13 Pages
English
Gain access to the library to view online
Learn more

A user friendly statistical system for polarimetric SAR image classification

-

Gain access to the library to view online
Learn more
13 Pages
English

Description

Resumen
En este artículo se presenta un sistema para la clasificación de imágenes SAR polarimétricas. Este sistema utiliza información contextual a través de un modelo Markoviano para las clases, además de modelos estadísticos para los datos. El sistema fue desarrollado pensando en el usuario y, por lo tanto, está íntegramente basado en interfaces gráficas. Toda vez que el usuario trata de activar una operación inválida, el sistema le informa la secuencia correcta de pasos. La funcionalidad del sistema se verifica clasificando áreas de cultivo, en una ima-gen SIR-C/X-SAR.
Abstract
This article presents a system for polarimetric SAR image classification. This system uses contextual information through a Markovian model for the classes, besides a statistical model for the data. It is developed with the user in mind and, therefore, it is solely based on graphic user interfaces. The user is prompted with the correct sequence of steps when-ever an invalid option is invoked. The functionality of the system is checked classifying a SIR-C/X-SAR image, where mainly crops are observed.

Subjects

Informations

Published by
Published 01 January 1998
Reads 17
Language English
Document size 1 MB

Exrait

Revista de Teledetección. 1998
A user friendly statistical system for polarimetric
SAR image classification
A. H. Correia*, c. da Costa Freitas*, A. c. Frery** and S. J. S. Sant' Anna*
*Instituto Nacional de Pesquisas Espaciais, Divisao de Processamento de lmagens, CP 525, 12201-970 Sao José dos
Campos, SP - Brazil ({correia;corina;sidnei}@dpi.inpe.br)
**Universidade Federal de Pemambuco, Departamento de Informática, CP 7851,50732-970 Recife, PE - Brazil
(frery@di.ufpe.br)



RESUMEN ABSTRACT
En este artículo se presenta un sistema para la This article presents a system for polarimetric SAR
clasificación de imágenes SAR polarimétricas. Este image classification. This system uses contextual
sistema utiliza información contextual a través de information through a Markovian model for the
un modelo Markoviano para las clases, además de classes, besides a statistical model for the data. It is
modelos estadísticos para los datos. El sistema fue developed with the user in mind and, therefore, it is
desarrollado pensando en el usuario y, por lo tanto, solely based on graphic user interfaces. The user is
está íntegramente basado en interfaces gráficas. prompted with the correct sequence of steps when-
Toda vez que el usuario trata de activar una opera- ever an invalid option is invoked. The functionality
ción inválida, el sistema le informa la secuencia of the system is checked classifying a SIR-C/X-
correcta de pasos. La funcionalidad del sistema se SAR image, where mainly crops are observed.
verifica clasificando áreas de cultivo, en una ima-
gen SIR-C/X-SAR.

PALABRAS CLAVE: Clasificación, Contexto, KEY WORDS: Classification, Context, Synthetic
Estadística, Radar de apertura sintética. aperture radar, Statistics.



