Professional Documents
Culture Documents
Singh]
On: 15 July 2013, At: 02:56
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
To cite this article: Geomatics, Natural Hazards and Risk (2013): Fusion of polarimetric channel
information of PALSAR data for land cover classification, Geomatics, Natural Hazards and Risk, DOI:
10.1080/19475705.2013.802260
Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
Geomatics, Natural Hazards and Risk, 2013
http://dx.doi.org/10.1080/19475705.2013.802260
Fully polarimetric radars are capable of preserving detailed information about the
targets because of the amplitude and phase information they contain which helps
in distinguishing different scattering mechanisms. Therefore, nowadays it is
needed to use this information for various applications, i.e., classification, target
identification etc. In polarimetry, circular (L: left-hand circular polarization and
R: right-hand circular polarization) as well as linear (H: horizontal polarization
and V: vertical polarization) polarizations have their own advantages. In this
paper, an attempt has been made to highlight the effect and importance of both
types of polarizations for land-cover classification using ALOS (Advanced Land
Observing Satellite) PALSAR (Phased Array-type L-band Synthetic Aperture
Radar) quad-polarimetric data. Polarization responses from certain targets in lin-
ear (i.e., HH, HV, VV) and circular (i.e., LL, LR, RR) bases have been extensively
analysed. Supervised classification using minimum distance classifier is used to
examine the performance of both the polarization bases for terrain classification.
It has been observed that circular basis polarization may be suitable for better
estimation of water class unlike linear basis polarization and inclusion of circular
basis enhances the classification results. This type of study will be quite helpful in
the near future to enhance the special features of terrain covers.
1. Introduction
Radar, an acronym for Radio Detection And Ranging has been used in a wider
sense, encompassing all the active microwave sensors applied for detecting physical
attributes of remotely located objects. Conventional imaging radar, operating in sin-
gle or dual polarization mode, measures the scattered wave as scalar quantity, i.e.,
amplitude. Therefore, it contains limited information about the target (Van Zyl et al.
1987). However, with the advent of radar polarimetry (polar: polarization, metry:
measure), the task of target identification and classification underwent substantial
improvement over conventional radars. Fully polarimetric radars are capable of
measuring scattering matrix for every small resolution element in an image and thus,
enabling the calculation of scattering cross-section of a scatterer for any assumed
transmit and received polarization combination (Henderson et al. 1998). This infor-
mation may be quite useful for land-cover classification.
Polarimetric Synthetic Aperture Radar (SAR) data provides better classification
results as compared to classification results based on single features, like intensities
categories in data (Lillesand & Kiefer 2000). As stated in review paper of Touzi et al.
(2004), supervised classification method for polarimetric data was based on the use
of multivariate complex Gaussian distribution of the three complex polarimetric
components, i.e., HH, HV and VV. Further works in the same direction led to the
formulation of algorithms which were applicable only for single look processed SAR
data. Covariance Wishart distribution was used by Lee et al. (1994) to extend the
maximum likelihood classifier for multilooked SAR data.
Each polarization effectively describes scattering from certain land cover type
due to dependence of backscattering coefficient on polarization. Fully polarized
data has the advantage that it can easily be converted into any other orthonormal
polarization basis (circular, elliptical etc.) by using special unitary transformation
(Boerner et al. 1992; Henderson et al. 1998). Linear polarization is restricted to a
single plane containing the direction of propagation while for circular polarization
wave radiation is in horizontal plane, vertical plane and all the planes between
them (Raney 2007). Due to this fact, circular polarization helps in investigating tar-
gets having different orientations or alignment with respect to radar line of sight
(LOS). The ability of circular polarizations to enhance the even bounces with right
circular transmit-right circular receive (RR) polarization and odd bounces with
right circular transmit-left circular receive (RL) polarizations has been effectively
in precipitation studies (Antar et al. 1992). Coherence of the RR and left circular
transmit-left circular receive (LL) polarization has been found to be extremely
sensitive to surface roughness (Mattia et al. 1997). Thus, circular polarization can
be an additional source of information if used for classification (Mittal & Singh
2008; Mittal 2010). Limited works are reported on the use of circular polarization
in this context. The present work is thus centred on the use of circular and linear
polarizations for land-cover classification. Information from both linear and circu-
lar bases have been fused to form feature sets. Supervised minimum distance classi-
fier was used for checking the performance of each of the feature sets for land-
cover classification. In order to describe the uncertainty of classification, confusion
matrix was calculated. This matrix provides certain error indicators like producer
accuracy, user accuracy, kappa coefficient, errors of commission and omission.
