MINISTRY OF EDUCATION AND TRAINING

MINISTRY OF NATIONAL DEFENCE

ACADEMY OF MILITARY SCIENCE AND TECHNOLOGY

BUI NGOC THUY

Research ON THE ACCURACY IMPROVEMENT OF THE TARGET
PARAMETERS IDENTIFICATION AND DETERMINATION USING
POLARIMETRIC SYNTHETIC APERTURE RADAR AND
POLARIMETRIC INTERFEROMETRIC IMAGES

Specialization: Electronic engineering
Code:

9 52 02 03

SUMMARY OF DOCTORAL THESIS IN ENGINEERING

Hanoi, 2019


This thesis has been completed at:
ACADEMY OF MILITARY SCIENCE AND TECHNOLOGY

Scientific Supervisors:
1. Assoc. Prof, Ph.D Le Vinh Ha
2. Ph.D Pham Minh Nghia

Reviewer 1: Prof, Ph.D Bach Gia Duong

has been researched and developed and applied in varied fields of socio-economic
life. Technique of Polarimetric Synthetic Aperture Radar and Polarimetric
Interferometric (PolSAR and PolInSAR) is a topic that is drawing the attention of
scientists all over the world. This technology contributes to deal with the problems
of the object‘s identification and measuring target parameters. Therefore, the
thesis has its scientific and practical significance.
Therefore, the PhD student has chosen the topic: Research on the accuracy
improvement of the target parameters identification and determination using
polarimetric synthetic aperture Radar and polarimetric interferometric images.
2. The objective
Research scientific basis and solutions to improve the accuracy of
identification and identify parameters of natural targets, targets and forest height
estimates based on PolSAR and PolInSAR images.
3. The subject and scopes
PolSAR's target decomposition technique is based on Freeman's threecomponent scattering model and Yamaguchi's four-component scattering model.
The PolInSAR decomposition technique is based on the coherence set and the
three-component decompositon technique for PolInSAR images. From above
researchs, proposing and developing solutions and simulations for the accuracy
improvement of interpretation, identification and determination of target
parameters are studied by the thesis.
4. Research methodology
- Methods of collecting information, documents, general analysis of scientific
works and articles published in the world and in the country. Collect PolSAR and
PolInSAR image data sources related to the test area.
- Researching target decomposition techniques and building algorithm models
to improve the accuracy of identification and determination of target parameters
based on PolSAR and PolInSAR images.
- Programming technology and application of informatics technology in
building a program to perform calculations and simulations using MATLAB tool


target itself, whereas the rest is reradiated as a new electromagnetic waves. Due to
the interaction with the target, the properties of the reradiated waves can be
different from those of the incident ones. From this change, they can help us
describe or identify targets. In reality, we are interested in the changes concerning
the polarization of the wave.

Figure 1.1: The interaction of electromagnetic waves with a target


3

The radar equation represents as the following:
P G ( , ) A ( , )
(1.1)
PR  T T 2  ER 2
4 rT
4 rR
1.2. The formation of Synthetic Aperture Radar image
SAR image systems allow for monitoring the earth on a global scale at all
times and in all weather conditions. The basic geometry of a SAR system is shown
in Figure 1.3.

Figure 1.3. Basic geometric
Figure 1.4. Ground range to slant
structure of a SAR space
range projection
One of the most important criteria for assessing the quality of SAR image
systems is its spatial resolution. The spatial resolution describes the visibility of
the image as much as possible so that two scattered objects can be separated in
term of spatial meaning.


To be able to extract physical information from the 2×2 coherent
backscattering matrix or Sinclair matrix which achieved through the construction
of system vectors, we represent the Sinclair matrix with the vector V(.) as follows:
S
S   HH
 SVH

S HV 
SVV 

1
 k  V  S   Trace  S  
2

(1.13)

Where ψ is a set of 2×2 complex basic matrices which are constructed as an
orthogonal set under the Hermitian inner product.
1.3.3. Coherency matrix [T] and covariance matrix [C] polarimetric
The polarimetric Pauli coherency matrix [T] and the Lexicographic
covariance matrix [C] are generated from outer product of the associated target
vector with its conjugate transpose as.

T  

k3P .k 3*TP

and C   k3L .k 3*TL



1.7. Interferometry Synthetic Aperture Radar
Interferential Synthetic Aperture Radar (InSAR) takes advantage of the phase
difference between two SAR complex images described from differences at
locations or at different times.

