IEEE Geoscience and Remote Sensing Magazine - December 2022 - 44

images suffer from another tough problem: speckle noise in
the magnitude term [58]. As a result, these unique properties
of SAR images greatly increase the difficulty of sparse representation.
To the best of our knowledge, no single dictionary
can be commonly accepted as an ideal option for SAR images.
This is still an open question for the sparse representation
of SAR images, and it is also a big challenge that hinders
practical applications.
To alleviate the aforementioned issues, there are some
powerful solutions. One potential approach is based on the
hypothesis of physical scattering behavior [30], [166]. For
high-resolution SAR images, the scattering response of an
object can be well approximated by decomposing it into several
components from individual reflectors. With respect to
each canonical reflector, the featured scattering response can
be modeled from the specific parameters of frequency, incident
angle, polarization, and so on. In this view, one object
is sparse by consisting of limited reflectors. Implicitly, this
model is flexible for specified objects, such as human-made
targets; however, it can hardly handle complicated scenes.
Hence, this method has some limitations and is not suitable
for dense SAR images.
The alternative approach is to sparsely represent only
the magnitude part by separating the magnitude from the
phase term [18], [20], [21], [33], [34], [43], [109]. In this
way, the random phase can be overcome to some degree,
increasing the compressibility of the SAR image. Then, the
conventional dictionary can be applied to sparsely represent
the magnitude part of the SAR image. In particular, the
redundant dictionary with overcomplete structure admits
specified superiority [119], [121]. Objectively speaking,
the aforementioned methods can moderately relieve the
compressibility of the SAR image. Actually, the achieved
sparsity might not be optimal under a predesigned dictionary
because the SAR image is usually nonstationary
and heterogeneous.
Essentially, the well-known speckle noise is a prime culprit
that adversely affects the compressibility of a SAR image.
In particular, the speckle noise is characterized by signal dependence.
From this perspective, the predesigned dictionary
is not adaptive and cannot handle the nonstationariness and
heterogeneity of a SAR image. Alternatively, the dictionary
learning approach is applied for improved sparsity in a datadriven
manner [115], [167], [168]. Although there are some
preliminary studies on dictionary learning in the context of
SAR imaging, this has not been deeply investigated by fully
exploiting the potential benefits. In fact, dictionary learning
comes from the community of ML. It is very meaningful to
incorporate CS with ML during sparse SAR imaging [79]. The
exploitation of learning methods in ML can be adopted and
developed in sparse SAR imaging, including statistical methods
[114], nonlocal-patch-based processing [139], tensor
representation [136], and DNNs [168].
A more detailed discussion of this issue with a review
of the existing work is presented in the " Feature Enhancement "
section.
44
NONIDEAL FACTOR AND MODEL
MISMATCH CORRECTION
Conventional sparse SAR imaging is based on the hypothesis
that the measurement matrix of SAR observation H can be
accurately or approximately constructed, which is important
to the stable solution of image reconstruction. However, this
precondition cannot always be guaranteed:
◗ For airborne SAR imaging, it is impossible to collect the
ideal synthetic aperture data due to the fact that the radar
platform may not fly in a straight line affected by
the atmospheric disturbance and other nonideal factors
[169], known as motion errors [1], [2]. In this case, the
motion errors tend to produce additional modulation
on the echo data in both the envelope and phase terms.
◗ For ISAR imaging, the motion of a noncooperative target
also introduces undesirable migration through range cells
(MTRC) in the envelope term and high-order phase errors
[170]. In particular, the maneuvering targets usually
exhibit nonuniform rotation and time-variant rotational
axis, which break down FFT-based imaging techniques
[26]. For these aforementioned issues, there are inevitable
model errors and basis mismatch during sparse SAR
imaging, which will severely affect the sparse image reconstruction.
In the case of data compressive sampling,
the traditional algorithms of motion compensation (i.e.,
range cell migration correction and phase adjustment),
such as phase gradient autofocusing (PGA), may no longer
be valid [171].
To alleviate these issues, the embeddable motion compensation
into sparse SAR imaging is suggested, known as sparsity-driven
autofocusing [81], [123], [172]. Meanwhile, several
existing algorithms are proposed to address the sparse maneuvering
target imaging and other important applications
[114], [124], [125], [126].
A more detailed discussion of this issue with a review of
the existing work is presented in the " Motion Estimation and
Parameter Inversion " section.
MULTIDIMENSIONAL DATA APPLICATION
Sparse SAR imaging can also be applied to multidimensional
data, including multiangle, multitemporal, multipolarimetry,
multifrequency, etc. In fact, there is high coherence or correlation
between the multidimensional data. To maintain
the coherence or correlation, it is suggested that a joint processing
strategy be employed on the multidimensional data
during sparse SAR imaging. There has been some interesting
preliminary work to address multidimensional SAR imaging,
such as InSAR/InISAR [113], TomoSAR [116], and so forth.
Compared with independent processing, the proposed joint
multidimensional data processing algorithms, including joint
sparsity and structured sparsity enforcement, can effectively
exploit the correlation and redundant information to improve
the imaging performance, such as robustness to the
limitations of data quality and quantity.
Essentially, apart from a sparse prior, there is an inherent
multimodal structure among the multiway data to
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2022
IEEE Geoscience and Remote Sensing Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2022
