1

Aerial Computing: A New Computing Paradigm,

Applications, and Challenges
Quoc-Viet Pham, Member, IEEE, Rukhsana Ruby, Member, IEEE, Fang Fang, Member, IEEE,
Dinh C. Nguyen, Member, IEEE, Zhaohui Yang, Member, IEEE, Mai Le,
Zhiguo Ding, Fellow, IEEE and Won-Joo Hwang, Senior Member, IEEE

Abstract—In existing computing systems, such as edge com- power, and processing capacity. To facilitate data transmission
puting and cloud computing, several emerging applications and from IoT devices worldwide, several studies have investigated
practical scenarios are mostly unavailable or only partially im-
arXiv:2204.02005v1 [cs.NI] 5 Apr 2022

aerial access networks (AANs), which can offer line-of-sight

plemented. To overcome the limitations that restrict such applica-
tions, the development of a comprehensive computing paradigm (LoS) communication links, favorable channels, and improved
has garnered attention in both academia and industry. However, coverage. An AAN generally consists of three primary compo-
a gap exists in the literature owing to the scarce research, and nents: low-altitude platforms (LAPs), high-altitude platforms
a comprehensive computing paradigm is yet to be systematically (HAPs), and low-Earth-orbit (LEO) satellites [2]. These plat-
designed and reviewed. This study introduces a novel concept, forms fully complement the conventional terrestrial access net-
called aerial computing, via the amalgamation of aerial radio
access networks and edge computing, which attempts to bridge work to create a future access network in 6G that can provide
the gap. Specifically, first, we propose a novel comprehensive com- wireless services with global coverage and diverse quality of
puting architecture that is composed of low-altitude computing, service (QoS). Moreover, to better support emerging compute-
high-altitude computing, and satellite computing platforms, along intensive applications, such as fully autonomous vehicles,
with conventional computing systems. We determine that aerial flying taxis, holographic communications, and virtual reality
computing offers several desirable attributes: global computing
service, better mobility, higher scalability and availability, and si- (VR), edge computing (e.g., fog computing and multi-access
multaneity. Second, we comprehensively discuss key technologies edge computing (MEC)) is a promising concept owing to the
that facilitate aerial computing, including energy refilling, edge availability of powerful computing and storage resources at the
computing, network softwarization, frequency spectrum, multi- edge of the network [3]. The amalgamation of AANs and edge
access techniques, artificial intelligence, and big data. In addition, computing introduces a novel concept referred to as aerial
we discuss vertical domain applications (e.g., smart cities, smart
vehicles, smart factories, and smart grids) supported by aerial computing. Aerial computing is expected to provide advanced
computing. Finally, we highlight several challenges that need to services, such as communication, computing, caching, sensing,
be addressed and their possible solutions. navigation, and control, at a global scale.
Index Terms—6G, Aerial Access Networks, Aerial Computing, Aerial computing is considered a promising paradigm that
Edge Computing, Enabling Technologies, Vertical Applications. enables local data analysis and real-time service provisioning
in the air. Facilitated by advantages, such as high mobility,
I. I NTRODUCTION fast deployment, global availability, scalability, and flexibil-
Research communities have started to define roadmaps for ity, aerial computing complements conventional computing
sixth-generation (6G) networks [1]. As various Internet-of- paradigms (e.g., cloud computing, fog computing, and MEC)
Things (IoT) applications are set to emerge in 6G, additional and is thus considered a pillar of the comprehensive computing
data need to be collected and transmitted. However, IoT de- infrastructure in future 6G networks [4]. For example, in [5],
vices are constrained by factors, such as batteries, transmission IoT data over small areas could be collected and processed
cooperatively by a set of unmanned aerial vehicles (UAVs) at
Quoc-Viet Pham is with the Korean Southeast Center for the 4th Industrial low altitudes, whereas IoT data over large-scale areas could
Revolution Leader Education, Pusan National University, Busan 46241, Korea.
Rukhsana Ruby is with College of Computer Science and Software Engi- be executed by HAPs at high altitudes. In [6], UAVs were
neering, Shenzhen University, China. leveraged as intermediate points between vehicles and the
Fang Fang is with the Department of Electrical and Computer Engineering network edge to enable data collection and task preprocessing
and the Department of Computer Science, Western University, London, ON
N6A 3K7, Canada. and thus facilitate data transfer in smart vertical domain
Dinh C. Nguyen is with the School of Electrical and Computer Engineering, environments. Meanwhile, in the race toward global Internet
Purdue University, USA. connectivity among tech giants, such as SpaceX, Google,
Zhaohui Yang is with the Centre for Telecommunications Research, De-
partment of Engineering, King’s College London, WC2R 2LS, UK. Amazon, and Facebook, several LEO satellite systems have
Mai Le is with the Department of Information Convergence Engineering, been deployed to provide Internet services to users, particu-
Pusan National University, Busan 46241, Korea. larly in rural and hard-to-reach areas, with a performance that
Zhiguo Ding is with the School of Electrical and Electronics Engineering,
The University of Manchester, Manchester M13 9PL, UK. is comparable to that of the current terrestrial mobile network
Won-Joo Hwang is with the Department of Biomedical Convergence [7]. However, these systems have not yet been integrated into
Engineering, Pusan National University, Yangsan 50612, Korea. current computing systems as key components.
E-mails: vietpq@pusan.ac.kr, ruby@szu.edu.cn, fang.fang@uwo.ca,
nguye772@purdue.edu, yang.zhaohui@kcl.ac.uk, maile2108@gmail.com, The above observations emphasize the importance of a
zhiguo.ding@manchester.ac.uk, wjhwang@pusan.ac.kr. comprehensive computing architecture for future 6G systems.
 2

As a native constituent, aerial computing is composed of communications [20], HAP networks [21], airborne communi-
low-altitude and high-altitude computing platforms as well as cations [22], satellite communications [23], terrestrial-satellite
satellite computing platforms. It can carry its full complement integrated networks [24], and future aerial networks [25].
of conventional terrestrial computing paradigms. Conceptually, Recently, the primary aspects of aerial radio access networks,
aerial computing can facilitate various smart industrial appli- such as architectures, network design, enabling technologies,
cations, such as smart cities, smart vehicles, smart factories, and applications, were reviewed in [26]. Owing to excellent
and smart grids. Moreover, it is expected to provide a reference features and numerous applications, many researchers believe
6G computing architecture for future studies. that AI techniques are a key component of many 6G systems.
For example, machine learning (ML) is considered a double-
edged sword for privacy in 6G [27]. Moreover, AI is an integral
A. Visions Toward a Comprehensive Computing Infrastructure
part of future networks, such as sensing AI at receivers, on-
In recent years, several surveys on edge computing and device AI, access AI (i.e., transmission at PHY, MAC, and
mobile networks have been conducted. For example, enabling network layers), and data-provenance AI [28].
MEC technologies, including virtual machines (VMs), net- The above surveys mainly concentrate on the KPIs, enabling
work function virtualization (NFV), software-defined network- technologies, potential applications, and major challenges of
ing (SDN), and network slicing were reviewed in [8]. The use 6G. The existing reviews emphasize the importance of a
of MEC for IoT applications, such as critical and massive IoT, comprehensive computing infrastructure for the foundation of
wearable IoT, smart energy, and IoT automotive, was discussed 6G in the next 10 years. However, despite promising studies
in [9]. Computation offloading, resource management, and and an urgent need for a comprehensive computing infrastruc-
optimal placement problems in edge computing systems were ture, a large gap exists in the existing studies because aerial
discussed in [10]–[12]. The amalgamation of MEC and AI, computing has not been systematically reviewed. This paper
referred to as edge intelligence and intelligent edge, has been attempts to bridge this gap by introducing a new computing
reviewed in several studies [13], [14]. Further, certain reviews architecture that can be employed in future network systems.
focused on the integration of edge computing with emerging This paper also presents a comprehensive review of aerial
technologies, such as blockchain [15] and fifth-generation (5G) computing. Table I summarizes the existing surveys on edge
and beyond networks [16]. More recently, privacy/security computing and aerial communications. It also highlights the
issues and solutions of MEC services were reviewed [17]. contributions of the present study.
MEC is a key enabler of many wireless services offering
better QoS. Moreover, MEC has the potential to be closely
integrated with emerging technologies and systems [15], [16]. B. Contributions and Research Methodologies
In addition, these studies elaborate on the various challenges of As discussed above, aerial computing is a promising concept
MEC that need to be examined in beyond-5G networks, such for providing computing resources and wireless services on
as resource optimization, security and privacy, and real-time a global scale. This paper attempts to bridge the gap in the
implementation. With the expected extraordinary populariza- existing literature by presenting an in-depth review of aerial
tion of the next-generation Internet and the emergence of killer computing with a vision toward a comprehensive computing
applications, such as VR, space tourism, fully autonomous infrastructure in future 6G networks. The review covers the
vehicles, holographic communications, and deep-sea sight, the following aspects.
current MEC infrastructure needs to be further expanded both Fundamentals and Designs (Section II): First, to enrich
vertically and horizontally to complement existing edge com- the full understanding of the concept of aerial computing,
puting systems. However, the aforementioned reviews [8]–[17] we present the system architecture, which is illustrated in
primarily focus on 5G network scenarios, technologies, and Fig. 1. This comprehensive computing architecture is expected
applications, but the new requirements for a comprehensive to be the key enabler of many computationally intensive
6G computing infrastructure have not been reviewed yet. applications that will be available in 6G systems. Following
Future 6G systems have been envisioned in recent studies. the analysis of the proposed architecture, we elaborate upon
For example, the authors of [18] discussed their vision for the important features of aerial computing, including ubiquity,
use cases, enabling technologies, and overcoming the chal- mobility, availability, simultaneity, and scalability, and further
lenges of 6G networks. In [19], a set of key performance differentiate between aerial computing from other computing
indicators (KPIs) were defined for 6G, such as peak data rate paradigms. Further, we present important design problems for
>100 Gbps, traffic density >100 Tb/s/km2 , network density an aerial computing system, including computation, commu-
>10 million connections per km2 , mobility >1000 km/h, nication, energy consumption, and latency models.
centimeter-level positioning accuracy, reliability >99.999%, Enabling Technologies (Section III): To realize aerial com-
receiver sensitivity <-130 dBm, 3-times spectral efficiency, puting in 6G systems, we discuss a set of key enabling
and 10-times energy efficiency. Six technical trends were technologies, which include network softwarization, energy
discussed in [19], namely, coverage expansion, new resources refilling, frequency spectrum, multi-access techniques, and AI
and bandwidth spectrum, new modulation techniques, enhanc- and big data. In particular, energy refilling is crucial for main-
ing the system capacity, adding computing and intelligence taining the sustainable operations of aerial components. With
capabilities, and three-dimensional (3D) network architecture. regard to the operational management of aerial computing
Regarding 3D networks, several studies have reviewed UAV systems, we review state-of-the-art studies on softwarization
 3

TABLE I
S UMMARY OF EXISTING SURVEYS ON EDGE COMPUTING AND FUTURE WIRELESS NETWORKS .

Main theme
Reference Key Contributions Limitations
6G/AI MEC AAN
[8] X Enabling technologies of MEC systems, such as VM, contain- This paper does not focus on aerial comput-
ers, NFV, SDN, and network slicing, are reviewed. ing and its role.
[9] X MEC applications for IoT service realization in 5G networks. The role and applications of aerial comput-
ing have not been presented.
[10]–[12] X Reviews of seminal MEC architectures, computation offloading The role and applications of aerial comput-
issues, resource management, and optimal placement problems. ing are not discussed.
[13], [14] X Interactions between AI techniques and MEC systems; that is, The role of aerial computing in enabling a
intelligent edge and edge intelligence. comprehensive computing infrastructure in
6G is not considered.
[15] X Integration of blockchain and MEC systems. Only the integration of blockchain into edge
computing systems is presented.
Integration of MEC with 5G technologies, such as cloud radio Only the integration of MEC with 5G tech-
[16] X
access networks, massive multi-input multi-output (MIMO), nologies is reviewed, while the role and ap-
unmanned aerial vehicle (UAV) communications, and AI. plications of aerial computing are ignored.
[17] X Comprehensive survey on privacy/security issues and solutions The paper reviews only the security and
in MEC systems. privacy aspects of MEC systems.
[18] X A speculative study on promising applications and major tech- Only the vision of future 6G networks is
nologies in future 6G networks. discussed.
[19] X A vision of KPIs, technologies, technical trends, and challenges Only a brief discussion of future 6G net-
of 6G networks. works is presented.
3D networking, such as UAV communications [20], HAP These papers primarily focus on the com-
[20]–[25] X
systems [21], airborne communications [22], satellite commu- munication aspects of AAN, while the com-
nications [23], terrestrial-satellite integrated networks [24], and puting aspects and the role of aerial comput-
wireless communications for future AANs [25]. ing are ignored.
Technical aspects of AANs in the 6G context and key enablers. The computing aspects and applications of
[26] X X
aerial computing are not presented.
[27] X A comprehensive survey of the alliance between ML and 6G The paper focuses only on privacy aspects
privacy. and its alliance with ML.
Four important domains of AI in 6G are reviewed, namely, The paper focuses on the role and applica-
[28] X
sensing AI, on-device AI, access AI, and data-provenance AI. tions of AI in future networks.
A comprehensive survey of aerial computing. In particular,
This paper X X X
• We introduce a novel architecture of aerial computing,
which complements the existing computing infrastruc-
tures. Important features of aerial computing, such as mo-
bility, availability, scalability, flexibility, and simultaneity,
have also been analyzed in detail.
• We provide in-depth discussions of key enablers of aerial
computing and the use of aerial computing in important
vertical applications, including smart cities, smart vehi-
cles, smart factories, and smart grids.
• Key challenges of aerial computing are presented along
with potential solutions and future research directives.

techniques, including VM, network slicing, NFV, and SDN. this regard, we discuss the manner in which aerial computing
These softwarization techniques are important for facilitating can support vertical domain applications.
flexible and programmable operations and the deployment Research Challenges and Potential Directions (Section V):
acceleration of aerial computing systems. In addition, we To further drive research into aerial computing in the future,
discuss frequency spectrum and multi-access techniques (e.g., we discuss the major challenges, possible solutions, and
massive MIMO, intelligent reflecting surface (IRS), and non- potential directives. Aerial computing systems face critical
orthogonal multiple access (NOMA)), used in aerial comput- challenges, such as energy efficiency, efficient resource man-
ing systems. Finally, AI and big data analytics are considered agement, and network stability. Further, large-scale network
important technologies for exploiting massive data generated optimization is another critical issue for embedding distributed
from aerial computing, thus improving system performance. and learning approaches such that aerial computing systems
Vertical Domain Applications (Section IV): The integration can efficiently operate on a large scale. Finally, security,
of wireless technologies into industrial automation systems has privacy, and trust are important aspects of aerial computing
shown great potential in creating more efficient and productive but have not been adequately addressed by existing studies.
vertical domain applications (i.e., smart cities, smart vehicles,
smart factories, and smart grids). Further, we present an II. F UNDAMENTALS AND D ESIGN
extensive review of state-of-the-art studies on these vertical In this section, we present the system architecture and
domain applications in the context of 6G aerial computing. In fundamental features of aerial computing.
 4

Network softwarization

Smart city Smart vehicle Smart factory Smart grid

Vertical domain applications

Satellite computing

Energy harvesting

Frequency spectrum

HAC platform

LAC platform Multi-access techniques

Underground
Terrestrial
computing Marine
communications
AI and Big Data

IoT Devices Key-enabling technologies

IoT Devices IoT Devices

Fig. 1. A novel comprehensive 6G computing architecture.

