Compressive Sensing and Generalized Likelihood Ratio Test in … · 2016-10-07 · CS and GLRT in SAR Tomography Compressive Sensing and Generalized Likelihood Ratio Test in SAR Tomography

CS and GLRT in SAR Tomography

Compressive Sensing and Generalized LikelihoodRatio Test in SAR Tomography

G. Fornaro1,2, A. Pauciullo1, D. Reale1, M. Weiss3, A. Budillon2, G. Schirinzi2

1 - Institute for Electromagnetic Sensing of the Environment (IREA)

National Research Council (CNR), Naples, Italy

2 - University of Napoli “Parthenope”

Department of Engineering, Naples, Italy

3- Fraunhofer FHR

Wachtberg, Germany.

CS and GLRT in SAR Tomography 2

Outline

� Synthetic Aperture Radar (SAR) Tomography and motivation for super-resolution

� Compressive Sensing in Tomo-SAR

� Scatterer detection problem Tomo-SAR . Generalized Likelihood Ratio Test (GLRT) schemes: the GLRT with cancellation (GLRT-C) and the support based GLRT (sup-GLRT) approaches

� Application to real data: CS, GLRT-C vs sup-GLRT and sup-GLR vs CS

� Conclusions and future works

CS and GLRT in SAR Tomography 3

SAR Tomography for full 3D Imaging

� SAR Interferometry and Differential SAR Interferometry has important applicationsin Digital Elevation Model (DEM) reconstruction and monitoring of deformation.

� SAR Tomography extends interferometric approaches for application to complexscenarios.

� By synthesizing an antenna also in the slant height direction (orthogonal tothe line of sight) it is possible to analyze the vertical structure of the scatteringthus extending SAR imaging form 2D (azimuth-slant range) to 3D (azimuth-slantrange-slant height).

CS and GLRT in SAR Tomography 4

3D SAR Imaging

( )∫−

−=max

max

2

s

s

sjn dsesx nπξγ ( ) Nnrbnn ,....12 == λξ

r

s FOURIER INVERSION FROM IRREGULAR SAMPLES:

� BeamForming (BF)

� Regularized inversion (SVD)

� Adaptive Beamforming (Capon)

� Compressive sensing (CS)1,2

signal to the n-th antenna

backscattering distribution

along the slant height

N acquisitions with spatial (orthogonal) baseline distribution Nbb ......1

( )Brs 2λ=∆

RAYLEIGH

RESOLUTION

minmax bbB −=

1. A. Budillon, A. Evangelista, G. Schirinzi. Three-Dimensional SAR Focusing From Multipass

Signals Using Compressive Sampling. IEEE Trans. Geosci. Remote Sens., 49 (1):488-499,

2011

2. X. X. Zhu, and R. Bamler. Tomographic SAR Inversion by L1 -Norm Regularization—The

Compressive Sensing Approach. IEEE Trans. Geosci. Remote Sens., 48(10):3839-3846, 2010.

CS and GLRT in SAR Tomography

INVERSION

5

[ ]∫ −= − dssjsNx nn πξγ 2exp)(21

Beamforming – Matched filter

( ) xaxAγHH ˆˆmms =⇒= γ

[ ]Nxx ....1T =x N measurements

Aγγx =

=−− LNN sjsj

eeN πζπζ 22 1

111

DISCRETIZATION

[ ]LaaA ...1= sensing matrix

{ }Ll ssss ,..,,...,1∈ L (bins): L≥N typically L>>N

[ ] Ne lN sjl

πξ2T ...1−=a

steering vector: response or “firm” of a scatterer at a given height

CS and GLRT in SAR Tomography 6

4D SAR Imaging (Differential SAR Tomography)

( )( )

∫−

−−=max

max

,4

2

s

s

tsdjsj

n dseesxn

n λπ

πξγ

Signal to the n-th antenna

Deformation term

r

s

N acquisitions with spatial baseline distribution and temporal distributionNbb ......1 Ntt ......1

( )∫ ∫− −

−−=max

max

max

max

22,

s

s

v

v

vjsjn dvdsevsax nn πηπξ

γ

( )rbt nnnn λξλη 22 ==

( )( )∫

−

−−=

max

max

2,

4

,

v

v

vjtsdj

dvevsae nn πηλ

π

λη nn t2=

CS and GLRT in SAR Tomography 7

Compressive Sensing

Result: if γγγγ is a sparse signal (can be well represented as few non zero contributions) then,

under the condition that a sufficient number of samples are acquired, the unknown can be

well reconstructed via the compressive sensing technique.

