urban areas
TRANSCRIPT
In SAR images
Urban areas
Outline
Introduction: Urban applications
Polarimetry: specificity of the urban context
Interferometry
POLINSAR
Applications
Conclusion
INTRODUCTION
SAR in urban areas
Urban applications
The world's population is rapidly increasing, especially in urban regions to which many rural inhabitants are migrating
Results in the need for a efficient method of monitoring cities both in developing and developed countries
Present monitoring techniques are inefficient and unable to maintain up-to date information
Demand for settlement detection, urban classification and population estimation.
source: Population Division, UN: World
Population Prospects
Why radar remote sensing?
Traditionnaly, sources of population data are obtained through national census and related statistics. Time consuming and expensive method. Developped countries only conduct a population census at five or ten year interval.
During its early stages of development, remote sensing could not be effectively used due to its low spatial resolution. However, improvement in ground resolution.
A new context volume
variety
veracity
velocity
Spatial resolution
Temporal resolution
Spectral resolution
Airborne, satellite (TerrarSAR-X)
Stripmap, spotlight
L-band, C-band, X-band, P-
band…
Full pol, dual pol
Different resolutions
ALOS RS2 AIRSAR TSX UAVSAR
• Big Data in remote sensing
(Copernicus)
• New opportunity for urban
monitoring
• More and more polarimetric systems
A particular geometry
shadow
forshortening
layover
Range axis
The main mechanisms in urban areas
Mixture of single mechanisms
First case: h/tan θ < L
layover roof shadow ground Ground
a b
c d
a
e
a a+c+d b d e a
θ
h
L
The main mechanisms in urban areas
Mixture of single mechanisms
First case: h/tan θ > L
layover roof shadow ground Ground
a b
c d
a
e
a a+c+d a+c d e a
θ
Some other well known buildings
Question:
Where is the range axis ?
How can we estimate building elevation?
POLARIMETRY
SAR in urban areas
Content
Main mechanisms
Lack of azimuthal symmetry
Orientation angle
Understanding the HV polarization over urban
Results of main classical decompositions
San Francisco images
13
X-band, 1m x 1m X-band, 2m x 6m TerraSAR-X TerraSAR-X
HH-VV
HH+VV
HV-VH
14
X-band, 2mx2m
C-band, 8mx8m
L-band, 10mx10m
HH-VV HV+VH
HH+VV
San Francisco images
AIRSAR
RADARSAT-2 TerraSAR-X
HH VV
ALOS PALSAR
L-band, ?mx?m
Toulouse images
15 RAMSES
TerraSAR-X
HH VV
HH-VV
HH+VV
HV+VH
Polarimetry over urban: two cases
All mechanisms have comparable
amplitudes
Rotation cannot change this feature
Spatial entropy is high
Streets not aligned with the
trajectory
Double bounce effects are higher
than other contributions.
strong intensity makes the double
bounce mechanism dominant :
spatial entropy is low
Streets aligned with the trajectory
The lack of azimuthal symmetry
The orientation angle
Induced either by:
- tilted roof
- dihedral effects non aligned with the azimuth
Azimuth line
Tilted roof Vertical wall
Vertical wall
φ
φ
α
Polarization
orientation angle
shifts computed
from the
polarimetric
SAR data
Fully polarimetric
image of a built-up
housing area
Street pattern. Areas surrounded by
the same colored lines possess similar
alignment
Orientation angle of PiSAR L-band (3/3)
(Sendaï)
Examples of polarizationa angles over San
Francisco images
-40
-30
-20
-10
0
10
20
30
40
X-band TerraSAR-X L-band ALOS-PALSAR
Noise level linked to the frequency bandwidth
X-band: very noisy over vegetation and ocean
L-band: very flat over ocean, noisy over vegetation
21/34
deterministic targets
non-determinitic targets
Coherent
Backscattered signal
Man-made targets:
vehicles, buildings, roads
……
Radar image: bright points
stochastic
Backscattered signals
Natural targets: forest,
meadows, rough surface
…..
