Aurlie Voisin, Vladimir A.

Krylov, and Josiane Zerubia

Synthetic Aperture Radar (SAR): active sensor used for
different kinds of applications:
Security, Epidemiology, Environment, Risk management
Very high resolution (VHR) SAR data (up to 1 meter)
Speckle [Oliver 04]
Heterogeneity [Cheney 09]
Supervised classification of VHR single-channel SAR
amplitude images
Methods : neural networks [Jacob 02], bags-of-features [Yang 09], etc.
Proposed method: statistical modeling + Markov random fields (MRFs)
2
3
Supervised classification
1 learning image
1 test image
M classes (e.g.: water, vegetation, urban)
Original SAR
amplitude image
4
Learning
Test
PDF SAR amplitude of
each class
MRF parameter
estimation
Energy minimization to
find the class of the test-
image pixels
5
Learning
Test
PDF SAR amplitude of
each class
MRF parameter
estimation
Energy minimization to
find the class of the test-
image pixels
Dictionary-based Stochastic Expectation Maximization
Heterogeneity of SAR images
Single probability density function (PDF) doesnt accurately model
SAR amplitude statistics
Model the SAR amplitude PDF by a FMM (Finite Mixture Model)
6

=
=
K
k
k k k
r p P r p
1
) ( . ) ( u u
[Moser 10]
Estimation of:
K: number of components
p
k
(r|
k
): k
th
PDF family component
Parameters of the k
th
component (
k
and P
k
)
with
Data incompleteness
belongs to which statistical population?
Unsupervised context
7

=
=
K
k
k
P
1
1
1 0 s s
k
P
} {
N
r r r ,...,
1
=

=
=
K
k
k k k
r p P r p
1
) ( . ) ( u u
8
Log Normal
Weibull
Fisher
Generalized Gamma
Nakagami
K-Root
Generalised Gaussian Rayleigh
H
e
u
r
i
s
t
i
c
T
h
e
o
r
e
t
i
c
a
l
2
) ) (ln(
1
2
1
) , (
o
t o
o
m r
e
r
m r f

=
n
u n
n
u
n
u n
|
|
.
|

\
|

=
r
e r r f
1
2
) , (
M L
L
M
Lr
M
Lr
M
L
M L
M L
M L r f
+

+
I I
+ I
=
) 1 (
) (
) ( ) (
) (
) , , (
1
3
u
u
u
u
v
o
kv
o k o
v
k o v
|
.
|

\
|

|
.
|

\
|
I
=
r
e
r
r f
1
4
) ( .
) , , (
|
|
.
|

\
|

|
|
.
|

\
|
I
=

1 2
5
) (
2
) , (
u
u
u
Lr
L
L
e r
L
L
L r f
|
|
.
|

\
|
|
|
.
|

\
|
|
|
.
|

\
|
I I
=

+
+
2 / 1
1
2
6
2
) ( ) (
4
) , , (
u u
u
LM
r K r
LM
M L
M L r f
L M
M L
M L
( ) ( )
u



t
u u

d e
r
r f
r

+
I
=
2 /
0
) sin( ) cos(
7
/ 1 / 1 / 1
) (

) , (
09/21/2010 SPIE Remote Sensing Toulouse 2010 - A. Voisin et al.
[Moser 06]
[Krylov 09]
E step:
S step: stochastic labeling step leading to
M step:
9

=
=
K
l
t
l i l
t
l
t
k i k
t
k
t
ik
r p P
r p P
1
) , ( .
) , ( .
u
u
t
t
ik
w
( )|
.
|

\
|
=

=
+
N
i
k i k
t
ik
t
k
r p w
k 1
1
) , ( ln max arg u u
u
image size
k pixels nb
P
t
k
_
_
1
e
=
+
[Celeux 95]
ML not feasible for some PDF distributions
Method of Log-Cumulants [Moser 06, Tison 04]:
Mellin transform of the PDF [Sneddon 72]
v
th
order second kind cumulant
10

= = u
0
1
). ( ) )( ( ) ( du u u p s p M s
s
u u u
) 1 ( )) (ln(
) (v
u v
u = k
Family MoLC equations
Log-Normal k
1
=m, k
2
=o
Weibull k
1
=ln(u)++(1)/n, k
2
=+(1,1)/n
Nakagami 2k
1
=+(L)-ln(L/), 4k
2
=+(1,L)
Generalized Gamma k
1
=+(k)/v+ln(o), k
2
=+(1, k)/v
k
3
=+(2, k)/v
3
E - step: (posterior proba)
S step: sample the label for each greylevel according to
MoLC step:
K step: for k=1,,K
t
, if P
k
t+1
< threshold, eliminate the k
th
component. K
t+1
= K
t
-1
Model Selection step: for each k, compute the log-likelihood,
and define p
k
t+1
(.) as the PDF yielding the highest value
11

