Progress in Aerospace Sciences 144 (2024) 100960

Contents lists available at ScienceDirect

Progress in Aerospace Sciences

journal homepage: www.elsevier.com/locate/paerosci

Artificial Intelligence for Trusted Autonomous Satellite Operations

Kathiravan Thangavel a, b, c, f, Roberto Sabatini a, b, c, f, *, Alessandro Gardi a, b, f,
Kavindu Ranasinghe b, g, Samuel Hilton b, f, Pablo Servidia d, Dario Spiller e, b
a
Department of Aerospace Engineering, Khalifa University of Science and Technology, PO Box 127788, Abu Dhabi, United Arab Emirates
b
School of Engineering, STEM College, RMIT University, Melbourne, VIC, 3000, Australia
c
Sir Lawrence Wackett Defence & Aerospace Centre, RMIT University, Melbourne, VIC, 3083, Australia
d
CONAE, Comisión Nacional de Actividades Espaciales, Av. Paseo Colón 751, C1063ACH, Buenos Aires, Argentina
e
School of Aerospace Engineering, Sapienza University of Rome, 00138, Rome, Italy
f
SmartSat Cooperative Research Centre, North Terrace, Adelaide, SA 5000, Australia
g
Insitec Pty Ltd, Melbourne, VIC 3000, Australia

A R T I C L E I N F O A B S T R A C T

Keywords: Recent advances in Artificial Intelligence (AI) and Cyber-Physical Systems (CPS) for aerospace applications have
Avionics brought about new opportunities for the fast-growing satellite industry. The progressive introduction of con­
Space systems nected satellite systems and associated mission concepts is stimulating the development of intelligent CPS (iCPS)
Satellite systems
architectures, which can support high levels of flexibility and resilience in an increasingly congested near-Earth
Machine learning
Artificial intelligence
space environment. The need for higher levels of automation and autonomy in satellite operations has stimulated
Cyber-physical systems numerous research initiatives in recent years, focusing on the progressive enhancement of systemic performance
Distributed space system (e.g., addressing safety, integrity and cyber-physical security metrics) and associated monitoring/augmentation
Distributed satellite systems approaches that can support Trusted Autonomous Satellite Operations (TASO). Despite these advances, in most
Human-machine interactions contemporary satellite platforms, autonomy is restricted to a specific set of rules and cases, while the transition to
Trusted autonomous systems TASO requires a paradigm shift in the design of both space vehicles and ground-based systems. In particular, the
use of AI is seen as an essential enabler for TASO as it enhances system performance/adaptability and supports
both predictive and reactive integrity augmentation capabilities, especially in Distributed Satellite Systems
(DSS). This article provides a critical review of AI for satellite operations, with a special focus on current and
likely future DSS architectures for communication, navigation and remote sensing missions. The aim is to
identify key contemporary challenges and opportunities associated with space iCPS design methodologies to
enhance the performance and resilience of satellite systems, supporting the progressive transition to TASO. A
comprehensive review of relevant AI techniques is presented to critically assess the potential benefits and
challenges of each method for different space applications. After describing the specificities of DSS and the
opportunities offered by iCPS architectures, the co-evolution of space and control (ground and on-board) seg­
ments is highlighted as an essential next step towards enabling TASO. As an integral part of this evolutionary
approach, the most important legal and regulatory challenges associated with the adoption of AI in TASO are also
discussed.

1. Introduction operations, remain to be explored due to the increasing complexity of

hardware/software components and associated safety, integrity and
Satellite systems provide a wide variety of services, which can be cyber-physical security concerns [2]. Present day autonomous systems
easily accessed from almost any location on the globe. These systems are capable of executing intelligent functions (e.g., decisions and/or
have rapidly evolved over the last few decades and have become actions traditionally performed by humans) using various
essential in various application domains, such as communications, computer-based algorithms, often referred to Artificial Intelligence (AI).
navigation, Earth Observation (EO) and astronomy [1]. However, This requires the ability to gather real-time data from the external
certain aspects of satellite technology, such as trusted autonomous operational environment (i.e., sensing), to perform inference and/or

* Corresponding author. Department of Aerospace Engineering, Khalifa University of Science and Technology, PO Box 127788, Abu Dhabi, United Arab Emirates.
E-mail address: [email protected] (R. Sabatini).

https://doi.org/10.1016/j.paerosci.2023.100960
Received 21 August 2023; Accepted 12 November 2023
Available online 27 December 2023
0376-0421/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

decision-making functions and to execute proper actions if and when parameters and configuration depend on its intended purpose. The
required. Despite the significant progress made in hardware and soft­ classification of spaceflight systems adopted in this article is presented
ware technologies, Trusted Autonomous Satellite Operations (TASO) are in Fig. 1. Broadly, spaceflight systems can be grouped into three cate­
still largely a research topic and significant investments are needed to gories: (1) Space exploration systems [5]; (2) Earth orbital/sub-orbital
fully exploit the anticipated safety, efficiency and sustainability benefits transport system [10]; (3) Earth orbit satellite systems. Earth-orbiting
that such operations would bring, possibly leading to the progressive satellite systems can be further divided into the following categories:
removal of present-day socio-political barriers such as AI ethics, liability (i) Monolithic satellite systems, (ii) Distributed Satellite Systems (DSS)
and public trust [3]. In many applications, fully autonomous satellite which are discussed broadly in the following sections.
operations are either impractical or undesirable, mainly because a minor
error can result in the loss of millions of dollars and, in some cases, lead 2.1. Monolithic Satellite Systems
to human casualties (point-to-point suborbital space transport,
Earth-orbiting inhabited space stations, etc.). Therefore, an acceptable If a satellite system with its modules or subsystems is physically in­
level of trust is required for near-Earth operations, especially consid­ dependent from other space assets, it is classified as a monolithic sat­
ering the steady increase of Resident Space Objects (RSO) in Low-Earth ellite system. Monolithic systems are still a large fraction of spacecraft
Orbits (LEO) and Geostationary Orbits (GEO) [4,5]. Furthermore, to being deployed in missions such as deep space exploration, technology
facilitate further progress in TASO research, it is essential to address the demonstration, universities and research centres [11,12]. The need for
implications of trusted autonomy and AI in the evolution of self-contained hardware and the required redundancy increases the
Cyber-Physical Systems (CPS) for space applications, including the system’s overall weight and volume, making it more expensive. A
co-evolution of system-level requirements (i.e., communication, control typical monolithic satellite system has the following modules PR: Pro­
and computing) and human-autonomy interactions. Current research cessor, PL: Payload, DL: Downlink, CM: Communications Module, BUS,
trends in this area show that Cyber-Physical-Human (CPH) architectures which integrates all the modules as depicted in Fig. 2. Monolithic sat­
are evolving with the widespread adoption of Machine Learning (ML) ellite systems are comprehensively reviewed in Refs. [13,14].
and hybrid AI techniques (e.g., neuro-fuzzy inference engines) and
becoming progressively more capable of modulating both the levels of 2.2. Distributed Satellite Systems
autonomation and the human command/control functions towards
achieving specific goals. In this context, we are participating to an DSS consist of multiple spacecraft working together to achieve one or
evolutionary process where human operators are progressively tran­ more common objectives. DSS therefore adopt a satellite architecture in
sitioning to a high-level supervisory role [6]. which the functional capabilities are shared among many space assets
Over the last few years, Distributed Satellite Systems (DSS) have that communicate via wireless networks [15]. The DSS concept is
been at the forefront of this transformation and it is now clear that the gradually migrating the physical connectivity of various components in
use of AI in DSS will play a significant role in easing the transition to conventional satellite systems into wireless connections using either
TASO. To meet the requirements of future trusted autonomous space radiofrequency or optical communication methods, i.e., Inter Satellite
vehicles and intelligent operation in highly integrated and information- Links (ISL). Therefore, DSS mission architectures shift away from
rich environments, a radical departure from conventional system design monolithic systems towards multiple spacecraft/modules of elements
and development approaches is required. Going forward, explainability that communicate, interact and cooperate with one another [16]. In
and certification of AI based systems will be critical, particularly in outer addition to these new system-wide properties, subdividing the modules
space operations where there is a need to simultaneously address safety, over many launches reduces risk, ensuring that the core system is not
security and legal requirements (e.g., liability for the damages these lost when a launch fails. This approach also offers the flexibility to
systems may cause). As a result, there is a need to understand the progressively deploy the system in orbit, thereby allowing the addition
associated technical, ethical and legal challenges that come with these of different DSS elements at successive stages. A study by the Research
evolving systems. and Development (RAND) corporation showed that distributed systems
This article starts with a high-level classification of spaceflight sys­ have the potential to [12]:
tems (section 1), with a focus on the various categories of Earth orbiting
satellites. Then, it moves on to discuss and compare monolithic and 1. Weigh less and be less expensive to launch.
distributed satellite configurations (section 2). After that, section 3 fo­ 2. Perform better during deployment and before the full DSS is
cuses on spacecraft autonomous operations, providing an overview of completely operational.
the most promising concepts applied to human and machine collabo­ 3. Show better tolerance to single and multiple failures, with no or
ration in space operations. Section 4 provides a critical examination of graceful degradation of performance.
the many AI techniques proposed for space system applications and 4. Be more survivable in the event of both cyber and physical attacks.
section 5 discusses how these techniques can be applied in practice to
various DSS architectures. The required evolution of spaceflight systems Due to these distinctive advantages, DSS can deliver a more
infrastructure is discussed in section 5, which highlights the most responsive and resilient solution to meet the expanding demands of the
promising applications of AI in the different system segments and the scientific community and also the defence sector, for instance by
need for a co-evolution of DSS space and ground segments. Section 6 improving the quality/quantity of measurements and associated data
discusses current and likely future applications of AI in near-Earth space analytics in EO [17], Space-Based Space Surveillance (SBSS) tasks
missions and section 7 addresses safety and security aspects associated [18–21] and Interplanetary mission [22]. Some notable examples
to the adoption of AI in DSS. Section 8 concludes the critical review include PRISMA [23], GRACE [24], TerraSAR-X and TanDEM-X mis­
highlighting the open technological, ethical and legal challenges, as well sions [25].
as the ongoing efforts to overcome these challenges and to facilitate the
uptake of AI technology in next-generation satellite systems. 2.2.1. Modularity
A distinctive attribute of DSS is modularity, Within the broader scope
2. Spaceflight systems of systems engineering, modularity is a feature of systems that quantifies
the degree to which a system’s functionalities can be subdivided into
Thousands of active satellites are currently orbiting Earth and, in distinct modules or clusters which interact with each other [26,27].
recent years [7], there has been an exponential growth of RSO [8], Damage to one module can cascade to subsequent modules in a highly
especially in the LEO environment [9]. Each satellite’s size, orbital interconnected system with minimal modularity, enhancing the risk of a

2
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 1. Classification of spaceflight systems.

used in conjunction with M+ operators, the spectrum can help designers

choose appropriate parameters and put together a system-specific
computational tool using a number of pre-existing tools and ap­
proaches [14,26,30–32].
The transition from M0 to M1 is referred to as Decomposition, from M1
to M2 as Splitting, from M2 to M3 as Fractionation and from M3 to M4 as
Resource Sharing. In satellite architecture, M0 , M1 andM2 are considered
Fig. 2. Monolithic satellite system. Adapted from Ref. [14]. monolithic systems, whereas M3 and M4 represent distributed archi­
tecture systems. While M0 , M1 andM2 cover all instances of modular­
isation for monolithic systems (systems with only one physical unit), M3
system-wide failure [28]. A disturbance to one component, on the other
and M4 cover systems with multiple units (distributed systems) and the
hand, may be best controlled in a system with a high level of modularity.
possibility of communication between them [33]. M+ operators effec­
Modularity is often explored as a spectrum of several levels and forms of
tively transform the conceptual framework into a computational
a system that exist as a continuum within the system and not a binary
decision-making tool, which allows to implement the best level of
property [29,30]. Further, continuous modularity can be intuitively and
modularity for a certain system’s functionality in a given mission envi­
methodically represented and quantified for some satellite systems that
ronment. The decision-making paradigm entails both the modularity
are now being introduced for a subset of elements in a network system
phase and the design implementation inside that phase. By including a
[27,31].
set of operators (M+ operators) for calculating the transition value from
In accordance with the broader concept, we can introduce a frame­
work consisting of five DSS modularity levels identified from M0 to M4 in one stage of modularity (Mx ) to its next immediate phase (Mx+1 ), we
focus on the former. By computing the probability distribution of the
Fig. 3. This discretization allows for better computational tractability
and comprises fully integral architectures (M0 ), integral yet decompos­ difference in value between two consecutive phases, the suggested
decision-making operators evaluate the performance of the system
able architectures (M1 ), modular yet monolithic architectures (M2 ),
static distributed architectures (M3 ) and dynamic distributed architec­ before and after operation [33]. This will allow decisions to be taken
based on an average value difference as well as the level of risk toler­
tures (M4 ). A set of value operators quantify the net operator (M+ ),
ance. For most engineering systems, M1 is the lowest modularity, so the
which shifts between two neighbouring levels in this spectrum. When
splitting operation suggests the changeover from M1 to M2 through the

3
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 3. A five-stage modularity with distributed architecture spectrum and four M+ operations. Adapted from Ref. [32].

development and use of proper standard interfaces. Fractionation analysis), while taking the subsequent criteria into account [32,33]:
operation by shifting one or more of its subsystems to other fractions
takes a system from M2 to M3 . Although M+ evaluation specifics a a) Technical Parameters: For instance, the probability density of a
procedural algorithm which is dependent on particular systems and its failure, the time required for an upgrade to become available, the
parameters, which acts as a decision-making evaluation engine [32]. highest number of modules allowable and the maximum trans­
The M+ value is measured by comparing the system’s value prior and mission bandwidth permitted.
post its operations. Such assessment involves knowledge of the system b) Economical Parameters: For instance, the user demand in terms of
and its settings [14,26,27,30,32,33]. Fig. 4 shows the input and output number of modules at a given time, the cost of launching and oper­
characteristics for evaluating M+ operators. At each level of modularity, ating a module and the rate at which distinct module types generate
the system’s value is determined using one of the common system value.
assessment methods (e.g., discounted cash flow analysis, scenario

Fig. 4. Quantification of the M+ operation value. Adapted from Ref. [32].

4
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

c) Life Cycle Parameters: Total time required for operation, budget, as

well as maximum time required for initial deployment.

A high-level process for calculating the decentralisation values from

M3 to M4 is depicted in Fig. 5. Because of the underlying network
structure, designers are recommended to rely on multi-agent techniques
that blend system dynamics and evolution with autonomous behaviour
[34].

2.3. DSS classification

DSS are categorized based on the type of mission and function they
perform. Activities required to meet local objectives (i.e., those specific
to each module) or small bits of a global objective’s functioning (i.e.,
particular to the infrastructure) may be included in modules performing
activities in a distributed infrastructure, whether it be in independent
satellite systems or distributed spacecraft. As a result, the function type
is measured in terms of how dispersed the mission’s goals are, ranging
from no collaboration between modules (i.e., local functionality) to a
fully functional symbiosis (i.e., distributed functionality). As a result,
different DSS missions are characterised according to their degree of
distribution in terms of the system’s capabilities or goals and resource
interdependence between modules. A bi-dimensional classification can
be introduced by considering these two domains, as shown in Fig. 6,
with values in the range [0,1]. The x-axis shows the degree of mission Fig. 6. DSS classification. Adapted from Ref. [36].
goal distribution, which ranges from missions in which satellite modules
work together to advance a single global function to goals in which each 2.3.1. Constellation
satellite module develops its own local activity. The y-axis shows the A satellite constellation is a collection of homogeneous or hetero­
degree of fractionation among scenarios where modules are totally geneous spacecraft that operate as a unified system, with the purpose of
reliant on one another and cases where nodes are completely resource providing continuous global or near-global coverage as shown in Fig. 7
self-sufficient [35,36]. Both classification axes are independent. The (a). Satellites are usually positioned among a set of orbital planes that
following are more specific classifications of DSS which will be discussed are complementary to one another and are connected directly or via
in the following subsections: other satellites to one or multiple ground stations across the world. The
literature provides wide coverage of satellite constellations [22,37–40].
• Constellations
• Fractionation 2.3.2. Fractionation
• Federated In fractionated systems, a spacecraft is divided into smaller units
• Modular collaborating to achieve a common mission objective. The satellite
• Swarms consists of co-dependent modules that require system resources to be
• Formation exchanged in order to function as shown in Fig. 7 (b) [41]. While all
• Constellation of formations fractionated systems need a common infrastructure consisting of data
• Hybrid Missions processing, power, communication link, etc., to complete the functions
calling for dedicated fractions to provide these services, two extremes
can be thought of based on task achievement. At one end of the spec­
trum, distinct spacecraft tasks are carried out by its units. Though there

Fig. 5. Calculation of the decentralisation operation (M3 to M4) value. Adapted from Ref. [32].

5
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 7. DSS types (a) constellation (b) fractionated (c) formation (d) cluster (e) swarms (f) constellation of formation (g) federated.

is minimal cooperation between them, each unit is still highly dependent needed (i.e., not a fraction) to operate and so being completely self-
on the infrastructure. At the other end of the spectrum, fully fractionated contained. Independent satellites are built and placed in orbit for spe­
systems have modules which collaborate on accomplishing the same cific objectives, allowing them to employ their resources and capabilities
task towards the global mission objectives. There is a considerable for an opportunistic distributed mission [48]. Federated satellites are
resource dependency in this scenario and functionalities of the modules simply another example of fractionated spacecraft, as they combine
[36,42–47]. some of their capabilities and resources for a global mission [48–51].
Because the transferred resources are always underutilised in a module’s
2.3.3. Federated system primary mission, the nodes are complete and form heterogeneous sys­
In a federated system, a group of satellites work together to provide a tems, allowing for a new category of distributed satellite missions to be
specific service, but each satellite operates independently, with its own categorized, as shown in Fig. 8 [36].
mission and communication capabilities. A Federated Satellite System
(FSS) is a network of satellites that coordinate by exploiting the potential 2.3.4. Modular system
of their resources, with each satellite having all of the infrastructure Modular systems are more far-fetched DSS characterized by

Fig. 8. Concept of SMSRS (a) and deployment stages: (b) folded state; (c) unfolding; (d) unfolded state; and (e) working configuration. Adapted from Ref. [52].

6
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

physically joined modules as shown in Fig. 7(d). For instance, based on maintain a predetermined relative angular separation in their orbit.
the CubeSat standard, Jiping et al. [52] presented a new type of DSS It is worth noting that the relative angular spacing in circular orbits
with a reconfigurable construction and customizable function, dubbed remains constant throughout. The relative angular spacing in elliptic
Space Modular Self-Reconfigurable Satellite (SMSRS) design concept. orbits, on the other hand, changes depending on the satellite’s
The following are some of the features of SMSRS: (1) modularity; (2) location. Normally, these angles are determined when a particular
scalability; (3) structural reconfigurability; (4) fault tolerance; and (5) satellite (denoted as primary) is at the perigee [59].
functional adjustability. Fig. 8 shows the SMSRS configuration and the • Leader-Follower formation: This arrangement occurs when one
various deployment stages from a folded state to the working state. spacecraft is designated as leader and one or multiple other space­
Optical cameras, Synthetic Aperture Radar (SAR), communication and craft are forced to fly in formation with the leader. It should be noted
other payloads are among the ones which can best exploit this concept: that the literature uses a variety of terms when referring to this
SMSRS arranges and reorganises these payloads in a variety of space arrangement, including: (i) chief-deputy; (ii) master-slave; (iii)
orientations through structural reconfiguration, allowing it to carry out mother-daughter ships; (iv) primary-secondary; and more [23,59,
a variety of missions [52]. 60].