These indicators are used to define the overall accuracy of classification. The kappa
coefficient has widely been used as a measure of classification excellence. The kappa
coefficient accounts for random chance in accuracy assessment and is used to deter-
mine the significant difference among various confusion matrices (Cohen 1960;
Congalton et al. 2008; Shi 2009). In order to estimate classification accuracies on a
class-by-class basis, a measure called conditional kappa coefficient was evaluated.
It determines the measure of agreement for each individual class within confusion
matrix (Senseman et al. 1995; Congalton et al. 2008; Shi 2009). Thus, in the present
paper kappa coefficient for the entire confusion matrix was calculated for determin-
ing the quality of classification based on measures provided by Landis & Koch
Fusion of polarimetric channel information 3
(1977) and the conditional kappa coefficients were measured to determine which
classes are well mapped (Rossiter 2004).
This paper is divided into six sections. The description of the experimental dataset
is provided in section 2 and section 3 deals with the theoretical formulations involved
for the analysis. Section 4 describes the classification methodology. Results are dis-
cussed in section 5 and conclusion is finally reported in section 6.
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
2. Experimental dataset
2.1. Study area
Roorkee town, which is located on the banks of upper Ganga canal, 30 km from the
district of Haridwar in the Indian state of Uttarakhand, has been chosen as the study
area. The location map and topographic sheet of Roorkee are shown in figure (1).
The area is relatively flat with elevations ranging from 245.5 m to 289.9 m. An image
subset of 300 300 pixels, extending from 29 500 39.98” to 29 550 38.77” N and from
77 510 42.65” to 77 540 42.75” E has been studied.
Figure 1. (a) Location map of Roorkee area (b) Topographical map of the region under
study.
4 G. Mittal et al.
Organization (JAROS). Data used for the study was acquired on 26th March 2007,
at centre frequency of 1270 MHz (23.6 cm) and 23.989 as the incidence angle.
Specifications of the data are mentioned in table 1. The observation mode is fully
polarimetric with nine channels (three real and three complex). Polarimetric data
used for the study is geocoded and externally calibrated using trihedral corner
reflectors (Tri-CRs) and polarimetric active radar calibrators (PARC) as the stan-
dard calibration targets (Shimada et al., 2009). Data is compressed in the azimuth
direction with seven looks. Data is in the form of cross products generated during
the conversion of scattering matrix to Stokes matrix. These cross products are
jSHHj2, jSHVj2, jSVVj2, SHHSHV, SHHSVV and SHVSVV. Despite of expressing the
polarization dependency of the scattering characteristics of the target, scattering
matrix is not suitable for SAR processing as it is inefficient of being processed addi-
tively (ERSDAC 2006). Scattering matrix can be employed to characterize deter-
ministic targets. However, in the absence of dominant scatterers, the statistics of
the response are given by the complex Gaussian scattering model which gives rise
to a phenomenon called speckley. On the contrary, scattering matrix cannot be
employed from polarimetric point of view to characterize distributed targets. This
type of scatterers can only be characterized, statistically, due to the presence of
speckle noise. Since speckle noise must be reduced, only second-order polarimetric
representations can be employed to analyse distributed scatterers (Putignano
2009). Due to this reason, data is to be converted to Stokes matrix. The advantage
of Stokes matrix is that it can be processed additively and expresses the polariza-
tion dependence of the scattering characteristics of the target similar to the scatter-
ing matrix (ERSDAC 2006).
small (22.5 –24 ), therefore, a constant incidence angle is used in the calculation of
sigma naught. During the calculation of the backscattering coefficient s , a mean fil-
ter of window size 3 3 was applied in order to reduce the effect of speckle.