=
=
t
K
l
t
l
t
l
t
l
t
k
t
k
t
k
t
k
z p P
z p P
1
) , ( .
) , ( .
u
u
t
) (z s
t
( )

e
=
kt
Q z
t
kj j
t
kj
z f z h L ) ( ln ). ( u
t
k
t
{ } 1 ,..., 0 = Z z

=
e +
=
1
0
1
) (
) (
Z
z
Q z t
k
z h
z h
P
kt

e
e +
=
kt
kt
Q z
Q z t
k
z h
z z h
) (
) ln( ). (
1
1
k

e
e

=
kt
kt
Q z
Q z
b
t
k
t
bk
z h
z z h
) (
) ) ).(ln( (
1
k
k
{ } k z s z Q
t
kt
= = ) ( :
[Moser 06]
[Krylov 09]
{ } 3 , 2 = b
Initialization
K
max
= 6 chosen by trial and error
Randomly chosen labels
For each iteration t, the global log-likelihood is
computed, if it is > (max log-likelihood), the parameters
are saved
Stopping criterion
Maximum number of iterations reached
Number of components = 1
12
13
Model KS distance
GGamma 0.022
DSEM4: GGamma,
Naka, LogN, Weib.
0.007
T
e
r
r
a

S
A
R
-
X

i
m
a
g
e

o
f

R
o
s
h
e
n
h
e
i
m
(
G
e
r
m
a
n
y
)

(

I
n
f
o
t
e
r
r
a
,

2
0
0
8
)
C
O
S
M
O
-
S
k
y
m
e
d
i
m
a
g
e

o
f
C
a
v
a
l
l
e
r
m
a
g
g
i
o
r
e
(
I
t
a
l
y
)

(

I
S
A
,

2
0
0
8
)
Model KS distance
Weibull 0.053
DSEM4: LogN,
Naka (2), Weib.
0.011
Statistics of the
whole image
Statistics of the
vegetation class
14
Learning
Test
PDF SAR amplitude of
each class
MRF parameter
estimation
Energy minimization to
find the class of the test-
image pixels
15
MRF: robustness against speckle and contextual
information [Besag 86, Dubes 89, Fjortoft 03]
Anisotropic second-order neighborhood system
Gibbs distribution [Besag 74, Geman 84]: Local
characteristic (conditional proba) for each class m
Potts model:
Pseudo-logLikelihood
| | M ; 1 e
( )
( )

=
=
=
=
=
= = =
M
j
x x H
x x H
s
s
m s
s
m s
s
m s
s
j s
s
m s
e
e
x P
x x P
x x P x p
1
) , (
) , (
) (
) (
) ( ) (
) (
) (
,
) ( ) (
e
e
e
e e
{ }

e e
=
|
|
.
|

\
|
=
S s C s s s
x x
s
s
s s
PDF x x H
' , : '
) (
'
. ) ln( ) , , ( o | |

=
I
s cliques
s
m s
x p x PL
_
) (
) ( ln )) ( ln( e |
16
Problematic: Single pol. CSK images
Find a 2
nd
channel to improve the accuracy
Textural features
C
O
S
M
O
-
S
k
y
m
e
d
i
m
a
g
e

o
f

C
a
v
a
l
l
e
r
m
a
g
g
i
o
r
e
(
I
t
a
l
y
)

(

I
S
A
,

2
0
0
8
)
Semi-Variogram[Chen 04]
Grey-Level Co-occurrence
Matrix (GLCM) variance
[Haralick 73]
17
PDF SAR amplitude of
each class and each
channel
Copulas
MRF parameter
estimation
Energy minimization to
find the class of the test-
image pixels
Learning
Test
[Moser 10]
18
PDF SAR amplitude of
each class and each
channel
Copulas
MRF parameter
estimation
Energy minimization to
find the class of the test-
image pixels
Learning
Test
[Moser 10]
( )
2 1
2 2 1 1
*
2 2 1 1
) ( ), (
) ( ). ( ) (
y y
y F y F C
y p y p y p
m m m m m
m m m m m m
c c
c
=
e e
e e e
19
Water Urban Vegetation Overall
DSEM-MRF
99.14 % 98.88 % 84.65 % 94.22 %
K-NN-MRF
96.72 % 96.09 % 99.92 % 97.58 %
CoDSEM(Semivar.)
98.37 % 98.91 % 100 % 99.09 %
CoDSEM(GLCM)
98.62 % 98.42 % 100 % 99.01 %
DSEM-MRF CoDSEM(GLCM)
C
O
S
M
O
-
S
k
y
m
e
d
i
m
a
g
e

o
f

C
a
v
a
l
l
e
r
m
a
g
g
i
o
r
e
(
I
t
a
l
y
)