Figure 1.21: Geometric structure of interference radar
When building the complex product s1s2* for interferometry, we could cancel the
scattering phase terms and keep the geometrical phase. In effect, we now obtain a
signal phase which depends only on the difference in range between the two positions
R  R1 R 2 as follows:




 j 2 2 R1s1  
 4


j

R


 
 
 

s1  a1e
 *

k 
 3 P1   k *T
T6   
  3P1
k
 3P2 

 
k3*T
P 

k 
 3L1   k *T
C6   
  3L1
k
 3L2 

 
k3*T
L 

 T1 

2



  *T



  arg  1k3P k 3*TP *2T
1

2

  arg 

*T
1

 2



(1.80)

By using Eq. (1.85), the complex interferometric coherence as a function of
the polarization of the two images may be written as:
12*
 1*2
  1, 2  

(1.81)
11* 2 2*
1*T T1  1 *2T T2   2

Figure 1.22: PolInSAR acquisition geometry
1.9. Conclusion of Chapter 1
Through an overview of the target decomposition technique based on PolSAR

 S     
(2.1)



cos   SVH SVV   sin cos 
 sin
Using a unitary transformation matrix according to the parameters of the Pauli
rotation matrix, we can express them in a form of unitary vector as follows:
1

1 0
k   
2 0
0


0
cos 2
sin2
0

0
 sin2
cos 2
0

0  S HH  SVV 




0

 f  *  f  *  fV
D
 S
8

0
2 fV 8
0


8

0

3 fV

fS  fD 
8 

f S   f D 

fV

(2.17)


8


*
SVV SVV
 fS  fD

The fS, fD coefficients and  or  parameters can be used to determine the
properties of the target. Finally, the contribution of each scattering mechanism can
be determined for the following spans:

Span  SHH  2 SHV  SVV  PS  PD  PV
2

2



PS  f S 1  

With:

2

;

2



PD  f D 1  


0 

0 

0 d

0

0

b

0

 2


0   fD  0

 *
1




0

a




1 

In which, fS, fD, fV, fH indicate the determined scattering coefficients,
corresponding to the surface scattering, double bounce, volume and helix
scattering mechanisms.
The contribution of each scattering mechanism can be estimated as:
2
2
PS  f S 1   ; PD  f D 1   ; PV  fV  a  b  c  ; PH  f H
(2.29)
2
2
2
Pt  PS  PD  PV  PH  SHH  2 SHV  SVV
2.2. Target identification based on the three-component scattering
decomposition with an adaptive volume modelling
Adaptive algorithm solves the problem of general eigenvalues analysis and
non-negative power constraints, thereby determining the unique minimum value
for volume scattering. Finally, we determine the power of two remaining
scattering components. The generalized double-bounce scattering component
reflects the interaction of electromagnetic waves in both natural and urban areas.










 *
2


 *
 *
 

 
 a d 0

P
 *   V  d b 0 (2.40)

2 
 0 0 c 
 2 

=1+    ;   a  b  c
(2.41)
With , ,  are parameters of the surface scattering and double bounce
scattering models
When  is known, we can calculate the remaining power Pv, and PS.
.P 2

P P 
2 
PD   Pt ; PV  (T33  t .  ); PS  1    T11  D  V a 
(2.49)




T23

T33

 

;



PD 



T33
*

;

PD  0; PV   c



PS  1  

2



Calculate PS., PD and PV using parameters ,  and 
Figure 2.5: The flowchart of proposed algorithm
To evaluate the proposed decomposition method, the experiments were
performed using the full PolSAR data of ESAR (Experimental Synthetic Aperture
Radar) with 3x3m resolution. By testing the area near Oberpfaffenhofen,


10

Germany, Figure 2.6 (a) shows the optical image of this area. The  is the ratio of
asymmetric scattering power to total power. Figure 2.6 (c) shows the color image
of the  coefficient.

