Satellite Computing Vision and Challenges

This article has been accepted for publication in IEEE Internet of Things Journal.
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2023.3303346
IEEE INTERNET OF THINGS JOURNAL, VOL. XX, NO. XX, 2023 1
Satellite Computing: Vision and Challenges

Shangguang Wang, Senior Member, IEEE, Qing Li
Abstract—The space industry experiences a rise in low-earth- raw data, identify the features of interest, and transmit only
orbit satellite mega-constellations to achieve universal connectiv- the interesting data, improving the transmission efficiency and
ity. At the same time, cloud firms (such as Google, Microsoft, reducing the ground infrastructure costs [1]. Third, advanced
and AWS) also have ambitions for computing in space to offer
public cloud services in orbit. Satellite computing, as a new computing platforms onboard transform constellations into
emerging concept, is promising to enable new paradigms in on- sophisticated data processing infrastructure and enable public
orbit autonomy, remote sensing, edge computing, and other areas cloud services in space as they exist today on the ground [3].
by empowering satellites with computing resources. However, The LEO mega-constellation is a new type of infrastructure
LEO mega-constellations bring inherent challenges for satellite and brings inherent challenges for satellite computing which
computing in satellite networking, computing, and others due to
the moving core infrastructure, reduced system power budget, is fundamentally different from cloud computing and edge
and harsh space environment. This paper presents vision and computing on the ground. LEO mega-constellations consist of
challenges for satellite computing based on a brief survey of thousands of satellites and each moves at high speed relative
the very recent literature in the “NewSpace” era and gives a to the Earth and other satellites. For example, a satellite at
case study of an open research platform on real satellites named an altitude of 550km must maintain a speed of 27,000km/h to
Tiansuan constellation. This paper aims to call researchers to
collaboratively undertake the research of satellite computing and maintain its orbit [4]. Besides, satellites in mega-constellations
provide some insights for the research community. usually have very limited weight and volume due to the
reduced manufacturing and launching cost. Moreover, the
Keywords-Satellite Computing, Satellite Internet, Testbed
space environment is harsh due to deep vacuum conditions,
radiation, strong vibrations, and higher temperature ranges [5].
I. I NTRODUCTION These factors make networking, computing, and other research
Recently, multiple commercial powers are gearing up to in satellite computing challenging. First, satellite network has
deploy LEO mega-constellations (with hundreds to tens of generated tremendous interest among networking researchers.
thousands of LEO satellites) to provide global low-latency They have highlighted the new opportunities and challenges
high-bandwidth Internet. Dated to July 2023, SpaceX has of LEO mega-constellations [6]–[8] and explored network
launched more than 4700 Starlinks and London-based OneWeb topology [4], intra-constellation routing [9], inter-domain rout-
has launched 620 internet satellites.1 With the exciting devel- ing [10]. But, it is unclear how to integrate the frequently
opment of satellite networks, satellites are equipped with more changed satellite networks with the Internet’s Border Gate-
powerful computing resources to supports in-orbit data pro- way Protocol and how to achieve congestion control. At the
cessing [1]. Cloud firms (Google, Microsoft, and AWS) also same time, some researchers propose orbit edge computing to
have ambitions for computing in space by forging significant process the space-native data [1] and provide public service for
alliances with leading satellite companies. Mega-constellations ground users [11]. However, it is still unsettled how to design
offer public cloud services in orbit so they are another hybrid space environment adaptive computing hardware, construct
cloud option for global users. computing platforms, and develop organization paradigms for
As satellite computing is still a concept, it deploys com- applications due to the moving service infrastructure, reduced
puting resources at the satellites to enable new paradigms system power budget, and harsh space environment. There are
in on-orbit autonomy, remote sensing, edge computing, and also many open questions about how to realize reliable and
other areas [2]. First, traditional satellite systems adopt a bent- secure satellite computing.
pipe architecture, where ground stations send human-operated This paper aims to present vision and challenges for satellite
commands to orbit and satellites reply with raw data (just like computing in the ”NewSpace” era. We first take a quick
a bent pipe). This human-in-the-loop architecture breaks down look at the current development of satellite constellations
in face of the mega-constellation scale [2]. Satellite computing (Section II). We then review the recent work in satellite
can enable on-orbit self-operation and reduce the reliance computing (Section III), discuss the challenges, and present
on the ground segment. Second, space-native raw data are several potential research topics (Section IV). We also give
increasing explosively with the constellation size and cannot a case study of an open research platform on real satellites
be downloaded in time due to the limited satellite-ground link named Tiansuan constellation and show some experiments
bandwidth. Satellites with computing power can process the deployed on Tiansuan constellation (Section V). Finally, we
conclude the paper and call researchers to collaboratively
Shangguang Wang is with the State Key Laboratory of Networking and
Switching Technology, Beijing University of Posts and Telecommunications, undertake the research on satellite computing (Section VI).
Beijing, China, 100876. E-mail: [email protected].
Qing Li is with the School of Computer Science, Peking University, II. A G LANCE AT THE E XISTING C ONSTELLATIONS
Beijing, China, 100871. E-mail: [email protected]. Qing Li is the
corresponding author. Recent years witness the surge of LEO mega-constellations,
1 https://en.wikipedia.org/wiki/List of Starlink and Starshield launches where hundreds or thousands of LEO satellites move around
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Internet of Things Journal. This is the author's version which has not been fully edited and
the earth at the altitude of about 500-2000 km over the sea work into three categories: networking, computing, and others
level at a high speed. Most LEO satellites are miniaturized as follows.
at lower cost, such as minisatellites, microsatellites, and
nanosatellites. Their orbits are usually determined by several
A. Satellite Network
key parameters, such as orbit plane number, satellite number of
each plane, inclination, altitude, and phase shift. For example, Back in the 1990s, Geosynchronous Orbit (GEO) and small
the planned Starlink Gen1 comprises 4,425 satellites. The first LEO satellites networks were heavily researched, but more
phase uses 24 orbits, each with 66 satellites, for a total of 1,584 recently, position papers have shed light on the exciting
satellites. Orbit inclination is 53° and the altitude is about 550 new possibilities and obstacles that come with LEO mega-
km. constellations. These papers [6]–[8] have pointed out that
In this section, we show the top 20 constellations (in terms satellite networks have immense potential for low-latency
of the launched satellite number) in TABLE I mainly according long-distance communication, and that there is still a lot of
to the statistics from NewSpace Index2 and update the data unexplored territory when it comes to network topology, intra-
dated by July 2023. Only about 4% of the constellations constellation routing, inter-domain routing, reordering, and
recorded by NewSpace Index have been fully launched and congestion control.
8% are currently being launched, so the information of the 1) Physical Layer: Compared with bent-pipe architecture
top 20 constellations tells a lot. From the launched satel- without inter-satellite links, Singla et al. [12] analyze the
lite number, we can observe that the large constellations benefit of inter-satellite links in improving satellite network
are dominated by only several organizations. This gap may latency, throughput, and reliability. Besides, to enable the
become even larger in the future. From the first launch research in the NewSpace era, they develop a packet-level
date, 19 of the 20 constellations are launched after 2010, satellite network simulation tool for simulating and visualizing
which implies the surge of the NewSapce era. From the form the network behavior of mega-constellations [9]. They design
factor of satellites, we can tell that small satellites (weight a new network topology to maximize network bandwidth and
lower than 500Kg) dominate the LEO mega-constellations. minimize latency after analyzing the limitation of the existing
From the applications, constellations are widely applied for +Grid topology [4]. The network topology is generalized from
Internet, Earth observation, IoT/M2M, AIS (Automatic Identi- +Grid pattern by repeating a motif for each satellite in the
fication System), ADS-B (Automatic Dependent Surveillance- network, all connected in exactly the same way, where the
Broadcast), GNSS (Global Navigation Satellite System), PTN connection remains stable over time. Kur et al. [13] conduct a
(Positioning, Timing, Navigation) and so on. strategic evaluation of a simulated inter-satellite link in Galileo
The industry has a new expectation of mega-constellations satellites, utilizing variance component estimation of varying
in application fields of the Internet, Earth observation, and complexity to test performance. The results of the evaluation
IoT fields. Satellite Internet is promising to provide high- indicate that incorporating different estimation methods in
speed, low-latency broadband internet across the globe. For inter-satellite link measurements could enhance link accuracy.
example, Starlink is available for Internet service for well over However, the paper does not provide clarity on the deviation
1.5 million users globally3 . Crisis on the earth such as the war range and calibration method of the inter-satellite link. Arslan
in Ukraine, climate change, and the covid pandemic fuel the et al. [14] use signal attenuation of satellite links to estimate
demand for earth observation. For example, Earth observation rainfall accumulation. The results of their study, which focused
companies like BlackSky or Planet use their constellation on Ku-band link measurements, indicate that satellite link
to monitor the activities interesting for customers such as attenuation is highly sensitive to heavy rainfall. Therefore, it
investors and financial firms4 . Satellite IoT helps modern-day is necessary to improve the sensitivity and prediction accuracy
businesses and organizations to track, monitor, and manage to small rainfall.
assets anywhere on the Earth, improving remote operations. 2) Network Layer: Routing is a fundamental function and
For example, the Iridium network is uniquely qualified to core task of communication networks, and satellite networks
provide global satellite IoT services by 66 LEO satellites are no exception. As such, the development of routing tech-
blanketing the earth with reliable and ubiquitous coverage5 . nologies in satellite networks is a major focus, aimed at en-
hancing user service quality and improving business efficiency.
In order to deal with the high cost of two-point communica-
III. L ITERATURE R EVIEW tion in satellite networks, Lan et al. [15] propose a traffic
While the industry is gearing up to deploy LEO mega- scheduling strategy based on multiple routing planes, which
constellations, the research community is also interested in the flexibly utilizes the diversity of network links to simplify the
new satellite infrastructure. Recent work has explored the new path selection problem. The results show that this method
research filed from many aspects. We summarize the related can significantly improve network throughput and reduce
latency. Highly dynamic and time-varying topology are typical
2 https://www.newspace.im characteristics of LEO satellite networks. To address this issue,
3 https://www.starlink.com
Lai et al. [16] propose a hybrid data transmission architecture
4 https://spacenews.com/a-boom-in-earth-observation-satellites-creating-
that combines LEO constellations and ground distributed base
new-demands-for-intelligence/
5 https://www.iridium.com/blog/2021/04/01/what-is-satellite-iot-and-how- stations. This architecture aims to achieve stable communica-
is-it-used/ tion and routing in satellite networks. The authors develop a
TABLE I: Global Top 20 Constellations Information (Dated by July 2023).

