Intelligent Spectrum Management in Satellite Communications

TL;DR

This paper surveys the use of AI/ML-driven dynamic spectrum management techniques in satellite communication networks, addressing challenges, regulations, and future research to improve spectrum efficiency and network adaptability.

Contribution

It provides a comprehensive review of intelligent DSM methodologies, regulatory issues, and AI/ML applications in CogSat networks, highlighting future research directions.

Findings

AI/ML methods enhance spectrum utilization in SatCom

Regulatory and standardization challenges are identified

Performance metrics for adaptive resource management are discussed

Abstract

Satellite Communication (SatCom) networks represent a fundamental pillar in modern global connectivity, facilitating reliable service and extensive coverage across a plethora of applications. The expanding demand for high-bandwidth services and the proliferation of mega satellite constellations highlight the limitations of traditional exclusive satellite spectrum allocation approaches. Cognitive Radio (CR) leading to Cognitive Satellite (CogSat) networks through Dynamic Spectrum Management (DSM), which enables the dynamic adaptability of radio equipment to environmental conditions for optimal performance, presents a promising solution for the emerging spectrum scarcity. In this survey, we explore the adaptation of intelligent DSM methodologies to SatCom, leveraging satellite network integrations. We discuss contributions and hurdles in regulations and standardizations in realizing intelligent DSM in SatCom, and deep dive into DSM techniques, which enable CogSat networks. Furthermore, we extensively evaluate and categorize state-of-the-art Artificial Intelligence (AI)/Machine Learning (ML) methods leveraged for DSM while exploring operational resilience and robustness of such integrations. In addition, performance evaluation metrics critical for adaptive resource management and system optimization in CogSat networks are thoroughly investigated. This survey also identifies open challenges and outlines future research directions in regulatory frameworks, network architectures, and intelligent spectrum management, paving the way for sustainable and scalable SatCom networks for enhanced global connectivity.

Figures

Tables

Table 1. TABLE I : Related literature surveys and contributions

Reference	Regulations & standardizations	Network Architectures	CR & DSM	AI & ML	Performance Metrics	Key Contributions
[fontanesi2025artificial]	N	L	M	H	N	Discuss the limits and constraints of AI & ML integration for SatCom onboard operations, along with different use cases and evaluate possible hardware solutions.
[liu2020review]	N	L	H	N	L	Elaborate on CR and application scenarios of CogSat communication. Summarises work on Sectrum Sensing (SS), spectrum allocation and power control.
[jasim2021survey]	H	M	H	H	N	Spectrum management regulations, architectures, approaches, and tools for Unmanned Aerial Vehicles (UAVs) communication.
[kodheli2020satellite]	H	H	L	L	N	Technological analyses on SatCom. Access control and networking challenges of satellite networks with testbed outcomes.
[zhang2022spectrum]	M	M	H	N	N	Spectrum sharing in aerial/space networks, with techniques, spectrum utilization rules and associated key technologies.
[heydarishahreza2024spectrum]	N	H	M	M	M	Explore LEO-Terrestrial network integration in the context of interference in different network deployment scenarios.
[ortiz2022machine]	N	N	H	H	L	ML for radio resource management in GEO satellites
[al2023artificial]	N	M	L	H	N	AI techniques for integrated terrestrial-massive satellite networks.
Ours	H	H	H	H	H	Explore regulation and standardization with existing DSM techniques along the line of AI & ML for SatCom networks. Discuss the network architectures that enable CogSat in SatCom networks and detail KPIs for such networks.
N - No impact, L - Low impact, M - Medium impact, H - High impact

Table 2. TABLE II : DSM in SatCom: State-of-the-Art

Ref.	Network Setup	Core Functionalities				DSM Technique					Key Contributions
		OSA	CSA	SS	Database/REM	Frequency Reuse	Power Allocation	Beam Pointing	Beam Hopping	Beam Forming
[gu2021dynamic]	GEO-LEO	-	$✓$	-	-	$✓$	$✓$	-	-	-	Cooperative service method to address the co-linear interference in GEO-LEO coexisting network.
[bohai]	GEO-LEO	-	$✓$	-	-	$✓$	-	-	-	-	DSA approach amid GEO-LEO system interference.
[yixuan2022underlay]	Satellite-Terrestrial	-	-	$✓$	-	-	-	-	-	-	Outage performance analysis of terrestrial users under the interference temperature constraint.
[benedetto2021unauthorized]	GEO-non GEO	-	-	$✓$	$✓$	-	-	-	-	-	A higher order moments-based SS approach for detecting unauthorized users.
[ren2021novel]	GEO-LEO	-	-	-	-	$✓$	-	-	-	-	Improved algorithms for frequency reuse in satellite communication.
[yao2023green]	Satellite-Terrestrial	$✓$	$✓$	$✓$	-	-	$✓$	-	-	-	A two-way-really aided model and a novel power allocation scheme for CSTN.
[feng2023radio]	Satellite-UAV-Terrestrial	-	$✓$	-	$✓$	-	-	-	-	-	6G-NTN network integration approach with coordinated spectrum sharing among satellite-UAV platforms to enhance coverage.
[zhang2021intelligent]	Satellite-Terrestrial	-	-	$✓$	$✓$	-	-	-	-	-	An analytical framework for a cloud-based satellite-terrestrial integrated network to improve spectrum utilization.
[zhao2024joint]	GEO-LEO	-	-	-	-	-	$✓$	$✓$	-	-	An optimization framework for spectrum sharing and interference minimizing in a GEO-LEO coexisting environment. A joint model-based and model-free DRL framework for the proposed framework.
[weththasinghe2023band]	GEO-LEO	-	-	$✓$	-	$✓$	-	-	-	-	A measuREM ent apparatus for frequency reuse opportunities in L-band. A spectrum analysis using Inmarsat data.
[weththasinghe2024optimising]	LEO- Terrestrial	-	-	-	-	$✓$	-	-	-	-	Investigation of multiple frequency reuse schemes and beam size optimization approach for LEOs in multi-beam 6G-LEO integrated network
[weththasinghe2024cognitive]	GEO-LEO	-	-	$✓$	-	-	-	-	-	-	SS platform on LEO satellites to investigate multiple GEO spot beam detections.
[li2023joint]	Satellite-Terrestrial	-	-	-	-	-	-	-	$✓$	-	Beam hopping and adaptive dynamic multiple access scheme to optimize beam scheduling.
[li2023pattern]	Satellite-Terrestrial	-	-	-	-	-	-	$✓$	$✓$	-	QoS guaranteed pattern design and power management scheme for LEO-Terrestrial beam hopping network setup.
[lin2020robust]	Satellite-Terrestrial	-	-	-	-	-	-	-	-	$✓$	A secure beamforming approach millimeter wave band sharing satellite-terrestrial network

Table 3. TABLE III : Comparison of OSA and CSA DSM schemes.

	OSA	CSA
SU status	On and Off	Always On
Environment awareness	geo-location database, SS	Channel estimations, interference prediction
PU precedence techniques	Terminate SU transmission through PU detection	Interference control through performance margins

Table 4. TABLE IV : Literature on leveraging ML for Intelligent Spectrum Management.

Ref.	ML Categorization & Algorithms				Network Setting	Focus Area				Key Contributions
	SL	USL	RL	Other		SS	Interference Detection	Spectrum Allocation	Resource Management
[9910435]	CNN-LSTM	-	-	-	GEO	$✓$	-	-	-	Realistic data collected from Tiantong-1 GEO satellite is used as training data.
[10000832]	LSTM	-	-	-	GEO	$✓$	-	-	-	Realistic data collected from Tiantong-1 GEO satellite is used as training data.
[9387707]	CNN	-	-	-	Satellite-Terrestrial	$✓$	-	-	-	Blind threshold algorithm eliminating the impact of noise uncertainty. Performance is evaluated against simulated and real-world data.
[ngo2023machine]	SR, LR, DT, SVM	-	-	-	GEO-LEO	$✓$	-	-	-	Multiple supervised learning approaches evaluated with cyclostationary feature detection.
[9293140]	CNN-BiLSTM	-	-	-	GEO-LEO	$✓$	-	-	-	Spectrum prediction evaluated under LEO using shared spectrum. Results were evaluated against several supervised learning algorithms.
[ngo2023multi]	-	-	MADRL DQN	-	Satellite-Terrestrial	$✓$	-	-	-	MADRL approach for reconfigurable intelligent surfaces assisted cognitive satellite-terrestrial networks.
[ferreira2018multiobjective]	-	-	DQN	-	Satellite-HAP	-	-	-	$✓$	Proposed approach is a part of the PoC for International Space Station’s CR engine.
[liu2021gnss]	Standard NN & CNN	-	-	-	Satellite-Terrestrial	-	$✓$	-	-	Explore GNSS interference events at airports that can affect airplane landings.
[saifaldawlaconvolutional]	-	CAE	-	-	GEO-NGEO	-	$✓$	-	-	CAE-based method for non-GEO satellite interference detection at GEO users.
[he2024collaborative]	-	-	PPO	-	GEO-LEO	-	$✓$	-	$✓$	DRL approach to resolve co-channel interference in GEO-LEO coexisting setup.
[9273081]	-	-	DQN	-	NTN	-	-	$✓$	-	Spectrum access approach accounting NTN BS dynamics.
[9047860]	-	-	DeepCA	-	LEO-Satellite Internet of Things (SIoT)	-	-	$✓$	-	A novel DRL-based approach dubbed DeepCA for optimal channel allocation considering energy constraints.
[8372935]	-	-	DQN	-	GEO-SIoT	-	-	$✓$	-	DSA method for multibeam satellite systems leveraging image-like tensors to extract environment information.
[Liu2018]	-	-	DQN-CNN	-	GEO-SIoT	-	-	$✓$	-	DSA algorithm for multibeam satellite systems.
[9387703]	-	-	-	AI	Satellite-Terrestrial	-	-	$✓$	-	SDN & AI integrated approach toward intelligent spectrum management.
[guo2022multi]	-	-	MADDP	-	Satellite-HAP	-	-	-	$✓$	Trajectory and power optimization to reduce latency.
[mishra2024traffic]	-	-	-	RNF	Satellite-Terrestrial	-	-	-	$✓$	SDN-based spectrum sharing/traffic offloading using RNF and feed-forward NNs.
[ML6]	-	-	Actor Critic-DQN	-	Heterogeneous SAT	-	-	$✓$	-	DRL and MADRL approaches to optimize resource utilization in SDN/NFV-enabled networks.
[abdu2022deep]	SMDL & MMDL	-	-	-	GEO SAT	-	-	$✓$	-	SMDL & MMDL-based accelerated method for bandwidth/power allocation.
[Xin2018]	-	-	CNN-DQN	-	GEO-SIoT	-	-	-	$✓$	Image tensor-based state reformation approach for spatial/temporal feature capture.
[Liao2020]	-	-	MADRL DQN	-	GEO-SIoT	-	-	$✓$	$✓$	Cooperative dynamic MADRL approach for distributed intelligence.
[ortiz2022machine]	-	-	CNN & DQN	-	GEO-SIoT	-	-	$✓$	$✓$	Power, bandwidth, and beam hopping across SL and RL layers.
[ortiz2022supervised]	LR	-	-	-	High throughput-SIoT	-	-	$✓$	-	Power and bandwidth allocation.
[duong2023machine]	DQN	-	-	-	Satellite-Terrestrial	-	-	-	$✓$	Multi-beam approach using Deep Q-Network (DQN) optimization and game theory.
[saifaldawla2024genai]	-	-	-	GenAI	GEO-LEO	-	$✓$	-	-	VAE and TrID-based Generative AI to mitigate LEO-to-GEO interference.
[zhao2024joint]	-	-	JMB-ML	-	GEO-LEO	-	-	-	$✓$	Joint model-based/model-free DRL for beam and resource management.
[sadique2025link]	RNN & CNN	-	-	-	LEO	-	-	-	$✓$	Link scheduling in over Riemannian Manifolds
[yang2024csl]	-	-	-	CNN, self-attention, LSTM, & soft fusion	GEO-LEO	$✓$	-	-	-	CSS model to improve detection performance
[tang2024dynamic]	-	-	DDQN & DDPG MADRL	-	satellite-terrestrial	-	-	$✓$	-	Multichannel LEO spectrum sharing framework for terrestrial users

