This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1

A New Algorithm for Processing Interferometric

Data-Stacks: SqueeSAR
Alessandro Ferretti, Member, IEEE, Alfio Fumagalli, Fabrizio Novali,
Claudio Prati, Fabio Rocca, and Alessio Rucci

Abstract—Permanent Scatterer SAR Interferometry (PSInSAR) cal decorrelation, often correspond to point-wise scatterers and
aims to identify coherent radar targets exhibiting high phase are typically characterized by high reflectivity values generated
stability over the entire observation time period. These targets by dihedral reflection or simple (deterministic) single-bounce
often correspond to point-wise, man-made objects widely available
over a city, but less present in non-urban areas. To overcome the scattering [11]. Man-made structures, boulders, and outcrops
limits of PSInSAR, analysis of interferometric data-stacks should can all generate good PS.
aim at extracting geophysical parameters not only from point-wise Displacement time-series can be retrieved also for natural
deterministic objects (i.e., PS), but also from distributed scatterers targets where no dominant scatterer can be identified within a
(DS). Rather than developing hybrid processing chains where two certain resolution cell, but typically the estimated time-series
or more algorithms are applied to the same data-stack, and results
are then combined, in this paper we introduce a new approach, of deformation is noisier compared to the signal retrieved for
SqueeSAR, to jointly process PS and DS, taking into account their a point-wise bright scatterer and the spatial density of these
different statistical behavior. As it will be shown, PS and DS can measurement points is typically very low (< 10 PS/sqkm,
be jointly processed without the need for significant changes to depending on the coherence threshold used to define a PS),
the traditional PSInSAR processing chain and without the need to compared to the PS density in urban areas (> 100 PS/sqkm for
unwrap hundreds of interferograms, provided that the coherence
matrix associated with each DS is properly “squeezed” to provide multi-temporal C-band satellite data).
a vector of optimum (wrapped) phase values. Results on real SAR A more thorough analysis shows that, rather than corre-
data, acquired over an Alpine area, challenging for any InSAR sponding to point-wise PS, these measurement points often
analysis, confirm the effectiveness of this new approach. correspond to image pixels belonging to areas of moderate
Index Terms—Coherence matrix, distributed scatterers (DS), coherence in some interferometric pairs of the available data-
InSAR, permanent scatterers, space-adaptive filtering. set, where many neighboring pixels share similar reflectivity
values, as they belong to the same object. These targets, referred
I. I NTRODUCTION to as distributed scatterers (DS), usually correspond to debris
areas, non-cultivated land with short vegetation or desert areas.
S INCE its introduction in the late 1990s, the Permanent
Scatterer (PS) technique [1]–[4] for the processing of
multi-temporal radar acquisitions has gained increasing atten-
Although the average temporal coherence of these natural radar
targets is typically low, due to both temporal and geometrical
decorrelation phenomena [12], the number of pixels sharing the
tion within the scientific community involved in the devel-
same statistical behavior can be high enough to make it possible
opment of interferometric synthetic aperture radar (InSAR)
for a few of them to exceed the coherence threshold and
algorithms. Indeed, PSInSAR1 was the first of a family of tech-
become PS.
nologies now referred to as Persistent Scatterer Interferometry
The need for a data fusion algorithm to properly combine PS
(PSI) [5]–[10].
and DS to increase the density of measurement points has been
PSInSAR aims to identify coherent radar targets exhibiting
already highlighted by many authors [13]–[15]. Theoretically,
high phase stability over the whole time period of observation.
one would like to increase the spatial density of measurement
These targets, only slightly affected by temporal and geometri-
points over areas characterized by DS, while preserving the
high-quality information obtained using the PS technique over
deterministic targets. More precisely, one would like to spatially
average the data over statistically homogeneous areas, increas-
Manuscript received November 12, 2010; revised January 14, 2011; accepted ing the signal-to-noise ratio (SNR), without compromising the
February 20, 2011.
A. Ferretti is with Tele-Rilevamento Europa (T.R.E.), 20149 Milano, Italy identification of coherent point-wise scatterers and possibly
(e-mail: [email protected]). without the need to carry out time-consuming phase unwrap-
A. Fumagalli, F. Novali, and A. Rucci are with the R&D, Tele-Rilevamento
Europa (T.R.E.), 20149 Milan, Italy (e-mail: [email protected];
ping procedures on hundreds of interferograms (as requested in
[email protected]; [email protected]). the conventional SBAS approach [13], [16]).
C. Prati was with the Politecnico, Elettronica ed Informazione, Milan, Italy. Rather than developing hybrid processing chains where two
He is now with DEI Polimi, 20133 Milano, Italy (e-mail: [email protected]).
F. Rocca is with the Elettronica ed Informazione, Politecnico, 20133 Milano,
or more algorithms are applied to the same data-stack, and
Italy (e-mail: [email protected]). results are then combined, in this paper we introduce a new
Digital Object Identifier 10.1109/TGRS.2011.2124465 approach, SqueeSAR2 [17], [18], to jointly process PS and DS,
1 Trademark registered by Politecnico di Milano (POLIMI). Patent granted to
POLIMI. 2 Trademark registered by TRE. Patent pending.