Index Organization Launched /Planned Number First Launch Form Factor Application
1 SpaceX (Starlink Gen1) 4700 / 4408 2018 Smallsat Internet
2 OneWeb 620 / 648 2019 Smallsat Internet
3 Planet (Flock/ Dove/SuperDove) 555 / 150 2013 CubeSat Earth Observation
4 Swarm Technologies 189 / 150 2018 CubeSat IoT / M2M
5 Spire (Lemur / Minas) 166 / 150 2013 CubeSat Weather, AIS, ADS-B
6 Chang Guang (Jilin-1) 89 / 300 2015 Satellite Earth Observation
7 Aireon 75 / 75 2017 Hosted ADS-B
8 DDK Positioning 75 / 75 2017 Hosted GNSS
9 Iridium (NEXT) 75 / 75 2017 Satellite Internet, IoT / M2M
10 ExactEarth 68 / 67 2008 Hosted, Microsat AIS
11 Satelles 66 / 66 2017 Satellite GNSS, PNT
12 Orbcomm (OG2) 51 / 52 2012 Microsat, Smallsat, CubeSat IoT / M2M, AIS
13 Satellogic 39 / 90 2016 Microsat Earth Observation
14 Globalstar (Second-Generation) 25 / 42 2010 Satellite Internet, IoT / M2M
15 SES (O3b / mPOWER) 24 / 70 2013 Satellite Internet
16 BlackSky 20 / 16 2016 Microsat Earth Observation
17 Astrocast 20 / 80 2018 CubeSat IoT / M2M
18 Kepler Communications 19 / 140 2018 CubeSat IoT / M2M, Internet
19 Spacety 18 / 480 2018 CubeSat Earth Observation
20 Planet (Terra Bella / Skybox) 15 / 24 2013 Smallsat Earth Observation
simulation platform, driven by public constellation informa- 3) Transport Layer: Satellite communication systems are
tion, to verify the performance of the proposed scheme in crucial in providing broadband services globally and are
reducing latency. However, the data transmission architecture becoming increasingly significant in terms of strategic impor-
is unable to preprocess remote sensing data, which increases tance. However, congestion control remains a concern when
the pressure on data transmission. The instability of satellite it comes to satellite transmissions. Page et al. [21] suggest a
links poses a challenge to the direct application of the shortest distributed probabilistic congestion control scheme that utilizes
path algorithm on the ground to satellite networks. Zhang a datagram routing algorithm to compute the minimum delay
et al. [17] propose a scalable two-layer routing architecture, path between any two satellite nodes. The scheme’s implemen-
which supports the implementation of two algorithms: delay- tation assumes a fixed topology of the constellation. However,
bounded routing and delay-aware routing. The results indicate since satellites are in constant motion, the satellite network
that both algorithms can significantly optimize the average topology is subject to continuous changes. As a result, the
network latency. However, the authors do not combine the proposed scheme lacks practicality. Bui et al. [22] introduce a
advantages of the two algorithms, leading to the requirement congestion control strategy for high-throughput satellite com-
of a larger forwarding table. Handly et al. [18] suggest using munication that is based on power control. By utilizing a multi-
ground relays instead of inter-satellite links to reduce access objective optimization framework, a trade-off is made between
latency. They implement an enhanced routing algorithm for the system speed and the number of users satisfying the QoS
large networks, with the result that lower latency can be requirement. Liu et al. [23] explore the impact of various
achieved. But it ignores the case where the link changes TCP congestion control algorithms on commercial satellite
due to satellite movement. Wang et al. [19] investigate the networks, highlighting how performance enhancing proxies
routing issue of link failure in LEO satellite networks to can significantly improve throughput. Meanwhile, Claypool et
enhance network survivability. In satellite network design, al. [24] compare different congestion control algorithms for
ensuring routing security is of paramount importance. Due satellite networks and not only find differences in throughput
to the resource-constrained and rapidly changing nature of but also in round-trip times. However, the unique conditions
satellite networks, implementing simple encryption algorithms of the satellite network may lead to a decline in TCP perfor-
to meet demand can be challenging. Zhao et al. [20] introduce mance, particularly when dealing with network jitter and signal
a lightweight risk-averse routing algorithm to mitigate routing interference. Therefore, it is important to consider alternative
risks, which proves to be a successful endeavor in improving protocols to support satellite communications. In this regard,
satellite network routing security. However, packet forwarding Dai et al. [25] suggest a distributed congestion control routing
that avoids high-risk areas comes at the cost of increased link protocol for LEO satellite network traffic classification that
overhead. achieves distributed congestion control performance through
traffic classification. Nevertheless, the artificially divided traf- [34] propose an IP address management scheme for satel-
fic types may not be standardized. lites in mega-constellations. The authors incorporate feature
4) Core Network on Satellites: Deploying a core network information and geographic locations of the two satellites
in aircraft and even in space is a cutting-edge topic that into IPv6 addresses. The geographic address management
presents many unknown challenges. However, extending the mechanism only conducts duplicate address detection on the
mobile core function to remote areas can help avoid the high corresponding satellite routers, which can significantly prolong
costs of ground station construction and operation, as well the lifespan of the satellite’s IPv6 addresses. However, the
as prevent various unpredictable geological disasters. Qazi et study does not include the development of trust mechanisms
al. [26] initially propose extending the evolved packet core for IPv6 addresses or address security issues such as identity
for deployment in IoT and designing an evolved packet core authorization. Furthermore, Lai et al. [35] develop a cost-
architecture to achieve optimal performance and scalability. effective content distribution framework based on mega con-
However, there has been no indication of any trend towards stellations. The authors store a copy of the content on the
deployment at high altitudes yet. Moradi et al. [27] have made LEO satellite or cloud and dynamically allocate user requests
significant progress in deploying the core network on non- to a cache server based on relevant information, significantly
terrestrial networks, successfully decomposing the evolved reducing content access latency. They also create a simulation
packet core into independent, lightweight entities to address platform geared towards mega constellations for performance
the challenges of deploying in resource-constrained and fast- testing [36]. The tool’s effectiveness is demonstrated by eval-
moving scenarios such as UAVs. The UAV-based core network uating and comparing the performance of several typical LEO
can communicate effectively with LTE base stations and constellations. However, the simulation platform lacks access
smartphones. Li et al. [28] conduct a comprehensive study of to traffic data from real satellites, limiting its ability to assess
the onboard core network, focusing on the challenges posed the actual benefits of performance. The LEO satellite network
by the extreme mobility of satellites to the mobile core. The architecture based on software-defined networks (SDN) is
authors highlight issues such as signaling storms, intermittent a promising network deployment architecture that has gar-
failures, and malicious attacks on the stateful core. To address nered attention from researchers. Tang et al. [37] propose
the inherent contradiction between mobility and stateful core, an SDN-based satellite-ground integrated network architecture
they develop a decoupling mechanism that separates the state and investigate the dynamic cooperative transmission problem.
and functions of the core network by saving the core network’s The study yields optimal network transmission characteristics.
state in UEs. They also design and verify corresponding The deployment and assignment of controllers becomes more
strategies through typical core network processes. Finally, difficult due to the highly dynamic and topological randomness
the efficacy of the space core is validated through extensive of the LEO satellite networks. In addition, Chen et al. [38]
simulations. define an adaptive controller configuration problem and then
One limitation of the aforementioned studies is the lack propose to use the control relation graph to measure the
of verification of the proposed schemes or architectures on control overhead of the LEO satellite networks. The controller
actual satellites. However, we have made significant strides assignment and placement algorithm based on the control
by successfully deploying a 5G core network on the Tiansuan relation graph can effectively reduce the cost of network man-
constellation satellite, making history as the first instance of agement and shorten the response time of satellite networks.
this achievement. Through thorough discussions, we demon- However, the model does not take into account factors such
strate the feasibility of deploying the core network on satel- as satellite movement and topology changes. Klenze et al.
lites. Finally, we validate the proper functioning of the primary [39] discuss business and interconnection models for space-
functions of the onboard core network [29]. operating ISPs and study satellite-ground integrated routing
5) Satellite-Ground Integration Network: Scholars and in- [10] to integrate the satellite networks into today’s Internet
dustries have carried out work on satellite-ground integration backbone. To ensure consistent high performance between the
under 5G and B5G [30], [31]. Li et al. [32] propose to use satellite and the ground, congestion control is necessary for
a handover synchronization method to reduce the network the highly heterogeneous satellite-ground integrated network.
convergence time and improve the availability of space-ground Li et al. [40] propose an adaptive congestion control scheme
integrated networks. However, the authors do not discuss based on multi-objective reinforcement learning, training the
the suitability of different routing protocols for the current reinforcement learning agent to adapt to the network envi-
architecture of satellite-ground integrated networks. Ji et al. ronment and balance congestion control objectives. However,
[33] design the network control structure by constructing an as the scale of the satellite-ground network increases, the
optimal spatial control network to achieve the best arrange- performance of the proposed scheme requires improvement.
ment of controllers in mega satellite networks. With the aim The satellite ground station is a crucial element of the
of enhancing the effectiveness and reliability of wireless com- satellite network system architecture, making it possible to op-
munication services, the authors utilize geometric topology timize the satellite network by improving the satellite ground
analysis to determine the optimal deployment conditions for station. However, real-time reception of satellite imagery from
spatial control networks, ultimately striking a balance between LEO satellite ground receivers is challenging due to receiver
network size, number of controllers, and transmission latency. scarcity. Singh et al. [41] propose a satellite receiver system
Although the study takes into account satellite coverage and that maximizes the diversity of LEO satellites by stitching
latency, it does not account for satellite mobility. Chen et al. together noisy images to create a clear image of the Earth.
RF signals can be transformed and aligned based on different there is no complete and effective protocol system to support
tracks, viewing angles, and wireless channel quality. However, the integration of network segments in space, air, ground and
the system does not currently support high-frequency real-time even ocean.
access to satellite images. Additionally, the authors propose
a community-driven distributed reception scheme for LEO
satellite signals to address the high cost of renting ground B. Computing
station infrastructure [3]. The scheme synergistically leverages 1) Satellite Edge Computing: Traditional satellite systems
satellite trajectories and other environmental signals to achieve typically have limited computing capabilities for specific mis-
a new receiver synchronous orientation technique, which can sions and rely on a bent-pipe architecture, which can lead
recover signals required by handheld receivers. Vasisht et to excessive energy loss, system failures, and other issues
al. [42] propose a hybrid distributed satellite ground station when the number of constellations increases. Enhancing the
design to improve the download process of satellite images, computing power of satellites has garnered significant attention
utilizing low-cost commodity hardware with a hybrid deploy as a potential solution. Edge computing offers services that are
model to achieve low latency and stable downlinks. However, located close to the user, addressing the flexibility demands
the study does not account for the effects of weather, ground of satellite network deployment while reducing latency and
station failures, and other factors on data transmission perfor- providing robust computing capabilities. Cassara et al. [50]
mance. The authors of [43] take into account the traffic char- explore the potential advantages of integrating LEO satellite
acteristics and LEO satellite topology to develop an iterative constellations with edge computing, such as enhancing system
satellite ground station deployment scheme that maximizes performance by delegating computing tasks. They also un-
revenue. This solution deploys each satellite ground station derscore the significance of machine learning in ensuring the
at the geographical location that has the largest marginal quality of satellite services. Denby et al. [2] suggest leveraging
benefit, significantly improving system throughput. However, edge computing to overcome the challenges posed by tradi-
the global traffic demand is unbalanced, and the authors do tional centralized architecture. They present a novel approach
not consider the system throughput when selecting the ground for spatially parallel computing and illustrate that an edge
station deployment location. Guo et al. [44] employ UAVs computing-based architecture for satellites can substantially
as repeaters to transmit signals to satellites, taking hardware reduce ground infrastructure. However, their simulation plat-
impairments into consideration. They use a closed expression form is only applicable to satellite remote sensing services and
of the downtime probability to assess the effects of key param- cannot be utilized for other applications. Pfandzelter et al. [51]
eters on the system. Xia et al. [45] propose an SDN-based LEO investigate the feasibility of implementing edge computing
satellite network communication architecture. Since a large in LEO satellite networks. Given the characteristics of LEO
number of routing requests need to be forwarded by satellites satellites, a serverless deployment approach is a promising
with limited processing power, the authors design a layered solution, as determined through theoretical analysis. However,
ground controller architecture. The architecture reuses satellite the authors do not compare the performance of the serverless
ground stations and adopts online network view integration paradigm with other application paradigms. Furthermore, the
to avoid the shortcomings of small ground station coverage. team presents a microVM-based LEO edge virtual platform to
The routing method based on this architecture can realize load simulate the practical effects of deploying edge computing on
awareness and improve system flexibility. The study discusses LEO satellites [52]. Nevertheless, the platform may be subject
GEO-based SDN deployments and does not consider the more to hardware device performance limitations when simulating
prevalent LEO satellites. large LEO constellations. Additionally, simulating the impact
Challenges: The inherent dynamics of LEO satellites create of environmental changes on satellite-ground links and inter-
many challenges at different network layers, such as the satellite links is challenging. Moreover, the authors delve into
network addressing at the physical layer [46], routing at the service placement in satellite edge computing, leveraging the
network layer [47], congestion control at the transport layer topological characteristics of LEO satellite networks to meet
[48], and task scheduling at the application layer [49]. It can the QoS requirements of service placement [53].
be found that there are still deficiencies in the research on Empowering satellites with computing capabilities will im-
satellite networking. Firstly, the mobility modeling of satellites prove the processing efficiency onboard, while it is non-
is insufficient, and some even ignore the mobility. Addressing, trivial to design edge computing architecture and schedule
routing, and congestion control are completely transformed heterogeneous network resources. Xie et al. [54] present a
into static problems, which do not conform to real space satellite-ground edge computing framework that deploys com-
scenarios. Secondly, there is a lack of real and complete puting resources in multi-layer heterogeneous edge clusters.
satellite data sets to support experiments. The experimental The authors highlight that efficient and trustworthy protocol
results obtained through the simulation environment and arti- design, mass user access, and reliable reception of con-
ficially manufactured data sets often deviate from the actual current signals are potential research concerns. Meanwhile,
application. Additionally, various network architectures pro- Tong et al. [55] propose a satellite-ground network resource
posed in the above studies are verified in different simulation allocation and computation offloading decision-making ap-
environments. Although ideal performance is achieved, there is proach based on mobile edge computing. In this scheme,
a lack of deployment experiments in actual satellite networks, resource allocation and computation offloading decisions are
i.e., no real satellite platform to support the conclusion. Finally, decomposable optimization problems, and techniques such
as potential games and the Lagrange multiplier method are vice coverage and satellite energy consumption optimization.
employed to obtain optimal decisions. However, the authors In the first work, we investigate how to deploy services flexibly
do not consider computational offloading between satellites. and efficiently at satellite edge nodes to provide service
Additionally, Gao et al. [56] propose a potential game-based coverage and propose an online service placement algorithm
virtual network function placement method for satellite edge using Lyapunov optimization and Gibbs sampling [11]. The
computing to optimize request deployment costs. The decen- results show that it can significantly improve service coverage.
tralized resource allocation algorithm based on potential games Since the energy-consuming computing component deepens
solves the virtual network function placement problem by battery discharge and impacts battery life, we attempt to reduce
searching for Nash equilibrium. However, this paper does not energy consumption by coordinating the sensing, computing,
examine the consequences and countermeasures of satellite and communication processes for an earth observation mission
topology changes on the service function chain. Tang et al. [68]. Specifically, an energy scheduling algorithm based on
[57] propose a hybrid cloud edge computing LEO satellite online convex optimization is proposed to reduce the depth of
network based on a three-layer computing architecture and battery discharge.
aim to provide diverse computing services for ground users. 2) AI in Space: With the significant advantages that people
The authors design a distributed algorithm to transform a non- have reaped from various AI applications on the ground, there
convex problem into a linear programming problem using a has been a growing interest in extending AI capabilities into
binary-variable relaxation method, optimizing computational space. Kothari et al. [69] concentrate on the implementation of
offloading decisions. Nonetheless, this paper also does not deep learning in satellites and the advantages it offers for space
account for the impact of satellite movement on computing data processing. However, their comparative experiments may
offload, and security during computational offloading also not accurately reflect the real satellite network environment
warrants attention. due to the excessive restrictions that are artificially imposed.
Satellite-ground IoT represents a crucial service paradigm Given that the majority of satellites currently in orbit are
for satellite edge computing [58]–[62]. Wei et al. [63] examine remote sensing satellites, image processing is a fundamental
the impact of edge computing and machine learning on target task in satellite computing. The primary focus of mainstream
image detection in satellite IoT and suggest an intelligent research on AI in space also centers around target detection
application strategy for edge computing in satellite physical or image processing. While satellite image data sets are
networks. However, this approach is only tested in a small- undoubtedly crucial, they also present a bottleneck in this
scale satellite network and necessitates verification on a larger field. Ding et al. [70] create a satellite data set containing
scale with more satellites and satellite data. Song et al. [64] 11,268 images annotated with 18 categories. The authors use
separate the computing offloading of IoT mobile devices and this data set to establish a baseline that consists of 10 advanced
LEO satellites into two components: ground and space. Using algorithms and 70 configurations. This data set provides re-
a unique edge computing framework, the authors perform searchers with extensive opportunities for further research.
computing offloading and resource allocation for satellite- Hoeser et al. [71] conduct research on the impact of con-
ground IoT from two perspectives, including minimizing space volutional neural networks on earth observation applications,
segment delay via Lagrangian dual decomposition methods. providing a theoretical foundation for the application of deep
The findings demonstrate that the proposed approach can learning in satellite image processing. The efficiency of image
significantly lower energy consumption. Nonetheless, the au- processing can be affected by the computing, memory, and
thors do not take into account the involvement of multiple power constraints of satellites. Lofqvist et al. [72] study the
terrestrial satellite terminals in computing offloading. Zhou performance of convolutional neural networks using various
et al. [65] present a handover algorithm for satellite edge image compression methods. Furthermore, they examine the
computing that enables flexible handover and scheduling of inference time and memory consumption of tasks performed
services for fast mobile terminals. The authors suggest that on different hardware. In addition to deep learning, deep
utilizing machine learning-based methods to predict satellite reinforcement learning has also been utilized in the object
trajectories can effectively enhance the algorithm’s efficiency. detection of large satellite images. Uzkent et al. [73] propose
Wang et al. [66] develop a profit maximization model for an adaptive detection method for image resolution based on
satellite-ground edge computing systems to ensure the QoE the reinforcement learning agent, which significantly enhances
of IoT devices. The authors optimize the offloading strategy operational efficiency. Cube satellites have gained increasing
and resource allocation from multiple perspectives, and the attention from manufacturers and commercial companies in
results confirm that the service provider’s profit improves recent years. However, they are severely limited in terms of
while maintaining the QoE. However, the authors overlook power and downlink capabilities. Maskey et al. [74] propose
the possibility that satellite movement may cause computing a downlink image selection scheme based on the lightweight
offloading terminals, thereby impacting the quality of service. convolutional neural network for classification before transmit-
Gost et al. [67] propose a collaborative optimization algo- ting the data to the ground station. Their research has shown
rithm for joint communication and edge computing resource that even deploying a small neural network model on the cube
management. The algorithm enables satellites to select edge or satellites can improve data reception quality.
cloud server computing tasks, reducing satellite system energy Challenges: It is obvious that the current research on satel-
consumption while satisfying end-to-end latency constraints. lite edge computing mainly focuses on computing offloading
We have explored satellite edge computing in terms of ser- [75], resource allocation [76], service placement [38], and
service optimization. It is also because most of them study requirements for satellite hardware and software. Moreover, it
satellite edge computing from the perspective of simulation, is extremely urgent to design and implement a real satellite
so they often ignore the inherent characteristics of satellites. experiment platform to test new payloads and devices.
In particular, the satellite network topology is set to a fixed
state according to the time slot, which will hide many practical IV. R ESEARCH O PPORTUNITES
problems. On the other hand, many of the proposed optimiza-
tion algorithms and applications lack a real satellite platform In this section, we envision several main research directions
for verification, which is also an urgent problem facing the in satellite computing as follows.
academic community.
Satellite computing faces several significant challenges.
A. Networking
Firstly, the computing components in LEO satellites typically
operate under extreme power conditions due to the limited en- It has become a consensus in the industry that satellite con-
ergy generation and storage onboard. This necessitates energy- stellations provide ubiquitous Internet services. Therefore, it is
efficient data processing, as well as position control and necessary to design a new satellite network architecture that
communication. Satellite computing requires new power and can provide efficient, flexible, and customized user services,
energy management systems, computer architecture, and hard- including the physical layer, network layer, transport layer,
ware specialization. Secondly, satellites in mega-constellations and even control plane. In our previous work [5], we propose
have limited computing capabilities and require collaboration the idea of deploying a 5G or even a 6G core network on
to distribute computing tasks for multiple application work- satellites, which is an essential and significant step forward.
loads. The constellation can be viewed as a moving distributed Deploying the core network on satellites would be of
computing infrastructure, and it requires cluster orchestra- great benefit. On the one hand, LEO satellites are made
tion and resource management. Thirdly, the organizational of general-purpose hardware and utilize onboard computing,
paradigm of applications on LEO constellations is still unset- network, and storage resources based on custom software. By
tled. Lastly, it is urgent and necessary for satellite computing to interconnecting the core network with the access network and
be deployed on the real satellite test environment, particularly data network, real-time and reliable customized services can
for AI algorithms. However, the above studies do not involve be provided to users. On the other hand, given the alarming
deployment on real satellites as LEO edge computing testbeds rate at which the number of satellites in space is growing, a
on real satellites are still missing. flexible and efficient unified management method for satellites
is urgently needed. By deploying a lightweight core network
on satellites, onboard resource scheduling and satellite unified
C. Others management can become more convenient and flexible.
There are several new hardware designs for LEO small The development of the satellite core network and the
satellites, including antennas, GPS receivers, and FPGA ac- decision-making process on which essential network elements
celerators. Global positioning is a crucial service provided are to be deployed on the satellite are crucial to the success-
by LEO satellites. However, the high power consumption and ful implementation of the satellite network. Thus, building
mobility characteristics of satellites present many challenges satellite networks comes with numerous challenges. Neverthe-
for satellite positioning services. Delamotte et al. [77] dis- less, these obstacles indicate promising avenues for research
cuss the application of various differentiated multi-antenna in communication [80], [81], virtualization, lightweight core
access schemes in satellites to support anytime, anywhere network deployment, and other aspects of satellite network
connectivity. The time to first fix is significantly increased construction.
due to the relative movement and Doppler shift between GPS 1) Satellite Communication: An important difference be-
satellites and small satellites. The cold start of GPS satellites tween satellite networks and ground networks is the lack
will cause a lot of extra energy consumption. Narayana et of stable communication links and channels. Satellite links
al. [78] design a low-power GPS receiver suitable for small are vulnerable to various factors such as weather, signal-
satellites and propose an energy optimization algorithm to to-noise ratio, power, and antenna gain. Therefore, reducing
shorten the satellite’s first positioning time. The CloudScout path loss is a crucial consideration in satellite communication.
project [79] demonstrated by the European Space Agency on Beamforming is a common method used to improve the quality
Earth observation satellites, claims that deep learning can bring of service and expand coverage. However, inter-satellite and
many advantages to satellite autonomy, including alleviating satellite-ground communications face challenges such as long
downlink pressure and reducing operating costs. The first distance and interference, and path loss is more significant.
instance of CloudScout deployment can be optimized for cost, Therefore, more effective access solutions are urgently needed
size, and power efficiency. to reduce path loss.
Challenges: The new computing demands have brought Another important difference between satellite and ground
about a revolution in satellite systems, requiring new system networks is the limited resources. Onboard computing, net-
architecture and hardware (such as sensors, antennas, and work, and storage resources are often limited by the satellite’s
ground stations) to support the deployment and operation of design and operation. Effectively utilizing the limited onboard
LEO mega-constellations. Reducing costs is vital for the ad- resources is key to improving the performance of satellite
vent of the New Space era, while also presenting strict design networks, in which spectrum resources play a critical role.
Spectrum is a limited natural resource, and several fixed fre- satellite networks is not keeping up with the trend of network
quency bands have been allocated for satellite communications virtualization.
[82]. However, in the era of 5G, B5G, and even 6G, massive The separation of the SDN control plane and data plane
terminal access, ultra-high data speed, and significant capacity helps to achieve efficient and flexible network management.
requirements are bound to put more pressure on the existing The SDN-based LEO satellite constellation can adopt the
spectrum. Cognitive radio and millimeter waves are effective design concept of a control plane and data plane. Deploy
technologies to improve the efficiency of spectrum utilization network controller in the control plane to realize data pro-
and deserve further research. cessing and transmission. Deploy infrastructure such as satel-
Communication delay is a crucial factor affecting user lites, gateways, and switches on the data plane. The control
service experience. Satellite communication is affected by the plane performs integrated management of the whole satellite
objective environment and channel quality, and the delay is network, and the data plane is used to collect various basic
usually substantial. Therefore, it is of great significance to equipment and terminal data. The separation of the control
reduce the delay of satellite networks further. Satellite commu- plane and the data plane of the LEO satellite constellation
nication delay is usually caused by transmission, propagation, makes computing offloading, routing decisions, and resource
queuing, and processing. The high dynamic and time-varying allocation more independent and flexible.
characteristics of satellite topology increase the instability Since satellites are made of general-purpose hardware, the
of communication delay. Improving the onboard computing invocation and implementation of onboard functions depend
power will enable user services to be processed on the satellite, on the dedicated interfaces of specific suppliers, which limits
significantly reducing the delay. the flexibility of satellite architecture and the programmability
The frequent movement of satellites raises the problem of of onboard resources. NFV shifts the realization of network
service migration. The ground area covered by LEO satellites functions from hardware dependence to a software program-
is not fixed, and the high-speed movement of LEO satellites ming paradigm. One view is that SDN can effectively address
results in frequent switching of satellite networks. Therefore, some of the challenges associated with NFV. The extensive
mobility management is essential to ensure the continuity of application of virtualization technology in cloud computing
satellite communication and meet the QoS requirements of platforms can inspire the construction of NFV-based satellite
users. Computing offloading is also a problem caused by satel- networks.
lite movement and limited onboard resources. Satellites need Network resource allocation is a core business of network
to offload some computing tasks to other eligible satellites virtualization, called virtual network embedding satellite net-
or ground base stations to improve QoS and nearby service works [85]. Differences in the design, structure, and charac-
capabilities. Dynamic unloading schemes are the focus of teristics of satellite networks and ground networks determine
research on satellite communication. In addition, congestion that the management of satellite network resources is not easy.
control and delay are also challenges faced in the computation Onboard computing resources are extremely limited. Network
offloading process. resources such as storage, spectrum, and energy also need to
Access to LEO satellites by a large number of users be properly orchestrated. SDN/NFV-based satellite network
raises security concerns. Some user data and information are resource management is an effective solution to the above
security-sensitive and private, which brings inevitable security challenges and deserves further research.
problems. The design of the security mechanism has to face 3) Lightweight Core Network [29]: The integration and
the challenges of an open satellite environment and highly evolution of 5G and satellite networks have become an
dynamic topology. Inter-satellite communications are prone to inevitable development trend [86]. To improve the access
interference or eavesdropping, which is also the main security capability of satellite networks, the core network functions
challenge facing satellite communications today. Mature data can be deployed on satellites. However, directly deploying the
encryption techniques have been developed. However, in order huge and complex core network functions on the satellite may
to further adapt to satellite communication scenarios and avoid limit the flexibility of the satellite and increase the difficulty
the increase in delay caused by complex keys, distributed key of management. Additionally, the large amount of signaling
calculation and management have attracted the attention of that cannot be handled directly over the satellite networks
researchers. Additionally, the authentication mechanism and increases network latency. As mentioned above, it is necessary
the security routing mechanism should be redesigned [83]. to deploy core network functions on satellites and tailor them
2) Virtualization of Satellite Networks: As mentioned appropriately to provide flexible management control services,
above, onboard resources and energy are typically limited. To as shown in Fig. 1. The reason for tailoring the core network
fully utilize the potential of satellites, onboard resources and is that the ground core network architecture is too large and
network functions should be arranged flexibly and efficiently. complex, and it is unrealistic to directly deploy it on resource-
Network virtualization technology is an effective means to constrained satellites. Deploying the lightweight core network
improve network flexibility and points towards the direction directly on the satellite can endow the satellite network with
for the construction of future network architecture. Network basic core capabilities such as access management, mobility
function virtualization (NFV) and software-defined network- management, and session management, thereby optimizing
ing (SDN) are typical applications of network virtualization, network management and reducing latency [87], [88]. In
which can provide a reference for the architecture design addition, the onboard core network is expected to provide more
of satellite constellations [84]. However, the development of key technologies and solutions for the vertical industry [89]. It
>&)?4:&$12"%&
!"##$%&'()&"%*+ /0&-1+2(-)3+ save communication costs and improve responsiveness by
*,-.&',* (224&'()&"%*
processing data onboard instead of transmitting it to the ground
FG$K
!"#$%&'"#$%()*+'
station. However, due to the limited size of a single satellite,
3&%#)"245&6$78&)$ FG$HIH ;%494&6$:(%#)(2$ FG$J ;%494&6$6"#"$ FG$P ."%"/&0&%#$
the computing ability of the deployed models is also limited.
97%:#4(%+3;<- 97%:#4(%+;3<- 97%:#4(%+;!<- 12"%&
FG$LIH FG$LIJ
Processing high-resolution earth imagery in real-time is a
!"#"$
%&#'()*
FG$O
;8&)$12"%& FG$HIJ 3(%#)(2$12"%& typical application scenario in LEO satellite applications, but
+!,-
=6/&$78&)$ 3&%#)"245&6$78&)$ FG$M @AB.C