Equations

$$U=B.S.T$$

$$SUE=\frac{M}{U}=\frac{B^{\prime }.S^{\prime }.T^{\prime }}{B.S.T}$$

$$RSE=\frac{SU{E}_a}{SU{E}_{std}}$$

$$S_U=\frac{N_s-N_c}{N_T}$$

$$\text{Service Retainability}=1-FTP$$

Full text

I Introduction

Satellite networks have emerged as a cornerstone of global communication, providing extensive coverage and reliability with enhanced bandwidth across numerous applications. Operating in three primary orbits—Low Earth Orbit (LEO), Medium Earth Orbit (MEO), and Geostationary Equatorial Orbit (GEO), each satellite system offers distinct characteristics tailored for specific use cases. Mega LEO constellations, such as OneWeb, SpaceX’s Starlink and Amazon’s Kuiper, are transforming global connectivity by delivering low-latency, high-speed internet to remote and underserved regions [pachler2021updated]. Meanwhile, MEO satellites play a vital role in Global Positioning Systems (GPS), offering accurate navigation and timing services while playing a prominent role in civilian and military communication. GEO satellites operate in a fixed position relative to the Earth, remain essential for broadcasting, weather monitoring, and latency-resilient communication over vast areas.

The growing number of satellites and rising user demand highlight the need for efficient spectrum utilization in Satellite Communication (SatCom)networks, as limited communication spectrum remains a primary barrier for new SatCom operators to enter the market. This scarcity also intensifies competition for spectrum access, inadvertently contributing to higher service costs for users. Dynamic Spectrum Management (DSM)through Cognitive Satellite (CogSat), the integration of Cognitive Radio (CR)to SatCom networks emerges as a solution to this problem, leveraging intelligent and advanced solutions to improve the spectrum utilization. Traditional spectrum management methods, often static, constrained and pre-defined, are no longer sufficient to cope with the distributed, heterogeneous, and congested nature of modern satellite networks. In recent years, Artificial Intelligence (AI)and Machine Learning (ML)have emerged as powerful enablers for addressing these challenges by enabling data-driven adaptive decision-making. Through real-time traffic patterns, spectrum occupancy, and environmental conditions, AI/ML techniques can optimize spectrum utilization, mitigate interference, and support CogSat systems across multiple orbital and terrestrial networks [liang2020dynamic, zhang2022spectrum].

I-A Motivation

Technological advancements, including reusable rockets and ride-share programs, have accelerated the growth of modern satellite networks, spurred by the increasing demand for high-capacity broadband and resilient communication systems. The exponential growth of the SatCom industry is illustrated in Fig. 1. Reports indicate there are 12,149 active satellites are orbiting the Earth in mid-2025, and this number is growing exponentially, while communication satellites account for 79% of satellites in space [bryce]. The mega LEO satellite constellation boom started with OneWeb and Starlink launches has skyrocketed the number of active satellites in orbit and with Guowang and Amazon Kuiper, it is expected to multiply. With this augmentation, the need for optimizing satellite operations under environmental and operational constraints is paramount. Specifically, the SatCom spectrum should be utilized optimally as it is one of the most sought-after finite resources in this growing industry. However, the existing exclusive licensing approach has led to an artificial scarcity of the radio communication spectrum. This growing mismatch between regulatory rigidity and operational complexity underscores the critical need for dynamic and intelligent spectrum management frameworks that can be integrated into SatCom networks.

To address these challenges, the concept of CogSat networks has emerged as a transformative approach, integrating CR into the SatCom environment, enabling intelligent spectrum access and autonomous resource control. CR enables sharing frequency, space, time, and power information capacity dimensions through dynamically altering radio transmission strategies. Through CR, DSM enables to leverage of unutilized or underutilized spectrum, improving spectral efficiency. CogSat systems extend these into the unique constraints of space, incorporating native SatCom parameters such as orbital dynamics, multi-orbital network coordination, and delay-sensitive network configurations [kodheli2020satellite]. Through the integration of cognitive functionalities into the satellite architecture, CogSat networks aim to significantly elevate the resilience and efficiency of space-based communication systems, thereby addressing the pressing demands of modern commercial and defence SatCom applications [gupta2015cognitive].

Traditional spectrum governance are characterized by rigid, long-term, and exclusive allocations, which poses a significant bottleneck to the adoption of dynamic and cognitive spectrum access strategies to SatCom. Without a coordinated regulatory framework that embraces flexibility, interoperability, and real-time spectrum sharing, the benefits of intelligent spectrum management cannot be fully realized. Standardization efforts led by bodies such as the International Telecommunication Union (ITU), 3rd Generation Partnership Project (3GPP), Institute of Electrical and Electronics Engineers (IEEE) and European Telecommunications Standards Institute (ETSI) are therefore essential to establish harmonized protocols, interoperability requirements, and policy guidelines that accommodate cognitive functionalities, orbital diversity, and cross-network coordination. On the other hand, the performance evaluation metrics for these dynamic and intelligent spectrum management approaches are yet to be fully explored. Therefore, the discussions on regulations and standardizations alongside DSM performance indicators is not only timely but also fundamental to ensuring that CogSat networks can operate efficiently, securely, and equitably within an increasingly congested and contested spectrum environment [centenaro2021survey].

AI and ML, known for their proficiency in pattern recognition and decision-making, are key enablers of this vision. In CR, ML techniques are already applied to optimize resource allocation, interference management, and spectrum access [al2023artificial]. In the context of SatCom networks, ML can address the challenges of atmospheric effects, propagation delays, and dynamic interference levels, facilitating the realization of CogSat systems [fontanesi2025artificial]. Software Defined Radio (SDR) brings this flexibility to radio networks through allowing the control of radio parameters through software programs. On the other hand, Software Defined Network (SDN), the paradigm of decoupling control and data planes enables new avenues towards flexible networking. The combination of SDN and SDR is a key enabler of CR and associated DSM techniques, and they are widely utilized in modern network integrations. Network Function Virtualization (NFV) enables network functions such as routers and firewalls to run as Virtual Network Functions (VNFs), allowing them to operate beyond vendor proprietary boundaries. These technologies are widely leveraged in terrestrial and non-terrestrial network deployments, as they provide unprecedented flexibility. CogSat networks empowered through AI and ML and upheld through SDN, NFV, and privacy-preserving techniques like blockchain, have the potential to meet the growing demands of data-intensive applications in future satellite networks.

Therefore, it is paramount to explore how such technologies can be leveraged to enable intelligent DSM in SatCom networks to improve spectral efficiency. Table I presents an overview of the existing literature surveys on AI & ML integration, DSM, regulation, and standardization of satellite and high altitude networks with their key contributions. The majority of work on intelligent spectrum management along CR methodologies focuses on terrestrial communication networks; while some survey work targets specific directions of CR, such as SS and Radio Environment Maps (REMs). The work discusses AI & ML approaches for satellite networks, generally deep dive into the limitations and challenges of integrating them in SatCom networks without discussing the intelligent DSM and associated challenges. Contributions and gaps in the regulation and standardization of intelligent spectrum management are rarely discussed alongside performance metrics for SatCom in the literature.

I-B Contributions

A summary of the main contributions of this paper is as follows:

• •

We extensively discuss the enablers of intelligent DSM in SatCom along the lines of satellite network integrations, CR, AI/ML, SDN, NFV and edge computing.

• •

We explore the existing regulatory and standardization bodies on SatCom networks in the context of spectrum management and network integrations, with their contributions towards the advancement of the SatCom industry.

• •

We further explore Opportunistic Spectrum Access (OSA), Concurrent Spectrum Access (CSA), SS, and database techniques for SatCom along with an extensive evaluation of satellite network architectures leading to CogSat networks. In addition, we categorize literature on DSM for SatCom based on core functionalities and DSM techniques.

• •

We investigate AI and ML methods leveraged in SS, spectrum allocation, interference mitigation and resource management. We further discuss ML model training and operational resilience, while extensively categorizing the state-of-the-art ML methods on satellite spectrum management.

• •

Accurate performance evaluation methods are essential for adaptive resource management and overall system performance optimization in CogSat networks. We focus on such metrics and discusses their evaluation criteria.

• •

Finally, we highlight the challenges and future directions in regulatory, architectural, and ML implementations in the context of realizing CogSat systems toward sustainable and scalable SatCom networks for global connectivity.

I-C Paper Organization

The remainder of this article is organized as follows. Section II elaborates on the key enablers of intelligent spectrum management in SatCom. Section III discusses the regulations and standardizations established by the prominent authorities on satellite spectrum management. Section IV focuses on DSM for CogSat, and Section V discusses the state-of-the-art ML techniques proposed in the literature for satellite spectrum management and categorizes them based on their primary focus areas. Section VI details the performance evaluation metrics for intelligent spectrum management in SatCom and Section VII extensively discusses the challenges in realizing DSM through CogSat networks within the existing framework. Finally, Section VIII concludes the paper.

II Enablers of Intelligent Spectrum Management in Satellite Communications

II-A Satellite Network Integration

II-A1 Intra-Satellite Network Integration

GEO, MEO, and LEO satellite networks leverage the strengths of different orbits to provide enhanced connectivity. Due to the unique capabilities of these satellite networks, network operators are leaning toward utilizing the SatCom spectrum harmoniously to enhance service delivery and widen the subscriber base with hybrid satellite networks. Although the concept of integrating different satellite networks has been discussed, connectivity is typically established through gateways, and networks are operated separately. Therefore, a higher level of integration between these satellite networks is required to leverage the full scale of capabilities. Fig. 2 illustrates a GEO-LEO integrated network architecture where control and data traffic are separated, which leverages SDN and NFV technologies [10211619]. A similar network setup where GEO satellites operate as the control layer is discussed and evaluated in [afzali2023integrating].

In addition to the theoretical analysis of inter-satellite network integrations, Intelsat, a GEO satellite operator, and OneWeb demonstrated broadband connectivity through an integrated GEO and LEO network for the U.S. Army. The demonstration exemplified a throughput increase in twofold as the users connect to both LEO and GEO networks simultaneously and route blended traffic, leveraging both satellite networks [intelsat_oneweb]. Eutelsat and OneWeb combined their resources in 2023 to become the world’s first LEO-GEO integrated satellite operator. They provide back-haul and corporate network connectivity for fixed users and mobile connectivity for maritime and in-flight users [eutelsat]. In September 2023, MEO satellite operator SES announced broadband connectivity, integrating its O3b mPOWER MEO satellites and Starlink’s LEOs for cruise ship passengers. The service is expected to facilitate data and voice over the internet for a guaranteed throughput of up to 3 Gbps per ship located anywhere in the world [ses].

