support the discrimination of snow using fractional
polarization value. Thus, Radar Snow Index (RSI) has
been developed using eigenvalues of coherency matrix
of fully polarimetric SAR data and these eigenvalues
based generated polarization fraction (PF).
have been generated and based on these images other
parameters namely entropy (H), anisotropy (A),
scattering mechanism angle (alpha), (1-H) and H (1-A),
Lunenburg anisotropy, Radar Vegetation Index (RVI),
SERD, DERD, PPD, co-polarization coherence have
been studied. These parameters have been analyzed to
find out which one can support the results of snow
discrimination obtained through polarization fraction.
Entropy, H(1-A), Lunenburg anisotropy, and RVI show
low values over the snow cover due to pure target and
high value over the other distributed targets (e.g.
ablation area of debris covered glacier and vegetation).
On the other hand anisotropy and asymmetry give high
value over snow area. All of these parameters are also
capable of discriminating snow from other targets but
do not have a proper range to suppress layover affected
snow free area from snow. Shannon entropy shows the
capability to suppress distorted area in the classified
image but it failed to discriminate snow from other
targets in some parts of image.
1.0
1.0
ALOS PALSAR Quad
Polarization SLC Data
0.50.
5
Extract Scattering Matrix(S)
Multi-Looked (6×1) in (Azimuth
× Range)
and make Coherency Matrix (T3)
(HV§VH)
0.00.
(a)
Fig.2 (a) PF (b) Normalized Ȝ3
Polarimetric Speckle Filtering
IV. RESULTS
Snow cover area has high polarization fraction value as
compared to other targets. The user accuracy of snow
classes is also higher than other classes. Hence full
polarimetric ALOS-PALSAR data is capable of snow
cover monitoring [4]. RSI, H-A-Alpha-Wishart
supervised classifier and 4-component decomposition
and Wishart supervised classifier scheme has been
applied on full polarimetric PALSAR data of May 12,
2007. Comparison between RSI, H-A-Alpha-Wishart
Supervised and 4-Component Decomposition-Wishart
Supervised Classifier has been done for the capability
of snow discrimination. Fig. 3 represents total snow
covered and snow free area. It has been noticed that RSI
based snow discrimination well agrees with snow
discrimination through H/A/Alpha Wishart supervised
classifier, and four scattering component and Wishart
supervised classifier.
The important aim of this exercise is to compare the
snow cover maps obtained through RSI, H-A-AlphaWishart supervised classifier and 4-component based
Wishart supervised classifier with AVNIR-2 image
derived snow cover map. For this purpose a subset of
Generate Eigenvalues Image
(Ȝ1, Ȝ2, Ȝ3)
Generate PF value image
0 d PF 1
3O3
d1
O1 O2 O3
PF•0.5
&& Ȝ3<0.015
YES
(b)
NO
Snow Free Area
Snow Cover Area
Fig.1. Flow chart of RSI development.
Eigenvalues Ȝ1, Ȝ2, Ȝ3 of coherency also show sufficient
range for suppressing this unwanted area from the snow
1777
about 34 km2 area in the vicinity of Chirbitiya glacier
has been chosen as this area is totally free from cloud
cover in AVNIR-2 image. However, the optical data
and PALSAR data have a temporal difference of 6
days. As this study area lies in the higher reaches of
Himalayas (Tibet), it can be safely assumed that there
may not be any significant difference in the snow
cover extent, particularly in early summer period
when melting would not have started in these high
altitudes.
RSI, H-A-Alpha-Wishart supervised classifier and 4component decomposition based Wishart supervised
classifier scheme have been applied on full
polarimetric PALSAR data of the study area of May
12, 2007. ALOS-AVNIR-2 image of May 6, 2007
covering Chirbitiya glacier is shown in Fig. 4(a).
Snow classified map of AVNIR-2 is presented in Fig.