(

I
S
A
,

2
0
0
8
)
20
Water Urban Vegetation Overall
DSEM-MRF
92.95 % 98.32 % 81.33 % 90.87 %
K-NN-MRF
90.56 % 98.49 % 94.99 % 94.68 %
CoDSEM(GLCM)
91.28 % 98.82 % 93.53 % 94.54 %
CoDSEM(GLCM) K-NN-MRF
TerraSAR-X image of Rosenheim
(Germany) (Infoterra, 2008)
21
Water Urban Vegetation Overall
DSEM-MRF
99.81 % 99.03 % 99.99 % 99.61 %
CoDSEM(GLCM)
98.66 % 99.56 % 99.27 % 99.16 %
CoDSEM(GLCM)
COSMO-Skymed image of Port-au-Prince
(Haiti) (ISA, 2009)
DSEM-MRF
22
Algorithm validated in the application:
urban/land/water separation on several single-pol. SAR
images
Smoothing effects at spatial borders
Possible improvements:
More sophisticated texture-extraction techniques
Taking into account urban geometry via hierarchical/multiscale
MRFs
23
French defense agency (DGA)
Italian Space Agency (ISA)
Infoterra (Astrium Services)
Dr. G. Moser and Prof. S. Serpico from Univ. of Genoa
for their helpful comments and fruitful collaboration
24
[Oliver 04] Oliver, C. and Quegan, S., [Understanding Synthetic Aperture Radar images], SciTech Publishing (2004).
[Jacob 02] Jacob, A. M., Hemmerly, E. M., and Fernandes, D., SAR image classification using a neural classifier based on Fisher criterion, in
[Medical Imaging: Image Processing], Proc. of the VII Brasilian Symposiumon Neural Networks (SBRN), 168172 (2002).
[Yang 09] Yang, W., Dai, D., Triggs, B., and Xia, G.-S., Semantic labeling of SAR images with hierarchical Markov aspect models, HAL Research
Report , hal00433600 (2009).
[Cheney 09] Cheney, M. and Borden, B., [Fundamentals of radar imaging], Phladelphia: Society for industrial and applied mathematics (2009).
[Moser 10] Moser, G., Krylov, V., Serpico, S. B., and Zerubia, J., High resolution SAR image classification by Markov random fields and finite
mixtures, Proc. of SPIE 7533, 753308 (2010).
[Moser 06] Moser, G., Serpico, S. B., and Zerubia, J., Dictionary-based Stochastic Expectation Maximization for SAR amplitude probability
density function estimation, IEEE Trans. Geosci. Remote Sens. 44(1), 188199 (2006).
[Krylov 09] Krylov, V., Moser, G., Serpico, S. B., and Zerubia, J., Dictionary-based probability density function estimation for high-resolution
SAR data, Proc. of SPIE 7246, 72460S (2009).
[Celeux 95] Celeux, G., Cheveau, D., and Diebolt, J., On stochastic versions of the EMalgorithm, INRIA Research Report 2514 (1995).
[Tison 04] Tison, C., Nicolas, J.-M., Tupin, F., and Maitre, H., A new statistical model for Markovian classification of urban areas in high-
resolution SAR images, IEEE Trans. Geosci. Remote Sens. 42(10), 20462057 (2004).
[Sneddon 72] Sneddon, I., [The use of integral transforms], McGraw-Hill, New York (1972).
[Fjortoft 03] Fjortoft, R., Delignon, Y., Pieczynski, W., Sigelle, M., and Tupin, F., Unsupervised classification of radar images using hidden
Markov chains and hidden Markov random fields, IEEE Trans. Geosci. Remote Sens. 41(3), 675686 (2003).
[Besag 86] Besag, J., On the statistical analysis of dirty pictures, Journal of the Royal Statistical Society 48, 259302 (1986).
[Dubes 89] Dubes, R. C. and Jain, A. K., Random field models in image analysis, Journal of Applied Statistics 16(2), 131164 (1989).
[Besag 74] Besag, J., Spatial interaction and the statistical analysis of lattice systems, Journal of the Royal Stat. Soc. 36(2), 192236 (1974).
[Geman 84] Geman, S. and Geman, D., Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images, IEEE Patt. Anal.
Mach. Intell. 6(6), 721741 (1984).
*Chen 04+ Chen, Q. and Gong, P., Automatic variogramparameter extraction for textural classification of the panchromatic IKONOS imagery,
IEEE Trans. Geosci. Remote Sens. 42(5), 11061115 (2004).
[Haralick 73] Haralick, R. M., Shanmugam, K., and Dinstein, I., Textural features for image classification, IEEE Trans. on Systems, Mans and
Cybern. 3(6), 610621 (1973).