A. System Architecture data collection efficiency, the UAV offloads the collected data
to a terrestrial MEC server and/or other more powerful UAVs.
As shown in Fig. 1, we propose a novel aerial computing- Terrestrial Computing: This layer comprises conventional
based 6G architecture wherein the aerial components are terrestrial computing paradigms, such as fog computing, MEC,
positioned in a hierarchical manner. In addition to conventional and cloudlet. Typically, computing nodes are deployed at
computing infrastructure, the proposed aerial computing archi- preset locations. According to [16], the potential deployment
tecture consists of four main entities: IoT devices, low-altitude locations of an MEC server include routers, IoT gateways,
computing (LAC) platforms, high-altitude computing (HAC) macro base stations (BSs), small cell BSs, optical network
platforms, and satellite computing platforms, along with ver- units, radio network controller sites, and wireless fidelity
tical domain applications and key enabling technologies. (WiFi) access points, whereas the common locations of a fog
IoT: This layer refers to any IoT device that generates data node are gateways, intermediate nodes (between the cloud and
and comprises compute-intensive applications that need to be end devices), and network elements (e.g., routers and switches)
executed by aerial computing systems. For example, wearable [29]. The collaborative and hierarchical MEC framework uti-
devices, such as smart watches, eyeglasses, fitness trackers, lizes the advantage of many applications and services, as edge
and body-mounted sensors are responsible for monitoring and devices with different capabilities can collaboratively perform
collecting the health information (e.g., heart rate, physical data processing and task execution [30].
activity, and blood pressure). These devices may offload col- Low-Altitude Computing: The LAC platform is positioned
lected data to an edge node for further processing because in the lowest tier of aerial computing systems with an altitude
they are typically resource-constrained and have a relatively of 0–10 km above the Earth’s surface. This platform offers
small size. Aerial components (e.g., UAVs and drones) can be several advantages, such as cost-effectiveness, rapid deploy-
a part of this IoT layer when they are dispatched to collect ment, and LoS communication links. Hence, LAC systems are
data from ground sensors. To maximize the flight time and highly effective for emergency scenarios, military surveillance,
 5

temporary events, and IoT data collection and processing Industrial Applications: Aerial computing supports and
[31]. The authors of [32] demonstrated that the UAV can enables (new) industrial applications owing to its distinctive
be leveraged for data collection, aerial relaying, and aerial features, such as high mobility, always-available computing,
computing of the tasks offloaded from ground IoT devices. and scalability (discussed in Section II-B). For example, aerial
The primary components of the LAC platform are drones and computing systems can support large-scale vehicle systems
UAVs equipped with computing capabilities, which typically and/or advanced traffic signal infrastructures to improve the
have limited battery capacities, low computing resources, and experience of in-car infotainment applications and reduce
low endurance. Consequently, several solutions have been pro- traffic congestion by deploying massive MEC on roadside
posed to address these challenges, such as wireless- and solar- units, power systems, buildings, and highway infrastructure.
powered UAV-MEC and task offloading to more powerful In Section IV, we comprehensively discuss the use of aerial
terrestrial and HAC-MEC servers [33]. There exist two types computing for four primary industrial applications: smart
of LAC servers: fixed-wing and rotary-wing. The former does cities, smart vehicles, smart factories, and smart grids.
not hold up in the air and needs to move forward continuously, Enabling Technologies: The realization of aerial computing
whereas the latter can take off and land vertically. is facilitated by several key enabling technologies, including
High-Altitude Computing: The HAC platform comprises network softwarization, energy refilling, frequency spectrum,
both manned and unmanned aerial components (e.g., airplanes, multi-access techniques, and AI and big data. A thorough
airships, and balloons) that operate at an altitude ranging from discussion of these technologies in the context of aerial
17 to 50 km. The advantages of HAC systems over other computing is presented in Section III.
platforms include large-area coverage with the cell size of up
to 10 km, adaptability to traffic demands, rapid deployment
B. Desirable Features
compared to terrestrial computing and satellite computing,
high endurance compared to LAC platforms, and green op- Aerial computing retains the base features of the concept
eration capability with solar power [34]. At the World Radio of edge computing, including on-premises, proximity, lower
Conference (WRC)-19, several frequency bands were allocated latency, location awareness, and network contextual infor-
to HAC systems by the ITU Radio communication Sector, mation. In addition, it has certain fundamental features that
including the 31–31.3 and 38–39.5 GHz bands for worldwide differentiate it from other computing paradigms (e.g., fog
use, 47.2–47.5 and 47.9–48.2 GHz for administration use, computing and cloudlet), which are discussed as follows.
and 21.4–22 and 24.25–27.5 GHz for fixed services in ITU • Mobility: Aerial platforms can be deployed quickly to
Regions 2 [35]. In recent years, several industrial projects have support on-demand computing applications. While LAC
been observed in LAC platforms for wireless and computing platforms can enable the rapid roll-out of computing
services, such as flying cells on wings [36], Google Skybender services, the deployment of a terrestrial computing system
[37], and ApusDuo [38]. Owing to the quasi-stationary feature, is time-consuming owing to planning procedures and
HAC platforms can form a collaborative HAC-MEC server civil works involved. From the perspective of the user,
with more powerful capabilities and higher time endurance, the overlaying architecture of LAC, HAC, and satellite
aiding in the execution of the computation tasks that are computing helps aerial computing systems flexibly satisfy
offloaded from LAC and/or satellite systems. different computation tasks from end IoT devices.
Satellite Computing: This layer is composed of LEO • Availability: Aerial computing systems are usually avail-
satellites with computing capabilities. LEO satellites typically able as they are not much affected by natural disas-
operate at an altitude of approximately 80–2000 km above ters, unlike terrestrial computing, and can be easily and
the Earth’s surface. For example, SpaceX launched 60 LEO quickly deployed for particular computation tasks. The
satellites at altitudes not greater than 580 km in May 2019. availability of aerial computing also relies on its hierar-
These satellites can provide Internet services globally with a chical architecture with different computing platforms at
data rate ranging from 50 to 150 Mbps and a latency ranging different altitudes and coverage distances.
from 20 to 40 ms (www.starlink.com/). Amazon also initiated • Scalability: Aerial computing systems can provide on-
a satellite project called Kuiper, which plans to launch a demand computing services as well as applications with
constellation of 3200 satellites at an altitude of approximately geographically distributed users. As the aerial compo-
610 km [39]. These LEO satellites are expected to provide nents can collaboratively form computing clusters, mas-
various wireless and computing services, particularly when sive IoT devices can be served, thereby improving user
other computing platforms are temporarily unavailable and/or satisfaction and guaranteeing service continuity.
satellite computing is the only option. Unlike in LAC and • Flexibility: Unlike terrestrial computing, wherein MEC
HAC platforms, users are not usually connected with satellite and fog nodes are typically deployed at pre-specified
computing systems because of the excessive hardware cost locations, aerial computing entities can be changed easily
incurred when MEC servers are embedded in LEO satellites. to flexibly suit different situations. Further, enhanced
Instead, LEO satellites can act as relays to receive and transmit computing services may be provided via additional aerial
computation tasks/results between users and MEC servers. computing platforms, such as LAC and HAC.
However, computing directly at the satellite edge server is • Simultaneity: Different aerial computing platforms can
beneficial for applications with sparse users; that is, users are provide computing services to geographical users simul-
located at different geographical locations [24], [40]. taneously. Thus, users in different countries can request
 6

TABLE II
C OMPARISON OF AERIAL COMPUTING WITH OTHER EDGE COMPUTING PARADIGMS .

Cloudlet Fog Computing MEC Aerial Computing

Introduced Satyanarayanan, 2009 [41] Cisco, 2012 ETSI, 2014 This work, 2021
by
Purpose Moving cloud computing capabilities to the network edge Making computing capabilities
available at all times, at both
the network edge and air
Deployment Business premises (e.g., shop- Strategic locations both indoors Various locations within radio Locations in the air and within
ping malls, companies, and and outdoors (e.g., IoT gate- access networks (RANs) (e.g., RANs (e.g., radio towers,
personal computers) ways, routers, and switches) radio towers and gateways) UAVs, and LEO satellites)
Application Locally resource-intensive and Smart vertical domain appli- Content caching, autonomous On-demand applications and
Examples interactive applications cations (e.g., smart cities and vehicle, augmented reality applications with sparse user
smart health) and video surveil- (AR), data analytics distribution
lance
Internet Autonomous operation (e.g., Autonomous operation or inter- Autonomous operation or mo- Autonomous operation or mo-
Connectivity WiFi) mittent Internet connectivity bile Internet bile Internet
Service Local Less global Less global Global
Coverage
Latency Low Low Low Medium
User Low Relatively low Low Various, from low to long
Proximity
Architecture Localized Distributed/hierarchical Localized/hierarchical Localized/hierarchical/on-
demand
Storage Dependent on deployment scenarios [29], [42] High High
Capacity
Computational Dependent on deployment scenarios [29], [42] High High
Power
Power Low [43] Low [43] High [43] Low
Consumption
Availability High [44] Very high
Scalability Low High [44] Very high
Mobility Low [44] High [44] Very high
Security Medium High
Context Medium High High High
awareness
Local High [17]
Awareness
Standard NIST OpenFog Consortium ETSI, 3GPP, ITU-T -
Organizations
Virtualization VM VM, NFV, network slicing, and other virtualization technologies
support
Operation Standalone only Cloud-connected support Both cloud-connected and stan- Both cloud-connected and stan-
Mode dalone operation dalone operation

computing services to LEO satellite computing services routers) for further processing. The integration shows great
concurrently. Therefore, aerial computing is regarded as potential to respond to heterogeneous network scenarios and
a full complement of terrestrial computing infrastructure, application requirements in 6G, but several challenges and
which is usually used for localized computing services. limitations should be addressed [45].
After analyzing these features, we draw a comparison be-
tween aerial computing and other edge computing paradigms, C. Network Design
as presented in Table II. Based on the above explanations,
This part explains the important models for designing and
aerial computing can support other computing concepts, such
optimizing an aerial computing system.
as in-network computing and cloud-native computing. How-
ever, resource and operation constraints of aerial computing 1) Computation Model: The first component of aerial
limit its capability to enable high-performance services in in- computing involves computing models with the following
network computing and scalable applications in cloud-native fundamental aspects:
computing. Therefore, the integration of aerial computing with • computation offloading decision,

edge computing, in-network computing, and AI techniques is • deployment locations of edge servers,

necessary to provide advanced services and applications in the • collaborative and hierarchical computation.

future. For example, network elements in aerial computing, as Computation offloading decision: A computation task I
edge devices, can partially perform computation tasks before can be modeled by a tuple I = (C, D, T ), where C is the
pushing them into the network elements (e.g., switches and computation workload (CPU cycles per bit), D is the task
 7

LAPs, which are mainly responsible for data collection from

ground IoT users [33], [46].
2) Communication Model: The communication model
Layer 2: plays an important role in aerial computing systems for
HAC platform guaranteeing efficient communications among computing plat-
HAC platform: processes the LAC platform: offloads the forms (e.g., between IoT devices and LAC servers). Several
computation tasks received tasks to the HAC platform communication models have been proposed for this purpose.
from the LAC platform for remote execution
In the literature, orthogonal multiple access (OMA), NOMA,
and MIMO have been typically adopted for the achievable
rate formula. When OMA (e.g., time division multiple access
UAV
(TDMA), frequency division multiple access (FDMA), and
Layer 1: LAC platform orthogonal frequency-division multiple access (OFDMA)) are
widely used for ease interference management. In addition,
LAC platform: data IoT devices: sensory data there are studies on hybrid OMA-NOMA computing systems
collection from IoT devices
[47]. They demonstrated that hybrid NOMA systems are
highly advantageous compared to conventional OMA systems.
Thus, the use of such hybrid OMA-NOMA approaches in
aerial computing systems is a promising approach.
3) Energy Consumption Model: Energy efficiency is an
important design requirement for aerial computing systems.
Layer 0: Ground IoT users
Energy is consumed in several aspects, including local compu-
tation, communication, remote computing, and operation. The
other part is the energy consumed by the aerial components
Fig. 2. Collaborative computation between LAC and HAC platforms.
(i.e., aerial users and servers), which is important for designing
sustainable operations of aerial computing systems, particu-
larly LAC platforms with limited onboard batteries. However,
size (bits), and T is the completion time deadline (seconds). various factors, such as flying modes, flying speed, payload,
Fundamentally, an offloading decision results in either local and external conditions (e.g., weather and wind speed) can
computing at the devices, full execution at the edge server, or affect the energy model of aerial components. In an effort
partial offloading at both. In the case of partial offloading, the to incorporate this, the authors of [48] modeled the power
task size D is divided into smaller parts for local execution and consumption of multirotor helicopters in hover and vertical
remote execution at the edge servers and/or nearby devices. moving modes. The power consumed by the UAV to remain
Deployment location: Computation in aerial computing may afloat in the horizontal moving mode is modeled in [49]. These
have the following deployment location possibilities: 1) local models have been studied to improve the energy and power
devices, 2) terrestrial edge servers, 3) LAC servers, 4) HAC efficiency of aerial computing [50]. Moreover, the possibility
servers, 5) LEO satellites, and 6) centralized clouds. Each of aerial computing in terms of energy and power consumption
deployment location of the edge servers has both advantages can be realized by integrating aerial computing with emerging
and disadvantages. For example, the first scheme (i.e., local technologies in 6G, as discussed in Section III.
computing at the end devices) does not incur any transmission 4) Latency Model: A benefit of aerial computing is the
latency, but poor computing resources render it unsuitable incorporation of LEO satellite computing platforms, which
for many emerging applications and new services, such as provide a global coverage but have lower latency compared
extended reality (XR) and smart cities. Further, computation to geostationary Earth orbit (GEO) and medium Earth orbit
and content caching at the satellite computing platforms (e.g., (MEO) platforms. As reported in [51], the OneWeb LEO
LEO satellites) is highly beneficial to applications with users satellite system can provide Internet services for 1 billion
in remote areas (e.g., mountains and islands) but might result users with an average latency of 30 ms. The Starlink system
in a higher latency and hardware cost compared to those of promises a latency between 25 and 35 ms, while the Telesat
terrestrial computing and LAC/HAC platforms [24], [40]. system aims to achieve latency from 30 to 50 ms, and the
Collaborative and hierarchical computation: However, re- average latency of the SpaceX system is less than 20 ms.
liance on only an aerial computing node/platform (i.e., LAC Moreover, positioned in the lowest tier of the comprehensive
or HAC only) to perform emerging compute-intensive applica- aerial computing architecture, LAC platforms are expected to
tions is highly impractical. This is due to the onboard energy provide much lower latency and thus are suitable for latency-
limitations of LAC platforms, intermittent harvested energy critical services. Further, in the middle tier, HAC platforms
of HAC platforms, and hardware cost of satellite computing aim to balance latency and computing service availability.
platforms for big data applications. The synergy between LAC Therefore, aerial computing can provide computing services
and HAC platforms in aerial computing systems is discussed with different latency requirements while offering desirable
in [33]; herein, it is illustrated in Fig. 2. In such a scenario, the features, such as mobility, availability, scalability, flexibility,
more powerful HAC platform and fixed-wing LAPs can assist and simultaneity.
the long-term operation and computation tasks of rotary-wing Other latency sources in aerial computing include com-
 8

putation, queuing, and fronthaul and backhaul transmission. equipment via virtualized network functions (VNFs). A typical
Latency also occurs when the server executes the computation NFV architecture consists of three main elements: infrastruc-
task remotely and broadcasts the processed results to the ture, management, and orchestration. The NFV infrastructure
users in the downlink. Consequently, aiming to minimize represents computing, storage, and network hardware as well
the maximum latency among users, [52] investigated a joint as software (abstractions of physical resources) resources
problem of two-dimensional LAC placement, fronthaul and constituting the environment where VNFs are deployed. In
backhaul bandwidth allocation, computing resource allocation, addition, a VNF is an implementation of a network function,
and caching decision. Further, in [53], the joint optimization for example, a firewall, deployed on virtual resources such
of task allocation, scheduling, power control, and LAC server as VMs. Furthermore, NFV-Management and Orchestration
trajectory was studied to minimize the total energy consump- (MANO) is the management and orchestration framework
tion. To manage the latency requirement, the LAC server must needed to provide the VNFs.
execute a minimum number of bits offloaded from each user. NFV provides many opportunities for network service pro-
visioning, e.g., independent deployment and maintenance of
D. Summary software and hardware, flexible and agile service deployment,
and dynamic resource allocation. Therefore, NFV can benefit
This section presents the system architecture of aerial aerial computing. For instance, NFV can allocate additional re-
computing, fundamental features, and key design aspects. sources from another LAC/HAC/satellite computing platform
Complementing conventional terrestrial computing systems, to alleviate computing congestion in one platform.
the proposed aerial computing framework provides computing 2) Software-Defined Networking: SDN is an emerging net-
services to massive IoT users worldwide. Compared to other work architecture that intelligently controls the network by
edge computing paradigms, aerial computing offers several decoupling the control plane and efficiently managing the
desirable features, such as mobility, availability, scalability, network via programming operations [54]. Driven by its key
flexibility, and simultaneity. Moreover, four key design aspects features, including separated control plane and data plane,
of aerial computing have been reviewed, namely, computation, centralized controller, open interfaces, and programmable op-
communication, energy consumption, and latency. Because erations, an SDN can intelligently operate the network at a
many emerging applications will be available in the near low operating cost. Thus, it can support intelligent applications
future, more thorough investigations are required to explore through simplified hardware, software, and management [8].
the potential advantages of aerial computing. Despite these advantages, SDN has several key challenges in
the 6G network: 1) efficient and intelligent maintenance of
III. E NABLING T ECHNOLOGIES the dynamic network topology [55], 2) intelligent network
Enabling technologies of aerial computing are presented in management and orchestration via AI/ML, and 3) traffic
this section. These technologies include network softwariza- engineering with heterogeneous QoS requirements [56].
tion, energy refilling, edge computing, frequency spectrum, In aerial computing, wireless networks support heteroge-
multi-access techniques, AI, and big data. To enrich the neous IoT devices/users, LAC/HAC platforms, 6G technolo-
understanding of the integration of aerial computing with gies, and configuration interfaces. SDN can provide logically
enabling technologies, we illustrate the manner in which centralized control by abstracting the underlying network
enabling technologies support aerial computing in Fig. 3. infrastructure. For instance, SDN controllers can handle aerial-
computing-related VNFs and VMs as another type of resource
that can be dynamically managed. Therefore, SDN can provide
A. Network Softwarization an efficient solution for managing aerial computing systems
Virtualization has enabled network operators to design, and services.
implement, and manage systems and network services with 3) Network Slicing: Network slicing is a virtualization
improved efficiency. NFV and SDN are two promising tech- technology that allows multiple logical networks (slices) to run
nologies that facilitate virtualization. NFV decouples net- on a unified physical network infrastructure [57]. Each logical
work functions from proprietary hardware (e.g., firewalls and network is independently configured to satisfy the required
routers), thereby providing equivalent network functionality network characteristics, such as bandwidth, delay, and capac-
with general-purpose servers. SDN separates the control plane ity, to provide diverse services of expected scenarios. Each
from the forwarding plane, thereby enabling network operators logical network also contains computing and storage resources
to configure and manage network functions in a centralized that are capable of realizing specific network functions via
manner through the software. Other commonly used virtu- NFV or service function chains. In contrast, VM can provide
alization technologies include network slicing and VM. We fine-grained control to instantiate and terminate tasks and
briefly introduce the fundamentals of these technologies and processes at any time without affecting the hardware on aerial
thereafter discuss their application in aerial computing. platforms. Each VM shares computing, storage, memory, and
1) Network Function Virtualization: The concept of NFV network resources from aerial computing systems, while its
was first proposed in 2012, with an aim to reduce the capital operation is entirely isolated from that of the host and guest
and operating expenses of telecom networks, as well as to VMs. Further, AI/ML can be employed to intelligently allocate
facilitate the deployment of new services. Its primary idea is virtual resources of VMs in aerial computing.
to decompose the network functions from physical network In aerial computing, network slicing can be utilized to
 9

