IEEE Geoscience and Remote Sensing Magazine - March 2020 - 99

classical SBAS algorithm [5], the original PSI [4] method
was later upgraded to allow for the monitoring of the socalled distributed and partially decorrelating scatterers. The
SqueeSAR technique [109] exploits contextual information
to estimate the sample data covariance matrix through
spatial multilooking: in fact, the scattering mechanism impinges specific characteristics to such a matrix.
To distinguish between the scattering typology (i.e.,
persistent or distributed partially decorrelating scatterers),
SqueeSAR implements a test for the selection of statistical
homogeneous pixels to find regions where the pixel neighbors share similar scattering properties. Following the measurement of the covariance matrix, a maximum likelihood
estimation of the PS-equivalent phase is carried out [110].
Generally, SqueeSAR processing results in an improvement
in the density of monitored targets that is especially remarkable in areas with sparse vegetation.
The exploitation of the covariance matrix was further
revisited in the context of SAR tomography. In [111] and
[112], an approach called component extraction and selection SAR (CAESAR), a decomposition of the covariance
matrix in orthogonal scattering mechanisms, is carried
out to extract principal components. The approach proposed relates multipass interferometric processing and a
signal processing methodology known as principal component analysis.
Inspired by the SqueeSAR selection of multiple looks,
that is, the availability of statistically homogeneous backscattering distributed over multiple pixels, the single-look
GLRT was also extended to the multilook case (MGLRT)
[113]. Similar to SqueeSAR, MGLRT and CAESAR can be
applied to improve monitoring in the interferometric/tomographic processing of partially decorrelating areas characterized by sparse vegetation as well as in urban environments to increase the density of monitored structures and
better discern the distortion effects caused by the layover
effect [113]. In the latter case, a limited number of multilooking factors is used to avoid significantly compromising
the spatial resolution.
The aspect related to the ability to provide results with
a spatially variable scale of analysis is of particular interest for the investigation of landslides and slope instabilities, which, typically, are characterized by the presence of
buildings or infrastructure over vegetated areas. On one
hand, such applications require a wide-area investigation
for rapid detection at the basin scale. On the other hand,
it is necessary to access spatially dense information to
achieve a detailed mapping of the slope instability process
as well as the modeling and understanding of spatial patterns (movement/activity), including kinematic models,
their relationships to causative or triggering factors, and
the investigation of the elements exposed to the hazard.
Scientific developments aimed at limiting the effects of
spatial resolution losses, carried out in the context of adaptive averaging and nonlocal means, were recently reported
[109], [113], [114].
MARCH 2020

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

SEQUENTIAL ANALYSIS AND DECISION STRATEGIES
As previously pointed out, the A-DInSAR literature distinguishes two main processing method classes: PSI and SCI
[4], [5]. Regardless of the class to which it belongs, any processing approach may, in principle, be schematized in a sequential analysis composed of two main steps, as in Figure 2.
The first step of the data analysis is aimed at extracting
the disturbance (APS) and evaluating a first guess of the
NLD, possibly with a nonzero MDV, component. This step
can be implemented at a lower resolution to filter the signal
and, in addition, reduce the computational costs: a hybrid
(spatially adaptive) choice can be considered, as in the case
of [109]. Whatever the resolution of the analysis (i.e., full
or reduced), the first stage
performs a spatial differentiation, separating the low-spaA DECOMPOSITION OF THE
tial-frequency APS from the
COVARIANCE MATRIX IN
(generally small-scale) NLD
ORTHOGONAL SCATTERING
component, based on their
MECHANISMS, IS CARRIED
different temporal behaviors.
OUT TO EXTRACT PRINCIPAL
After data calibration (i.e.,
COMPONENTS.
removal of APS and NLD), a
second step is implemented,
typically at full or close to
full resolution (large scale), to provide the refined NLD
and RT. A pixel-by-pixel (temporal) coherence, that is, the
matching with a model at full or close to full resolution,
is implemented in the complex data domain, as specified
in the analytical description provided in the following
in the context of the decision tests. As already stated, the
exploited model generally accounts for RT and residual
MDV. On this occurrence, depending on the data, the
model might be updated to introduce other contributions
relevant for the phase, such as the azimuth offset of the
scatterer in the presence of high variations of the imaging
aspect angle and the seasonal or quasiperiodic components, that is, for thermal dilation of elements of the built
environment [115], [116].
In all of the described steps, decision tests based on
proper coherence thresholds are used to discriminate between reliable and unreliable measurements. The coherence measurements provided in association with the output product (a result of a first-guess/low-frequency analysis
or a model match at full resolution) can, indeed, be related
to rather different concepts. Accordingly, the annotated coherences may not be comparable to each other at all.
All detectors essentially perform a decision test exploiting a normalized cross correlation:
	

1
K = F A, f ^Ch (5)

where, generally, A is a matrix comprising a model; f is
a transformation of the data-dependent matrix C, the latter being typically associated with the sample (single or
multilook) covariance matrix; and $ , $ is the inner product operator between matrixes defined by the Frobenius
99
IEEE Geoscience and Remote Sensing Magazine - March 2020

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2020