the electromagnetic signal in both horizontal and INTRODUCCION
vertical polarisation and, thus, they mar carry a
The intensification of remote sensing studies in larger amount of information than that available
the field of Synthetic Aperture Radar (SAR) imag- from a single component. Though there is cur-
ing sensors is leading towards a better understand- rently no sensor operating in different bands and
ing of the scattering mechanisms of terrestrial polarisations, studies in this area are useful.
targets in the microwaves spectrum. Besides this, it Several works are devoted to the statistical char-
has led to more dependable applications of SAR acterisation of single-look polarimetric SAR data.
imagery and products to geology, cartography, and The reader is referred to DeGrandi et al. (1992),
other fields of knowledge. Kong, (1988), Lim et al. (1989), Quegan and Rho-
One of the most useful products of digital im- des (1995), Yueh et al. (1989), to name a few.
ages is the result of automatic or semiautomatic The potential of multilook polarimetric data,
data classification. This product is becoming more where each value is the mean over several observa-
and more precise since the Gaussian hypothesis tions, is notorious as presented in Lee and GcuDes
was weakened, and since better suited distributions (1994) and in Lee et al. (1995), for instance. The
for SAR data were incorporated into the process statistical properties of this kind of data have not
(Nezry et al., 1996; Frery et al., 1997a). been fully exploited yet They have the advantage
In Vieira (1996) this improvement becomes evi- of exhibiting a speckle noise reduction as well as
dent: it is shown that for monospectral SAR data, data reduction. The disadvantage is the resolution
the simultaneous use of proper distribution for loss.
each class, along with contextual information, Given the potentiality of polarimetric data for
leads to better classifications than those obtained image classification, there is a need for systems
either by Gaussian fitting and/or by pointwise that use all the information of polarimetric data, in
classification. On the other hand, the use of mono- a manner that the user can handle it easily without
spectral SAR data has its limitations. knowing too much about the complexity of the
The number of studies and applications involv- underlying theory. The authors of this paper have
ing polarimetric SAR data is increasing steadily. no knowledge of such a system implemented in
These data are formed by sending and receiving commercial software.
Nº 10 – Diciembre 1998 1 de 13 A. H. Correia, c. da Costa Freitas, A. c. Frery and S. J. S. Sant' Anna
The objective this paper is to present a system STATISTICAL PROPERTIES OF
for multilook polarimetric SAR image classifica- POLARIMETRIC SAR DATA
tion which was developed to assess the potential of
Data obtained with coherent illumination, as is this kind of data. The system is strongly based on
the case of SAR data, are corrupted by a signal-the statistical properties of the data, and it uses a
dependent noise called speckle. A usual model for Maximum Likelihood (ML) classification as the
the signal and this noise is the Multiplicative initial configuration for a contextual Markovian
Model. It states that, under certain conditions (Tur classification technique: the Iterated Conditional
et al., 1982) the observed value in every pixel is Modes (ICM for short), presented in Vieira (1996).
the outcome of the random variable Z = XY, where The system present in here allows the analysis of
X is the random variable that models the backscat-intensity, phase difference, ratio of intensities and
ter and Y is the one that models the speckle noise, intensity-phase data. These data formats are de-
and fuese last two variables are independent. rived from multilook polarimetric SAR imagery,
Statistical models for multilook polarimetric data and their distributional properties are here recalled.
are derived from the covariance matrix, which The system is based on graphic user interfaces, and
exhibit a complex Wishart distribution (Lee and was developed as an extension of the ENVI (Envi-
Grunes, 1992; Du and Lee, 1996). ronment for Visualizing Images) image processing
Ullaby and Elachi (1990) show that, for satellites system (ENVI, 1996).
that transmit and receive through the same antenna
(which is the usual case), it is possible to suppose POLARIMETRIC SAR SYSTEMS
that S = S Therefore, the matrix presented in HV YH
Conventional SAR systems operate in a single equation (1) can be reduced, without loss of infor-
frequency, with a single antenna of fixed polarisa- mation to
tion for both the transmitted and received signals.
Usually only the intensity or the amplitude data is  S 1 supplied to the user and, as a consequence, any (3)  Z= S2information carried in the phase of the complex 
S electromagnetic signal is lost. 3 
When polarimetric SAR sensors are used, the
full complex signal is recorded and, thus, the re-
where S , 1 ≤ i ≤ 3 denotes S , S and S ini HH HV VVturn in all the configurations (HH, H~ VH and VV)
any convenient order. are fully recorded (intensities or amplitudes and
When the number of elementary scatterers (de-
relative phases). In order to accomplish this for
noted N in equation (2)) is very large, it can be every resolution cell the complex scattering ma-
assumed that the vector Z in equation (3) obeys a
trix, denoted as multivariate complex Gaussian distribution

(Goodman, 1963). This is true if the backscatter X
 S S VV VH  (1) is constant, independently of the imaged area,
S= S S since the speckle Y is assumed to obey a  HV HH
  mu1tivariate complex Gaussian law.
In this work mu1tilook data are considered and,
is measured. Subscripts p, q ∈ {H, V} denote the in order to derive their distributional properties,
transmission and reception components of the vector Z in equation (3) will be, thus, considered
signal, respectively, and elements S are called pq the k-th single-look observation and denoted as
complex scattering amplitude. Sarabandi (1992) Z(k). A fixed number, n, of independent outcomes
shows that of Z are averaged to form the n-1ooks covariance
matrix, given by (Lee et al., 1995)
N
n iφ iφpq n pq (2) S =|S |e = |s | epq pq ∑ pq n
n=1 1(n) * TZ = Z(k) Z (k)∑ n k=1 (4)
where N is the number of scatterers of each reso-
n Tlution element, each having amplitude | s | and where Z*(k) denotes the transposed conjugate pq
of Z(k).
nphase φ . pq
The advantage of working with the covariance Other ways of representing polarimetric data are
(n)matrix, defined as A = nZ , is that it exhibits a the Stokes matrix, the modified Stokes matrix, the
multivariate complex Wishart distribution covariance matrix and the Mueller matrix (Ulaby
(Srivastava, 1963). Its density is given by and Elachi, 1990).