Energy harvesting

AI and Big data

THz / THz /Sub 6GHz
Sub 6GHz
AI/ML AI/ML

AI
Internet IRS
NFV Infrastructure

NOMA NOMA
AI/ML
Power

Vertical Domain
Massive Applications
Frequency
NOMA MIMO

NOMA

Fig. 3. Integration of aerial computing with enabling technologies.

slice the entire network into individual networks considering 2) Wireless Power Transfer: Another popular application of
computing services and IoT services, which can be optimized energy harvesting is that energy is usually harvested from am-
by specific requirements and services. However, an AI-enabled bient radio frequency electromagnetic signals. In simultaneous
intelligent management and orchestration framework should be wireless and information power transfer (SWIPT), energy can
designed to efficiently support network slicing in 6G networks. be harvested, usually partially, from the transmitted signals via
power switching or time switching techniques. A strong user
with a high channel gain can forward data to a weak user with
B. Energy Refilling a low channel gain by using the harvested energy from the
transmitted signals. Wireless power communication networks
1) Energy Harvesting: Energy harvesting is a novel ap- (WPCNs) are used to harvest energy from radio transmission,
proach to utilize renewable energy sources. For example, solar wherein sequential energy is transferred wirelessly to the
energy can be converted into electric energy by deploying intended users, which supports dedicated wireless charging
photovoltaic (PV) cells on UAVs, and the electric energy [61]. Aerial components at different computing platforms (e.g.,
can be stored in rechargeable batteries [58]. However, natural UAVs at the LAC platform and balloons at the HAC platform)
energy sources are highly dependent on climate variability. For have been proposed to be integrated with WPCN. This flexible
instance, solar power harvesting can be severely affected by implementation allows aerial components to be charged or
weather conditions. to charge a set of ground users, thus providing sustainable
Some non-electromagnetic field-based charging technolo- solutions for aerial computing systems [62], [63]. Besides,
gies have been used, including gust soaring, PV arrays, laser aerial components can be equipped with solar-powered cells
beaming, and battery dumping [59]. Dynamic soaring involves and large energy storage to store energy harvested from solar
harvesting energy from wind and airflow by adjusting the powers in the environment.
trajectory of UAVs to prolong their flying duration [60]. For
PV arrays, solar irradiation energy is harvested by PV cells to Wireless power transfer (WPT) has been proven to provide
power a drone or recharge its battery, which is used for the power for UAVs wirelessly without landing for energy re-
operation during the night. This technology can be applied fueling [64]. Moreover, the coordination of LAC servers at
to fixed-wing drones because a certain amount of space is the LAC platform and edge servers on different platforms
needed for PV panels. In laser-beaming technology, a laser (i.e., LAC, HAC, and satellite) is a promising approach
fed by an external source of energy produces a concentrated to better provide computing services over large-scale areas
and streamlined beam of light with a certain frequency to the [63]. Since WPT from the air is quite limited due to the
specific PV cells on a drone. The PV cell can then convert the severe path loss, many factors (e.g., antenna size) need to be
laser beam into energy to recharge the battery of the drone. considered for practical implementation. To further support
 10

low-power IoT/mobile devices, emerging technologies, such learn, and predict various communication-related parameters,
as backscatter communications and IRS, can be deployed. For such as wireless channels, traffic patterns, user context, and
example, [65] considered IoT devices as backscatter-assisted device locations. In general, ML algorithms include supervised
users, which harvest energy from the ground gateway to learning, unsupervised learning, reinforcement learning (RL),
perform computation offloading. Similarly, aerial components deep learning (DL), deep reinforcement learning (DRL), and
can assist low-power users in performing necessary data federated learning (FL) [69].
transmissions and computation tasks. 1) Supervised Learning/Unsupervised Learning: In super-
vised learning algorithms, both the input and desired outputs of
C. Frequency Spectrum the used datasets are available. However, supervised learning
algorithms can only be employed in scenarios with sufficient
1) Sub-6 GHz and mmWave: To meet the extremely high
labeled data, for example, classification. In contrast, in unsu-
data rate requirements of 5G systems, high-frequency spec-
pervised learning, the dataset used for training does not include
trum bands are exploited. In particular, millimeter waves
labeled output or target values. The purpose of unsupervised
(mmWave) can provide gigabit-per-second data rates owing
learning is to extract key features of the data for better
to the radio frequency spectrum in the range of 30–300 GHz,
prediction. Further, unsupervised learning algorithms can be
whereas the sub-6 GHz range can provide large coverage with
applied to scenarios, such as clustering based on the available
high cost efficiency with a spectrum band below 6 GHz [66].
data. In aerial computing, supervised learning/unsupervised
Driven by the increase in IoT devices, such as XR services en-
learning can be used for user clustering/grouping to provide
compassing AR, mixed reality, and VR, some studies proposed
better service.
the integration of mmWave and Sub-6 GHz to balance the
2) Deep Learning: DL algorithms have been developed
high data rate and large coverage in next-generation wireless
to deal with complex input–output mappings. DL consists
communication systems. Specifically, mmWave can be used in
of multiple layers for feature extraction and transformation.
tiny cells for mobile and fixed access, whereas sub-6 GHz can
DL can be used for in-depth analysis in a complex scenario
be used in small cells. For example, the use of the mmWave
with massive data and to realize different control schemes for
band in aerial computing was considered in [67], wherein a
different protocol layers. For example, in [70], DL was used
hierarchical architecture of a satellite computing server, two
to classify cloud images captured by satellite clouds in aerial
terrestrial MEC servers, and a set of vehicles was considered.
computing systems.
2) THz: Terahertz (THz) band, in the frequency range
3) Reinforcement Learning: The main idea of RL is to train
of 0.1–10 THz, has been explored in the next-generation
the agent to generate actions according to the current environ-
wireless communication systems to achieve the terabits per
ment. In RL, the problems are solved by employing a sequence
second data rate with low latency. Potential use cases for
of actions that use the trial-and-error rule. RL algorithms
THz communications include close proximity communication,
have been extensively used in wireless network optimization
extremely high data rates for indoor communications, and
to obtain the optimal policy, for example, user grouping
wireless backhauling and fronthauling technologies. Thus,
decisions or actions. DRL is a DL framework developed
THz can be utilized in LoS and high data transmission among
based on RL; it relies on updated samples in practice instead
aerial components in aerial computing systems [68].
of the ideal transition probability in theory. DRL involves
3) Other High-Frequency Bands: At WRC-12, the ITU
learning from the feedback that evaluates the actions taken
identified the C-band for the use of control and non-payload
rather than learning from the correct actions. DRL algorithms
communication. The C-band refers to a frequency range of
have been applied to wireless networks for multiple aspects,
4–8 GHz and is used primarily for satellite communications.
including mobile networking, resource allocation, schedular
Moreover, the K-band in the frequency range of 14–27 GHz
design, and routing [71]. Furthermore, it can promptly make
is typically used for short-range communications owing to
a decision under dynamically changing network conditions,
the high atmospheric attenuation. The K-band can be used
such as channel state information [72].
in satellite communications, astronomical observations, and
4) Federated Learning: FL is an AI approach that enables
radar. Radars in this frequency range provide a short range
users to collaboratively learn a shared model with data main-
and high throughput.
tained on their own devices [73]. In contrast to a standard
ML algorithm that requires a centralized training dataset in the
D. Artificial Intelligence and Big Data Analytics data center, FL allows devices to train a learning model locally
AI technology aims to train machines to perform human and transmit the training parameters instead. Thus, it addresses
tasks. It has been applied to various areas, such as image concerns, such as user privacy and limited data transmission
recognition, robotic vehicles, machine translation, and game resources (e.g., bandwidth) [62], [73]. A potential application
AI. Further, ML, which is a promising subset of the AI of aerial computing for FL-enabled wireless networks is aerial
technique to learn from the data and impart intelligence aggregation [74]; that is, a computing server (e.g., LAC,
to existing systems, has been employed to render wireless HAC, and satellite computing) in aerial computing acts as the
communication and networks highly efficient and adaptable. aerial server and performs model aggregation in the air, as
Intelligent wireless network can provide efficient support for shown in Fig. 4(a)]. Moreover, as shown in Fig. 4(b), aerial
aerial computing. The advantage of applying ML in wireless computing can extend the learning coverage of terrestrial FL-
network operation is that it enables the network to monitor, enabled networks, particularly when certain obstacles prevent
 11

LAC Satellite Terrestrial BS

platform computing
Aerial platform
Model
aggregation server
MEC
in the air
Server
Aerial relay
HAP
Ground
station

Local
training

Local
FL
data Obstacles
Users Local data Local data

a) FL model aggregation in the air b) FL model relaying in the air

Fig. 4. Applications of aerial computing for FL-enabled wireless systems [2].

model updating and broadcasting between FL users and the nications and signal processing computation task control for
aggregation server. Further, aerial components (e.g., LAC, increasingly complex and heterogeneous networks.
HAC, and satellites) can be FL users, which train local AI
models and share updates with the central cloud [75]. E. Other Wireless Techniques
5) Big Data: The features of big data have been commonly 1) Non-orthogonal Multiple Access: NOMA is considered
described by five “V”s: volume, variety, velocity, veracity, and a promising technology in 5G and beyond owing to its high
value. Certain attempts have been undertaken to apply big spectrum efficiency and massive connectivity. It is broadly
data analytics to wireless communications [76]. The increas- classified into two categories: power-domain NOMA and
ing complexity of networks and complicated wireless traffic code-domain NOMA. Power-domain NOMA uses the power
patterns make big data analytics appealing. ML algorithms domain for multiple access, whereas code-domain NOMA
coupled with edge computing can be used to process big data exploits sparse code for multiple access. NOMA can support
in 6G [77]. MEC has become the primary computing method more users than the number of available subcarriers, thereby
for big data analysis in heterogeneous 6G. improving wireless communication performance with multiple
AI and big data, combined with other enabling technologies features, including massive connectivity, low latency, high
mentioned in this section, have been widely introduced to spectral efficiency, and high energy efficiency.
aerial computing in 6G owing to the advantages of intelligent • First, as NOMA enables multiple users to be multiplexed
management and automatic self-improvement. In particular, on the same channel simultaneously, it is suitable for a
AI/ML can be used in intelligent UAV/drone trajectory design, scenario with a large number of connections (e.g., IoT
virtual resource management, data processing, task computa- applications). Therefore, NOMA can be applied in aerial
tion, and channel estimation. In addition, AI and big data can computing to support multiple devices/users to offload
be used in NFV and SDN-based networks to achieve intelligent their tasks to aerial servers, as shown in Fig. 3.
network management and optimization [78]. RL approaches • Second, compared with OMA where users must wait until
have also been conducted to improve resource utilization and a resource block is available to transmit or receive data,
service provision [79]. Further, advantages of the other ML NOMA can provide grant-free transmission with flexible
techniques, such as meta learning and transfer learning, can be scheduling. Therefore, the offloading/downloading delay
exploited. For example, the benefits of transfer learning with of the devices/users in aerial computing can be signifi-
knowledge preprocessing in LAC systems have been demon- cantly reduced.
strated in [80]. However, major challenges should be answered • Third, NOMA can achieve higher spectral efficiency and
to further improve the effectiveness of transfer learning in user fairness than OMA [86]. Further, NOMA users
heterogeneous aerial computing systems. More specifically, we can utilize the entire frequency bands for transmission,
need to determine the source agent and transferred knowledge, whereas OMA users can only use a fraction of the entire
and optimize scheduling and resources. spectrum for communication. In aerial computing, high
Certain challenges need to be addressed to realize aerial spectrum efficiency can also be achieved in the communi-
computing in 6G, including data security/privacy, learning effi- cations between LAC/HAC platforms when using a high-
ciency, communication cost, and the tradeoff between learning frequency spectrum [67].
accuracy and convergence. AI/ML is recognized as the most 2) Massive MIMO: The concept of massive MIMO has
promising solution for providing intelligent wireless commu- been proposed to drastically increase data rates, spec-
 12

TABLE III
S UMMARY OF ENABLING TECHNOLOGIES OF AERIAL COMPUTING .

Category Enabler Advantages Key Challenges

Network NFV High flexibility and low operational cost Service heterogeneity
Softwarization VM Sufficient computing resources Intelligent management of virtual resources
[54]–[56], Network slicing Flexibility, resource efficiency and Security Intelligent management and orchestration framework design
[81]–[83] SDN High flexibility and low operational cost SDN controller placement
Energy Refilling Energy harvest- Ambient natural energy source Highly dependent on weather conditions
[58], [59], [61], ing
[63], [84], [85] WPT (SWIPT Harvest energy from radio transmission Trade-off between harvested energy and consumed power and
and WPCNs) joint resource allocation
Frequency Sub-6 GHz High bandwidth and high data rate Severe attenuation and blockage and low range
Spectrum [66], THz High bandwidth and high data rate Severe attenuation and blockage and low range
[68] C-Band K-Band Satellite communications and high through- Atmospheric attenuation
put
Multiple Access NOMA High spectral efficiency and massive con- Decoding complexity for multiple users
Techniques nectivity
[86]–[90] Massive MIMO High statistical multiplexing gain and high High implementation cost, intelligent environment adaption
spectral efficiency and channel prediction
IRS High channel gain and low implementation Intelligent optimization of IRS
costs
AI/Big Data ML Automatic features, predictions Computational complexity
Analytics [69], DL Automatic features, predictions Computational complexity
[71]–[73], RL Generate optimal decisions in a dynamic en- High dimensionality of the states and action spaces
[76]–[79] vironment and require the Markovian model
and high computation
FL Protects user data privacy Communication cost and tradeoff between learning accuracy
and convergence
Big data Prediction of user preference distribution Feature extraction and data modeling
analytics

tral/energy efficiency, and coverage of wireless networks by wireless environment, and 3) limitation in channel prediction.
aggressive spatial multiplexing. In massive MIMO, a BS Equipped with full-dimensional large arrays, ground BSs can
equipped with a few hundred antenna arrays serves tens of apply adaptive fine-grained 3D beamforming to mitigate the
users simultaneously. It is considered a promising technology strong interference between high-altitude UAVs and low-
in 6G owing to the following advantages [87]. altitude terrestrial users.
• Multiplexing gain: Aggressive spatial multiplexing can 3) Intelligent Reflecting Surface: IRS has been proposed as
increase the data rate dramatically in the spatial domain. a promising technology in 6G to alter the wireless channel by
Further, the offloading delay of the computation task is intelligently adjusting the amplitude and/or phase shift of each
significantly reduced for remote aerial computing. element of the IRS [90]. Extensive deployment of IRSs in the
• Energy/spectral efficiency: The energy efficiency can be wireless network and smart coordination of the reflections can
improved via massive MIMO as it is inversely propor- result in the flexible reconfiguration of the wireless channels
tional to the number of antennas at the BS. between transmitters and receivers to expand the commu-
• Increased robustness: Due to a large number of antennas, nication coverage and improve the wireless communication
the propagation channel provides additional diversity capacity and reliability. IRS has various advantages.
gains, which also increases the link reliability. Thus, • Low cost: The reflecting elements of IRS are low-cost
MIMO can provide reliable links in aerial computing, printed dipoles, which only passively reflect the signal
where the uncorrelated noise and intra-cell interference without transmitting the signal.
can vanish with an increasing number of antennas. • Full-duplex mode: Different from the traditional relay
Various multiple access technologies, such as OMA and system with half-duplex relay, IRS can work in a full-
NOMA, can be combined with massive MIMO to further duplex mode.
improve the communication performance [91]. In massive • Auxiliary device: IRS has great flexibility and compati-
MIMO-OMA systems, the maximum ratio combing and zero- bility to be integrated into the existing wireless networks
forcing can be utilized to achieve a high spectral efficiency in such as WiFi or cellular.
underloaded systems. However, some studies proposed mas- Thus, IRS can be massively deployed in wireless networks
sive MIMO-NOMA to support a large number of users; that is, and combined with other promising technologies, such as
the overloaded system, and to meet the massive connectivity NOMA [92], massive MIMO, and UAVs. However, IRS is
requirement of 6G [89]. However, there exist challenges for the still in its nascent stages, and certain potential challenges
realization of massive MIMO, which are the 1) deployment of need to be addressed: 1) The deployment of IRS, especially
extremely large antenna arrays, 2) adaptation to the intelligent for air computing, the location IRS can significantly affect
 13