it is challenging to process such large resolution images with
97%:#4(%+=;<- 97%:#4(%+3;<- (?&)2"D
a small model. Therefore, deploying a single AI model in a
single satellite is unrealistic, and transmitting all images to the
FG$NIL FG$NIJ FG$NIH
ground station is not feasible. Instead, a distributed learning
@::&88$%&#'()* ;8&)$&E7410&%#
+@,- +;=- framework deployment on LEO is promising [92].
However, traditional distributed learning deployment is not
Fig. 1: Architecture Reference Model of Lightweight Core practical in LEO satellites because inter-satellite communi-
Network. cation is currently not possible, and even if satellites could
communicate in the future, the amount of memory and power
cannot be ignored that the research on core network satellites
required to do so would quickly become the bottleneck.
is still in its infancy. Which network functions need to be
Therefore, Federated Learning (FL) [93]–[95] as a special dis-
deployed on satellites? How to simplify network functions?
tributed learning paradigm has great prospects for widespread
What impact will the deployment of the core network on
deployment in satellites. In this way, we can directly process
satellites bring to the satellite network’s architecture? These
and analyze the collected information on the satellite while
questions need to be answered.
protecting the privacy of each satellite’s data [96]. We can also
4) Others: The combination of satellite constellations and
deploy a large model in the ground station and use this model
other fields provides researchers with numerous research op-
to assist a shallow model deployed on the satellite. However,
tions. Hardware platforms and distributed system design are
FL, like traditional deep learning systems, often demonstrate
all starting points for future satellite network construction.
incorrect or unexpected corner-case behaviors, especially in
Considering the inherent characteristics of satellite networks,
the harsh space environment. Therefore, a systematic testing
the optimization of satellite network performance will be
tool is necessary for automatically detecting erroneous be-
more complex than ground network optimization, including
haviors of FL-driven satellites that can potentially lead to
resource management, cost, energy efficiency, etc. For ex-
data invalidation analysis [97]. This tool can automatically
ample, inter-satellite or inter-satellite-ground communication
generate test cases leveraging real-world changes in the space
links are susceptible to interference and become unstable.
environment (such as meteorites, lighting conditions, shooting
The commonly used method is to combine beamforming
angle, etc.) to retrain the corresponding FL model and improve
and distributed control strategy to suppress communication
the model’s robustness.
interference. Another issue worth discussing is that with the
wide deployment of satellite constellations, it is necessary to
carry out experiments in real space environments based on C. Satellite and Smart City
satellite platforms, not just at the simulation level. Therefore, it With the increasing availability of remote sensing (RS) data
is necessary to strengthen the hardware platform construction from satellites, urban analysis and management have seen
of satellite constellations, such as intelligent reflecting surfaces significant progress. For instance, meteorological satellites
[90]. The satellite network also needs to provide services such can capture climate change, improving weather prediction
as network spatiotemporal change information (data traffic, accuracy, while synthetic aperture radar imaging satellites can
location, ephemeris) and target positioning. identify urban buildings and roads, aiding in urban planning.
However, utilizing RS data effectively and efficiently for
B. AI in Space urban computing faces several challenges. Firstly, the single
With the construction of Starlink and OneWeb, our focus view/modality of RS data cannot fully represent the entire city,
has shifted towards space exploration [2]. Remote sensing necessitating fusion with data from other sources that has not
satellites capture vast amounts of space images from various been extensively studied. Secondly, the low spatio-temporal
angles to explore unknown information in space. However, resolution of RS images limits their applicability in real-time
traditional methods of transmitting these images to the ground scenarios such as urban traffic perception and prediction. We
for analysis face two challenges [42], [91]: limited uplink and discuss research opportunities that address these issues using
downlink bandwidth, especially since the uplink bandwidth AI technologies and data management.
is only in the range of tens to hundreds of kb/s and the 1) Urban Satellite-Ground Data Fusion: Ground data pri-
transmission process has a large latency, and the transmission marily collected from sensors and GPS devices do not cover
process is relatively fragile, and the transmission may be all the spatial and temporal dimensions of a city. Satellite data
interrupted, leading to loss of data. can, therefore, enrich the available information. For instance,
Satellites have become more powerful with the development Wu et al. [98] used the attention mechanism to fuse RS images
of satellite technology, and they can process information and GPS trajectories for automatic road network generation.
onboard. Therefore, deploying AI models on satellites can However, consolidating satellite-ground data faces two main
challenges. Firstly, the quality of urban satellite-ground data process. However, SmartSat does not take into account the
may not meet application requirements due to data errors, use of commercial components for assembling the on-board
deficiencies, and noise. Hence, data cleaning is a primary computer. Additionally, it does not provide any solutions for
problem to be solved. To do this, we can utilize satellite data fault tolerance, improving operating system performance, and
to verify ground data and vice versa, and leverage advanced scheduling resources.
data management technologies such as active learning and HPE has equipped Space-born Computer 7 with a server
crowdsourcing to correct errors. Secondly, urban satellite- operating system that considers the performance improvement
ground data have different modalities such as texts, images, and fault tolerance of servers using commercial devices. How-
and videos, which requires different models for encoding data ever, this system is not specifically designed for the satellite
in different modes. Additionally, there are two ways to fuse environment, and the cost of fault tolerance is quite high,
them based on the order of data integration in the whole which puts it far behind the actual deployment of satellite
process. Pre-fusion involves concatenating encoding results operating systems.
(e.g., vector representations) for downstream applications, In the past, SpaceX has launched Starlink satellites that
while post-fusion independently applies each mode of data carry a stripped-down version of the Linux operating system.
for each task, then fuses task results as the final result of This solution takes into account multiple factors such as
downstream applications. compatibility with the ecosystem, improving the performance
2) Data Augmentation for Urban RS: The low spatio- of the operating system, and maintaining its real-time perfor-
temporal resolution of urban RS images limits their applicabil- mance. However, the system relies on radiation-resistant CPUs
ity in real-time scenarios. Enhancing RS data with data aug- and does not use commercial devices, resulting in a relatively
mentation technology presents a promising solution. Existing limited system flexibility.
methods mainly focus on mapping from low spatial resolution We believe that the future satellite operating system should
to high spatial resolution, such as detecting changes via possess the following characteristics. Firstly, in the past, on-
mapping multi-spatial-resolution RS images [99]. However, board computers were designed for stability and real-time per-
mapping-based methods require sufficient low-high resolution formance. However, with the new requirements for computing
pairs, making them impractical in urban applications. We power and the use of commercial devices to build the on-board
can convert the data augmentation problem into a generation computer, the satellite operating system needs to consider
problem solved using generative models like generative ad- performance factors as well. It should also be compatible with
versarial networks [100]. Prior knowledge, such as the road the existing Linux software ecosystem, allowing for smooth
network topology and the first theorem of geography, should migration of existing software to satellites. Secondly, the
be considered in the generation process. Additionally, we can on-board computers are built with radiation-resistant special
utilize urban data collected from other domains to enhance RS devices, and the operating system is designed to be reliable
data. For example, detecting road lines in street view photos in that environment. However, after using commercial devices
is useful when the related RS image has low resolution. to improve performance, the hardware can no longer provide
security and reliability guarantees. Therefore, providing a
D. Satellite Operation System fault-tolerant mechanism at the software level of the operating
Satellite software is advancing rapidly, and it needs inno- system is a key issue [102]. Additionally, the traditional
vative hardware to support it. The satellite components are satellite operating system has a long application development
becoming more generalized. They no longer rely solely on and deployment cycle. Once an application is deployed, it
radiation-resistant devices like the CPU, DRAM, and FPGA. is difficult to update, and development between different
The use of commercial computer devices to build the satellite satellites cannot be migrated, leading to extremely high devel-
platform has provided a solid foundation for the develop- opment costs. Therefore, providing a set of mechanisms for
ment of satellite operating systems [101]. However, over the development, testing, deployment, and dynamic adjustment on
past few decades, existing satellite operating systems have the satellite operating system is of great significance. Lastly,
mainly focused on stability and real-time performance, with in the scenario where a multitude of tasks are performed on
no changes to basic functions and structures. As a result, these satellites, such as AI image inference, video transmission, and
systems are unable to keep up with the new trend of satellite resource monitoring programs, it is also necessary to consider
software upgrades and hardware generalization. ensuring the isolation between different tasks and providing
Currently, there has been some industry research on the new scheduling capabilities based on the importance of each task.
generation of satellite operating systems. Lockheed Martin has
come up with SmartSat, an operating system that is CubeSat- V. C ASE S TUDY
based 6 . Satellites that are equipped with SmartSat will cre- In order to promote research on satellite computing, we
ate a space cloud computing platform with on-board data established the Tiansuan constellation as an open research
processing capabilities. The satellite capabilities are updated platform [5], illustrated in Fig. 2. Tiansuan supports on-
through software that is uploaded from the ground. SmartSat board computing services, satellite operating systems, 6G
is centered around the application store and provides a rela- core network systems, and federated learning as well as AI
tively comprehensive solution for the application development acceleration. A total of 300 satellites will be launched in
6 https://www.lockheedmartin.com/ 7 https://www.hpe.com/us/en/home.html
;EF3GD
!"#$%&'()*+,-1(),""*),
B,C'D
455"*6()*#0
)$2*#$. #")($#. /"#$%%&#$.
/$,/$ )$1*(,&#&*, 1*22+,&1"#&*,
>$$-?"+,-$)@"A+&'-$ 1,#)"'2"#$%%&#$.3+,4#&*, 1,#$)'2"#$%%&#$.3+,4#&*,