II-A2 Satellite-Terrestrial Network Integration

Fifth Generation (5G) and beyond cellular networks are expected to have seamless integration with satellite networks, thus taking another step toward solving the challenging problem of global connectivity [latva2019key]. Coverage through satellite networks to remote areas is identified as a cost-effective alternative for expensive terrestrial network expansions [Telstra_5gNTN]. On the other hand, satellite networks provide a unique advantage for maritime and aviation communication requirements. However, even though satellite networks are equipped to facilitate remote connectivity requirements, satellite and terrestrial cellular networks operate as separate entities connected through network gateways. Therefore, a tight integration between the two networks is required to facilitate low latency and high bandwidth requirements that modern user applications demand. The following are concrete examples of 5G terrestrial and satellite network integrations to deliver connectivity over the hybrid network setup.

Ericsson, Qualcomm, and Thales Alenia Space: Collaboration of these three companies successfully initiated the world’s first publicly announced integration of 5G Non-Terrestrial Networks (NTN)-based networks set up on 3GPP standards and tested a 5G standard call through LEO satellite channels. Their test-bed has accounted for inherited challenges such as delay and Doppler effects while ensuring seamless satellite handovers [Ericsson_5gNTN]. 2. 2.

Vodafone and AST SpaceMobile: The two companies collaboratively delivered the world’s first space-based 5G voice call using an unmodified regular 5G enabled smartphone. The direct-to-device test call was made from Hawaii to José Guevara through AST SpaceMobile’s BlueWalker 3 LEO [Vodafone_5gNTN]. 3. 3.

Sateliot and Amazon Web Services (AWS): A LEO satellite constellation operator, Sateliot and AWS have partnered to deliver cloud native 5G Narrowband Internet of Things (NB-IoT) service. This will enable Sateliot LEOs to act as 5G mobile transceivers and connect unmodified NB-IoT devices to the 5G network globally [Sateliot_5gNTN]. 4. 4.

Telstra and OneWeb: The two companies deliver LEO satellite-based cellular backhaul throughout Australia. Voice calls have been made through the OneWeb LEO constellation successfully, and the target of this collaboration to deliver 25 Gbit/s service to the remote mobile customers across Australia [Telstra_5gNTN]. 5. 5.

European Space Agency (ESA), Telesat and Amarisoft: The alliance between the three parties has successfully materialised the world’s first 3GPP (Release 17) NTN link over LEO, between the ESTEC 5G laboratory and the Telesat LEO-3 satellite, taking a step forward towards delivering 5G over NTN. The 5G bidirectional link between gNB and the user supports 3 bits/s/Hz over adaptive modulation selection from Quadrature Phase Shift Keying (QPSK) to 64-Quadrature Amplitude Modulation (QAM) [esa_5gNTN]. 6. 6.

Japan-Europe long-distance 5G network over satellites: This test network has incorporated Ku-band over GEO satellites to connect 5G networks between Japan and Europe, and they have evaluated 4K video, Internet of Things (IoT) data, and network control signals over this link. Progressing towards unified satellite-terrestrial networks, their findings provide evidence of successful 5G networks over satellite, with tolerable latency [japan_5gNTN].

II-B Cognitive Radios (CR)

The ITU-R SM.2152 defined the CR systems as a radio network setup which leverages technology to extract operational and environmental knowledge, and dynamically and autonomously adjusts its operational parameters and protocols based on the obtained information to achieve predefined objectives. This approach can be leveraged to enhance spectral efficiency, especially in congested bands. In the literature, CR systems often follow the hierarchical classification of Primary User (PU) and Secondary User (SU), aiming to grant SUs access to under-utilized spectrum without disrupting PU s. The initial phase of the cognitive process is to observe the spectrum and gather information about spectrum utilization. The second phase is spectrum analysis, where CR evaluates the observed signal characteristics. Based on the analysis, the optimal unused frequency band is selected for transmission according to the feature requirements [hassan2021survey]. CR process then reconfigures the software parameters to change transceiver frequency, switch methodology, and energy transfer approach [SS_survey1]. Literature has explored the concept of CR along multiple axes such as regulation [zhang2016cognitive, CR4], applications [CR6, CR7] and technology [hoyhtya2012application, 9003405] with the leap of AI/ML and with the adoption of SDR in the past decade [Wang2019].

Due to the inherited complexity in CR techniques, excessive standardization is paramount. The IEEE 802.22 standard, developed by the IEEE 802 LAN/MAN Standards Committee, is the first published standard on leveraging IEEE in licensed bands for Wireless Regional Area Network (WRAN) [802_22]. It elaborates on utilizing geo-location and SS to enable CR within dynamic radio environments, by identifying and using unoccupied frequency channels without interfering PU communications. Geographical locations are facilitated through a database of local licensed transmitters or PUs, while the SS techniques are leveraged to detect unused frequency channels for SUs to transmit. Further, CR methodologies are extended to efficient use of unlicensed spectrum through IEEE 802.15 standard [802_15], which focuses on the coexistence of Wi-Fi and Bluetooth networks. In addition, the Dynamic Spectrum Alliance was formed to oversee the advancement of CR standardization and deployments in 2014 [ds_alliance]. It has contributed significantly towards Television White Spaces (TVWS) utilization for broadband and developing CR standards in collaboration with regulatory authorities.

The following is a summary of materialized implementations of CR approaches:

TVWS in Rural Broadband: The widely adopted CR technology for commercial deployments. According to the Dynamic Spectrum Alliance, more than 10 successful commercial or pilot projects have been completed worldwide, including the United States of America (USA), the United Kingdom (UK), Japan, Canada, and Ghana [ds_alliance]. Another test project to utilize TVWS was successfully implemented in India, which aimed to provide broadband connectivity to rural areas at a low cost [khaturia2018tv]. Further, preliminary surveys have been done in the USA to regularize the use of TVWS [nekovee2009survey], thus providing broadband internet access in rural and underserved areas. 2. 2.

Defense Advanced Research Projects Agency (DARPA) XG Program: The Next Generation (XG) program initiated by DARPA primarily focused on developing CR technologies for military and public safety applications [darpa]. The program successfully demonstrated CR capabilities under challenging radio conditions and developed an architectural framework and protocols for Dynamic Spectrum Allocation (DSA), enabling military communications systems to use unutilized spectrum more efficiently, avoiding interference with other users. 3. 3.

Nokia’s CR Networks: Nokia has successfully carried out several CR implementations for testing purposes [mukherjee2015cognitive]. These tests focused on leveraging CR techniques to enhance spectral usage in high-demand urban environments. Furthermore, Nokia has developed a self-organizing network solution to dynamically change radio parameters to meet the user demand, thereby improving mobile network performance and user experience [8368999]. 4. 4.

NASA SCaN: John H. Glenn Research Center in NASA, led by the Space Communications and Navigation (SCaN) group, places a research platform on the International Space Station (ISS). This SCaN testbed, equipped with three SDRs, enables CR research on orbital communication platforms [ferreira2018multiobjective, reinhart2014space]

CogSat refers to the concept of adapting CR capabilities to SatCom systems, enabling real-time SS, dynamic access, and intelligent decision-making based on the environmental context and spectral occupancy. This enables satellites and heterogeneous space-terrestrial networks to identify underutilized spectrum bands and intelligently reuse them without causing harmful interference to primary/license users. Furthermore, enhanced CogSat techniques tailored to encapsulate spatial, spectral, and temporal conditions allow adaptive link configuration, interference avoidance, and load balancing across beams and frequency bands.

II-C AI & ML

ML is a subset of AI that enables machines or systems to learn and improve from experience without being explicitly programmed for specific tasks. The taxonomy of ML is complicated, considering the novel approaches branched out from the preliminary algorithms. Classifying ML algorithms is a complex task [mahesh2020machine], however, in this work, we are detailing well-explored ML methods and ML approaches that carry the potential to significantly impact satellite spectrum management.

II-C1 Supervised Learning

Refers to a type of ML methodology targeted at predicting or classifying outcomes using labeled data, where both input features and corresponding outputs are known. It is particularly effective for tasks requiring prediction or classification in dynamic environments such as satellite and terrestrial networks, for identifying signal patterns, resource allocation, or anomaly detection [burkart2021survey]. Supervised Learning (SL) plays an important role in enhancing decision-making and optimizing operations in dynamic and complex communication environments such as satellite networks [fourati2021artificial]. Key SL algorithms explored in the literature are as below:

• •

Linear and Logistic Regression: Used for predicting continuous outcomes (linear regression) or binary classifications (logistic regression). These are interpretable methods suited for simple relationships between features and targets [burkart2021survey].

• •

Decision Trees: Contrive a tree-like model to split data into branches/sections based on feature thresholds. Accountable and effective for both classification and regression [navada2011overview].

• •

Random Forests: An ensemble method that builds multiple decision trees and aggregates their predictions for robust and less over-fitted results [liu2012new].

• •

Support Vector Machine (SVM): Separates classes by identifying the hyperplane with the maximum margin between different categories. Effective for high-dimensional data but less interpretable [tang2013deep].

• •

Regularization Techniques: Introduces penalties to regression models to reduce over-fitting and handle multicellularity among features. LASSO, Ridge, and Elastic Net are a few regularization techniques found in literature [nusrat2018comparison].

• •

Ensemble Learning: Also referred to as Super Learning, combines multiple algorithms to optimize prediction performance by leveraging the strengths of each individual method [dong2020survey].

• •

K-Nearest Neighbours (KNN): Classifies a sample based on the majority class of its nearest neighbors in feature space, suitable for simpler, smaller datasets [ML3].

II-C2 Unsupervised Learning

Refers to a ML approach used to identify patterns and structures in unlabeled data [sindhu2020survey], which is particularly beneficial for environments such as satellite networks where vast amounts of data are generated without pre-classification. An overview of key unsupervised learning algorithms is as follows:

• •

Clustering: Group data points based on similarity. A widely used algorithm that falls under this is K-Means. This approach is often used to identify patterns in unlabeled datasets [camino2021milp].

• •

Association Rule Learning: As the name implies discovers relationships, associations, and patterns between variables. Apriori algorithm is a well-explored method, which can be categorized under this method [al2014improved].

• •

Anomaly Detection: Identifies irregular patterns in data, useful for detecting signal interference or system faults [gunn2018anomaly].

• •

Autoencoders: Neural networks specifically designed for dimensionality reduction and feature extraction. This approach has unique advantages in efficiently processing large datasets [saifaldawlaconvolutional].

• •

Principal Component Analysis (PCA): Reduces data dimensionality while retaining key information. Further enabling faster and more accurate processing for resource optimization and fault analysis [vazquez2021machine].

In the context of satellite communication, clustering algorithms can group satellite transmission data or identify spectral usage patterns. Anomaly detection approaches can monitor signal behavior irregularities or identify jamming attempts. Adapting these Unsupervised Learning (USL) methods enables autonomous decision-making, allowing systems to adapt instantly to dynamic environments without human intervention [saifaldawlaconvolutional, li2021scalable]. The flexibility and scalability inherited by unsupervised learning techniques make the methods well-suited for diverse applications, such as resource management, fault detection, and data optimization across various domains [vazquez2021machine, sindhu2020survey, mahesh2020machine]. Considering the vast array of data complex network environments generate, advanced techniques like dimensionality reduction with PCA can enhance data processing efficiency by ensuring that models focus on the most relevant features of the data.