2 de 13 Nº 10 – Diciembre 1998 A user friendly statistical system for polarimetric SAR image classification
qn (n−q) −1 where Z is a normalising constant, 1 is the in-n |z| exp[−nTr(C z)] β A (5) p (z) =(n)Z n dicatar function of the set A, and (s,t) denotes that K(n,q)|C|
co-ordinates s and t are neighbours, then it is said
that M obeys the Potts-Strauss model with parame-where q denotes the dimension of the vector Z,
ter β. It is important to notice that for every β> 0 q(q-l)/2 K(n,q) = π Γ(n)... Γ(n-q + 1), Tr denotes the
*r this model favours those configurations that ex-trace of the matrix, C = E[ZZ ], and Γ is the Eu1er
hibit clusters of same-class pixels. Garnma function. Using equation (5) it is possible
Once defined this distribution, it can be used as to derive the densities for situations of particular
the prior for the classes in a Bayesian framework. interest, as presented in Lee et al. (1995). They are
Every class ξ ∈ Ξ, 1≤ i ≤ l, will be associated to a l addressed here for a better understanding of the
certain type of target. implemented classifiers. The following situations
For a discussion of the possible ways to obtain were implemented in the system here considered: a
estimators of η (the true map) given the data, the pair of intensities, phase difference, ratio of inten-
reader is referred to Besag (1989). The system sities and pair intensity-phase.
here presented implements one of fuese estimation
techDiques: the Iterated Conditional Modes CONTEXT, THE POTTS-STRAUSS
(ICM).
MODEL AND THE ICM ALGORITHM Assuming that the classes can be described by
the Potts-Strauss model, the problem of classifica-The use of Markovian distributions (also known
tion consists of finding an estimator of the true as Markov random fie1ds) for the parametric mod-
class configurationη given the data. It will also be elling of context dates back to the 70s, but their
assumed that the distribution of the data given the use became widespread after the work by Geman
classes is known, after the training steps required and Geman (1984).
by the ML procedure. Markov random fields, are a mu1tidimensional
The ICM algorithm consists of the iterative im-extension of the index of Markov chains, where
provement of the classification of the co-ordinates, the concept of future given past is transformed into
using the information of its return and the classes spatial conditioning. The interest in this kind of
of its neighbouring sites. Denotingη(k) the avail-distributions dates back to the beginning of the
able classification after the k-th iteration, this clas-century, since the well-know Ising model for mag-
sification will be improved replacing the class netism is one of its most famous particular cases.
The reader is referred to (Besag, 1989) for more observed in every site s by the class ξ' ∈ Ξ that
information about their use in image analysis. For maximises theexpression
the purpose of this paper, it will suffice to define
the underlying distribution for the classes: the (7) L(ξ') = f (z ) exp (β#{t∈∂ :ξ = ξ'))ξ' s s t
Potts-Strauss modelo
Denote η=[η , s ∈ S] a particular configuration s where f ,is the density associated to class ξ'and ξ
of classes, with S the set of co-ordinates of the δ is the set of neighbouring co-ordinates around s
image. Any η will be regarded as the outcome of site s. The relevant densities for the problem at
sthe random variable defined as W: Ω→ Ξ , where hand are presented in the next section. The process
Ω. is a sample space and Ξ={ ξ – ξ } is the set of 1 2 iterates until there is evidence of convergence.
sall possible classes for each co-ordinate, and Ξ is Equation (7) is the conditional likelihood of
the set of all possible maps (completely classified class ξ' given the data and the neighbouring
images). classes. The first term alone would have yielded to
Markov random fields are specifications of the ML classification scheme, while the second
Sprobabilities to every map η=Ξ , satisfying some alone leads to the mode filter (the replacement of
mild conditions. These probabilities can be chosen the current class by the most frequently observed
in arder to model spatial interaction. This model- one in its neighbourhood).
ling would be attained by associating higher
(lower, resp.) probability values to more (less, EQUIVALENT NUMBER OF LOOKS
resp.) ordered -smooth, less varyingmaps. For the
ESTIMATION definition of the techniques embedded in the sys-
tem here considered, fuese probabilities depend on The equivalent number of looks n is one of the
η and on a single real parameter β. parameters of the distributions arising from the
If the random variable M (for "map") obeys the multiplicative modelo This parameter could be
distribution given by estimated only once for the whole image, using
samples selected ayer homogeneous region. This
~ -1 (6) means, for linear detection (amplitude data) that Pr (M = η) = Z exp (β 1 (η ))β ∑ η st 1/2(s,t) fuese should be samples from Γ distributions
(Frery et al., 1997a).
Nº 10 – Diciembre 1998 3 de 13 A. H. Correia, c. da Costa Freitas, A. c. Frery and S. J. S. Sant' Anna
The interactive procedure implemented to esti- ICM intensity bivariate
mate the equivalent number of looks here de-
This option applies the ML and ICM classifica-scribed was proposed in Vieira (1996).
2 tions to a pair of intensity images, either two po-Several samples may be selected, and χ good-
larimetric components or the result of two pas-1/2ness-of-fit test for the Γ distribution is performed
sages of the same monospectral sensor (such as
for each sample. The sample with low likelihood JERS-l, ERS-l, etc.). 1/2of belonging to the Γ distribution (low p-values) After the input of the initial data the interface
may be discarded, and the final estímate is com- shown in Figure 4 is presented. It exhibits the 2-D
puted as the mean of the remaining estimated val- histogram of the pair of bands, along with the 2-D
ues. The user has the option of the decorrelating estimated density, both in perspective and in con-