*,6*")-.789.("#$4": /"#$%%&#$'()*+,-.
1*%%"6*)"#&*,
*,6*")-./$)0&1$./+33%:
;"-)6/$ 1*23+#"#&*,"%.;+/&*,
8',&' *+"+,-$ *,6*")-./$)0&1$."($,#
!"#"$%&'$( $-($."+#*,*2:
!"#$%%&#$'()*+,-./$)0&1$.
*+"+,-$ 2"33&,( %&(=#4$&(=#.1*23*,$,#
6/"$(/"$(
1(),""*),&./#$0%-',)2#/3
.,/0/"$
., "-"3#&0$. %&,5. ,$#4*)5. =$#$)*($,$*+/.
H-=:" *+"+,-$ 3,"$2/"$ ;',<,$() ,$#4*)5&,( "-"3#"#&*, /$1+)&#: -$0&1$."11$//
*+
*+"+,-$
1-$+=-:)1'$+'=
3,"$2/"$ 4"2+
1&,$" *+"+,-$ !"#$%&'()*+,-./#$0%-1)()*#0
56"$+",7 /"#$%%&#$'()*+,-.%&,5.",-.##<1 )$/*+)1$.2","($2$,#
3#&4"!1#01%35 4$#!'&$() +(45$)'&$() +((!-$)'&$() !(/&$)#