II-C3 Reinforcement Learning

Refers to a feedback-based methodology of ML, in which a learning agent takes actions in an environment based on the rewards offered for the action. For each preferred action which the agent takes in a particular state, it gets positive feedback, and bad actions gets penalized. The agent’s transition from one state to another depends on the previous state, the action taken, and the next state. There is no labeled training data in Reinforcement Learning (RL), therefore, the agent is bound to learn from its experience. Hence, the agent reacts with the environment and explores by itself while trying to maximize cumulative rewards [kaur2022comprehensive]. When complete information about the system is available, the dynamic programming approach can be used to determine the optimal policy. The formal model of ML is the Markov Decision Process (MDP), however, when complete system information is not available other than the sequence of past states, actions, and rewards, the Monte-Carlo method can be applied to get the optimal policy. Temporal difference learning takes a different approach to the above by not forming a system model [graesser2019foundations].

Deep Reinforcement Learning (DRL) adds RL and Deep Learning (DL) techniques together, enabling agents to make decisions in the state space using unstructured input data. Due to the capability of DRL to take large inputs, and the multiple neural network architecture makes DRL an ideal methodology to exploit real-world problems and achieve resolutions beyond human capabilities [graesser2019foundations]. Multi-Agent Deep Reinforcement Learning (MADRL), focuses on the behavior of multiple learning agents in a shared environment. In MADRL, each agent is responsible for its own actions and rewards, often pursuing different and sometimes conflicting objectives, leading to complex group dynamics. Compared to conventional DRL methods, MADRL is compatible with information sharing between agents. This helps in accelerating the learning of similar tasks and enables the achievement of better overall performance. In addition, when an agent or several agents fail, the remaining agents can take over their tasks, demonstrating the inherent robustness of this approach.

II-C4 Distributed Learning

Distributed learning is a paradigm of ML, where the workload is spread and shared across multiple nodes, which enables the processing of large data sets under time and computational resource constraints [verbraeken2020survey]. Parallelism is the key concept behind distributed learning, which facilitates data splitting across multiple nodes where each node processes a portion of the data. In addition, model parallelism enables dividing the ML model across different nodes where each node computes parts of the model. Under realistic settings, both terrestrial and non-terrestrial radio networks are wide-area networks with distributed computational resources. Processing large datasets generated through these networks demands higher computational resources, and the end nodes are inherently computationally constrained, thus making distributed learning an ideal approach for SatCom networks. Further, utilizing the limited communication resources to transmit sensory data to build the ML model affects the total system efficiency. Federated Learning (FL) is a well-established distributed learning approach that can counter the disadvantages of centralized learning approaches. In FL, data remains decentralized and only the ML model updates are shared with a central server, thus reducing the communication overhead while enhancing privacy and security [ML1, razmi2024board].

II-C5 Generative AI and Large Language Models

Generative AI and Large Language Models (LLMs) represent groundbreaking advancements in AI, enabling machines (essentially large deep neural networks) to understand, process, and generate human-like content with remarkable accuracy and coherence [yao2024survey]. LLMs, such as OpenAI’s Generative Pre-trained Transformer (GPT) series [roumeliotis2023chatgpt] and Meta’s llama [touvron2023llama] are a class of DL models built on the transformer architecture. These models consist of a large neural network framework that efficiently encapsulates long-range dependencies in sequential data through mechanisms like self-attention. On the other hand, Generative AI, a broader category encompassing LLMs, focuses on creating new content imitating the style and data structure it was trained on. Generative AI models extend to producing images, music, video, and even synthetic datasets, utilizing models such as Generative Adversarial Networks (GANs) [goodfellow2020generative] and Variational Autoencoders (VAEs) [kingma2019introduction], expanding beyond text generation. Moreover, as these LLMs continue to evolve, their impressive performance is adapted to robotics in generating control commands [driess2023palm], in biology for predicting protein structures [lin2023evolutionary], and in networking [wu2024netllm]. Such cross-domain adaptations have initiated a promising avenue for LLMs, which has significant potential to enhance operations in unexplored areas such as SatCom.

II-D Software-Defined Radio and Networking

II-D1 Software-Defined Radio (SDR)

SDR is a versatile radio communication approach that uses software to define and control radio frequency functionalities such as modulation, demodulation, signal processing, transmission power, and frequency selection [jondral2005software]. Unlike traditional hardware-based radios with fixed parameters, SDR relies on reprogrammable components, such as Field-Programmable Gate Arrays (FPGAs) and General-Purpose Processors (GPPs), to adapt its operation dynamically according to the environment of operation. This flexibility in SDR makes it an essential enabler of CR, empowering radios to modify their parameters in real-time based on environmental conditions, user demands, and regulatory requirements [jondral2005software]. SDR’s ability to reconfigure its transmission characteristics paves the way for advanced spectrum management techniques, thus making it foundational for enabling adaptive and intelligent capabilities that are vital for CR systems.

In the context of CogSat networks, SDR serves as the hardware platform that supports DSA and real-time environmental sensing, enabling environmental dynamic-based decision making and automated changes. For instance, CRs requires the ability to detect and exploit spectrum holes (discussed in Section IV) to dynamically adjust transmission parameters and switch between frequency bands under pre-programmed guidelines. SDR’s programmability and reconfigurability enable these functions, allowing seamless transitions across access communication protocols and frequencies. For example, a CR with an SDR platform can enable flexible transmission from a Wi-Fi band to a cellular band when spectrum congestion and other anomalies are detected, ensuring uninterrupted communication [gudipati2013softran]. Furthermore, SDR integration enables realizing advanced signal processing algorithms in SS and modulation recognition, which are crucial for identifying opportunities and mitigating interference in shared spectrum environments.

II-D2 Software-Defined Networking (SDN)

SDN introduces a paradigm shift by decoupling the control and data planes, thereby centralizing decision making [jiang2023software]. Integrating these approaches in SatCom systems play a pivotal role in enabling intelligent and DSM [li2016using, qi2022sdn]. SDN brings indispensable qualities to SatCom environments in managing current mega satellite constellations, which operate under a highly dynamic and time-varying network topologies. By maintaining a global view of the network, including channel allocations, Inter-Satellite Link (ISL) conditions, and traffic congestion, SDN enables adaptive and efficient spectrum allocation policies, fulfilling key intelligent spectrum management objectives. Moreover, SDN is a primary enabler of seamless integration of satellite and terrestrial networks into a unified, programmable infrastructure, enabling cross-domain policy enforcement, such as coordinated spectrum reuse and terrestrial to satellite network handovers [LP5]. Its programmable interfaces support the integration of AI/ML-driven spectrum management, getting the advantage of real-time telemetry for context-aware decision making. Furthermore, the multi-domain and hierarchical controller architecture of SDN ensures scalability and supports agile reconfiguration in response to satellite mobility and link disruptions [jiang2023software].

II-E Network Function Virtualization (NFV)

Traditional satellite infrastructures are characterized by monolithic, vendor-specific hardware, which constrains flexibility, increases operational costs, and limits the integration of new services and protocols. NFV addresses these limitations by separating network functions, such as routers, firewalls, performance-enhancing proxies, and load balancers, from proprietary hardware, allowing them to run as VNFs on commodity computing platforms [jia2021vnf]. These precedents can be leveraged to expand mega satellite constellations catering to dynamic demands [maity2024traffic]. Especially in satellite-terrestrial hybrid networks, NFV enhances operational efficiency by allowing service providers to dynamically instantiate and orchestrate VNFs that can adaptively manage spectrum usage based on traffic, interference temperatures, and Quality of Service (QoS) metrics. Through such virtualization, intelligent spectrum management becomes feasible via on-demand resource allocation, fast reconfiguration, and improved interoperability with terrestrial networks. NFV, in cooperation with SDN, facilitates distributed control, multi-tenancy, and service chaining across satellite and terrestrial segments, thus enabling end-to-end resource management intelligence. Such architectures support the introduction of use cases such as on-demand bandwidth allocation and edge processing, which are key to adaptive and efficient spectrum use in dynamic operational environments [LP9].

II-F Edge Computing

This methodology pushes computation and decision-making closer to where data is generated. Multi Access Edge Computing (MEC), an extension of edge computing, enables data processing with reduced hops in communication networks. In SatCom networks, edge computing clusters can be located in satellites, distributed earth stations, or they can take a hybrid form, scattering and processing data between satellite-earth station hybrid systems [zhang2019satellite]. Satellite networks relying on centralized cloud-based processing introduce significant latency and bandwidth overhead, especially for delay-sensitive applications, due to the inherited signal traveling distances. In contrast, through the satellite-terrestrial integrated edge computing networks and LEO-based edge computing paradigms, edge computing reduces the number of data traveling hops, thus minimizing the latency [xie2020satellite]. These advantages can be leveraged for spectrum analytics, local policy enforcement, and DSM directly as they demand real-time configuration changes, considering dynamics in SatCom networks. Therefore, edge computing is an important facilitator in reducing processing latency and back-hauling traffic, while enhancing spectrum responsiveness and autonomy in SatCom systems.

II-G Blockchain for privacy and security

Blockchain is emerging as a transformative enabler of intelligent spectrum management, as it addresses challenges related to security, decentralization, transparency, and automation of spectrum transactions [liang2020dynamic]. The inherent security within blockchain architecture offers a decentralized and tamper-resistant database, such as REM, for spectrum usage records, enabling transparent and secure sharing of spectrum access among stakeholders. In addition, blockchain facilitates DSM policy enforcement through smart contracts without relying on centralized authorities [chen2022blockchain]. This facilitates the creation of a self-organized spectrum market that supports real-time trading and leasing of unused spectrum, optimizing utilization across satellite networks. Moreover, blockchain secures SS, spectrum auctions, and dynamic access processes, enhancing trust among participants in a decentralized CogSat ecosystem [perera2024survey].

III Regulations and standardizations in Satellite Spectrum Management

This section discusses the existing regulatory and standardization bodies on SatCom networks in the context of spectrum management and satellite network integrations, with their contributions towards the advancement of the SatCom industry.

III-A IEEE

As a leading standardization body, the IEEE has played a significant role in advancing SatCom networks by developing technical standards that promote efficient spectrum utilization and interoperability. The frequency bands allocated to SatCom generally fall within the 1–40 GHz range, although application-specific satellites may operate outside this spectrum. The IEEE categorized them into seven frequency bands with their own characteristics and properties, making them suitable for specific satellite operations. These SatCom bands and their applications are discussed extensively in [gagliardi2012satellite]. The lower part of the spectrum has higher propagation qualities and, therefore, is utilized in applications with extensive coverage and low throughput requirements. Higher frequencies can facilitate more bandwidth and require higher transmission power to compensate for signal degradation. Therefore, most LEO satellites utilize Ku (12–18 GHz) and Ka (26.5–40 GHz) frequency bands to facilitate high-throughput connections, as they offer higher bandwidth and data rates. Additionally, the shorter wavelengths allow for smaller, lighter antennas and more efficient beam forming, which are advantageous for both satellite payload design and user terminals.

IEEE has played a crucial role in integrating SatCom with 5G and beyond next-generation wireless networks [fu2023satellite]. Established IEEE 5G and 6th-Generation (6G) working groups focus on defining the coexistence of terrestrial and satellite networks, particularly in shared and mmWave spectrum bands. IEEE ’s drive on research for beam-forming, Multiple Input Multiple Output (MIMO), and AI-driven spectrum management, enhancing the spectral efficiency of satellite-based NTN. Furthermore, IEEE’s collaboration with regulatory and standardization bodies such as ITU and 3GPP ensures that its efforts align with global telecommunications frameworks for seamless network integration. One of the most notable contributions towards intelligent spectrum management that can be adapted to SatCom platforms is the IEEE 1900 series through Dynamic Spectrum Access Networks (DySPAN) Standards Coordinating Committee 41 (SCC41), which focuses on DSA and CR systems critical for spectrum efficiency in satellite and terrestrial networks. It further defines key concepts in CR, policy-based radio, adaptive radio, and SDR, ensuring efficient spectrum utilization and interoperability, with technical guidelines for analyzing coexistence and mitigating interference between radio systems operating in overlapping or adjacent frequency bands. By adopting real-time spectrum monitoring, interference prediction, and adaptive mitigation strategies, IEEE 1900 further enhances spectrum-sharing efficiency for applications such as CogSat communications and terrestrial-satellite hybrid networks.

