IEEE Geoscience and Remote Sensing Magazine - June 2021 - 15

clustering characteristic of the residue; then, each subinterferogram
obtained by the tiling strategy will be processed
independently of one another [6], [11], [19]. In other words,
it is a clustering-analysis-based preprocessing step before
the branch-cut placement.
For example, the authors in [5] designed an a-approximation
TS-L1 tiling strategy, making use of the clustering
characteristic of the residue distribution, which can
save the computing source. Figure 7(a) shows the residue
distribution of Figure 2(a). Figure 7(b) presents the residue-clustering
result (achieved by an a-approximation
TS-L1) of Figure 7(a), where the neighboring clusters are
distinguished from each other using their boundaries.
Because the clustering analysis prejudges and groups together
the residues that need to be connected by branch
cuts, the preprocessing step can significantly save computing
resources (e.g., the peak memory consumption
and execution time).
In [5], an example that examines the computational
performance of the a-approximation TS-L1 tiling strategy
compared with the global PU processing was presented,
illustrating how much the computing source can be saved
when the size of the interferogram increases. Figure 8(a)
and (b) shows the curves of execution time and peak
memory consumption of the PU processing using the tiling
strategy and global PU processing, respectively. The
behavior of these curves in Figure 8 shows that, when the
AI-based tiling strategy is adopted, the computational
and peak memory consumption of PU processing is much
lower than that of the global PU processing, especially in
a large interferogram.
The methods in the second group resolve the branchcut
deployment problem by using a traditional AI technology
at early stages, e.g., genetic algorithms [20], [21],
simulated annealing [22]-[24], ant colonies [25], and
so on. Some of these methods are robust, but they have
heavy computational burdens. A deep learning-based
branch-cut deployment method recently presented [26]
transforms the residue-balancing problem as a semanticsegmentation
problem [that is, as a branch-cut Network
-2
2
(a)
-1
1
(b)
(c)
FIGURE 6. (a) The simulated wrapped phase image with a 0.5
coherence coefficient. (b) The residues obtained by the Itoh condition.
(c) The residues obtained by PGNet [16].
0.5
1
-0.5
-1
(a)
(b)
FIGURE 7. (a) The residue distribution of Figure 2(a). (b) The clustering
result of (a) using an a-approximation, TS-L1 [5].
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
FIGURE 8. (a) The curves of the global and tiling execution times. (b)
The curves of the global and tiling peak memory consumptions [5].
15
-1
1
200
400
600
800
02 46 810
Number of Pixels
(a)
2,000
4,000
6,000
8,000
02 46 810
Number of Pixels
Global Peak Memory Consumption
Tiling Peak Memory Consumption
(b)
×106
×106
Global Execution Time
Tiling Execution Time
Peak Memory
Consumption (MB)
Execution Time (min)
IEEE Geoscience and Remote Sensing Magazine - June 2021

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