1&"$(2&") !"#$%&"'
!"#$%&!'&$() %"!*$+",-$%+(*"!. %"!*$+",$)*(+'&$()
*+"+,-$ !())*+
!"#$%%&#$'()*+,-./&,0
,-. /01"20+ '/&0(!$1'&$() '++"%% %"+/!$&. &!'22$+,+()&!(3
1,#$)'!"#$%%&#$ /&,0
3,"$2/"$ *-/+&
5)6&# 7%",$ !"#$%&'()*+,-./#$'%-0)()*#'-1$'!)*#'
1&,$" *+"+,-$
1,#$),$# 59/,:,$7
Fig. 2: Tiansuan Constellation Platform Architecture.
Fig. 3: The picture of Baoyun, BUPT-1, and Innovation Raytheon (from left to right).
TABLE II: Main parameters of Tiansuan Constellation Phase 1.

Number Orbital Altitude Mass Battery Capacity Spectrum Uplink Rate Downlink Rate ISLs Processors
1 500±50km ≤ 30kg 118Wh – 236Wh X-band 0.1Mbps – 1Mbps 100Mbps – 600Mbps NO CPU/NPU
4 > 500km > 50kg > 360Wh X, Ku, Ka bands ≥ 200Mbps ≥ 1Gbps YES CPU/NPU/GPU
three phases. This will include the first phase with 6 satellites, with remote sensing applications are tested. Communication
the second phase with 24 satellites, and the third phase with capabilities with inter-satellite links will be explored with
300 satellites. Table II lists the parameters of satellites in the last three satellites. The on-board payloads provide the
the first phase. As of July 2023, the Tiansuan constellation majority of the computing power as listed in TABLE II. As
has launched five satellites, Baoyun, Innovation Raytheon, shown in Fig. 2, ground stations are gateways for satellites
Lize1, Yuanguang, and BUPT-1 as shown in Fig. 3. Another to transmit data. By integrating with the private data centers,
satellite, Wangqizhou, will also be launched this year. In ground stations can offer the received data through the Internet
Tiansuan, satellites are manufactured according to [103]. Most to users. We have deployed many ground stations, distributed
satellites will be placed in sun-synchronous orbit. For the first in Hunan, Xinjiang, and Heilongjiang Provinces. Next, we
three satellites in the first phase, edge computing capabilities introduce several use cases deployed in the Tiansuan constel-
789'"%,+*0/$%5$-/01$5%6+,+#$6$,* .$/01$%6+,+#$6$,*
3!'-#"/'%/)1#"2 4567'
;*)/..+)/'
='07$#"$1$'
!'(&-(. *+%,'#&)(-(. /)001(-2#"-)(& !"#$%&'(&)$ :;!%$'2'-<'$
&'(&)$
!"#$%&'$
-(,#''.,#%/$+0)1 '.)2
!"#$" !/$,*%4'5
8/""/()"+*.'3!'9"+:*)/'%/)1#"2 )$*+%)+,+#$- .$/01$%230,

!"#$%&'()# *+),$+'%&'()# 3#"#%4#&'
,#-*.'/&0/'
#-*. /&0 !"#$%&'(
%/)1#"2 !"#$%&'()*)+#% 56.'%*778-2#"-)(
9: 0"#$ %(&)"%&'$
,#$8.)('"
!"#$%&'(
3.4#15(66#""%
1 9:%;("#%
#70.&8#), ",(,.+) .$/01$%79,*-988$-
!"#
$%&
!"#$%79,*-988$-
Fig. 4: Satellite-Borne B5G Core Network. *+,+-%$.,$'/ 9&'$&
Fig. 5: Satellite-Ground Collaboration of the Cloud-Native