Additionally, IEEE 802.16 for wireless broadband access has been adapted for SatCom applications [IEEE80216], enhancing internet accessibility in remote areas. Protocols such as IEEE 802.22 for WRAN [IEEE80222] have also been instrumental in utilizing TV white spaces, which can be leveraged for satellite-terrestrial hybrid networks. Furthermore, IEEE has also made significant contributions to the global rise of IoT over satellite through standardizing low-power communication protocols [centenaro2021survey]. IEEE’s 802.11 AI/ML topic interest group focuses on the application of AI in wireless networks, as such developments carry the potential to optimize connectivity through intelligent spectrum allocation and interference mitigation.

III-B ITU

The ITU is the United Nations specialized agency for digital technologies. ITU coordinates global spectrum and satellite orbit usage via the radio regulations treaty and develops the technical standards that ensure networks and technologies connect seamlessly [ITU_regulations, sharma2013satellite]. As discussed above, satellite transmission frequency offers unique characteristic advantages that can be leveraged for different use cases. To this end, ITU has significantly contributed to characterizing the environmental factors, channel fade models, and dynamics, improving SatCom service efficiency. Some of the key points ITU that have been addressed in satellite link modeling are as below:

• •

Absorption, scattering, and depolarization by water, ice drops, clouds, and other hydrometeors in the atmosphere.

• •

Signal loss due to the refraction in the atmosphere.

• •

Antenna gain decreases due to the phase decorrelation caused by irregularities in the refractive index.

• •

Slow fading caused by beam bending, and rapid fading due to refractive index variations.

• •

Bandwidth limitations caused by multipath scattering.

• •

Varying elevation angle of LEO satellites.

• •

Faraday rotation and ionospheric scintillation.

In addition to discussing the above factors extensively in [ITU2], [ITU1] studies tropospheric and ionospheric effects, shadowing in satellite-to-land channels with measurements up to 20 GHz. Furthermore, the ITU studies present multi-path channel models for clear Line of Sight (LoS) conditions, a statistical model for mixed propagation conditions, and a physical statistical wide band model for mixed propagation conditions.

Through its radio regulations for international frequency management, the ITU establishes global policies that prevent harmful interference between satellite services by assigning specific frequency bands to various satellite applications/ operators [ITU_regulations]. To manage frequency assignment and satellite characteristics, the ITU categorizes the existing satellite services into three divisions as Fixed Satellite Service (FSS), Mobile Satellite Service (MSS), and Broadcasting Satellite Service (BSS). Furthermore, the ITU maintained Master International Frequency Register (MIFR), ensuring the global recognition of frequency assignments, facilitating structured and interference-free spectrum management across borders [mifr]. In addition, the World Radiocommunication Conference (WRC), hosted by the ITU every four years, acts as the primary forum for updating regulations to address emerging technological needs. In the last WRC held in 2023, coexistence between Non-Geostationary Satellite Orbits (NGSO) and GEO satellites with established power limits for NGSO satellites, exploring new frequency bands for mobile satellite services, and equitable access to spectrum for developing countries were among the key points discussed [ITU_wrc23].

The ITU constitution acknowledges that radio frequencies and satellite orbits are limited natural resources, which necessitates rational, efficient, and economic usage to benefit both developed and developing nations. The coordination and regulatory framework ITU has established for satellite networks further strengthens the fair usage of SatCom spectrum, balancing the efficient utilization of spectrum resources with equitable access for all countries. In order to materialize this fair usage policy, ITU leverages a “first-come, first-served” coordination procedure, ensuring the orbital and spectral resources are allocated based on actual needs, thus improving spectrum utilization. Additionally, ITU has introduced planning mechanisms that reserve frequency allocations for future use, particularly safeguarding access for nations that do not have a presence in orbit and SatCom, thus ensuring that the spectrum is equitably distributed. To this end, ITU’s radio regulations framework establishes clear rules for frequency coordination, advance publication, and notification of SatCom networks, ensuring fair access while encouraging further advancements.

The ITU has progressively adapted its regulatory framework to support AI-driven spectrum management as it believes AI approaches to facilitate resource management in SatCom, such as coverage adjustments, capacity, and spectrum allocation. With ITU reports such as ITU-R S.2357 and ITU-R S.2361, the organization contributes to providing guidelines for naval AI aspects in FSS communications with mobile platforms and broadband access, respectively. Beyond allocation and coordination, the ITU has been instrumental in defining regulations for emerging satellite technologies, such as NGSO satellites, 5G and beyond NTN, and satellite-based IoT, in collaboration with other standardizing bodies, such as the 3GPP. The ITU ’s role in space sustainability is evident in its efforts to prevent spectrum congestion, reduce signal interference, and implement space debris mitigation policies. By continuously evolving its regulations, ITU remains at the forefront of global satellite spectrum governance, ensuring its sustainability and accessibility for enhanced SatCom networks.

III-C 3GPP

3GPP is the prominent standardization body for terrestrial communication; however, with the recent advancements in the SatCom sector, it also has significant contributions towards satellite and NTNs. Consequently, 3GPP has been instrumental in standardizing frequencies and related technologies for terrestrial networks, thereby enabling the seamless integration of SatCom into terrestrial mobile communication systems. S, L, and Ka are among the key spectrum bands the 3GPP has identified for NTN communication, considering user requirements and spectral characteristics. These standardized frequency allocations align with ITU regulations, ensuring global spectrum harmonization and preventing interference. 3GPP Release 18 is a key reference guideline to overcome hurdles in standardizing terrestrial and non-terrestrial networks, creating a common platform for both architectures. With Release 18, the 3GPP has encapsulated interference mitigation techniques like power control, beam-forming, and DSA into its NTN specifications, aiming to manage interference effectively within satellite-based communication systems, particularly in the context of 5G and beyond networks [3GPP_r18, 3GPP]. Furthermore, enhancements in waveforms, timing synchronization, and power management are introduced in 3GPP Release 18 to address challenges such as Doppler shift and latency, making SatCom networks more compatible with terrestrial network infrastructure.

3GPP Release 18 also makes significant contributions towards the NTN-terrestrial network integration, marking the first formal standardization support for mobile cellular network extension through NTN [3GPP_r18]. In reference to integrations with GEO and LEO constellations, 3GPP identified several implementation scenarios based on the payload, which are extensively discussed in  [vanelli20205g]. Primarily, these 3GPP architecture options can be categorized as transparent and regenerative. Transparent payloads can support multiple functions and user equipment without physical modification of the data packets, on the other hand, regenerative payloads are altered to match the transport network. Thus, focus has shifted to transparent payload options since they can harness the advantage of ISLs. In addition, user equipment can be connected to a relay node, rather than directly connected to the satellite network, which can be a popular adaptation considering power and radio connectivity limitations.

Fig. 3 illustrates the proposed 3GPP network architecture of satellite-terrestrial network integration for the direct access scenario. In the case of regenerative traffic, satellites should have the gNB functionality built onboard; thus, the gNB establishes an air interface with the gateway to route traffic to the Next Generation Core network (NGC). Satellites can handle transparent traffic without inbuilt gNB functionality, and the received traffic is transported through a gateway to the terrestrial network. In this case, the next node can be a gNB. Regardless of the payload type, user equipment should access the satellites using an NR-Uu air interface to establish a proper data link layer connection.

3GPP contributions are extended in IoT and Machine-Type Communications (MTC) over satellite networks, enabling NB-IoT and LTE-M to function in NTN environments. 3GPP has ensured spectrum efficient deployments of satellite-based IoT solutions by defining technical specifications for low-power, wide-area IoT communications over satellites. This enables IoT applications to extend beyond terrestrial coverage to support applications in maritime, remote agriculture, and disaster response. These industries benefit from the real-time monitoring and connectivity SatCom networks provide through NB-IoT Furthermore, with the adaptive modulation and coding techniques, 3GPP ’s standards dynamically adjust to spectrum conditions, optimizing data transmission efficiency. Through these policy developments and standardizations, 3GPP has significantly contributed to the integration of satellite services into mobile networks, expanding connectivity to remote and underserved areas while maintaining efficient spectrum utilization, thus paving the way towards global connectivity [3GPP_NTN].

III-D ETSI & Europe

The ETSI is another primary standardization body that plays a pivotal role in the context of SatCom systems, services, and applications, including fixed, mobile, broadcasting and satellite navigation systems. Its contributions towards satellite-terrestrial integration architecture can be considered as a key enabler in industrial satellite-5G collaborative networks. Through its Satellite Earth Stations and Systems (SES) technical committee, ETSI develops standards for SatCom networks, as well as satellite navigation and earth station equipment. ETSI has further contributed to standardizing Broadband Satellite Multimedia (BSM) communications, enabling efficient IP-based satellite access networks. The modular BSM architecture, as outlined in ETSI TS 102 292 and TR 101 984, allows the integration of satellite-dependent transmission technologies with common satellite-independent IP networking functions, such as QoS, multicasting, and security. ETSI contributions in this context enhance the efficiency of satellite communications through optimizing IP interoperability and compatibility between satellite and terrestrial networks, thus supporting next-generation broadband and NTN advancements.

ETSI TR 103 611 recommends the fundamental standards in the integration of satellite and terrestrial networks within the 5G ecosystem, enabling seamless connectivity. Within it, ETSI defines architectural models and access scenarios, such as direct, indirect, and mixed 3GPP access of NTN integration with terrestrial 5G systems. The report further ensures interoperability between the two networks of interest. By integrating High-Altitude Platform Stations (HAPS) and satellites into 5G, ETSI promotes flexible, multi-layered network architectures capable of handling diverse communication needs, from Enhanced Mobile Broadband (eMBB) to Massive Machine Type Communications (mMTC) and Ultra-Reliable Low-Latency Communications (URLLC). In addition, ETSI TR 103 124 focuses on defining satellite-terrestrial network integrations. In that recommendation, ETSI emphasizes the rationale behind integrated networks and the value additions it can bring to the identified combined satellite-terrestrial network integrations. These integration standards open doors for co-existence of NTN and terrestrial networks, thus highlighting the importance of efficient resource utilization in terms of spectral and computational assets. DSM approaches allowing adaptive spectrum sharing between satellite and terrestrial systems while minimizing interference can play a pivotal role in these co-existing network setups.

Apart from the ETSI, Europe has several other authorized bodies that regulate and standardize SatCom operations. The European Conference of Postal and Telecommunications Administrations (CEPT) is one such organization, and they work closely with the ITU and ETSI in developing harmonized spectrum policies for European countries. CEPT’s electronic communication committee defines policies for satellite spectrum management, including frequency coordination and interference mitigation for the European region. The European Broadcasting Union (EBU) is another regulatory body that works within the European region, and it sets standards and policies related to the use of satellite networks for data transmission within its jurisdiction. EBU is responsible for standardizing satellite-based broadcasting services, including Direct-to-Home (DTH) TV, satellite radio, and multimedia distributions. In addition, EBU also works on satellite frequency coordination for broadcasting applications across Europe.