the channel condition for data transmission; 2) The reflection computing on relevant domains: smart cities, smart vehicles,
of IRS is sensitive to the angle of arrival, which affects the manufacturing, and smart grids.
channel model of the IRS in practice [90].
IRS can be applied in various scenarios to improve the
communication performance [68]. The most important task A. Smart City
for aerial computing is to prolong the battery life of UAVs, The IoT involves all devices that are connected to the
LAC servers, and IoT devices. Energy harvesting or WPT Internet on Earth. Due to the potential of these devices in
techniques can be implemented using IRS to overcome the improving the quality of human lives, they are sensitive to
high-power loss over long distances. For example, a highly latency, storage, bandwidth, and security. With the increasing
efficient energy charging zone can be created by deploying popularity of the IoT, cities are becoming smarter, and the
IRSs in the proximity of devices or UAVs. In addition, if concept of smart cities is gaining traction and a new dimen-
IRSs are deployed on UAVs, the UAVs can flexibly establish sion. Aerial computing can aid in better understanding the
strong LoS links with the ground nodes to improve the potential of the IoT in developing new strategies for smart
communication quality for aerial computation task offloading cities, thus reducing costs and improving safety. Furthermore,
and downloading. new endeavors, such as 5Gcity [95] and SynchroniCity [96],
can help realize the idea of a future city. According to [95],
F. Summary and Discussion there are four possible major themes in smart cities, which
are data communication and processing, MEC support, urban
In this section, we have presented a set of enabling tech-
planning and management, and surveillance and security. In
nologies of aerial computing. Table III presents a summary of
Fig. 5, we provide an example of a smart city through the use
the enabling technologies and their advantages and challenges.
of aerial computing technologies.
These technologies can be flexibly implemented in future
1) Data Communication and Processing: With embedded
6G networks to efficiently provide fast computing services,
multiple sensors, LACs, HACs, and satellite computing plat-
better mobility, and higher scalability and availability. Further,
forms have great potential for sensing data in IoT environ-
these technologies can facilitate the process of addressing
ments and providing social services to smart cities. In [97],
challenges, such as intelligent communication and computa-
the authors proposed a joint scheme of 3D placement, com-
tion resource allocation design, joint optimization of UAV
putation, and communication resource allocation for multiple
trajectory and placement design, user association and grouping
HACs in an uplink IoT network, where the task distribution
design, and secure computing and communications. However,
among HACs and communication resources were considered.
certain challenges need to be addressed, as highlighted below.
The solution scheme was developed based on the K-means
• Intelligent Control: The cooperation of enabling tech-
method, a Hungarian-based algorithm, and an iterative method.
nologies should allow for intelligence in aerial comput- The simulation results illustrate that the total transmission
ing. The joint intelligent optimization of computing re- power of the IoT nodes is significantly minimized through
sources (LAC, HAC, and satellite computing platforms), the proposed algorithm. In [98], a technique referred to as
communication link design with different multi-access AirComp was developed to receive data from multiple sen-
techniques, and network control/scheduling can be inves- sors simultaneously by utilizing the superposition property of
tigated in future studies. multi-access channels, as shown in Fig. 5. A near-optimal
• Secure Control: For aerial computing, the user data need
equalizer was derived using differential geometry to facilitate
to be transmitted and shared with the aerial server, which multimodal sensing technology in the proposed technique.
leads to security concerns and data transmission over- Finally, an efficient channel feedback mechanism was designed
head. Therefore, FL and blockchain can be exploited to to facilitate the acquisition of the entire channel information
protect data privacy and improve security by transmitting from many sensors simultaneously. In [99], a novel and
the training parameter to the server for global aggregation efficient control algorithm for LACs was proposed to manage
[93], [94]. sensing and their movement using DRL techniques. The pro-
• Aerial Intelligence: AC, HAC, and satellite computing
posed method extracts features using the convolutional neural
platforms can enable aerial intelligence by collecting network (CNN) and infers decisions following the multi-agent
surrounding data and then executing onboard intelligent deep deterministic policy gradient (DDPG) method. However,
algorithms. The aims of intelligent algorithms include swarms of LACs and HACs may face difficulties owing
autonomous collision avoidance, adaptive flight gesture to their limited storage and computation capacities during
adjustment, and trajectory optimization for data collec- crowd sensing. To resolve this issue, in [100], edge/cloud
tion. Therefore, the design of intelligent algorithms can computing technologies were introduced to the swarm for
efficiently support aerial computing in terms of energy enhancing their QoS. Based on a case study of latency-
savings, large coverage, delay minimization, etc. critical applications, their simulation results illustrated that
the proposed approach could effectively improve the QoS of
IV. V ERTICAL D OMAIN A PPLICATIONS the swarm servicing entity. Furthermore, to enhance the poor
Aerial computing is expected to serve different vertical do- coverage of information-centric IoT networks, the authors of
main applications that require service provision from different [101] optimized the coverage of such a network via LACs and
perspectives. In this section, we explore the impact of aerial HACs at a low cost. The proposed scheme was designed based
 14

a fairness-aware task distribution and trajectory optimization

scheme in LAC systems, in which a fixed-wing LAC played
Road Network
Extraction the role of a flying computing server for mobile terminals.
MEC server Under fairness consideration, the authors aimed to minimize
MEC the maximum energy consumption of all mobile terminals.
Services Video The non-convex problem was converted into a convex one,
Surveillance
Multimodal-
and an iterative algorithm was proposed to solve the problem.
Sensing Further, under the assumption that the communication between
AirComp the LAC and mobile terminals followed the NOMA technol-
Collaborative
Service using Air Quality ogy, the authors in [106] proposed an iterative algorithm to
Blockchain Monitoring minimize the maximum energy consumption of all mobile
Urban Tree
Inventory terminals by controlling the trajectory, task data, and com-
Monitoring
puting resources in an adaptive manner. The simulation results
Terrestrial
BS
illustrated that the proposed algorithm could effectively reduce
the highest energy consumption among mobile terminals while
considering the fairness issue. The work in [107] introduced
a highly diversified urban scenario, which was served by
multiple LAC servers. Two non-convex programming prob-
lems were sequentially formulated to optimize the delay and
energy performance of uplink and downlink communication
scenarios for AR applications, which were solved by convex
approximation techniques. Through experiments, the authors
Fig. 5. Illustration of a smart city using aerial computing. verified that the proposed network model could serve cellular
users in a satisfactory manner.
3) Urban Planning and Management: Extracting building
on an improved rolling horizon strategy (IRHS). Through anatomy through remote sensing technologies is crucial for
comprehensive experiments compared to previous studies, the urban planning and management. Considering diverse scales
authors demonstrated that the proposed scheme can improve and appearances, assisted by HACs, the authors of [108]
the coverage ratio (by 21.42%) and reduce the cost ratio (by proposed an automatic method to extract building instances.
13.335% to 34.32%). The proposed HAC unit consisted of an improved hybrid
Aiming to track a moving target using synthetic aperture task cascade method, which had three components for three
radar (SAR) based on compressive sensing technology, the different consecutive tasks: high-resolution representation, def-
authors of [102] considered applying the multi-channel pro- inition of guided anchor, and formation of focal loss. Com-
cessing technique to the SAR computation unit. Because of the prehensive experiments on a real dataset revealed that the
joint consideration of the magnitude and phase, the proposed proposed method was shown to perform better compared to the
weight-adjusted sparse algorithm in the computing unit could conventional R-CNN method in terms of extracting building
improve the performance of clutter suppression. In another instances. As of the LAC server, in [109], a target tracking
work [103], the same authors determined a different method to algorithm to locate LAC-captured targets, such as pedestrians
track moving targets with the assistance of compensated target and vehicles, was proposed using the sparse representation
motion in the SAR computing unit. The motion compensation theory. Upon implementation with real data, the results prove
procedure in the SAR computing unit consisted of two steps: that the proposed tracker can achieve better performance com-
the SPECAN processing was first applied to the frequency pared with existing tracking algorithms. Further, to monitor
and azimuth domain, and then the correction of the residual urban environments, the authors of [110] proposed the use of
component was incorporated into the sparse imaging outcome. heterogeneous aerial computing (e.g., LACs, HACs, and satel-
Further, the authors also overcame the challenges associated lite computing) and investigated a collaborative methodology
with the motion of the target and the imaging of multiple mov- to share the workload. A simple illustration picture of this
ing targets using a novel parametric sparse imaging approach collaboration feature is provided in Fig. 5. For a particular
and an orthogonal matching pursuit method, respectively. set of degrees of freedom (e.g., a number of tasks and
2) MEC Support: In [104], a novel online HAC-assisted vehicles), an efficient algorithm was proposed to maximize the
edge server management scheme was proposed to provide overall weight of completed tasks. Through simulations, the
flexible edge computing services. Several hot-spot areas were authors determined that the proposed algorithm significantly
extrapolated by geographically merging the tasks, and the outperformed the other heuristic algorithms.
corresponding HACs were dispatched to the appropriate lo- With the growing trend of industrialization, air pollution
cations for the necessary service. A simple illustration of this has reached a significant level in urban areas. Consequently,
use case is provided in Fig. 5. Through extensive simulation monitoring air pollution has become a matter of great concern
with real-world data, the authors showed that the proposed in recent years. To this end, the authors of [111] presented
mobile server dispatching scheme could serve more users with an efficient and cost-effective air quality monitoring system
a high utilization of resources. The authors of [105] proposed using LACs and HACs. The three major aspects of the system
 15

proposed in [111] were data processing, deployment, and The results from comprehensive experiments indicated that the
power control. The system was deployed at Peking University proposed algorithm achieves centimeter-level accuracy with a
and Xidian University, and about 100,000 samples have been precision of 84.8%.
collected since February 2018. A partial illustration of this 4) Surveillance and Security: The concept of a smart city is
system is provided in Fig. 5. Using a 360-degree panoramic vulnerable to security and privacy issues since the collection,
camera in HACs, the authors of [112] designed a novel air dissemination, management, and processing of data in aerial
quality indicator (AQI) monitoring system, AQ360, to detect computing platforms can be interrupted by malicious users.
the air quality level (shown in Fig. 5). Upon solving the Thus, evidenced by certain existing reports, we discuss the
recognition problem of images, they optimized the placement way in which aerial computing platforms (e.g., LACs, HACs,
of the corresponding LAC to achieve the optimal outcome. and satellites) could be the means of resolving security and
Through the implementation of the system under realistic privacy issues in smart cities.
scenarios, they showed that the system can perform better With the assistance of 3D simulation and a set of robot
compared to existing studies in terms of AQI recognition error operating system-equipped LACs, the authors of [120] demon-
and energy consumption. Further, using microwave downlink strated an autonomous navigation system to control the phys-
sensor networks, the authors of [113] studied a rainfall moni- ical security of smart buildings. The results from the ex-
toring system via a satellite computing platform. The proposed periments indicated that the proposed approach achieves an
method was implemented in the satellite computing unit, and it acceptable level of accuracy in terms of mapping the indoor
was compared with the Ligurian regional tipping-bucket rain environment of smart buildings. In [121], a novel framework
gauge (TBRG) network at a city in Italy. The performance was proposed to monitor suspicious links by utilizing the
of these two types of monitoring systems was found to be characteristics of LACs. Particularly, the system consists of
similar, based on the rain events that occurred over the time a suspicious transmission link and an LAC unit for mon-
period between January 2017 and December 2018. itoring. Further, in the LAC system, the wireless resource
Fine-grained driving lines are crucial components of high- allocation problems, including trajectory planning and energy
definition maps for the smooth operation of autonomous minimization, can be solved using the popular successive
vehicles. To this end, the authors of [114] presented a semi- convex approximation method.
automated driving line generation method utilizing a mobile Video surveillance applications are associated with several
laser scanning computing system, as shown in Fig. 5. Through challenges in smart cities, such as scalability, integrity, and
simulations on real datasets, the authors demonstrated that latency. In [122], a series of optimal scheduling and control
the proposed method achieved an average recall, precision, algorithms were designed to address these challenges. A
and F1-score of 90.79%, 92.94%, and 91.85%, respectively. network with full coverage was established via an LAC cluster,
Further, using the same mobile laser point clouds, the authors and the scheduling problem was solved using the bi-objective
of [115] proposed an image-translation-based method to ob- fragile bin packing technique. Thereafter, extensive simula-
tain the 3D vectors of typical road markings, as shown in tions with realistic parameter settings were conducted, and the
Fig. 5). Another method for extracting road networks using a effectiveness of the proposed scheme was verified in terms of
satellite computing platform was proposed in [116], wherein many systems and video-specific performance metrics. The
a novel aperiodic directional structure measurement (ADSM) authors of [123] introduced a surveillance model for a multi-
technique was adopted. In contrast, to facilitate city planning domain IoT environment using heterogeneous smart HACs and
via swarms of fully automated LACs, the authors of [117] pro- LACs (shown in Fig. 5). A heuristic method was used to
posed a holistic distributed framework equipped with various minimize the maximum movement of LACs while avoiding
effective and efficient skills, which could be adjusted based collisions among them. Extensive simulations were conducted
on the requirements of the LACs and environments. Through considering heterogeneous scenarios with multiple HACs and
experiments, the authors demonstrated that a swarm of LACs LACs to verify the merits of the proposed method.
equipped with the proposed framework could effectively sat- In addition to limited storage and computation power,
isfy the requirements of diverse applications. The authors of security and privacy are two of the alarming concerns for data
[118] presented an efficient framework for extrapolating urban dissemination schemes in aerial environments. To this end, the
tree inventories by using 3D point clouds acquired by a HAC- authors of [124] presented a low-cost computationally efficient
borne laser scanning system (shown in Fig. 5). The high-level short proxy signature (CB-PS) scheme for the LAC server,
processing steps of the proposed HAC framework are individ- which was oblivious to secret key information in conventional
ual tree cluster extraction, geometric parameter estimation, and ID-based schemes. To further enhance the security of such
tree species classification. Through experiments, the authors tasks, the authors of [125], presented a novel blockchain-based
demonstrated that the detection accuracy of the roadside tree strategy to facilitate multi-party authentication among multiple
was over 93% with an average error of approximately 5%, LACs and HACs. Specifically, this strategy is useful for P2P
and the overall classification accuracy was approximately and group communications among LACs and HACs while
78%. Another application of the point clouds collected by ensuring the efficiency of data dissemination. Through com-
the computing system is the modeling of an underground prehensive experiments and simulations, the authors verified
parking lot [119]. The method utilized in the HAC consists of the merit of the proposed technique in terms of authentication
two parts: a joint localization and mapping algorithm based among multiple LACs and HACs. Subsequently, to satisfy the
on sparse point clouds and a semantic modeling algorithm. connectivity, data, and service demands of an exponentially
 16

detection algorithm (shown in Fig. 6). Implemented at the

Vehicle Flow HAC server, the detection method considers sparsity in data
Statistics
as well as multiple features. Further, through experiments, the
authors verified the utility of the proposed algorithm in terms
Infotainment
of detecting a large number of vehicles in metropolitan cities.
2) Safety and Security: Based on the data collected by
the LAC server, the authors of [132] proposed the Siamese
Interaction Long Short-Term Memory (SILSTM) network to
Collision-prone
Vehicle Detection identify the behavior of collision-prone vehicles. Upon learn-
ing the interaction scenarios of a vehicle with its neighbors,
MEC
the SILSTM network can predict accident-prone interactions
Services in advance. Subsequently, through extensive experiments with
a real dataset, the authors verified the merit of the proposed
MEC
approach in terms of detecting collision-prone trajectories at
Terrestrial
server
BS
four different intersections. A collision-prone scenario and its
resolution are illustrated in the intersection of Fig. 6. Traffic
Fig. 6. Illustration of smart vehicles using aerial computing. surveillance is one of the core components of next-generation
ITS. LACs or HACs can act as relays to facilitate quick data
delivery between vehicles and edge nodes, which is associated
growing number of IoT devices in smart cities, the authors with a high risk of information leakage. To this end, for the
of [126] envisioned a 5G network environment supported by LAC server, the authors of [133] proposed an optimization
blockchain-enabled LACs and HACs. The objective of this model based on a probabilistic data structure (PDS)-based
network was to provide Internet connectivity to IoT devices strategy with a triple Bloom filter that can detect cyber threats
via blockchain-enabled multiple LACs and HACs and to of smart vehicles in advance.
facilitate decentralized service delivery and routing facilities 3) MEC Support: During unexpected events, e.g., adverse
in a reliable and secure manner. weather conditions and extreme traffic congestion, LAC/HACs
can act as flying RSUs. Such a scenario was described in
B. Smart Vehicle [134], providing the required infrastructure for diverse traf-
Over the last 30 years, intelligent transportation systems fic applications for the improvement of the QoS (shown in
(ITS) have been investigated to improve the safety and effi- Fig. 6). Through simulation, the authors demonstrated that the
ciency of the transportation industry. Along this line, IEEE proposed method could achieve satisfactory performance in
developed a protocol called IEEE 802.11p [127]. However, terms of network coverage and latency. In the context of social
this standard is not sufficiently competent to meet the real- Internet of Vehicles (SIoV) to strengthen social relationship
time requirements of the emerging vehicular industry. Conse- among vehicles, an LAC platform for SIoV, consisting of a
quently, existing studies have emphasized the merits of aerial three-layer integrated architecture, was adopted in [135]. The
computing for efficient ITS [128], [129] to achieve reliable aim of the architecture was to jointly optimize the transmission
and low-latency, smart, and seamless vehicle-to-everything power of vehicles and the LAC trajectory under realistic con-
communication. Along with benefits, several research chal- straints. Through simulations compared with existing schemes,
lenges still exist in aerial vehicular computing, which can be the authors demonstrated that the proposed architecture in the
overcome by leveraging emerging technologies, such as IRS LAC platform could effectively maintain social relationships
and AI. For example, [128] showed that AI-enabled LAPs among vehicles.
help improve the service performance of vehicular networks 4) Infotainment: To disseminate information (e.g., news
compared with ones without AI and LAP deployment. and entertainment), [136] proposed a new LAC-based wire-
Furthermore, in recent years, the 5G Automotive Asso- less access infrastructure using the RaptorQ-protected content
ciation (5GAA) defined a new concept, namely cooperative diffusion technique. A simple illustration of this function
intelligent transportation systems (C-ITS), which considers is shown at the right-hand side of Fig. 6. Through exper-
aerial computing as an enabling technology for V2X com- iments using real vehicles and LAC platforms, the authors
munications [130]. According to this standard, we highlight demonstrated that the RaptorQ-based content dissemination
a few main use cases in the context of C-ITS, such as mechanism in LAC platforms was highly efficient and cost-
traffic management, safety and security, MEC provision, and effective in terms of transmitting information to multiple
infotainment. Fig. 6 illustrates a few applications served by moving receivers simultaneously. For the same purpose, in
aerial computing platforms in smart vehicles. [137], a 3D LAC-based simulation model was constructed
1) Traffic Management: With the growing number of vehi- using the OMNET++ simulator with the objective of passing
cles in recent years, vehicle detection has become a vital part information from the LACs to cars. Various path loss and
of the ITS, which assists in resolving various information and elevation models were considered to render the simulation
management issues, such as traffic and vehicle flow statistics more realistic.
and road network management. By acquiring high-resolution Disseminating data efficiently is one of the core tasks in
images using HACs, the authors of [131] designed a vehicle V2X networks. To this end, the authors of [138] proposed
 17

Crop Row becoming increasingly automated with the development of

Detection
robots and smart tractors. However, to guide these machines
in an automated manner, the accurate detection of crop rows
is crucial. In [142], based on images acquired by LACs,
Collaborative Factory
Monitoring the authors proposed a learning CRowNet model to detect
crop rows (shown in Fig. 7). The corresponding LAC server
was equipped with a model formed using SegNet (S-SegNet)
and a CNN-based Hough transform (HoughCNet) technique.
Through extensive numerical simulations compared with tra-
ditional methods, the authors demonstrated that the CRowNet
model achieved 93.58% accuracy.