2.3.5. Swarms 2.3.7. Constellation of formation

Swarm intelligence enables natural (and artificial) multi-agent sys­ A constellation of formation combines the characteristics of forma­
tems decentralised control and self-organization capabilities as shown in tion flying satellites and constellations. It is therefore a set of formations,
Fig. 7 (e). Bloom [53] coined the term while researching complex each of which has a flight coordination between neighbouring satellites.
adaptive systems and it is made up of several principles (distributed In the constellation view, each formation can be described by the centre
parallel processing, super organism, group selection, apoptosis). A of mass of each formation, flying far away from each other but with a
typical swarm system has specific characteristics, such as a large number common mission goal as shown in Fig. 7 (f). Within the formation there
of homogenous agents that interact with one another via fundamental is typically relative navigation, guidance and control to acquire and
rules that exploit only local information. Information is exchanged reconfigure the layout. On the other hand, the constellation objective is
either directly with another agent or indirectly through the environ­ defined in terms of the orbit of a reference satellite (sometimes called
ment. Stigmergy is the name given to this indirect coordination mech­ chief) or some weighted position average as the centre of mass of each
anism [54]. The system’s overall behaviour results from the coordinated formation in the constellation. It shall be noted that each satellite must
interaction of the various agents. In these systems, individual behaviour obey to two objectives: to keep relative the formation and to remain in
is commonly described in probabilistic terms, as a result of local the constellation.
neighbourhood perception [55]. These characteristics allow a
multi-objective optimization of the swarm that maximizes scalability, 2.3.8. Hybrid mission
parallelization and fault resistance. Swarms are very adaptable while Hybrid mission architectures are the theorized new frontiers of the
also being extremely resilient (the system continues to work even if DSS concept and therefore entail a mix of distributed systems, which
certain components fail) and completely decentralised. It works whether bridge the gaps between the previously discussed concepts and are
they are being used to describe natural or human-made agents. Satellite therefore located around the centre of the bidimensional classification in
swarms are distributed missions in which the infrastructure modules are Fig. 6. For instance, fractionated satellites are capable of forming a
autonomous satellites performing their own functions/tasks without the constellation with other satellites (fractionated satellite swarm) or
interchange or collaboration of resources (such as data). A distributed cooperating with other units in more heterogeneous and complex situ­
satellite of this sort is made up of homogenous modules [56]. By ations (federation of fractionated satellites). It is worth mentioning that
increasing the number of modules dedicated to a certain task (i.e., by some DSS can change their typology in certain situations, depending on
adding redundancy), the set of constellation-conforming modules in­ the mission goals set by the ground segment [35,36]. The objectives of
creases the system’s usability, which benefits the system’s robustness. hybrid missions may alter because of technical issues (e.g., unit main­
For example, a deteriorating sensor in one of the modules of an EO tenance, repair and replacement, research potential) or for commercial
mission, does not prevent the operators from obtaining images. How­ reasons (exploitation of modules, sporadic provision of services).
ever, the amount of resources transferred (i.e., power, computational Federated satellite systems with modules that can function individually
resources) is almost minimal in this situation. This type of distributed or in formation in flight are a good example of this dynamism [36].
spacecraft can still communicate with one another to preserve the
intended formations and to relay critical trajectory information (e.g., to 2.3.9. DSS architectures and classifications
avoid collisions within the swarms) [35,56–58]. Nonetheless, their Table 1 provides a detailed description of different types of DSS ar­
functions are limited to local neighbours and their activities are done chitecture. Within the table, homogeneity is defined as the degree of
autonomously without transferring any resources altogether [35,36]. similarity between satellites within a DSS. On the other hand, the level
of the operational independence of a satellite or a fraction of distributed
2.3.6. Formation spacecraft is characterised as Operational/Functional Independence [12,
The coordinated motion of multiple satellites is known as satellite 16,35,61].
formation flying. In an effort to match the user’s requirements, different A Distributed Spacecraft Mission (DSM) is a mission in which
formations are possible. Satellite formation flight entails many archi­ numerous spacecraft/modules work together to achieve one or more
tectures depending on geometric configuration, operational arrange­ common objectives. This broad definition of DSM intends to avoid
ment and other characteristics [59,60]. The main operational specifying whether the multiple modules/spacecraft are launched
arrangements are: simultaneously, achieve common goals by design or ad hoc (i.e.,
application-driven), or if the common goals are scientific. Jacqueline
• Cluster formation: A cluster configuration occurs when a set of et al. [62] studied a variety of DSM attributes, classified them according
satellites are organised in a close formation and positioned in orbits to the taxonomy shown in Fig. 9 and defined all the concepts used in this
that keep them nearby each other. Satellites in a cluster normally taxonomy.
travel close together; which is different from trailing formations [46,
59]. 2.4. DSS functional architecture
• Trailing formation: Satellites share the same orbit and therefore
follow one another on the same path. The satellites are configured to A methodology for bottom-up design of a distributed architecture is

7
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Table 1
Summary of distributed architecture classifications. Adapted from Ref. [16].
DSS architecture Mission goals Cooperation Homogeneity Operational/Functional
Independence

Constellation Mission goal shared Cooperation is required to support the In general, homogeneous components, some Autonomous
(Iridium, GPS) mission goals differences possible (GPS generations)
Formation Trains Mostly Independent, but Cooperation from optional to required Heterogeneous components Autonomous
could be shared
Clusters Mission goal shared Cooperation is required to support mission Homogeneous components From autonomous to
goals completely co-dependent
Leader- Mission goal shared Cooperation from optional to required Heterogeneous components From autonomous to
Follower completely co-dependent
Swarms Mission goals shared Cooperation required to support mission From homogeneous to heterogeneous From autonomous to
goals components completely co-dependent
Fractionated Shared mission goals From optional (service areas) to required Heterogeneous components From autonomous to
(distributed critical spacecraft functions) completely co-dependent
Federated Independent mission Ad-hoc, Optional Heterogeneous components Autonomous
goals
Modular Mission goal shared Cooperation is required to support mission From homogeneous to heterogeneous From autonomous to
goals components completely co-dependent
Hybrid Mostly Independent, but Ad-hoc, Optional Heterogeneous components From autonomous to
could be shared completely co-dependent
Constellation of Mostly shared but could Cooperation is required to support mission From homogeneous to heterogeneous From autonomous to
formations be independent goals components completely co-dependent

presented, where elements of each layer are built up to reach the desired Consequently, DSS control architectures are also evolving to include
distributed architecture. The basic units arise from the bottom layer’s dynamic management policies [64]. For instance, the Local-Global
objects and elements. At the top, there is a launch plan that shows which paradigm just discussed is a mixed management policy, which is
vehicle will launch each module [47]. DSS hardware and software ar­ intended to provide an adaptive planning solution to an arbitrary
chitecture is discussed in the following sections. number of autonomous heterogeneous distributed spacecraft modules (i.
e., payloads, computational capabilities, ISL, etc.) [36]. In a DSS with
2.4.1. Hardware and software elements dynamic management policy, the top-level “multiple-tasks multi­
DSS hardware includes service avionics and mission-specific sys­ ple-modules” mission problem is translated into a “multiple-tasks sin­
tems, which typically combine payload instruments and supporting gle-module” arrangement by decomposition, as depicted in Fig. 13.
infrastructure [11,12,35]. Hardware architectures are modular and Once the allocated sub-tasks are completed by the individual modules,
expandable due to the inherent characteristics of the DSS. A possible their results are combined to synthesize the final solution.
space segment hardware arrangement is shown in Fig. 10 [14], while a
modern EO DSS control segment is shown in Fig. 11. In this particular 3. Autonomous operation in space
configuration, cloud-based mission control is used for planning and
scheduling purposes. Payload schedules and task allocations can be Autonomy is generally associated to a system’s ability to function
defined by direct ISL data sharing, ground control uploads, or both [20, without direct human interference, though it is a spectrum with several
21]. levels and grey areas. Therefore, in a system’s context, autonomy can be
Historically, spacecraft implemented a centralized data-bus archi­ defined as the ability to make informed, reasonable, self-reliant and self-
tecture, including a single On-Board Computer (OBC), normally redun­ determined decisions. A system should be able to sense, think and act
dant. Modern spacecraft, on the other hand, employ a number of OBC within its surrounding environment in order to be deemed autonomous,
modules, interconnected through either distributed, federated or therefore necessitating the capacity to detect its surroundings as well as
modular data networks. Software architecture for DSS is characterized some awareness of one’s own powers and how they affect one’s envi­
by integrating some level of system autonomy, where the components ronment and internal states. The autonomous inferences and conclu­
interact to: sions shall be drive by its own goals and necessary actions to achieve
them [65]. Additionally, an autonomous system shall be capable of
i. Distribute tasks between modules/components of satellites; reacting to non-nominal conditions by adjusting its behaviour to fulfil its
ii. Allocate infrastructure resources; goal while remaining safe and secure. The degree of autonomy that a
iii. Perform task scheduling in a distributed manner as per system achieves can be defined by the degree of off-nominality that it
requirements. can handle and the level of abstraction of its objectives [66]. Some
autonomous systems in aerospace carry out predetermined acts that do
An exemplary DSS software architecture is shown in Fig. 12, where not alter in response to the environment (automatic). Other systems
the corresponding hardware modules are not homogenous, indicating (automated) initiate or modify their behaviour or output in response to
they have different computational capabilities and availability times (i. environmental feedback, while more advanced systems (autonomous)
e., system encapsulation). The system consists of various autonomy combine environmental feedback with the system’s own interpretation
management entities (i.e., task planners) that interact to operate each of its current situation. Due to the need for reasoning regarding own and
spacecraft in synergy with one another. The Distributed System Layer environmental states, increased autonomy is usually associated to
(DSL) provides a common communication link between global and local increased “intelligence” or even “AI” for a specific mission and equated
entities [36]. The entire architecture entails two control levels in a with greater capability to adapt to the environment.
master-slave hierarchical relationship: (1) the global control level,
which is mainly relative to the software infrastructure domain; and (2)
the local control level, which is relative to each module domain. 3.1. Sense-think-act and layered architecture
In recent times, software architectures have evolved to suit multiple
and dynamically evolving operational environments [63]. A closed-loop (“sense-think-act”) system, as illustrated in Fig. 14,
describes an autonomous (machine) device or function for a layered

8
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 9. DSM classification. Adapted from Ref. [62].

architecture where the fundamental tasks are: The conventional layered architecture involves planning, task
sequencing and reactive capabilities. Deliberative, executive and func­
• Sensors (“sense”): gathering data from lower levels or hardware and tional layers are all terms used to describe these processes and con­
transferring it to a representation that the software can understand; ventional thinking is that autonomy shall imbue all of these. The layers
• Control (“think”): weighing sensory data, spacecraft information and are defined by their abstraction from the real environment and the time
desired outcomes before deciding which actions should be enacted; they take to complete an iteration, with clear response-time constraints
• Actuation (“act”): execute the operation that was determined by the as a function of the time horizon that can be considered. The functional
control analysis process (without further human interference). layer has fast turnaround requirements since this must maintain pace
with hardware sensors and actuators, and each component normally

9
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 10. DSS architecture. PR: Processor, DL: Downlink, PL: Payload are fractionated and distributed and connected through ISL. Adapted from Ref. [14].

Fig. 11. DSS architecture entailing cloud-based control segment. Adapted from Ref. [20].

only considers one task or series of tasks. The executive layer manages a technologies can be deemed genuinely autonomous, according to the
collection of tasks at the moment and it only needs to reply fast enough criteria in Table 2, whereas levels E1 to E3 relate with human-operated
to meet task action potentials and terminations. Finally, the deliberative or automated systems. Unlike autonomous systems, automated systems
layer takes into account many tasks and multiple possibilities, as well as can only deal with scenarios that their designers have predicted. These
future repercussions. It just needs to reply quickly enough to offer the systems will react to these conditions using so-called on-board control
executive extra job sets or plans when necessary. On the other hand, the procedures, which are pre-programmed sequences of operations (i.e.,
layers do not just indicate a boost in capabilities; trade-offs do exist. The events). In order to operate the entire mission autonomously, there is a
functional layer has access to detailed data about the hardware and need for autonomy in mission data management and mission fault
frequently performs complex numerical calculations to decide responses management. ECSS defines these capability levels as in Table 3 and
or provide data to the layers above. The executive layer usually includes Table 4 [68]. Both Earth-orbiting and deep space missions require the
contingency management and control skills that the deliberative layer use of robotics and autonomy. Most spacecraft operations’ control
lacks. Each layer executes a variation of the sense-think-act cycle in an functions and procedures are transmitted for immediate execution by
autonomy system. The overarching system of autonomy has a sense-act- telecommand or, more commonly, in precisely organised sequences at
think cycle as well [67]. specified times. Almost all remote sensing satellites, for example, gather
When looking at autonomy in space, the European Cooperation for images and downlink these to Earth at predetermined geographic areas
Space Standardization (ECSS) has defined four degrees of capability, while retaining the correct attitude using on-board sensors and reaction
with level E4 being the most autonomous. Only level E4 compatible wheels. On the other hand, Astronauts operate robotic systems in space,

10
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 12. Distributed software architecture. Adapted from Ref. [36].

Fig. 13. Task execution in a distributed system. Adapted from Ref. [36].

such as the Canada arm Remote Manipulator. Few autonomous aero­ making for remote sensing for both defence and civil applications, as
space systems make the decisions without human intervention in order well as future human space habitation, which could include both space
to attain high-level objectives. Variable autonomy, as defined by Proud tourism and deep space colonisation, are all plausible scenarios [66,
et al. [69] and Novaes [70], revolves around the concept of selecting 71–73].
desirable levels of autonomy while operating a space system. This allows Fig. 15 shows an evolutionary roadmap of space system capabilities
the autonomous system or the human user to alter the level of autonomy with a growing degree of autonomy over time. The four categories are
as needed by the situation. Autonomous space systems provide excellent defined as follows: (1) teleoperation (operated from Earth); (2) auto­
sensing and are therefore necessary if human usage and exploration of matic operation (pre-programmed self-controls); (3) semi-autonomous
space are to expand in terms of both reach and complexity. operation (start with predefined command sequence, where machine
Trusted autonomous space systems will allow such activity to adapts to the external environment); and (4) fully autonomous opera­
continue with confidence. For crucial space systems, several scenarios tion (autonomous decision making (goal-oriented)) [74]. Autonomy can
can be predicted. Some of these are already in the development and be incorporated in various segments of the satellite infrastructure. With
demonstration stages. For example, autonomous on-board data pro­ recent trends, TASO for space applications is becoming more popular. AI
cessing, on-orbit satellite servicing/repairing, analysis and decision- applications in the control and space segments have the potential to

11
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Table 3
Levels of autonomy for Mission fault management according to ECSS [68].
Level Description Functions

F1 Establish a safe space segment Identify anomalies and report to

configuration following an on- ground segment Reconfigure on-board
board failure systems to isolate failed equipment or
functions Place space segment in a safe
state
F2 Re-establish nominal mission As F1, plus reconfigure to a nominal
operations following an on-board operation configuration Resume
failure execution of nominal operations
Resume generation of mission
products

Table 4
Levels of autonomy for mission data management according to ECSS [68].
Level Description Functions

D1 Storage on-board of essential mission data Storage and retrieval of event

following a ground outage or a failure reports Storage management
situation
D2 Storage on-board of all mission data, i.e., As D1 plus storage and
the space segment is independent from retrieval of all mission data
the availability of the ground segment

adapting to machines”, whereas after WWII, ergonomics and human

factors engineering studies succeeded in establishing that “machines
shall adapt to humans”. This has proven often challenging especially in
the case of inherently non-adaptive systems because the varying
complexity and pace of tasks was entirely absorbed by the human
operator. The further step in this evolution is cognitive systems, where a
two-way adaptive interaction between humans and the machine are
possible.
For decades, the HCI community has used a human-centred
approach. The current generation is transitioning from human-centred
design to human-centred AI, which is not a novel idea. While a
technology-centric approach has dominated the development of AI
technology, academics have studied a range of human-centred ways to
address the particular difficulties highlighted by AI technology. Fig. 16
expands on the approach and the reader is referred to Ref. [75] for
further details.
Fig. 14. The “sense-think-act” cycle autonomous system layered architec­
ture [67]. Human-Machine Systems (HMS) incorporate the functions of human
operator (individual or a group) and a machine, which can be treated as
a single entity interacting with the external environment (see Fig. 17).
Table 2 An autonomous system or function is, by definition, out of human
Levels of autonomy for mission execution according to ECSS [68]. control to some extent. Humans can, however, exert some control during
Level Description Functions
the design and development stage at the point of task initiation and
during service, for example, by interrupting the system’s operation.
E1 Mission execution underground Real-time control from the ground
HMS can be controlled in a variety of ways:
control with limited on-board for nominal operations Execution of
capability for safety issues time-tagged commands for safety
issues 1. Direct control: Requires continuous interaction by a human oper­
E2 Execution of pre-planned, ground Capability to store time-based ator to control the system’s functions directly or indirectly, making it
defined, mission operations on- commands in an on-board
non-autonomous.
board scheduler
E3 Execution of adaptive mission Event-based autonomous 2. Shared control: Certain tasks are controlled directly by the human
operations on-board operations Execution of on-board operator, while others are controlled by the machine under the op­
operations control procedures erator’s supervision. The aim of shared control is to:
E4 Execution of goal-oriented mission Goal-oriented mission re-planning a. Combine the advantages of human control (global situational
operations on-board
awareness and decision) and computer control (high-speed, high-
accuracy actions).
increase the value of both ground and space operations. b. Partially overcome human control limitations (attention period and
field of vision limitations, tension and fatigue) and machine control
limitations (limited situational awareness and decision-making ca­
3.2. Human-machine interaction at increasing levels of autonomy pacity, sensing uncertainties).
3. Supervisory control: a device executes tasks autonomously while a
Since the industrial revolution, machines have been typically used as human operator supervises and provides guidance, intervenes and
a tool to aid humans in executing tasks. Before World War II (WWII), the reclaims control as required.
human-machine interaction design paradigm was based on “humans

12
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 15. Evolution of autonomy in space systems. Adapted from Ref. [74].

3.2.1. Human-on-the-loop control

Autonomous control is challenging in most real-world environments
because the operating environment is complex, unpredictable and dy­
namic in nature. Human supervisory control, on the other hand, allows
operators to exert some control through “human-on-the-loop” moni­
toring and intervention (as opposed to “human-in-the-loop”, which
conventionally refers to direct control). There may be several loops in
which the operator may intervene, each with different outcomes,
ranging from lower-level roles (monitoring and amending machine
resolutions) to higher-level strategic tasks (planning evolving mission
goals and operational constraints). In any case, successful human-on-
the-loop monitoring and intervention necessitate constant situational
awareness as well as sufficient time to intervene (i.e., override, deacti­
vate, or take other actions) and a way to interfere, such as a permanent
contact connection (for remotely operating systems) and/or direct
physical controls that allow the user to regain control or deactivate the
machine. Unfortunately, even though the human-on-the-loop model
meets all the above requirements, it is not a silver bullet for maintaining
successful control over autonomous systems due to well-known human-
machine interaction issues. These include:

1. Over-trust in the system, or automation bias, occurs when humans

tend to put too much confidence in the operation of an autonomous
Fig. 16. Framework for human-centred AI. From Ref. [75].
machine;
2. The diminished awareness by the human operator(s) of operational
Supervisory control is often used in applications where direct or
and environmental conditions due to not being directly tasked to
shared control of a machine is not feasible due to communications de­
address these;
lays between both the operator’s commands and the system’s corre­
3. The ethical buffer, in which the human operator delegated moral
sponding operation, such as in systems working in outer space or deep-
obligation and accountability to the system, which is viewed as a
sea areas.
valid authority.
The autonomy of a machine influences the number of tasks it can
complete as a result of the demand and regularity of human-machine
3.2.2. Adaptive and cognitive HMS
interaction. From the human interaction perspective, the levels of au­
There is no universally accepted models and optimality criteria for
tonomy vary from teleoperation to complete autonomy. Beer [76] pro­
human interaction with autonomous systems. The need for human su­
poses a structure for categorising levels of autonomy and guidelines for
pervision and the acceptable level of autonomy is related to the
choosing and maximising the appropriate level of human-machine
complexity of the environment wherein the system operates as well as
interaction centred on the machine’s intended purpose. The catego­
the complexity of the role it executes. It is common understanding that
risation is shown in Table 5. Humans are involved at all levels except the
the higher the complexity of the task and/or the environment, the
final stage of autonomy where in the aerospace context it is more
greater the challenges for autonomy design and verification, particularly
appropriate to refer to trusted autonomy [77].
for safety-critical tasks and environments in which a system failure may
cause fatalities, injuries or property damage. This is one of the critical

13
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 17. CHMI2 framework [85].

aspects that will be addressed in evolving safety certification standards, beneficial for space applications.
such as the evolutions of SAE and MIL-STD-822F [77,78]. When a poorly In space applications, CHMI2 functionalities can improve the oper­
designed autonomous system is used in an unpredictable, uncontrolled ational effectiveness of spacecraft operation as well as improve the
environment, there is a high risk of failure with unforeseen conse­ overall safety and effectiveness of operations [80–82]. CHMI2 supports
quences. Nonetheless, recent technological advancements in complex human-machine teaming, whereby a system senses and adapts to the
control software, such as AI techniques, improve the adaptiveness of the mission environment and the cognitive state of the operator. The CHMI2
system and therefore increase the degree of autonomy that can be suc­ concept supports Trusted Autonomous Operations (TAO) in both
cessfully implemented for more complex tasks in complex mission-critical and safety-critical applications [18]. The definition of
environments. CHMI2 builds on and capitalises on significant developments in aero­
Cognitive Human-Machine Interfaces and Interactions (CHMI2) is a space avionics human factors [83,84], which are detailed in Refs. [83,
new concept to human factors engineering in aerospace that in­ 84]. The CHMI2 framework’s primary feature is an expansion of a device
corporates adaptive functionalities into the design of the operator’s monitoring approach that assesses a HMS entire integrity by including
command, control and display capabilities [79,80]. A CHMI2 system both cognitive (human) and automated (machine) components. It
evaluates human cognitive states built on critical psycho-physiological conceived to characterise the operator’s actions that resulted in a certain
observables being measured [80]. The cognitive states have been uti­ mission outcome by detecting specific features that can affect cognitive
lised to estimate and improve the operator performance in the accom­ states (for better or worse). This closed-loop input helps to improve the
plishment of tasks to improve the efficiency and effectiveness of the HMS trustworthiness in important areas like the initial design process. It
overall human-machine teaming. Moreover, the result in the literature supports cognitive system engineering activities, such as the creation of
[80] indicates that higher levels of CHMI2 supported automation are system automation methodology based on operator policies and online

14
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Table 5 1. A foundational layer that covers traditional methodologies like sta­

Levels of automation. From Ref. [76]. tistics and econometrics, as well as complexity theory and game
Level Name Description theory.
2. A behavioural layer that comprises operational procedures including
1 Manual Human performs all mission aspects
2 Tele-operation Machine assists in task execution automated processes, machine translation, as well as collaborative
3 Assisted Tele-operation Machine assists in sensing and task execution and adaptive systems, among others.
and intervenes when needed 3. A sensory layer that provides language, audio and visual information
4 Advisory Execution Human formulates mission and machine to the model.
executes the task
5 Skilled Execution Machine assists in sensing and planning and
4. The “intelligence” is provided by a cognitive layer that incorporates
executes the task ML, reasoning and information representation.
6 Shared Control with Autonomous Machine operations with human
Human Initiative oversight These definitions are helpful in thinking about the purpose of the
7 Shared Control with Autonomous Machine operations with human
strategies being used. A combination of these would be used for the most
Machine Initiative assistance
8 Supervisory Control Autonomous Machine operations with a advanced AI systems.
human directive
9 Executive Control Autonomous Machine operations with human 4.1. AI techniques
override
10 Full Autonomy Autonomous Machine operations without
human intervention
AI, in contrast to natural intelligence, is the study of intelligence as
manifested in computer systems and observed in people and other life
forms [87]. To be considered intelligent, a computer system must be
adaption of the HMS based on the cognitive state of the person and the capable of making reasonable judgments based on experiences and ob­
operating environment, during the early design process [82]. servations of the world (or a simplified model of it) and a set of objec­
The evolution that leads to introducing the CHMI2 over the time is tives to be met. By applying suitable AI techniques, satellite systems can
shown in Fig. 18. Similar to what has happened in the computer era, AI make decisions in real-time without the need for explicit instructions. A
technology has enabled a new sort of CHMI2 collaborative interaction plethora of studies coupled with many tests are underway to implement
that would eventually lead to a paradigmatic shift in CHMI2 application AI-based technology in space systems, with various projects beings
areas in the AI era, resulting in new design thinking and approaches to carried out [71,73]. Some of the most commonly used AI methods are
AI system development. Subsequently, by using the human-machine co- shown in Fig. 19. AI can also be classified into two distinct groups,
operative relationship, it may be possible to optimise the benefits while strong AI and weak AI, based on the given task. Strong or general AI is
limiting the potential safety risks of utilising AI technology. concerned with the replication or outperformance of human brains,
including sensitivity, consciousness, mind and feelings. Weak or applied
4. AI techniques for space systems AI, on the other hand, is focused on completing a single task or resolving
a specific problem. Because most research in the space domain is limited
The term “intelligent space system” refers to space systems that to weak AI, this paper focuses only on applied AI.
operate independently using intelligent methods. To achieve autonomy,
AI approaches are used. Tasks are completed without the need for 4.1.1. Metaheuristics techniques
human interaction in this system. Not only can AI assist in physically Most conventional optimization methodologies use a deterministic
speeding up the process of manufacturing satellites, but it can also be rule to switch from a single point in the decision hyperspace to another.
used to analyse the process to determine if there are any ways to The main disadvantage of this method is the local convergence limita­
enhance it. Furthermore, AI may examine previous work to make sure tion which may prevent reaching the global optimum. Since stochastic
everything is completed correctly. Furthermore, including collaborative algorithms are designed to find the global best solution to problems with
robots (“cobots”) into the production process decreases the requirement multiple local minima (usually nonconvex problems), they typically
for human workers in clean rooms. It improves the consistency of pro­ overcome this issue. There are two kinds of stochastic algorithms,
duction processes that are prone to errors. AI, unlike humans, does not namely heuristic and metaheuristic, though the distinction is minor.
require rest or sleep in order to digest large amounts of data quickly. The Stochastic optimization is sometimes the second-best way to get a so­
basic objective of the techniques utilised can also be used to classify AI, lution. Conventional techniques such as linear programming, as well as
resulting in the following four layers [65,73,86]: specialised approaches that take full advantage of problem under­
standing, should be investigated first. Classical and specialised methods,
on the other hand, are often naive, whereas heuristic and metaheuristic
paradigms can be utilised to various conditions. One key advantage of
heuristic and metaheuristic paradigms is their robustness. In this
context, robustness refers to an algorithm’s ability to solve a wide range
of problems and even multiple sorts of problems, with only slight
changes to account for each problem’s specific properties. A stochastic
algorithm typically requires the length of the problem-solution vectors,
certain information about their encoding and an evaluation function,
with the remainder of the programme, remaining unchanged. A heu­
ristic algorithm is a strategy that uses a rule (or a set of rules) to find (or
try to find) appropriate solutions at a low cost of computing. Theoreti­
cally, a heuristic provides (eventually) a decent answer with relatively
little effort, but this does not ensure optimality. Heuristics are a
straightforward way of showing which of many options appears to be
the best [88,89]. The so-called metaheuristic algorithms are an exten­
sion of heuristic algorithms. Meta signifies “beyond” or “higher level,”
Fig. 18. The development of the human-machine relationship across time. and metaheuristics outperform simple heuristics. Heuristics use
Adapted from Ref. [75]. problem-specific information to identify a “good enough” solution to a

