Persistent Scatterer InSAR and

StaMPS

Andy Hooper
University of Leeds

UNAVCO InSAR Course

August 1--5, 2016
Good Interferogram
2015 M8.3 Illapel earthquake,
Chile

• Good correlation (low

noise)

• Signal is dominated by
deformation

Sentinel--1 data supplied by ESA:

each colour fringe represents 2.8
cm of displacement away from
satellite

2
Unwrapped Good Interferogram

• Can be easily unwrapped

• Deformation dominates

Integrated phase cycles giving

1.5 m displacement

2
 Typical interferograms

Signal dominated by
amosphere, orbit and
DEM errors

(larger than
deformation for low
strains and short
intervals)

100 km Bekaert MSc thesis

2
 Typical interferograms

Signal dominated by High

amosphere, orbit and Decorrelation
DEM errors
(especially for
(larger than long intervals)
deformation for low
strains and short
intervals)

100 km

2
 Persistent Scatter (PS) InSAR

Motivation!
• Allows better selection of coherent
pixels

• DEM error estimation possible

• More reliable phase unwrapping

possible (3--D)

• Atmosphere/orbit errors can be

reduced by filtering in space and
time
A time series analysis approach
• Sub--pixel resolution possible
2
 Improvement of coherence
InSAR (80 looks) Persistent Scatterer InSAR

Bekaert MSc thesis

7
 After unwrapping separation of non-
deformation signals

Signal due to Master DEM error

slow slip on atmosphere
subduction
interface

Bekaert et al, JGR, 2015

8
High resolution PS Processing

Barcelona Olympic Port (Institut de Geomatica)

8
Main Categories of Algorithms

Time Series
InSAR

Persistent Scatterer Small Baseline

Methods Methods

Combined
Methods

10
Persistent Scatterer Methods

Time Series
InSAR

Persistent Scatterer Small Baseline

Methods Methods

Combined
Methods

10
 Cause of Decorrelation

If scatterers move with respect

to each other, the phase sum
changes

Distributed scatterer pixel

(similar effect if incidence angle changes)
 Persistent Scatterer (PS) Pixel

Distributed scatterer pixel “Persistent scatterer” (PS) pixel

13
 PS Interferogram Processing

• All interferograms formed with same “master” image

• No spectral filtering applied (to maximise resolution)

• Oversampling is preferred, to avoid PS being at edge of pixel

• Coregistration can be difficult -- use DEM/orbits or slave--slave

coregistration

• Reduction of interferometric phase using a priori DEM to

minimize ambiguities

13
Interferograms formed

13
Example: single-master interferograms

Sentinel--1, Mexico City, Master date: 3 Oct 2014

13
 Interferometric Phase

For each pixel in each interferogram (after correction for

topography):

int = W{defo + atmos + orbit + topo+ noise}

Atmospheric DEM
Delay Error

Deformation Orbit Error “Noise”

in LOS

W{} = wrapping operator

13
 PS Processing Algorithms

PS
Methods

Temporal Spatial
Model Correlation

• Relying on model of deformation in time: e.g. “Permanent

Scatterers” (Ferretti et al. 2001), Delft approach (Kampes et al.,
2005)

• Relying on correlation in space: StaMPS (Hooper et al. 2004)

18
 PS Processing Algorithms

PS
Methods

Temporal Spatial
Model Correlation

• Relying on model of deformation in time: e.g. “Permanent

Scatterers” (Ferretti et al. 2001), Delft approach (Kampes et al.,
2005)

• Relying on correlation in space: StaMPS (Hooper et al. 2004)

19
“Permanent Scatterer” Technique

San Francisco Bay Area

Ferretti et al, 2004

19
 Double-difference phase

For each pair of pixels in each interferogram:

int = defo + atmos + orbit+ topo+ noise

Atmospheric DEM
Delay Error

Deformation Orbit Error “Noise”

in LOS

19
 Double-difference phase

If pixel pairs are nearby:

int = defo + atmos + orbit+ topo+ noise

Atmospheric DEM
Delay Error

Deformation Orbit Error “Noise”

in LOS

19
 Double-difference phase

If pixel pairs are nearby:

int ≈ defo + topo+ noise

DEM
• model these two termsError
Deformation “Noise”
in LOS

19
Preliminary Network

19
 Initial selection

• Initial selection based on amplitude dispersion (Ferretti et al.,

2001)