Figure 2.6: Survey area, (a) Optical image of Oberpfaffenhofen area,
(b) Pauli image, (c) Image color of correlation coefficient 
A

B

The proposed
decomposition

(a)

(b)

Freeman decomposition

Figure 2.7: Decomposition image of the test area, (a) Color image of the three


Urban

The proposed method

(a)

(d)
Surface scattering

(b)
Freeman decomposition

(e)
Double bounce scattering

(c)

(f)
Volume scattering

Figure 2.9: Pie chart of three components of scattering in surveyed areas.
(a,b,c) the proposed decomposition method, (d,e,f) Freeman decomposition


12

It could be seen that the stability of the proposed decomposition method
compared to the Freeman dicomposition method is more clearly in comparing the
ratio of pixels with non-negative power components to filtering windows of

all parameters related to asymmetric scattering information, the proposed
decomposition model is presented as follows:
C  f S CS  f DCD  fV CV  f asymCasym


 2

f S  0
 *


0
0
0

 2
 

0   f D  0

 *
1 



0
0
0

a


*

2 *
*

2

2

2 *




2 

1 




(2.54)

2
2
where     2  1;   a  b  c; Pasym is the asymmetric scattering
power component; 2   , 2  are both complex numbers corresponding to C12
and C23, The scattering terms can be utilized to describe the general case of
nonreflection symmetric scattering, generally, C12  0 and C23  0 and a  f are

in Figure 2.12 (b). From the results of the proposed decompositon method, the
target identification and determination of the proposed method are more precise
than that in the Yamaguchi decomposition technique.
The proposed method utilise an asymmetric scattering component in order to
improve the limitation of Yamaguchi decomposition in terms of the interpretation,
estimation, and classification of the terrain of the targets. From the above results,
comparisons and analysis, the proposed method shows better improvement of
Yamaguchi's four-component decomposition method. Although the proposed
method still has the disadvantage is the forest areas located bottom right image of
Figure 2.12 (a) misidentification due to asymmetric scattering component exceeds
the power level. In the future, the thesis will continue to research further
improvement and hope for better results aimed at improving the effectiveness of
the proposed method.


14

Forest

Agricultural land
The proposed method

Urban

(a)

(b)
Yamaguchi decomposition

(c)



15

The implementation of the three-stage inversion method is summarized in
Figure 3.1.

Figure 3.1: Three-stage inversion method
The forest height estimation is usually determined after eliminating the
phase component of the surface in complex interferometric coherence coefficients
of the volume scattering component in terms of an inversion space in the unit
circle. A  vol value in the inversion space can be used to recover the pair of value
of height and extintion coefficient.
The definition of average height of tree is:
1 N
hV   hV i
(3.2)
N i1
The root mean square error is defined as:
N

RMSE 

 hV i  hV

i 1

N

2

improved by a coherence method. First, the phase and the parameters of the
scattering object in the tree canopy are determined by the target analysis algorithm
by the adaptive scattering model, then removing the scattering component from
the covariance matrix and the phase of the terrain is determined by the ESPRIT
algorithm. Finally, the forest height is estimated and compensated by the
coherence compensation method. Experimental results show that the accuracy of
forest parameter estimation is significantly improved.
3.2.1. Polarization interference coherence
The data received from the PolInSAR system are usually represented by a
complex coherence matrix 6×6, and are represented in (3.28).


17

T   kk

 T1

 *T


*T




T2 

k 
1


  11  12

  21  22

 0
0


0 


0 

 33 

(3.30)

3.2.2. Scattering mechanisms for PolInSAR data
- Surface scattering
- Double-bounce scattering
- Volume scattering
3.2.3. Estimating scattering parameters of trees
The contracted coherence matrix PolInSAR's is analyzed into the sum of three
sub-matrices corresponding to the three scattering components: volume scattering,
double bounce scattering and surface scattering:
j
  f S e S TS   f D e jD TD   fV e jV TV 
(3.49)
where f S , f D .and fV represent the scattering power coefficient of single bounce,

  0 ; f S  0 ; f D   22  33   11 ; D  arg  22  33   11

(3.50)



Conversely, if  11   33 , then parameters of random scattering
components from tree canopy are directly determined from (3.49)


fV  33 ; V  arg  33
T
TV 33
 v33
FS  f S e jS 
FD  f D e

jD



jV
 ; FV  fV e ;   arcos 




11   22   TV 11  TV 22  FV 




2

  TV 11  TV 22  FV



2

 4  12  TV 12 FV



2

(3.51)