15
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 19. Most common AI techniques suitable for space systems.

given problem, but metaheuristics, such as design patterns, are broad • Relying on direct fitness data rather than function derivatives or
algorithmic notions that may be applied to a variety of problems. other similar details;
Importantly, all metaheuristic algorithms use some sort of randomiza­ • Using probabilistic, rather than deterministic, transition rules.
tion and a trade-off between local and global search. Practically every
metaheuristic algorithm strives to be appropriate for global optimiza­ Population-based algorithms adopt a similar approach, regardless of
tion [90]. The following features are shared by almost all metaheuristic the applied paradigm and follow from the algorithm below.
algorithms [91]:
1. Initialise the population;
• They are nature-inspired, relying on physics, biology, or ethology 2. Fitness is calculated for each individual in the population;
principles; 3. Produce a new population-based on some rules that strictly depend
• They use stochastic components (incorporating random variables); on the fitness of each individual;
• They do not use the objective function’s gradient or Hessian matrix; 4. Repeat steps 2–4 until a condition is met.
And
• They have multiple parameters that must be adapted to the nature of 4.1.2. Machine learning techniques
the problem. ML approaches are a subset of AI techniques that allows for the
creation of analytical models to be automated. It is a branch of AI based
Metaheuristic optimization algorithms can solve complex problems on the idea that computers can learn through data, identify patterns and
over several iterations. Because of their inherent versatility and make judgments with small or no human intervention. A ML process is
simplicity, metaheuristic algorithms have recently attracted a lot of shown in Fig. 20. A model that can be queried by an application is
attention. Metaheuristics can be broadly classified into four different trained based on a data or knowledge base. Regardless of suitable con­
types; the first one is ancient inspired, mainly based on the Giza pyramid ditioning, data selection, or overfitting, the model improves with a
construction. Mutation, reproduction, recombination and selection are larger database and longer durations of training. If the model can learn
the fundamental processes involved in evolutionary algorithms, which in the field, the application can add data to the knowledge base during
are based on the survival fitness of candidates in a population (i.e., a set
of solutions) for a specific environment. The idea of population-based
metaheuristics is to construct a solution that combines components of
good solutions. Trajectory-based metaheuristics are based on the idea of
developing a solution and iteratively refining it (moves). The reader is
referred to Refs. [88–96] to get a complete understanding of these
concepts. A population-based metaheuristics approach, i.e.,
nature-inspired, as indicated in Ref. [97] are, distinguished by:

• Their use of a population of points (potential solutions) in their

search;
Fig. 20. Generalized machine learning process [98].

16
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

runtime and train the model with it. between the training data points’ features or characteristics and the
A brief review of ML methods and an exposure to commonly used labels that were allocated to them. The model would then be put to the
domain terms are provided in the preceding s sections. Different tech­ test to see if it would generalise to new information points or ’incidents’.
niques for classifying ML methods have been taken in the literature. The Before being deployed to service, trained models were fine-tuned based
most prevalent taxonomy in which techniques are classified according to on the assessment findings to create a model that extrapolates well with
the type of learning system used. The main variants of ML techniques the new data. In most supervised approaches, the learning method is to
are: keep track of the difference between both the model prediction and the
Supervised learning: the algorithm is supplied with labelled label and use it as an error term to drive model updating [3,81].
training data, i.e., appropriate labels are included in the desired result. Unsupervised Learning: the training data provided to the algorithm
During the training phase, a model is constructed that specifies the link is unlabelled and the relationship model is created solely on the data

Fig. 21. Neural Network Categories [81].

17
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

attributes. These methods include clustering, dimensionality reduction though the concepts seem to vary. In addition, they integrate within
approaches and association rule-learning methods [3,81]. projects, as the models are structured to execute the task in the most
Semi-supervised Learning: is a technique that involves training efficient way possible, not to stick to a pure type. So, what distinguishes
using a small amount of labelled data and a large amount of unlabelled ML, DL and RL precisely is actually a difficult question to address, which
data. Semi-supervised learning is a sort of learning that falls somewhere can be addressed by following the standards shown in Fig. 22.
between unsupervised (in which there is no labelled training data) and
supervised learning (with labelled training data) [81]. 4.2. AI in space operations
Reinforcement learning: a model is trained to iteratively learn a
behavioural policy through a large number of simulations, which is The field of data-driven AI has a wealth of valuable and adaptable
called the training set. The agent learns how to attain a goal in an un­ tools that can be used in various applications with minimal enhance­
certain and potentially complex environment through trial and error. In ments. While AI has been used successfully in space, it is still constrained
Reinforcement Learning (RL), AI is presented with a game-like scenario. to offline data processing but has not yet been utilised fully “on the
The machine uses the method of trial and error to find a solution to the edge” within spacecraft. Space Applications of AI has the potential to
problem at hand. AI gains either rewards or penalties for the acts it takes significantly affect human and robotic space exploration missions in
in persuading the system to perform the actions the programmer desires. several different ways.
Its aim is to increase the overall award. As a result, the AI agent learns to As time progresses, AI will complement the space exploration ac­
do the best set of actions or rules to achieve a user-defined reward tivities in a variety of ways, as seen in Fig. 23. Some of the main ap­
function. To find the optimal system settings autonomously, sophisti­ plications of AI and intelligent systems in the near-Earth region and
cated learning and decision-making functionalities must be adopted, for multi-planetary exploration in outer space are:
which ML algorithms offer several advantages [99].
Deep learning is a subset of ML that learns from data using artificial • Remote sensing data analytics.
neural networks. Artificial Neural Networks (ANN) are based on the • Satellite trajectory planning and collision avoidance.
human brain and consist of layers of interconnected nodes. Each node in • Satellite health monitoring.
a Neural Network (NN) can learn to do something simple, like recognise • Satellite communications.
edges in an image or identify patterns in text. By combining these simple • Deep-space and multi-planetary exploration.
tasks, neural networks can learn to execute complicated tasks such as
object recognition in images and language translation. Deep learning Table 6 gives a summary of AI techniques utilised in different
(DL) is a powerful tool, but it must be used with caution. DL models are spaceflight operational tasks.
prone to bias and can be exploited to generate damaging or misleading
resources. It is critical to be aware of these hazards and to take pre­ 4.2.1. AI for remote sensing data analytics
cautions to mitigate them. ANN is a type of AI that tries to replicate the EO and astronomical satellites nowadays process about 150 giga­
way the human brain works. The processing units are ANN that are bytes of observation data per day or more. The autonomous acquisition
composed of inputs and outputs. ANN are a kind of ML technology that is and processing of images introduces a number of opportunities where AI
inspired by biology and is supposed to work in way similar to the brain can assist greatly. Without AI, humans are largely responsible for
(loosely). Fig. 21 depicts the main types of NN types [3,81]. interpreting, comprehending and analysing imagery [86]. By the time a
There are some distinctive attributes to be considered with specific human arrives around to evaluate an image, the satellite may have
reference to the adopted neuron model network, learning method and moved back to the same place, requiring more refinement of the image
topology. Convolutional Neural Networks (CNN) are regularised ver­ analysis. The AI-enabled recognition gives the researcher a lot of power
sions of the multilayer perceptron. Such perceptron are typically fully when it comes to image analysis and reviewing the millions of images
connected networks, in which each neuron in one layer is connected to produced by satellites. On the other extreme, AI can analyse images as
all neurons in the next layer [99]. CNN are generally used for segmen­ they can be captured and identify whether they have any anomalies
tation, classification, image processing and other auto-correlated data [102,109,110]. The use of AI to process satellite images also eliminates
processing. They are also utilised for speech recognition. Convolution is
the process of applying a filter to an input signal as it is being played
back. When looking for specific elements in a picture, it may be more
productive to look at little sections of the image rather than the entire
image at once. Among the most common applications of CNN is image
classification, such as discriminating between satellite images that
feature roadways and those that do not. The use of CNNs for other
standard functions such as image segmentation and signal processing is
also a good fit for them. Each layer of a CNN model learns a collection of
convolutional kernels throughout the training operation, which is
essentially what happens during the training phase. During the
deployment of the model, the trained kernels extract spatial informa­
tion. Each convolutional layer is made up of a collection of filters known
as convolutional kernels, which work together to create the final result.
Filtering is accomplished by applying a subset of the input pixel values
to a matrix of integers that has the same dimensions as the kernel
[100–103]. The CNN thus far seen are classic feed-forward networks, in
which activations travel from the input to output layers at a pre­
determined rate. The network output is distinct from the outputs of
previous timesteps at any given timestep. Recurrent Neural Networks
(RNN), on the other hand, keep track of previous outputs at each epoch
by integrating feedback loops. RNN are better at learning temporal re­
lationships in data sequences than NN, which are meant to learn spatial
patterns [81]. There is no sharp divide between these subtypes, even Fig. 22. Classification Standard [104].

18
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 23. AI-augmented space exploration.

minimising communication, lowering battery use and accelerating the

Table 6
image collection process [109].
Summary of AI techniques used in spaceflight tasks.
Spaceflight Tasks Techniques References 4.2.2. Satellite mission planning and collision avoidance using AI
Remote sensing • On-board data processing [102,103, Satellites are complex systems of systems and many issues could
• Image processing for precision 105–118] affect their operation, ranging from equipment malfunctions to crashes
agriculture and agroindustry
with other satellites/debris. AI can be used to maintain track of satellite
• Hyperspectral image
classification
health and ensure that it continues to function properly. AI can keep
Communication • Satellite communication [119–124] track of sensors and equipment in real-time, generating integrity flags (i.
• Intersatellite communication e., both predictive and reactive alerts) and, whenever appropriate,
• Space communications cognitive taking corrective action. AI can also be utilised to control satellite orbital
engine
parameters and to perform advanced navigation tasks (e.g., DSS relative
Automated control and • Satellite attitude control [125–133]
navigation. • Autonomous proximity navigation and formation flight). AI can also enhance the surveillance
operations and docking of tasks, such as tracking of other spacecraft and debris. Once AI has
spacecraft. inferred the state vector of tracked objects, orbital propagation can be
• AI-based control systems
accomplished to identify collision risks and, when necessary, to imple­
Satellite Health • Automatic anomaly detection [134–140]
Monitoring techniques
ment an optimal manoeuvre for collision avoidance [134,136,138,139,
• Intelligent health monitoring 164,165].
systems
• Spacecraft structural health 4.2.3. Intelligent Health and Mission Management
monitoring
Satellites have intricate subsystems and equipment that are required
Deep space and Multi- • Exoplanet detection [119,
Planetary Exploration • Interplanetary trajectory design 141–148] for operation. Malfunctions in this equipment have the potential to lead
• Deep space communication to several on-orbit failures, such as attitude control malfunctions, bat­
Satellite Mission Planning • Autonomous planning and [15,133, tery and solar-array failures [166]. Conventionally, satellite operations
scheduling 149–155]
involve very simplistic on-board Fault Detection, Isolation and Recovery
• Trajectory optimization of the
space launch vehicle
(FDIR), and fault diagnostics and prognostics carried out by human
• Spacecraft trajectory optimization operators on the ground from telemetry data. This arrangement is
Space Traffic Management • Collision avoidance [6,156–163] resource-intensive and time-consuming, therefore inadequate for mis­
• Separation assurance sions involving large and complex DSS. AI can address this challenge by
• Space Based Space Surveillance
monitoring the health of all satellite subsystems and providing advanced
• Space domain awareness
diagnosis and prognosis, including both predictive and reactive integrity
features supporting TASO. Additionally, a properly developed AI system
the need for a lot of communication to and from Earth to analyse images can reconfigure DSS resources to mitigate its impact, thereby imple­
and decide whether or not a new one should be acquired. AI therefore menting the so-called Intelligent Health and Mission Management
saves computing power and radiofrequency spectrum utilization by (IHMM), which allows to drastically reduce the demand for human

19
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Table 7
Summary of AI techniques and their specific function in different space missions and operations.
Mission/Operation Type of AI Specific Function (performed by AI)

Earth Observation (Satellite Image Analysis) Machine Learning Object detection and classification in satellite imagery [98,175,176]
Land cover classification and mapping [177–179]
Change detection for monitoring environmental dynamics [180–182]
Anomaly detection for identifying environmental hazards [181,183–185]
Computer Vision Image segmentation, object detection [186,187]
Data Mining Detecting changes in Earth’s gravity field, studying water resources [188–190]
Neural Networks Soil moisture mapping, drought monitoring [191–193]
Deep Learning Image classification and segmentation for land cover analysis [194–196]
Change detection for monitoring environmental changes [197–200]
Object detection and tracking in satellite imagery [201,202]
Anomaly detection for identifying irregularities [203,204]
Data fusion and integration for combining heterogeneous data [205–209]
Super-resolution and image enhancement for improved analysis [202,210,211]
Automated feature extraction for object identification [212,213]
Data compression and transmission optimization [214–217]
Data quality assessment for ensuring reliable data [218,219]
Space Debris Monitoring and Mitigation Machine Learning Object tracking and prediction of space debris movement [220,221]
Collision prediction and avoidance maneuvers [222,223]
Debris removal mission planning and optimization [224,225]
Astronomy Machine Learning Image classification, data processing [226,227]
Neural Networks Exoplanet detection, data analysis [148,228–230]
Deep Learning Star classification, astrometric measurements [231–234]
Computer Vision Image processing, object detection [235,236]
Data Mining Image reconstruction, black hole imaging [227,237,238]
Pattern Recognition Galaxy classification, dark matter detection [239,240]
Satellite Operation Autonomous Navigation Machine Learning Trajectory optimization and path planning for spacecraft [241–243]
Collision avoidance and space debris detection [222,223,244]
Adaptive control systems for spacecraft maneuvering [245–247]
Spacecraft Health Monitoring Anomaly detection for identifying system failures [134,136,248,249]
Predictive maintenance for optimizing maintenance schedules [167,250,251]
Decision support for diagnosing and troubleshooting issues [134,136,167]
Mission Planning and Scheduling Optimization Algorithms Resource allocation optimization considering mission objectives [252–255]
Real-time adaptation and adjustment of mission plans [256–258]
Communication NLP Algorithms Voice command recognition and spacecraft control [259–261]
Intelligent responses and communication with astronauts [262,263]

involvement and is potentially much timelier and more effective in 4.2.5. AI for deep space and multi-planetary exploration
addressing anomalies [167]. After immediate automated mitigation and Even satellites on other planets or in interplanetary space, like the
recovery measures, human operators can still implement offline inves­ Curiosity rover currently on the red planet, use AI to operate. The Mars
tigation and forensics to obtain further information on newly discovered rover is using AI to assist it in navigating the planet. The computer may
issues and resolve them. Since only non-nominal situations necessitate make several modifications to the rover’s trajectory every minute. The
operator intervention, the “human-on-the-loop” concept is promoted. Mars rover’s technology is quite similar to that used by self-driving
Increased on-board autonomy would allow for more complicated sat­ automobiles. The key difference is that the rover should cross more
ellite applications missions, as well as reduce human operator workload difficult terrain without having to worry about other vehicles or pe­
[168]. destrians. The rover’s computer vision systems analyse the difficult
terrain as it traverses. If an issue with the terrain is detected, the
4.2.4. Satellite communications using AI autonomous system adjusts the rover’s navigation or modifies its tra­
It can be challenging to communicate between Earth and space, in jectory to avoid it [22,71,142–144,174]. A summary of AI techniques
addition to keeping spacecraft operational. Interference with other sig­ and their specific function in space missions and operations is presented
nals and the environment depends on the state of the atmosphere and on in Table 7.
neighbouring entities. A satellite may have communication difficulties
to overcome as a result of uncertainties in the environment. AI is now
being utilised to control satellite communication in order to circumvent 4.3. On-board AI
any transmission issues. These AI-enabled technologies can figure out
how much power and what frequencies are needed to send data back to Currently, AI is employed in space to improve monitoring and di­
Earth or to other satellites. The satellite does this on a regular basis with agnostics, make predictions and analyse images. AI has not yet been
an on-board Al to allow signals to pass through as it travels through implemented on-board spacecraft. Image processing, instrument control
space [169–173]. In particular, AI can greatly support several commu­ and satellite navigation are just a few of the potential applications for AI
nication system tasks, including beam-hopping, anti-jamming, detecting on-board spacecraft. Training models on the ground and then uploading
ionospheric scintillation, network traffic forecasting, channel modelling, them to spacecraft is one method to implement AI on-board of the
telemetry mining, interference management, remote sensing, behaviour spacecraft. This would make spacecraft more autonomous and increase
modelling, space air-ground integrating and energy management. AI the value of the data they collect. Even the most computationally
should be used to produce more effective, reliable, consistent and demanding AI models can now be executed on mobile devices such as
high-quality communication systems in the future [122]. smartphones. This indicates that AI can be utilised on even the smallest
spacecraft. Table 8 provides examples of specific AI applications on on-
board of the spacecraft. Space exploration has the potential to be

20
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Table 8
A brief overview of AI on-board missions. Adapted from Refs. [264,265].
Mission Specific Function (performed by AI)

In-orbit servicing: Debris removal, Docking, i.Feature extraction

ii.Identification against 3D mesh model
iii.Obstacle avoidance
EO missions (to be scaled to mission size, criticality, i.Band co-registration for push-broom multispectral and hyperspectral images
duration) ii.Change detection in time series of Earth Observation images, various resolutions
iii.Cloud detection algorithms (F-mask or Sen2Cor, however, the whole Sen2Cor is quite big, maybe some essential parts of
it)
iv.Fire/flares detection
v.Image compression (jpeg/CCSDS), (preferably Earth Observation-like picture)
vi.Increase resolution of all Sentinel-2 bands to 10 m/pixel
vii.Monitoring of forest distribution
viii.Monitoring of ice at poles
ix.Open sea objects detection and monitoring
x.Reconstruction involving multiple Images alignment using SURF equivalent, like BRISK or ORB (SURF is patented) and
RANSAC
xi.Super-resolution (increase resolution using series of images) through compressive sensing methods, like over-determined
equations
xii.Supervised NN Image Classification of Multi-Spectral Images Based on Statistical Machine Learning (TBD if learning
speed should be measured as the benchmark as well)
xiii.Template matching (scale and rotation invariant) in Earth Observation-type image (e.g., from Sentinel-2)
xiv.Vessel detection/identification, integration and data fusion with AIS receivers - identification of piracy
Visible spectrum EO mission (For generic Imaging i.Active/adaptive optics: wave front analysis + actuation
Instrument calibration) ii.Auto-exposure
iii.Flat field dynamic correction
iv.Focal plane adjustment and calibration
v.Geometric calibration
vi.Top of Atmosphere calibration
Visible spectrum astronomy (For on board platform i.On-board platform imagers processing
imagers processing) ii.Identification of fast-moving meteoroids/disturbance/radiation
iii.Star tracing and multiple sensor data fusion
iv.Orbital propagation
Planetary Exploration i.Camera/LIDAR fusion processing.
ii.Identification of craters, boulders, obstacle avoidance, automatic path discovery
iii.Vehicle GNC
iv.Autonomous Landing
v.Robotic Exploration
Reconfigurable platforms/on-board telemetry analysis, i.Adapt platforms to change in requirements or new standards
FDIR ii.Autonomous failure prognostic and detection
iii.Autonomous Safe mode management.
iv.AI-based FDIR
Satellite guidance applications i.Autonomous AOCS management for constellations
ii.Autonomous collision avoidance
iii.Autonomous navigation
iv.Autonomous pointing and/or acquisition (AOCS-in-the-loop)
v.Payload-in-the-loop visual-based navigation
vi.SDR/Beamforming/Adaptive Coding and Modulation
vii.Smart FDIR/failure prediction/smart HKTM
New missions that are candidates for the use of AI i.Reconfigurable science (several missions with the same Hardware/Instrument)
ii.Servicing/Non-cooperative approach and rendezvous
iii.Debris detection and removal
iv.On-board feature extraction/mapping Raw data downlink only On-Demand basis or Added-Value basis
v.Rapid alert: fire, flood, earthquake detection

revolutionised by AI. By incorporating AI into spacecraft, scientists can systems are increasingly relying upon suitable forms of AI to autono­
collect more data, make better judgements and conduct space explora­ mously check for faults, detect/isolate them and recover optimal func­
tion in a safer manner [264,265]. tionality [266–269]. Concerning the control segment, various functions
are seeing an increasing adoption of AI [270–273]. Ground control
5. Space infrastructure evolution systems, in particular, is being extensively automated to reduce the
reliance on human operators and associated support personnel [268,
The benefits offered by AI algorithms will be fully achieved upon 269,274,275]. These systems can handle numerous spacecraft at the
integration in all space mission operational segments (Fig. 24): space, same time, normally requiring minimal human supervision and inter­
control, user, link segments and inter-vehicle communications. vention. For example, NASA’s Near-Earth Network and Space Network
use time-sharing to manage many spacecraft connections [276,277].
5.1. Space and control segments Table 9 summarizes the applications of AI envisaged for the evolu­
tionary space and ground segments.
The space segment involves a variety of specialised payloads, such as On-board control functionalities, have recently been the focus of
remote sensing, navigation and communications, to carry out the increasing levels of automation, as in the case of system health man­
intended missions [21]. In addition to other more mission-specific as­ agement, which evolved from traditional FDIR practices. The capability
pects extensively covered in other sections of this article, spaceborne to adapt and maintain dynamic plans for individual spacecraft prompts