lation. Satellite.
A. Satellite-borne B5G Core Network As the first cloud-native satellite, Baoyun leverages the
On August 9, 2021, we conducted the first deployment test space computing resources and central cloud on the ground
of a 5G core network system on the experimental satellite station and consequently they have AI and multi-tasking
(TY20) of Tiansuan. The experimental network consists of capabilities. For on-orbit AI inference application, low-quality
the onboard B5G core network and the terrestrial private image data may be discarded to lower transmission overhead
5G network. The on-board B5G core network implements due to over 50% coverage by clouds. After useful image
three network functions (AMF, SMF, and UPF). The network data is transmitted back, a high-precision model deployed
functions support essential system procedures such as user at the ground station with abundant computing resources is
registration and session establishment. By incorporating the used to the following computing, which further shortens the
terrestrial full-fledged 5G network, we test the signaling inter- processing time from usual about one day to even one hour.
action between the control plane and the user plane. Downlink Besides, these models in space can be updated when needed.
telemetry showed that the three network functions were oper- Cloud-native satellites will be an integral part in our daily
ating normally. It also showed that the control signaling was lives. For example, we compare the images before and after
generated correctly. The control signaling was then transmitted rainstorms to identify the risk of mountain collapse through
to the terrestrial private 5G network. It triggered the local on-orbit AI inference, which guides us to discover hidden
computation offloading controlled by satellites. We conducted danger and prevent catastrophe in advance.
further tests such as video calls. As shown in Fig. 4, we BUPT-1, which is the first primary satellite of Tiansuan,
have successfully deployed lightweight B5G core networks on was successfully launched and operated on January 15th,
both the Baoyun and Innovation Raytheon satellites. This core 2023. Over a period of 15 days, we conducted several ver-
network is the updated version of the former 5G core network. ifications to confirm and quantify the performance gains of
It improves signaling interactions and can be used to set up our cloud-native system on BUPT-1. During this time, we
video calls based on the Session Initiation Protocol. After the tested remote sensing image AI inference, real-time video
two satellites are launched, functional and performance tests streaming transmission between the satellite and ground, and
will be conducted. More details about this work can refer to other key functionalities. Based on telemetry from BUPT-1,
[29]. we found that the average response time for the end-to-end
measurement and control service between satellite and ground
B. Cloud-Native Satellite was 8 seconds. Additionally, telemetry data collected from the
Baoyun was successfully launched on 7 December 2021, on-board program running status field revealed that a signif-
which carried the computing payload of Tiansuan Constella- icant improvement in the utilization rate of onboard services
tion. Fig. 5 shows the first cloud-native satellite that integrates compared to non-cloud-native platforms. We package the ser-
satellite and ground computing capabilities to compete the vice programs and relevant dependencies into containers, and
space tasks. The satellites run smoothly and provide service deploy them uniformly on satellites. Containerized deployment
in orbit. In addition, Huawei Cloud is the first time to make ensures the elasticity and scalability of the services, allowing
“Cloud-Edge Synergy” come true in space by cloud-native for dynamic resource adjustments based on actual needs and
satellite. In the future, the “Cloud-Edge Synergy” scheme automatic addition or removal of instances when necessary. On
will be deployed in 6 satellites in the first phase of the the ground, the required services can be flexibly launched on
Tiansuan constellation to form a unified computing network demand through remote control commands. In contrast, non-
collaboratively. We expect cloud-native satellites to create new cloud-native platforms can only provide customized services,
capabilities for emergency communication, ecological moni- occupying limited on-board resources for extended periods
toring, disaster prevention and mitigation, urban construction, during program runtime, resulting in a significant decrease in
etc. service utilization. More details about this work can refer to
!"#$%&'())*'(
remote control command was initiated by the ground station
-%.'!#&'
in Changsha, Hunan Province, and the onboard client started to
2+3%&'()"*"+,
initiate transmission. The ground station in Tongchuan, Shanxi
('%)*
/!
-(.()."+,/0+-(1
Province received the data at high transmission speed and
('%)*0/12
analyzed the QUIC traffic. For the QUIC protocol based real-
&%'(
time transmission of the satellite-ground link, we developed
!"#$%
a customized QUIC server logic to achieve compatibility and
+,-./0$1-/',-))(,
a data frame analysis system following the AOS (Advanced
!"#$%!#&'
!"#$%&'()"*"+,
!#)3#&"0%*3* On-orbit System) protocol issued by the CCSDS (Consultative
('%)*
+,
-(.()."+,/
0+-(1
455"16*31#)783*3$8
%*3*
Committee for Space Data Systems) to achieve QUIC protocol
!"#$%&'$
parsing enhancement.
Fig. 6: Satellite-Ground Collaborative Image Inference based
on KubeEdge. VI. C ONCLUSION
LEO satellite constellations are experiencing rapid devel-
[104].
opment and satellite computing is promising to address the
limitation of traditional satellite bent-pipe architecture and
C. Satellite-Ground Collaborative Image Inference
provide computing services for ground users. In this paper, we
We have implemented a satellite-ground collaborative in- first analyze the current development of LEO constellations.
ference application based on KubeEdge and its AI extension, Then, we survey recent work in the field of satellite computing
Sedna, on both Baoyun and Innovation Raytheon as shown and discuss the research challenges. We also put forward
in Fig. 6. To communicate with the satellite, we modified research opportunities that are worth working on, naming
the KubeEdge runtime (CloudCore) to build a central ground networking, computing, smart city, satellite operating system.
controller upon a Linux server, which communicates with the Finally, we introduce our open research platform on real satel-
satellite intermittently according to the position of the satellite. lites named Tiansuan constellation and several experiments
We later deployed two image detection models with different deployed on it. Our next step is to establish a larger open-
precisions, a low-precision model and a high-precision model. source satellite constellation, allowing more people to access
The low-precision image detection model is deployed inside and study satellite technology. Our ultimate aim is to serve the
the EdgeCore of the satellite and used to detect whether the entire human population by democratizing access to satellite
captured image is of interest. Once the detection on the low- computing.
precision model achieves high confidence, the satellite will use
the concrete result to facilitate later processing (e.g., calculate
total the area of agricultural lands). When the confidence is R EFERENCES
low, the satellite will download the captured image to the [1] D. Bhattacherjee, S. Kassing, M. Licciardello, and A. Singla, “In-orbit
ground controller. The ground controller will detect the objects computing: an outlandish thought experiment?” in Proceedings of the
ACM Workshop on Hot Topics in Networks, 2020, pp. 197–204.
using the high-precision model and then upload the exact [2] B. Denby and B. Lucia, “Orbital edge computing: nanosatellite con-
result to the satellite. By fully leveraging the computation stellations as a new class of computer system,” in Proceedings of
capacity on the satellite, we can reduce the image detection International Conference on Architectural Support for Programming
Languages and Operating Systems, 2020, pp. 939–954.
latency from days to minutes in such a collaborative computing
[3] V. Singh, A. Prabhakara, D. Zhang, O. Yağan, and S. Kumar, “A
paradigm without accuracy cost. In addition, we propose an community-driven approach to democratize access to satellite ground
on-board image segmentation solution to mitigate the occur- stations,” GetMobile: Mobile Computing and Communications, vol. 26,
rence of redundant images due to cloud occlusion. Performing no. 1, pp. 35–38, 2022.
[4] D. Bhattacherjee and A. Singla, “Network topology design at 27,000
image segmentation before on-orbit inference helps reduce the km/hour,” in Proceedings of the International Conference on Emerging
computational load and save bandwidth. Networking Experiments and Technologies, 2019, pp. 341–354.
[5] S. Wang, Q. Li, M. Xu, X. Ma, A. Zhou, and Q. Sun, “Tiansuan
constellation: An open research platform,” in Proceedings of IEEE
D. Quic Protocol based Real-Time Transmission of the International Conference on Edge Computing, 2021, pp. 94–101.
Satellite-Ground Link [6] D. Bhattacherjee, W. Aqeel, I. N. Bozkurt, A. Aguirre, B. Chan-
drasekaran, P. B. Godfrey, G. Laughlin, B. Maggs, and A. Singla,
On August 25, 2022, Tiansuan successfully completed the “Gearing up for the 21st century space race,” in Proceedings of ACM
world’s first QUIC (Quick UDP Internet connections) protocol Workshop on Hot Topics in Networks, 2018, pp. 113–119.
based real-time transmission test of the satellite-ground link. [7] M. Handley, “Delay is not an option: low latency routing in space,” in
Proceedings of ACM Workshop on Hot Topics in Networks, 2018, pp.
This test is mainly based on the satellite-ground integrated 85–91.
distributed network verification platform jointly developed by [8] D. Bhattacherjee, W. Aqeel, G. Laughlin, B. M. Maggs, and A. Singla,
Huawei Cloud and our scientific research team. The test “A bird’s eye view of the world’s fastest networks,” in Proceedings of
the ACM Internet Measurement Conference, 2020, pp. 521–527.
was started on May 18, 2022. At 11:23 on June 18, 2022, [9] S. Kassing, D. Bhattacherjee, A. B. Águas, J. E. Saethre, and A. Singla,
multiple real-time transmission tests of the QUIC protocol “Exploring the” internet from space” with hypatia,” in Proceedings of
of the satellite-ground link were completed. On August 24, ACM Internet Measurement Conference, 2020, pp. 214–229.
[10] G. Giuliari, T. Klenze, M. Legner, D. Basin, A. Perrig, and A. Singla,
2022, the data analysis was completed. This test realized end- “Internet backbones in space,” ACM SIGCOMM Computer Communi-
to-end data transmission based on the QUIC protocol. The cation Review, vol. 50, no. 1, pp. 25–37, 2020.
[11] Q. Li, S. Wang, X. Ma, Q. Sun, H. Wang, S. Cao, and F. Yang, Proceedings of IEEE Communications Magazine, vol. 60, no. 4, pp.
“Service coverage for satellite edge computing,” IEEE Internet of 18–24, 2022.
Things Journal, vol. 9, no. 1, pp. 695–705, 2021. [32] J. Li, H. Li, J. Liu, Z. Lai, Q. Wu, and X. Wang, “A timeslot division
[12] Y. Hauri, D. Bhattacherjee, M. Grossmann, and A. Singla, “”internet strategy for availability in integrated satellite and terrestrial network,”
from space” without inter-satellite links,” in Proceedings of ACM in Proceedings of IEEE Wireless Communications and Networking
Workshop on Hot Topics in Networks, 2020, pp. 205–211. Conference, 2021, pp. 1–7.
[13] T. Kur and T. Liwosz, “Simulation of the use of variance component [33] S. Ji, D. Zhou, M. Sheng, and J. Li, “Mega satellite constellation
estimation in relative weighting of inter-satellite links and gnss mea- system optimization: from a network control structure perspective,”
surements,” Remote Sensing, vol. 14, no. 24, p. 6387, 2022. IEEE Transactions on Wireless Communications, vol. 21, no. 2, pp.
[14] C. H. Arslan, K. Aydin, J. V. Urbina, and L. Dyrud, “Satellite-link at- 913–927, 2022.
tenuation measurement technique for estimating rainfall accumulation,” [34] Y. Chen, H. Li, J. Liu, Q. Wu, and Z. Lai, “Gams: an ip address
IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 2, management mechanism in satellite mega-constellation networks,” in
pp. 681–693, 2018. Proceedings of International Wireless Communications and Mobile
[15] T. Lan, H. Li, Q. Wu, Z. Lai, and J. Liu, “Exploiting path diversity to Computing, 2021, pp. 229–234.
increase system performance in mega-constellations,” in Proceedings [35] Z. Lai, H. Li, Q. Zhang, Q. Wu, and J. Wu, “Cooperatively constructing
of IEEE Wireless Communications and Networking Conference, 2021, cost-effective content distribution networks upon emerging low earth
pp. 1–7. orbit satellites and clouds,” in Proceedings of International Conference
[16] Z. Lai, Q. Wu, H. Li, M. Lv, and J. Wu, “Orbitcast: exploiting mega- on Network Protocols, 2021, pp. 1–12.
constellations for low-latency earth observation,” in Proceedings of [36] Z. Lai, H. Li, and J. Li, “Starperf: characterizing network performance
IEEE International Conference on Network Protocols, 2021, pp. 1–12. for emerging mega-constellations,” in Proceedings of International
[17] S. Zhang and K. L. Yeung, “Scalable routing in low-earth orbit satellite Conference on Network Protocols, 2020, pp. 1–11.
constellations: architecture and algorithms,” Computer Communica- [37] F. Tang, “Dynamically adaptive cooperation transmission among
tions, vol. 188, pp. 26–38, 2022. satellite-ground integrated networks,” in Proceedings of IEEE Inter-
[18] M. Handley, “Using ground relays for low-latency wide-area routing national Conference on Computer Communications, 2020, pp. 1559–
in megaconstellations,” in Proceedings of the 18th ACM Workshop on 1568.
Hot Topics in Networks, 2019, pp. 125–132. [38] L. Chen, F. Tang, and X. Li, “Mobility- and load-adaptive controller
[19] S. Wang, Y. Zhao, and H. Xie, “Improving survivability of leo satellite placement and assignment in leo satellite networks,” in Proceedings of
network with guaranteed based approaches,” in Proceedings of IEEE IEEE International Conference on Computer Communications, 2021.
Symposium on Computers and Communications, 2020, pp. 1–6. [39] T. Klenze, G. Giuliari, C. Pappas, A. Perrig, and D. Basin, “Networking
[20] Z. Zhao, Q. Wu, H. Li, Z. Lai, and J. Liu, “Lrar: a lightweight in heaven as on earth,” in Proceedings of ACM Workshop on Hot Topics
risk-avoidance routing algorithm for leo satellite networks,” in 2021 in Networks, 2018, pp. 22–28.
International Wireless Communications and Mobile Computing, 2021, [40] X. Li, F. Tang, J. Liu, L. T. Yang, L. Fu, and L. Chen, “Auto: adaptive
pp. 223–228. congestion control based on multi-objective reinforcement learning for
[21] P. S. Page, K. S. Bhargao, H. V. Baviskar, and G. S. Kasbekar, the satellite-ground integrated network,” in Proceedings of the USENIX
“Distributed probabilistic congestion control in leo satellite networks,” Annual Technical Conference, 2021, pp. 611–624.
in Proceedings of the International Conference on COMmunication [41] V. Singh, O. Yagan, and S. Kumar, “Selfiestick: towards earth imaging
Systems and NETworks, 2023, pp. 335–339. from a low-cost ground module using leo satellites,” in Proceedings of
[22] V.-P. Bui, T. V. Chien, E. Lagunas, J. Grotz, S. Chatzinotas, and B. Ot- the ACM/IEEE International Conference on Information Processing in
tersten, “Robust congestion control for demand-based optimization in Sensor Networks, 2022, pp. 220–232.
precoded multi-beam high throughput satellite communications,” IEEE [42] D. Vasisht, J. Shenoy, and R. Chandra, “L2d2: low latency distributed
Transactions on Communications, vol. 70, no. 10, pp. 6918–6937, downlink for leo satellites,” in Proceedings of the Conference on
2022. Applications, Technologies, Architectures, and Protocols for Computer
[23] M. Liu, Y. Liu, Z. Ma, Z. Porter, J. Chung, S. Claypool, F. Li, J. Tutlis, Communication, 2021, pp. 151–164.
and M. Claypool, “The effects of a performance enhancing proxy on [43] S. Liu, T. Wu, Y. Hu, Y. Xiao, D. Wang, and L. Liu, “Throughput
tcp congestion control over a satellite network,” in Proceedings of evaluation and ground station planning for leo satellite constellation
the IEEE International Performance, Computing, and Communications networks,” in Proceedings of International Conference on Space Infor-
Conference, 2022, pp. 325–331. mation Network, 2020, pp. 3–15.
[24] S. Claypool, J. Chung, and M. Claypool, “Comparison of tcp conges- [44] K. Guo and K. An, “On the performance of ris-assisted integrated
tion control performance over a satellite network,” in Proceedings of the satellite-uav-terrestrial networks with hardware impairments and inter-
International Conference on Passive and Active Network Measurement, ference,” IEEE Wireless Communications Letters, vol. 11, no. 1, pp.
2021, pp. 499–512. 131–135, 2021.
[25] S. Dai, L. Rui, S. Chen, and X. Qiu, “A distributed congestion [45] D. Xia, X. Zheng, P. Duan, C. Wang, L. Liu, and H. Ma, “Ground-
control routing protocol based on traffic classification in leo satellite station based software-defined leo satellite networks,” in Proceedings
networks,” in Proceedings of the IFIP/IEEE International Symposium of IEEE International Conference on Parallel and Distributed Systems,
on Integrated Network Management, 2021, pp. 523–529. 2019, pp. 687–694.
[26] Z. A. Qazi, M. Walls, A. Panda, V. Sekar, S. Ratnasamy, and [46] L. Han, A. Retana, C. Westphal, R. Li, T. Jiang, and M. Chen, “New
S. Shenker, “A high performance packet core for next generation ip based semantic addressing and routing for leo satellite networks,”
cellular networks,” in Proceedings of the Conference of the ACM in Proceedings of the IEEE International Conference on Network
Special Interest Group on Data Communication, 2017, pp. 348–361. Protocols, 2022, pp. 1–6.
[27] M. Moradi, K. Sundaresan, E. Chai, S. Rangarajan, and Z. M. Mao, [47] H. Qi, Y. Guo, D. Hou, Z. Xing, W. Ren, L. Cong, and X. Di,
“Skycore: moving core to the edge for untethered and reliable uav- “Sdn-based dynamic multi-path routing strategy for satellite networks,”
based lte networks,” in Proceedings of the Annual International Con- Future Generation Computer Systems, vol. 133, pp. 254–265, 2022.
ference on Mobile Computing and Networking, 2018, pp. 35–49. [48] P. Zhao, B. Peters, J. Chung, and M. Claypool, “Competing tcp conges-
[28] Y. Li, H. Li, W. Liu, L. Liu, Y. Chen, J. Wu, Q. Wu, J. Liu, and tion control algorithms over a satellite network,” in Proceedings of the
Z. Lai, “A case for stateless mobile core network functions in space,” IEEE Annual Consumer Communications and Networking Conference,
in Proceedings of the Conference of the ACM Special Interest Group 2022, pp. 132–138.
on Data Communication, 2022, pp. 298–313. [49] X. Cao, B. Yang, Y. Shen, C. Yuen, Y. Zhang, Z. Han, H. V. Poor,
[29] R. Xing, X. Ma, A. Zhou, S. Dustdar, and S. Wang, “From earth and L. Hanzo, “Edge-assisted multi-layer offloading optimization of
to space: A first deployment of 5g core network on satellite,” China leo satellite-terrestrial integrated networks,” EEE Journal on Selected
Communications, vol. 20, no. 4, pp. 315–325, 2023. Areas in Communications, vol. 41, no. 2, pp. 381–398, 2023.
[30] T. Darwish, G. K. Kurt, H. Yanikomeroglu, M. Bellemare, and G. La- [50] P. Cassará, A. Gotta, M. Marchese, and F. Patrone, “Orbital edge
montagne, “Leo satellites in 5g and beyond networks: a review from a offloading on mega-leo satellite constellations for equal access to
standardization perspective,” IEEE Access, vol. 10, pp. 35 040–35 060, computing,” IEEE Communications Magazine, vol. 60, no. 4, pp. 32–
2021. 36, 2022.
[31] B. Al Homssi, A. Al-Hourani, K. Wang, P. Conder, S. Kandeepan, [51] T. Pfandzelter, J. Hasenburg, and D. Bermbach, “Towards a computing
J. Choi, B. Allen, and B. Moores, “Next generation mega satellite net- platform for the leo edge,” in Proceedings of International Workshop
works for access equality: opportunities, challenges, and performance,” on Edge Systems, Analytics and Networking, 2021, pp. 43–48.
[52] T. Pfandzelter and D. Bermbach, “Celestial: virtual software system [74] A. Maskey and M. Cho, “Cubesatnet: ultralight convolutional neural
testbeds for the leo edge,” in Proceedings of the ACM/IFIP Interna- network designed for on-orbit binary image classification on a 1u
tional Middleware Conference, 2022, pp. 69–81. cubesat,” Engineering Applications of Artificial Intelligence, vol. 96,
[53] Pfandzelter, Tobias and Bermbach, David, “Qos-aware resource place- p. 103952, 2020.
ment for leo satellite edge computing,” in Proceedings of International [75] R. Wang, W. Zhu, G. Liu, R. Ma, D. Zhang, S. Mumtaz, and
Conference on Fog and Edge Computing, 2022, pp. 66–72. S. Cherkaoui, “Collaborative computation offloading and resource
[54] R. Xie, Q. Tang, Q. Wang, X. Liu, F. R. Yu, and T. Huang, allocation in satellite edge computing,” in Proceedings of the IEEE
“Satellite-terrestrial integrated edge computing networks: architecture, Global Communications Conference, 2022, pp. 5625–5630.
challenges, and open issues,” IEEE Network, vol. 34, no. 3, pp. 224– [76] P. Li, Y. Wang, and Z. Wang, “A game-based joint task offloading
231, 2020. and computation resource allocation strategy for hybrid edgy-cloud
[55] M. Tong, X. Wang, S. Li, and L. Peng, “Joint offloading decision and cloudy-edge enabled leo satellite networks,” in Proceedings of
and resource allocation in mobile edge computing-enabled satellite- the IEEE/CIC International Conference on Communications in China,
terrestrial network,” Symmetry, vol. 14, no. 3, p. 564, 2022. 2022, pp. 868–873.
[56] X. Gao, R. Liu, and A. Kaushik, “Virtual network function placement [77] T. Delamotte, M. G. Schraml, R. T. Schwarz, K.-U. Storek, and
in satellite edge computing with a potential game approach,” IEEE A. Knopp, “Multi-antenna-enabled 6g satellite systems: roadmap,
Transactions on Network and Service Management, vol. 19, no. 2, pp. challenges and opportunities,” International ITG Workshop on Smart
1243–1259, 2022. Antennas, pp. 1–6, 2021.
[57] Q. Tang, Z. Fei, B. Li, and Z. Han, “Computation offloading in [78] S. Narayana, R. V. Prasad, V. Rao, L. Mottola, and T. V. Prabhakar,
leo satellite networks with hybrid cloud and edge computing,” IEEE “Hummingbird: Energy efficient gps receiver for small satellites,” pp.
Internet of Things Journal, vol. 8, no. 11, pp. 9164–9176, 2021. 1–13, 2020.
[58] T. Kim, J. Kwak, and J. P. Choi, “Satellite edge computing architecture [79] E. Rapuano, G. Meoni, T. Pacini, G. Dinelli, G. Furano, G. Giuffrida,
and network slice scheduling for iot support,” IEEE Internet of Things and L. Fanucci, “An fpga-based hardware accelerator for cnns inference
Journal, vol. 9, no. 16, pp. 14 938–14 951, 2021. on board satellites: benchmarking with myriad 2-based solution for the
[59] X. Fang, W. Feng, T. Wei, Y. Chen, N. Ge, and C.-X. Wang, “5g cloudscout case study,” Remote Sensing, vol. 13, no. 8, p. 1518, 2021.
embraces satellites for 6g ubiquitous iot: Basic models for integrated [80] A. Gaber, M. A. ElBahaay, A. M. Mohamed, M. M. Zaki, A. S. Abdo,
satellite terrestrial networks,” IEEE Internet of Things Journal, vol. 8, and N. AbdelBaki, “5g and satellite network convergence: Survey for
no. 18, pp. 14 399–14 417, 2021. opportunities, challenges and enabler technologies,” in Proceedings of
[60] J. Chu and X. Chen, “Robust design for integrated satellite–terrestrial Novel Intelligent and Leading Emerging Sciences Conference, 2022,
internet of things,” IEEE Internet of Things Journal, vol. 8, no. 11, pp. pp. 366–373.
9072–9083, 2021. [81] X. Zhu and C. Jiang, “Integrated satellite-terrestrial networks toward
[61] X. Zhu and C. Jiang, “Integrated satellite-terrestrial networks toward 6g: Architectures, applications, and challenges,” IEEE Internet of
6g: Architectures, applications, and challenges,” IEEE Internet of Things Journal, vol. 9, no. 1, pp. 437–461, 2021.
Things Journal, vol. 9, no. 1, pp. 437–461, 2022. [82] ESA, https://www.esa.int/Applications/Telecommunications Integrated
[62] L. Zhen, A. K. Bashir, K. Yu, Y. D. Al-Otaibi, C. H. Foh, and P. Xiao, Applications/Satellite frequency bands, 2022.
“Energy-efficient random access for leo satellite-assisted 6g internet [83] K. Gai, Y. Wu, L. Zhu, K.-K. R. Choo, and B. Xiao, “Blockchain-
of remote things,” IEEE Internet of Things Journal, vol. 8, no. 7, pp. enabled trustworthy group communications in uav networks,” IEEE
5114–5128, 2021. Transactions on Intelligent Transportation Systems, vol. 22, no. 7, pp.
[63] J. Wei and S. Cao, “Application of edge intelligent computing in 4118–4130, 2021.
satellite internet of things,” in Proceedings of IEEE International [84] G. Gardikis, H. Koumaras, C. Sakkas, and V. Koumaras, “To-
Conference on Smart Internet of Things, 2019, pp. 85–91. wards sdn/nfv-enabled satellite networks,” Telecommunication Systems,
[64] Z. Song, Y. Hao, Y. Liu, and X. Sun, “Energy-efficient multiaccess edge vol. 66, pp. 615–628, 2017.
computing for terrestrial-satellite internet of things,” IEEE Internet of [85] P. Zhang, H. Yao, and Y. Liu, “Virtual network embedding based on
Things Journal, vol. 8, no. 18, pp. 14 202–14 218, 2021. computing, network, and storage resource constraints,” IEEE Internet
[65] P. Zhou, Y. Liu, L. Guo, N. Lu, J. Wu, and H. Jiang, “Handoff of Things Journal, vol. 5, no. 5, pp. 3298–3304, 2017.
of satellite network for high-speed mobile terminals based on edge [86] L. Ni, B. Hu, C. Wang, S. Gu, M. Meng, and J. Zhang, “Research
computing,” in Proceedings of IEEE International Conference on on the evolution of 5g and satellite network integration,” Mobile
Cyber Security and Cloud Computing, 2021, pp. 167–175. Communication, vol. 46, no. 01, pp. 51–57, 2022.
[66] B. Wang, X. Li, D. Huang, and J. Xie, “A profit maximization [87] Z. Han, C. Xu, K. Liu, L. Lu, G. Zhao, and S. Yu, “A novel mobile
strategy of mec resource provider in the satellite-terrestrial double edge core network architecture for satellite-terrestrial integrated network,” in
computing system,” in Proceedings of IEEE International Conference Proceddings of the IEEE Global Communications Conference, 2021,
on Communication Technology, 2021, pp. 906–912. pp. 01–06.
[67] M. M. Gost, I. Leyva-Mayorga, A. Perez-Neira, M. A. Vazquez, [88] J. Kim, J. Lee, H. Ko, T. Kim, and S. Pack, “Space mobile networks:
B. Soret, Moretti, and Marco, “Edge computing and communication for Satellite as core and access networks for b5g,” IEEE Communications
energy-efficient earth surveillance with leo satellites,” in Proceedings Magazine, vol. 60, no. 4, pp. 58–64, 2022.
of the IEEE International Conference on Communications Workshops, [89] X. Zhu and X. Qu, “Research on 5g lightweight core network technol-
2022, pp. 556–561. ogy for vertical industries,” in Proceddings of the International Wireless
[68] Q. Li, S. Wang, X. Ma, A. Zhou, and F. Yang, “Towards sustainable Communications and Mobile Computing, 2021, pp. 1499–1505.
satellite edge computing,” in Proceedings of IEEE International Con- [90] S. Alfattani, W. Jaafar, Y. Hmamouche, H. Yanikomeroglu,
ference on Edge Computing, 2021, pp. 1–8. A. Yongaçoglu, N. D. DJào, and P. Zhu, “Aerial platforms with recon-
[69] V. Kothari, E. Liberis, and N. D. Lane, “The final frontier: deep learning figurable smart surfaces for 5g and beyond,” IEEE Communications
in space,” in Proceedings of the International Workshop on Mobile Magazine, vol. 59, no. 1, pp. 96–102, 2021.
Computing Systems and Applications, 2020, pp. 45–49. [91] Y. Li, H. Li, L. Liu, W. Liu, J. Liu, J. Wu, Q. Wu, J. Liu, and
[70] J. Ding, N. Xue, G.-S. Xia, X. Bai, W. Yang, M. Y. Yang, S. Belongie, Z. Lai, “”internet in space” for terrestrial users via cyber-physical
J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Object detection in aerial convergence,” in Proceedings of the ACM Workshop on Hot Topics
images: a large-scale benchmark and challenges,” IEEE Transactions in Networks, 2021, pp. 163–170.
on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. [92] J. So, K. Hsieh, B. Arzani, S. Noghabi, S. Avestimehr, and R. Chandra,
7778–7796, 2022. “Fedspace: An efficient federated learning framework at satellites and
[71] T. Hoeser, F. Bachofer, and C. Kuenzer, “Object detection and image ground stations,” arXiv preprint arXiv:2202.01267, 2022.
segmentation with deep learning on earth observation data: a review - [93] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas,
part ii: applications,” Remote Sensing, vol. 12, no. 10, p. 1667, 2020. “Communication-efficient learning of deep networks from decentral-
[72] M. Lofqvist and J. Cano, “Accelerating deep learning applications in ized data,” in Proceedings of Artificial intelligence and Statistics, 2017,
space,” arXiv preprint arXiv:2007.11089, 2020. pp. 1273–1282.
[73] B. Uzkent, C. Yeh, and S. Ermon, “Efficient object detection in large [94] T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V. Smith,
images using deep reinforcement learning,” in Proceedings of the IEEE “Federated optimization in heterogeneous networks,” Proceedings of
Winter Conference on Applications of Computer Vision, 2020, pp. the Conference on Machine Learning and Systems, vol. 2, pp. 429–
1824–1833. 450, 2018.
[95] K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan,