3. Theoretical formulation
3.1. Polarization signature
The polarization signature is a graphical representation of the variation of the
received scattering cross-section as a function of the transmitting antenna polariza-
tion. The polarization response is a plot of received intensity as a function of the
ellipse polarization angles (tilt angle ‘c’ and ellipticity angle ‘x’). Fully polarimetric
radar system measures 2 2 complex scattering matrix [S] samples for each resolu-
tion cell in the image. The matrix S can be used to define 4 4 Kennaugh matrix [K].
PALSAR 4.1 data is in the form of cross products (jSHHj2, jSHVj2, jSVVj2, SHHSHV,
SHHSVVand SHVSVV) and can be used directly for construction of [K] matrix using
the equations as given by Boerner et al. (1992) and Mott (2007). Matrix [K] could be
used further to produce co- and cross-pol signatures using equation (2) (Qong 2004):
4p r
s ¼ g~ ½Kgt ð2Þ
k2
where s is the backscattering coefficient, k is the wave number and g is the normal-
ized Stokes vector defined as
2 3 2 3
1 1
6 cos xr coscr 7 6 cos xt cosct 7
gr ¼ 6 7
4 cos xr sincr 5; gt ¼ 6 7
4 cos xt sinct 5 ð3Þ
sincr sinct
where subscript (or superscript) ‘t’ denotes ‘transmit’, while ‘r’ denotes ‘receive’ chan-
nels. For co-polarized signatures the conditions (xr ¼ xt) and (cr ¼ ct) while for
cross-polarized signature the conditions (xr ¼ xt) and (cr ¼ ct þ 90 ) would be
satisfied.
circular basis using basis transformation (Boerner et al. 1992). Elements of scattering
matrix [S] in (L, R) basis is related to matrix elements in (H, V) basis as given by
Hajnsek (2001):
1
SLL ¼ ðSHH SVV þ j2SHV Þ ð4Þ
2
1
SLR ¼ ½jðSHH þ SVV Þ ð5Þ
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
2
1
SRR ¼ ð SHH þ SVV þ j2SHV Þ ð6Þ
2
4. Classification methodology
The flow chart given in figure (2) shows the various steps involved in the supervised
classification. Geo-coded and calibrated PALSAR data is used as the input for classi-
fication purpose. Geo-correction of the image is done prior to radiometric calibra-
tion. The backscattering coefficient for the input images is computed after geo-
referencing. Linear basis components, i.e., HH, HV and VV are directly extracted
from the sigma naught images. Circular basis components, i.e., LL, LR and RR are
extracted from the Kennaugh equations in circular basis. In linear basis, HH compo-
nent enhances dihedral scattering due to small correlation lengths, VV contributes to
Bragg scattering whereas HV contributes to diffuse scattering (Pierce et al. 1994;
Baronti et al. 1995; Henderson et al. 1998; Hallberg et al. 2008). In circular basis,
LL is sensitive to surface roughness when used in coherence with RR (Mattia et al.
1997). Double bounce scattering is enhanced by RR component and RL enhances
single bounce scattering (Touzi et al. 2004; Hallberg et al. 2008). These six elemen-
tary bases (HH, HV, VV, LL, LR and RR) are used to form 11 different combina-
tions (HH–HV; HH–VV; HV–VV; LL–LR; LL–RR; HH–HV–VV; LL–LR–RR;
HH–HV–VV–LL; HH–HV–VV–LL–LR; HH–HV–VV–LL–LR–RR). The combi-
nations thus obtained are used as 11 different feature sets and a supervised minimum
distance classifier is used for segmenting each of the 11 images into three distinct clas-
ses as urban, water and agriculture.
of 180 between the two electric-field components. It closely resembles the theoretical
co-polarized and cross-polarized signatures of an odd bounce scatterer, a dielectric
plate as shown by Dong et al. (1998). However, in LR basis, peaks are observed at x
¼ 0 and c ¼ 90 (figure 3b1). Co-pol signature in (H, V) basis for urban area
shows a ridge-like shape (figure 3a3), with a peak at x ¼ 0 and c ¼ 10 , while
cross-pol signature in circular basis shows maximum power return at x ¼ 0
and c ¼ 45 which signifies that scatterers are inclined at an angle of 45 with the
radar LOS (figure 3b4). The signature closely resembles the theoretical cross-pol sig-
nature of the double bounce scatterer (Dong et al. 1998). Circular basis suggests that
even bounce scattering is dominant in the urban area, however, it was not interpreted
from the signature in linear basis. Polarization signatures for sugarcane, a low-