21
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

the adoption of on-board autonomy for both mission planning and connect with one another. As a result, the authenticity and integrity of
execution [279]. Nowadays, most satellite operators/owners prefer these communications are critical. Space Traffic Management (STM),
on-board control due to the limitation in bandwidth and the ability to efficient and easier maintenance, are all benefits of vehicular commu­
perform autonomous operations [105–108]. These on-board control nication, which ultimately lead to safer space operations. Wireless and
systems are able to process mission and link segment data such as EO optical fibre communications are the most common forms of IVC [275].
data [271], space weather data, etc. without human intervention and Different ISL topologies are the following 1) Ring, 2) Star, 3) Mesh, or 4)
employing intelligent algorithms. Table 10 summarizes the comple­ Hybrid, depending on the communication links between the DSS. These
mentary nature of ground and space control segment components. topologies are depicted in Fig. 25, with the ISL represented by the ar­
rows. Liz Martinez et al. provides the various strategies that are suited
for DSS [211]. Technology based on RF, has traditionally been utilised
5.2. Inter-vehicle communications for inter-satellite communication; however, increasingly, modern sat­
ellites are turning to technology based on lasers and optics in order to
The term Inter-Vehicular Communication (IVC) refers to data ex­ communicate with one another [281–283]. Fig. 26 illustrates the
change within the DSS framework which can be achieved with ISL. The different links between space and ground assets, which can support
different topologies are shown in Fig. 25. command and control functionalities such as initiate equipment di­
Mission tasks, including redundancy of operations, command and agnostics, reset the state of equipment, and/or begin the vehicle’s pro­
control, mission activities, tracking, networking of computing capabil­ pulsion systems. These commands can be transmitted to a space vehicle
ities and communications, are all performed when functioning satellites via a variety of communication mechanisms. For instance, they can be
sent by optical fibre to a remote ground station, where they are sent to
the satellite through a direct Radio Frequency (RF) or optical link.
Alternatively, a space relay system can be established, in which initial
commands are communicated from the ground through RF or optical to
a relaying satellite, which will then be transmitted via radio frequency to
the target satellite. Finally, mobile devices and technologies that are not
tied to a specific ground operation region, like IVC, can be used to send
commands to a satellite or its payload [275].

5.3. User segment

The user segment includes civil and military equipment that receives
and processes satellite signals. Specialised receivers and/or transceivers
are required in most applications such as communication, navigation,
positioning, time dissemination and research (such as measurement of
atmospheric parameters). As the term suggests, receivers have no re­
striction on the number of users they may serve because the signals are
sent in the service volume to each suitable receiver, making communi­
cation one-way, meaning the user does not broadcast but only receives.
The variety of transport, surveying, agricultural, industrial, defence and
recreational applications currently found for Global Navigation Satellite
Fig. 24. Example of modern satellite system operational segments. Systems (GNSS) receivers, has created a large and diverse user com­
munity base. Today, most mobile and personal communication devices
Table 9 feature embedded GNSS receivers, with an increasing number of ap­
AI application in space and ground segments. Adapted from Ref. [278]. plications depending on satellite navigation signals. The aviation com­
Space segment Smart Payloads • Ground target analysis
munity strongly supports GNSS evolutions because they provide
application segmentation, detection and tracking dependable aircraft’s positioning and navigation [284,285]. Satellite
•Cloud segmentation communication services are also increasingly more accessible and
•Vegetation detection and health extensively used for beyond Line-of-Sight (BLOS) aeronautical commu­
monitoring
nication networks, including those serving uninhabited/autonomous air
•Multispectral and Hyperspectral
information extraction and surface vehicles. Other common applications include Earth Obser­
Smart Spacecraft •Automated planning and reasoning vation (EO), where service owners/users access various kind of raw,
•Adaptive on-board operations semi-processed and processed datasets including weather, agricultural,
•Event-driven and goal-driven geospatial, etc. [286–288].
autonomy
•Autonomous Moon/Mars/Asteroid
landing
Ground segment Trend Analysis •Anomaly detection on multivariate
applications time-series
•Intelligent health and mission
management Table 10
•Correlation on multivariate time- Comparison of ground and space control segment. Adapted from Ref. [279].
series
Ground Control On-board Control
•Automated root cause analysis
Augmented •Satellite automated tasking and •CPU power available, •Reactive to the environment,
Operations planning •Software flexibility, •Processing data without
•Automated debris avoidance •The testing procedure not impacting the communication delay,
•Virtual assistants for ground operators mission, •Reduced communication to ground,
Generative AI •Dataset augmentation •Interactions with operators and experts in •Human intervention is limited.
•Super-resolution imaging a short loop,
•Data quality enhancement •Lower cost of software development.

22
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

and responding to anomalies can be lowered [71,106,108]. Increasing

the level of autonomy within the space segment is therefore critically
important [71]. However, autonomous data analysis and mission plan­
ning decisions taken on-board can challenge the human understanding
and situation awareness to a point where even a supervisory role would
be completely compromised. A co-evolution of space and ground control
segments is therefore necessary to allow a safe adoption of on-board
autonomy and thus achieve trusted autonomy [6]. Contemporary
research in human factors engineering is investigating how to evolve
Human-Machine Interfaces and Interactions (HMI2) in a way which best
matches the complementary human and machine capabilities but also
considering the ever-changing performance of both. This variability in
space and time calls for adaptive and cognitive forms of HMI2, which can
maximise human situation awareness and correct interpretation of a
variety of situations in a supervisory (human-on-the-loop) control
paradigm. It is increasingly evident that AI can be a critical asset in both
segments to endow the overall CPH system with the necessary cognitive
adaptation capabilities and therefore achieve trusted autonomy [85,
289]. The heuristic and unpredictable character of some operational
decisions aboard spacecraft and the difficulty of algorithmically
encoding elements such as expert judgment in such systems create
Fig. 25. Different ISL topologies [280].
considerable challenges on this front [143,268,269]. Hence the
co-evolution of these two segments will be essential to enable TASO.
5.4. Co-evolution of space and control segments Fig. 27 depicts a possible approach to address the challenges of space
and control segment co-evolution. In particular, the proposed architec­
Trusted autonomy is envisioned to increase efficiency and cost- ture supports autonomous goal-based mission planning based on both
effectiveness in both space and control segments, especially by SBSS data and ground-based surveillance data from the Consultative
lowering the significant recurring costs of operating space missions. The Committee for Space Data Systems (CCSDS), [6].
increasing amounts of data processing and interpretation performed The accessibility of the information necessary to make actions is a
within the space segment will lead to less raw data downlinked and critical consideration. The availability of such data should then define
process on the ground, thereby reducing the requirements for spectrum, the chosen paradigm. For example, on-board autonomy is the obvious
human operators and ground-based computational resources. By auto­ option if the goal is to make the system more resilient to failure, but if
mating the processing and interpretation of operational information the intent is to maximise the system responsiveness to user requests,
from different sources, the technical expenses of operating a spacecraft

Fig. 26. Links between spacecraft and ground segment.

23
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

more complexity should be handled by the ground segment. Both op­ working at normal levels and its trusted autonomous reconfiguration
tions are generally required; hence, autonomy must be flexibly managed capabilities allow it to redistribute workload without interference from
by either control or the space segment with variable levels of autonomy the ground. Therefore, iDSS enables real-time or near-real-time opera­
and associated task redistribution. Distributed mission scenarios may tions for time-critical events such as natural disasters (in this case,
not always be adaptive to single spacecraft solutions. This is the case wildfires). In this context, ISL allows for data sharing and reactive AI
because distributed missions introduce a new level of complexity to algorithms perform on-board data processing so that the iDSS can
mission planning, such as the spacecraft coordination mechanism. This reconfigure rapidly [290,291].
is completely dependent on IVC and computing power, which are both
finite resources. Determining the right degree of responsibility between 6. Current and future applications
both the control and space parts is one of the primary difficulties for
future missions [279]. Ultimately, even goal-based operational concepts The following subsections examine relevant use cases of AI in space
alone cannot address the requirements of trusted autonomy without operations. These examples illustrate the significant potential of AI to
resorting to cognitive/adaptive forms of HMI2 such as the CHMI2 enhance a variety of real-world space operations. From autonomous
concept overviewed in satellite control to on-board data processing, AI has proved to be a
section 3.1.2. potent tool for improving the efficiency, dependability and safety of
space missions. These e demonstrate the potential for AI to revolutionise
space operations and bring us closer to understanding the complexities
5.5. Intelligent DSS operations
of the Earth and deep space.
On-board data processing (i.e., reducing data link load and relaxing
ground-based processing tasks) requires both AI algorithms (software) 6.1. PhiSat-1 mission
and suitable hardware (such as accelerators). TASO is therefore possible
only with cutting-edge AI enabled astrionics. Astrionics are generally The European satellite PhiSat-1 (φ-Sat-1) is the first to demonstrate
subdivided in two categories: (i) Service astrionics, which includes on- how on-board AI might improve the efficiency of EO data transmission
board GNC, communication, mission reconfiguration, etc.; and (ii) back to Earth. Enhanced imaging capabilities employing a hyperspectral
Mission astrionics, which deals with mission-specific satellite-related camera and HyperScout-1 imager, as well as advancements in a Feder­
tasks such as fusion, analysis, georeferencing and other data-related ated Satellite System (FSSCat) mission using AI, have resulted in PhiSat-
tasks. Sharing information about the acquired data is made possible 1. The presentation of a cutting-edge AI accelerator, including the first
by iDSS, allowing for maximum scientific output to be achieved through AI algorithms for cloud screening, was the highlight of the show.
the use of opportunistic research. iDSS also offer real-time event man­ Applying ML algorithms to HyperScout-2 data in real-time allows for
agement i.e., disaster and rare events. The operational requirements can considerable improvements in inferring and presenting EO-based in­
be lowered with iDSS autonomy, allowing for human-in-the-loop oper­ formation at high temporal precision and in a short amount of time.
ations to be converted into human-on-the-loop activities. Humans will Scanning data before it has been downloaded is one possible use case.
be responsible for overseeing the operations in some capacity. Despite This was the case in the first experiment, which took place in orbit as
the loss of one spacecraft, intelligent DSS (iDSS) is able to continue part of the φ-Sat-1 mission and involved AI being utilised to retrieve

Fig. 27. Concept of space and control segment co-evolution by goal-based programming. Adapted from Ref. [6].

24
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

cloud coverage data from the gathered picture [272]. benefit of a fully trustworthy autonomous mission is lower operational
costs. This is especially relevant for small satellite missions, where costs
6.2. Disaster management of operations have been shown not to scale with spacecraft size and
constellations, where a large number of operators for each spacecraft
Climate action (SDG-13) is one of the United Nations (UN) Sustain­ simply cannot be afforded with hundreds of thousands of satellites. At
able Development Goals and wildfire is one of the catastrophic phe­ Technology Readiness Levels (TRL) 4 to 6, there are few mission man­
nomena that both impact and are exacerbated by climate change. Large- agement technologies available in the industry, MIRAGE [278] and
scale wildfires have increased in frequency and size in Australia and Antelope [298] are the notable ones.
other nations in recent years. The capacity to change the spacecraft
cluster/formation and progressively add new or upgrade existing sat­ 6.5. Space Based Space Surveillance
ellites in the formation has surely increased mission value in recent
improvements in DSS. These qualities have intrinsic advantages such as The effectiveness of space object surveillance in the rapidly evolving
greater mission efficacy, multi-mission capability, design flexibility and space domain is challenged on multiple fronts and is particularly known
so on. The predictive and reactive integrity characteristics given by AI, to be limited in the size and optical magnitude of the objects being
including both on-board satellites and in the ground control segments, monitored. Additionally, ground-based observations are prone to
make TASO practical. The DSS must be able to reconfigure indepen­ disruption and disturbance by atmospheric weather effects. SBSS is
dently in order to efficiently monitor and manage time-critical events therefore widely considered as the long-term complement to ground-
such as disaster relief operations. To achieve TASO, the DSS architecture based RSO tracking as it is not affected by atmospheric weather and is
should support reconfiguration and spacecraft should communicate able to detect much smaller and fainter objects. The integration of SBSS
with one another via an ISL. Recent breakthroughs in AI, sensor and has the potential to enhance tracking accuracy, RSO predictability and
computing technologies have created new promising concepts for the weather independence, thus potentially supporting the development of
DSS’s safe and efficient functioning. Combining these technologies en­ an integrated Space Domain Awareness (SDA) framework for safe, sus­
ables trusted autonomy in iDSS operations, providing a more responsive tainable and unrestricted LEO operations.
and resilient approach to Space Mission Management (SMM) data col­ iDSS characteristics offer a lot of potential for SBSS missions as the
lecting and processing, particularly when using cutting-edge optical number of observations and RSO potentially tracked make the downlink
sensors. This study investigates the potential uses of iDSS by proposing a of data and a ground-based mission planning highly impractical. Lagona
network of LEO satellites for near-real-time wildfire management. Sat­ et al. [299] proposed AI algorithms for autonomous planning and
ellite missions must have sizable coverage, return intervals and recon­ reconfiguration onboard a SBSS DSS. For on-board trajectory genera­
figuration capabilities to continually monitor Areas of Interest (AOI) in a tion, the proposed approach provided a viable autonomous manoeuvre
dynamically changing environment, which iDSS can provide. Our recent planning solution employing PSO requiring common computing power.
works established the viability of AI-based data processing employing Because of the exceptional parallelization, robustness and efficiency
cutting-edge on-board astrionics hardware accelerators [292]. Based on features of PSO-based methodologies, high accuracy results were ob­
these first findings, AI-based software for wildfire detection on iDSS tained within acceptable calculation timeframes for the desired SBSS
satellites has been developed in stages. To show the applicability of the mission. The suggested trajectory optimization methodology not only
proposed iDSS architecture, simulation case studies in various enables timely and responsive manoeuvring to support next-generation
geographic regions have been carried out and substantiate the viability DSS missions, but it also provides a rationale for ground control segment
of this mission concept [290,291,293–295]. progression towards trusted autonomous space operations. Despite these
obvious advantages, metaheuristic solutions are stochastic in nature,
6.3. Maritime ISR posing a barrier to the interpretability and explainability of the
computed manoeuvre from the standpoint of the human operator. The
In the twenty-first century, space based EO systems have undergone long-term goal of this research is to achieve a flexible and scalable RSO
continuous development. Their application across the world’s water­ detection, SDA and collision avoidance capability. A strategy to extend
ways, among others, has risen significantly with the help of space-based the concept to cooperative tracking was proposed aimed at accurately
Maritime Domain Awareness (MDA), particularly Automatic Identifi­ measuring and estimating the kinematic states of RSO [300–303]. These
cation Systems (AIS). This study investigates the possible utility of studies examined an intruder (debris) being monitored by multiple
Synthetic Aperture Radar (SAR) and DSS for MDA operations. A resilient space-based assets in this circumstance. For collision avoidance and
multi-baseline Along-Track Interferometric Synthetic Aperture Radar separation assurance, the effects of manoeuvring the satellite to an
(AT-InSAR) formation flying concept is developed to efficiently combine appropriate level can be achieved by either performing an orbit raising
many along-track baseline observations for single-pass interferometry. or lowering operations [299]. The ultimate goal is to introduce a Space
The simulation findings show that it is possible to implement this Traffic Management (STM) framework that complements conventional
acquisition mode with autonomous orbit control utilising low-thrust Air Traffic Management (ATM) in ensuring safe and efficient near-Earth
actuation appropriate for electric propulsion. A constellation of these space operations, including rocket launch/re-entry, satellite orbital
formation concept is also proposed to improve repeatability and incor­ deployment and manoeuvring, as well as new forms of point-to-point
porate the benefits of the DSS. An MDA application is a hypothetical suborbital and orbital transportation [304,305].
mission that will be solved using this integrated technique [296,297].
7. Safety and security considerations
6.4. Mission management
Since space assets provide their operators a significant strategic and
Consistently with the IHMM concept, to facilitate trusted autono­ technological advantage; an adversary will almost certainly aim to
mous space mission management, the operational tasks to be performed degrade, deny, or disrupt access to space system capabilities [306].
and their schedule shall be defined and updated using suitable AI Fig. 28 depicts a list of the most common hazards/threats, along with the
planning techniques that consider data collected throughout the mission severity of each. Although cyberattacks are reversible to some extent in
(telemetry payload and external data). The satellites become aware of outer space, the TASO in space requires cyber resilience against cyber
their environment, their health and the Mission Control Centre’s even­ threats. Cybersecurity must be considered early in the design process,
tual directives. During a mission, AI provides trusted autonomous de­ particularly in the context of TASO in iDSS. In the event of planetary
cisions customised to the demands of end-users. The true, long-term threats such as debris or satellites, for example, autonomous orbit

25
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

replanning is required. management approach for different industries. It establishes a common

terminology and methodology that organisations can use in line with
their available resources and operational needs. Five functions make up
7.1. Cyber-space
the Cybersecurity Framework: identify, protect, detect, respond and
recover. To emphasise to users that cybersecurity is an active process
Cyber-space is the notional environment in which computer net­
that enables a company to navigate the constantly evolving spectrum of
works interact. The Fourth Industrial Revolution of the twenty-first
cybersecurity threats, the functions are represented in a circular pattern
century, propelled by the rapid and disruptive rapid rise of DL and
[275,317].
ramped up by the synergy of cyberspace and AI, has assumed critical
importance in modern defence and security [307–310]. Cyberspace is a
7.1.1. Inter-vehicle cybersecurity
complex hierarchical structure of interconnected technological and se­
The ability of the satellite vehicle to protect itself from cyberthreats,
mantic layers (physical, logical, information and human). Humans use
recognise threat actions, respond to cyberattacks, and, if required,
increasing amounts of information in their daily lives, thereby making it
recover is referred to as inter-vehicle cybersecurity. These capabilities
one of the most important resources nowadays. Satellites have histori­
have to be included in the cybersecurity development cycle and applied
cally benefited from “security by obscurity,” which prevents all but the
early in the life cycle of the iDSS system. Small commercial satellite
most sophisticated cyber-attackers due to the system’s complexity and
owners and operators are frequently responsible for inter-vehicle
expensive equipment costs. Due to the widespread use of COTS com­
cybersecurity, while the majority of the remaining infrastructure is
ponents and constellations with thousands of identical satellites, oper­
contracted to certain other suppliers and providers [275].
ational complexity and diversity are unlikely to provide long-term
security [311]. The danger to satellites is generally acknowledged and
obvious. Space systems play a significant role in military C4ISR (Com­ 7.2. Satellite resilience
mand, Control, Communications, Computer, Intelligence, Surveillance
and Reconnaissance) capabilities [312]. In theory, competitors with The aerospace industry of an increasingly multi-polar world will
strong motivations to “level the playing field” with big powers, are have to negotiate a future that is getting more and more crowded with
increasingly motivated to target satellites [313]. Public participation is familiar, unfamiliar and unfriendly players all vying for technological
dependent on space services, whether it be positioning information for and strategic superiority. Threats to the space realm and its sustaining
everyday logistics and transportation or meteorological systems that infrastructure have escalated as a result of this change [318]. Ground
protect millions from disasters. As a “single point of failure” in vital and aerospace system architectures must offer a high level of resilience
infrastructures, satellites may appeal to individuals who are looking to in order to ensure mission success. Resilience is therefore a crucial
upend civilisation [314]. Cyber-threats exist in cyberspace, which must design factor that should be traded off against cost and capabilities when
be taken into account and handled during all design, implementation making decisions [319]. States now recognise their reliance on
and operational phases. Extra caution and system resiliency should be space-based services due to their growing significance for military ac­
employed to reduce cyber dangers by implementing intelligent agents tivities. For these reasons, the majority of governments consider any
for TASO [18,315,316]. The framework for cybersecurity is shown in operation to have space support. However, few individuals legitimately
Fig. 29. comprehend its contribution, let alone the effects of any space service
The framework is based on a customizable cybersecurity risk suspension or significant degradation [306,307,319–321]. A taxonomy

Fig. 28. Outer space threats. Adapted from Ref. [306].

26
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

(especially in DSS), although a number of technical and legal challenges

arise with the adoption of these techniques, both during the develop­
ment and operational phases of a space system’s lifecycle [325–331].