the samples (i.e. of resampling in lines and col- tour plot. The estimated parameters are presented
umns) by defining the horizontal and vertical lags. at the bottom of the plots.
To help the user in this task, the system calculates As every interface presented in this work, that
the autocorrelation function of the data. The inter- presented in Figure 4 is fully interactive with the
face s for the calculating the autocorrelation func- user The user can specify the interval the plots will 2tion of samples for the χ test, and the estimation be drawn, any desirable rotation, the number of
of n, are shown in Figure 1, Figure 2 and Figure 3, contour levels to be used, etc. This feature greatly
respectively. stimulates the interaction of the user with the data.
The input values affect all the sub-windows, since
THE SYSTEM they are connected in order to help the visualisa-
tion. The system behaves as an extension of fue ENVI
This interface has to be used for every class of v. 2.5 system, and it uses its native functions and
interest. Once this is performed, the ML classifica-others from IDL (lnteractive Data Language). In
tion is performed, and the interface shown in Fig-this inanner, several functions such as those for
ure 5 is presented to the user. The user can interac-data management, processing and analysis were
tively choose the classes for which the estimated reused.
densities are presented (in perspective and as a Both classifications implemented are supervised
contour plot). The user can specify the viewpoint and, thus, require the specification of training sets
and number of (evenly spaced) contour levels. for parameter estimation. These sets are informed
Each class is associated to an unique colour. through regions of interest, previously defined by
The ML classification is produced, and used as the used with ENVI utilities. The equivalent num-
initial configuration by the ICM algorithm. This ber of looks (n in equations (4) and (5)) is also an
iterative technique stops according to the number input parameter; it can be estimated within the
of co-ordinates whose classification changes from system presented in Vieira (1996) and described in
one iteration to the next (Vieira, 1996). the previous section.
Denoting as R , R the pair of intensities, their 1 2The ICM classification method is a contextual
joint density under the model characterised by procedure that, in order to classify every pixel,
equation (5) is uses both the observed value in the corresponding
coordinate and the classification of the surrounding
 R R 1 2 n +sites. In order to incorporate this context within a  (n−1) H H11 22 n+1(R R )  1 2 2n exp( )  −statistical framework, the Potts-Strauss is used for 2  1−|ρ | 2n |ρ | R Rc c 1 2p(R ,R ) = I  
1 2 n−1(n+1) 2 H H 1−|ρ | 11 22 the classes. 2 n−1 c2  (H H ) Γ(n)(1−|ρ | )|ρ |11 22 c c  
The system here presented uses an inference
technique called pseudolikelihood, in order to where H = E[R ] and H = E[R ], I denotes 11 1 22 2 n-1estimate the required parameter of the Markovian the modified Bessel function of order n-1, and
model (β in equation (6)). This technique alleviates
the user from the need of choosing parameters in a *E[S S ] i j iθρ = =|ρ |etrial-and-error basis, a major drawback of most c c
2 2E[|S | ]E[|S | ]i jadvanced classification algorithms. Details are
available in Vieira (1996), Vieira et al. (1997) and
in Frery et al. (1997b). The current implementation The parameter |ρ | can be estimated by selecting c
uses any existing classification as starting point, a sample of size m and computing
being the ML the default.
The following subsections describe the function- m
ality of the system, in every case for n looks inten- [(R -R )(R -R )]∑ 1i 1 2i 2
i=1ˆsity data. The densities and parameter estimators ρ=
m m
are as presented in Lee et al. (1995). 2 2[(R -R ) ] [(R -R ) ]∑ 1i 1 ∑ 2i 2
i=1 i=1
4 de 13 Nº 10 – Diciembre 1998 A user friendly statistical system for polarimetric SAR image classification
WhereR andR denote the sample means of R Though, from the theoretical point of view the 1 2 1
denominator in R /R will take positive values and R , respectively. 2 1 2
with probability 1, the fact of dealing with discrete
ICM phase difference data imposes the use of a "safe" ratio. The system
works with R /max{R ,l}, which eliminates the 1 2This option applies the ML and ICM classifica-
possibilit y of an overflow. tions to Ψ, the difference between the phases of
two complex images. These images are derived Intensity and phase ICM SAR
from two components S (k) and S (k) of single-look i j
This option calculates both the ML and ICM images (equation (3)) in the following manner:
classification, using a multilook intensity image R 1
(n)
Rij and a phase difference ψ. The input data for this
6474 484
 m (n) processing are two multilook bands Ri and R , and jℑ[R ] 1 * −1 ij  (n)ψ =arg| S (k)S (k)|= tan∑ i j the corresponding multilook complex image R  (n)  ijin ℜ[R ]k=1 ij  (see equation (8)).  
The rest of the process is as presented in previ-
ous sections, namely the same as for classification where ℜ and ℑ denote, respectively, real and
using a pair of intensities. imaginary parts.
In order to derive the joint density of R and ψ, After the required parameters have been intro- I
intensity and phase difference data obtained from duced, the interface shown in Figure 6 is pre-
two components S and S of the scattering ma-sented, with the histogram of the data, the fitted i j
trix, consider the image density and estimated phase difference parameters.
2 n n−1When every class has been checked with this inter- τ*Γ(2n)(1−|ρ | ) (τ+ w)w(n) c p (w) =
2 2 (2n+1)/2face, Figure 7 is shown. This interface presents the Γ(n)Γ(n)[(τ+ w) −4τ|ρ | w]c
estimated densities of the phase difference for
every considered class, allowing the visual as-
The joint density of B and ψ is given by 1sessment of their separability throughout this fea-