0196-2892/$26.00 © 2011 IEEE

 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

taking into account their different statistical behavior. As it will A. Space Adaptive Filtering: The DespecKS Algorithm
be shown, PS and DS can be jointly processed without the need
Space adaptive processing or variable support filtering is well
for significant changes in the traditional PSInSAR processing
known within the SAR community working on amplitude SAR
chain.
data. Indeed, most speckle filtering techniques aimed at improv-
The paper is organized as follows: Section II provides a
ing radiometric data quality while preserving image details are
statistical characterization of DS and an algorithm which allows
based on a space adaptive filter [24]. Space adaptive filters have
them to be jointly processed with PS. Section III introduces and
also been applied to interferogram filtering [25], as well as to
explains in detail the SqueeSAR algorithm. Real data results
the estimation of complex coherence values. The latest can be
are provided, along with a comparison with PSInSAR results
severely biased whenever the estimation window is comprised
in Section IV, while conclusions and future research efforts are
of pixels belonging to different radar targets, characterized by
drawn in Section V.
different radar signatures [26]. Statistically homogeneous areas
also have to be identified to accurately estimate covariance or
coherency matrixes in PolSAR and PolInSAR data [27], [28]
II. D ISTRIBUTED S CATTERERS and effectively apply target decomposition theorems [28]. In
general, without a proper estimation of local statistics, geophys-
To overcome the limits of PSInSAR and increase the density
ical parameters associated to radar targets cannot be estimated
of measurement points and the quality of the time-series over
effectively.
non-urban areas, the analysis of interferometric data-stacks
The key element of the space adaptive filtering procedure
should aim at extracting geophysical parameters not only on
presented here is the definition of a statistical test capable of dis-
point-wise deterministic objects (i.e., PS), but also on dis-
criminating whether two image pixels belonging to an interfer-
tributed scatterers (DS), whose statistical description can vary
ometric data-stack can be considered statistically homogeneous
significantly as a function of the parameters of the sensor used
or not. Once the proper estimation window has been defined
for the acquisitions and the kind of objects illuminated by the
for each image pixel, by carefully selecting SHP families,
radar signal.
amplitude data can be despeckled, interferometric phase values
The exploitation of “temporary PS” or “semi-PS,” i.e., deter-
can be filtered and coherence values can be estimated properly.
ministic scatterers behaving as PS only on a sub-set of images
It should be noted that, despite the fact that this paper is
of the data-stack [19], although useful and effective in some
focused on the analysis of multi-temporal InSAR data-sets, the
cases, does not significantly change of the performance of the
algorithm described in this section can be applied to general
PSInSAR approach over non-urban areas.
multi-dimensional SAR systems.
Unlike PS, DS corresponding to natural targets are affected
Given a set of N SAR images, supposed properly re-sampled
by temporal, geometrical and volumetric decorrelation [12],
on the same master-grid, let d be the complex data vector
[14], [20] leading to lower SNR values for the interferomet-
ric phase. Consequently, the estimates of the parameters of d(P ) = [d1 (P ), d2 (P ), . . . , dN (P )]T
interest are also noisier (e.g., average target velocity, eleva-
tion, displacement time-series), at least as long as data are where T indicates transposition, P is a generic image pixel,
processed independently pixel-by-pixel, as with conventional di (P ) is the complex reflectivity value of the ith image of the
PS analyzes. data-stack corresponding to pixel P .
If we compare the results of PSInSAR analysis with the For point scatterers, d is an N -dimensional deterministic
differential interferograms generated using an available data- vector, while for distributed radar targets, where no dominant
set, it is often clear that the PSInSAR approach struggles to scatterer can be identified within a resolution cell, d is a
extract all available information for areas characterized by DS, (complex) random vector.
where, however, simple moving average filters can strongly Given two data vectors d(P1 ) and d(P2 ) the two image
improve the SNR of the interferometric fringes at the cost pixels P1 and P2 will be defined as statistically homogenous if
of a decrease in resolution—a well known fact by those who the null hypothesis that the two vectors are drawn from the same
typically apply interferogram stacking techniques [21], [22] or probability distribution function (p.d.f.) cannot be disproved,
SBAS [16], [23]. to a certain required level of significance. Practically, for each
Therefore, under the assumption that radar returns and the image-pixel, statistical tests are applied to all pixels within a
geophysical parameters of interest (e.g., the displacement vec- certain estimation window centered on the pixel under analysis,
tor) are common to all pixels belonging to a certain area, it to carefully select a homogeneous statistical population.
should be desirable to process them jointly to enhance the SNR A possible option for this purpose is the well-known
and improve the quality of any parameter estimation procedure. two-sample Kolmogorov-Smirnov (KS) test [29]–[31]: easy to
Since information associated with point-wise PS should be implement, non parametric, not designed for a specific class
preserved in any filtering procedure (coherent PS should not of probability distribution functions and applicable with high
be averaged with possible low-SNR neighboring pixels), the reliability to data-stacks with as little as 8 images [29]. Since
spatial filter to be implemented should be space adaptive, i.e., the KS test can be applied to real data vectors rather than
able to average statistically homogeneous pixels (SHP) only, complex reflectivity values, and following the conventional PS
without compromising the point-wise information associated to approach where the statistics of amplitude data is used as a
point targets. proxy for phase stability [4], the test is applied to vectors of
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERRETTI et al.: ALGORITHM FOR PROCESSING INTERFEROMETRIC DATA-STACKS: SqueeSAR 3

amplitude values of radar reflectivity, rather than to real and/or

imaginary parts.
More precisely, given a SAR data-stack of N images, the
sorted list of amplitude values (x = |d|) of a certain image-
pixel can be easily converted to an unbiased estimator SN (x)
of the cumulative distribution function (c.d.f.) of the p.d.f. from
which it is drawn that

0, if X < x1
SN (X) = Nk , if xk ≤ X < xk+1 (1)
1, if X ≥ xN
where xi is the ith element of the list of amplitude values.
To evaluate whether two image-pixels P1 and P2 are SHP
(i.e., have amplitude data drawn from the same p.d.f.), the
two-sample KS test measures the maximum value (DN ) of
the absolute difference between their cumulative distribution
P1 P2
functions, SN and SN Fig. 1. Multi-image reflectivity map (incoherent average of 65 RADARSAT
  P1  scenes) of the Alpine region used as a test site. The extent of the area is about
DN = N/2 supx∈R SN (x) − SNP2
(x) . (2) 20 × 25 km.

