IEEE Geoscience and Remote Sensing Magazine - June 2021 - 25

TABLE 3. THE CHARACTERISTICS OF THE AI-BASED SB AND MB PU METHODS.
AI PU
CATEGORY
LEARNING
MODE
AI-based SB PU Supervised
learning
AI-based MB PU Unsupervised
learning
USE OF DATA
Makes the link between input interferogram
and ground-truth data
Does not need ground-truth data
LEARNING
TRUSTWORTHY EXECUTE TIME MAIN DRAWBACK
High
Moderate
Needs offline
training
Takes place in
near real time
Needs exertive work to build up
the complicated training data set
Needs enough interferograms with
suitable normal baseline lengths
kind of network architectures and learning procedures
that are the best fit for the PU problem seems destined
to remain unclear. As far as we are concerned, the reason
is the lack of a shared perspective for researchers in this
field, including a thorough comparative analysis of the
state of the art, which discusses the competing architectures
and learning procedures and allows for a clear understanding
of the main limitations to further improvement.
In addition, InSAR data are crucial for the deep
learning technique-based PU methods to train and test.
Under this condition, a natural question arises: What
kind of InSAR data should be made publicly available
for deep learning-based PU methods to ensure reproducibility
and a fair comparative evaluation? Therefore,
further study needs to focus on research reproducibility
in this field, from the availability of InSAR data sets
(which can be shared in IEEE DataPort) to the sharing
of executable codes (which can be submitted to the IEEE
Remote Sensing Code Library) to the definition of common
experimental protocols and sophisticated tricks
for improvement.
AI FOR PREPROCESSING OF PU
The presence of measurement bias increases PU difficulty
and even leads to failures in the PU process. Therefore, the
preprocessing of PU for enhancing interferogram quality is
a necessary step before PU. To date, some works have proposed
AI techniques for tackling this problem. For example,
[69] and [70] have applied deep learning on interferometric
phase filtering and coherence estimation. In addition, deep
learning is applied on atmospheric noise removal in [71].
Moreover, DCNNs can be used to generate the mask for assisting
in marking low-quality regions in the interferogram,
which will be excluded in the following PU process [72]. Although
some researchers have already looked into this direction,
there are still several scientific questions that need
to be answered, e.g., how well can the AI-based preprocessing
benefit the subsequent PU stage along the processing
pipeline? In addition, how best to establish a systematic AIbased
InSAR signal processing chain for DEM generation or
could deformation extraction could also be an interesting
future research direction.
PU MASTER
AlphaGo is the first program to defeat a world champion in
the game of Go, which is widely viewed as a grand challenge
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
for AI [73]. It was able to do this
using DCNN trained by reinforcement
learning, in which AlphaGo
became its own teacher
without human data, guidance,
or domain knowledge beyond
the game rules. As described
in the " AI in SB PU " section,
branch-cut deployment could
be an NP-hard problem similar
to the game of Go, implying
that it is impossible to efficiently
solve this problem
without losing optimality.
Under this condition, how best to apply the reinforcement
learning by shaping the appropriate reward function into
the PU domain for designing a self-learning AI PU master
program like AlphaGo could be another key future topic of
AI-based PU.
ACKNOWLEDGMENTS
The authors would like to thank the editors and the anonymous
reviewers for their valuable suggestions, which have
helped to improve the quality of this article. This article
was supported, in part, by the National Science Fund for
Distinguished Young Scholars under grant 61825105 and,
in part, by National Natural Science Foundation of China
under grant 42071438.
AUTHOR INFORMATION
Lifan Zhou (zhoulifan_rs@163.com) received his B.S. degree
in geographic information systems from Wuhan University,
Wuhan, China, and his Ph.D. degree in cartography and
geographic information systems from Zhejiang University,
Hangzhou, China, in 2006 and 2014, respectively. Since
2014, he has been a lecturer with the School of Computer
Science and Engineering, the Changshu Institute of Technology,
Suzhou, 215500, China, where he is the principal
of a scientific research project on learning single-multibaseline
phase unwrapping methods for the synthetic
aperture radar interferogram, which was supported by the
National Natural Science Foundation of China. In 2016, he
was a visiting scientist with the Chinese University of Hong
Kong, Hong Kong, China. His main research interests are
in the fields of phase unwrapping, algorithm design, and
machine learning.
25
THE PERFORMANCE OF
AI-BASED MB PU METHODS
IS DIRECTLY RELATED TO
THE CHOICE OF THE
NORMAL BASELINE LENGTH
IN THE INSAR SYSTEM,
WHICH IS SIMILAR TO THE
TRADITIONAL MODELBASED
MB PU METHODS.
IEEE Geoscience and Remote Sensing Magazine - June 2021

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