D. Smart Grid
Fig. 7. Illustration of smart factories using aerial computing. Smart grids have been widely used in recent years. A large
number of new power lines have been under real operation,
a novel LAC-based scheduling protocol that consists of a resulting in a significant growth of transmission line mileage
proactive caching policy and a file-sharing strategy. Subse- and power equipment. However, the monitoring, detection,
quently, through simulations, the authors demonstrated that the and maintenance of power transmission systems are essential
proposed scheduling strategy could enhance the performance to ensure uninterrupted operations. Having noticed the risk
of the system in terms of throughput and latency. Further, of failure in conventional manual monitoring methods, many
to enhance the security of V2X communications, in [139], a researchers have been working on the automation of power
few physical layer security (PLS) strategies were proposed for line inspection, monitoring, and management using the latest
LACs and HACs. Comprehensive simulations were conducted, technologies [143], [144], such as aerial computing, image
the outcome of which revealed that the proposed strategies processing, AI, and DL. Most of the existing studies emphasize
in LACs and HACs could effectively handle many emerging two crucial aspects of smart grid automation: 1) monitoring
security threats in V2X communications. and inspection and 2) control and management. We focus on
the computational aspect of aerial technologies in performing
these tasks. In Fig. 8, we briefly illustrate a few use cases of
C. Smart Factory the smart grid domain using aerial computing.
Among the many enabling technologies, 6G, industrial 1) Monitoring and Inspection: Insulators are prevalent in
IoT (IIoT) and aerial computing have been integrated with high-voltage transmission lines and play a key role in electrical
cognitive skills and innovation to aid industries in increas- insulation and conductor conjunction. Insulator faults (e.g.,
ing production and delivering customized products more glass insulators self-blast) pose a grave threat to power systems
quickly [140]. Such enabling technologies constitute Industry as they can cause cascade failure. Considering the difficulty
5.0, an advanced production model focusing on the interac- of manual inspection, [145] proposed a new DL framework to
tion between machines and humans. This collaborative work detect the location of broken insulators by capturing images
facilitates human capabilities, realizing higher productivity using HACs (shown in Fig. 8). Specifically, the proposed DL
and exceptionally easy automation for individuals and small model in the HAC server works in the low-SNR regime with
businesses. Such a feat was not possible a few years back. the assistance of two modules: R-CNN to detect objects and
Here, we provide a few examples of manufacturing models U-net to classify pixels. Using a real dataset, via experiments
with the assistance of LACs and HACs, the corresponding compared with the other methods, the authors showed that
schematic of which is illustrated in Fig. 7. the proposed approach achieves real-time accuracy. After
With the increasing trend of automation, IIoT has become capturing images of transmission lines using HACs, [146]
prevalent in smart factories for data communication and pro- also proposed an improved insulator defect detection method
duction management. Wireless sensor networks (WSNs), the based on an improved ResNeSt and region proposal network.
most important components of the IIoT, are used for numerous Through suitable experiments on a real dataset, the authors
tasks in indoor industrial environments, such as irrigation and showed that the proposed method deployed in the HAC server
the inspection of machines. However, these networks have a can achieve 98.38% accuracy in detecting insulator defects.
low computational capability. Therefore, in [141], a WSN- Further, aerial computing is a feasible solution to inspect
based safe navigation algorithm was proposed for micro aerial power transmission lines. Consequently, while solving the
robots (i.e., LACs) in the IIoT. A WSN with a 3D ranger battery limitation problem of aerial computing platforms (e.g.,
was used as a computational tool for micro flying robots HACs and LACs), a new idea was proposed in [147], suggest-
to detect obstacles in an indoor industrial environment, as ing the use of a smart hangar as an assistant to achieve full
shown in Fig. 7. Through extensive simulations with multiple power automation, a path planning model for the LAC server
micro flying robots, the authors verified the merit of the was developed. Through simulations, the authors demonstrated
proposed algorithm in terms of obstacle-aware monitoring that the proposed scheme in the LAC server is effective in
in dynamic factory environments. Meanwhile, agriculture is solving existing inspection problems.
 18

Insulator Defect satellite, and a hybrid solar forecasting method was proposed
Detection
Electrical Theft to run the smart grid in a cost-effective manner. For the core
Dectection
computing unit, desirable features and cloud cover factors
Connect Between Two were extracted from the images using certain intermediate
Power Distribution
Networks
mathematical formulae and a modified CNN, respectively.
Solar Forcasting
Coupled with meteorological information and cloud cover
Method factors, a final forecasting model was developed.
Wind Turbine
Management

E. Summary and Discussion

In this section, important vertical domain applications sup-
ported by aerial computing, including smart cities, smart
vehicles, smart factories, and smart grids, have been discussed.
A summary of the applications is presented in Table IV.
Data communication, processing, and management of
Fig. 8. Illustration of smart grid using aerial computing.
ground sensors and IoT nodes are crucial tasks to facilitate
applications in smart cities. Aerial computing platforms (e.g.,
LACs, HACs, and satellites) are effective in performing these
With the growing popularity of wind power, more turbines tasks owing to their flexible deployment and LoS communica-
are being deployed in remote areas. [148] considered the tion features in air-to-air or air-to-ground links. However, even
deployment of HACs and LACs for inspecting turbines (shown with the use of aerial computing, several research challenges
in Fig. 8). Using the forecasted wind conditions obtained require to be solved to obtain optimal performance, such as 3D
by the HAC-borne edge intelligence, two algorithms were deployment, joint computation and communication resource
developed to minimize the flight time of the HAC and to management, coverage, and trajectory design. Several existing
optimize power generation by the turbine. Through simulation studies have addressed these challenges by adopting cutting-
using a real dataset, it was shown that the proposed method edge techniques, such as convex optimization, graph theory,
can generate 44% more power and reduce the flight time by and AI/ML tools. Because of security breaches, smart city
25% compared to an existing scheme. With the growth of applications are associated with several challenges, such as
rooftop solar PV arrays in power generation, [149] designed a designing low-cost and efficient signature and authentication
new system called SolarFinder, which can be implemented in protocols. Few studies have adopted the CB-PS scheme and
a satellite computing platform to detect distributed solar PV blockchain to solve these problems to a certain extent. Further,
arrays over an area in a cost-effective and autonomous manner, aerial computing is a promising solution for realizing MEC
as shown in Fig. 8. The intelligence of SolarFinder is built on with other communication technologies, which is associated
the support vector machine (SVM) model and a deep CNN with several challenges as well, such as task data allocation,
approach jointly. Over 269,632 images captured by satellite, computation and communication resource management, and
the performance of SolarFinder was evaluated, and the results aerial device deployment issues. City planning and manage-
showed that the pre-trained SolarFinder achieved a Matthews ment are also crucial aspects of smart cities. Certain useful
correlation coefficient (MCC) of 0.17. tasks in the context of this use case are target tracking to facili-
2) Control and Management: According to the require- tate city design, air pollution monitoring, and city map designs.
ments of a smart grid network, controlling non-technical losses In addition to providing certain functional requirements, aerial
(e.g., electrical theft) is one of the most alarming concerns computing may aid in achieving surveillance and security
in developing countries. In [150], the authors discussed a in smart cities. The effective utilization of aerial computing
possible communication network and a framework using LAC is associated with certain challenges, such as the HAC/LAC
platforms to control non-technical losses (i.e., tampering of navigation strategy, deployment, and resource management.
electrical devices) in energy distribution systems. An example In the domain of smart vehicles, aerial computing platforms
of this use case is shown in Fig. 8. For the effective man- can be effectively utilized for various purposes, such as
agement of power distribution networks, which consist of nu- traffic management, safety and security, flying RSUs, and
merous sparsely distributed grids, LEO satellites were utilized infotainment. Despite serving all these use cases, several
as the key components of the communication infrastructure research challenges need to be solved, such as computation
to manage active networks [151]. Such a scenario is shown and communication resource management, trajectory design of
at the left corner of Fig. 8. The effectiveness of the proposed aerial computing platforms, data analysis, and routing strategy.
LEO satellite-equipped framework in terms of managing future Several existing studies have addressed these issues via AI/ML
power supply networks was established through extensive ex- techniques, probabilistic approaches, and optimization tools.
periments. In [152], the feasibility of satellite communication Finally, monitoring, inspection, and management are the
for connecting a smart grid with remotely located actuators primary applications of aerial computing platforms in smart
was also studied. In addition to the routing and channel manufacturing and smart grid domains. However, certain tech-
access mechanism, the possible optimization of computation nical problems must be solved to obtain optimal outcome
resources was studied. In [153], images were captured using a from aerial computing platforms in these domains, such as
 19

TABLE IV
S UMMARY OF VERTICAL DOMAIN APPLICATIONS .

References Use Cases Analytical Tools Relevant Lessons

– [97]–[103] – Data collection, high-mobility mul- – K-means method, Hungarian – Minimize the total transmission power
– [104]–[107] timodal sensing, control and manage- algorithm, differential geometry, of IoT nodes, outperforms three existing
Smart Cities

– [108]–[119] ment of sensing, enhancement of QoS CNN and DDPG, edge/cloud schemes, improves QoS, improves the cover-
– [120]–[126] in crowd sensing, improvement of in- computing, IRHS, sparse age ratio of information by 21.42%, and can
formation coverage, and tracking mov- algorithms based on compressive track moving targets accurately.
ing targets. sensing technology, and a novel – Provide energy-efficient AR services to
– Serving IoT devices with effi- parametric sparse imaging mobile users under tight delay constraints.
cient resource utilization, task data al- approach. – Enhance performance compared with the
location, and trajectory optimization – Geographic merging of tasks, mainstream Mask R-CNN method; achieve
with/without NOMA, and serving as optimization algorithms, and good precision compared with state-of-the-
cache-enabled edge computing nodes. convex approximation. art works; maximize the total weight of
– Automatic extraction of building – Sparse representation theory, completed tasks; achieve a lower AQI recog-
instances, target tracking, monitor- collaborative workload sharing nition error; has close agreement with other
ing urban environments, air quality algorithm, system development, existing methods; achieve an average recall,
monitoring, rainfall monitoring, semi- placement optimization, satellite precision, and F1-score of 90.79%, 92.94%,
automated driving line automation, microwave downlink sensor and 91.85%, respectively; achieve an overall
building road networks, city planning, networks, inverse distance accuracy of 78%; and achieve centimeter-
extraction of urban tree inventory, and weighted method, a novel level accuracy with a precision of 84.8%.
modeling an urban parking lot. ADSM technique, decentralized – Achieve an acceptable accuracy for indoor
– Physical security of smart buildings, agent-based control architecture, environments, effectively monitor a suspi-
monitoring suspicious links, mitigat- classification-based method, and cious link, make video surveillance applica-
ing challenges of video surveillance SLAM algorithm. tions work efficiently, enhance the security
applications, constructing surveillance – Jamming signal, an of the LAC server, perform authentication
models for multi-domain IoT, ensur- optimal scheduling algorithm, tasks effectively, and proves that blockchain
ing security in data dissemination, and optimization, low-cost CB-PS is effective in collaborative security among
trustworthy authentication. scheme, and blockchain. multiple LACs and HACs.

– [131] – Vehicle detection for traffic manage- – ML technique. – Achieve satisfactory performance in vehi-
Smart Vehicles

– [132], [133] ment and road network planning. – SILSTM network and a PDS- cle detection.
– [134], [135] – Traffic collision detection and cyber based scheduling technique. – Can detect collision-prone interaction tra-
– [136]–[138] threat detection. – Algorithm and optimization jectories and cyber threats effectively.
– Flying roadside units, and social tools. – Can achieve full network coverage under
relationships among vehicles. – RaptorQ-based content dissem- different scenarios and optimize the transmit
– Disseminating news and entertain- ination mechanisms and opti- power in vehicles and LAC trajectory.
ment information among vehicles and mized scheduling policies. – Can deliver infotainment service effec-
sharing files in V2X. tively.

– [141] – Monitoring industrial environments. – A navigation algorithm. – Can navigate accurately while detecting
Factories

– [142] – Detecting crop rows in agriculture. – A model formed with SegNet obstacles.
(S-SegNet) and HoughCNet. – Can detect crop row with the accuracy of
93.58%.
– [145]–[149] – Detecting insulator defects in trans- – DL and customized neural net- – Achieve 93.58% accuracy, achieve 44%
– [150]–[153] mission lines, inspecting power trans- works, optimization tools, and improvement in power generation and 25%
Smart Grids

mission lines, inspecting wind tur- integration of an SVM model reduction in flight time, achieve an MCC of
bines, and detecting distributed solar with deep CNNs. 0.17, which is 3 times better than a pre-
PV arrays. – Design of a suitable com- trained CNN-based approach, proves that a
– Controlling non-technical losses in munication network; integration smart hanger can effectively solve the limited
energy distribution systems, manag- of LEO satellites, optimiza- battery capacity problem of LACs/HACs.
ing active power distribution networks, tion tools, and CNN-based ap- – Can detect tampering of electrical devices,
connecting two power distribution net- proaches. show that LEO satellites are effective in man-
works, and forecasting solar energy for aging remote power generation sites, and can
the smooth operation of smart grids. forecast solar energy for smooth operation.