8.1. Technical challenges

The iDSS concept introduces strict requirement on AI for trusted

autonomy. The space vehicle and its components must be able to react
quickly, adapt to changing situations and coordinate their operation
with a potentially large number of other spacecrafts. A number of AI
techniques are presently available to suit these needs, but several of
them need further development and verification to attain full techno­
logical maturity and operational reliability. For instance, on-board AI is
allowing shared mission planning operations between the control
segment and the satellites. To implement sophisticate forms of data
mining on the telemetry and other sensor data, several low-level com­
mands must be aggregated as a single event. The human operator should
take on the role of supervisor, which requires a transition away from the
centralized command and control paradigm of legacy space systems and
an increasing adoption of AI reasoning. A critical challenge in this
process is the transition from deterministic to nondeterministic engi­
Fig. 29. Cybersecurity framework [275]. neering design, but if distributed missions are the future, this evolution
must be acknowledged and embraced. Of course, evolving the paradigm
developed by The Aerospace Corporation (Fig. 30) looks at a system’s from the present rigid and inflexible safety assurance process to a more
capacity to fulfil mission objectives throughout the course of its full responsive and probabilistic approach would take time and developing a
lifecycle. Once the mission’s requirements and specific functions have roadmap to make this process practicable will be one of the key mile­
been determined, the possible threats that can affect the mission should stones to advance in this direction [279]. Although several experts are
be identified, along with potential strategies. The threat’s goals must be concerned about the vulnerability or failure of AI algorithms on-board
also identified because the strategy is a tool for attaining them [322]. spacecraft, research has shown that they are critical for attaining
Several AI approaches should be examined, along with their possibilities mission goals in cybersecurity and satellite health monitoring. Despite
for practical adoption in the iDSS considering human supervision, with the progress made, some difficult difficulties remain, requiring scientific
the aim of achieving TASO. breakthroughs. To date, the majority of progress has been accomplished
in what is known as “narrow AI,” which involves developing ML algo­
rithms to fulfil specific and well-defined functions, such as natural lan­
7.3. Safety considerations guage processing. More difficult problems are left to be dealt with by
“artificial general intelligence,” in which sophisticated AI approaches
Dependence on AI systems is critical for technological frameworks to are developed to reason and solve problems in the same way that
be widely applicable and accepted. Standards, norms, accreditations and humans can. Despite the introduction of new approaches to address
certificates are examples of legal regulations commonly used to them, AI still faces a number of practical obstacles. The labelling of the
formalise societal tools for expressing reliance [323]. Freed et al. [324] training data required for supervised learning frequently necessitates a
offer a Verification and Validation (V&V) methodology for autonomous significant amount of human labour. Data sets that are large and thor­
space systems that aims to increase trust in the stability of complicated ough enough to train a model are likewise challenging to come by. The
software. This approach combines runtime analyses and model control “explainability” difficulty arises from the intricacy of deep learning
using software design architectures to enable traceable modular verifi­ techniques’ “black box” nature, although some success was achieved
cation activities and automated code generation while delivering auto­ with post-hoc explanation methods revealing what factors contribute to
matic formal V&V. Freed’s intelligent automation system guarantees a judgement or prediction and in what manner. Another obstacle is
that software is conceived, produced and verified by domain developing generalized learning strategies, as AI methods continue to
experts-engineers and scientists for space activities, which is another have difficulty transmitting their experiences from one set to the next. A
important part of creating confidence in autonomous software [324]. promising approach to addressing this challenge is transfer learning, in
Generally, the V&V system engineering approach is used for space which an AI model is trained on a given task and then applied to a
mission development. In order to verify and validate the system, it must similar but separate activity [327]. One of the most significant technical
undergo a certain certification process based on their criticality level. challenges is data security; as ML training relies on large amounts of
Table 11 shows the critical levels of these ML models. classified data, which is frequently sensitive and of a personal nature. As
Well-established certification processes for traditional deterministic a result, it is subject to data theft and identity theft [332]. Another
software applications cannot be applied in a straightforward manner in critical difficulty is data scarcity, as the accuracy and relevance of the
the context of ML, therefore major certification topics must be recon­ supervised and labelled datasets used for training and learning directly
sidered. A comprehensive certification approach for ML applications affect the ML application’s power and capabilities. Unfortunately,
would help to improve the overall quality and safe use of this technol­ labelled datasets are scarce, so attempts are underway to develop ap­
ogy, which would in turn increase public acceptance and trust. Winter proaches for making ML models learn despite the paucity of high-quality
et al. [323] proposed a certification standard for supervised learning labelled data using transfer learning, active learning, DL and unsuper­
with low-risk potential, illustrated in Fig. 31. vised learning [333]. Because several of these AI methods are based on
training datasets, bias is a very significant issue. The correctness of AI
8. Technical, ethical and legal challenges decision-making capacity is determined by how well data was trained
utilising real and unbiased facts. When data used for training is laced
A wide range of opportunities exists for the application of AI to space with racial, gender, community, or ethnic biases, unethical and unfair
systems. As discussed, AI can play a key role in achieving TASO effects are inherently present. As many AI systems continue to be trained

27
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Fig. 30. Overview of the resiliency taxonomy [321].

with limited data, such biases are expected to become increasingly se­ neural network models, according to research, can be compacted
vere [326,334–336]. Table 12 lists the some of the foreseeable techno­ without significant accuracy loss [340].
logical challenges associated to different application areas. 4. Another reason that may have slowed AI adoption for on-board ap­
As of now, AI in space is limited to offline data processing and has not plications is a lack of trust due to the inherent unpredictability. This
been adopted “on the edge” within the spacecraft [264]. When looking nondeterminism arises from the impossibility of de facto testing the
particularly at the most promising AI algorithms for spaceborne appli­ weights set due to training for all possible inputs using a finite
cations, several challenges are identified: amount of data. More predictable AI approaches are generally
preferred because the high cost of failure is a critical factor in space
1. There are several reasons for this, including the difficulty of porting applications [341].
DL networks to hardware that predates the algorithms themselves 5. AI applications could be limited to the payload level to reduce risks
and has insufficient performance to make even basic inferences. in EO missions, where object detection and classification could be
Many AI models and particularly DL require a large number of op­ performed locally on sensor data [342]. In this particular instance, AI
erations per second to meet the stringent real-time operational re­ failures would only affect the quality of data for a single payload
quirements of many on-board applications, making their inference rather than putting the satellite at risk [264].
computationally intensive [337]. 6. The final point to consider is Deep Neural Network (DNN) training.
2. In addition, quantization and pruning techniques can compress the Indeed, a significant issue is the lack of datasets for training and
model [338], potentially improving its accuracy over the original model evaluation, particularly for missions involving new equip­
[339]. Different hardware can be exploited depending on the arith­ ment, such as novel sensors, for which no DNN training dataset ex­
metic representation used. ists. DNN training will also be performed on-ground, utilising cloud-
3. By performing model compression, network selection and design based Graphics Processing Units (GPU) or more specialised training
strategies can be used to mitigate memory budget problems. Certain hardware such as Tensor Processing Unit (TPU) due to its

28
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

Table 11 Table 12
ML criticality levels for certification [323]. Technological challenges for space systems in the upcoming decades. Adapted
Criticality Impact Potential (Examples) ML Application Requirements
from Ref. [74].
Level (CL) Areas Goals AI Opportunities and
Challenges
1 There is no risk of harm to living Basic minimum requirements of
beings, no risk of loss of a competently developed ML Sensing & To provide situational •Algorithms for 3D
confidential data and no ethical application are fulfilled. perception awareness for autonomous perception, state estimation
or privacy concerns. space agents, explorers and and data fusion
2 Living beings could be harmed The ML application is assistants •On-board data processing
with limited, no permanent developed according to •Object, event or activity
damage. Temporarily industry standards and follows recognition
unavailability of non-critical best practices that are regarded Mobility To reach and operate at sites •Multi-environment and
data and services, violation of as state of the art. of scientific interest on multi-modal locomotion (e.
ethical concerns without extra-terrestrial surfaces or g., flying, walking, climbing,
identifiable harm to actual free space environments rappelling, tunnelling, swim­
persons. ming, sailing)
3 Living beings could die or be The ML application is •Intelligent manipulation
restricted for life; the developed and documented systems performing
environment could be damaged. with great care. Safety & intentional changes to the
Manipulation of data with Security is ensured with environment (e.g., placing,
severe financial consequences processes and techniques that assembling, digging,
and loss of control of the system go beyond traditional best trenching, drilling, sampling,
to malicious attackers. practices and industry grappling and berthing).
standards. High-level To provide robust and safe •High-integrity GNC
4 Many living beings could die or The ML application is autonomy for autonomous navigation, algorithms
could be restricted for life; the developed and documented system and sub- rendezvous and docking •Intelligent docking and
environment could be damaged with great care. Safety & systems capabilities and enable capture algorithms
permanently. Loss of Security is ensured with extended-duration •Autonomous mission
information which endangers processes and techniques that operations without human planning, scheduling and
the existence of the go beyond traditional best interventions control
organization. Long-term practices and industry • Multi-agent coordination
unavailability of critical data or standards. All components of and cooperative control
services without which the the ML application are formally • Automated data analytics
organization cannot function. secured and validated. for autonomous decision
making
Human-machine To enable humans to • Multi-modal interaction
interactions understand the machine’s (virtual and augmented
state accurately and rapidly reality)
in collaboration and act • Remote and supervisory
effectively and efficiently control
towards the goal state • Distributed collaboration
and coordination
• Adaptive human-machine
Interfaces
• Cognitive human-
autonomy interactions
System To provide a framework for • DDT&E lifecycle
engineering understanding and evolutions
coordinating the complex • Verification and validation
interactions of space CPS of complex/adaptive AI
and to achieve the desired based systems
mission requirements • Digital threads and digital
twins
• Safety certification, trust
and cyber-security

questions. Some researchers have identified the need for AI ethicists to

help navigate where advances in this technology could lead [328]. This
is more exciting because AI and space technologies offer a wide range of
opportunities. Nevertheless, the need to understand the ultimate
outcome of the technology remains unanswered. Not to mention that
Fig. 31. Ml certification Workflow [323]. even the scientific research community are unable to agree on a prece­
dent arising from the use of AI. Prominent scientists and industry leaders
complexity. These factors raise serious concerns about the usability argued that AI could radically transform the way we live and work,
of models developed before the launch of satellites and not based on potentially threating our civilisation and even human survival [343]. A
the original satellite data. However, this problem is mitigated by the report on robotics and AI published by the British House of Commons
ability to reconfigure models during the life of missions, which is highlighted specific ethical and legal issues, including transparent
enabled by the use of modern Commercial, Off-the-Shelf (COTS) and decision-making, minimising bias, accountability and privacy [325].
the smaller file sizes required for programming, which is becoming The first draft of the “Ethics Guidelines for Trustworthy AI” was pub­
compatible with the uplink bandwidth of small satellites [264]. lished by the European Commission’s High-Level Expert Group on
Artificial Intelligence (‘AI HLEG’) [344]. According to the guidelines,
8.2. Ethical and legal challenges trustworthy AI must adhere to the following principles:

The use of AI in space systems raises a number of ethical and legal

29
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

1. Ethical purpose: AI development, deployment and use should respect 9. Conclusion and future research
fundamental rights and applicable regulations as well as core prin­
ciples and values to ensure “ethical purpose"; Large and heterogeneous satellite constellations are becoming viable
2. Technical robustness: AI should be technically robust and reliable and advantageous options, eliciting an evolution of operational para­
since its use can cause unintentional harm, even in the presence of digms towards higher levels of autonomation and trusted autonomy.
good intentions [344]. Despite the fact that achieving full autonomy is still a relatively distant
goal, the demand for Artificial Intelligence (AI) is on the rise, creating
AI systems use large amounts of data, causing increasing concerns as new opportunities for expanded space system capabilities. While
more data is collected and used. Such high volumes and the level of increasingly higher levels of automation have been introduced in con­
dependence on such data will keep privacy at the forefront as one of the ventional satellite systems, AI-based trusted autonomy is becoming
most significant legal issues to be addressed in the future. For instance, essential in large Distributed Space Systems (DSS). Intelligent DSS
setting ethical parameters within which AI systems operate is para­ (iDSS) are thus an evolving paradigm, which relies on new reactive and
mount in tackling bias, considering the application of AI to data predictive functionalities, essential to accomplish Trusted Autonomous
generated in space and prospective on-board AI space sector Satellite Operations (TASO). The adoption of safe, efficient and reliable
developments. AI techniques in all DSS segments unleashes unique opportunities to
Another important issue is liability in closed-loop human-machine enhance the overall space mission lifecycle, and particularly the design,
systems: if AI is used in a Decision Support Tool (DST) and an accident development and operational phases. To support the transition to TASO,
takes place, the critical question being raised is who is responsible, the a new dedicated AI design, verification and certification framework has
operator or the designer of the AI-based DST? Based on the current legal to be established, addressing both autonomous cyber-physical systems
framework, if the human operator did not follow the DST recommen­ and closed-loop human-machine system architectures. The certification
dation and the DST was right, s/he will be blamed for having made the of truly intelligent and evolving space systems is a critical undertaking
wrong decision. If the operator, conversely, followed the DST and the and several aspects will have to be properly addressed to ensure the
DST was wrong, s/he will be blamed for having made the wrong deci­ safety and security of these systems, including: (1) integrity of both
sion. This is an obvious paradox, which will have to be properly system and data being exchanged; (2) consistency and resilience against
addressed. Since we are dealing with complex CPH systems employing adversarial threats; and (3) interpretability of the autonomous decisions.
AI (i.e., interconnected and closed-loop human-machine systems), there Based on our review, it is evident that to safely integrate AI into all
is general consensus that the liability issues must be addressed primarily aspects of space operations, further research is needed in several areas.
at the design and certification stage. This is because AI explainability First of all, AI-based functionalities are needed for: (1) Mission Planning
and trusted autonomy are design attributes of such systems and not and Scheduling (MPS); (2) on-board data collection and processing; (3)
measurable/observable parameters during operations. Clearly, this is a reconfigurable networking for optimized data exchange with other
major paradigm shift and requires radical changes in the current Design, space assets and ground terminals; and (4) autonomous detection,
Development, Test and Evaluation (DDT&E) industrial practices as well tracking and avoidance of orbital collision hazards. All these function­
as the development (and uptake) of new operational standards and alities will require efficient and intelligent data analytics, interpretation
regulations. The tort of negligence, which is a common legal principle and decision-making on-board satellites, both to minimize the data
applied to liability, is concerned with whether a party owed another a throughput requirements and to maximise mission responsiveness.
duty of care, whether that duty was breached and whether the breach Ideally, the associated AI techniques should have the ability to contin­
resulted in damages. In the case of negligence, reasonable foreseeability uously enhance the iDSS performance through learning and adaptation.
is a crucial concept [345]. As AI systems become less reliant on tradi­ In terms of on-board Mission Planning and Scheduling (MPS), the AI
tional algorithms, they will be able to exhibit behaviours that are not techniques shall consider operational uncertainties and constraints, to
only unexpected by their creators and possibly unpredictable. In a sce­ optimise objectives, resource allocation and trajectory planning/
nario where there is a lack of predictability, the law could replace its replanning. In relation to this, research in autonomous detection and
negligence-based approach with one based a more holistic liability avoidance of collision hazards has to explore the effectiveness of
concept, which clearly specifies the allocation of responsibility to both incorporating AI models into the Space Based Space Surveillance (SBSS)
AI/DST designers and operators. If, on the other hand, a for Space Domain Awareness (SDA). Particular emphasis should be
negligence-based approach is retained, it will be necessary to define the given to algorithms that can improve the performance of tracking debris
applicable duty of care requirements in a hybrid AI-human operational and other objects. This will, in turn, support the co-evolution of air and
context. Whichever the approach adopted, the regulation of AI will space traffic management (e.g., launch and re-entry provisions and
inevitably continue to pose significant challenges, which are exacer­ point-to-point suborbital operations). Concerning the resilience of AI-
bated by the inherent continuous evolution of AI algo­ based Cyber-Physical Systems (CPS) in space, the challenge is to
rithms/applications and the associated need to develop a more flexible develop robust, fault-tolerant and secure systems that ensure TASO
and responsive legal framework to keep pace with these technologies given the specific hardware constraints, the variability of mission re­
[344,345]. There are currently two main approaches to address quirements, the uncertainties in physical processes and the possibility of
lawmaking for AI [329]. The first approach stipulates that one must both cyber/physical attacks and human errors. In particular, the
solve problems only as they arise (i.e., responsive in nature). There is no development of Intelligent Health and Mission Management (IHMM)
need to introduce new regulation when there is no problem. The main functions on-board space assets, which exploit AI to provide predictive
rationale behind this approach is that early (normative) intervention can integrity which can track the state of health and operational efficiency of
prevent or even block certain innovation paths. The second approach is space systems, detect anomalies and implement timely reconfiguration
proactive and based on the premises that universal pre-emptive stan­ actions which minimize outages and risks of catastrophic failures. These
dards are necessary, without waiting for actual problems to occur. functions form part of the autonomous mission MPS capability. To
Whichever approach is pursued, it is important to note that the Outer achieve the full potential of iDSS for Earth observation and astronomy,
Space Treaty requires international monitoring of the application of the further research is needed into efficient AI algorithms capable of ana­
rules and the establishment of an international law enforcement agency lysing and interpreting large volumes of mission data in real-time,
to carry out such monitoring [346]. Differences in national legal prompting human review only when and where needed, therefore
frameworks and in the uptake of international regulations/standards reducing the time and resources necessary in these missions. Another
may challenge this process [347]. important aspect concerns the ethics and trustworthiness of AI-based
CPS. In this respect, research should investigate the legal

30
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

consequences of AI applications in the space sector and identify practical International Federated And Fractionated Satellite Systems Workshop, 2017. Rome,
Italy.
ways to ensure that AI-based decision-making systems are transparent,
[17] E.J.M. Angelita, C. Kelly, The A-train: NASA’s earth observing system (EOS)
fair and accountable by design. Explainability and interpretability will satellites and other earth observation satellites, in: 4th IAA Symposium on Small
be critical aspects of this endeavour and shall be explored in relation to Satellites for Earth Observation IAA-B4–1507P, 2003.
the duality of human-autonomy interactions (i.e., integrity monitoring [18] R. Sabatini, Aerospace Cyber-Physical and Autonomous Systems, IEEE Aerospace
and Electronic Systems Society (AESS) Distinguished Lectureed, 2020. Available
and augmentation of both human and autonomous agents). Once the at: https://www.researchgate.net/publication/341787434_Aerospace_Cyber-Ph
robustness, fault-tolerance, security and ethical aspects are properly ysical_and_Autonomous_Systems.
addressed, a comprehensive AI regulatory framework should be estab­ [19] J. Utzmann, A. Wagner, J. Silha, T. Schildknecht, P. Willemsen, F. Teston, et al.,
Space-based space surveillance and tracking demonstrator: mission and system
lished to support the necessary evolution of systems engineering and design, in: Proceedings of 65th International Astronautical Congress, 2014,
lifecycle management practices (i.e., design, integration, test, certifica­ pp. 1648–1655. Toronto, Canada.
tion and operations), which are still based on purely deterministic and [20] M.K. Ben-Larbi, K. Flores Pozo, T. Haylok, M. Choi, B. Grzesik, A. Haas, et al.,
Towards the automated operations of large distributed satellite systems. Part 1:
relatively inflexible paradigms. review and paradigm shifts, Adv. Space Res. 67 (2021/06/01/2021) 3598–3619.
[21] M.K. Ben-Larbi, K. Flores Pozo, M. Choi, T. Haylok, B. Grzesik, A. Haas, et al.,
Towards the automated operations of large distributed satellite systems. Part 2:
Declaration of competing interest classifications and tools, Adv. Space Res. 67 (2021/06/01/2021) 3620–3637.
[22] D. Wischert, P. Baranwal, S. Bonnart, M. Casanova Álvarez, R. Colpari Carrizo,
The authors declare that they have no known competing financial M. Daryabari, et al., Conceptual design of a mars constellation for global
communication services using small satellites, in: Presented at the 71st
interests or personal relationships that could have appeared to influence
International Astronautical Congress (IAC), United Arab Emirates, 2020. Dubai.
the work reported in this paper. [23] CNES, PRISMA: testing new formation-flying technologies, Available at: http
s://prismaffiord.cnes.fr/en/PRISMA/index.htm, 2016.
Data availability [24] B. Tapley, S. Bettadpur, J.C. Ries, P. Thompson, M. Watkins, GRACE
measurements of mass variability in the earth system, Science 305 (08/01 2004)
503–505.
No data was used for the research described in the article. [25] G. Krieger, A. Moreira, H. Fiedler, I. Hajnsek, M. Werner, M. Younis, et al.,
"TanDEM-X: a satellite formation for high-resolution SAR interferometry",
Geoscience and Remote Sensing, IEEE Transactions on 45 (12/01 2007) 3317–3341.
Acknowledgment [26] R. Langlois, Modularity in technology and organization, J. Econ. Behav. Organ.
49 (2002) 19–37.
The authors would like to thank Khalifa University (United Arab [27] M.E.J. Newman, Modularity and community structure in networks, Proc. Natl.
Acad. Sci. USA 103 (2006) 8577.
Eirates), Northrop Grumman Corporation (United States), and the [28] A. Kharrazi, Resilience, in: B. Fath (Ed.), Encyclopedia of Ecology, second ed.,
SmartSat Cooperative Research Centre (Australia) for their support of Elsevier, Oxford, 2019, pp. 414–418.
this work through the Grant No. FSU-2022-013, the Collaborative [29] K. Ulrich, The role of product architecture in the manufacturing firm, Res. Pol. 24
(1995/05/01/1995) 419–440.
Research Project No. RE-04143, and the Doctoral Research Project No. [30] K. Hölttä-Otto, O. de Weck, Degree of modularity in engineering systems and
2.13s, respectively. The authors would also like to thank Dr Andoh Afful, products with technical and business constraints, Concurr. Eng. 15 (2007/06/01
Dr Suraj Bijjahalli and Prof. Wei Xiang and Mr Thomas Fahey for their 2007) 113–126.
[31] D.A. Gianetto, B. Heydari, Network Modularity is essential for evolution of
insightful feedback, which helped to improve the quality of this article. cooperation under uncertainty, Sci. Rep. 5 (2015/04/07 2015) 9340.
[32] B. Heydari, M. Mosleh, K. Dalili, From modular to distributed open architectures:
References a unified decision framework, Syst. Eng. 19 (2016) 252–266.
[33] M. Mosleh, K. Dalili, B. Heydari, Distributed or monolithic? A computational
architecture decision framework, IEEE Syst. J. PP (2016), 08/02.
[1] R.M. Millan, R. von Steiger, M. Ariel, S. Bartalev, M. Borgeaud, S. Campagnola, et
[34] G. Weiss, Multiagent Systems: A Modern Approach to Distributed Artificial
al., Small satellites for space science: a COSPAR scientific roadmap, Adv. Space
Intelligence, MIT Press, 1999.
Res. 64 (2019/10/15/2019) 1466–1517.
[35] C. Araguz, E. Bou-Balust, E. Alarcón, Applying autonomy to distributed satellite
[2] K. Böhringer, M. Campbell, J. Vagners, Intelligent satellite teams space systems,
systems: trends, challenges, and future prospects, Syst. Eng. 21 (2018) 401–416.
in: Proceedings of AAAI Spring Symposium, 1999.
[36] C. Araguz, A. Alvaro, I.d. Portillo, K. Root, E. Alarcón, E. Bou-Balust, On
[3] K. Ranasinghe, R. Sabatini, A. Gardi, S. Bijjahalli, R. Kapoor, T. Fahey, et al.,
autonomous software architectures for distributed spacecraft: a Local-Global
Advances in integrated system health management for mission-essential and
Policy, in: Proceedings of 2015 IEEE Aerospace Conference, 2015, pp. 1–9.
safety-critical aerospace applications, Prog. Aero. Sci. 128 (2021), 100758, 11/
[37] R. Preston, Distributed satellite constellations offer advantages over monolithic
18.
systems, Available at: https://www.rand.org/pubs/research_briefs/RB92.html,
[4] E. Lagona, S. Hilton, A. Afful, A. Gardi, R. Sabatini, Autonomous trajectory
2004, July.
optimisation for intelligent satellite systems and space traffic management, Acta
[38] M.R. Schoeberl, The afternoon constellation: a formation of Earth observing
Astronaut. 194 (2022) 185–201.
systems for the atmosphere and hydrosphere, in: Proceedings of International
[5] N. Cramer, D. Cellucci, C. Adams, A. Sweet, M. Hejase, J. Frank, et al., Design and
Geoscience And Remote Sensing Symposium, 2002, pp. 354–356. Toronto, Canada.
testing of autonomous distributed space systems, in: Presented at the Small
[39] I. del Portillo, B. Cameron, E. Crawley, A technical comparison of three low earth
Satellite Conference, 2021.
orbit satellite constellation systems to provide global broadband, Acta Astronaut.
[6] S. Hilton, A. Gardi, S. Desai, A. Gardi, R. Sabatini, N. Ezer, et al., Human-Machine
159 (03/01 2019) 123–135.
System Design for Autonomous Distributed Satellite Operations, 2020.
[40] J.G. Walker, Satellite constellations, J. Br. Interplanet. Soc. (JBIS) 37 (1984) 559.
[7] M. Vasile, E. Minisci, K. Tang, Computational intelligence in aerospace science
[41] J. Guo, D. Maessen, E. Gill, Fractionated spacecraft: the new sprout in distributed
and engineering [guest editorial], IEEE Comput. Intell. Mag. 12 (2017) 12–13.
space systems, in: Presented at the 60th International Astronautical Congress,
[8] H. Klinkrad, Space Debris: Models and Risk Analysis, Springer Science & Business
2009. Daejeon, Republic of Korea.
Media, 2006.
[42] O. Brown, P. Eremenko, The value proposition for fractionated space
[9] J.N. Pelton, New Solutions for the Space Debris Problem, Springer, 2015.
architectures, Sciences 4 (09/01 2006) 23.
[10] S. Ramasamy, R. Sabatini, A unified approach to cooperative and non-cooperative
[43] O. Brown, P. Eremenko, Fractionated Space Architectures: A Vision for
sense-and-avoid, in: Proceedings of 2015 International Conference On Unmanned
Responsive Space, 2006.
Aircraft Systems, ICUAS 2015, Denver, CO, USA, 2015, pp. 765–773.
[44] J. Lafleur, J. Saleh, GT-FAST: a point design tool for rapid fractionated spacecraft
[11] G. Dubos, J. Saleh, Comparative cost and utility analysis of monolith and
sizing and synthesis, in: AIAA SPACE 2009 Conference & Exposition, American
fractionated spacecraft using failure and replacement Markov models, Acta
Institute of Aeronautics and Astronautics, 2009.
Astronautica - Acta Astronautica 68 (02/28 2011) 172–184.
[45] I.d. Portillo, E. Bou, E. Alarcón, M. Sanchez-Net, D. Selva, Á. Á, On scalability of
[12] R. Preston, Distributed Satellite Constellations Offer Advantages over Monolithic
fractionated satellite network architectures, in: Proceedings of 2015 IEEE
Systems, 2004.
Aerospace Conference, 2015, pp. 1–13.
[13] G.F. Dubos, J.H. Saleh, Comparative cost and utility analysis of monolith and
[46] J. Wang, S. Nakasuka, Cluster flight orbit design method for fractionated
fractionated spacecraft using failure and replacement Markov models, Acta
spacecraft, Aircraft Eng. Aero. Technol. 84 (2012) 330–343.
Astronaut. 68 (2011) 172–184.
[47] W. Yao, X. Chen, Y. Zhao, M. van Tooren, A fractionated spacecraft system
[14] M. Mosleh, K. Dalili, B. Heydari, Distributed or monolithic? A computational
assessment tool based on lifecycle simulation under uncertainty, Chin. J.
architecture decision framework, IEEE Syst. J. 12 (2018) 125–136.
Aeronaut. 25 (2012/02/01/2012) 71–82.
[15] S. Nag, Design and Evaluation of Distributed Spacecraft Missions for Multi-
[48] A. Golkar, I. Lluch i Cruz, The Federated Satellite Systems paradigm: concept and
Angular Earth Observation, Massachusetts Institute of Technology, 2015.
business case evaluation, Acta Astronaut. 111 (2015) 230–248.
[16] A. Poghosyan, I. Lluch, H. Matevosyan, A. Lamb, C. Moreno, C. Taylor, et al.,
Unified classification for distributed satellite systems, in: Proceedings of 4th