True phase here is zero

nA
Imag

    DA
σφ
σn≈ σA A A
Real
A≈μA
Phase noise

Reasonable proxy for small phase noise (<0.25 rad)

19
Preliminary Network

19
Phas Estimation in Time
e

Time

(for each arc between 2 points)

19
Simultaneous Estimation of DEM Errors

Phas
e

Perpendicular Baseline (B )

Constant for each
interferogram

θ is incidence angle, Δh is DEM error,

19
Preliminary Network

19
Integrated results (Las Vegas)

DEM error

Linear deformation rate

30
 Next steps…

• Estimation and interpolation of atmospheric delay

from initial network. This is subtracted from all pixels
• Testing of all other pixels by forming arcs to initial
network
• Filtering in time and space to try and separate
unmodelled deformation from atmosphere

30
Corner Reflector Experiment

30
Corner Reflector InSAR vs Leveling

Marinkovic et al, CEOS SAR workshop, 2004

30
 Results: Bay Area, California

San Francisco Bay Area (Ferretti et al., 2004)

 Works well in urban areas, but not so well in areas
without man--made structures. Why?

30
 Initial Selection

All pixels Best candidates Bad candidates

picked rejected using
e.g. Amplitude phase model
for pixel pairs

35
 Why few pixels picked in rural areas

All pixels Too few “best” Difference in atmospheric

candidates noise between pixels is
large, so unable to reliably
estimate velocity and DEM
error: All pixels rejected
• Lowering the bar for candidate pixels also leads to failure:
too many “bad” pixels for network approach.

35
 Results for Castagnola, Italy

Scarps

PS

Castagnola, Northern Italy (from Paolo Farina)

 Algorithm rejects pixels whose phase histories deviate too much
from a predetermined model for how deformation varies with time

35
 Why few pixels picked when
deformation rate is irregular

All pixels Best candidates Phase model

picked inadequate
e.g. Amplitude due to
deformation

35
 Example of rural area with irregular
deformation

California

Long Valley Volcanic 5km

Caldera

35
Using Temporal Model Algorithm

• 300 high--amplitude persistent scatterers

35
 StaMPS PS Approach
Developed for more general applications, to work:

a) in rural areas without b) when the deformation rate is

buildings (low amplitude) very irregular

35
 PS Processing Algorithms

PS
Methods

Temporal Spatial
Model Correlation

• Relying on correlation in space: StaMPS Hooper et al. (2004, 2007,

2012)

42
 Series of single-master interferograms
• Pre--Processing as for Temporal Model Algorothm

= “Master”

August 2, 43
2016 43
 Spatial Correlation PS Algorithm
Exploits spatial correlation of the deformation signal.

Interferometric phase terms as before:

int = defo + atmos + orbit + topo + noise

Atmospheric DEM
Delay Error

Deformation Orbit Error “Noise”

in LOS

44
 Spatial Correlation PS Algorithm
Exploits spatial correlation of the deformation signal.

Interferometric phase terms as before:

int = defo + atmos + orbit + topo + noise

44
 Spatial Correlation PS Algorithm
Exploits spatial correlation of the deformation signal.

Interferometric phase terms as before:

+  uncorr
topo
+ noise
int = defo + atmos + orbit
+  corr
topo

 Correlated spatially -- estimate by iterative spatial bandpass

filtering

46
46
Estimation of Spatially Correlated Terms

= crude low--pass filter

in spatial domain
(Hooper et al., 2004)
Frequency response

Better (Hooper et al., 2007)

• Low frequencies plus
dominant frequencies in
surrounding patch are
passed.
Example frequency response

e.g., low--pass + adaptive “Goldstein” filter (Goldstein and

Werner, 1998)