III-E National and Regional Authorities

In addition to the global framework established by the ITU, national and regional regulatory authorities impose guidelines on satellite spectrum management. They are often affiliated with national security interests and political agendas, considering the surveillance and monitoring capabilities satellites possess [sharma2013satellite]. Each country, through its national telecommunications authority, such as the Federal Communications Commission (FCC) in the United States, the Australian Communications and Media Authority (ACMA), or Ofcom in the United Kingdom, manages and licenses spectrum use within its jurisdiction. These authorities ensure that satellite operators comply with national laws and international obligations, often enforcing licensing conditions and monitoring for interference. On a regional level, organizations such as the CEPT, the Inter-American Telecommunication Commission (CITEL), and Asia-Pacific Telecommunity (APT) facilitate cooperation among neighboring countries and study related legal issues, coordinating spectrum use to minimize cross-border interference. Furthermore, they facilitate discussions, harmonize standards, and promote the development of SatCom technologies to ensure efficient spectrum usage and equitable access to satellite services across member countries. Regional agreements are significant in aligning national interests and obligations with international regulations, enabling smoother coordination for satellite operations across multiple national boundaries. These frameworks set by global and national authorities ensure that the growing demand for satellite services is managed efficiently and that spectrum resources are used responsibly across different geographical regions.

IV Dynamic Spectrum Management for Cognitive Satellite Communications

DSM has been widely recognized as a promising solution to address spectrum scarcity and has been extensively studied in the literature. DSM is primarily achieved through CR and vice versa, thus there exists a huge overlap between the two technologies [liang2020dynamic]. Keeping that in mind, in this section, we explore OSA, CSA, SS, and database techniques for SatCom along with an extensive evaluation of satellite network architectures leading to CogSat networks. In addition, we categorize literature on DSM for SatCom based on core functionalities and DSM techniques, as presented in Table II.

IV-A Opportunistic Spectrum Access

In DSM settings, PU s are the privileged user group, as the name implies, and the service provider/network operator has priority over the wireless transmission frequency, as they have bought the rights from a regulatory authority. SU s intend to communicate on the same frequency as the PU s with minimal interference between the two systems. The literature identifies PU s as licensed users and SU s as non-licensed users, which is a debatable fact, as SU s are also a responsible user group utilizing/sharing a specific frequency band and should also be licensed and recognized by a regulatory authority. This spectrum access policy, where SU s can transmit without a dedicated frequency band, is known as DSA. According to SU s’ spectrum access, DSA is categorized into two models: OSA i.e., spectrum overlay or interweave paradigm and CSA, also referred to as spectrum underlay. Table III provides a high-level comparison of OSA and CSA approaches [liang2020dynamic]. OSA in CogSat networks leverages the concept of dynamic utilization of spectrum holes or the portions of the frequency spectrum that are temporarily unoccupied by PU s, and allows SU s to transmit in the identified gaps. In the broader context of SatCom, where spectrum is often spatially and temporally underutilized due to the orbital movements of non-GEO satellites, OSA enables SU s to access these unused frequency bands opportunistically, as the name implies [gu2021dynamic]. This is enabled through real-time SS and geo-location databases, which are discussed in the latter part of this section.

Unique challenges in SatCom environments affect the implementation of OSA, thus the realization of CogSat systems. Therefore, advanced techniques empowered through wideband SS enabled through SDR s are critical for detecting spectrum holes across extensive frequency bands. Even though OSA policies have been discussed in the context of terrestrial networks [santivanez2006opportunistic], it is yet to be fully evaluated and discussed for SatCom networks. This highlights the importance of policy agility, the ability to dynamically adapt to varying regulatory and spectrum usage policies, as it is essential for OSA in a global satellite context. OSA-driven CogSat realization demands the integration of machine-readable policy frameworks that can be updated in real time, enabling seamless operation across different geopolitical regions and spectrum environments.

The primary commercial benefit of OSA in CogSat is the spectral efficiency improvement that SatCom operators can obtain through successful implementations. Such approaches are evaluated extensively in the literature, a OSA approach for agricultural sensor network deployment is discussed in [bayrakdar2020employing], while the work in [patil2024comprehensive] discussed OSA in terrestrial mobile networks. Furthermore, a ML deployment strategy for OSA is evaluated in [zhu2020machine]. In the broader context, OSA represents the majority of CR concepts evaluated in the state-of-the-art, as it enables the dynamic access of underutilized spectrum with non or minimal interference to PU s. Moreover, OSA fosters innovation in SatCom services, such as broadband internet, GPS, and remote sensing, by allowing the coexistence of multiple communication systems within the same spectral resources.

IV-B Concurrent Spectrum Access

CSA is another pivotal technique within CogSat networks to achieve efficient spectrum utilization. Unlike OSA, where SU s transmit only when the PU s are idle, CSA enables SU s to transmit concurrently with PU s on the same frequency band. This is achieved by ensuring that the interference generated by the SU transmitter at the primary receiver remains below a tolerable threshold, known as the interference temperature. This approach allows for continuous SU transmission, eliminates the need for constant SS, and ultimately supports higher spectrum reuse, especially in dense traffic environments such as satellite hotspots. A typical CSA model considering CogSat environment is illustrated in Fig. 4. The figure further shows how the SU transmitter inevitably generates interference to the PU, in frequency reuse scenarios, if the PU is within its transmission range. Therefore, to realize CSA, the SU transmitter has to deploy cognitive approaches to predict and minimize the interference at the PU to an acceptable level, so that the PU can achieve the required QoS level in its transmission.

CSA approaches are typically developed considering multi-user scenarios such as terrestrial cellular networks and satellite networks. Therefore, leveraging multi-user diversity to enhance the overall performance of the SU system by prioritizing SU s that generate minimal interference to PU s represents an extension of the CSA approach, as thoroughly investigated in [yixuan2022underlay, benedetto2021unauthorized]. Another CSA approach is to share the spectrum without an interference limitation to PU s. In these network setups, both primary and secondary users have equal priorities to access the spectrum. License shared access [chan2015feasibility], and spectrum sharing in unlicensed bands explored in [etkin2007spectrum] are the two main avenues explored.

It is important to note that secondary system throughput and the QoS guarantees for the primary system are inherently conflicting objectives. Therefore, making optimal spectrum allocation decisions is vital to reach a balance between the two networks. CSA relies on Channel State Information (CSI) estimation methods and interference prediction to share the spectrum with SU s while maintaining satisfactory service quality for PU s. The secondary system can make improved decisions when it has access to the primary system CSI and other critical network information. In [SMT2, bohai], the authors proposed network architectures where primary and secondary satellite systems share network information and user locations to improve spectrum allocation decisions in the secondary system, minimizing the interference to PU s.

IV-C Spectrum Sensing

SS in CR networks refers to the technique by which SU s detect and access unused spectrum bands by monitoring the radio environment without interfering with PU s. Prominent SS methods explored used in CR networks are as follows:

• •

Energy Detection: Measures the energy level of the received signal and compares it with a predefined threshold to measure the channel occupancy [ren2021novel].

• •

Matched Filter Detection: Uses a matched filter to maximize the Signal to Noise Ratio (SNR) for detecting a known signal. Requires prior knowledge of the PU’s signal [sharma2012satellite].

• •

Cyclostationary Feature Detection: Exploits the cyclostationary properties of signals, such as periodicity in their statistics (e.g., mean and autocorrelation) [ngo2023machine].

• •

Waveform-Based Sensing: Detects known patterns or pilot signals embedded in the primary user transmission [nasser2021spectrum].

• •

Radio Identification-Based Sensing: Required prior knowledge of PU transmission signal properties. Identifies the specific characteristics of the primary user’s signal to determine spectrum usage [yucek2009survey].

• •

Cooperative Spectrum Sensing (CSS): CR s share their SS information to improve detection accuracy and mitigate the effects of fading and shadowing [yao2023green].

• •

Compressive Sensing: utilizes the sparsity of spectrum occupancy to reconstruct signals using fewer samples than traditional methods [heimerl2024compressed].

• •

Eigenvalue-Based Detection: Uses the eigenvalues of the covariance matrix of received signals to detect the presence of PUs [chaurasiya2019fast].

• •

Machine Learning-Based Sensing: Applies ML to classify or predict spectrum occupancy based on training data [zheng2020spectrum].

• •

Hybrid Methods: Combines two or more SS techniques to improve performance [ngo2023machine].

However, the effectiveness of SS methods in SatCom networks is challenged by multi-path fading, large-scale shadowing, and the high variability of satellite transmission channels. CSS has emerged as a solution to overcome these limitations by allowing multiple SU s to exchange sensing information, enhancing detection accuracy, and mitigating the hidden PU problem. However, unlike terrestrial networks, where CR users may collaborate extensively, security and data integrity, along with the distributed nature of the SatCom network architecture, limit collaboration among SatCom networks, creating a barrier for CSS. Modern advancements exploit additional degrees of freedom, combining two or multiple SS techniques, leading to hybrid approaches. Furthermore, leveraging ML techniques brings an additional dimension to the table with improved prediction capabilities with past data and experience-driven approaches. These novel methodologies not only improve the detection accuracy but also minimize sensing time, ensuring faster adaptation to the dynamic satellite environment. Practical implementations of SS in CogSat radios must account for unique satellite-specific challenges and characteristics. For instance, the long transmission distances often result in low received signal strength for ground users in SatCom networks, which can inadvertently lead to lower Signal to Noise plus Interference Ratio (SINR) values considering the interferences and noise component. To counter this challenge, SatCom systems require low SINR sensing capabilities, with thresholds extending as low as -20 dB, as specified in standards like IEEE 802.22 [cordeiro2005ieee].

IV-D Database Technique

Database techniques, particularly those involving REM s, are pivotal in enabling DSM for CogSat networks. REM s function as comprehensive databases containing critical environmental information for CR networks, including frequency channel allocations, Received Signal Strength Indicator (RSSI), interference levels, geo-located user activities, and regulatory policies [sayrac2013cognitive, zhao2007applying], Fig. 5 further illustrates the data accumulated in a REM. REM s enable environment-aware radio resource management in CogSat networks by leveraging direct observations and accumulated network data to construct a precise map of the radio environment. Therefore, in the context of CogSat radios, REM s are invaluable tools for tasks such as predicting PU frequency usage patterns, enabling SU s to identify and access white spaces, and facilitating spectrum sharing in both licensed and unlicensed bands. Thus, minimizing interference towards PU and SU s while improving overall spectral efficiency. Key database deployment approaches for CR are as follows:

• •

Centralized Database: A centralized database manages all the data related to spectrum availability. A central authority could manage and could be deployed leveraging on-premises or cloud deployment options [vizziello2011system].

• •

Distributed Database: Spectrum information is distributed across multiple nodes, avoiding reliance on a central database and minimizing communication latency, thus optimizing the database access process [ezzati2018distributed].

• •

Hierarchical Database: Combines centralized and distributed approaches, leading to a hierarchy of databases [zhao2006overhead].

• •

Blockchain-based database: Leverage blockchain methods leading towards secure and decentralized spectrum management databases [chen2022blockchain].

In addition, hybrid approaches with combinations of multiple database management techniques can be utilized to enhance the overall efficiency.