ture.
Bn−1 1The density of the quantity defined above, under B exp(− )1 2 21−|ρ | 1 βcthe aforementioned model, is given by p(B ,ψ) = F [1; ; B ]1 1 11 22π Γ(n) 2 1−|ρ |c
2 n 2 n Γ(n +1/2)(1−|ρ | ) β (1−|ρ | )(n) 2c c 1 2p (ψ) = + F(n,1;1/2;β ) n−ψ B (1−β )2 n+1/2 12π 22 π Γ(n)(1−β ) βB exp(− )1 21−|ρ |c +
22π Γ(n) 1−|ρ |where -π < ψ ≤ π, b=|ρ |cos(ψ-θ), θ is the phase cc
of the complex coefficient of correlation and 2 2F(n,1;1/2;β )=2F (n,1;1/2;β ) is the Gaussian hy-1
pergeometric function. In the system this function where w = R /R and τ = H / H1 2 11 22 was implemented based on the algorithm described Though, from the theoretical point of view the
in Press et al. (1988), for any n > 0. denominator in R /R will take positive values 1 2
with probability 1, the fact of dealing with discrete ICM ratio of intensities
data imposes the use of a "safe" ratio. The system
Both the ML and ICM classification are ob- works with R /max{R ,l}, which eliminates the 1 2
tained, derived from the ratio between two multi- possibilit y of an overflow.
look in tensity bands, i.e., using data of the form
Intensity and phase ICM SAR R /R . i j
Analogously to the previous situation, namely This option calculates both the ML and ICM
for the classification using phase difference, after classification, using a multilook intensity image R 1
the required inputs the histogram, fitted densities and a phase difference ψ. The input data for this
and estimated parameters are shown for every processing are two multilook bands Ri and R , and jclass. Once the fittings have been checked for (n)the corresponding multilook complex image R ijevery class, the whole set of fitted densities is (see equation (8)).
shown. The rest of the process is as presented in previ-
The density that characterises this data is ous sections, namely the same as for classification
using a pair of intensities.
2 n n−1τ*Γ(2n)(1−|ρ | ) (τ+ w)w(n) c In order to derive the joint density of R and ψ, p (w) = I
2 2 (2n+1)/2Γ(n)Γ(n)[(τ+ w) −4τ|ρ | w]c intensity and phase difference data obtained from
two components S and S of the scattering ma- i j
trix, consider the image where w = R /R and τ = H / H1 2 11 22