The distribution of DN can be approximated by the KS

distribution, whose c.d.f. H(t) is given by [29]–[31]

P (DN ≤ t) = H(t)
∞

(−1)n−1 e−2n
2 2
=1 − 2 t
(3)
n=1

and does not depend on the specific c.d.f. of the data. The
KS test considers the two data vectors drawn from the same
statistical population if DN ≤ c, where the threshold c depends
on the fixed significance level α and can be found from the
condition

α = 1 − H(t) (4)

It should be noted that, since the KS test is invariant under

reparametrization of the random variates under analysis, the
same significance of the test would be obtained working on Fig. 2. Number of SHP identified by the KS test.
either a linear or logarithmic scale applied to amplitude data
and using amplitude or intensity values.
Step 3 reduces the number of SHP identified by the KS test
The space–variant algorithm developed for the identification
but is included to increase the probability that nearby pixels
of statistically homogeneous areas, DespecKS,3 can be pre-
belong to the same radar target and share the same geophysical
sented schematically as follows:
parameters.
1) For each image-pixel P , define an estimation window The defined space adaptive filter can be used as: (1) a
centered on P where pixels sharing the same statistics as despeckle filter to be applied either to each SAR image or to
P have to be identified. the average reflectivity map of the area, by simply averaging
2) By applying the two-sample KS test to amplitude data amplitude or intensity values of SHP. The result is a reduction
vectors, all pixels within the estimation window that can of speckle noise over homogeneous areas (such as agricultural
be considered as statistically homogeneous with P are fields, forests, etc.) without affecting point-wise targets (such
selected, given a certain level of significance. as buildings, boulders, and outcrops); and (2) an estimator of
3) Discard image-pixels that, though selected by the KS test, the complex coherence of any interferogram generated from the
are not connected to P directly or through other SHP. available SAR data-set, whenever the number of SHP surround-
4) Pixel P and all SHP within the estimation window ing the pixel under analysis is high enough (see next section).
connected to P are then considered as a homogeneous In Fig. 2 we show the number of SHP identified by the KS
statistical population (identified in the following as set Ω) test for the alpine region represented in Fig. 1 (see Section IV).
for further processing and analysis. The dimensions of the window used to identify SHP is about
a hectare, which corresponds in pixels to a window of 15 ×
3 Trademark registered by TRE. Patent pending. 21. Visual comparison between Figs. 1 and 2 shows how the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

mum number of SHP on which a spatial filtering can be carried

out, it is possible to preserve PS information.
It should be noted that the KS test is just one (and probably
the easiest) option to design the algorithm for SHP identifica-
tion, based on a data-stack of SAR scenes. Although simple and
effective, the KS test is almost insensitive to different extremal
values of the two data vectors under test. For instance, a very
high radar return (i.e., a spike) affecting just one of the two
data vectors will not be enough to differentiate the two pixels
under analysis, that will be considered SHP. Indeed, a limit
of the KS test is its poor sensitivity to deviations from the
hypothesized distribution that occur in the tails. To overcome
this limit, several modified KS tests have been proposed [31].
The basic idea is to introduce a weighting function into the
test to make it consistent with respect to local deviations that
may occur in the tails or in the middle of the distribution [32],
[33]. The drawback with this formulation is in determining
the reasonable criteria for the choice of optimum weighting
functions. Although the Anderson–Darling test [31], [34] can
Fig. 3. Comparison between: (a) a close-up of a reflectivity image (single represent another interesting option to consider.
SAR scene, amplitude data), and (b) its filtered version obtained by applying
the DespecKS algorithm.
B. Statistical Description of Distributed Scatterers
The DespecKS algorithm, described in the previous para-
graph, is an important tool to properly estimate the statistical
parameters of each DS. In fact, under the Gaussian scattering
assumption based on the central limit theorem [35], SAR data
vector can be described by a zero-mean, multi-dimensional,
complex Gaussian p.d.f. Thus, for a complete statistical char-
acterization of a DS it is sufficient to know the covariance (or
correlation) matrix.
After identification of the SHP of pixel P , it is possible to
estimate the sample covariance matrix given by
1  
C(P ) = E[ddH ] ≈ d(P )d(P )H = C (5)
|Ω|
P ∈Ω

where H indicates Hermitian conjugation and Ω is the set of

SHP used in the sample estimate of the covariance matrix.
Fig. 4. Histogram of the number of SHP for targets exhibiting a temporal It is worth noting how the principal diagonal of the covari-
coherence higher than 0.8.
ance matrix of any DS is actually a data vector of N despeckled
(because spatially averaged) intensity values of the N available
KS test indentifies a high number of SHP over homogenous scenes, while the phase values of the off-diagonal complex
areas, while in urban areas and ridges, the number of SHP is elements of Ĉ correspond to spatially filtered interferograms.
very low. A thorough analysis of the properties of covariance matrixes
The efficiency of the despeckle filtering is proven by a associated to single-polarization multi-temporal SAR data-sets
comparison between Fig. 3(a) and (b), in fact we notice a is beyond the scope of this paper [40]. Suffice it to say, however,
significant reduction in speckle noise over non-urban areas that since the covariance matrix and its eigen-decomposition
without blurring the point scatterers present in the image, as have been exploited successfully in a PolSAR and PolInSAR
can be seen in Fig. 3(b). Where, for example, the sidelobes of context for classification purposes [28], most of the algorithms
strong scatterers are preserved. developed for polarimetric data could be applied to Ĉ.
Analysis of the distribution of SHP applied to the PS popu- Whenever amplitude data are normalized such that
lation shows that most of the PS correspond to isolated pixels. E[|di |2 ] = 1, for all i, a coherence matrix Γ is obtained.
More precisely, results on real data (Section IV) have shown Working on coherence matrixes rather than covariance matrixes
that 95% of PS exhibiting a temporal coherence greater than can be beneficial to compensate for possible backscattered
0.8 have less than 20 SHP (see Fig. 4), corresponding to about power unbalances among all the images (similarly to the
8 independent looks, in SAR data interpolated by a factor of computation of the temporal coherence in PSInSAR, where
2 in range direction only. Therefore, by fixing 20 as the mini- all amplitude values are normalized). Absolute values of the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERRETTI et al.: ALGORITHM FOR PROCESSING INTERFEROMETRIC DATA-STACKS: SqueeSAR 5