18

3.2.3. Estimate the parameters of terrain using the coherence set
In order to reduce the computational complexity and to improve the
accuracy of the terrain parameter estimation, the thesis proposes the using of the
coherence set to directly estimate the parameters of the terrain.
Algorithm flowchart:
PolSAR
Dataset 1

PolSAR
Dataset 2


hV 

V  arg 

V 0
kZ

 V 0 





0  arg  2  31 L 

 R sin 
4 Bcos  

Figure 3.7: The flowchart of proposed algorithm
Based on the characteristics of contracted coherence matrix in (3.31), we have
a coherence set for PolInSAR data as follows:
(3.52)
app  w*T w : w*T w  1, w  3
Equation (3.52) has the same form as the numerical range of the square
3
matrix A . Therefore, the numerical range of the  matrix can also be
considered as the region of the coherence set.
The interferometric phase of the ground topography is determined as:
0  arg  2  3 1  L 

3
Table 3.2: Forest parameter estimation from simulation data
Parameters Real value Three-stage inversion method Proposed method
hV  m
18
16.0906
17.8397
0  rad 
0.0148
0.0442
0.0158
 dB / m
0.2
0.3168
0.1687
RMSE
0
3.5006
2.3264
Table 3.2 shows the results of comparing the estimated forest parameters
from the proposed method with the three-stage inversion model.

Figure 3.8: Pauli image on RGB coding (a) Pauli image of the survey area,
(b) Graph comparing the forest height of two algorithms
Therefore, from Figure 3.8 (b) and Table 3.2 it could be confirmed that the
proposed algorithm has a higher accuracy than the three-stage inversion model.
The tree height in the testing forest area estimated from the proposed method
is presented in Figure 3.9. For having a deeper evaluation of the effectiveness of
the proposed method, the thesis randomly took 200 pixels in azimuth in the test
area. The main parameters of the forest are estimated by the proposed method

can identify three types of topography: the red areas denoting areas of bare land or
agricultural and roads, the green areas represent forest areas.

4 Figure 3.13: The compared graph of the height results
The Figure 3.13 is the histogram of the height results estimated by the
proposed algorithm compared to the three-stage inversion algorithm through 495
pixels and most of the forest height ranges from 10 m to 25 m.
The tree height in the testing forest area has an average height of 19m. The
estimated height from the proposed method shows that the forest elevation
fluctuates at an elevation of 18.18 m and is greater than the height of 16.46 m of
the three-stage inversion method shown in Figure 3.14.


22

(a)
(b)
Figure 3.14: The estimated forest height of two methods
(a) Three-stage inversion method, (b) The proposed method
5
Table 3.3: Forest parameters estimated from two approachs
Parameters
Three-stage inversion method
Proposed method
hV  m
16.4633
18.1842
0  rad 
-0.3110
-0.1416

estimated from the three-component target decomposition technique for
PolInSAR images to extract the forest height.
3. Combining theoretical and empirical research. Based on the results
achieved with the simulation data, the proposed algorithm will be applied to the
experimental data received from airbonre and satellite remote sensing systems.
In addition, they could be used to directly recover other forest parameters such
as the anisotropy, the random orientation, wave attenuation in the environment.
CONCLUSION
1. The results of the thesis
The content of the thesis "Research on the accuracy improvement of the
target parameters dentification and determination using polarimetric synthetic
aperture Radar and polarimetric interferometric images" has solved the
problem of detecting and identifying targets in natural areas, urban areas and
determining forest elevation. The thesis has studied the Freeman's threecomponent scattering model and Yamaguchi's four component scattering
model, three-stage inversion algorithm and Coherence set theory. As a result,
three basic problems solved in the thesis are:
The first problem: The thesis has proposed two adaptive three-component
algorithms and four extensions with asymmetric scattering models to enhance
the capability of target identification including natural and urban areas.
The second problem: The thesis has proposed an algorithm to improve the
accuracy of forest height estimation using PolInSAR image based on the threestage inversion model, ESPRIT algorithm and coherence set.
The third problem: Three algorithms have been tested, evaluated and
compared to previous algorithms tested with the same PolSAR data set (the
observed area of Oberpfaffenhofen city of Germany) as well as simulation and
PolInSAR satellite data. The theoretical results of algorithms have been
simulated by the actual PolSAR and PolInSAR image data, the simulation
results showed the correctness of the proposed solutions and high applicability