Nº 10 – Diciembre 1998 5 de 13 A. H. Correia, c. da Costa Freitas, A. c. Frery and S. J. S. Sant' Anna
The quantitative comparison of results, when where F is the confluent hypergeometric func-1 1
bivariate intensity data are used, is presented in tion.
Table 4. The best classification achieved (namely In the system this function was implemented by
for the L HVVV data set) is highlighted. Similar means of an adaptation of the Gaussian hyper-
results, obtained with phase difference, with the geometric function algorithm described in the
ratio of intensities and with the pair intensity-phase Press et al. (1988).
are presented in Table 5, Table 6 and Table 7,
respectively. In all these tables the best results are
CASE STUDY highlighted.
The analysis of these tables shows that classifi-In order to assess the information content of po-
cations obtained with phase difference C HH-HV larimetric SAR data in crops areas, a SIR-C/X-SAR
data leads to the worst overall results, for both ML image was analysed.
and ICM methods. These two classifications are As presented in the previous section, the system
shown in Figure 9. offers a wide variety of options for input data. All
It is also noticeable that phase difference data
these possibilities were tested, namely bivariate from HH-HV and HVVV polarisations, from both
intensity, phase difference, ratio of intensities and L and C bands, do not carry useful information
the pair intensity-phase. For each of fuese tour about the studied areas and can, therefore, be dis-
options the ML classification was obtained and carded from the Test of the comparisons. Classifi-
used as initial configuration for the ICM algo- cations using these data sets were assessed in Ta-
rithm. ble 5, since the system allows it.
Taking into account the results obtained with the Data and preliminary analysis
other data sets for L band, it can be concluded that
The main parameters of the space shuttle image 1. The worst ML classification was obtained
under study are presented in Table 1. The central with the ratio of HHVV intensities, and the
co-ordinates of the area are 09(07' S, 40(18' W. best with the HVVV pair of intensity data
The imaged area corresponds to an irrigated region (see Figure 10). The improvement from the
where several types of crops are observed. worst to the best ML classification is, in this
Figure 8 presents two compositions of the data case, of 475,77%.
set under study. To the left (right, resp.) the red, 2. The worst ICM classification was obtained
green and blue channels were associated to the L when using ratio of HH-HV intensities, and
(C resp.) band and polarisations HH, HV and VV. the best was attained when the initial con-
This images also present the test (right) and train- figuration was the best ML classification,
ing (left) sets. i.e., when the HV-VV pair of intensity data
The classes analysed in this work are presented was used (Figure 11). The improvement, in
in Table 2, where their respective colour keys (the this case, is of314,74%.
colours with which they will be represented in the Discarding the HH-HV and HVVV phase differ-
classifications), number of training and test sam- ence data sets, and considering C band data, it can
ples and sizes of these samples are also shown. be concluded that
These classes are river (that will be depicted in 1. The worst ML classification was obtained
blue), a steppe vegetation type called caatinga with the ratio of HHVV intensities, and the
(green), prepared soil (red), soy (magenta), tillage best with the HV-HV pair of intensity data.
(cyan) and corn (yellow). The obtained improvement from the worst to
Though, as presented in Table 1, the nominal the best ML classification is of 523,22%.
number of looks is 4.7854018, the estimated quan- 2. The worst ICM classification was obtained
tity amounts to 2.97479. This value is the mean of when using ratio of HHVV intensities, and
the equivalent number of looks observed in each the best was attained when the initial con-
component (see Table 3). The estimation of this figuration was the best ML classification,
quantity was performed using samples as large as i.e., when the HH-HV pair of intensity data
possible, and taken over homogeneous areas. Be- was used (Figure 12). The improvement, in
fore using these samples, they were submitted to this case, is of 365,42%.
2χ goodness of fit test, and they all passed at the 1 Comparing the best ML and ICM results, for
% level of significance the hypothesis of being each band, it can be concluded that ICM improves
homogeneous areas. 28,14% ML the classification in L band, and
20,38% in C band. It is interesting to compare ML
Comparison of classifications and ICM results, since the latter is an improved
result that used the former as starting point. The All possible combinations of data for same band
improvement is notorious, both from the qualita-were used to generate classifications, as described
tive point of view (regions are smoother and better in Correia (1998). These precision of these results defined) and from the quantitative one.
was assessed with the use of the estimated Kappa
Though the aforementioned improvement is ˆ( k ) coefficient of agreement and its sample vari-
welcome in every application, it is not as dramatic 2ance ( ) over test areas (Landis and Koch, 1977). sk
6 de 13 Nº 10 – Diciembre 1998 A user friendly statistical system for polarimetric SAR image classification
as that obtained in Frery et al. (1997b) and Vieira tion was obtained with the band L, HVVV inten-
(1996). In that work improvements of the order of sity pair (band C, HH-HV phase difference, resp.)
500% were obtained when comparing ML and data.
ICM classifications. This might be due to, among Using the ICM algorithm yields to classifica-
other reasons, the fact that in those works a single tions not worse than the initial one. Since in this
component amplitude multilook data were used, study all initial classifications were obtained by the
which conveys les s information per pixel than that ML method, improvements from this pointwise
carried by polarimetric images. classification technique were expected. In fact, as
The analysis of the confusion matrices (not presented, significant improvements were achieved
shown here, but available. through the system) at the mere expense of CPU time.
shows that, using the ICM algorithm in all cases, The version of the system here presented only
the best results are allows the use of distributions associated to homo-
l. For the class River, the pair intensity HH-HV geneous regions. The authors are currently upgrad-
band C yielded to a 100% of correctly classi- ing the functionality of the procedures, in order to
fied pixels (Figure 12, right). allow the use of other distributions, more suited to
2. The class Caatinga was best classified (98% heterogeneous and extremely heterogeneous ob-
of correctly labelled pixels) with the pair in- servations.