Fig. 5. Two examples of absolute values of coherence matrices estimated from the data, showing how the decorrelation mechanisms can significantly differ
among DS. The shape of the first example of C is typical for those DS which are mainly affected by temporal decorrelation. The second kind presents a seasonal
behavior revealing a loss in coherence due to snowfall in the winter period.

off-diagonal elements of Γ̂ are actually an estimate of the Hermitian) are simply the difference in phase values of the N
coherence values [36] for all possible interferograms of the available SAR scenes, so the following equation holds:
data-stack (γkj ), In Fig. 5 we show two examples of C    
exhibiting different decorrelation mechanisms, while the loss ∠ {Γ}  nm · {Γ}
 nj = ∠ {Γ}  ∗
jm
in coherence of C in Fig. 5(a) is mostly due to temporal
decorrelation. In Fig. 5(b), we mainly notice a seasonal = θn − θj n, m, j = 1, . . . , N (9)
behavior in C, related to snowfall in the winter period.
The phase values of the coherence matrix, as for the sam-
pled covariance matrix, correspond instead to spatially filtered where the operator {·}jm indicates the element of the matrix at
interferometric phases (φkj ) raw j and column m. In other words, in case of deterministic
radar targets, the property of phase triangularity, or phase
consistency, is met by the construction
 = {γkj · eiφkj }
Γ (6)
φnj = φnm − φmj (10)
The correlation matrix is a Hermitian matrix and its eigen-
decomposition is just a sum of outer products of its eigenvectors This is no longer true for a DS. The coherence matrix is not
(un ), weighted by their eigenvalues (σn ) redundant (it is not rank 1) and, in general, (10) does not
hold. Indeed, the filtered interferogram generated by image n

N
and image j is not equal to the phase difference between the
Γ= σn un uH
n (7) filtered interferograms generated using the image pair (n, m)
n=1 and (m, j). Then, for a DS, we are somewhat forced to deal
with N (N − 1)/2 interferometric phase values and not simply
The coherence matrix associated with an ideal PS is equal, N , as with PS.
instead, to a singular matrix having just one non-zero eigen- Incidentally, this is reason why in the SBAS algorithm, even
value and eigenvector and all moduli equal to 1 after successfully phase unwrapping a sub-set of possibly high-
coherence interferograms connecting all images within a data
Γ(P S) = 1eiθ · (1eiθ )H (8) set, the Singular Value Decomposition (SVD) is applied to
the data as a tool to retrieve the time-series of deformation:
spatially filtered phase values are not consistent and 3-D phase
where θ = [θ1 , θ2 , . . . , θN ]T is simply the vector of the phase
unwrapping algorithms cannot be applied to filtered data.
values of the N available images in correspondence to the PS.
A key problem is then related to the estimation of a vector of
Therefore, for an ideal PS, no matter which image pairs are used
N phase values, ϑ = [ϑ1 , ϑ2 , . . . , ϑN ], matching those of the
to generate the interferogram, the coherence is always 1.
off-diagonal elements of Γ̂, properly taking into account the
associated coherence values (i.e., the moduli of Γ̂).
To create the right framework for this optimization, we
C. Phase Triangulation Algorithm
suppose that the coherence matrix of a generic pixel P can be
For a deterministic radar target, such as a PS, the phase expressed as follows:
values of all off-diagonal elements of the coherence matrix
are redundant: the N (N − 1)/2 phase values (the matrix is Γ(P ) = ΘΥΘH (11)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

where: “goodness of fit” measure is the following:

• Υ is an N × N symmetric real-value matrix whose el- N N
1
ements correspond to the coherence values of all the γPTA = eiφnk e−i(ϑn −ϑk ) (15)
interferograms. N 2 − N n=1
k=n
• Θ is an N × N diagonal matrix, Θ = diag{exp(jθ)},
containing the values of the “true” phase values of pixel which can be also written as
P , related to the optical path of the radar beam in each N 
N
2
acquisition. γPTA = Re eiφnk e−i(ϑn −ϑk ) . (16)
N2 − N n=1
Under the assumption that all pixels belonging to Ω are k=n+1

described by the same set of phase values θ, the p.d.f. of the Indeed, γPTA can be seen as an extension, for DS, to the
SHP can then be expressed as [37] temporal coherence [1], [2], [38] widely used in PS analyzes.
−1 H
It should be noted that, by applying the phase triangulation
p(dΩ |θ) ∝ exp −dH
P ΘΥ Θ dP algorithm (PTA) described above, the phase values of each DS
P∈Ω
are actually filtered before any phase unwrapping, taking into
= exp −trace(ΘΥ−1 ΘH Γ̂) (12) account all possible interferograms. Once N optimum phase
values are available, 3-D phase unwrapping algorithm [39] can
and so the maximum likelihood (ML) estimate of θ is obtained then be performed.
by maximizing this p.d.f. or minimizing the absolute value of
its logarithm. III. S QUEE SAR
Since only phase differences appear in Γ̂, phase values can
be estimated up to an arbitrary additive constant. Without loss In the previous sections, the two main concepts applied in
of generality, we can set the interferometric phase of the first the new SqueeSAR approach have been presented. The first
image to zero. one is the DespecKS algorithm, for the identification of SHP
The optimal estimate of the N − 1 phase values, λ = families used in the estimation of the sampled coherence matrix
[0, ϑ2 , . . . , ϑN ]T , is then given by and the statistical characterization of each DS, preserving the
  information of point-wise PS. The second one is the PTA, which
λ̂ = arg max exp −trace(ΦΥ−1 ΦH Γ̂) provides the bridge between PS and DS and making it possible
λ to characterize a DS through N phase values (as for PS), rather
 
= arg max exp −ΛH (Υ−1 ◦ Γ̂)Λ than N (N − 1)/2, at least for DS exhibiting high enough γPTA
λ values.
 