density crop, show (figure 3a5) maximum power at x ¼ 0 and c ¼ 5 . Co-pol sig-
nature plotted in circular basis has maximum power return at x ¼ 0 and c ¼ 0
8 G. Mittal et al.
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
Figure 3. Linear co- and cross-pol signatures are presented in Figure ai and aj and circular co
and cross-pol signatures are presented in Figure bi and bj, respectively (where i ¼ 1, 3, 5 and
j ¼ 2, 4, 6). Figure a1, a2, b1 and b2 represents water; figure a3, a4, b3 and b4 represents urban;
and figure a5, a6, b5 and b6 represents grassland sequentially.
(figure 3b5). Cross-pol signature for sugarcane shows the presence of dihedral scatter-
ing resulting from stem-ground scattering (figure 3b6).
Linear polarization is confined to a single plane containing the direction of prop-
agation while for circular polarization wave radiation is in both horizontal and ver-
tical planes and all the planes between them (Raney 2007). Thus, circular
polarization is found more effective in examining the targets having different orien-
tations or alignment with respect to radar LOS. As for the urban area, signatures
in linear basis could not identify the double bounce scattering due to a different
Fusion of polarimetric channel information 9
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
Figure 3. Continued
alignment with respect to LOS but that was clearly predicted through circular
polarization.
grassland areas under consideration are almost barren due to which backscattering
coefficient was found more for these areas with least standard deviation of 2.9 dB.
Backscattering coefficient values for grassland is lower for circular channels as com-
pared to linear ones. Values of backscattering coefficient for urban and agriculture
areas are approximately close to each other for image channels in linear basis. The
backscattering coefficient for an urban area is highest for RR channel. At channels
HH, HV and VV there is maximum of 1 dB difference between classes urban, sugar-
cane and wheat. For channel LL, urban class was found to be alienated from the
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
Figure 4. Response of backscattering coefficient with various land cover for different
polarization.
Fusion of polarimetric channel information 11
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
Figure 5. Input RGB composite image (a) Ortho-polar (b) Ortho-Stokes image.
sub-setted from the input images, are chosen for the purpose. There are two input
images as shown in figure 5(a) and 5(b). Image in figure 5(a) is a colour composite
image of jSHHj2 (red), jSHVj2 (green) and jSVVj2 (blue) and the image in figure 5(b) is
a colour composite image of SHHSHV (red), SHHSVV (green) and SHVSVV (blue).
Supervised classification using the minimum distance classifier is performed in order
to test the role of both the polarization bases (i.e., linear and circular) in the identifi-
cation and classification of various land covers. Total 1524 ground control points are
collected with the help of toposheet of the study area. Around 266 ground control
points are used for training of the minimum distance classifier. Remaining 1258
points are used for testing the performance of the classifier by calculation of overall
classification accuracy. Unclassified class has been masked in the classification proce-
dure. The overall accuracy of each of the six basis components, their kappa coeffi-
cients and accuracies in estimating individual classes of agriculture, water and urban
are summarized in table 2. According to Landis and Koch (1977), kappa coefficient
value less than 0 indicates no agreement, 0–0.20 as slight agreement, 0.21–0.40 as fair
agreement, 0.41–0.60 as moderate agreement, 0.61–0.80 as substantial agreement,
Table 2. Producer’s accuracy for each class along with overall accuracy at each of the
individual basis.
Producer accuracy
HH 75 0.63 79 71 76
HV 66 0.49 76 74 35
VV 73 0.61 78 75 64
LL 52 0.22 57 70 19
LR 73 0.62 78 69 70
RR 61 0.27 86 46 38
12 G. Mittal et al.
Table 3. Conditional kappa coefficient for each class at each of the individual basis.
and 0.81–1 as almost perfect agreement in classification. Based on this concept, it was
observed that all the classification results belong to slight-to-substantial-agreement
category (for kappa coefficient in between 0.2 to 0.8). HH component shows the
highest overall accuracy of 75% among the six elementary bases. Overall accuracy
for the VV and LR bases is approximately the same, i.e., 73%. However, the VV basis
appears to be more capable of classifying the urban class, with producer accuracy of
75%, as compared to the LR basis that has producer accuracy of 69% for the urban
class. LR basis is found to be more efficient in the classification of the agriculture
class, with producer accuracy of 70%, as compared to the VV basis having producer
accuracy of 64% for agriculture. Though the overall accuracy for the LL basis is only
52%, it has good performance in urban estimation, with producer accuracy of 70%.