[ ] γaaAγx L...1== [ ] Ne lN sj

l

πζ2T ...1−=a

Hypothesis: the number K of non-zero entries of γγγγ ( )

is sufficiently smaller than L,N (K<<N<<L)

K=0

γ Aγx =

1γ

2γ

ε<−=20

tosubjectminargˆ Aγxγγ

γ

L >> N i.e. few measurements (N) and many unknown (L)

NP Complete problem, i.e. requires the check of all L!/K!/(L-K)! combinations

{ }12

minargˆ γAγxγγ

δ+−=

Under “some” mild hypothesis the solution above is equal to that of one of the following (equivalent) convex problems:

Basis Pursuit De-Noising (BPDN)

ε<−=21

tosubjectminargˆ Aγxγγγ

CS and GLRT in SAR Tomography 8

Scatterer Detection Problem

Key elements: the False Alarm Probability (FAP) and the Detection Probability (DP).

waax

wax

wx

++=

+=

=

2211

11

:

:

:

γγ

γ

2

1

0

H

H

H21; γγ are the scatterers’ reflectivities

Generalized Likelihood Ratio Test (GLRT) is required.

The maximization at the numerator selects the highest peak of the beamforming and normalizes to the data vector norm.

( ) ( ) ( ) ( )T

H

H

0

1

2

2H

2

2H

2

2HH

2

2H

22

2H

min

1

maxmaxmaxmax

<>−====

⊥

x

xPx

x

Pxx

x

xpapax

x

xpa

xa

xpappppp

( ) HH1HaaIaaaaIP −=−=

−⊥

w is the additive noise

� A. De Maio, G. Fornaro, A. Pauciullo, Detection of Single Scatterers in Multidimensional SAR Imaging, IEEE Trans. Geosci. Remote

Sens.,Vol. 47, No. 7, pp. 2284-2297, Jul. 2009

Letting p to collect the unknown parameters (height, height/velocity, …) for a

Gaussian model the discrimination between the first two hypotheses is achieved as:

CS and GLRT in SAR Tomography 9

Sequential GLRT with cancellation (GLRT-C)

2

2

1H

2

2

1H

22

2

1H

11

1

1

min

1)(

max

)(max

)(x

xPx

x

xpa

xa

xpa

xp

p

p

⊥

−===L

Estimation from the data of the direction of the first scatterer and evaluation of the coherence L1(x)

Projection of the data in the complement orthogonal to the subspace spanned by the estimated direction

Estimation from the projected data of the direction of the second scatterer and evaluation of the coherence L2(x)

)ˆ(ˆ

ˆˆˆ

11

H111

paa

aaIP

=

−=⊥

2

c

2

2c2

2

c2H

c2

2

)(

)(max)(

2 xpa

xpax

p=L

Advantages: requires only one dimensional maximizations (computational time)

Disadvantages: no super-resolution

212c

1c

ˆ

ˆ

aPa

xPx

⊥

⊥

=

=

� A. Pauciullo, D. Reale, A. De Maio, G. Fornaro, Detection of Double Scatterers in SAR Tomography, IEEE Trans. Geosci.

Remote Sens., Vol. 50, No. 9, pp. 3567-3586, Sept. 2012

CS and GLRT in SAR Tomography

Scheme of the GLRT-C detector

22 )( TL >x 11 )( TL >xFalse False

True True

x0

H

2H

1H

The thresholds T1 and T2 are set based on the fixed values of PFA, on the first

and on the second scatterers

xpap

)(max 1H

1

1p̂

CS and GLRT in SAR Tomography 11

sup-GLRT: an advance over the GLRT-C

2

1H

2

21H

,

2

)(min

),(min

1

1

21

xpPx

xppPx

p

pp

⊥

⊥

−=gML estimation

Advantages: allows super-resolution (i.e. detection of targets below the Rayleigh resolution)

Disadvantages: computationally demanding

xx

xppPxpp

H

2

21H

,

1

),(min

1 21

⊥

−=g

11

0

0

TgH

H

<>

22

1

2

TgH

H

<>

2

21H

,

2

21H

,),(maxarg),(minarg

2121

xppPxxppPxpppp

=⊥

waax

wax

wx

++=

+=

=

2211

11

:

:

:

γγ

γ

2

1

0

H

H

H

2

1H

2

1H )(maxarg)(minarg

11

xpPxxpPxpp

=⊥

sup-GLRT

11 Tg >

� A. Budillon, G. Schirinzi, GLRT Based on Support Estimation for Multiple Scatterers Detection in SAR Tomography, IEEE

Journ. Select. Topic. Appl. Earth Observ. Remote Sens., Vol. 9, No. 3, pp. 1086-1094, March 2016

CS and GLRT in SAR Tomography 12

• 25 TerraSAR-X Spotlight acquisitions over the city of Las Vegas USA (from 2008. 02. 02 to 2009. 04. 06)

• Imaging Mode: HS (High Resolution Spotlight)

• Orbit Direction: Ascending

• Beam Identification: spot_042

• Orbit Number: 3522

• Incidence Angle: 35.8°

• Look Direction: Right

• Azimuth resolution: ∼ 1.1 meters

• Slant Range resolution: ∼ 0.6 meters

• Polarisation Mode: Single

• Polarisation: VV

The TERRASAR-X dataset over Las Vegas

CS and GLRT in SAR Tomography 13

Rayleigh resolution: 40m (slant height) 27m (vertical)

Acquisition distribution of the Las Vegas dataset

CS and GLRT in SAR Tomography

CS and GLRT in SAR Tomography 15

Application to high resolution data (The Mirage Hotel)

CS and GLRT in SAR Tomography 16

CS Detected Single Scatterers

CS and GLRT in SAR Tomography 17

CS Detected Double Scatterers (lower)

CS and GLRT in SAR Tomography 18

CS Detected Double Scatterers (higher)

CS and GLRT in SAR Tomography 19

CS Detected Double Scatterers: height difference

CS and GLRT in SAR Tomography 20

sup-GLRT Detected Single Scatterers

CS and GLRT in SAR Tomography 21

GLRT-C Detected Single Scatterers

CS and GLRT in SAR Tomography 22

sup-GLRT Detected Double Scatterers(lower)

CS and GLRT in SAR Tomography 23

GLRT-C Detected Double Scatterers(lower)

CS and GLRT in SAR Tomography 24

Sup-GLRT Detected Double Scatterers(higher)

CS and GLRT in SAR Tomography 25

GLRT-C Detected Double Scatterers(higher)

CS and GLRT in SAR Tomography 26

sup-GLRT Detected Double Scatterers: height difference

CS and GLRT in SAR Tomography 27

GLRT-C Detected Double Scatterers: height difference

CS and GLRT in SAR Tomography 28

Topographic Difference Between CS and sup-GLRT

Histogram of the difference between the height of double scatterers estimated by CS and sup-GLRT (mask of points detected by sup-GLRT)

CS and GLRT in SAR Tomography 29

Deformation Mean Velocity Difference between CS and sup-GLRT

Histogram of the difference between the deformation mean velocity of double scatterers estimated by CS and sup-GLRT (mask of points detected by sup-GLRT)

CS and GLRT in SAR Tomography 30

CS Post Detection Single Scatterers

CS and GLRT in SAR Tomography 31

Sup-GLRT Single Scatterers

CS and GLRT in SAR Tomography 32

CS Post Detected Double Scatterers(lower)

CS and GLRT in SAR Tomography 33

Sup-GLRT Double Scatterers (lower)

CS and GLRT in SAR Tomography 34

CS Post Detected Double Scatterers(higher)

CS and GLRT in SAR Tomography 35

Sup-GLRT Double Scatterers (higher)

CS and GLRT in SAR Tomography 36

Conclusions and future works

SAR Tomography allows implementing a radar scanner from the space to reconstruct 3D point clouds and monitor deformations.

Next generation VHR sensors (COSMO-SkyMED II Generation, HRWS) will allow further improving this technology for application to urban area and critical infrastructure monitoring.

Super-resolution in SAR tomography allows achieving improvements in the generation of 3D point clouds.

An open issue is the coupling between the reliability of the reconstruction and the computational performances. To this end a key point seems to be associated with the “assimilation” of proper detection schemes within computationally efficient L1 methods.

THANK YOU DANKE