= arg max ΛH (Υ−1 ◦ Γ̂)Λ (13) By applying these tools, it is possible to pre-process the
λ available data set so that both PS and DS can be exploited
successfully, using the standard PSInSAR processing chain.
where:
The SqueeSAR algorithm can be described as follows:
• Φ is an N × N diagonal matrix, Φ = diag{exp(iλ)} 1) Apply the DespecKS algorithm to identify, for each pixel
• Λ is an N -dimensional vector, Λ = exp(iλ) P , the family of SHP. Let NS be the number of SHP.
• ◦ represents the Hadamard (i.e., entry-wise) product. 2) Define DS all those pixels for which NS is larger than a
Since the “true” coherence matrix Υ is unknown, one can use certain threshold.
an estimate of Υ as the absolute value of Γ̂ 3) For all DS, estimate the sample coherence matrix taking
advantage of the SHP families identified in Step 1 above.
   
λ̂ = arg max ΛH |Γ̂|−1 ◦ Γ̂ Λ (14) 4) Apply the PTA algorithm to each coherence matrix asso-
λ ciated to each DS.
5) Select the DS exhibiting a γPTA value higher than a
The drawbacks of this solution are due to the fact that certain threshold and substitute the phase value of the
coherence estimates are biased [26] and the matrix |Γ̂| is not, in original SAR images with their optimized values.
general, positive and definite as with Γ̂. For this reason, before 6) Process the selected DS jointly with the PS using the
matrix inversion, it is usually necessary to insert a damping traditional PSInSAR algorithm for the estimation of the
factor to avoid small negative or null eigenvalues. displacement time-series of each measurement point.
The algorithm requires the minimization of a nonlinear As already mentioned, it is worth pointing out that, using
functional, implying the use of iterative methods. A possible the SqueeSAR approach, the N phase values that best fit the
solution is the BFGS (Broyden–Fletcher–Goldfarb–Shanno) phases of the sample coherence matrix are retrieved before any
algorithm, which is a quasi-Newton method for unconstrained phase unwrapping algorithm: this represents one of the most
nonlinear optimization [29]. Another option to deal with bad- relevant differences compared to many other algorithms, such
conditioned matrixes is to rely on the eigen-decomposition of as SBAS [23]. It is well known phase unwrapping is the most
|Γ̂| and use its generalized inverse or pseudoinverse. delicate and difficult phase of any InSAR processing chain, as
Once the optimal solution has been obtained, the quality of hundreds of interferograms should be successfully unwrapped
the estimated phase values θi should be assessed. A possible to effectively estimate the deformation field. In SqueeSAR, the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERRETTI et al.: ALGORITHM FOR PROCESSING INTERFEROMETRIC DATA-STACKS: SqueeSAR 7

grams to estimate the optimum phase values, not only a subset

of them, selected based on a model for target coherence. In fact,
any interferogram can provide useful information, according to
its coherence level. Since SqueeSAR is based on the analysis of
the correlation matrix, all possible interferograms are generated
and the associated coherence values are computed, no matter
the temporal or geometrical baseline.
The last important aspect concerns the implemented spatial
filtering technique. By means of the DespecKS algorithm, only
SHP-families are averaged, preserving the information associ-
ated to point-wise radar targets, which are identified as isolated
pixels by the test and, as a consequence, are not affected by the
filtering procedure.
The term SqueeSAR is related to the concept of “squeezing”
the information associated to the coherence matrix to retrieve a
vector of optimum phase values to be used for interferometric
analysis.

IV. R ESULTS ON R EAL DATA

To assess the improvements related to the new algorithm,
we selected an Alpine region in Italy, relatively challenging
for InSAR analysis due to its rough topography and vegetation.
65 descending RADARSAT images acquired from April 2003
to March 2008 were processed. All data were interpolated
by a factor of two in range direction and re-sampled on the
same master acquisition (November 19, 2005). In Fig. 1, the
incoherent average of all the SAR scenes (multi-image reflec-
tivity map) is reported. The area is approximately 20 × 25 km
wide.
The data-set was processed using both the standard PSInSAR
approach and the new SqueeSAR algorithm. To implement
the new processing, the DespecKS algorithm was applied to
identify the SHP for each pixel. Whenever the number of
SHP exceeded 20 pixels (NS = 20) the coherence matrix was
estimated and the PTA was applied to retrieve the N optimum
phase values. As one of the primary objectives of SqueeSAR
is to maintain the point-wise radar responses of PS, the value
of NS is based on the results shown in Fig. 4, which shows
how 95% of PS have less than 20 SHP. However, this value
should not be used for all sites as it is dependent on the spatial
resolution of the satellite providing the data.
After PTA, we generate N differential optimally filtered
interferograms. The PTA actually allow the reconstruction of
any differential interferogram of interest, based on the op-
Fig. 6. (a) Reflectivity map of the region considered to compare (b) a full timum phase values estimated from the coherence matrixes.
resolution interferometric phase, (c) the multi-looked interferogram filtered Fig. 6(a)–(c) show how effective the PTA can be. The ML
using a 13 × 25 boxcar kernel and (d) the interferogram reconstructed applying
the PTA algorithm. The normal baseline of the interferogram under analysis is estimation of the phase values can be seen as a spatial filter
600 m and the temporal baseline 200 days. (since SHP are considered, not only the single pixel P) and as a
temporal filter (as all InSAR pairs are properly combined).
optimum N phase values are properly combined before any The new differential interferograms, available after the
phase unwrapping occurs. This is an important procedure as DespecKS and the PTA steps, are used as input data for the
the interferometric phases “reconstructed” by applying the PTA traditional PSInSAR algorithm. By comparing Fig. 7(a) and (b),
are less noisier than the spatially filtered interferometric phases, it is clearly visible how the SqueeSAR approach significantly
making more robust the spatial phase unwrapping procedure increases the spatial density of measurement points (MP),
(see Section IV, Fig. 6). passing from 85 PS/km2 to 450 MP/km2 . It is now possible
Another significant difference with respect to other multi- to better define the footprints of sliding areas and even identify
interferogram algorithms is the use of all possible interfero- new ones, not previously detected by PSInSAR (see Fig. 8).
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 7. Comparison between the LOS velocity map [mm/yr] estimated applying the (a) PSInSAR and the (b) SqueeSAR algorithm. A positive velocity means a
movement of the target toward the satellite.