tensity HVVV band C (Figure 13, left).
3. Phase difference between HHVV band C data ACKNOWLEDGEMENTS
yielded to the best Soil classification (94%,
This work was partially developed with re-Figure 12 left).
sources from CNPq (Proc. 523469/96-9) and 4. The Soy class was best classified through the
FACEPE (APQ 0707-1.03/97). use of intensity HH-VV pair, from L band
(Figure 13, right).
BIBLlOGRAPHY 5. The pair phase-intensity, from HHVV L band,
yielded to the best Tillage classification
BESAG, J. 1989. Towards Bayesian image analysis. J.
(95%, Figure 14). App. Stat. 16(3):395-407.
6. The best (86%) Corn classification was ob- CORREIA, A.H. 1998. Desenvolvimento de classifica-
tained when the intensity par HV-VV data set dores de máxima verossimilhanr;a e ICM para ima-
from L band was used (Figure 10, right). gens SAR polarimétricas. (MSc in Remote Sensing)
Insti tuto de Nacional de Pesquisas Espaciais. Sao Jo-These results show that the overall performance
sé dos Campos, SP, Brazil. To be presented. ˆcriterion, namely the estimated Kappa ( ) coeffi-k
DEGRANDI, G.; LEMOINE, G. and SIEBER, A. 1992. cient of agreement, mar not inform how well indi-
Supervised fully polarimetric classification: an ex-vidual classes are classified. Besides this, these
perimental study on the MAESTRO-1 Freiburg data
results show the importance full polarimetric im- seto In: IGARSS'92 International Geoscience and
ages, since the information they convey is rather Remote Sensing Symposium'92, Houston. Intema-
specialised and in concentrated in different data tional Space Year: space remate sensing. IEEE, V. 1,
sets for different classes. p. 782-785.
DU, L.J. and LEE, J.S. 1996. Polarimetric SAR image
classification based on target decomposition theorem CONCLUSIONS AND FUTURE
and complex Wishart distribution. In: IGARSS'96 In-
WORK ternational Geoscience and Remote Sensing Sympo-
sium'96, Lincoln. Remote Sensing for a Sustainable The system here presented allowed the compari-
Future. IEEE, v. 1, p. 439-441.
son of classifications using four types of multilook ENVI 2.5 user's guide: The Enviromment for Visualiz-
polarimetric data: bivariate intensity, phase differ- ing Images, version 2.5. 1996. Lafayette, Better Solu-
ence, ratio of intensities and the pair intensity- tions Consulting, 1993-1996. Under contract of Re-
phase. search Systems Inc.
This system is user friendly and goal driven, and FRERY, A.C.; MÜLLER, H.J.; YANASSE, C.C.F and
SANT' ANNA, S.J.S. 1997a. A model for extremely it proved being easy to use. The users is only re-
heterogeneous clutter. IEEE Trans. Geosc. Rem. Sens. quired to know the basic ideas of maximum likeli-
35(3):648-659. hood classification, in order to be able to produce
FRERY, A.C.; YANASSE, C.C.F.; VIEIRA, P.R. improved results based on the Markovian model-
SANT' ANNA, S.J.S. and RENNÓ, C.D. 1997b. A
ling of classes and pseudolikehood estimation.
user-friendly system for synthetic aperture radar im-
The intense use of graphic interfaces eases the age classification based on grayscale distributional
modelling and understanding of data, a central properties and contexto Simp6sio Brasileiro de Com-
issue in SAR image analysis. puta~ao Gráfica e Processamento de Imagens, 10.,
A SIR-C/X-SAR image of mainly crops was 1997, p. 211-218. SIBGRAPI97. Los Alámitos, CA,
IEEE Computer Society. chosen as case study. Training and test samples
GEMAN, D. and GEMAN, S. 1984. Stochastic relaxa-were chosen and analysed within the system, and
tion, Gibbs distributions and the Bayesian restoration all possible data configurations for each band were
classified. The best (worst) ICM overall classifica-
Nº 10 – Diciembre 1998 7 de 13 A. H. Correia, c. da Costa Freitas, A. c. Frery and S. J. S. Sant' Anna
of images.IEEETrans. Patt. An. Mach.Int. 6(6): 721- cation of K -distributed SAR images of natural targets
741. and probability of error estimation. IEEE Trans.
GOODMAN, N. R. 1963. Statistical analysis based on a Geosc. Rem. Sens. 34(5):1233-1242.
certain multivariate complex Gaussian distribution. PRESS, W.H.; FLANNERY, B.P.; TEULOSKY, S.A.
Ann. Math. Stat. 34(1):152-177. and VE1TERLING, W.T. 1988. Numerical recipes in
KONG, J.A. 1988. Identification of terrain cover using C, Cambridge, Cambridge University Press.
the optimal polarimetric classifier. J. Electrom. Waves QUEGAN, S. and RHODES, l. 1995. Statistical models
Appl.2(2):171-194. for polarimetric data: consequences, testing and valid-
LANDIS, J. and KOCH, G.G. 1977. The measurements ity.Int. J. Rem. Sens. 16(7):1183-1210.
of observer agreement for categorical data. Biomet- SARABANDI, K. 1992. Derivations of phase statistics
rics. 33(3):159-174. from the Mueller matrix. Radio Sci. 27(5):553-560.
LEE, J. S. and GRUNES, M. R. 1992. Feature classifica SRIVASTAVA, M.S. 1963. On the complex Wishart
tion using multi-look polarimetric SAR imagery. In: distribution. Ann. Math. Stat. 36(1):313-315.
IGARSS'92 Intemational Geoscience and Remote TUR, M.; CRIN, K.C. and GOODMAN, J.W. 1982.
Sensing Symposium'92, Houston.Intemational Space When is speckIe noise multiplicative? Appl. Opto
Year: space remate sensing.IEEE, v. 1, p. 77-79. 21:1157-1159.
LEE, J.S.; HOPPEL, K.W. and MANGO, S.A. 1994. ULABY, F. T. and ELACHI, C. 1990. Radar po-
Intensity and phase statistics of multi-look polarimet- larimetriy for geoscience applications. Norwood,
ric and interferometric SAR imagery. IEEE Trans. Artech House, 364p.
Geosc. Rem. Sens. 32(5):1017-1028. VIEIRA, P.R. 1996. Desenvolvimento de
LEE, J.S.; DU, L.; SCHULER, D.L. and GRUNES, classl:ficadores de máxima verossimilhanr;a e ¡CM
M.R. 1995. Statistical analysis and segmentation of para imagens SAR. (MSc in Remote Sensing) Institu-
multilook SAR imagery using partial polarimetric to de Nacional de Pesquisas Espaciais. Sao José dos
data. In: IGARSS'95 Intemational Geoscience and Campos, SP, Brazi1, 251 p. (INPE-6124-TDI/585).
Remote Sensing Symposium'95, Firenze. Quantitative VIEIRA, P.R.; YANASSE, C.C.F.; FRERY, A.C. and
Remote Sensing for Science and Applications. Pisca- SANT' ANNA, S.J.S. 1996. Um sistema de aná1ise e
taway, IEEE, v. ni, p. 1422-1424. classificasao estatísticas para imágens SAR. In: Pri-
LEE, J.S. and GRUNES, M.R. 1994. Classification of meras Jornadas Latinoamericanas de Percepción Re-
multilook polarimetric SAR imagery based on com- mota por Radar, Buenos Aires, Dez. 1996. Técnicas
plex Wishart distribution. Int. J. Rem. Sens. 15(11): de Processamiento de Imágenes. Paris, ESA, p. 170-
2299-2311. 185.
LIM, H.H; SWARTZ, A.A.; YUEH, H.A.; KONG, J.A.; YUEH, S.H.; KONG, J.A.; JAO, J.K.; SHIN, R.T. and
SHIN, R.T. and VAN ZYL, J.J. 1989. Classifications NOVAK, L.M. 1989. K-distributition and polarimet-
of earth terrain using polarimetric SAR images. J. ric terrain radar clustter. J. Electrom. Waves Appl.
Geophys. Res. 94(B6):70497057. 3(8): 747-768.
NEZRY, E.; LOPÉS, A.; DUCROT-GAMBART, D.;
NEZRY, C. and LEE, J.S. 1996. Supervised classifi-


