IEEE Geoscience and Remote Sensing Magazine - June 2023 - 79

CREATE MORE STABLE AND RELIABLE
EXPLANATIONS FOR THE NETWORK
It has been demonstrated in [157] that existing interpretations
of the network are vulnerable to small perturbations.
The fragility of interpretations sends a message that designing
robust XAI methods will have promising applications
for adversarial attacks and defenses for EO.
PROVIDE HUMAN-UNDERSTANDABLE
EXPLANATIONS IN EO TASKS
Previous studies show that there is still a large gap between
the explanation map learned by XAI methods and human
annotations, thus, XAI methods produce semantically
misaligned explanations and are difficult to understand
directly. This problem sheds light on the importance of
deriving interpretations based on the specific EO task and
human understanding, increasing the accuracy of explanation
maps by introducing more constraints and optimization
problems to explanations.
CONCLUSIONS AND REMARKS
Although AI algorithms represented by deep learning theories
have achieved great success in many challenging tasks
in the geoscience and RS field, their related safety and security
issues should not be neglected, especially when addressing
safety-critical EO missions. This article provided
the first systematic and comprehensive review of recent
progress on AI security in the geoscience and RS field, covering
five major aspects: adversarial attack, backdoor attack,
FL, uncertainty and explainability. Although research
on some of these topics is still in its infancy, we believe that
all these topics are indispensable for building a secure and
trustworthy EO system, and all five deserve further investigation.
In particular, in this section, we summarize four
potential research directions and provide some open questions
and challenges. This review is intended to inspire
readers to conduct more influential and insightful research
into related realms.
SECURE AI MODELS IN EO
Currently, the security of AI models has become a concern
in geoscience and RS. The literature reviewed in this article
also demonstrates that either adversarial or backdoor
attacks can seriously threaten deployed AI systems for EO
tasks. Nevertheless, despite the great effort that has been
made in existing research, most of the studies focus only
on a single attack type. How to develop advanced algorithms
to defend the AI model against both adversarial
and backdoor attacks simultaneously for EO is still an
open question. In addition, although most of the relevant
research focuses on conducting adversarial (backdoor) attacks
and defenses in the digital domain, how effective
adversarial (backdoor) attacks and defenses in the physical
domain might be carried out, considering the imaging
characteristics of different RS sensors, is another meaningful
research direction.
JUNE 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
DATA PRIVACY IN EO
State-of-the-art AI algorithms, especially deep learningbased
ones, are usually data driven, and training these giant
models often depends on a large quantity of high-quality labeled
data. Thus, data sharing and distributed learning have
played an increasingly important role in training large-scale
AI models for EO. However, considering the sensitive information
commonly found in RS data, such as military targets
and other confidential information related to national defense
security, the design of advanced FL algorithms to realize
the sharing and flow of necessary information required
for training AI models while protecting data privacy in EO is
a challenging problem. Additionally, most of the existing research
focuses on horizontal FL, in which it is assumed that
distributed databases share high similarity in feature space.
Improving FL ability in cross-domain, cross-sensor, or crosstask
scenarios for EO is still an open question.
TRUSTWORTHY AI MODELS IN EO
The uncertainty in RS data and models is a major obstacle to
building a trustworthy AI system for EO. Such uncertainty exists
in the entire lifecycle of EO, from data acquisition, transmission,
processing, and interpretation, to evaluation, and
constantly spreads and accumulates, affecting the accuracy
and reliability of the eventual output of the deployed AI model.
Currently, most of the existing research adopts the deterministic
and Bayesian inference methods to quantify the uncertainty
in data and models, which ignores the close relationship between
data and models. Thus, finding a method to achieve uncertainty
quantification for data and models simultaneously
in EO deserves more in-depth study. Furthermore, apart from
uncertainty quantification, it is equally crucial to develop advanced
algorithms to further decrease uncertainty in the entire
lifecycle of EO so that errors and risks can be highly controllable,
achieving a truly trustworthy AI system for EO.
XAI MODELS IN EO
As an end-to-end, data-driven AI technique, deep learning
models usually work like an unexplainable black box. This
makes it straightforward to apply deep learning models
in many challenging EO missions, like using a point-andshoot
camera. Nevertheless, it also brings about potential
security risks, including vulnerability to adversarial (backdoor)
attacks and model uncertainty. Thus, achieving a
balance between tractability, explainability, and accuracy
when designing AI models for EO is worthy of further investigation.
Finally, considering the important role of expert
knowledge in interpreting RS data, finding a way to
better embed the human-computer interaction mechanism
into the EO system may be a potential research direction
for building XAI models in the future.
ACKNOWLEDGMENT
The authors would like to thank the Institute of Advanced
Research in Artificial Intelligence for its support. The corresponding
author of this article is Shizhen Chang.
79
IEEE Geoscience and Remote Sensing Magazine - June 2023

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