IEEE Geoscience and Remote Sensing Magazine - June 2021 - 55

and alunite and kaolin. For those optical broadband
imaging products (e.g., MS imagery), they can identify
only certain materials with observable differences in the
spectral signatures, e.g., water, trees, and soil.
2) The higher spectral resolution creates possibilities for some
challenging applications, e.g., parameter extraction in biophysics
and biochemistry, biodiversity conservation, monitoring
and management of the ecosystem, and the automatic
detection of food safety, that can hardly be achieved
by depending only on former imaging techniques, thus
providing new insights into RS and geoscience fields.
3) Due to limitations on image resolution, in either the
spectral or spatial domains, physical and chemical atmospheric
effects, and environmental conditions (e.g.,
the interference of soil background, illumination, the
uncontrolled shadow caused by clouds or building occlusion,
topography change, and complex noises), those
traditional RS imaging techniques were, to a great extent,
dominated by qualitative analysis. As HS RS arises, quantitative
or semiquantitative analysis becomes increasingly
plausible in many practical cases.
AN EVER-GROWING RELATION BETWEEN
NONCONVEX MODELING AND INTERPRETABLE
AI IN HS RS
In recent years, a vast number of HS RS missions [e.g.,
MODIS, the Hyperspectral Satellite for Earth Observation
(HypSEO), the German Aerospace Center Earth Sensing
Imaging Spectrometer (DESIS), Gaofen-5, the Environmental
Mapping and Analysis Program (EnMap), the Hyperspectral
Infrared Imager (HyspIRI), and so on] have been
launched to enhance our understanding of Earth and its
environment, contributing to a rapid and better development
of a wide range of relevant applications, such as land
cover land-use classification, SU, data fusion, image restoration,
and multimodal data analysis. With the ever-growing
availability of RS data sources from both satellite and
airborne sensors on a large and even global scale, expert
system-centric data processing and the analysis mode have
run into bottlenecks and cannot meet the demands of the
big data era. For this reason, data-driven signal and image
processing, ML, and AI models have garnered growing interest
and attention from researchers in the RS community.
Supported by well-established theory and numerical optimization,
convex models have been effective for a variety
of HS tasks under highly idealized assumptions. However,
there exist unknown, uncertain, and unpredictable factors
in complex real scenes. Due to these factors, which lead to
a lack of sound understanding and modeling capability,
convex models fail to work properly. The specific reasons
could be twofold. On the one hand, integrating the benefits
of 1D and 2D signals, 3D, structured HS images offer
greater potential and better solutions (compared to natural
images) for dealing with varying situations but simultaneously
increase the model's complexity and uncertainty to
some extent. On the other hand, due to unprecedented
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
spatial, spectral, and temporal resolutions of HS images
in remotely sensed HS imaging, the difficulties and challenges
in the sophisticated HS vision approaches are mainly
associated with the volume of HS data, complex material
(spectral) mixing behavior, and uncontrolled degradation
mechanisms in data acquisition caused by illumination,
noise, and atmospheric effects.
The aforementioned factors, to a great extent, limit convex
models to being intelligent approaches for fully understanding
and interpreting real-life
scenarios. Therefore, this naturally
motivates us to investigate
the possibility of processing
and analyzing HS data in a
nonconvex modeling fashion.
In the following, we make a
brief qualitative comparison
between convex and nonconvex
models to clarify that nonconvex
modeling might be an
optimally feasible solution toward
interpretable AI models
in HS RS.
◗ Convex models are theoretically guaranteed to converge
to the global optimal solution, yet most tasks related to
HS RS are, in reality, complex and hardly simplified to an
equivalent and perfect convex formulation. This, to some
extent, makes convex models inapplicable to practical tasks
due to the lack of interpretability and completeness for
problem modeling.
SUPPORTED BY WELLESTABLISHED
THEORY AND
NUMERICAL OPTIMIZATION,
CONVEX MODELS HAVE
BEEN EFFECTIVE FOR A
VARIETY OF HS TASKS
UNDER HIGHLY IDEALIZED
ASSUMPTIONS.
◗ Rather, nonconvex models can characterize the complex
studied scene in HS RS more finely and completely, thereby
tending to achieve more automatization and " intelligentization "
in the real world. Moreover, by excavating the
intrinsic properties from HS data to effectively yield physically
meaningful priors, the solution space of nonconvex
models can be shrunk to a good region bit by bit.
◗ Although nonconvex models are complex and consider
more complicated prior knowledge, possibly leading to
the lack of stable generalization ability, they have higher
potential than convex models, particularly in explaining
models, understanding scenes, and achieving intelligent
HS image processing and analysis. Furthermore,
this might be able to provide researchers with a broader
range of HS vision-related topics, making it applicable
for more real cases in a variety of HS RS-related tasks.
CONTRIBUTIONS
With the advent of the big data era, ever-increasing data bulk
and diversity bring rare opportunities and challenges for the
development of HS RS in Earth observation. Data-driven AI
approaches, e.g., ML and deep learning (DL) based, have occupied
a prominent place in manifold HS RS applications.
Nevertheless, how to open the models and give them interpretability
remains yet unknown. In this article, we raise a
bold and understandable viewpoint, i.e., that nonconvex
55
IEEE Geoscience and Remote Sensing Magazine - June 2021

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