RR basis shows an overall accuracy of 61%, but it gives the highest producer accu-
racy of 86% in the estimation of water. Table 3 shows conditional kappa coefficient
for each class. It helps in determining the accuracy by which each class is mapped by
certain feature. Class water is well mapped by all the features except LL. Class urban
is well mapped by all the features except LL and RR. Class agriculture is well
mapped by only two features HH and LR. A closer look at the performance of all
the bases makes it obvious that components with low overall accuracy are effective
in estimating certain class with higher producer accuracy as compared to compo-
nents with higher overall accuracy. Therefore, the six elementary bases are combined
together to form 11 feature sets with emphasis laid on their accuracy estimation for
individual class and not on the overall accuracy.
5.3.2. Classification results after fusion of different channels. The overall accuracy
of all the features, kappa coefficient and producer accuracy for individual classes are
given in table 4. By observing kappa coefficient of agreement it was found that the
classification result of all the feature sets are in good agreement with the estimated
kappa coefficient, being more than or equal to 0.4 for all the features. Overall accu-
racy changes with an increase in the number of features. The interesting results can
be interpreted by taking a closer look into the producer accuracies computed for
each of the classes. Feature set 1, i.e., HH–HV has overall accuracy of 76% and it
estimates all the three classes with moderate accuracy. Feature sets 2 and 3 show
approximately same overall accuracy, with 74% being for feature set 2 (HH–VV) and
75% for feature set 3 (HV–VV). However, agriculture class is not well mapped by fea-
ture set 3. By combining the three bases as feature set 7 (HH–HV–VV), an increase in
overall accuracy of 82% is observed. An increase in producer accuracy of 1%–2% is
Fusion of polarimetric channel information 13
Table 4. Producer’s accuracy for each class along with overall accuracy at each of the selected
basis combinations.
Producer accuracy
Feature Overall Overall
set no. Features accuracy kappa Water Urban Agriculture
1 HH–HV 76 0.65 79 73 75
2 HH–VV 74 0.65 77 71 74
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
3 HV–VV 75 0.64 80 79 60
4 LL–LR 66 0.48 62 77 57
5 LL–RR 68 0.51 81 50 47
6 LR–RR 65 0.40 89 50 47
7 HH–HV–VV 82 0.67 81 74 75
8 LL–LR–RR 71 0.56 83 55 73
9 HH–HV–VV–LL 75 0.61 70 82 73
10 HH–HV–VV–LL–LR 75 0.62 71 80 81
11 HH–HV–VV–LL–LR–RR 75 0.62 82 65 75
also observed for water and agriculture classes. Table 5 shows that all the classes are
well estimated by feature set 7. Feature set 4 (LL–LR), feature set 5 (LL–RR) and
feature set 6 (LR–RR) have an overall accuracy of 66%, 68% and 65%, respectively.
In spite of having low overall accuracy, feature set 6 (LR–RR) shows the highest pro-
ducer accuracy of 89% for the water class. Amongst the three mentioned feature sets,
feature set 4 (LL–LR) estimates urban class well with the producer accuracy of 77%.
However, agriculture class is estimated with low accuracy. By combining the three
circular bases as feature set 8 (LL–LR–RR), an increase in the overall accuracy as
compared to that of feature sets 4, 5 and 6 is observed. In addition to it, the agricul-
ture class is estimated with a good accuracy of 73% and water class with 83% (higher
than that of feature sets 1, 2, 3 and 7). On combining circular basis LL, LL–LR and
LL–LR–RR to feature set 7 as feature set 9, 10 and 11, respectively, an overall
Table 5. Conditional kappa coefficient for each class at each of the selected basis
combinations.
accuracy of 75% is observed in each case. However, a look at the producer accuracy
for each class reveals that feature sets 9 and 4 give the best estimation for the urban
class, feature set 10 for the agriculture class and water is best estimated by feature set
5. Figure 6(a)–(c) shows the classified images for HH–HV–VV, LL–LR–RR and
HH–HV–VV–LL–LR–RR, respectively.