Fig. 8. LOS velocity measurements [mm/yr], obtained by (a) PSInSAR and (b) SqueeSAR, for the Cancano lakes region.
Another advantage of the new approach concerns the quality starting from the focused and co-registered interferograms. The
of the displacement time-series of the DS. In some cases, some additional required processing time has so far been the main
DS are identified as PS by the PSInSAR processing chain, drawback of the new approach.
making it possible to compare the time-series of deformation
estimated by SqueeSAR and that available with PSInSAR. By
applying SqueeSAR, it is possible to significantly enhance the V. C ONCLUSION
SNR resulting in a less noisy time-series, as shown in Fig. 9. PSInSAR is a powerful tool for monitoring urban areas,
For the area under analysis, mainly a mountainous region, about characterized by a high density of point-wise stable targets orig-
6000 targets are identified as PS by the standard algorithm. Of inated by man-made structures. However, all PSI algorithms
these 6000 targets, 4000 are identified as DS by SqueeSAR. struggle to provide high-quality measurements over non-urban
Thanks to the new processing chain, we measured an average areas characterized by DS rather than PS. On the contrary, other
increase in DS coherence of about 0.1. approaches, like conventional DInSAR, interferogram stacking,
The improvement in quality of the time-series, above the SBAS, etc., can extract information from DS with more ease
optimum estimate of the interferometric phases provided by than from PS, paying in accuracy of the measurements. In
PTA, is also due to the fact that, by increasing the spatial density this paper, we have proposed a new approach, referred to as
of the MP, atmospheric effects can be better estimated and SqueeSAR, to jointly process PS and DS, taking into account
removed. their different statistical behavior. The results provided by the
From a computational point of view, the improvements ob- new algorithm clearly show how a proper synergistic analysis
tained using the SqueeSAR approach require a processing time of PS and DS can significantly improve the density and quality
of about four times longer than the standard PSInSAR analysis, of InSAR measurement points, over non-urban areas. The key-
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERRETTI et al.: ALGORITHM FOR PROCESSING INTERFEROMETRIC DATA-STACKS: SqueeSAR 9

Fig. 9. Comparison between the displacement time series obtained by (a) the PSInSAR and (b) the SqueeSAR approaches for the same target.
point is the statistical characterization of the DS; the study [7] A. Hooper, P. Segall, and H. Zebker, “Persistent scatterer interferometric
of the covariance or coherence matrix associated with the DS synthetic aperture radar for crustal deformation analysis, with application
to Volcán Alcedo, Galápagos,” J. Geophys. Res., vol. 112, p. B07407,
will, in our opinion, be the roadmap for an advanced multi- 2007, DOI: 10.1029/2006JB004763.
interferogram algorithm. The name SqueeSAR resembles the [8] C. Werner, U. Wegmuller, A. Wiesmann, T. Strozzi, G. Sensing, and
very fact that the information content of these matrix should be S. Muri, “Interferometric point target analysis with JERS-1 L-band
SAR data,” in Proc. IGARSS, Toulouse, France, Jul. 21–25, 2003,
actually “squeezed.” pp. 4359–4361.
Noteworthy it is the introduction of two-sample statistical [9] H. A. Zebker, A. P. Shanker, and A. Hooper, “InSAR remote sensing
over decorrelating terrains: Persistent scattering methods,” in Proc. IEEE
test for SHP identification, in a multi-image framework. Apart Radar Conf., 2007, pp. 717–722.
from allowing an effective statistical characterization of DS, the [10] D. A. Schmidt and R. Burgmann, “Time-dependent land uplift and subsi-
DespecKS approach proved to be a key-factor to preserve the dence in the Santa Clara valley, California, from a large interferometric
synthetic aperture radar data set,” J. Geophys. Res., vol. 108, no. B9,
high-quality information associated with PS. p. 2416, 2003, DOI: 10.1029/2002JB002267.
Further work will be devoted to the optimization of the [11] D. Perissin and A. Ferretti, “Urban target recognition by means of re-
DespecKS algorithm, to deal with long-tailed distributions and peated spaceborne SAR images,” IEEE Trans. Geosci. Remote Sens.,
vol. 45, no. 12, pp. 4043–4058, Dec. 2007.
reflectivity spikes, as well as the identification of robust and [12] H. A. Zebker and J. Villasenor, “Decorrelation in interferometric radar
fast optimization algorithm for carrying out the PTA. The echoes,” IEEE Trans. Geosci. Remote Sens., vol. 30, no. 5, pp. 950–959,
computational costs represent, in fact, the main drawback of Sep. 1992.
[13] A. Hooper, “A multi-temporal InSAR method incorporating both persis-
the new processing chain compared to the standard PSInSAR tent scatterer and small baseline approaches,” Geophys. Res. Lett., vol. 35,
analysis. p. L16 302, 2008, DOI: 10.1029/2008GL03465.
[14] F. Rocca, “Modeling interferogram stacks,” IEEE Trans. Geosci. Remote
Sens., vol. 45, no. 10, pp. 3289–3299, Oct. 2007.
ACKNOWLEDGMENT [15] H. A. Zebker and A. P. Shanker, “Geodetic imaging with time series per-
sistent scatterer InSAR,” presented at the American Geophysical Union,
Authors would like to thank all TRE staff for their support in San Francisco, CA, 2008.
[16] R. Lanari, O. Mora, M. Manunta, J. J. Mallorqui, P. Berardino, and
data processing and S. Tebaldini, F. De Zan and A. M. Guarnieri E. Sansosti, “A small-baseline approach for investigating deformations
for fruitful discussions. on full-resolution differential SAR interferograms,” IEEE Trans. Geosci.
Remote Sens., vol. 42, no. 7, pp. 1377–1386, Jul. 2004.
[17] A. Ferretti, A. Fumagalli, F. Novali, C. Prati, F. Rocca, and A. Rucci,
R EFERENCES “Exploitation of distributed scatterers in interferometric data stacks,” pre-
[1] A. Ferretti, C. Prati, and F. Rocca, “Non-uniform motion monitoring using sented at the Int. Geoscience Remote Sensing Symp. (IGARSS) Conf.,
the permanent scatterers technique,” in Proc. 2nd Int. Workshop ERS SAR Cape Town, South Africa, 2009.
Interferometry, FRINGE, Liège, Belgium, Nov. 10–12, 1999, pp. 1–6. [18] A. Ferretti, A. Fumagalli, F. Novali, C. Prati, F. Rocca, and A. Rucci,
[2] A. Ferretti, C. Prati, and F. Rocca, “Nonlinear subsidence rate estima- “The second generation PSInSAR approach: SqueeSAR,” presented at
tion using permanent scatterers in differential SAR interferometry,” IEEE the Int. Workshop ERS SAR Interferometry (FRINGE), Frascati, Italy,
Trans. Geosci. Remote Sens., vol. 38, no. 5, pp. 2202–2212, Sep. 2000. 2009.
[3] A. Ferretti, C. Prati, and F. Rocca, “Process for radar measurements of the [19] M. Basilico, A. Ferretti, F. Novali, C. Prati, and F. Rocca, “Advances
movement of city areas and landsliding zones,” International Application in permanent scatterers analysis: Semi and temporary PS,” in Proc.
Published Under the Patent Cooperation Treaty (PCT), Nov. 2000. Eur. Conf. Synthetic Aperture Radar, Ulm, Germany, May 25–27, 2004,
[4] A. Ferretti, C. Prati, and F. Rocca, “Permanent scatterers in SAR inter- pp. 349–350.
ferometry,” IEEE Trans. Geosci. Remote Sens., vol. 39, no. 1, pp. 8–20, [20] R. Hanssen, Radar Interferometry: Data Interpretation and Error Analy-
Jan. 2001. sis. Dordrecht, The Netherlands: Kluwer, 2001.
[5] B. Kampes, Radar Interferometry: Persistent Scatterer Technique. [21] D. T. Sandwell and E. J. Price, “Phase gradient approach to stacking
Dordrecht, The Netherlands: Springer-Verlag, 2006. interferograms,” J. Geophys. Res., vol. 103, no. B12, pp. 30 183–30 204,
[6] A. Hooper, H. Zebker, P. Segall, and B. Kampes, “A new method for Dec. 1998.
measuring deformation on volcanoes and other natural terrains using In- [22] S. Lyons and D. Sandwell, “Fault creep along the southern San Andreas
SAR persistent scatterers,” Geophys. Res. Lett., vol. 31, p. L23611, 2004. from InSAR, permanent scatterers, and stacking,” J. Geophys. Res.,
DOI: 10.1029/2004GL021737. vol. 108, no. B1, p. 2047, 2003, DOI: 10.1029/2002JB001831.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