S. Patel, D. Ramage, A. Segal, and K. Seth, “Practical secure aggrega-
tion for privacy-preserving machine learning,” in Proceedings of ACM
SIGSAC Conference on Computer and Communications Security, 2017.
[96] M. Yurochkin, M. Agarwal, S. Ghosh, K. Greenewald, N. Hoang, and
Y. Khazaeni, “Bayesian nonparametric federated learning of neural
networks,” in Proceedings of International Conference on Machine
Learning, 2019, pp. 7252–7261.
[97] Q. Yang, “Advances and open problems in federated learning,” Foun-
dations and Trends in Machine Learning, 2021.
[98] H. Wu, H. Zhang, X. Zhang, W. Sun, B. Zheng, and Y. Jiang, “Deep-
dualmapper: A gated fusion network for automatic map extraction using
aerial images and trajectories,” in Proceedings of the AAAI Conference
on Artificial Intelligence, vol. 34, no. 01, 2020, pp. 1037–1045.
[99] P. Zhang, M. Gong, L. Su, J. Liu, and Z. Li, “Change detection
based on deep feature representation and mapping transformation
for multi-spatial-resolution remote sensing images,” ISPRS Journal of
Photogrammetry and Remote Sensing, vol. 116, pp. 24–41, 2016.
[100] H. Zhang, I. Goodfellow, D. Metaxas, and A. Odena, “Self-attention
generative adversarial networks,” in Proceedings of the International
conference on machine learning, 2019, pp. 7354–7363.
[101] V. Bhosale, K. Bhardwaj, and A. Gavrilovska, “Toward loosely coupled
orchestration for the leo satellite edge,” in Proceedings of Workshop
on Hot Topics in Edge Computing, 2020.
[102] Y. Kimoto and H. Matsumoto, “Evaluation of radiation effects on
commercial-off-the-shelf (cots) parts for use on low-orbit satellite,”
COSPAR Scientific Assembly, vol. 43, p. 542, 2021.
[103] Cubesat design specification, California Polytechnic State University,
2014, rev. 13. Technical report.
[104] C. Wang, Y. Zhang, Q. Li, A. Zhou, and S. Wang, “Satel-
lite computing: A case study of cloud-native satellites,” url-
https://arxiv.org/pdf/2307.08530.pdf, 2023.
Shangguang Wang is a Professor at the School of