REM s can be further categorized into direct, indirect, and hybrid approaches based on the data collection and distribution methods, each offering unique insights into the radio environment [pesko2014radio]. Direct methods rely on real-time measurements, while indirect methods synthesize data through statistical models and historical information, and hybrid methods combine these approaches to enhance accuracy and predictive capabilities. These approaches highlight the role of REM in fostering adaptive and efficient spectrum management, even in environments where direct radio environment measurements are unavailable. In addition, at a structural level, REM s exist in two main forms, local REM s and global REM s. Local REM s, synchronized with their global counterparts, facilitate individual users or user groups with tailored CR environment insights. This hierarchical structure, which combines centralized and distributed databases, enhances the decision-making processes of CR s, enabling them to learn from accumulated experiences and adapt to changing conditions [zhao2006overhead]. By integrating REM data, CR networks can reduce processing overhead and adaptation times, making large-scale deployments such as CogSat networks more cost-effective in DSM. Furthermore, REM s enable advanced network functionalities such as situation awareness and multi-domain knowledge sharing, further enhancing the CR network’s overall intelligence and adaptability [4221484].

The practical deployment of REM is exemplified in systems like the WRAN proposed in [zhao2007applying], where sensing and measuring data are leveraged for realization. Using geo-location data and querying centralized REM-enabled databases, WRAN systems dynamically identify available spectrum resources, including operating parameters like channel availability, center frequencies, and power levels, ensuring PU protection while maximizing SU performance. Advanced REM architectures, enhanced through spatial statistical modeling and topology engines, further refine spectrum management capabilities. These innovations make database-driven approaches a cornerstone of DSM techniques, enabling CogSat radios to navigate complex and dynamic spectrum environments with enhanced efficiency and reliability.

IV-E Network architectures for Cognitive Satellites

CogSat networks can be broadly categorized into;

• •

Integrated CogSat, Hybrid CogSat, Cognitive Satellite Terrestrial Network (CSTN)- Coexistence of satellite and terrestrial network sharing the same spectrum

• •

Dual CogSat- Inter satellite system spectrum sharing,

based on the network architecture [SMT2].

IV-E1 Integrated Cognitive Satellite Networks

OSA in integrated CogSat or CSTN can be categorized based on the spectrum presidence, the first one being the satellite network taking priority over the terrestrial networks, where terrestrial users access the transmission frequency with minimal interference to the primary satellite users. A multi-beam satellite network operating as the primary system while sharing the frequency with randomly distributed terrestrial Base Stations (BS) is discussed and evaluated in [kolawole2017performance], where a BS thinning process is proposed to minimize the interference below a predefined primary system requirement. A time-splitting spectrum sharing approach is investigated in [9187830] where the primary satellite network shares the spectrum with the secondary terrestrial network. Similar network models have been investigated along different aspects in [zhu2021integrated, 9628097] and the references therein. The second approach is the inverse of the first scenario, where the terrestrial network gets the priority in accessing the dedicated frequency band, and the satellite network shares the terrestrial network frequency with minimal interference to the primary terrestrial users. Therefore, a transmission power and carrier allocation methodology is proposed in [7336495] for an integrated CogSat network where satellites exploit the microwave frequency band allocated to terrestrial networks.

CSA is also a possibility in integrated CogSat, where both satellite and terrestrial networks utilize the shared spectrum simultaneously maintaining a maximum interference threshold for PU s [8786872]. Performance of such networks, concerning interference power constraints imposed by terrestrial communication regulations, is evaluated with regard to bit error rate and network outage in [ruan2018performance]. An underlay CSTN where satellites operate in microwave frequency allocated to terrestrial use is proposed and evaluated in [ruan2017effective], and considering statistical delay QoS requirements, satellite network effective capacity is investigated under terrestrial imposed interference power limitations. Non-Orthogonal Multiple Access (NOMA) along the direction of spectral efficiency [le2021enabling] and signal relays in cooperative NOMA [singh2020underlay, 8949359] is investigated extensively for underlay CSTN in the literature.

Another integrated CogSat approach discussed in the literature is Cooperative Integrated-Cognitive Satellite Terrestrial Network (CI-CSTN), which is an additional step towards realizing satellite-terrestrial networks [7336495]. In CI-CSTN, the SatCom network connects remote terrestrial network users, while the terrestrial network connects the rest. These networks enable the utilization of a common frequency band with minimal interference between the satellite and terrestrial networks. Thus, improving the spectral utilization efficiency with minimal inter-network interference. A CI-CSTN network architecture is illustrated in Fig. 6. Considering the current direction of satellite-terrestrial network integrations, CI-CSTN are the most likely to be realized in a large-scale network.

IV-E2 Dual Cognitive Satellite Networks

Refers to the scenario where two satellite systems operate simultaneously over a coverage area utilizing the same spectrum band [gupta2015cognitive, bohai]. Spatial and spectral degrees of freedom are shared between the satellite systems in these networks. Based on the network architecture, literature on dual CogSat can be mainly categorized into the same type of satellites coexisting networks and GEO and non-GEO satellite coexisting networks, as illustrated in Fig. 7. The same kind of satellites can be either GEO, MEO, or LEO, sharing the frequency to serve a common area of interest. They can be from the same constellation or different constellations, but should utilize CR techniques in their deployments. Due to the growing demand for satellite-based services, multiple satellites can be deployed in close proximity, thus creating overlapping coverage areas. The fixed satellite services deployed to serve hot orbital zones like 13E and 19E with orthogonal frequency plans are conventional use cases for such networks [gupta2015cognitive]. Further, these satellite networks can leverage both mono-beam and multi-beam technologies to serve the users depending on the use case and capacity requirements. A study exploring the coexistence of multi-beam GEO satellites was done in [nguyen2017cognitive], in which the authors propose a cognitive beam-forming approach to mitigate the uplink co-channel interference.

In GEO and non-GEO satellite coexistence networks, in-line interference is an additional component to the interference generated from the co-located satellites. Further, orbital relative motions of non-GEO satellites add further complexities to these networks. To this end, Skybridge and Teledesic LEO satellite systems proposed to reuse the GEO frequency band for their transmission. Ideally, Skybridge LEOs proposed to terminate transmission within a certain distance from the equator to mitigate interference to GEO users [ashford2004non]. On the other hand, Teledesic planned to terminate transmission when the satellite coverage footprint intersects with the equator. Earth terminals would only transmit with the Teledesic LEO satellites if their latitude is northerly within the sub-satellite point in the northern hemisphere, and the relative sub-satellite point latitude has to be southerly in the southern hemisphere to initiate a successful transmission [stuart1998frequency]. However, neither of these cognitive frequency reuse approaches are materialized [gupta2015cognitive].

IV-F Dynamic Spectrum Management Techniques

Techniques such as frequency reuse, power allocation, beam pointing optimization, and beam hopping are integral in realizing DSM in SatCom systems. The key objectives of these DSM techniques in a CR setting are to improve spectral efficiency, maximize the PU and SU system throughput, and minimize or eliminate the PU interference while maintaining an operational level of interference for SU s.

IV-F1 Frequency Reuse

This is a widely adopted DSM technique where the same frequency band is reused in the non-overlapping coverage area, maximizing spectrum efficiency. This approach is already in operation in both terrestrial [xie2010frequency] and non-terrestrial [kisseleff2020radio] networks under general (non-cognitive) settings, as the licensed spectral bandwidth for each operator is limited and the user capacity requirements are high. For instance, the Ka-band in High Throughput Satellite (HTS) s employs a frequency reuse factor of 4 or higher [hasan2016ka], effectively dividing the available spectrum across a multi-beam environment, while minimizing inter-beam interference. This approach can be extended for CR through opportunistic frequency reuse techniques, as SU reusing PU spectrum with non or minimal interference to the PU network [luo2022research]. In the context of CogSat radios, intelligent frequency reuse schemes utilize advanced algorithms to dynamically adjust the reuse pattern based on user density, interference levels, satellite dynamics, and traffic demands [bohai]. Moreover, the 3GPP identifies frequency reuse as key for efficient spectrum utilization in NTN starting from Release 17 [giordani2020non], thus highlighting the importance of this technique going forward with the standardized implementations. Through advanced SS and REM approaches, the frequency reuse technique can be further enhanced to improve the spectral efficiency.

IV-F2 Power Allocation

A critical technique in enabling CR networks, which refers to dynamic power adjustment of SU transmission to optimize spectrum utilization while minimizing interference to PU s. This efficient power allocation approach ensures the SU operation with minimal disruptions to the PU network, achieving a balance between performance, energy efficiency, and spectrum fairness. In the context of CogSat networks, the power allocation technique ensures efficient distribution of transmission power across different beams, users, and frequency channels while optimizing link quality and spectrum utilization. Literature has explored several key approaches in realizing power allocation for CogSat systems, such as game theory [del2009power], optimization [ge2021joint], and ML-based [guo2022multi]. Furthermore, this technique can be extended to maintain the QoS levels, as CR refers to adapting radio network parameters according to the operational environment. For example, power allocation can adapt to rain fade conditions dynamically by increasing power in affected beams [destounis2011dynamic]. In multi-beam systems, power allocation-based cognitive mechanisms can be deployed to analyze interference patterns and adapt beam power levels to mitigate inter-beam interference, enhancing overall network capacity [hajipour2020computing].

IV-F3 Beam Pointing

Refers to dynamically adjusting the directions and shape of the radio beam, while maximizing the coverage and minimizing the interference to PU s, while improving spectral efficiency. This approach is also leveraged to cater to higher user demands, as the existing radio resources can be exhausted under unplanned scenarios. In SatCom radio networks, the beam pointing optimization approach uses advanced algorithms [he2024interference] and ML models [zhang2022deep] to predict traffic patterns, thus resulting in beam steering. For instance, due to dense deployment and high mobility, LEO satellite users in mega constellations can face interference issues, such problems can be addressed through situational aware beam pointing optimization approaches [he2024interference]. Furthermore, this approach can be extended into PU and SU operating CogSat environments to opportunistically utilize frequency bands, generating minimal interference to PU users [he2024joint].

IV-F4 Beam Hopping

This technique represents the dynamic allocation of beams to provide coverage for different geographical areas. Beam hopping enables service facilitation across multiple regions using a single beam by allocating time slots. CogSat networks, with the integration of SS and REM techniques, leverage this approach for adaptive resource allocation. For example, LEO satellites can schedule beam arrangements based on dynamic user demands. The available spectrum resources are also time-varying due to the dynamics of the satellites, as beam hopping facilitates time slot-based operations. This technique can facilitate agile beam allocations under constraints, and such approaches for CSTN are proposed in [li2023joint, li2023pattern]

IV-F5 Beam Forming

Refers to shaping the radiation pattern of the antenna array for concentrated energy transmission toward desired directions and suppressing interference, thus creating directional radiation patterns for a specific user or a region. This can be leveraged for CR networks, as it enables SU s connection amid the PU s in the radio environment. In the context of CogSat, advanced beam forming techniques are leveraged to adapt to varying spectrum and traffic demands in real time, thus directing the transmission energy of identified beams toward specific users or regions, maximizing signal quality, and minimizing interference of the overall communication system. To realize this technique, literature has explored advanced optimization schemes such as the discretisation and the Taylor expansion combined approach proposed in [lin2020robust] for CSTN, and the penalty function-based approach discussed in [zhao2021outage] for CogSat-Arial network setup. Furthermore, the potential of advanced ML methods towards this is also explored in [zhang2022deep], for MIMO in LEO satellite systems. Moreover, digital, analogue, and hybrid beam forming approaches are explored towards advancing this technique in [al2022review, rao20215g] and the references within provide a deep dive into those approaches.