31
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

[49] I. Lluch, A. Golkar, Design implications for missions participating in federated [80] Y. Lim, S. Ramasamy, A. Gardi, T. Kistan, R. Sabatini, Cognitive human-machine
satellite systems, J. Spacecraft Rockets 52 (2015) 1361–1374. Interfaces and interactions for unmanned aircraft, J. Intell. Rob. Syst. (2017)
[50] J.A. Ruiz-de-Azua, L. Fernandez, J.F. Muñoz, M. Badia, R. Castella, C. Diez, et al., 10–13.
Proof-of-Concept of a federated satellite system between two 6-unit CubeSats for [81] S. Bijjahalli, R. Sabatini, A. Gardi, Advances in intelligent and autonomous
distributed earth observation satellite systems, in: Proceedings of IEEE navigation systems for small UAS, Prog. Aero. Sci. 115 (2020), 100617.
International Geoscience And Remote Sensing Symposium 2019, 2019, [82] N. Pongsakornsathien, Y. Lim, A. Gardi, S. Hilton, L. Planke, R. Sabatini, et al.,
pp. 8871–8874. Yokohama, Japan. Sensor networks for aerospace human-machine systems, Sensors 19 (08/08 2019)
[51] R. Akhtyamov, R. Vingerhoeds, A. Golkar, Identifying retrofitting opportunities 3465.
for federated satellite systems, J. Spacecraft Rockets 56 (2019) 620–629, 05/01. [83] Y. Lim, A. Gardi, R. Sabatini, S. Ramasamy, R. Bolia, Avionics human-machine
[52] J. An, X. Li, Z. Zhang, G. Zhang, W. Man, G. Hu, et al., Path planning for self- Interfaces and interactions for manned and unmanned aircraft, Prog. Aero. Sci.
collision avoidance of space modular self-reconfigurable satellites, Aerospace 9 102 (2018) 1–46.
(2022) 141. [84] Y. Lim, V. Bassien-Capsa, S. Ramasamy, J. Liu, R. Sabatini, Commercial airline
[53] H. Bloom, D.S. Wilson, The lucifer principle: a scientific expedition into the forces single-pilot operations: system design and pathways to certification, IEEE Aero.
of history, Foreign Aff. 74 (1995), 01/01. Electron. Syst. Mag. 32 (09/19 2017) 4–21.
[54] F. Heylighen, Stigmergy as a universal coordination mechanism I: definition and [85] Y. Lim, N. Pongsakornsathien, A. Gardi, R. Sabatini, T. Kistan, N. Ezer, et al.,
components, Cognit. Syst. Res. 38 (2016) 4–13, 2016/06/01. Adaptive human-robot interactions for multiple unmanned aerial vehicles,
[55] S. Cannon, J. Daymude, D. Randall, A. Richa, A Markov chain algorithm for Robotics 10 (2021) 1–30.
compression in self-organizing particle systems, in: Proceedings of ACM [86] S. Chien, R. Morris, Space applications of artificial intelligence, AI Mag. 35 (12/
Symposium On Principles Of Distributed Computing, 2016, pp. 279–288. Chicago, 01 2014) 3–6.
Illinois, USA. [87] S. Russell, P. Norvig, Artificial Intelligence: A Modern Approach, Prentice Hall
[56] S. Engelen, E. Gill, C. Verhoeven, Systems engineering challenges for satellite Press, 2009.
swarms, in: Proceedings of Aerospace Conference, Ystems Engineering Challenges [88] S. Voß, MetaheuristicsMetaheuristics, in: C.A. Floudas, P.M. Pardalos (Eds.),
for Satellite Swarms, 2011, pp. 1–8. Encyclopedia of Optimization, Springer US, Boston, MA, 2009, pp. 2061–2075.
[57] Z. Chen, Y. Zeng, A swarm intelligence networking framework for small satellite [89] S. Voß, Meta-heuristics: the state of the art, in: Proceedings of Workshop On Local
systems, Commun. Netw. 5 (01/01 2013) 171–175. Search For Planning And Scheduling, 2000, pp. 1–23.
[58] N. Olsen, E. Friis-Christensen, R. Floberghagen, P. Alken, C.D. Beggan, [90] X.-S. Yang, Engineering Optimization: an Introduction with Metaheuristic
A. Chulliat, et al., The swarm satellite constellation application and research Applications, 2010.
facility (SCARF) and swarm data products, Earth Planets Space 65 (2013/11/22 [91] I. Boussaïd, J. Lepagnot, P. Siarry, A Survey on Optimization Metaheuristics, vol.
2013) 1. 237, Information Sciences, 2013/07/10/2013, pp. 82–117.
[59] S. Mathavaraj, R. Padhi, Satellite Formation Flying: High Precision Guidance [92] S. Almufti, Historical survey on metaheuristics algorithms 7 (2019) 1–12, 11/17.
Using Optimal and Adaptive Control Techniques, Springer Singapore Pte. [93] S. Harifi, J. Mohammadzadeh, M. Khalilian, S. Ebrahimnejad, Giza Pyramids
Limited, Singapore, 2021. Construction: an Ancient-Inspired Metaheuristic Algorithm for Optimization,
[60] D. Wang, B. Wu, E.K. Poh, S, in: Atellite Formation Flying : Relative Dynamics, Evolutionary Intelligence, 2020/07/13 2020.
Formation Design, Fuel Optimal Maneuvers and Formation Maintenance, Intelligent [94] G. Dhiman, A. Kaur, Optimizing the design of airfoil and optical buffer problems
Systems, Control and Automation: Science and Engineering, vol. 87, Springer using spotted hyena optimizer, Design 2 (08/01 2018) 28.
Singapore, Singapore, 2017. [95] A. Hegazy, M. Makhlouf, G. Eltaweel, Dimensionality reduction using an
[61] D. Selva, A. Golkar, O. Korobova, I. Cruz, P. Collopy, O. de Weck, Distributed improved whale optimization algorithm for data classification, Int. J. Mod. Educ.
earth satellite systems: what is needed to move forward? J. Aero. Inf. Syst. 14 Comput. Sci. 10 (07/08 2018) 37–49.
(08/08 2017) 1–26. [96] X.-S. Yang, Firefly algorithms for multimodal optimization, in: Proceedings, 2009,
[62] J.L. Moigne, J.C. Adams, S. Nag, A new taxonomy for distributed spacecraft pp. 169–178. Berlin, Heidelberg.
missions, IEEE J. Sel. Top. Appl. Earth Obs. Rem. Sens. 13 (2020) 872–883. [97] D. Spiller, Optimal Control Problems Solved via Swarm Intelligence, 2018.
[63] J. Kramer, J. Magee, Self-managed systems: an architectural challenge, in: [98] P. Miralles, K. Thangavel, A. Scannapieco, N. Jagadam, P. Baranwal, B. Faldu, et
Proceedings of Future Of Software Engineering, FOSE’07), 2007, pp. 259–268. al., A critical review on the state-of-the-art and future prospects of Machine
[64] C.-Y. Chong, S.P. Kumar, Sensor networks: evolution, opportunities, and Learning for Earth Observation Operations, Adv. Space Res. (02/17 2023).
challenges, Proc. IEEE 91 (2003) 1247–1256. [99] H. Shakhatreh, A.H. Sawalmeh, A. Al-Fuqaha, Z. Dou, E. Almaita, I. Khalil, et al.,
[65] J.-G. Meß, F. Dannemann, F. Greif, Techniques of Artificial Intelligence for Space Unmanned aerial vehicles (UAVs): a survey on civil applications and key research
Applications - A Survey, 2019. challenges, IEEE Access 7 (2019) 48572–48634.
[66] A. Jonsson, R. Morris, L. Pedersen, Autonomy in Space Exploration: Current [100] S. Indolia, A.K. Goswami, S.P. Mishra, P. Asopa, Conceptual understanding of
Capabilities and Future Challenges, 2007. convolutional neural network- A deep learning approach, Procedia Comput. Sci.
[67] N.J. Nilsson, "Shakey the Robot", AI Center, in: SRI International, vol. 323, 132 (2018/01/01/2018) 679–688.
Ravenswood Ave., Menlo Park, CA 94025, 1984, p. 333. [101] Govinda Rao Locharla, Jaya Prakash Allam, Y.V. Narayana, Yellapu Anusha,
[68] E. Vassev, M. Hinchey, Engineering requirements for autonomy features, in: Review of the convolution neural network architectures for deep learning,
M. Wirsing, M. Hölzl, N. Koch, P. Mayer (Eds.), Software Engineering for International Journal of Advanced Science and Technology 29 (2020)
Collective Autonomic Systems: the ASCENS Approach, Springer International 2251–2262, 05/13 2020.
Publishing, Cham, 2015, pp. 379–403. [102] Firdaus, Y. Arkeman, A. Buono, I. Hermadi, Satellite image processing for
[69] R. Proud, J. Hart, R. Mrozinski, Methods for Determining the Level of Autonomy precision agriculture and agroindustry using convolutional neural network and
to Design into a Human Spaceflight Vehicle: A Function Specific Approach, 2003, genetic algorithm, IOP Conf. Ser. Earth Environ. Sci. 54 (01/01 2017), 012102.
p. 15, 09/01. [103] W. Hu, Y. Huang, L. Wei, F. Zhang, H. Li, Deep convolutional neural networks for
[70] F. Kucinskis and M. Ferreira, "Taking the ECSS autonomy concepts one step hyperspectral image classification, J. Sens. (2015), 258619, 2015/07/30 2015.
further", in SpaceOps 2010 Conference, 25 - 30 April 2010, Huntsville, Alabama [104] F. Kucinskis, M. Ferreira, Taking the ECSS autonomy concepts one step further, in:
(USA). Proceedings of SpaceOps 2010 Conference, 2010. Huntsville, AL, USA.
[71] R. Boyce, D. Griffin, Future trusted autonomous space scenarios, in: H.A. Abbass, [105] A. Argan, M. Tavani, A. Trois, The on-board data processing of the AGILE satellite,
J. Scholz, D.J. Reid (Eds.), Foundations of Trusted Autonomy, Springer Rendiconti Lincei. Sci. Fis. Nat. 30 (2019), 10/19.
International Publishing, Cham, 2018, pp. 355–364. [106] P. Buist, B.-J. Vollmuller, On-board Data Processing for Earth and Atmospheric
[72] N. Steiner, P. Athanas, Hardware autonomy and space systems, in: Proceedings of Observations, vol. 12, Transactions Of The Japan Society for Aeronautical and
IEEE Aerospace Conference, Big Sky, MT, USA, 2009, pp. 1–13. Space Sciences, Aerospace Technology Japan, 01/01 2014, pp. Tn_25–Tn_30.
[73] J. Valasek, T.C. Lieuwen, Advances In Computational Intelligence and Autonomy for [107] T. Itoh, H. Oya, H. Yamamoto, Y. Masumoto, T. Ono, Y. Minami, et al., The On-
Aerospace Systems, Computational Intelligence and Autonomy for Aerospace Board Data Processing System for Scientific Satellite - EXOS-C, 09/01 1980.
Systems, American Institute of Aeronautics and Astronautics, Inc., Reston, VA, [108] M. Tragni, C. Abbattista, L. Amoruso, L. Cinquepalmi, F. Bgongiari, W. Errico, On-
2018. board Payload Data Processing from Earth to Space Segment, 2013.
[74] E. Allouis, Y. Gao, A. Kisdi, R. Ward, D. Jones, UK RAS White Papers - Space [109] W. G.G, K. Charalabos, Application of Artificial Intelligence in Land Use Mapping
Robotics & Autonomous Systems, Widening the horizon of space exploration, from Satellite Imagery, 1990.
2016. [110] D. Hemanth, Artificial Intelligence Techniques for Satellite Image Analysis, 2020.
[75] W. Xu, M.J. Dainoff, L. Ge, Z. Gao, From Human-Computer Interaction to Human- [111] G.J.-P. Schumann, L. Giustarini, M. Zare, B. Gaffinet, Call to action: pushing
AI Interaction: New Challenges and Opportunities for Enabling Human-Centered scientific and technological innovation to develop an efficient AI flood mapper for
AI, 2021 arXiv preprint arXiv:2105.05424. operational SAR satellites, in: Proceedings of EGU General Assembly Conference
[76] J. Beer, A. Fisk, W. Rogers, Toward a framework for levels of robot autonomy in Abstracts, 2021. EGU21-5943.
human-robot interaction, Journal of Human-Robot Interaction 3 (06/01 2014) [112] D.A. Zaidenberg, A. Sebastianelli, D. Spiller, B. Le Saux, S.L. Ullo, Advantages and
74. bottlenecks of quantum machine learning for remote sensing, in: Proceedings of
[77] JARUS Methodology for Evaluation of Automation for UAS Operations, Joint 2021 IEEE International Geoscience And Remote Sensing Symposium IGARSS, 2021,
Authorities for Rulemaking of Unmanned Systems, 2023. pp. 5680–5683.
[78] Department of Defense Standard Practice, System safety, in: The International [113] A. Nowakowski, D. Spiller, N. Cremer, R. Bonifacio, M. Marszalek, M. Garcia-
System Safety Society, 2021. MIL-STD-882F. Herranz, et al., AI opportunities and challenges for crop type mapping using
[79] Y. Lim, A. Gardi, R. Sabatini, UAS Human Factors and Human-Machine Interface sentinel-2 and drone data, in: Proceedings of 2021 IEEE International Geoscience
Design, 2020, pp. 23–48. And Remote Sensing Symposium IGARSS, 2021, pp. 258–261.

