IEEE Geoscience and Remote Sensing Magazine - June 2021 - 20

TABLE 1. THE CHARACTERISTICS OF THE REPRESENTATIVE PATH-FOLLOWING OF AI AND THE GLOBAL PU OF AI METHODS.
AI-BASED SB
PU CATEGORY
Path-following
PU of AI
-
-
AI-BASED SB PU
SUBCATEGORY
Residue estimation
of AI
-
Branch cut of AI
PU METHOD
PGNet [16]
CNN [18]
BCNet [26]
Global PU of AI One-step PU of AI DNN [33]
DLPU [36]
PUNet [37]
-
-
-
Two-step PU of AI
-
-
Segmentation
and denoised
networks [42]
DeepLabV3+based
PU [44]
-
-
DNN: deep NN.
The L1 loss is defined between the ground truth and the
unwrapped phase using the L1 norm and is given by
lossL =- t
1
/ |( )( )|.ss
s
}}
Therefore, the overall loss can be a sum of
loss =+ + lossL1
OverallCross
loss
loss
Residue
(9)
.
(10)
Overall, compared with path-following, AI-based PU,
which needs some existing processing steps in the classic
model-based PU method, the global AI-based PU has
direct, end-to-end characteristics, but it still has room for
improvement. Specifically, traditional PU methods usually
require quality maps (such as a coherence-coefficient
map or a gray-level cooccurrence matrix [4]) to guide PU.
However, the methods in this group cannot cooperate with
these assistant PU components that help in achieving better
unwrapping results. On the other hand, because the network
is trained for an interferogram with a fixed, small size
(e.g., smaller than 256 × 256), once the size of the interferogram
becomes extremely large, the PU result obtained by
the network will become unreliable. To solve the second issue,
Zhang et al. [44] suggested that the large interferogram
be partitioned into a number of subinterferograms of small
sizes. Then, each subinterferogram could be predicted separately
by the network, and, finally, an image-stitching strategy
could be used for merging into a large interferogram.
Table 1 summarizes the characteristics of the representative
path following of AI and the global PU of AI methods. Besides
these two groups of AI-based SB PU methods, a fuzzy logic
algorithm has been applied in SB PU for clustering purposes.
Considering that there is no PU method that can unwrap all
types of interferograms, fuzzy logic clustering algorithms can
20
choose which PU method (e.g., path-following or minimumnorm
method) is the most suitable for a certain cluster in the
input interferogram [48]. Moreover, some other AI-based algorithms
for SB PU that employ extended Kalman [49] and
particle filters [50] have been proposed as well.
AI IN MB PU
Different from SB PU, MB PU is well posed rather than ill
posed by taking advantage of the baseline diversity used in
MB InSAR systems so that it can completely break through
the constraint of the Itoh condition. However, most of the
MB PU methods suffer from poor noise robustness for two
main reasons. First, the mathematical foundation of the
MB PU methods, i.e., the CRT, is too noise-sensitive to use
directly. Second, the fundamental processing principle of
MB PU is pixel by pixel on the repeated observations, so
the information used for the MB PU of each pixel is independent
of the surrounding ones. To tackle this problem,
numerous AI technique-based MB PU methods have been
put forward to improve the noise robustness of MB PU due
to its powerful noise-immunity ability. In the following
sections, we first review some of the early machine learning
techniques that have been used in the MB domain, and
then we discuss recently developed deep learning-based
MB PU methods through an in-depth investigation.
MB PU METHODS USING EARLY MACHINE
LEARNING TECHNIQUES
The MB PU methods using early machine learning techniques
can be roughly divided into three groups: maximum
likelihood (ML), maximum a posteriori (MAP), and cluster
analysis (CA) methods. In the following sections, we offer
a detailed introduction of these three kinds of algorithms.
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021
-
-
LEARNING
OBJECT
Phase gradient
Phase gradient
Branch cut
QUALITY
MAP
Yes
Yes
Yes
-
-
Unwrapped phase No
Unwrapped phase No
Unwrapped phase No
Wrap count
No
Wrap count
PhaseNet [40] Wrap count
PhaseNet 2.0
[46]
Wrap count
No
No
No
FRINGE
CONGRUENCY
Yes
Yes
Yes
No
No
No
Yes
Yes
Yes
Yes
POSTPROCESSING
MCF
Wrap-count
reconstruction
Branch-cut
method
No
No
No
Phase-jump
identification
and
correction
Phase-jump
identification
and correction
Clustering-based
smoothness
No
INTERFEROGRAM
SIZE
Large
Large
Large
Small
Small
Small
Small
Small
Small
Small

IEEE Geoscience and Remote Sensing Magazine - June 2021

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