47
 Spatial Correlation PS Algorithm

+  uncorr
int = defo + atmos + orbit topo
+ noise
+  topo
corr

 Correlated spatially -- estimate by iterative spatial bandpass

filtering

48
 Spatial Correlation PS Algorithm

+  uncorr
int = defo + atmos + orbit topo
+ noise
+  topo
corr

 Correlated spatially -- estimate by iterative spatial bandpass

filtering

 Correlated with perpendicular baseline -- estimate by inversion

49
 Spatial Correlation PS Algorithm


int -- filtered


0


--1500 --1000Perpend
--5ic
0u0 lar Baseli0
ne (B ) 500 1000

• 1--D problem (as opposed to 2--D with temporal model approach)

Temporal coherence is then estimated from residuals

49
Re-estimation of Spatially Correlated
Terms

Contribution of each pixel weighted based

on its estimated temporal coherence

• Followed by restimation of DEM error and

temporal coherence

• Iterated several times

49
 Selecting PS

-6
x 10
4
PS candidates PDF
3 Random phase PDF

2 95% threshold

0
0 0.2 0.4 0.6 0.8 1
x

Where x is the temporal coherence

52
 Results in Long Valley

• 29,000 persistent scatterers

53
 Wrapped PS Phase

 Interferogram phase, corrected for topographic error

53
 Phase unwrapping

• With temporal model, phase is unwrapped by finding model

parameters that minimise the wrapped residuals between double
difference phase and the model

• If we do not want to assume a temporal model of phase evolution

we need another strategy

55
 2D phase unwrapping
(Points represent PS)
+
B

• Integrate phase differences

between neighboring pixels
A
• Avoid paths where phase
- difference > half cycle

Residues lie on branch cuts

55
 2D Inverse Problem
+ -
+

-
-
? -

+ +

• Connect residues to maximise probability or minimise some norm

55
3D Problem (Sparse)
-
+ - + -

+ -
+ +
-
+ residues in
+
space--space
- -
+ - +
residues in
+ space--time
+
+ - +

55
Unwrapped PS Phase

14 Phase --18

 Not linear in time

55
 Estimation of Atmospheric Signal
And Orbit Errors
 Filtering in time and space, as for temporal model approach

Estimate of atmospheric and orbit errors subtracted, leaving deformation

estimate (not necessarily linear).

55
 Comparison of approaches

Temporal model approach Spatial correlation approach

Long valley caldera

55
 Validation with Ground Truth

 PS show good agreement

55
 Eyjafjallajökull PS time series

T132
cumulative
line--of--sight
displacement

Earthquake
epicentres for each
epoch (Iceland Met
Office)

11.0 --9.7 (cm)

55
 Co/Post-eruptive phase

4
5

−5

−10

−15

3--4
phase difference

−20

−25

−30

−35

−40

−45
20−Mar 11−Apr 22−Apr 05−Jun 16−Jun 19−Jul 30−Jul 10−Aug 01−Sep
Images dates

55
Error estimation
• Because no temporal model is assumed, probability density
functions can be estimated by repeatedly fitting a temporal model
using the percentile bootstrapping method.

Subsidence rates in Bangkok Standard deviations of rates

55
 Comparison PS Algorithms
PS
Methods

Temporal Spatial
Model Correlation

• Spatial correlation algorithm works in more general case,

but may miss PS with non--spatially correlated deformation
• Temporal model algorithm more rigorous in terms of PS
reliability evaluation, but may not work in rural areas, or
where deformation is irregular in time.

55
Comparison PS Algorithms
(Sousa et al, 2010)

Temporal model approach

(DePSI, Ketelaar thesis, 2008)

Spatial coherence approach

(StaMPS, Hooper et al, JGR 2007)

Housing development near Granada, Spain 67

 Interpretation of PS observations
Consider what is actually moving

68
 Persistent Scatterer (PS) InSAR
Summary

• Relies on pixels that exhibit low decorrelation with time

and baseline

• Non--deformation signals are reduced by modelling and

filtering

• PS techniques work best in urban environments, but can

also be applied in rural environments

68