32
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

[114] A. Nowakowski, J. Mrziglod, D. Spiller, R. Bonifacio, I. Ferrari, P.P. Mathieu, et [145] D. Izzo, C. Sprague, D. Tailor, Machine Learning and Evolutionary Techniques in
al., Crop type mapping by using transfer learning, Int. J. Appl. Earth Obs. Geoinf. Interplanetary Trajectory Design, 02/01 2018.
98 (2021), 102313. [146] M. Agnan, J. Vannitsen, Scaling uncertainties on asteroid characteristics to
[115] D. Spiller, L. Ansalone, F. Carotenuto, P.P. Mathieu, Crop type mapping using prepare datasets for machine learning, Adv. Space Res. 68 (2021/10/15/2021)
prisma hyperspectral images and one-dimensional convolutional neural network, 3225–3232.
in: Proceedings of 2021 IEEE International Geoscience And Remote Sensing [147] B. Gaudet, R. Linares, R. Furfaro, Deep reinforcement learning for six degree-of-
Symposium IGARSS, 2021, pp. 8166–8169. freedom planetary landing, Adv. Space Res. 65 (2020/04/01/2020) 1723–1741.
[116] A. Sebastianelli, D.A. Zaidenberg, D. Spiller, B.L. Saux, S.L. Ullo, On Circuit-Based [148] F. Latorre, D. Spiller, S.T. Sasidharan, S. Basheer, F. Curti, Transfer learning for
Hybrid Quantum Neural Networks for Remote Sensing Imagery Classification, real-time crater detection on asteroids using a Fully Convolutional Neural
2021 arXiv preprint arXiv:2109.09484. Network, Icarus 394 (2023), 115434.
[117] M. Rosso, A. Sebastianelli, D. Spiller, P. Mathieu, S. Ullo, On-board volcanic [149] R. Sherwood, S. Chien, R. Castaño, G. Rabideau, Autonomous Planning and
eruption detection through CNNs and satellite multispectral imagery, Rem. Sens. Scheduling on the TechSat 21 Mission, 2002.
13 (09/02 2021) 3479. [150] A.A. Bataleblu, J. Roshanian, Robust trajectory optimization of space launch
[118] L. Shi, J. Zhang, D. Zhang, T. Igbawua, Y. Liu, Developing a dust storm detection vehicle using computational intelligence, in: Proceedings of 2015 IEEE Congress on
method combining Support Vector Machine and satellite data in typical dust Evolutionary Computation, CEC), 2015, pp. 3418–3425.
regions of Asia, Adv. Space Res. 65 (2020/02/15/2020) 1263–1278. [151] D.S. Kolosa, A Reinforcement Learning Approach to Spacecraft Trajectory
[119] P.V.R. Ferreira, R. Paffenroth, A.M. Wyglinski, T.M. Hackett, S.G. Bilen, R. Optimization, 2019.
C. Reinhart, et al., Reinforcement learning for satellite communications: from LEO [152] K. Parmar, D. Guzzetti, Interactive imitation learning for spacecraft path-planning
to deep space operations, IEEE Commun. Mag. 57 (2019) 70–75. in binary asteroid systems, Adv. Space Res. 68 (2021/08/15/2021) 1928–1951.
[120] P.V.R. Ferreira, R. Paffenroth, A.M. Wyglinski, T.M. Hackett, S.G. Bilén, R. [153] S. Yin, J. Li, L. Cheng, Low-thrust spacecraft trajectory optimization via a DNN-
C. Reinhart, et al., Multi-objective reinforcement learning-based deep neural based method, Adv. Space Res. 66 (2020/10/01/2020) 1635–1646.
networks for cognitive space communications, in: Proceedings of 2017 Cognitive [154] O. Jung, J. Seong, Y. Jung, H. Bang, Recurrent neural network model to predict
Communications For Aerospace Applications Workshop, CCAA), 2017, pp. 1–8. re-entry trajectories of uncontrolled space objects, Adv. Space Res. 68 (2021/09/
[121] T.M. Hackett, S.G. Bilén, P.V.R. Ferreira, A.M. Wyglinski, R.C. Reinhart, D. 15/2021) 2515–2529.
J. Mortensen, Implementation and on-orbit testing results of a space [155] I.P. Rodrigues, P.A.S. Oliveira, A.M. Ambrosio, R.A.J. Chagas, Modeling satellite
communications cognitive engine, IEEE Transactions on Cognitive battery aging for an operational satellite simulator, Adv. Space Res. 67 (2021/03/
Communications and Networking 4 (2018) 825–842. 15/2021) 1981–1999.
[122] F. Fourati, M.-S. Alouini, Artificial Intelligence for Satellite Communication: A [156] S. Hilton, F. Cairola, A. Gardi, R. Sabatini, N. Pongsakornsathien, N. Ezer,
Review, 2021 arXiv preprint arXiv:2101.10899. Uncertainty quantification for space situational awareness and traffic
[123] H. Dahrouj, R. Alghamdi, H. Alwazani, S. Bahanshal, A.A. Ahmad, A. Faisal, et al., management, Sensors 19 (10/09 2019) 4361.
An Overview of Machine Learning-Based Techniques for Solving Optimization [157] S. Hilton, A. Gardi, F. Cairola, R. Sabatini, A Sensor-Centric Approach to Space
Problems in Communications and Signal Processing, IEEE Access, 2021. Traffic Management, 2019.
[124] A. Anzagira, W.W. Edmonson, D.N. Amanor, LED-based visible light inter-satellite [158] S. Hilton, R. Sabatini, A. Gardi, H. Ogawa, P. Teofilatto, Space traffic
communication for distributed space systems, IEEE Journal on Miniaturization for management: towards safe and unsegregated space transport operations, Prog.
Air and Space Systems (2021). Aero. Sci. 105 (02/01 2019).
[125] P. Guan, X.-J. Liu, J.-Z. Liu, Adaptive fuzzy sliding mode control for flexible [159] E. Blasch, D. Shen, G. Chen, J. Sung, Multisource AI scorecard table analysis of
satellite, Eng. Appl. Artif. Intell. 18 (2005/06/01/2005) 451–459. AMIGO, in: Proceedings of Sensors And Systems For Space Applications XIV, 2021,
[126] W. MacKunis, F. Leve, P.M. Patre, N. Fitz-Coy, W.E. Dixon, Adaptive neural p. 117550B.
network-based satellite attitude control in the presence of CMG uncertainty, Aero. [160] S. Nag, D.D. Murakami, N.A. Marker, M.T. Lifson, P.H. Kopardekar, Prototyping
Sci. Technol. 54 (2016/07/01/2016) 218–228. operational autonomy for space traffic management, Acta Astronaut. 180 (2021)
[127] D. Lee, Guidance, Navigation and Control System for Autonomous Proximity 489–506.
Operations and Docking of Spacecraft, 01/01 2009. [161] S.J. Siddiqi, F. Naeem, S. Khan, K.S. Khan, M. Tariq, Towards AI-enabled traffic
[128] S. Lim, M. Stoeckle, B.J. Streetman, M. Neave, Markov neural network for management in multipath TCP: a survey, Comput. Commun. (2021).
guidance, navigation and control, in: AIAA Scitech 2020 Forum, American [162] J. Zhang, D. Ye, J.D. Biggs, Z. Sun, Finite-time relative orbit-attitude tracking
Institute of Aeronautics and Astronautics, 2020. control for multi-spacecraft with collision avoidance and changing network
[129] Q. Yao, Neural adaptive attitude tracking control for uncertain spacecraft with topologies, Adv. Space Res. 63 (2019/02/01/2019) 1161–1175.
preassigned performance guarantees, Adv. Space Res. (2021/10/18/2021). [163] K. Thangavel, A. Gardi, S. Hilton, A. Afful, R. Sabatini. Towards Multi-Domain
[130] Q. Yao, Adaptive fuzzy neural network control for a space manipulator in the Traffic Management, UAE, November 2021.
presence of output constraints and input nonlinearities, Adv. Space Res. 67 [164] T. Yairi, N. Takeishi, T. Oda, Y. Nakajima, N. Nishimura, N. Takata, A data-driven
(2021/03/15/2021) 1830–1843. health monitoring method for satellite housekeeping data based on probabilistic
[131] M. Zheng, J. Luo, Z. Dang, Feedforward Neural Network Based Time-Varying clustering and dimensionality reduction, IEEE Trans. Aero. Electron. Syst. PP (02/
State-Transition-Matrix of Tschauner-Hempel Equations, 2021. Advances in Space 20 2017), 1-1.
Research, 2021/10/20. [165] A. Young, C. Kitts, M. Neumann, I. Mas, M. Rasay. Initial Flight Results for an
[132] R. Lungu, M. Lungu, C. Efrim, Adaptive control of DGMSCMG using dynamic Automated Satellite Beacon Health Monitoring Network, Utah State University,
inversion and neural networks, Adv. Space Res. 68 (2021/10/15/2021) Utah (USA, 2010.
3478–3494. [166] G. Landis, S. Bailey, R. Tischler, Causes of power-related satellite failures, in:
[133] E. Sin, S. Nag, V. Ravindra, A. Li, M. Arcak, Attitude Trajectory Optimization for 2006 IEEE 4th World Conference on Photovoltaic Energy Conference, vol. 2,
Agile Satellites in Autonomous Remote Sensing Constellation, 2021. 2006, pp. 1943–1945.
[134] S. Abdelghafar, A. Darwish, A.E. Hassanien, Intelligent Health Monitoring [167] K. Ranasinghe, R. Sabatini, A. Gardi, S. Bijjahalli, R. Kapoor, T. Fahey, et al.,
Systems for Space Missions Based on Data Mining Techniques, 2020, pp. 65–78. Advances in integrated system health management for mission-essential and
[135] C.-P. Fritzen, Vibration-based structural health monitoring – concepts and safety-critical aerospace applications, Prog. Aero. Sci. 128 (2022).
applications, Key Engineering Materials - KEY ENG MAT 293 (01/01 2005). [168] P. Grandjean, T. Pesquet, A. M. M. Muxi, and M. C. Charmeau, "What on-board
[136] S. Fuertes, G. Picart, J.-Y. Tourneret, L. Chaari, A. Ferrari, C. Richard, Improving autonomy means for ground operations: an autonomy demonstrator conceptual
Spacecraft Health Monitoring with Automatic Anomaly Detection Techniques, design", in Space OPS 2004 Conference, 17-21 May 2004 Montreal, Quebec
2016. (Canada).
[137] D. E. Paris, L. C. Trevino and M. D. Watson, "A framework for integration of IVHM [169] A. Börner, H.-W. Hübers, O. Kao, F. Schmidt, S. Becker, J. Denzler, et al., Sensor
technologies for intelligent integration for vehicle management," 2005 IEEE Artificial Intelligence and its Application to Space Systems – A White Paper, 2020.
Aerospace Conference, Big Sky, MT, USA, 2005, pp. 3843-3852. [170] O. Kodheli, E. Lagunas, N. Maturo, S.K. Sharma, B. Shankar, J. Montoya, et al.,
[138] K. Penta, D. Khemani, Satellite Health Monitoring Using CBR Framework, vol. Satellite Communications in the New Space Era: A Survey and Future Challenges,
3155, 2004. 2020.
[139] M. Tipaldi, B. Bruenjes, Spacecraft health monitoring and management systems. [171] L. Pellaco, N. Singh, J. Jaldén, Spectrum Prediction and Interference Detection for
2014 IEEE Metrology for Aerospace (MetroAeroSpace), Benevento, Italy, 2014, Satellite Communications, 2019.
pp. 68–72, https://doi.org/10.1109/MetroAeroSpace.2014.6865896. [172] A. Vazquez, P. Henarejos, A. Perez-Neira, E. Grechi, A. Voight, J. Gil, et al., On the
[140] L. Elerin, C. Learoyd, B. Wilson, Applying neural networks and other AI Use of AI for Satellite Communications, 2020.
techniques to fault detection in satellite communication systems, in: Proceedings [173] M. Bappy, R. Huq, S. Siddique, Ai-Obc, Conceptual Design of a Deep Neural
of Neural Networks For Signal Processing VII. Proceedings of the 1997 IEEE Signal Network Based Next Generation Onboard Computing Architecture for Satellite
Processing Society Workshop, 1997, pp. 617–625. Systems, 2019.
[141] A. Malik, B. Moster, C. Obermeier, Exoplanet Detection Using Machine Learning, [174] V. Kothari, E. Liberis, N. Lane, The Final Frontier: Deep Learning in Space, 2020.
2020. [175] A. Tahir, H.S. Munawar, J. Akram, M. Adil, S. Ali, A.Z. Kouzani, et al., Automatic
[142] T. Estlin, F. Fisher, D. Mutz, S. Chien, Automated planning for a Deep Space target detection from satellite imagery using machine learning, Sensors 22 (2022)
Communications Station 2 (1999). 1147.
[143] F. Fisher, T. Estlin, D. Mutz, S. Chien, An AI Approach to Ground Station [176] W. Guo, W. Yang, H. Zhang, G. Hua, Geospatial object detection in high resolution
Automomy for Deep Space Communications 440 (07/31 1999) 191. satellite images based on multi-scale convolutional neural network, Rem. Sens. 10
[144] F. Fisher, R. Knight, B. Engelhardt, S. Chien, N. Alejandre, A Planning Approach (2018) 131.
to Monitor and Control for Deep Space Communications, vol. 2, 2000.

33
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

[177] S. Talukdar, P. Singha, S. Mahato, S. Pal, Y.-A. Liou, A. Rahman, Land-use land- [205] Y. Han, Y. Liu, Z. Hong, Y. Zhang, S. Yang, J. Wang, Sea ice image classification
cover classification by machine learning classifiers for satellite observations—a based on heterogeneous data fusion and deep learning, Rem. Sens. 13 (2021) 592.
review, Rem. Sens. 12 (2020) 1135. [206] S. Hartling, V. Sagan, P. Sidike, M. Maimaitijiang, J. Carron, Urban tree species
[178] C. Huang, L. Davis, J. Townshend, An assessment of support vector machines for classification using a WorldView-2/3 and LiDAR data fusion approach and deep
land cover classification, Int. J. Rem. Sens. 23 (2002) 725–749. learning, Sensors 19 (2019) 1284.
[179] D.K. McIver, M.A. Friedl, Estimating pixel-scale land cover classification [207] R.K. Vemuri, P.C.S. Reddy, B. Puneeth Kumar, J. Ravi, S. Sharma, S. Ponnusamy,
confidence using nonparametric machine learning methods, IEEE Trans. Geosci. Deep learning based remote sensing technique for environmental parameter
Rem. Sens. 39 (2001) 1959–1968. retrieval and data fusion from physical models, Arabian J. Geosci. 14 (2021)
[180] G.P. Petropoulos, P. Partsinevelos, Z. Mitraka, Change detection of surface mining 1230.
activity and reclamation based on a machine learning approach of multi-temporal [208] N.-B. Chang, K. Bai, Multisensor Data Fusion and Machine Learning for
Landsat TM imagery, Geocarto Int. 28 (2013) 323–342. Environmental Remote Sensing, CRC Press, 2018.
[181] G. Camps-Valls, Machine learning in remote sensing data processing, in: [209] G. Soldi, D. Gaglione, N. Forti, L.M. Millefiori, P. Braca, S. Carniel, et al., Space-
Proceedings of 2009 IEEE International Workshop on Machine Learning for Signal based global maritime surveillance. Part II: artificial intelligence and data fusion
Processing, 2009, pp. 1–6. techniques, IEEE Aero. Electron. Syst. Mag. 36 (2021) 30–42.
[182] W. Shi, M. Zhang, R. Zhang, S. Chen, Z. Zhan, Change detection based on artificial [210] H.M. Keshk, X.-C. Yin, Obtaining super-resolution satellites images based on
intelligence: state-of-the-art and challenges, Rem. Sens. 12 (2020) 1688. enhancement deep convolutional neural network, International Journal of
[183] S. Zhong, K. Zhang, M. Bagheri, J.G. Burken, A. Gu, B. Li, et al., Machine Aeronautical and Space Sciences 22 (2021) 195–202.
Learning: New Ideas and Tools in Environmental Science and Engineering, vol. [211] Z. Wang, J. Chen, S.C. Hoi, Deep learning for image super-resolution: a survey,
55, Environmental Science & Technology, 2021, pp. 12741–12754. IEEE Trans. Pattern Anal. Mach. Intell. 43 (2020) 3365–3387.
[184] M.A. Dias, E.A.d. Silva, S.C.d. Azevedo, W. Casaca, T. Statella, R.G. Negri, An [212] H. He, H. Xu, Y. Zhang, K. Gao, H. Li, L. Ma, et al., Mask R-CNN based automated
incongruence-based anomaly detection strategy for analyzing water pollution in identification and extraction of oil well sites, Int. J. Appl. Earth Obs. Geoinf. 112
images from remote sensing, Rem. Sens. 12 (2019) 43. (2022), 102875.
[185] R. Sommer, V. Paxson, Outside the closed world: on using machine learning for [213] Z. Xie, U.K. Haritashya, V.K. Asari, B.W. Young, M.P. Bishop, J.S. Kargel,
network intrusion detection, in: Proceedings of 2010 IEEE Symposium on Security GlacierNet: a deep-learning approach for debris-covered glacier mapping, IEEE
and Privacy, 2010, pp. 305–316. Access 8 (2020) 83495–83510.
[186] V. Khryashchev, L. Ivanovsky, V. Pavlov, A. Ostrovskaya, A. Rubtsov, Comparison [214] L. Liao, J. Xiao, Y. Li, M. Wang, R. Hu, Learned representation of satellite image
of different convolutional neural network architectures for satellite image series for data compression, Rem. Sens. 12 (2020) 497.
segmentation, in: Proceedings of 2018 23rd Conference of Open Innovations [215] V. Kothari, E. Liberis, N.D. Lane, The final frontier: deep learning in space, in:
Association, FRUCT), 2018, pp. 172–179. Proceedings of Proceedings Of the 21st International Workshop on Mobile Computing
[187] R. Gupta, M. Shah, Rescuenet: joint building segmentation and damage Systems and Applications, 2020, pp. 45–49.
assessment from satellite imagery, in: Proceedings of 2020 25th International [216] V.A. de Oliveira, M. Chabert, T. Oberlin, C. Poulliat, M. Bruno, C. Latry, et al.,
Conference On Pattern Recognition, ICPR), 2021, pp. 4405–4411. Satellite image compression and denoising with neural networks, Geosci. Rem.
[188] W. Feng, M. Zhong, J.M. Lemoine, R. Biancale, H.T. Hsu, J. Xia, Evaluation of Sens. Lett. IEEE 19 (2022) 1–5.
groundwater depletion in North China using the Gravity Recovery and Climate [217] Y. Dua, R.S. Singh, K. Parwani, S. Lunagariya, V. Kumar, Convolution neural
Experiment (GRACE) data and ground-based measurements, Water Resour. Res. network based lossy compression of hyperspectral images, Signal Process. Image
49 (2013) 2110–2118. Commun. 95 (2021), 116255.
[189] R. Houborg, M. Rodell, B. Li, R. Reichle, B.F. Zaitchik, Drought indicators based [218] B. Herfort, H. Li, S. Fendrich, S. Lautenbach, A. Zipf, Mapping human settlements
on model-assimilated Gravity Recovery and Climate Experiment (GRACE) with higher accuracy and less volunteer efforts by combining crowdsourcing and
terrestrial water storage observations, Water Resour. Res. 48 (2012). deep learning, Rem. Sens. 11 (2019) 1799.
[190] E. Ibrahim, L. Lema, P. Barnabé, P. Lacroix, E. Pirard, Small-scale surface mining [219] A. Brand, A. Manandhar, Semantic segmentation of burned areas in satellite
of gold placers: detection, mapping, and temporal analysis through the use of free images using a U-net-based convolutional neural network, Int. Arch. Photogram.
satellite imagery, Int. J. Appl. Earth Obs. Geoinf. 93 (2020), 102194. Rem. Sens. Spatial Inf. Sci. 43 (2021) 47–53.
[191] V.S. Sardar, K. Yindumathi, S.S. Chaudhari, P. Ghosh, Convolution neural [220] J. Allworth, L. Windrim, J. Bennett, M. Bryson, A transfer learning approach to
network-based agriculture drought prediction using satellite images, in: space debris classification using observational light curve data, Acta Astronaut.
Proceedings of 2021 IEEE Mysore Sub Section International Conference, 181 (2021) 301–315.
MysuruCon), 2021, pp. 601–607. [221] B. Wang, S. Li, J. Mu, X. Hao, W. Zhu, J. Hu, Research advancements in key
[192] A.F. Marj, A.M. Meijerink, Agricultural drought forecasting using satellite images, technologies for space-based situational awareness, Space: Sci. Technol. 2022
climate indices and artificial neural network, Int. J. Rem. Sens. 32 (2011) (2022).
9707–9719. [222] T. Mahendrakar, A. Ekblad, N. Fischer, R. White, M. Wilde, B. Kish, et al.,
[193] L. Hassan-Esfahani, A. Torres-Rua, A. Jensen, M. McKee, Assessment of surface Performance study of YOLOv5 and faster R-CNN for autonomous navigation
soil moisture using high-resolution multi-spectral imagery and artificial neural around non-cooperative targets, in: Proceedings of 2022 IEEE Aerospace
networks, Rem. Sens. 7 (2015) 2627–2646. Conference, AERO), 2022, pp. 1–12.
[194] N. Kussul, M. Lavreniuk, S. Skakun, A. Shelestov, Deep learning classification of [223] C.E. Oestreich, R. Linares, R. Gondhalekar, Autonomous six-degree-of-freedom
land cover and crop types using remote sensing data, Geosci. Rem. Sens. Lett. spacecraft docking with rotating targets via reinforcement learning, J. Aero. Inf.
IEEE 14 (2017) 778–782. Syst. 18 (2021) 417–428.
[195] V. Alhassan, C. Henry, S. Ramanna, C. Storie, A deep learning framework for land- [224] B.B. Virgili, T. Flohrer, H. Krag, K. Merz, S. Lemmens, CREAM-ESA’s proposal for
use/land-cover mapping and analysis using multispectral satellite imagery, collision risk estimation and automated mitigation, in: Proceedings of First
Neural Comput. Appl. 32 (2020) 8529–8544. International Orbital Debris Conference, 2019, p. 6031.
[196] K. Karra, C. Kontgis, Z. Statman-Weil, J.C. Mazzariello, M. Mathis, S.P. Brumby, [225] M. Shan, J. Guo, E. Gill, Review and comparison of active space debris capturing
Global land use/land cover with Sentinel 2 and deep learning, in: Proceedings of and removal methods, Prog. Aero. Sci. 80 (2016) 18–32.
2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, 2021, [226] S. Sen, S. Agarwal, P. Chakraborty, K.P. Singh, Astronomical big data processing
pp. 4704–4707. using machine learning: a comprehensive review, Exp. Astron. 53 (2022) 1–43.
[197] S.H. Khan, X. He, F. Porikli, M. Bennamoun, Forest change detection in [227] C.J. Fluke, C. Jacobs, Surveying the reach and maturity of machine learning and
incomplete satellite images with deep neural networks, IEEE Trans. Geosci. Rem. artificial intelligence in astronomy, Wiley Interdisciplinary Reviews: Data Min.
Sens. 55 (2017) 5407–5423. Knowl. Discov. 10 (2020) e1349.
[198] A. Varghese, J. Gubbi, A. Ramaswamy, P. Balamuralidhar, ChangeNet: a deep [228] S. Zucker, R. Giryes, Shallow transits—deep learning. I. Feasibility study of deep
learning architecture for visual change detection, in: Proceedings of Proceedings learning to detect periodic transits of exoplanets, Astron. J. 155 (2018) 147.
Of the European Conference on Computer Vision, ECCV) workshops, 2018, 0-0. [229] A. Dattilo, A. Vanderburg, C.J. Shallue, A.W. Mayo, P. Berlind, A. Bieryla, et al.,
[199] A. Mullapudi, A.D. Vibhute, S. Mali, C.H. Patil, Spatial and seasonal change Identifying exoplanets with deep learning. ii. two new super-earths uncovered by
detection in vegetation cover using time-series landsat satellite images and a neural network in k2 data, Astron. J. 157 (2019) 169.
machine learning methods, SN Computer Science 4 (2023) 254. [230] J. Kremer, K. Stensbo-Smidt, F. Gieseke, K.S. Pedersen, C. Igel, Big universe, big
[200] M. Ahangarha, R. Shah-Hosseini, M. Saadatseresht, Deep learning-based change data: machine learning and image analysis for astronomy, IEEE Intell. Syst. 32
detection method for environmental change monitoring using sentinel-2 datasets, (2017) 16–22.
Environmental Sciences Proceedings 5 (2020) 15. [231] A. Druetto, M. Roberti, R. Cancelliere, D. Cavagnino, M. Gai, A deep learning
[201] A. Groener, G. Chern, M. Pritt, A comparison of deep learning object detection approach to anomaly detection in the gaia space mission data, in: Proceedings of
models for satellite imagery, in: Proceedings of 2019 IEEE Applied Imagery Pattern Advances In Computational Intelligence: 15th International Work-Conference On
Recognition Workshop, AIPR), 2019, pp. 1–10. Artificial Neural Networks, IWANN 2019, Gran Canaria, Spain, 2019, pp. 390–401.
[202] T. Lu, J. Wang, Y. Zhang, Z. Wang, J. Jiang, Satellite image super-resolution via June 12-14, 2019, Proceedings, Part II 15.
multi-scale residual deep neural network, Rem. Sens. 11 (2019) 1588. [232] M. Hon, D. Stello, J. Yu, Deep learning classification in asteroseismology using an
[203] S.M. Erfani, S. Rajasegarar, S. Karunasekera, C. Leckie, High-dimensional and improved neural network: results on 15 000 Kepler red giants and applications to
large-scale anomaly detection using a linear one-class SVM with deep learning, K2 and TESS data, Mon. Not. Roy. Astron. Soc. 476 (2018) 3233–3244.
Pattern Recogn. 58 (2016) 121–134. [233] M. Hon, D. Stello, J. Yu, Deep learning classification in asteroseismology, Mon.
[204] K.A. da Costa, J.P. Papa, L.A. Passos, D. Colombo, J. Del Ser, K. Muhammad, et Not. Roy. Astron. Soc. 469 (2017) 4578–4583.
al., A critical literature survey and prospects on tampering and anomaly detection [234] M. Ansdell, Y. Ioannou, H.P. Osborn, M. Sasdelli, J.C. Smith, D. Caldwell, et al.,
in image data, Appl. Soft Comput. 97 (2020), 106727. Scientific domain knowledge improves exoplanet transit classification with deep
learning, Astrophys. J. Lett. 869 (2018) L7.