obstacle-aware monitoring a factory environment [141] and AI/ML, and optimization tools.
crop row detection (using ML tools) in agriculture field
[142]. Furthermore, a few more problems in smart grids V. C HALLENGES AND F UTURE D IRECTIONS
using aerial computing, such as the detection of insulator In this section, we identify interesting research challenges
defects, management of solar PV array, and inspection of and highlight possible future directions for aerial computing.
power line and wind turbine, need to be resolved. In terms of
control and management, the main problems are electrical theft A. Energy Efficiency
detection, active and power distribution network management,
and forecasting solar energy for the smooth and continued In aerial computing, achieving sustainable energy manage-
operation of smart grids, which can be addressed by equipping ment at LAC and HAC platforms and energy-efficient satellite
special hardware and software in aerial computing platforms, computing is a major concern. Data communications and
service delivery in satellite environments and space travel with
 20

airplanes, require large energy resources to ensure network optimization, and hardware design were used to accelerate on-
operations. In addition, the power consumption of aerial com- device data training [159]. A streamlined slimming framework
ponents may vary depending on the operation mode. For exam- was developed and combined with a consecutive tensor layer
ple, average power consumption of the UAV is 8.2412 W in the to improve the training rates. Simulation results show that the
idle mode and 245.2815 W in the horizontal movement mode, proposed method can enhance the training rate by up to 30%
and average power consumption is 8.2618 W and 8.2637 W compared to traditional approaches without compromising
when the UAV performs communication transmissions with learning accuracy. Thus, this approach provides opportunities
WiFi and GPS, respectively [154]. Thus, we need to focus for designing intelligent flying devices in aerial computing.
on energy-efficient designs to realize sustainable and green
aerial computing. For example, the authors of [155] suggested C. Network Stability
a space-air-ground architecture with a focus on maximizing
The topology of aerial computing fluctuates with the number
the system energy efficiency enabled by the joint optimization
of nodes, battery levels, and varying communication condi-
of uplink transmission power control, sub-channel selection,
tions. The trajectories and speeds of aerial components and
and deployment of aerial relays. This can be achieved by
flying devices also vary owing to the terrestrial application
dividing the original problem into two sub-problems, optimal
requirements and environmental dynamics, e.g., different al-
subchannel selection and power control policy, which are ob-
titudes of buildings in smart cities in the LAC platform and
tained by available aerial relay deployment. Moreover, energy
different orbital altitudes of multiple LEO satellites. Further,
refilling techniques to exploit renewable energy resources can
the operational complexity and unpredictability of aerial com-
be useful for building sustainable aerial computing systems.
ponents in space can make the involved aerial computing
Aerial devices and satellites can harvest power from ambient
system unstable. In fact, aerial nodes can join and leave
environments, such as wind, solar, vibration, and thermal
unpredictably, or their flight can be stationary, slow, or fast.
power, at the LAC platform to support their operations, e.g.,
Thus, achieving network sustainability in aerial systems is
data communications over aerial links.
critical. As a promising method, an optimal tracking policy
was constructed in [160] for each aerial device in an aerial-
B. Resource Management based network to mitigate the varying network topology issues
in aerial computing operations, e.g., flying trajectories of
Aerial computing accommodates data collection and com-
UAVs. The key focus was on achieving an adaptive surround-
putation tasks to serve end IoT devices via different computing
ing network configuration for varying channel quality and
platforms (e.g., LAC and HAC) and supports edge services
communication bandwidth resources. Subsequently, a particle
in future 6G networks. Compared to terrestrial computing
swarm optimization algorithm was developed to optimally
infrastructures, such as cloud data centers and cloudlet, com-
schedule the energy allocation among a set of aerial nodes,
puting platforms in aerial computing possess limited storage
while the prediction error of the surrounding node locations
resources and battery capabilities, which would hinder the de-
was minimized. Another study in [161] built an observer,
ployment of aerial computing services, such as airplane-based
which could monitor surrounding unmanned aerial nodes in
civil monitoring for terrestrial IoT environments [156]. Thus,
an aerial environment using Kalman filtering with respect to
resource management strategies are urgently required to satisfy
the maximum number of parallel targets, measurement time,
seamless service delivery in space. Moreover, in the future 6G
measurement success rate, and measurement noise. Thus, the
era, intelligent aerial computing is expected to be a dominant
optimal measurement policy was obtained for network topol-
research area, where AI functions can be integrated into aerial
ogy monitoring, aiming to achieve an accurate prediction of
devices at the LAC platform to enable self-controlled and
trajectory topology in autonomous aerial computing systems.
autonomous aerial systems. In this context, training DL models
Moreover, blockchain can help build scalable aerial comput-
on these aerial devices with massive datasets may be infeasible
ing networks thanks to its decentralized feature for enabling
because of the high demands of computation and memory
large-scale interconnection among servers and devices, e.g.,
resources, particularly when training with large-scale audio
unmanned aerial components in the HAC platform [162]. Each
and image data [157]. Therefore, optimizing on-device AI/DL
aerial device works as a blockchain node to build decentralized
models is of paramount importance for solving the computa-
data services, such as data sharing, data communications, and
tional burden posed on aerial devices in the air. Consequently,
data storage, without the need for a central server for better
several solutions have been proposed to facilitate resource
networking scalability.
management in aerial computing. For example, the authors of
[158] considered a resource management solution for aerial
computing systems wherein a joint optimization of task of- D. Large-Scale Network Optimization
floading, resource allocation mechanism, and the trajectory of In aerial computing, aerial devices operate on a large scale,
aerial devices in the air was derived with respect to terrestrial and cooperative optimization is needed to utilize the advan-
users’ latency requirements. Thus, the energy of the aerial tages from multiple and distributed network datasets, such as
devices was minimized, while a longer flight duration was diverse channel features and environment properties, while
achieved. Moreover, it is also necessary to develop on-device improving the network quality. For example, performing an
learning solutions to support self-learning aerial computing optimal trajectory control policy for all manned and unmanned
systems. For example, improved network architecture, training aerial components is a challenge on the HAC platform if only
 21

the characteristics of an aerial device are obtained [163]. More VI. C ONCLUSION
importantly, the datasets of future intelligent aerial networks Edge computing has become an indispensable component of
are distributed over large-scale networks rather than being cen- the present network infrastructure, but it also has various lim-
trally located. Therefore, there is an urgent need for distributed itations owing to the emergence of new services and applica-
and large-scale optimization approaches to enable scalable tions and the expansion of the network. We highlighted the fact
and intelligent aerial computing applications. A large-scale that the current computing infrastructure does not meet new
trajectory optimization solution was proposed in [164] for the demands and requirements. In this regard, the concept of aerial
Internet of aerial devices, for example, UAVs and balloons, computing was introduced and comprehensively reviewed in
through the use of a multi-agent DRL algorithm. This enables this study. First, we presented an overview of aerial computing,
aerial devices to collaboratively develop a distributed sense- starting from the system architecture to the reference model
and-send protocol in an aerial computing setting, facilitating and fundamental features, and compared it with conventional
large-scale sensing and data task transfer in cellular networks. computing paradigms, such as cloudlet, fog computing, and
MEC. Next, we presented a set of key enabling technologies
and discussed the use of aerial computing in vertical domain
E. Security, Privacy, and Trust applications. Finally, we discussed the research challenges and
promising future directions pertaining to aerial computing. As
Although 6G-based aerial computing can offer global cov- the development of aerial computing is still in a preliminary
erage and diverse QoS provision to industrial applications, stage and there are many unexplored issues, we believe that
critical issues related to security, privacy, and trust must be this paper has revealed certain important lessons and key ideas
solved. Adversaries may attack aerial communication chan- that will drive further research and unlock the full potential of
nels and deploy data breaches in the flying BSs on the a comprehensive 6G computing infrastructure in the future.
HAC platform, as the management of lower altitude-based
servers is limited owing to the physical distance [165]. AI
techniques are extensively used to enable intelligent aerial R EFERENCES
computing, but they often require centralized data collection [1] H. Tataria, M. Shafi, A. F. Molisch, M. Dohler, H. Sjöland, and
F. Tufvesson, “6G wireless systems: Vision, requirements, challenges,
for training, which raises potential privacy issues owing to insights, and opportunities,” Proceedings of the IEEE, vol. 109, no. 7,
the exposure of sensitive information in the air. Moreover, the pp. 1166–1199, Jul. 2021.
deployment of satellite-terrestrial communications in space can [2] Q.-V. Pham, M. Zeng, T. Huynh-The, Z. Han, and W.-J. Hwang, “Aerial
face critical challenges caused by untrusted environments, as access networks for federated learning: Applications and challenges,”
IEEE Network, 2022, in press.
third parties and attackers can compromise the data exchange [3] F. Jiang, K. Wang, L. Dong, C. Pan, W. Xu, and K. Yang, “AI
among aerial devices, BSs, and terrestrial IoT users. There- driven heterogeneous MEC system with UAV assistance for dynamic
fore, blockchain is a promising solution [166] for building environment: Challenges and solutions,” IEEE Network, vol. 35, no. 1,
pp. 400–408, Jan./Feb. 2021.
trust and establishing secure decentralized satellite-ground [4] M. Ke, Z. Gao, Y. Huang, G. Ding, D. W. K. Ng, Q. Wu, and J. Zhang,
communications for aerial computing systems. Blockchain, “An edge computing paradigm for massive IoT connectivity over high-
with its traceability and trustworthiness, can provide enhanced altitude platform networks,” IEEE Wireless Communications, vol. 28,
no. 5, pp. 102–109, Nov. 2021.
security for aerial computing systems [167]. The data from [5] Z. Yu, Y. Gong, S. Gong, and Y. Guo, “Joint task offloading and
aerial communications, such as device location and channel resource allocation in UAV-enabled mobile edge computing,” IEEE
information, can be stored in the data ledger where blocks Internet of Things Journal, vol. 7, no. 4, pp. 3147–3159, Jan. 2020.
[6] S. Garg, A. Singh, S. Batra, N. Kumar, and L. T. Yang, “UAV-
are built based on immutable transactions. This makes the empowered edge computing environment for cyber-threat detection in
information in the blockchains resistant to data modifications smart vehicles,” IEEE Network, vol. 32, no. 3, pp. 42–51, May/Jun.
or threats. The elimination of centralized authority of the 2018.
[7] A. Borrett, “Satellite broadband is the future
blockchain also helps mitigate single-point failures in involved of the $1trn space economy,” Mar. 2021. [On-
aerial networks. In addition, smart contracts can be used as line]. Available: https://techmonitor.ai/internet-of-things/connectivity/
self-executing software running on blockchain for providing satellite-broadband-future-1trn-space-economy
[8] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, and D. Sabella,
automatic authentication and verification to ensure reliable “On multi-access edge computing: A survey of the emerging 5G
aerial communications [166]. This technique attains further network edge cloud architecture and orchestration,” IEEE Communica-
relevance in the 6G era because aerial computing systems tend tions Surveys & Tutorials, vol. 19, no. 3, pp. 1657–1681, Thirdquarter
2017.
to be decentralized and deployed on a large scale, which can [9] P. Porambage, J. Okwuibe, M. Liyanage, M. Ylianttila, and T. Taleb,
be realized using the decentralization feature of blockchain. To “Survey on multi-access edge computing for internet of things realiza-
preserve data privacy in aerial computing, perturbation tech- tion,” IEEE Communications Surveys & Tutorials, vol. 20, no. 4, pp.
2961–2991, Fourthquarter 2018.
niques, such as differential privacy and dummy operations, are [10] P. Mach and Z. Becvar, “Mobile edge computing: A survey on archi-
particularly helpful in protecting data leakage against external tecture and computation offloading,” IEEE Communications Surveys &
threats during data exchange. For example, differential privacy Tutorials, vol. 19, no. 3, pp. 1628–1656, Thirdquarter 2017.
[11] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey
was adopted in [168] to realize privacy-enhanced intelligent on mobile edge computing: The communication perspective,” IEEE
aerial communications by introducing artificial noise into the Communications Surveys & Tutorials, vol. 19, no. 4, pp. 2322–2358,
trained gradients at aerial devices. This solution helps hide Fourthquarter 2017.
[12] B. Sonkoly, J. Czentye, M. Szalay, B. Németh, and L. Toka, “Survey
sensitive information while guaranteeing convergence, such on placement methods in the edge and beyond,” IEEE Communications
that adversaries cannot retrieve useful data samples. Surveys & Tutorials, vol. 23, no. 4, pp. 2590–2629, Fourthquarter 2021.
 22

[13] X. Wang, Y. Han, V. C. Leung, D. Niyato, X. Yan, and X. Chen, Resource allocation and trajectory optimization,” IEEE Transactions on
“Convergence of edge computing and deep learning: A comprehensive Wireless Communications, vol. 21, no. 3, pp. 1621–1637, Mar. 2022.
survey,” IEEE Communications Surveys & Tutorials, vol. 22, no. 2, pp. [32] X. Hu, K.-K. Wong, and Y. Zhang, “Wireless-powered edge computing
869–904, Secondquarter 2020. with cooperative UAV: Task, time scheduling and trajectory design,”
[14] D. Xu, T. Li, Y. Li, X. Su, S. Tarkoma, T. Jiang, J. Crowcroft, and IEEE Transactions on Wireless Communications, vol. 19, no. 12, pp.
P. Hui, “Edge intelligence: Empowering intelligence to the edge of 8083–8098, Dec. 2020.
network,” Proceedings of the IEEE, vol. 109, no. 11, pp. 1778–1837, [33] J. Liu, L. Li, F. Yang, X. Liu, X. Li, X. Tang, and Z. Han, “Minimiza-
Nov. 2021. tion of offloading delay for two-tier UAV with mobile edge computing,”
[15] R. Yang, F. R. Yu, P. Si, Z. Yang, and Y. Zhang, “Integrated blockchain in 15th International Wireless Communications & Mobile Computing
and edge computing systems: A survey, some research issues and Conference (IWCMC), Tangier, Morocco, 2019.
challenges,” IEEE Communications Surveys & Tutorials, vol. 21, no. 2, [34] T. Tozer and D. Grace, “High-altitude platforms for wireless communi-
pp. 1508–1532, Thirdquarter 2019. cations,” Electronics & Communication Engineering Journal, vol. 13,
[16] Q.-V. Pham, F. Fang, V. N. Ha, M. J. Piran, M. Le, L. B. Le, W.-J. no. 3, pp. 127–137, Jun. 2001.
Hwang, and Z. Ding, “A survey of multi-access edge computing in [35] “HAPS – high-altitude platform systems.” [Online].
5G and beyond: Fundamentals, technology integration, and state-of- Available: https://www.itu.int/en/mediacentre/backgrounders/Pages/
the-art,” IEEE Access, vol. 8, pp. 116 974–117 017, Jun. 2020. High-altitude-platform-systems.aspx
[17] P. S. Ranaweera, A. D. Jurcut, and M. Liyanage, “Survey on multi- [36] A. Pregler, “When COWs fly: AT&T sending LTE signals from drones.”
access edge computing security and privacy,” IEEE Communications [Online]. Available: https://about.att.com/innovationblog/cows fly
Surveys & Tutorials Computing, vol. 23, no. 2, pp. 1078–1124, Sec- [37] “Project Skybender: Google’s secretive 5G
ondquarter 2021. Internet drone tests revealed.” [Online]. Avail-
[18] F. Tariq, M. R. Khandaker, K.-K. Wong, M. A. Imran, M. Bennis, and able: https://www.theguardian.com/technology/2016/jan/29/
M. Debbah, “A speculative study on 6G,” IEEE Wireless Communica- project-skybender-google-drone-tests-internet-spaceport-virgin-galactic
tions, vol. 27, no. 4, pp. 118–125, Aug. 2020. [38] T. Cozzens, “UAVOS joins HAPS for high-altitude un-
[19] S. Chen, Y.-C. Liang, S. Sun, S. Kang, W. Cheng, and M. Peng, manned platforms.” [Online]. Available: https://www.gpsworld.com/
“Vision, requirements, and technology trend of 6G: How to tackle the uavos-joins-haps-for-high-altitude-unmanned-platforms/
challenges of system coverage, capacity, user data-rate and movement [39] E. Howell, “The FCC has approved Amazon’s plan for its
speed,” IEEE Wireless Communications, vol. 27, no. 2, pp. 218–228, kuiper satellite constellation.” [Online]. Available: www.space.com/
Apr. 2020. amazon-kuiper-satellite-constellation-fcc-approval.html
[20] A. Fotouhi, H. Qiang, M. Ding, M. Hassan, L. G. Giordano, A. Garcia- [40] Z. Zhang, W. Zhang, and F.-H. Tseng, “Satellite mobile edge comput-
Rodriguez, and J. Yuan, “Survey on UAV cellular communications: ing: Improving QoS of high-speed satellite-terrestrial networks using
Practical aspects, standardization advancements, regulation, and secu- edge computing techniques,” IEEE Network, vol. 33, no. 1, pp. 70–76,
rity challenges,” IEEE Communications Surveys & Tutorials, vol. 21, Jan./Feb. 2019.
no. 4, pp. 3417–3442, Fourthquarter 2019. [41] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, “The case for
[21] G. Karabulut Kurt, M. G. Khoshkholgh, S. Alfattani, A. Ibrahim, VM-based cloudlets in mobile computing,” IEEE pervasive Computing,
T. S. J. Darwish, M. S. Alam, H. Yanikomeroglu, and A. Yongacoglu, vol. 8, no. 4, pp. 14–23, Oct.-Dec. 2009.
“A vision and framework for the high altitude platform station (HAPS) [42] Q.-V. Pham, T. Huynh-The, M. Alazab, J. Zhao, and W.-J. Hwang,
networks of the future,” IEEE Communications Surveys & Tutorials, “Sum-rate maximization for UAV-assisted visible light communications
vol. 23, no. 2, pp. 729–779, Secondquarter 2021. using NOMA: Swarm intelligence meets machine learning,” IEEE
[22] X. Cao, P. Yang, M. Alzenad, X. Xi, D. Wu, and H. Yanikomeroglu, Internet of Things Journal, vol. 7, no. 10, pp. 10 375–10 387, Oct.
“Airborne communication networks: A survey,” IEEE Journal on 2020.
Selected Areas in Communications, vol. 36, no. 9, pp. 1907–1926, Sep. [43] A. Yousefpour, C. Fung, T. Nguyen, K. Kadiyala, F. Jalali, A. Ni-
2018. akanlahiji, J. Kong, and J. P. Jue, “All one needs to know about fog
[23] O. Kodheli et al., “Satellite communications in the new space era: computing and related edge computing paradigms: A complete survey,”
A survey and future challenges,” IEEE Communications Surveys & Journal of Systems Architecture, vol. 98, pp. 289–330, Sep. 2019.
Tutorials, vol. 23, no. 1, pp. 70–109, Firstquarter 2021. [44] R. Roman, J. Lopez, and M. Mambo, “Mobile edge computing, fog et
[24] R. Xie, Q. Tang, Q. Wang, X. Liu, F. R. Yu, and T. Huang, “Satellite- al.: A survey and analysis of security threats and challenges,” Future
terrestrial integrated edge computing networks: Architecture, chal- Generation Computer Systems, vol. 78, pp. 680–698, Jan. 2018.
lenges, and open issues,” IEEE Network, vol. 34, no. 3, pp. 224–231, [45] D. Zeng, N. Ansari, M.-J. Montpetit, E. M. Schooler, and D. Tarchi,
May/Jun. 2020. “Guest editorial: In-network computing: Emerging trends for the edge-
[25] A. Baltaci, E. Dinc, M. Ozger, A. Alabbasi, C. Cavdar, and D. Schupke, cloud continuum,” IEEE Network, vol. 35, no. 5, pp. 12–13, Sep./Oct.
“A survey of wireless networks for future aerial communications 2021.
(FACOM),” IEEE Communications Surveys & Tutorials, vol. 23, no. 4, [46] Z. Xie, J. Liu, M. Sheng, N. Zhao, and J. Li, “Exploiting aerial
pp. 2833–2884, Fourthquarter 2021. computing for air-to-ground coverage enhancement,” IEEE Wireless
[26] N.-N. Dao, Q.-V. Pham, N. H. Tu, T. T. Thanh, V. N. Q. Bao, Communications, vol. 28, no. 5, pp. 50–58, Oct. 2021.
D. S. Lakew, and S. Cho, “Survey on aerial radio access networks: [47] L. Liu, B. Sun, Y. Wu, and D. H. Tsang, “Latency optimization for
Toward a comprehensive 6G access infrastructure,” IEEE Communica- computation offloading with hybrid NOMA-OMA transmission,” IEEE
tions Surveys & Tutorials Computing, vol. 23, no. 2, pp. 1193–1225, Internet of Things Journal, vol. 8, no. 8, pp. 6677–6691, Apr. 2021.
Secondquarter 2021. [48] K. Dorling, J. Heinrichs, G. G. Messier, and S. Magierowski, “Vehicle
[27] Y. Sun, J. Liu, J. Wang, Y. Cao, and N. Kato, “When machine learning routing problems for drone delivery,” IEEE Transactions on Systems,
meets privacy in 6G: A survey,” IEEE Communications Surveys & Man, and Cybernetics: Systems, vol. 47, no. 1, pp. 70–85, Jan. 2017.
Tutorials, vol. 22, no. 4, pp. 2694–2724, Fourthquarter 2020. [49] R. Austin, Unmanned aircraft systems: UAVs design, development and
[28] D. C. Nguyen, P. Cheng, M. Ding, D. Lopez-Perez, P. N. Pathirana, deployment. John Wiley & Sons, 2011, vol. 54.
J. Li, A. Seneviratne, Y. Li, and H. V. Poor, “Enabling AI in future [50] B. Zhu, E. Bedeer, H. H. Nguyen, R. Barton, and J. Henry, “UAV
wireless networks: A data life cycle perspective,” IEEE Communica- trajectory planning in wireless sensor networks for energy consumption
tions Surveys & Tutorials, vol. 23, no. 1, pp. 553–595, Firstquarter minimization by deep reinforcement learning,” IEEE Transactions on
2021. Vehicular Technology, vol. 70, no. 9, pp. 9540–9554, Aug. 2021.
[29] M. Mukherjee, L. Shu, and D. Wang, “Survey of fog computing: [51] Y. Su, Y. Liu, Y. Zhou, J. Yuan, H. Cao, and J. Shi, “Broadband LEO
Fundamental, network applications, and research challenges,” IEEE satellite communications: Architectures and key technologies,” IEEE
Communications Surveys & Tutorials, vol. 20, no. 3, pp. 1826–1857, Wireless Communications, vol. 26, no. 2, pp. 55–61, Apr. 2019.
Thirdquarter 2018. [52] Y. Zhou, C. Pan, P. L. Yeoh, K. Wang, M. Elkashlan, B. Vucetic,
[30] B. Wu, K. Xu, Q. Li, S. Ren, Z. Liu, and Z. Zhang, “Toward and Y. Li, “Communication-and-computing latency minimization for
blockchain-powered trusted collaborative services for edge-centric net- UAV-enabled virtual reality delivery systems,” IEEE Transactions on
works,” IEEE Network, vol. 34, no. 2, pp. 30–36, Mar./Apr. 2020. Communications, vol. 69, no. 3, pp. 1723–1735, Nov. 2021.
[31] D.-H. Tran, V.-D. Nguyen, S. Chatzinotas, T. X. Vu, and B. Ottersten, [53] T. Zhang, Y. Xu, J. Loo, D. Yang, and L. Xiao, “Joint computation
“UAV relay-assisted emergency communications in IoT networks: and communication design for UAV-assisted mobile edge computing
in IoT,” IEEE Transactions on Industrial Informatics, vol. 16, no. 8,
pp. 5505–5516, Oct. 2019.
 23

