IEEE Geoscience and Remote Sensing Magazine - March 2023 - 61

early HS RS data analysis of Bioucas-Dias [4], HS subspace
identification of Bioucas-Dias [233], and recent
interpretable HS artificial intelligence of Hong [3].
2) Parameter adjustment: In the algorithmic solution, parameter
adjustment is an indispensable portion to
achieve significant HS data processing performance. Parameters
can be gradually tuned via extensive simulated
experiments, while sometimes, they should be reset for
various datasets, due to the uncertainty of data size. In
practice, users are most likely to be nonprofessional,
with little knowledge of a special algorithm, leading to
improper parameter setting and unsatisfactory processing
results. Therefore, in the future, efforts should be
made to design a fast proper-parameter search scheme
and reduce the number of parameters to increase algorithmic
practicability. Based on this overview, beginners
are encouraged to make good use of open source codes
listed in Table 1 and HS RS data given in Table 2. After
long-term learning and actual operations, beginners
will be proficient in properly adjusting parameters and
increasing algorithmic feasibility.
3) Computational consumption: Tensor decomposition-based
methods have achieved satisfactory results in HS data
processing, yet they sometimes cause high computational
consumption. For instance, for an NLR tensor
denoising model, TDL spends more than 10 min for a
dataset of
200 200 80## As the image size increases,
.
the increasing number of nonlocal FBPs will cause greater
time consumption. Thus, there exists vast room for
promotion and innovation in improving the optimization
efficiency of HS data processing. Parallel and distributed
implementations of big tensor decompositions
are beneficial to accelerate the optimization process. A
classic tutorial [244] is presented for practitioners to understand
the parallel and distributed computation of big
tensors from the perspective of mathematical analytics.
Several state-of-the-art techniques [245], [246], [247],
[248] related to this architecture have emerged in recent
years. For example, Rolingera et al. [245] extended
three parallel tools to implement CP decomposition using
ALS fitting: SPLATT, DFacTo, and ENSIGN. Tensor
products can be reduced to multiple sparse matrix operations,
and how such operations are used to execute
CP-ALS has a significant practical influence on total
memory usage and parallelizability.
4) Methodological feasibility: Unlike deep learning-based
methods, designing handcrafted priors is the key to
tensor decomposition-based methods. Existing methods
exploit the structure information of the underlying
target image by implementing various handcrafted priors,
such as LR, TV, and nonlocal similarity. However,
different prior assumptions apply to specific scenarios,
making it challenging to choose suitable priors according
to the characteristics of HS images to be processed.
Deep learning-based methods automatically learn prior
information implicitly from datasets themselves, withMARCH
2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
out the trouble of manually designing a manual regularizer.
As an advisable approach, deep learning can be
incorporated into tensor-based methods to mine essential
multifeatures and enhance methodological feasibility.
For instance, recently, transform-based TNN minimization
approaches have been developed to capture
underlying LR tensor structures in HS RS applications.
A nonlinear transform is learned by a nonlinear multilayer
neural network under self-supervision [249], [250].
5) Multimission applications: The extremely broad field of
HS imagery makes it impossible to provide an exhaustive
survey of all the promising HS RS applications. It is
certainly of significant interest to develop tensor decomposition-based
models for other noteworthy processing
and analysis chains in future work, including classification,
change detection, large-scale land cover mapping,
and image quality assessment. Some HS tasks serve as
a preprocessing step for high-level vision. For example,
the accuracy of HS classification can be improved after
an HS denoising step. How to apply tensor decomposition
for high-level vision and even multimission frameworks
may be a key challenge. To try tensor decomposition-based
models for multimission applications, we
should confirm the significance of specific multimission
research, which needs to master the whole HS RS
imaging process and consequent data analysis. Simultaneously,
we recommend several existing multimission
techniques, such as coupled denoising and unmixing
[251], denoising and classification [252], unmixing using
AD [253], and so on.
ACKNOWLEDGMENT
This work was supported by the National Natural Science
Foundation of China, under grants 42271350, 62201552,
and 62201553, and China Postdoctoral Science Foundation,
under grant 2022M713223. This work was also supported
by the Multidisciplinary Institute in Artificial Intelligence at
Grenoble Alpes (project ANR-19-P3IA-0003) and AXA Research
Fund. The corresponding author is Danfeng Hong.
AUTHOR INFORMATION
Minghua Wang (wangmh@aircas.ac.cn) received her B.S.
degree in 2016 from the School of Automation, Harbin Institute
of Technology, Harbin, China, where she received
her Ph.D. degree in control science and engineering in
2021. She was also a visiting Ph.D. student at the University
of Grenoble Alpes, Grenoble, France, in 2019-2020.
She is currently a postdoc with the Aerospace Information
Research Institute, Chinese Academy of Sciences, Beijing
100094 China. Her research interests include hyperspectral
image denoising, anomaly detection, tensor learning, and
low-rank representation. She is a Member of IEEE.
Danfeng Hong (hongdf@aircas.ac.cn) received his M.Sc.
degree (summa cum laude) in computer vision from the
College of Information Engineering, Qingdao University,
Qingdao, China, in 2015 and his Dr.-Ing degree (summa
61
IEEE Geoscience and Remote Sensing Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2023