[23] P. Berardino, G. Fornaro, R. Lanari, and E. Sansosti, “A new algorithm Alfio Fumagalli was born in Bergamo, Italy, in
for surface deformation monitoring based on small baseline differential 1977. He received the B.S. and M.Sc. degrees
SAR interferograms,” IEEE Trans. Geosci. Remote Sens., vol. 40, no. 11, in telecommunication engineering from POLIMI,
pp. 2375–2383, Nov. 2002. Milano, Italy, in October 2002, working on Digital
[24] R. Touzi, “A review of speckle filtering in the context of estimation elevation model (DEM) generation using differential
theory,” IEEE Trans. Geosci. Remote Sens., vol. 40, no. 11, pp. 2392– SAR interferometry.
2404, Nov. 2002. In 2003, he joined TRE where he worked on
[25] J. S. Lee, S. R. Cloude, K. P. Papathanassiou, M. R. Grunes, and algorithms, software development, and data inte-
I. H. Woodhouse, “Speckle filtering and coherence estimation of po- gration in GIS environment. His research interests
larimetric SAR interferometry data for forest applications,” IEEE Trans. concern nonlinear systems inversion involved in PS
Geosci. Remote Sens., vol. 41, no. 10, pp. 2254–2263, Oct. 2003. Technique and phase unwrapping algorithms. He has
[26] R. Touzi, A. Lopes, J. Bruniquel, and P. Vachon, “Unbiased of the co- been involved in many different projects related to ground surface movements
herence from multi-look SAR data,” IEEE Trans. Geosci. Remote Sens., detection and monitoring.
vol. 68, pp. 662–664, May. 1996.
[27] G. Vasile, E. Trouve, J.-S. Lee, and V. Buzuloiu, “Intensity-driven
adaptive-neighborhood technique for polarimetric and interferometric
SAR parameters estimation,” IEEE Trans. Geosci. Remote Sens., vol. 44,
no. 6, pp. 1609–1621, Jun. 2006.
[28] J. Lee and E. Pottier, Polarimetric Imaging: From Basics to Applications. Fabrizio Novali was born in Bergamo, Italy, in 1975.
Boca Raton, FL: CRC Press, 2009. He received the B.S. and M.Sc. degrees (cum laude)
[29] M. A. Stephens, “Use of the Kolmogorov-Smirnov, Cramér-Von Mises in telecommunication engineering at POLIMI,
and related statistics without extensive tables,” J. R. Stat. Soc. Ser. B Milano, Italy, in April 2000, working on atmospheric
(Methodological), vol. 32, no. 1, pp. 115–122, 1970. effects estimation and removal in differential SAR
[30] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numer- interferometry.
ical Recipes in C: The Art of Scientific Computing. Cambridge, U.K.: In 2001, he joined TRE where he developed
Cambridge Univ. Press, 1988, ch. 14. algorithms and associated software. His research
[31] P. Kwam and B. Vidakovic, Nonparametric Statistics With Applications interests include the assessment and estimation of
to Science and Engineering. Hoboken, NJ: Wiley, 2007.
atmospheric distortions and the detection of non-
[32] D. Manson and J. Shuenemeyer, “A modified KS test sensitive to tails
linear motion within SAR interferogram, using the
alternatives,” Ann. Statist., vol. 11, no. 3, pp. 933–946, Sep. 1983.
PSInSAR approach. He is project team leader of many international projects
[33] P. Révész, “A joint study of the Kolmogorov-Smirnov and the Eicker-
related to ground surface movements detection and monitoring. In 2004, he
Jaeschke statistics,” Statist. Decisions, vol. 1, pp. 57–65, 1982.
was appointed the Research and Development Manager of TRE.
[34] F. W. Scholz and M. A. Stephens, “K-sample Anderson-Darling tests,”
J. Amer. Stat. Assoc., vol. 82, no. 399, pp. 918–924, Sep. 1987.
[35] R. Bamler and P. Harti, “Synthetic aperture radar interferometry,” Inverse
Probl., vol. 14, no. 4, pp. R1–R54, Aug. 1998.
[36] F. De Zan and F. Rocca, “Coherent processing of long series of SAR
images,” in Proc. IGARSS, Seoul, Korea, 2005, pp. 1987–1990.
[37] A. M. Guarnieri and S. Tebaldini, “On the exploitation of targets statistics Claudio Prati was born in Milano, Italy, on March
for SAR interferometry applications,” IEEE Trans. Geosci. Remote Sens., 20, 1958. He is a Full Professor of Telecommunica-
vol. 46, no. 11, pp. 3436–3443, Nov. 2008. tions at the Electronic Department of the POLIMI,
[38] P. Shanker and H. Zebker, “Persistent scatterer selection using maximum Milano, Italy. He presently chairs the Telecommu-
likelihood estimation,” Geophys. Res. Lett., vol. 34, no. 22, p. L22301, nications Study Council at POLIMI. He holds five
Nov. 2007. patents in the field of SAR and SAS data processing.
[39] A. Hooper and H. Zebker, “Phase unwrapping in three dimensions with He published more than 100 papers on SAR data
application to InSAR time series,” J. Opt. Soc. Amer. A, Opt. Image Sci., processing and interferometry. He is also co-founder
vol. 24, no. 9, pp. 2737–2747, Sep. 2007. of Tele-Rilevamento Europa (TRE), a RS spin-off
[40] A. M. Guarnieri and S. Tebaldini, “Hybrid Cramér-Rao bounds for company of POLIMI.
crustal displacement field estimators in SAR interferometry,” IEEE Signal Mr. Prati has been awarded with two prizes
Process. Lett., vol. 14, no. 12, pp. 1012–1015, Dec. 2007. from the IEEE Geoscience and Remote Sensing Society (IGARSS’89 and
IGARSS’99).