[54] D. Wu, X. Huang, X. Xie, X. Nie, L. Bao, and Z. Qin, “LEDGE: [75] Z. M. Fadlullah and N. Kato, “On smart IoT remote sensing over
Leveraging edge computing for resilient access management of mobile integrated terrestrial-aerial-space networks: An asynchronous federated
IoT,” IEEE Transactions on Mobile Computing, vol. 20, no. 3, pp. learning approach,” IEEE Network, vol. 35, no. 5, pp. 129–135,
1110–1125, Mar. 2021. Sep./Oct. 2021.
[55] S. Khan, A. Gani, A. W. Abdul Wahab, M. Guizani, and M. K. [76] A. Imran, A. Zoha, and A. Abu-Dayya, “Challenges in 5G: how to
Khan, “Topology discovery in software defined networks: Threats, empower SON with big data for enabling 5G,” IEEE Network, vol. 28,
taxonomy, and state-of-the-art,” IEEE Communications Surveys & no. 6, pp. 27–33, Nov./Dec. 2014.
Tutorials, vol. 19, no. 1, pp. 303–324, Firstquarter 2017. [77] Y. Liu, S. Bi, Z. Shi, and L. Hanzo, “When machine learning meets
[56] S. Tomovic and I. Radusinovic, “Toward a scalable, robust, and QoS- big data: A wireless communication perspective,” IEEE Vehicular
aware virtual-link provisioning in SDN-based ISP networks,” IEEE Technology Magazine, vol. 15, no. 1, pp. 63–72, Mar. 2020.
Transactions on Network and Service Management, vol. 16, no. 3, pp. [78] A. Guo, C. Yuan, G. He, and L. Xu, “Research on SDN/NFV network
1032–1045, Jul. 2019. traffic management and optimization based on big data and artificial
[57] L. Nkenyereye, J. Y. Hwang, Q.-V. Pham, and J. S. Song, “Virtual intelligence,” in 18th International Symposium on Communications and
IoT service slice functions for multi-access edge computing platform,” Information Technologies (ISCIT), Bangkok, Thailand, Dec. 2018.
IEEE Internet of Things Journal, vol. 8, no. 14, pp. 11 233–11 248, Jul. [79] S. I. Kim and H. S. Kim, “A research on dynamic service function
2021. chaining based on reinforcement learning using resource usage,” in
[58] J. Huang, Y. Zhou, Z. Ning, and H. Gharavi, “Wireless power transfer Ninth International Conference on Ubiquitous and Future Networks
and energy harvesting: Current status and future prospects,” IEEE (ICUFN), Milan, Italy, 2017.
Wireless Communications, vol. 26, no. 4, pp. 163–169, Aug. 2019. [80] K. Zhang, D. Si, W. Wang, J. Cao, and Y. Zhang, “Transfer learning
[59] M. Lu, M. Bagheri, A. P. James, and T. Phung, “Wireless charging for distributed intelligence in aerial edge networks,” IEEE Wireless
techniques for UAVs: A review, reconceptualization, and extension,” Communications, vol. 28, no. 5, pp. 74–81, Oct. 2021.
IEEE Access, vol. 6, pp. 29 865–29 884, May 2018. [81] S. Sezer, S. Scott-Hayward, P. K. Chouhan, B. Fraser, D. Lake,
[60] P. K. Chittoor, B. Chokkalingam, and L. Mihet-Popa, “A review on J. Finnegan, N. Viljoen, M. Miller, and N. Rao, “Are we ready for
UAV wireless charging: Fundamentals, applications, charging tech- SDN? implementation challenges for software-defined networks,” IEEE
niques and standards,” IEEE Access, vol. 9, pp. 69 235–69 266, 2021. Communications Magazine, vol. 51, no. 7, pp. 36–43, Jul. 2013.
[61] M. Vaezi, G. A. Aruma Baduge, Y. Liu, A. Arafa, F. Fang, and [82] I. Afolabi, T. Taleb, K. Samdanis, A. Ksentini, and H. Flinck, “Network
Z. Ding, “Interplay between NOMA and other emerging technologies: slicing and softwarization: A survey on principles, enabling tech-
A survey,” IEEE Transactions on Cognitive Communications and nologies, and solutions,” IEEE Communications Surveys & Tutorials,
Networking, vol. 5, no. 4, pp. 900–919, Aug. 2019. vol. 20, no. 3, pp. 2429–2453, Thirdquarter 2018.
[62] Q.-V. Pham, M. Zeng, R. Ruby, T. Huynh-The, and W.-J. Hwang, “UAV [83] C. Benzaid and T. Taleb, “AI-driven zero touch network and service
communications for sustainable federated learning,” IEEE Transactions management in 5G and beyond: Challenges and research directions,”
on Vehicular Technology, vol. 70, no. 4, pp. 3944–3948, Apr. 2021. IEEE Network, vol. 34, no. 2, pp. 186–194, Mar./Apr. 2020.
[63] L. Xie, X. Cao, J. Xu, and R. Zhang, “UAV-enabled wireless power [84] P. L. Richardson, “Upwind dynamic soaring of albatrosses and UAVs,”
transfer: A tutorial overview,” IEEE Transactions on Green Communi- Progress in Oceanography, vol. 130, pp. 146–156, Jan. 2015.
cations and Networking, vol. 5, no. 4, pp. 2042–2064, Dec. 2021. [85] L. Xie, J. Xu, and R. Zhang, “Throughput maximization for UAV-
[64] Y. Yixin, S. Wan, and X. Zhang, “Design of UAV wireless power enabled wireless powered communication networks,” IEEE Internet of
transmission system based on coupling coil structure optimization,” J. Things Journal, vol. 6, no. 2, pp. 1690–1703, Apr. 2019.
Wireless Com. Network, vol. 2020, no. 1, pp. 1–13, Mar. 2020. [86] Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, I. Chih-Lin, and H. V.
[65] P. X. Nguyen, D.-H. Tran, O. Onireti, P. T. Tin, S. Q. Nguyen, Poor, “Application of non-orthogonal multiple access in LTE and 5G
S. Chatzinotas, and H. V. Poor, “Backscatter-assisted data offloading networks,” IEEE Communications Magazine, vol. 55, no. 2, pp. 185–
in OFDMA-based wireless-powered mobile edge computing for IoT 191, Feb. 2017.
networks,” IEEE Internet of Things Journal, vol. 8, no. 11, pp. 9233– [87] O. Elijah, C. Y. Leow, T. A. Rahman, S. Nunoo, and S. Z. Iliya, “A
9243, Jun. 2021. comprehensive survey of pilot contamination in massive MIMO—5G
[66] C. de Alwis, A. Kalla, Q.-V. Pham, P. Kumar, K. Dev, W.-J. Hwang, system,” IEEE Communications Surveys & Tutorials, vol. 18, no. 2,
and M. Liyanage, “Survey on 6G frontiers: Trends, applications, pp. 905–923, Secondquarter 2016.
requirements, technologies and future research,” IEEE Open Journal [88] E. Björnson, E. G. Larsson, and T. L. Marzetta, “Massive MIMO: ten
of the Communications Society, vol. 2, pp. 836–886, Apr. 2021. myths and one critical question,” IEEE Communications Magazine,
[67] J. Jiao, Y. He, S. Wu, and Q. Zhang, “Intelligent hybrid non-orthogonal vol. 54, no. 2, pp. 114–123, Feb. 2016.
multiple access relaying for vehicular networks in 6G,” IEEE Internet [89] K. Senel, H. V. Cheng, E. Björnson, and E. G. Larsson, “What role
of Things Journal, vol. 8, no. 19, pp. 14 773–14 786, Oct. 2021. can NOMA play in massive MIMO?” IEEE Journal of Selected Topics
[68] Z. Chen, X. Ma, C. Han, and Q. Wen, “Towards intelligent reflecting in Signal Processing, vol. 13, no. 3, pp. 597–611, Jun. 2019.
surface empowered 6G Terahertz communications: A survey,” China [90] J. C. B. Garcia, A. Sibille, and M. Kamoun, “Reconfigurable intelligent
Communications, vol. 18, no. 5, pp. 93–119, May 2021. surfaces: Bridging the gap between scattering and reflection,” IEEE
[69] N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y.-C. Journal on Selected Areas in Communications, vol. 38, no. 11, pp.
Liang, and D. I. Kim, “Applications of deep reinforcement learning 2538–2547, Nov. 2020.
in communications and networking: A survey,” IEEE Communications [91] M. Zeng, W. Hao, O. A. Dobre, Z. Ding, and H. V. Poor, “Massive
Surveys & Tutorials, vol. 21, no. 4, pp. 3133–3174, Fourthquarter 2019. MIMO-assisted mobile edge computing: Exciting possibilities for com-
[70] C. Bai, M. Zhang, J. Zhang, J. Zheng, and S. Chen, “LSCIDMR: Large- putation offloading,” IEEE Vehicular Technology Magazine, vol. 15,
scale satellite cloud image database for meteorological research,” IEEE no. 2, pp. 31–38, Jun. 2020.
Transactions on Cybernetics, 2021, in press. [92] F. Fang, Y. Xu, Q.-V. Pham, and Z. Ding, “Energy-efficient design of
[71] C. She, C. Sun, Z. Gu, Y. Li, C. Yang, H. V. Poor, and B. Vucetic, IRS-NOMA networks,” IEEE Transactions on Vehicular Technology,
“A tutorial on ultrareliable and low-latency communications in 6G: vol. 69, no. 11, pp. 14 088–14 092, Nov. 2020.
Integrating domain knowledge into deep learning,” Proceedings of the [93] W. Y. B. Lim, N. C. Luong, D. T. Hoang, Y. Jiao, Y.-C. Liang, Q. Yang,
IEEE, vol. 109, no. 3, pp. 204–246, Mar. 2021. D. Niyato, and C. Miao, “Federated learning in mobile edge networks:
[72] L. Qian, Y. Wu, F. Jiang, N. Yu, W. Lu, and B. Lin, “NOMA assisted A comprehensive survey,” IEEE Communications Surveys & Tutorials,
multi-task multi-access mobile edge computing via deep reinforcement vol. 22, no. 3, pp. 2031–2063, Thirdquarter 2020.
learning for industrial internet of things,” IEEE Transactions on Indus- [94] D. C. Nguyen, M. Ding, Q.-V. Pham, P. N. Pathirana, L. B. Le,
trial Informatics, vol. 17, no. 8, pp. 5688–5698, Aug. 2021. A. Seneviratne, J. Li, D. Niyato, and H. V. Poor, “Federated learning
[73] M. Chen, Z. Yang, W. Saad, C. Yin, H. V. Poor, and S. Cui, “A joint meets blockchain in edge computing: Opportunities and challenges,”
learning and communications framework for federated learning over IEEE Internet of Things Journal, vol. 8, no. 16, pp. 12 806–12 825,
wireless networks,” IEEE Transactions on Wireless Communications, Aug. 2021.
vol. 20, no. 1, pp. 269–283, Jan. 2021. [95] C. Colman-Meixner et al., “Deploying a novel 5G-enabled architecture
[74] Q.-V. Pham, M. Le, T. Huynh-The, Z. Han, and W.-J. Hwang, “Energy- on city infrastructure for ultra-high definition and immersive media
efficient federated learning over UAV-enabled wireless powered com- production and broadcasting,” IEEE Transactions on Broadcasting,
munications,” IEEE Transactions on Vehicular Technology, 2022, in vol. 65, no. 2, pp. 392–403, Jun. 2019.
press.
 24