4(b). Fig. 4(c) represents RSI classified total snow
cover area over Chirbitiya glacier. Total
discriminated snow-covered area through 4component decomposition based Wishart supervised
classification in the vicinity of Chirbitiya glacier is
presented in Fig. 4(d).Snow cover classified by H-AAlpha-Wishart supervised classification is shown in
Fig. 4(e).
is a difference of about 6.3 km2 snow cover area
between AVNIR-2 and RSI maps. AVNIR-2 data
discriminates more snow cover area than RSI map.
This may be attributed to the intrinsic character of
SAR sensor which causes significant layover and
shadow effects due to the incident angle and
topographic relief of the terrain.
Fig. 4 (a). ALOS AVNIR-2 image of Chirbitiya glacier for
06-05-2007
Chirbatiya
Glacier Area
Snow
Snow Cover
Snow Free
(a)
(b)
(c)
Fig. 3 Total snow discriminated based on (a) H/A/AlphaWishart supervised (b) RSI and (c) 4-component
decomposition and Wishart supervised classification
techniques
Layover-Shadow
/Snow Free (LS/SF)
(b)
(c)
Fig. 4.(b). Snow cover map from ALOS-AVNIR-2 image
(c) RSI based snow cover map
Table 1 shows the classification performance of H-AAlpha-Wishart supervised classifier, 4-component
decomposition based Wishart supervised classifier
and RSI with reference to the mapping of snowcovered area. RSI provides a slightly higher snow
cover area as compared to other two techniques.
Based on visual analysis, it is observed that RSI snow
cover map seems to be more or less matching with
the AVNIR-2 snow cover map. Thus, RSI algorithm
based on PF and normalized third eigenvalue of
coherency matrix, proves to be more reliable and
robust than the supervised classification methods.
RSI method does not require training sets and
topographic information for snow discrimination
from other targets. RSI map exhibits more snow
covered area than the other two classifications. There
Snow Cover
Layover-Shadow
/Snow Free (LS/SF)
(d)
(e)
Fig. 4 (d) 4-component decomposition using Wishart
supervised classification based snow cover map (e) H-AAlpha-Wishart supervised classification based snow cover
map
1778
Table 4.9 Area statistics of snow-cover in the vicinity of
Chirbitiya glacier
cover mapping. In addition, RSI does not require any
training sample and topographic information for
snow discrimination from other targets.
ACKNOWLEDGEMENT
Authors would like thank to JAXA for providing data of
ALOS-PALSAR and ALOS-AVNIR under project JAXA342.
REFERENCES
[1].
S. R. Cloude and E. Pottier, “An entropy based
classification scheme for land applications of
polarimetric SAR,” IEEE Transactions on
Geoscience and Remote Sensing, Vol 35(1), pp.
68-78, 1997.
[2]. A. Freeman and S.L Durden, “A threecomponent scattering model for polarimetric
SAR data,” IEEE Trans. Geosci. Remote Sens.,
vol. 36 (3), pp. 936–973, 1998.
[3]. Y. Yamaguchi, Y. Yajima, and H. Yamada, “A
four-component decomposition of POLSAR
images based on the coherency matrix,” IEEE
Geosci. Remote Sens. Lett., vol. 3 (3) pp. 292–
296, 2006.
[4]. G. Singh , G. Venkataraman , V. kumar, Y.S.
Rao and Snehmani, “The H/A/Alpha polarimetric
decomposition theorem and complex wishart
distribution for snow cover monitoring,” Proc. of
IEEE IGRASS’08, vol.4, pp1081-1084, 2008.
V. CONCLUSIONS
In this Investigation, a novel Radar Snow Index (RSI)
is introduced by integrating decomposed parameters.
It has been observed that RSI maps a slightly more
snow covered area over Chirbatiya glacier as
compared to other two supervised techniques viz. HA-Alpha-Wishart supervised classifier and 4component
decomposition-Wishart
supervised
classifier. Thus, this novel RSI classification
algorithm of full polarimetric SAR data based on PF
and normalized third eigenvalue of coherency matrix
(which represents or highlights the more distributed
targets or statistical disorder of targets), is observed
to be more robust and simple to implement than
supervised methods like H-A-Alpha-Wishart
supervised
classifier
and
4-component
decomposition-Wishart supervised classifier for snow
1779