34
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

[235] F.I. San Martín, C.A. Perez, J.E. Tapia, S. Virani, M.J. Holzinger, Automatic space [266] S.D. Ilčev, Space segment, in: Global Satellite Meteorological Observation
object detection on all-sky images from a synoptic survey synthetic telescope (GSMO) Theory, vol. 1, Springer International Publishing, Cham, 2018,
array, Adv. Space Res. 65 (2020) 337–350. pp. 71–143.
[236] S. Khlamov, I. Tabakova, T. Trunova, Recognition of the astronomical images [267] X. Vuong, Efficient Utilization of the Space Segment Resources, 2008, 10/01.
using the Sobel filter, in: Proceedings of 2022 29th International Conference On [268] C.-J. Fong, W. D, E. Yang, K. Cook, V. Chu, S. B, et al., Space and ground segment
Systems, Signals And Image Processing, IWSSIP), 2022, pp. 1–4. performance of the FORMOSAT-3/COSMIC mission: four years in orbit,
[237] K. Borne, Scientific Data Mining in Astronomy, 2009 arXiv preprint arXiv: Atmospheric Measurement Techniques Discussions 4 (2011) 1–24.
0911.0505. [269] E. Gill, O. Montenbruck, H. Kayal, K. Briess, A Concept for an Autonomous Space
[238] J.-L. Starck, F. Murtagh, Astronomical Image and Data Analysis, 2007. and Ground Segment Applied to the BIRD Satellite Mission, 2001.
[239] S. Varma, M. Fairbairn, J. Figueroa, Dark Matter Subhalos, Strong Lensing and [270] M. Esposito, B. Carnicero Domíguez, M. Pastena, N. Vercruyssen, S. Conticello,
Machine Learning, 2020 arXiv preprint arXiv:2005.05353. C. Dijk, et al., Highly Integration of Hyperspectral, Thermal and Artificial
[240] L.E. Strigari, Galactic searches for dark matter, Phys. Rep. 531 (2013) 1–88. Intelligence for the Esa Phisat-1 Mission, 2019.
[241] A. Brandonisio, M. Lavagna, D. Guzzetti, Reinforcement learning for [271] M. Pastena, ESA Earth Observation on Board Data Processing Future Needs and
uncooperative space objects smart imaging path-planning, J. Astronaut. Sci. 68 Technologies, 2019.
(2021) 1145–1169. [272] M. Pastena, Φsat-1: the first ESA Earth Observation Directorate cubesat mission,
[242] S. Silvestrini, M. Lavagna, Model-based reinforcement learning for distributed in: Presented at the 4th Cubesat Industry Days, 2019. Netherlands.
path planning, in: Proceedings of 15th Symposium On Advanced Space Technologies [273] M. Esposito, S. Conticello, M. Pastena, B. Carnicero Domínguez, In-orbit
In Robotics And Automation, 2019. Demonstration of Artificial Intelligence Applied to Hyperspectral and Thermal
[243] J. Broida, R. Linares, Spacecraft rendezvous guidance in cluttered environments Sensing from Space, in: SPIE Optical Engineering + Applications, vol. 11131,
via reinforcement learning, in: Proceedings of 29th AAS/AIAA Space Flight SPIE, 2019.
Mechanics Meeting, 2019, pp. 1–15. [274] P. Safarik and S. Schuenemann, "Ground segment as a service", in SpaceOps 2016
[244] J.L. Gonzalo, C. Colombo, On-board collision avoidance applications based on Conference.Daejeon, South Korea, 16 - 20 May 2016.
machine learning and analytical methods, in: Proceedings of 8th European [275] M. Scholl, Introduction to cybersecurity for commercial satellite operations, Draft
Conference On Space Debris, ESA/ESOC, Darmstadt, Germany, Virtual Conference, NISTIR 8270 (2021).
2021, pp. 20–23. [276] L. Koerschner, Ground Segment Preparation for NPSAT1, 09/01 2007, p. 77.
[245] M. Shirobokov, S. Trofimov, M. Ovchinnikov, Survey of machine learning [277] E. Romero, M. Oglietti, Cooperative Space Mission Operation Planning by
techniques in spacecraft control design, Acta Astronaut. 186 (2021) 87–97. Extended Preferences, 2020, 08/21.
[246] Q. Qu, K. Liu, W. Wang, J. Lü, Spacecraft proximity maneuvering and rendezvous [278] aiko, MIRAGE -the Future of Autonomous Operations in Space, 2021.
with collision avoidance based on reinforcement learning, IEEE Trans. Aero. [279] C. Iacopino, P. Palmer, Autonomy, in: M. D’Errico (Ed.), Distributed Space
Electron. Syst. 58 (2022) 5823–5834. Missions for Earth System Monitoring, Springer New York, New York, NY, 2013,
[247] J. Elkins, R. Sood, C. Rumpf, Adaptive continuous control of spacecraft attitude pp. 309–329.
using deep reinforcement learning, in: Proceedings of 2020 AAS/AIAA [280] L. Martínez, J.C. Merlano Duncan, J. Querol, S. Kumar, J. Krivochiza, S.
Astrodynamics Specialist Conference, 2020, pp. 420–475. K. Sharma, et al., Architectures and Synchronization Techniques for Distributed
[248] H. Ahn, D. Jung, H.-L. Choi, Deep generative models-based anomaly detection for Satellite Systems, A Survey, 2022.
spacecraft control systems, Sensors 20 (2020) 1991. [281] V.W. Chan, Optical satellite networks, J. Lightwave Technol. 21 (2003) 2811.
[249] S.K. Ibrahim, A. Ahmed, M.A.E. Zeidan, I.E. Ziedan, Machine learning techniques [282] H. Hemmati, D. Caplan, Optical Satellite Communications, Optical Fiber
for satellite fault diagnosis, Ain Shams Eng. J. 11 (2020) 45–56. Telecommunications, 2013, pp. 121–162.
[250] G. Makridis, D. Kyriazis, S. Plitsos, Predictive maintenance leveraging machine [283] A. Chaudhry, H. Yanikomeroglu, Free Space Optics for Next-Generation Satellite
learning for time-series forecasting in the maritime industry, in: Proceedings of Networks, 2020.
2020 IEEE 23rd International Conference On Intelligent Transportation Systems, [284] J.J. Dougherty, H. El-Sherief, D.J. Simon, G.A. Whitmer, GPS modeling for
ITSC), 2020, pp. 1–8. designing aerospace vehicle navigation systems, IEEE Trans. Aero. Electron. Syst.
[251] M. Nacchia, F. Fruggiero, A. Lambiase, K. Bruton, A systematic mapping of the 31 (2) (1995) 695–705.
advancing use of machine learning techniques for predictive maintenance in the [285] R. Sabatini, T. Moore, S. Ramasamy, Global navigation satellite systems
manufacturing sector, Appl. Sci. 11 (2021) 2546. performance analysis and augmentation strategies in aviation, Prog. Aero. Sci. 95
[252] S. Mitrovic-Minic, D. Thomson, J. Berger, J. Secker, Collection Planning and (2017) 45–98.
Scheduling for Multiple Heterogeneous Satellite Missions: Survey, Optimization [286] R. Rifandi, S. Assagaf, Y.D.W.K. Ningtyas, An Insight about GPS, 2013.
Problem, and Mathematical Programming Formulation", Modeling And [287] A. Boda, Cycle Slip Detection Using PVA Kalman Filter, York University, Thesis,
Optimization In Space Engineering: State Of the Art And New Challenges, 2019, 2017.
pp. 271–305. [288] F. Major, Satellite Navigation: the User Segment, 2014, pp. 349–366.
[253] S.-w. Baek, S.-m. Han, K.-r. Cho, D.-w. Lee, J.-s. Yang, P.M. Bainum, et al., [289] L.J. Planke, A. Gardi, R. Sabatini, T. Kistan, N. Ezer, Online multimodal inference
Development of a scheduling algorithm and GUI for autonomous satellite of mental workload for cognitive human machine systems, Computers 10 (2021).
missions, Acta Astronaut. 68 (2011) 1396–1402. [290] K. Thangavel, D. Spiller, R. Sabatini, P. Servidia, P. Marzocca, H. Fayek, et al.,
[254] W. Martin, Satellite image collection optimization, Opt. Eng. 41 (2002) Trusted Autonomous Distributed Satellite System Operations for Earth
2083–2087. Observation", Presented at the 17th International Conference on Space
[255] H. Kim, Y.K. Chang, Mission scheduling optimization of SAR satellite Operations, United Arab Emirates, 2023.
constellation for minimizing system response time, Aero. Sci. Technol. 40 (2015) [291] K. Thangavel, D. Spiller, R. Sabatini, S. Amici, N. Longepe, P. Servidia, et al.,
17–32. Trusted autonomous operations of distributed satellite systems using optical
[256] J. Bonnet, M.-P. Gleizes, E. Kaddoum, S. Rainjonneau, G. Flandin, Multi-satellite sensors, Sensors 23 (2023) 3344.
mission planning using a self-adaptive multi-agent system, in: Proceedings of [292] K. Thangavel, D. Spiller, R. Sabatini, S. Amici, S.T. Sasidharan, H. Fayek, et al.,
2015 IEEE 9th International Conference On Self-Adaptive And Self-Organizing Autonomous satellite wildfire detection using hyperspectral imagery and neural
Systems, 2015, pp. 11–20. networks: a case study on Australian wildfire, Rem. Sens. 15 (3) (2023).
[257] L.T. Bui, Z. Michalewicz, E. Parkinson, M.B. Abello, Adaptation in dynamic [293] K. Thangavel, D. Spiller, R. Sabatini, P. Marzocca, "On-board data processing of
environments: a case study in mission planning, IEEE Trans. Evol. Comput. 16 earth observation data using 1-D CNN", presented at the SmartSat, CRC
(2011) 190–209. Conference (2022) 12–13. September 2022.
[258] S. Chien, A generalized timeline representation, services, and interface for [294] K. Thangavel, D. Spiller, R. Sabatini, P. Marzocca, M. Esposito, Near real-time
automating space mission operations, in: SpaceOps 2012, 2012, 1275459. wildfire management using distributed satellite system, Geosci. Rem. Sens. Lett.
[259] G.A. Salazar, Considerations for Implementing Voice-Controlled Spacecraft IEEE PP (12/01 2022), 1-1.
Systems through a Human-Centered Design Approach, 2018. [295] D. Spiller, K. Thangavel, S.T. Sasidharan, S. Amici, L. Ansalone, R. Sabatini,
[260] G. Salazar, Development Considerations for implementing a voice-controlled Wildfire segmentation analysis from edge computing for on-board real-time alerts
spacecraft system, in: Proceedings of 2019 IEEE International Symposium On using hyperspectral imagery, in: Proceedings of 2022 IEEE International Conference
Measurement And Control In Robotics, ISMCR), 2019. B3-1-1-B3-1-20. On Metrology For Extended Reality, Artificial Intelligence and Neural Engineering
[261] T. Vega, A. Bueno, C. Cerutti, H. Chang, M. Garvin, C. Garza, et al., Enabling a (MetroXRAINE), 2022, pp. 725–730.
voice management system for space applications. IEEE Aerospace Conference, Big [296] K. Thangavel, P. Servidia, R. Sabatini, P. Marzocca, H. Fayek, S.H. Cerruti, et al.,
Sky, Montana. March 5-12, 2022. A distributed satellite system for multibaseline AT-InSAR: constellation of
[262] S.S. Heikkilä, A. Halme, A. Schiele, Affordance-based indirect task formations for maritime domain awareness using autonomous orbit control,
communication for astronaut-robot cooperation, J. Field Robot. 29 (2012) Aerospace 10 (2023) 176.
576–600. [297] K. Thangavel, P. Servidia, R. Sabatini, P. Marzocca, H. Fayek, D. Spiller,
[263] I.L. Nunes, Advances in human factors and systems interaction, in: Proceedings of Distributed Satellite System for Maritime Domain Awareness", Presented at the
Proceedings Of the AHFE 2020 Virtual Conference On Human Factors And Systems Australian International Aerospace Congress, AIAC20), Melbourne, Australia,
Interaction, July, 2017, pp. 16–20. 2023.
[264] G. Furano, G. Meoni, A. Dunne, D. Moloney, V. Ferlet-Cavrois, A. Tavoularis, et [298] KP Labs, Antelope, Accessed: Jul. 14, 2022, https://kplabs.space/wp-conte
al., Towards the use of artificial intelligence on the edge in space systems: nt/uploads/Antelope_technical-sheet.pdf.
challenges and opportunities, IEEE Aero. Electron. Syst. Mag. 35 (2020) 44–56. [299] E. Lagona, S. Hilton, A. Afful, A. Gardi, R. Sabatini, Autonomous trajectory
[265] G. Furano, A. Tavoularis, M. Rovatti, AI in space: applications examples and optimisation for intelligent satellite systems and space traffic management, Acta
challenges, in: Proceedings of 2020 IEEE International Symposium On Defect And Astronaut. (01/25 2022).
Fault Tolerance In VLSI and Nanotechnology Systems (DFT), 2020, pp. 1–6.

35
K. Thangavel et al. Progress in Aerospace Sciences 144 (2024) 100960

[300] K. Hussain, K. Thangavel, G. Alessandro, S. Roberto, Autonomous optical sensing [325] H. o. C. S. a. T. Committee, Robotics and Artificial Intelligence, 2016.
for space-based space surveillance, in: Presented at the 44th IEEE Aerospace [326] D. Harkut, K. Kasat, Introductory Chapter: Artificial Intelligence - Challenges and
Conference, 2023. Montana, United States of America. Applications, 2019.
[301] K. Hussain, K. Thangavel, A. Gardi, R. Sabatini, Autonomous tracking of Resident [327] J.B. James Manyika, The Promise and Challenge of the Age of Artificial
Space Objects using multiple ground-based Electro-Optical sensors, in: Presented Intelligence, 2018.
at the 17th International Conference on Space Operations, 2023. United Arab [328] A. Pavaloiu, U. Köse, Ethical artificial intelligence - an open question, Journal of
Emirates. Multidisciplinary Developments 2 (04/01 2017) 15–27.
[302] K.F. Hussain, K. Thangavel, A. Gardi, R. Sabatini, Passive electro-optical tracking [329] M.U. Scherer, Regulating artificial intelligence systems: risks, challenges,
of resident space objects for distributed satellite systems autonomous navigation, competencies, and strategies, Harv. J. Law Technol. 29 (2015) 353.
Rem. Sens. 15 (2023) 1714. [330] L. Soroka, K. Kurkova, Artificial intelligence and space technologies: legal, ethical
[303] K. Hussain, K. Thangavel, A. Gardi, R. Sabatini, Autonomous Tracking of Resident and technological issues, Advanced Space Law 3 (05/01 2019).
Space Objects Using Multiple Ground-Based Electro-Optical Sensors, 2023. [331] J. Stephenson, C. Acklam, Artificial intelligence in care: where does responsibility
[304] K. Thangavel, A. Gardi, S. Hilton, A.M. Afful, R. Sabatini, Towards multi-domain lie? Nurs. Resid. Care 21 (05/02 2019) 281–283.
traffic management, in: Proceedings of 72nd International Astronautical Congress, [332] P. Vähäkainu, M. Lehto, Artificial Intelligence in the Cyber Security Environment
IAC 2021, UAE, Dubai, 2021. Artificial Intelligence in the Cyber Security Environment, 2019.
[305] S. Hilton, R. Sabatini, A. Gardi, H. Ogawa, P. Teofilatto, Space traffic [333] M. Hutchinson, E. Antono, B. Gibbons, S. Paradiso, J. Ling, B. Meredig,
management: towards safe and unsegregated space transport operations, Prog. Overcoming Data Scarcity with Transfer Learning, 2017, 11/02.
Aero. Sci. 105 (Feb 2019) 98–125. [334] E. Ntoutsi, P. Fafalios, U. Gadiraju, V. Iosifidis, W. Nejdl, M.E. Vidal, et al., Bias in
[306] A. Console, Space resilience – why and how? JAPCC Journal (2018). data-driven artificial intelligence systems—an introductory survey, WIREs Data
[307] C. Atkinson, Hybrid Warfare and Societal Resilience Implications for Democratic Mining and Knowledge Discovery (02/03 2020).
Governance, vol. 39, Information & Security, 2018, pp. 63–76. [335] D. Roselli, J. Matthews, N. Talagala, Managing Bias in AI, 2019.
[308] N. Nikolic, P. Anastasov, G. Tsvetkov, M. Street, I.I. Mestric, R. Birger, Connecting [336] C. Yavuz, Machine Bias Artificial Intelligence and Discrimination, Lund
conflict concepts: hybrid warfare and warden’s rings, Inf. Secur. Int. J. 41 (2018). University, Thesis, 2019.
[309] G. Treverton, in: F.H.S. Försvarshögskolan ( (Ed.), The Intelligence Challenges of [337] G. Benelli, G. Meoni, L. Fanucci, A low power keyword spotting algorithm for
Hybrid Threats: Focus on Cyber and Virtual Realm, 2018. memory constrained embedded systems, in: Proceedings of 2018 IFIP/IEEE
[310] E. Meyrick, A. Pickard, T. Rahloff, S. Bonnart, A. Carlo, K. Thangavel, Ground International Conference On Very Large Scale Integration, VLSI-SoC), 2018,
station as a service: a space cybersecurity analysis, in: Presented at the 72nd pp. 267–272.
International Astronautical Congress (IAC), United Arab Emirates, Dubai, 2021. [338] M. Blott, T.B. Preußer, N.J. Fraser, G. Gambardella, K. O’brien, Y. Umuroglu, et
[311] J. Pavur, I. Martinovic, SOK: Building a Launchpad for Impactful Satellite Cyber- al., FINN-R: an end-to-end deep-learning framework for fast exploration of
Security Research, 2020, 10872. ArXiv, vol. abs/2010. quantized neural networks, ACM Trans. Reconfigurable Technol. Syst. (TRETS)
[312] M.E. Grant, Space Dependence - A Critical Vulnerability of the Net-Centric 11 (2018) 16.
Operational Commander, Naval War College Joint Military Operations Dept, [339] G. Dinelli, G. Meoni, E. Rapuano, G. Benelli, L. Fanucci, An FPGA-based hardware
Newport RI, 17 May 2005. accelerator for CNNs using on-chip memories only: design and benchmarking
[313] J. Pavur, I. Martinovic, The cyber-ASAT: on the impact of cyber weapons in outer with intel movidius neural compute stick, International Journal of Reconfigurable
space, in: Proceedings of 2019 11th International Conference On Cyber Conflict, Computing (2019), 7218758, 2019/10/22 2019.
CyCon), 2019, pp. 1–18. [340] G. Chen, W. Choi, X. Yu, T. Han, M. Chandraker, Learning efficient object
[314] G. Falco, The vacuum of space cyber security, Proceedings (2018). detection models with knowledge distillation, in: Presented at the Proceedings of
[315] S. Gálik, S. Gáliková Tolnaiová, Cyberspace as a New Existential Dimension of the 31st International Conference on Neural Information Processing Systems,
Man, 2019. 2017. Long Beach, California, USA.
[316] C.P. Gonçalves, Cyberspace and Artificial Intelligence: the New Face of Cyber- [341] V. Kothari, E. Liberis, N. Lane, The final frontier: deep learning in space, in:
Enhanced Hybrid Threats, 2019. Proceedings of the 21st International Workshop on Mobile Computing Systems
[317] K. Thangavel, J. Plotnek, A. Gardi, R. Sabatini, Understanding and investigating and Applications, 2020.
adversary threats and countermeasures in the context of space cybersecurity, in: [342] W. Li, M. Liewig, A Survey of AI Accelerators for Edge Environment, World
Presented at the IEEE/AIAA 41st Digital Avionics Systems Conference, DASC), Conference on Information Systems and Technologies (WorldCIST 2020): Trends
2022. Portsmouth, VA, USA. and Innovations in Information Systems and Technologies, pp. 35–44.
[318] E.L. Pawlikowski, Doug, Tom Cristler, Space: disruptive challenges, new [343] R. Cellan-Jones, Stephen Hawking warns artificial intelligence could end
opportunities, and new strategies 6 (2012). mankind, Available at: https://www.bbc.com/news/technology-30290540#:~:
[319] H. Report, Space Domain Mission Assurance: A Resilience Taxonomy, 2015. text=Prof%20Stephen%20Hawking%2C%20one%20of,end%20of%20the%20h
[320] R. Peldszus, Resilience of space systems: principles and practice, in: Handbook of uman%20race.%22, 2014.
Space Security: Policies, Applications and Programs, Springer International [344] E. Commission, Have your say: European expert group seeks feedback on draft
Publishing, Cham, 2019, pp. 1–17. ethics guidelines for trustworthy artificial intelligence, Available at: https://ec.
[321] A.T.O. Report, Resilience for Space Systems: Concepts, Tools and Approaches, europa.eu/digital-single-market/en/news/have-your-say-european-expert-gro
2017. up-seeks-feedback-draft-ethics-guidelines-trustworthy, 2018.
[322] Z. Zhang, H.A. Hamadi, E. Damiani, C.Y. Yeun, F. Taher, Explainable artificial [345] S.T. Todd, J. Burke, Emerging legal issues in an AI-driven world, Available at:
intelligence applications in cyber security: state-of-the-art in research, IEEE https://gowlingwlg.com/en/insights-resources/articles/2019/emerging-legal-iss
Access 10 (2022) 93104–93139. ues-in-an-ai-driven-world/, 2019.
[323] P.M. Winter, S. Eder, J. Weissenböck, C. Schwald, T. Doms, T. Vogt, et al., Trusted [346] Outer Space Treaty, Treaty on Principles Governing the Activities of States in the
Artificial Intelligence: towards Certification of Machine Learning Applications, Exploration and Use of Outer Space, Including the Moon and Other Celestial
2021 arXiv preprint arXiv:2103.16910. Bodies, 1966.
[324] M. Freed, P. Bonasso, M. Ingham, D. Kortenkamp, B. Pell, and J. Penix, "Trusted [347] T. Graham, K. Thangavel, A.-S. Martin, New challenges for international space
autonomy for spaceflight systems", in 1st Space Exploration Conference: law: artificial intelligence and liability, Proceedings of the 17th International
Continuing the Voyage of Discovery.AIAA 2005-2521. 1st Space Exploration Conference On Space Operations, Dubai (United Arab Emirates), March 2023.
Conference: Continuing the Voyage of Discovery. January 2005.

36