Alessandro Ferretti (M’03) was born in Milan, Fabio Rocca (Specialized Graduate in Electronic
Italy, in 1968. He received the B.S. degree (cum Engineering, 1962) is Professor of Digital Signal
laude) in electronic engineering at the Politecnico Processing at POLIMI, Milano, Italy. He researched
di Milano (POLIMI) Technical University, Milano in digital signal processing for television bandwidth
and the M.Sc. degree in information technology from compression, emission tomography, seismic data
CEFRIEL, Milano, working on digital signal pro- processing, and SAR. From 1978 to 1988, he was a
cessing, in 1993 and the D.Sc. degree in electrical Visiting Professor at Stanford University, where he
engineering from the POLIMI, Milano, in July 1997. became the Department Chair from 1975 to 1978
In May 1994, he joined the radar group of the Po- and Commissione d’Ateneo from 1980 to 1993. He
litecnico working alongside professors Fabio Rocca was also a cofounder of two small technological
and Claudio Prati on satellite radar interferometry companies, Telerilevamento Europa and Aresys, a
and digital elevation model reconstruction. After devoting most of his research spin-off of POLIMI. He was the Past President of the EAEG, an Honorary
efforts on multi-temporal SAR data stacks at the Department of Electronics, Member of SEG (1989) and EAGE (1998), and President of OGS Trieste from
he developed together with Prof. Rocca and Prof. Prati what is now called the 1982 to 1983. He was the Coordinator of the first EEC research program in
“Permanent Scatterer Technique,” a technology patented in 1999 that can over- Geosciences and a member of the Scientific Council of IN-OGS and of the
come most of the difficulties encountered in conventional SAR interferometry. SAR Advisory Group of ESA.
In March 2000, he founded, together with Prof. Rocca, Prof. Prati, and POLIMI Mr. Rocca received the following awards: HUSPI (1979), Schlumberger
the company “Tele-Rilevamento Europa” (TRE), first POLIMI spin-off, where (1990), Italgas for Telecommunications (1995), Special SEG Commendation
he is currently the CEO. His research interest include radar data processing, (1998), Eduard Rhein Foundation for Technology (1999), Doctor Honoris
optimization algorithms, data fusion, and use of remote sensing information for Causa in Geophysics, Institut Polytechnique de Lorraine (2001). He was an
Civil Protection applications. Since January 2008, he is Chairman of the Board Associate Editor of Journal of Seismic Exploration and a member of the
of TRE Canada Inc. Editorial Committee of the Oil & Gas Science and Technology.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERRETTI et al.: ALGORITHM FOR PROCESSING INTERFEROMETRIC DATA-STACKS: SqueeSAR 11

Alessio Rucci was born in Milano, Italy, in 1982. He

received the M.Sc. (cum laude) and D.Sc. degrees
in electronic engineering from the POLIMI, Milano,
in 2007 and 2011, respectively. His thesis is on
advanced PInSAR technique and its application in
reservoir monitoring and modeling.
In 2008, he was with Lawrence Berkeley National
Lab as a Visiting Scholar to work on the InSalah
Project for Carbon Capture Sequestration (CCS). In
2011, he joined the R&D division of TRE, working
on the assessment and estimation of atmospheric
effects and surface deformation in non-urban areas.