[96] F. Cirillo, D. Gómez, L. Diez, I. Elicegui Maestro, T. B. J. Gilbert, [117] K. Kuru, “Planning the future of smart cities with swarms of fully
and R. Akhavan, “Smart city IoT services creation through large-scale autonomous unmanned aerial vehicles using a novel framework,” IEEE
collaboration,” IEEE Internet of Things Journal, vol. 7, no. 6, pp. 5267– Access, vol. 9, pp. 6571–6595, Jan. 2021.
5275, Jun. 2020. [118] Y. Chen, S. Wang, J. Li, L. Ma, R. Wu, Z. Luo, and C. Wang, “Rapid
[97] Y. Liu, K. Liu, J. Han, L. Zhu, Z. Xiao, and X. G. Xia, “Resource urban roadside tree inventory using a mobile laser scanning system,”
allocation and 3-D placement for UAV-enabled energy-efficient IoT IEEE Journal of Selected Topics in Applied Earth Observations and
communications,” IEEE Internet of Things Journal, vol. 8, no. 3, pp. Remote Sensing, vol. 12, no. 9, pp. 3690–3700, Sep. 2019.
1322–1333, Feb. 2021. [119] Z. Gong, J. Li, Z. Luo, C. Wen, C. Wang, and J. Zelek, “Mapping
[98] G. Zhu and K. Huang, “MIMO over-the-air computation for high- and semantic modeling of underground parking lots using a backpack
mobility multimodal sensing,” IEEE Internet of Things Journal, vol. 6, LiDAR system,” IEEE Transactions on Intelligent Transportation Sys-
no. 4, pp. 6089–6103, Aug. 2019. tems, vol. 22, no. 2, pp. 734–746, Feb. 2021.
[99] C. H. Liu, Z. Chen, and Y. Zhan, “Energy-efficient distributed mobile [120] E. Ertugrul, U. Kocaman, and O. K. Sahingoz, “Autonomous aerial
crowd sensing: A deep learning approach,” IEEE Journal on Selected navigation and mapping for security of smart buildings,” in 6th Inter-
Areas in Communications, vol. 37, no. 6, pp. 1262–1276, Jun. 2019. national Istanbul Smart Grids and Cities Congress and Fair (ICSG),
[100] W. Chen, B. Liu, H. Huang, S. Guo, and Z. Zheng, “When UAV swarm Istanbul, Turkey, 2018.
meets edge-cloud computing: The QoS perspective,” IEEE Network, [121] S. Hu, Q. Wu, and X. Wang, “Energy management and trajectory
vol. 33, no. 2, pp. 36–43, Mar./Apr. 2019. optimization for UAV-enabled legitimate monitoring systems,” IEEE
[101] T. Li, K. Ota, T. Wang, X. Li, Z. Cai, and A. Liu, “Optimizing the Transactions on Wireless Communications, vol. 20, no. 1, pp. 142–
coverage via the UAVs with lower costs for information-centric Internet 155, Jan. 2021.
of Things,” IEEE Access, vol. 7, pp. 15 292–15 309, Jan. 2019. [122] Y. Jin, Z. Qian, and W. Yang, “UAV cluster-based video surveillance
[102] G. Xu, Y. Liu, and M. Xing, “Multi-channel synthetic aperture radar system optimization in heterogeneous communication of smart cities,”
imaging of ground moving targets using compressive sensing,” IEEE IEEE Access, vol. 8, pp. 55 654–55 664, Mar. 2020.
Access, vol. 6, pp. 66 134–66 142, Jul. 2018. [123] H. Kim and J. Ben-Othman, “A collision-free surveillance system using
[103] G. Xu, Y. Liu, J. Sheng, G. Jing, X. Lin, and J. Chen, “Parametric smart UAVs in multi domain IoT,” IEEE Communications Letters,
sparse SAR imaging of ground moving targets integrated with motion vol. 22, no. 12, pp. 2587–2590, Dec. 2018.
compensation,” IEEE Access, vol. 6, pp. 73 457–73 464, Dec. 2018. [124] G. K. Verma, B. B. Singh, N. Kumar, and D. He, “CB-PS: An
[104] J. Wang, K. Liu, and J. Pan, “Online UAV-mounted edge server efficient short-certificate-based proxy signature scheme for UAVs,”
dispatching for mobile-to-mobile edge computing,” IEEE Internet of IEEE Systems Journal, vol. 14, no. 1, pp. 621–632, Mar. 2020.
Things Journal, vol. 7, no. 2, pp. 1375–1386, Feb. 2020. [125] K. Gai, Y. Wu, L. Zhu, K.-K. R. Choo, and B. Xiao, “Blockchain-
[105] X. Diao, J. Zheng, Y. Cai, Y. Wu, and A. Anpalagan, “Fair data enabled trustworthy group communications in UAV networks,” IEEE
allocation and trajectory optimization for UAV-assisted mobile edge Transactions on Intelligent Transportation Systems, vol. 22, no. 7, pp.
computing,” IEEE Communications Letters, vol. 23, no. 12, pp. 2357– 4118–4130, Jul. 2021.
2361, Dec. 2019. [126] M. Aloqaily, O. Bouachir, A. Boukerche, and I. A. Ridhawi, “Design
[106] X. Diao, J. Zheng, Y. Wu, Y. Cai, and A. Anpalagan, “Joint trajectory guidelines for blockchain-assisted 5G-UAV networks,” IEEE Network,
design, task data, and computing resource allocations for NOMA-based vol. 35, no. 1, pp. 64–71, Jan./Feb. 2021.
and UAV-assisted mobile edge computing,” IEEE Access, vol. 7, pp. [127] S. A. A. Shah, E. Ahmed, M. Imran, and S. Zeadally, “5G for vehicular
117 448–117 459, 2019. communications,” IEEE Communications Magazine, vol. 56, no. 1, pp.
[107] Z. Tan, H. Qu, J. Zhao, S. Zhou, and W. Wang, “UAV-aided edge/fog 111–117, Jan. 2018.
computing in smart IoT community for social augmented reality,” IEEE [128] J. Hu, C. Chen, L. Cai, M. R. Khosravi, Q. Pei, and S. Wan, “UAV-
Internet of Things Journal, vol. 7, no. 6, pp. 4872–4884, Jun. 2020. assisted vehicular edge computing for the 6G internet of vehicles: archi-
[108] X. Liu, Y. Chen, M. Wei, C. Wang, W. N. Gonçalves, J. M. Junior, and tecture, intelligence, and challenges,” IEEE Communications Standards
J. Li, “Building instance extraction method based on improved hybrid Magazine, vol. 5, no. 2, pp. 12–18, Jun. 2021.
task cascade,” IEEE Geoscience and Remote Sensing Letters, vol. 19, [129] Y. Cao, S. Xu, J. Liu, and N. Kato, “Toward smart and secure V2X
pp. 1–5, Mar. 2022. communication in 5G and beyond: A UAV-enabled aerial intelligent
[109] M. Wan, G. Gu, W. Qian, K. Ren, X. Maldague, and Q. Chen, reflecting surface solution,” IEEE Vehicular Technology Magazine,
“Unmanned aerial vehicle video-based target tracking algorithm using vol. 17, no. 1, pp. 66–73, Mar. 2022.
sparse representation,” IEEE Internet of Things Journal, vol. 6, no. 6, [130] A. Sharma, P. Vanjani, N. Paliwal, C. M. W. Basnayaka, D. N. K.
pp. 9689–9706, Dec. 2019. Jayakody, H.-C. Wang, and P. Muthuchidambaranathane, “Communi-
[110] P. Li, X. Wu, W. Shen, W. Tong, and S. Guo, “Collaboration of cation and networking technologies for UAVs: A survey,” Journal of
heterogeneous unmanned vehicles for smart cities,” IEEE Network, Network and Computer Applications, vol. 168, p. 102739, Oct. 2020.
vol. 33, no. 4, pp. 133–137, Jul./Aug. 2019. [131] Z. Chen, C. Wang, H. Luo, H. Wang, Y. Chen, C. Wen, Y. Yu, L. Cao,
[111] Z. Hu, Z. Bai, Y. Yang, Z. Zheng, K. Bian, and L. Song, “UAV aided and J. Li, “Vehicle detection in high-resolution aerial images based on
aerial-ground IoT for air quality sensing in smart city: Architecture, fast sparse representation classification and multiorder feature,” IEEE
technologies, and implementation,” IEEE Network, vol. 33, no. 2, pp. Transactions on Intelligent Transportation Systems, vol. 17, no. 8, pp.
14–22, Mar./Apr. 2019. 2296–2309, Aug. 2016.
[112] J. Gao, Z. Hu, K. Bian, X. Mao, and L. Song, “AQ360: UAV-aided [132] D. Roy, T. Ishizaka, C. K. Mohan, and A. Fukuda, “Detection of
air quality monitoring by 360-degree aerial panoramic images in urban collision-prone vehicle behavior at intersections using siamese interac-
areas,” IEEE Internet of Things Journal, vol. 8, no. 1, pp. 428–442, tion LSTM,” IEEE Transactions on Intelligent Transportation Systems,
Jan. 2021. 2020, in press.
[113] M. Colli, F. Cassola, F. Martina, E. Trovatore, A. Delucchi, S. Maggi- [133] S. Garg, A. Singh, S. Batra, N. Kumar, and L. T. Yang, “UAV-
olo, and D. D. Caviglia, “Rainfall fields monitoring based on satellite empowered edge computing environment for cyber-threat detection in
microwave down-links and traditional techniques in the city of genoa,” smart vehicles,” IEEE Network, vol. 32, no. 3, pp. 42–51, May/Jun.
IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 9, 2018.
pp. 6266–6280, Sep. 2020. [134] H. El-Sayed, M. Chaqfa, S. Zeadally, and D. Puthal, “A traffic-aware
[114] L. Ma, Y. Li, J. Li, Z. Zhong, and M. A. Chapman, “Generation approach for enabling unmanned aerial vehicles (UAVs) in smart city
of horizontally curved driving lines in HD maps using mobile laser scenarios,” IEEE Access, vol. 7, pp. 86 297–86 305, Jun. 2019.
scanning point clouds,” IEEE Journal of Selected Topics in Applied [135] L. Zhang, Z. Zhao, Q. Wu, H. Zhao, H. Xu, and X. Wu, “Energy-aware
Earth Observations and Remote Sensing, vol. 12, no. 5, pp. 1572– dynamic resource allocation in UAV assisted mobile edge computing
1586, May 2019. over social internet of vehicles,” IEEE Access, vol. 6, pp. 56 700–
[115] L. Liu, H. Ma, S. Chen, X. Tang, J. Xie, G. Huang, and F. Mo, 56 715, Oct. 2018.
“Image-translation-based road marking extraction from mobile laser [136] S. Ortiz, C. T. Calafate, J. Cano, P. Manzoni, and C. K. Toh, “A UAV-
point clouds,” IEEE Access, vol. 8, pp. 64 297–64 309, Apr. 2020. based content delivery architecture for rural areas and future smart
[116] Y. Zang, C. Wang, L. Cao, Y. Yu, and J. Li, “Road network extraction cities,” IEEE Internet Computing, vol. 23, no. 1, pp. 29–36, Jan./Feb.
via aperiodic directional structure measurement,” IEEE Transactions 2019.
on Geoscience and Remote Sensing, vol. 54, no. 6, pp. 3322–3335, [137] S. A. Hadiwardoyo, C. T. Calafate, J. Cano, Y. Ji, E. Hernández-
Jun. 2016. Orallo, and P. Manzoni, “3D simulation modeling of UAV-to-car
communications,” IEEE Access, vol. 7, pp. 8808–8823, Dec. 2019.
 25

[138] R. Lu, R. Zhang, X. Cheng, and L. Yang, “UAV-assisted data dissem- [153] Z. Si, Y. Yu, M. Yang, and P. Li, “Hybrid solar forecasting method using
ination with proactive caching and file sharing in V2X networks,” in satellite visible images and modified convolutional neural networks,”
IEEE Global Communications Conference (GLOBECOM), Waikoloa, IEEE Transactions on Industry Applications, vol. 57, no. 1, pp. 5–16,
HI, USA, 2019. Jan./Feb. 2021.
[139] B. Shang, L. Liu, J. Ma, and P. Fan, “Unmanned aerial vehicle meets [154] H. V. Abeywickrama, B. A. Jayawickrama, Y. He, and E. Dutkiewicz,
vehicle-to-everything in secure communications,” IEEE Communica- “Empirical power consumption model for UAVs,” in IEEE 88th Vehic-
tions Magazine, vol. 57, no. 10, pp. 98–103, Oct. 2019. ular Technology Conference (VTC-Fall), Chicago, IL, USA, Apr 2018.
[140] J. Müller, “Enabling technologies for Industry 5.0,” Directorate- [155] Z. Li, Y. Wang, M. Liu, R. Sun, Y. Chen, J. Yuan, and J. Li, “Energy
General for Research and Innovation (European Commission), Sep. efficient resource allocation for UAV-assisted space-air-ground internet
2020. of remote things networks,” IEEE Access, vol. 7, pp. 145 348–145 362,
[141] H. Li and A. V. Savkin, “Wireless sensor network based navigation Oct. 2019.
of micro flying robots in the industrial internet of things,” IEEE [156] R. Verdone and S. Mignardi, “Joint aerial-terrestrial resource manage-
Transactions on Industrial Informatics, vol. 14, no. 8, pp. 3524–3533, ment in UAV-aided mobile radio networks,” IEEE Network, vol. 32,
Aug. 2018. no. 5, pp. 70–75, Sep./Oct. 2018.
[142] M. D. Bah, A. Hafiane, and R. Canals, “CRowNet: Deep network for [157] U. Challita, A. Ferdowsi, M. Chen, and W. Saad, “Machine learning for
crop row detection in UAV images,” IEEE Access, vol. 8, pp. 5189– wireless connectivity and security of cellular-connected UAVs,” IEEE
5200, Dec. 2020. Wireless Communications, vol. 26, no. 1, pp. 28–35, Feb. 2019.
[143] L. Yang, J. Fan, Y. Liu, E. Li, J. Peng, and Z. Liang, “A review on [158] Y. K. Tun, Y. M. Park, N. H. Tran, W. Saad, S. R. Pandey, and C. S.
state-of-the-art power line inspection techniques,” IEEE Transactions Hong, “Energy-efficient resource management in UAV-assisted mobile
on Instrumentation and Measurement, vol. 69, no. 12, pp. 9350–9365, edge computing,” IEEE Communications Letters, vol. 25, no. 1, pp.
Dec. 2020. 249–253, Jan. 2021.
[144] A. Meloni and L. Atzori, “The role of satellite communications in the [159] Q. Zhou, Z. Qu, S. Guo, B. Luo, J. Guo, Z. Xu, and R. Akerkar, “On-
smart grid,” IEEE Wireless Communications, vol. 24, no. 2, pp. 50–56, device learning systems for edge intelligence: A software and hardware
Apr. 2017. synergy perspective,” IEEE Internet of Things Journal, vol. 8, no. 15,
[145] Z. Ling, D. Zhang, R. C. Qiu, Z. Jin, Y. Zhang, X. He, and H. Liu, pp. 11 916–11 934, Aug. 2021.
“An accurate and real-time method of self-blast glass insulator location [160] A. Razi, F. Afghah, and J. Chakareski, “Optimal measurement policy
based on faster R-CNN and U-net with aerial images,” CSEE Journal for predicting UAV network topology,” in 51st Asilomar Conference
of Power and Energy Systems, vol. 5, no. 4, pp. 474–482, Dec. 2019. on Signals, Systems, and Computers, Pacific Grove, CA, USA, 2017.
[146] X. Tao, D. Zhang, Z. Wang, X. Liu, H. Zhang, and D. Xu, “Detection [161] A. Razi, “Optimal measurement policy for linear measurement systems
of power line insulator defects using aerial images analyzed with with applications to UAV network topology prediction,” IEEE Trans-
convolutional neural networks,” IEEE Transactions on Systems, Man, actions on Vehicular Technology, vol. 69, no. 2, pp. 1970–1981, Feb.
and Cybernetics: Systems, vol. 50, no. 4, pp. 1486–1498, Apr. 2020. 2020.
[147] Z. Liu, X. Wang, and Y. Liu, “Application of unmanned aerial vehicle [162] E. Ghribi, T. T. Khoei, H. T. Gorji, P. Ranganathan, and N. Kaabouch,
hangar in transmission tower inspection considering the risk probabil- “A secure blockchain-based communication approach for UAV net-
ities of steel towers,” IEEE Access, vol. 7, pp. 159 048–159 057, Oct. works,” in IEEE International Conference on Electro Information
2019. Technology (EIT), Chicago, IL, USA, 2020.
[148] H.-M. Chung, S. Maharjan, Y. Zhang, F. Eliassen, and T. Yuan, “Edge [163] H. Qiu and H. Duan, “A multi-objective pigeon-inspired optimization
intelligence empowered UAVs for automated wind farm monitoring approach to UAV distributed flocking among obstacles,” Information
in smart grids,” in IEEE Global Communications Conference, Taipei, Sciences, vol. 509, pp. 515–529, Jan. 2020.
Taiwan, 2020. [164] J. Hu, H. Zhang, L. Song, R. Schober, and H. V. Poor, “Cooperative
[149] Q. Li, Y. Feng, Y. Leng, and D. Chen, “SolarFinder: Automatic internet of UAVs: Distributed trajectory design by multi-agent deep re-
detection of solar photovoltaic arrays,” in ACM/IEEE International inforcement learning,” IEEE Transactions on Communications, vol. 68,
Conference on Information Processing in Sensor Networks (IPSN), no. 11, pp. 6807–6821, Nov. 2020.
Sydney, NSW, Australia, 2020. [165] W. Li, Z. Su, R. Li, K. Zhang, and Y. Wang, “Blockchain-based data
[150] P. Rengaraju, S. R. Pandian, and C. Lung, “Communication networks security for artificial intelligence applications in 6G networks,” IEEE
and non-technical energy loss control system for smart grid networks,” Network, vol. 34, no. 6, pp. 31–37, Nov./Dec. 2020.
in IEEE Innovative Smart Grid Technologies - Asia (ISGT ASIA), Kuala [166] W. Sun, L. Wang, P. Wang, and Y. Zhang, “Collaborative blockchain
Lumpur, Malaysia, 2014. for space-air-ground integrated networks,” IEEE Wireless Communica-
[151] Q. Yang, D. I. Laurenson, and J. A. Barria, “On the use of LEO satellite tions, vol. 27, no. 6, pp. 82–89, Dec. 2020.
constellation for active network management in power distribution [167] K. Lei, Q. Zhang, J. Lou, B. Bai, and K. Xu, “Securing ICN-based UAV
networks,” IEEE Transactions on Smart Grid, vol. 3, no. 3, pp. 1371– ad hoc networks with blockchain,” IEEE Communications Magazine,
1381, Sep. 2012. vol. 57, no. 6, pp. 26–32, Jun. 2019.
[152] M. De Sanctis, E. Cianca, G. Araniti, I. Bisio, and R. Prasad, “Satellite [168] Y. Wang, Z. Su, N. Zhang, and A. Benslimane, “Learning in the
communications supporting internet of remote things,” IEEE Internet air: Secure federated learning for UAV-assisted crowdsensing,” IEEE
of Things Journal, vol. 3, no. 1, pp. 113–123, Feb. 2016. Transactions on Network Science and Engineering, vol. 8, no. 2, pp.
1055–1069, Apr.-Jun. 2021.