Image identifier P-11534
thAcquisition date April 14 1994
Size o fue considered image 407 x 370 pixels
Frequency L (1.254 GHz) and C(5.304G Hz)
Polarisation HH,HV, WandVH
Incidence angle 49.496°
Platform height 216.14 km
Orbit direction Descending
Type of product Multilook Complex (MLC)
Nominal number of looks 4.7854018
Geometric representation Ground range
Pixel spacing 12.5m in e and 12.5m in azimuth

Table 1. SIR-C/X-SAR image main parameters


Training samples Test samples
Classes Colour Number Pixels Number Pixels
River Blue 2 4949 2 3844
Caatinga Green 5 5177 5 3585
Prepared soil Red 1 3221 1 2101
Soy Magenta 2 2609 2 2128
Tillage Cyan 1 635 1 360
Corn Yellow 2 3505 2 1946

Table 2. Classes of interest, colour keys. training and test samples
8 de 13 Nº 10 – Diciembre 1998 A user friendly statistical system for polarimetric SAR image classification

Band
Polarisation L C
HH2.6688 2.6713
HV 3.18362.9723
W3.5340 2.8188
Mean 2.97479

Table 3. Estimated equivalent numbert of looks in all the avail-
able bands and polarisations, and overall mean


-3ˆInput data Variance (x 10 ) k
(band - polarisation)
ML ICM ML ICM
L HH-HV 0.596980 0.710471 2.38412 1.99384
L HH-VV 0.575008 0.694241 2.48444 2.14506
L HV-VV 0.606424 0.777114 2.35719 1.64955
C HH-HV 0.575344 0.692635 2.50954 2.10982
C HH-VV 0.482214 0.571004 2.61262 2.72449
C HV-VV 0.566975 0.614171 2.68826 2.90562

Table 4. Assessment of classification precision using bivariate intensity data


-3ˆInput data Variance (x 10 ) k
(band - polarisation)
ML ICM ML ICM
L HH-HV 0.020904 0.025786 2.56678 3.39736
L HH-VV 0.193186 0.372729 2.68909 3.12572
L HV-VV 0.008927 0.014442 2.66556 5.07250
C HH-HV -0.001943 -0.008580 2.06280 2.23533
C HH-VV 0.092317 0.148819 3.19603 4.39456
C HV-VV 0.005144 0.007090 2.62186 3.37758

Table 5. Assessment of classification precision using phase difference data


-3ˆ Variance (x 10 ) Input data k
(band - polarisation)
ML ICM ML ICM
L HH-HV 0.105324 0.187371 3.12436 3.90025
L HH-VV 0.272400 0.360187 2.74605 3.37850
L HV-VV 0.318741 0.455631 2.57741 2.70515
C HH-HV 0.116141 0.165443 3.19497 4.82030
C HH-VV 0.181810 0.274525 3.71316 4.92802
C HV-VV 0.203089 0.272691 2.76586 3.36486

Table 6. Assessment of classification precision using the ratio of intensities data


-3ˆInput data Variance (x 10 ) k
(band - polarisation)
ML ICM ML ICM
L HH-HV 0.469581 0.633668 2.63720 2.34108
L HH-VV 0.325979 0.424540 2.77930 2.85456
C HH-HV 0.418908 0.521662 2.59719 2.55940
C HH-VV 0.416235 0.565799 2.65869 2.74780

Table 7. Assessment of classification precision using the the pair intensity-phase data







Nº 10 – Diciembre 1998 9 de 13 A. H. Correia, c. da Costa Freitas, A. c. Frery and S. J. S. Sant' Anna






10 de 13 Nº 10 – Diciembre 1998