The conditional kappa coefficient for each class is listed in table 5. It indicates the
classification accuracy of individual class (Senseman et al. 1995). It can be visualized
from table 5 that class water is well mapped with conditional kappa estimate of 0.79
as compared to classes urban and agriculture, with conditional kappa estimate of
0.57 for both the classes. Like Senseman et al. (1995), we can state that for feature
set 1, assignment of pixels to class water is 79% more accurate as compared to that of
random assignment of pixels to one of the three classes. This table shows that the
water class is well mapped by all the features as compared to classes urban and agri-
culture. Classification of urban is less accurate as compared to class water with the
conditional kappa coefficient of more than 0.4 for all the feature sets except feature
sets 5 and 6. Class agriculture is well mapped by all the feature sets except feature
sets 3, 4 and 6, which have conditional kappa coefficient of less than 0.4.
6. Conclusion
In the present work, critical analysis of circular as well as linear bases is done for the
fully Polarimetric PALSAR data. PALSAR data is initially in linear basis. It is trans-
formed to circular basis using mathematical formulation. Polarization signatures are
obtained using Kennaugh matrices in both the bases. Circular basis provides better
estimation of the water class and linear basis provides better estimation of the urban
class. By combining both the bases as feature set 11 (i.e., HH–HV–VV–LL–LR–
RR), good results for agriculture, urban and water were observed. It may thus be
concluded that circular basis plays a significant role in land-cover classification.
Future work will be focused on incorporating circular polarization for classifying
more land covers.
Fusion of polarimetric channel information 15
Acknowledgement
Authors are thankful to Space Application Centre, Ahmedabad, India and Depart-
ment of Earth Sciences, India for providing funds for this study.
References
Antar YMM, Hendry A, McCormick GC. 1992. Circular polarization for remote sensing of
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
Pierce LE, Ulaby FT, Sarabandi K, Dobson MC. 1994. Knowledge-based classification of
polarimetric SAR images. IEEE Trans on Geoscience and Remote Sens. 32(5):
1081–1086.
Putignano C. 2009. PolSOM and TexSOM in polarimetric SAR classification [PhD thesis].
Rome (Italy): TOR Vergata University.
Qong M. 2004. Polarization state conformation and its application to change detection in
polarimetric SAR data. IEEE Geoscience and Remote Sens Lett. 1(4):304–308.
Raney RK. 2007. Hybrid polarity SAR architecture. IEEE Trans on Geoscience and Remote
Sens. 45(11):3397–3404.
Downloaded by [Indian Institute of Technology Roorkee], [Dharmendra Singh] at 02:56 15 July 2013
Rossiter DG. 2004. Technical note: statistical methods for accuracy assesment of classified the-
matic maps. Enschede (NL): International Institute for Geo-information Science &
Earth Observation (ITC).
Senseman GM, Bagley CF, Tweddale SA and Construction Engineering Research Laborato-
ries (US). 1995. Accuracy assessment of the discrete classification of remotely-sensed
digital data for landcover mapping. Champaign (IL): US Army Corps of
Engineers/Construction Engineering Research Laboratories.
Shimada M, Isoguchi I, Tadono T, Isono K. 2009. PALSAR radiometric and geometric cali-
bration. IEEE Trans on Geoscience and Remote Sens. 47(12):3915–3932.
Shi W. 2009. Principles of modeling uncertainties in spatial data and spatial analyses. Boca
Raton, FL: CRC Press.
Touzi R, Boerner WM, Lee JS, Lueneburg E. 2004. A review of polarimetry in the context of
synthetic aperture radar: concepts and information extraction. Can J Remote Sens.
30(3):380–407.
Van Zyl JJ, Zebker HA, Elachi C. 1987. Imaging radar polarization signatures: theory and
observation. Radio Sci. 22(4):529–543. Available from: http://earth.esa.int/polsarpro/
Manuals/4_polarimetric decompositions.pdf