IEEE Geoscience and Remote Sensing Magazine - June 2021 - 12

in the statistical sense that we can design the best SB PU
algorithm. Under this condition, accumulating SB PU
processing experiences from different study cases will
be important for the algorithm design. AI techniques,
particularly deep learning technology, provide us with
the potential framework to accumulate the processing
experience, and a flood of valuable data coming from
different InSAR sensors enables the learning-based PU
method outside of the traditional model-based one. Under
this situation, the SB PU problem has been tackled by
a number of researchers who have attempted to solve it
using different AI techniques.
MB PU, using stacks of interferograms of the same
scene, can be used for restoring the unique PU result,
which refines the traditional SB PU problem from the
ill to the well posed [4], [7]. Nevertheless, many MB PU
methods are not robust enough in the presence of phase
noise due to their mathematical foundation, i.e., their
Chinese remainder theorem (CRT), so they are inapplicable
in practical cases. However, AI technology and machine
learning algorithms can be fed millions of noisy
and clear images simultaneously until they actually learn
what noise is and how best to remove it. In this case, some
researchers have already utilized AI technology and machine
learning methods to improve the noise robustness
of MB PU.
Figure 3 shows the statistics for journal and conference
publications covering AI-based PU from 1995 to 2020.
From Figure 3, we observe no obvious increase in the number
of AI-based PU publications before 2018. The number
of publications began to accelerate after 2018, however,
suggesting a surge of interest in AI-based PU fields. Deep
learning technology has continued to occupy the PU domain
in recent years. These statistics imply that research of
AI-based PU has become a hot spot in InSAR.
In this article, we survey the innovative AI-based PU
methods related to SB PU and MB PU developed during
the past decades. For each method, the experimental results
from simulated and real InSAR data sets are provided
to give readers an overview of AI-based PU processing
potential and limitations. In addition, future prospects
and suggestions in the AI PU domain are pointed out.
Overall, this article aims to provide a comprehensive
road map of the AI-based PU techniques in InSAR for new
researchers and students who wish to familiarize themselves
this topic.
AI IN SB PU
OVERVIEW OF SB PU
The measured result of an object by SB InSAR is given by
() () () ,ss ks2
{} r=(3)
where
()s{ is the wrapped phase of the sth pixel, ()s} is
the unknown unwrapped phase of the sth pixel, and k(s) is
the unknown wrap count (also called the ambiguity number)
of the sth pixel. Obviously, if there is no constraint given
to the solution, SB PU is an ill-posed inverse problem, i.e.,
there is an infinite number of different solutions to the SB
PU problem, so we cannot use only (3) to uniquely achieve
the unwrapped phase. To find the unique PU solution to
the SB PU problem, the Itoh condition (also known as
the phase continuity assumption) is employed by most of
the SB PU methods [8]. In the case that the Itoh condition
holds, the unwrapped phases of all the pixels can be
obtained through a simple flood-fill integration process
using the absolute phase gradients in which the integration
result is independent of the integration path. The
absolute phase gradient between neighboring pixels can
be determined by
t
D -=
---
-
-- +
} ss
(,
*
14
12
10
8
6
4
2
1)
{{
{{
{{
t
where D ss 1- ) is the absolute phase gradient between
neighboring pixels s and pixel s 1- . Usually, there are
} (,
two gradient directions (vertical and horizontal) for the
SB PU problem. Unfortunately, the Itoh condition may
not work very well when it is destroyed by a great deal of
strong phase variation and noise, for which the PU result
depends on the integration path. Under this condition,
it is difficult for SB PU to obtain the correct result.
To solve this problem, PU researchers have designed
and invented some mature concepts and techniques for
SB PU. In the " Residue " and " Branch Cut " sections, we
introduce two mature components in SB PU.
FIGURE 3. The statistics for AI-based PU publications from 1995 to
2020.
12
RESIDUE
Residue is a very classic concept in SB PU and is derived
from input wrapped phase [4], [9]. The presence of residue
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021
() (),
() () ,
() () ,
ss
ss
ss
1
12
12
r
r
|( )( )|
() ()
() ()
{{
{{
{{
ss
ss
ss
----
-
1
1
1
2
1
# r
r ,
r
(4)
The Total Number of Publications
1995
1997
1999
2001
2003
2005
2007
2009
2011
2013
2015
2017
2019
IEEE Geoscience and Remote Sensing Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2021