Stochastic parameters
Coherent
decomposit
ions
Incoherent
decompositi
ons
Coherent or incoherent targets
different polarimetric decompositions
• Coherent decompositions
• Pauli
• Krogager
• Cameron
• Touzi criterion
• Incoherent decomposition
• Based on eigenspace: Huynen, Barnes and Holm, Cloude Pottier, Holm
• « physical decomposition »: Freeman Durden, Yamaguchi, Van Zyl, Neumann
• Multiplicative decomposition: Lu and Chipman
Application of some of the decompositions and
limitations
- Influence of resolution and wavelength
- Behavior of a 45° tilted builduing block:
classical decompositions fail to indentify it as
urban
- mixing of several orientation effects:
difficult to identidy them
Entropy – alpha - span
RADARSAT 2
TERRASAR-X AIRSAR
ALOS PALSAR
Yamaguchi versus Freeman Durden
Yamaguchi better reduces the volume component
But still fails to identify the 45° tilted block
INTERFEROMETRY
SAR in urban areas
Ambiguity height
27
Can you give an
estimation of the
ambiguity height?
Interferometry over San Francisco
28
Subpixellic coregistration
Orbital fringes removal
Hue : interferometric phase,
Intensity : span:
Saturation : coherence level
Details of interferogram over a building
Comparison single pass – multi pass at X-band
30
Question:
Which images are acquired in repeat pass mode? In
single pass mode?
San Francisco Washington
Comparison single pass – multi pass at X-band
Information is avalaible
only on buildings
Hue : interferometric phase
Intensity : span:
Saturation : coherence level
Interferometric phase at X-band over San Francisco
Interferograms for different baselines
32
Pass 1-Pass 2 22 days Pass 2-Pass 3 11 days
Pass 1-Pass 3 11 days
Question:
Why are these patterns different?
Use of interferometric coherence
33
Intensity
Seems to be an interesting
parameters to
discriminate deterministic
targets! Correlation
After sub pixellic coregistration
Interferometric coherences in the Pauli basis
34
100 200 300 400 500 600 700 800 900 1000
100
200
300
400
500
600
700
800
900
1000
A 1000x1000 pixel
sub image
Interferometric phase Interferometric level in HH+VV
Interferometric level in HH-VV Interferometric level in HV
Discrimination of buildings
35
Simple threesholding on the optimal
interferometric coherence
Intensity after equalization Optimal coherence level after optimization
INTERFEROMETRY AND
POLARIMETRY
A general simple model of bright points
Modelling of the coherence set
Multipass interferometry over urban
Monopass interferometry over urban
SAR in urban areas
Generalized coherence
37
hH
vVhV
Applications :
-coherence optimization for
the estimation of heights
- target analysis: how to get
the maximum information
- separation and
interpretation of different
heights
2211
12)(TT
T
22221111
212121 ),(
TT
T
2112
2222
1111
kkT
kkT
kkTCoherence
matrices
3 x 3
• generalized coherence
« N bright points » Model
38
NA jxy
NN
jxy
AA
xyescescE
11 44
1
NA jxy
NN
jxy
AA
xyescescE
22 44
2
cSE
1
cDSE
2
Nj
j
j
e
e
e
D
0
02
1
Matrix expression:
Backscattered field of N bright points within a resolution cell
AA sc
A1
A2
cSk1 cDSk2
Nj
j
j
e
e
e
0
02
1
D
Selection of one mechanism
1
0
0
A
002
12
100
2
11
2
1
S
0
1
1
2
1Cs
1
0
1
2
1As
0
0
1
Bs
012
210
0101H
)(SM
1
1
1
B
0
1
0
C
C12C
B12B
A12A
ωTω
ωTω
ωTω
H
C
H
B
H
A
arg
arg
arg
To estimate the interferometric height of one mechanism:
Choose it in the space orthogonal to the space spanned by the other scattering vectors:
Application to a resolution cell with 3 bright points
40
0
1
1
2
1Cs
1
0
1
2
1As
0
0
1
Bs
)(2
1)(
2211
12
TT
T
SDccDSSccS
SDccS
S: diffusion vector
c: amplitudes
D: interferometric angles
coherence set:
white noise is assumed
Noise with Wishart distribution
HccC ● Assumption about the statistical fluctuation: only amplitude noise:
● Resulting coherence set:
● coherence optimization method makes possible the scattering phase centers
separation
)(2
1)(
ωSDccDSωωSccSω
ωSDccSω
2211
12
T
HHHH
T
HH
T
HHHH
Statistics State of the Art
Limitations
On the statistical hypothesis
Does not allow to simulate only one scatterer or two scatterer cell
(coherence matrices are not full rank)
In practice, coherence sets do not intersect the unitary circle. How to
"inverse" such coherence shapes ?
Correlated Statistical
Variations of D and S:
- coherence levels decrease
- the estimates of interferometric
phases are biased.