Computer Science and Engineering, Beijing Univer-
sity of Posts and Telecommunications, China. He
received his Ph.D. degree at Beijing University of
Posts and Telecommunications in 2011. He has pub-
lished more than 150 papers. His research interests
include service computing, mobile edge computing,
and satellite computing. He is currently serving as
Chair of IEEE Technical Committee on Services
Computing (2022-2023), and Vice-Chair of IEEE
Technical Committee on Cloud Computing (2020-).
He also served as General Chairs or Program Chairs of 10+ IEEE conferences.
He is a Fellow of the IET, and Senior Member of the IEEE. For further
information on Dr. Wang, please visit: http://www.sguangwang.com.
Qing Li received her Ph.D. degree in Beijing Uni-

versity of Posts and Telecommunications, Beijing,
China, in 2022. She is currently a research assistant
in the School of Computer Science at Peking Univer-
sity, Beijing, China. Her research interests include
mobile edge computing and satellite computing.

Satellite Computing Vision and Challenges

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Satellite Computing Vision and Challenges

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in IEEE Internet of Things Journal.

IEEE INTERNET OF THINGS JOURNAL, VOL. XX, NO. XX, 2023 1

Satellite Computing: Vision and Challenges

TABLE I: Global Top 20 Constellations Information (Dated by July 2023).

=6/&$78&)$ 3&%#)"245&6$78&)$ FG$M @AB.C

>$$-?"+,-$)@"A+&'-$ 1,#)"'2"#$%%&#$.3+,4#&, 1,#$)'2"#$%%&#$.3+,4#&,

3#&4"!1#01%35 4$#!'&$() +(45$)'&$() +((!-$)'&$() !(/&$)#

Fig. 2: Tiansuan Constellation Platform Architecture.

TABLE II: Main parameters of Tiansuan Constellation Phase 1.

8/""/()"+.'3!'9"+:)/'%/)1#"2 )$*+%)+,+#$- .$/01$%230,

Fig. 5: Satellite-Ground Collaboration of the Cloud-Native

[95] K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan,

Shangguang Wang is a Professor at the School of

Qing Li received her Ph.D. degree in Beijing Uni-

You might also like