Example of cases
A
B
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
Towards more sophisticated
models
One type of scatterer
Example : ground (1000 samples)
Mathematical modelling
Observation on a real case of the different contributions:
ii
11 xki
1 iiji
iAe 222 xk
-1 -0.5 0 0.5 1
-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
20
40
60
80
30
210
60
240
90
270
120
300
150
330
180 0
A
ii
21
Amplitudes are
perfectly correlated
0 0.5 1 1.50
10
20
30
40
50
60
70
80
90
2
1
x
xX
HXXM
c
b
a
00
00
00
,mmm
mmM
Correlation between polarimetric pair
Statistical hypothesis
Associated coherence
Shape
With no correlation between polarimetric vectors (1)
With two equal polarimetric vectors (maximum interferometric correlation) (2)
With a covariance matrix for and non zero extradiagonal elements (3)
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
ii
21 xx
021 Hji
xx
021 Hji
xx
2
1
x
xX
(1) (2)
(3)
021 Hji
xx
021 Hji
xxii
21 xx
Modelling of a 2 bright point cell
Mathematical modelling
Mixture of two different polarimetric statistics on points A and B:
Results of simulations:
Very low influence of statistics on C and D
Predominance of statistics on S
i
B
i
Ai
B
i
A
1
111 ,
xxk
i
1
i
B
i
A
j
ji
B
i
A iB
iA
e
e
2
2222
0
0,
xxki
11 cS22 cDS
A
A
A
2
1
x
xX
H
AAA XXM H
BBB XXM
B
B
B
2
1
x
xX
iB
i
A 11 ,xxS1
iB
i
A 22 ,xxS2
Point A Point B
Global level of coherence
Internal description of and
,
0000
0000
0000
0000
0000
0000
,
csc
bsb
asa
scc
sbb
saa
BAM
BA MM
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
Increasing s
Estimation of coherence on real data
Ground segments and building segments
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
estimation classique
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
-1 -0.5 0 0.5 1-1
-0.5
0
0.5
1
0.2
0.4
0.6
0.8
1
30
210
60
240
90
270
120
300
150
330
180 0
Re
Im
Building
segments with
polarimetric
diversity
Bare soil
segments
Building
segments without
internal
polarimetric
diversity
49/34
Polarimetric
composition
aerial photography
Parking lot
buildings
Calibration
trihedral
corners
trees red=hh+vv,
green=hh-vv,
blue= 2 hv
rang
e
azimuth
Examples of first Results
50/34
separation of different scatterers
heights in the optimal polarimetric
basis
Estimated height using the first
mechanism of the optimal basis
(ground)
Estimated heights using the second
mechanism of the optimal basis (roof)
Better accuracy of the DEM
Polarimetric
coherence
optimization
Average in the
dual space
Examples of first Results
Example
Example : ground + roof
A
B
Double bounce
Roof scattering
ground
Roof diffraction
The generalized coherence set
APPLICATIONS
Classification
3D rendering
Subsidence
Classification
Using classical polarimetric parameters: covariance
matrix elements, and H/a/A decomposition
qualitative performances …
55
San Francisco Toulouse
Is 3D rendering possible using multipass ?
56
Hue : interferometric phase,
Intensity : span: Saturation :
coherence level
The problem for 3D rendering is that
we can estimate top heights of buildings,
but not the elevation of ground !
3D rendering
Height estimation
Segmentation and 3D estimation
Google Data
Root mean squared error
Tomography
subsidence
Why subsidence in urban areas ? may be caused by factors including
• groundwater extraction
• load of constructions
• natural consolidation of alluvium soil
• geotectonic subsidence
Monitoring of land subsidence in suspected cities is required for
groundwater extraction regulation,
effective flood control and seawater intrusion,
conservation of environment
construction of infrastructure, and spatial development planning in general.
Bologne
Ast
riu
m G
EO-I
nfo
rma
tio
n S
ervi
ces
Example over Venice
Contribution of polarimetry to PSI: increase the number of PSC
by optimising the quality criteria
Example in Murcia (Spain), with 45 TerraSAR-X images HHVV
Coherence: 60%
Amplitude: 170%
Increase in number of
pixels over single-pol:
Contribution of POLSAR to PSI
Urban updating
Conclusion
• SAR remote sensing interesting in urban areas for:
– subsidence and deformation
– Change detection
– 3D updating
• Perspectives:
– Better resolutions
– Increased revisit times