V Machine Learning in Satellite Communication and Spectrum Management

This section explores ML and AI methods leveraged in SS, spectrum allocation, interference mitigation and resource management. We further discuss ML model training and operational resilience, while extensively categorizing the state of the art ML methods on satellite spectrum management in Table IV.

V-A Machine Learning for Spectrum Sensing

SS can be characterized as a binary classification problem that can be solved using supervised and unsupervised ML algorithms, in which the classifier has to determine the availability and unavailability of a radio channel of interest. Energy and probability vectors can be used as features in ML algorithms to predict spectrum hole availability [janu2022machine]. Long Short-Term Memory (LSTM) is a classification of Recurrent Neural Network (RNN), with the capability to effectively learn and remember long-term dependencies in sequential data. A study of SS for SatCom, taking the propagation delay into calculations, is presented in [10000832]. The authors highlight that neglecting the propagation delay in satellite links can cause co-frequency interference at ground level, and to mitigate this, they propose a DL-based joint LSTM and autoregressive moving average (LSTM-ARMA) SS scheme. Reliable SS under low SNR conditions, otherwise identified as SNR-wall, is a key challenge in NTN s. To this end, a SS scheme using a combined convolutional neural network and long short-term memory (C-CNN-LSTM) to mitigate the effect of low SNR is proposed in [9910435]. Further, a Convolutional Neural Network (CNN)-based approach to improve SS under low SNR conditions in space-air-ground integrated networks is evaluated in [9387707], where the authors derive a likelihood test for SS under the Neyman-Pearson lemma using CNN.

In [ngo2023machine], the authors have explored a cyclostationary feature detection method in the context of dual satellite networks and proposed an ML approach to improve the SS. Further, they have evaluated the performance of the SVM, Decision Tree, Logistic Regression, and Softmax Regression supervised learning algorithms for dual satellite network SS. DQN is an RL algorithm that combines Q-learning, which maximizes the expected future rewards of taking a given action in a given state, of a defined RL environment, with deep neural networks. DQN enables the agent to learn optimal policies directly from high-dimensional sensory inputs. In [kumar2022deep], a DQN-based SS approach is proposed and evaluated against an energy detection algorithm and a CNN algorithm. The authors highlight the performance improvement of the proposed DRL methodology, even with the quantity of relatively small training data. However, the full potential of DRL against SS for SatCom networks is yet to be fully explored in the literature.

V-B Machine Learning for Spectrum Allocation

Spectrum allocation in satellite networks under CR settings itself is a complex problem governed by the temporal, spectral, and frequency characteristics of the environment. Conventional spectrum allocation methods often rely on static rules or simplistic models that fail to adapt to the rapidly changing conditions of SatCom networks. However, ML offers robust solutions to this complex problem, taking multiple factors such as user demand, signal strength, and interference levels into account, and makes real-time spectrum allocation decisions. Particularly, DRL algorithms with the inherited ability to interact with the environment and learn from the feedback have shown superior performance in solving multi-dimensional problems similar to spectrum allocation in SatCom networks. The adaptability and predictive capabilities based on historical data and real-time inputs of DRL algorithms make it a powerful tool for enhancing the spectral efficiency of SAT systems.

A Dynamic Channel Allocation (DCA) methodology for multi-beam satellites leveraging DQN is proposed in [8372935], where the authors introduce an image-like tensor to represent the state, thus encapsulating spatial and temporal features of the SatCom environment. LEO satellites are designed with power constraints to reduce production and deployment costs; thus, power efficiency is the dominant factor in highly dynamic LEO satellite constellations. Therefore, a power-efficient channel allocation approach empowered by DRL is presented in [9047860] contemplating a satellite and IoT environment. In addition, a DQN-driven multi-user access control approach for NTN is presented in [9273081], where the authors propose a methodology to improve the long-term throughput of the ground users by minimizing frequent handovers. DCA methods can be used to minimize co-channel interference in SatCom systems. Therefore, mitigating the flows in DCA based on beam traffic load and user terminal distribution, an improved DRL-based DCA algorithm for multi-beam satellite systems is proposed in [Liu2018], to minimize service blocking probability.

V-C Machine Learning for Interference Mitigation

In order to facilitate high throughput requirements in modern applications, and to improve spectral efficiency frequency reuse techniques are leveraged, and it is a widespread methodology deployed in almost every wireless Wide Area Network (WAN) s. Frequency sharing and reuse within and between satellite networks generate co-channel and inline interference, which is an identified challenge [SMT2]. Interference minimization in satellite networks can be achieved primarily through adjusting parameters such as Effective Isotropic Radiated Power (EIRP), antenna direction, and frequency planning. The capabilities of ML techniques facilitate improved solutions in predicting these communication parameter adjustments amid the highly topological and dynamic characteristics of SatCom. ML models analyze vast amounts of real-time data from satellite sensors, ground stations, and user terminals to detect patterns and anomalies, enabling proactive measures to prevent or reduce interference, thus leading to optimized SAT networks facilitating improved service to the users.

Autoencoders are neural networks designed to learn efficient and compressed data representations by encoding input data into a lower-dimensional latent space, reconstructing the original data from this compressed representation. A convolutional autoencoder-based interference detection approach is proposed in [saifaldawlaconvolutional] for GEO and non-GEO coexisting satellite networks. Further, a Generative AI (GenAI) methodology for interference management for GEO frequency sharing with non-GEO satellites is present and evaluated in [saifaldawla2024genai]. A dynamic interference management methodology for LEO downlink is presented in [yun2023dynamic], where the authors elaborate on DQN performance over other ML algorithms on downlink throughput maximization. A CNN based approach to detect and mitigate interference in the Global Navigation Satellite System (GNSS) was proposed in [liu2021gnss]. In contrast, an LSTM algorithm for interference detection in SatCom networks is evaluated in [9384473]. In addition, a collaborative interference avoidance method for GEO-LEO coexisting satellite systems leveraging the Proximal Policy Optimization (PPO) algorithm is proposed in [he2024collaborative].

V-D Machine Learning for Resource Management

In SatCom networks, ML techniques are widely used to optimize key resource utilization, including transmission power, bandwidth, and computational capacity, enabling more efficient and adaptive network operations [ML6]. ML algorithms facilitate dynamic bandwidth allocation based on traffic demand and optimize onboard computational resources for tasks such as data compression and routing. Given the high cost and limited feasibility of satellite maintenance due to their altitude, ML can also play a critical role in predictive maintenance, helping to detect potential issues before they escalate. These capabilities collectively enhance overall network performance and support real-time adaptation of communication parameters, ensuring seamless and reliable service even in challenging environmental conditions.

Scheduling bandwidth to improve transmission efficiency and coverage in satellite networks is a challenging problem considering the environmental dynamics. A MADRL approach is proposed to solve this problem, considering a GEO satellite environment in [Liao2020]. Very High Throughput Satellites (VITS) is a satellite-terrestrial integration approach discussed in literature [giambene2018satellite]; however, the traffic demand for VITS is not uniformly distributed, and this initiates the requirement for flexible payload architectures. Hence, [9448341] presents a dynamic resource management methodology leveraging DRL. Further, a resource management framework for SatCom compatible with SDN/NFV-based management structure is studied in [ML6], which supports intercommunication with different satellite systems. The authors then use DRL for resource allocation in the proposed method. Demand-based dynamic resource allocation in satellite networks is a high computational task due to the variation and complexity of the parameters governing the problem. Authors in [abdu2022deep] identified this as a barrier to the practical deployment of such approaches and present a methodology combining conventional optimization and DL techniques. Through simulations, they show the proposed approach takes less time to optimize the parameters, resulting in less use of satellite resources.

V-E Training

One of the critical architectural questions needed to be addressed in adapting ML for SatCom is training location; should training occur onboard the satellite or on the ground? This distinction is more than semantic, as it fundamentally affects system decision latency, adaptability, computational requirements, and overall responsiveness [al2023artificial]. As illustrated in Fig. 8, there can be multiple solutions to this problem, as it is directly affiliated with the QoS satellite system delivered to its ground users. Onboard ML training, often conflated with online learning, entails training models directly on the satellite using data collected during its operation. This approach supports real-time adaptation to evolving environmental conditions, enhances autonomy, and eliminates additional latency caused by communication with ground stations [ren2021interaction]. However, it imposes significant challenges due to the constrained power, memory, and computational capabilities of the modern compact satellites [hussein2021reducing]. Despite these limitations, recent advances in low-power AI accelerators and edge ML chips (e.g., Intel’s Movidius and NVIDIA Jetson) are acting as key facilitators for onboard ML deployment in satellite systems [fontanesi2025artificial].

The other option is to leverage the computational abundance of terrestrial infrastructure for ML model training and off-board the trained model to satellite systems. Also referred to as on-ground training, this methodology can also benefit from the extensive data sets to develop large-scale models [ortiz2023onboard, furano2020towards]. This approach is typically associated with offline learning, where models are pre-trained on static datasets before deployment. The on-ground strategy simplifies model development and version control while enabling thorough testing prior to mission deployment, leading to more robust models. However, it also introduces latency in dynamic decision making and lacks the adaptability to adapt to orbital anomalies and mission changes unless frequent model updates are transmitted to the satellite, which can be bandwidth and time-constrained. While this approach is more practical for complex and large ML models such as LLM adaptations, its reliance on predefined training data means that unforeseen operational scenarios may lead to performance degradation.

Hybrid strategies are gaining traction to address the trade-offs between onboard and on-ground ML training for SatCom networks. These architectures combine the strengths of both domains by performing online learning on the ground using near real-time or recorded satellite telemetry, then transmitting updated models or decision parameters to the satellite [fontanesi2025artificial]. Incremental learning is another hybrid approach, where a model pre-trained on the ground is fine-tuned onboard with local data, thus customizing the model with unique characteristics. Such configurations enable systems to adapt over time while managing satellite constraints. Furthermore, distributed learning also materializes solutions, allowing multiple satellites or a whole constellation to collaboratively train a global model without sharing raw data, preserving bandwidth and privacy. Ultimately, the choice between onboard, on-ground ML and hybrid training must align with mission requirements, computational budgets, latency tolerance, and operational resilience, which are unique characteristics for different satellite constellations and use cases.

V-F Operational Resilience

SatCom networks are prone to unexpected changes due to environmental changes and space debris [ya2016mitigation, ren2021interaction]. Therefore, reliability and resilience of ML for SatCom systems are critical attributes that complement performance and ensure consistent operation under those uncertain, resource-constrained, and dynamic conditions. Unlike performance, which measures how well an ML agent executes its intended task, resilience incorporates and depicts the ML agent’s ability to generalize across varying scenarios and resist failures. In addition to the environmental uncertainties, SatCom networks pose unique challenges in operational resilience due to limited onboard computational capacity, delayed feedback loops, safety-critical operations, and partially observable high-dimensional state spaces. For instance, deploying DRL models directly onboard for real-time inference can be hazardous, as misinformed actions may cause irreversible satellite damage and degrade user QoS. To address these risks, recent studies explore model-based and offline DRL approaches, where agents learn from historical data or simulations, avoiding costly real-time experimentation [xie2024computation]. Hardware advances, such as radiation-tolerant AI accelerators and compact edge ML processors, now enable limited onboard training and inference, further enhancing robustness while maintaining safety margins [ya2016mitigation, fontanesi2025artificial].

Another significant concern in the robustness of ML models in SatCom is the presence of system delays, both in communication and reward feedback. Traditional DRL assumes immediate feedback, which is impractical in real-world satellite networks, due to the model training location and network architecture constraints. Delays in observation or reward reception due to long propagation distances can lead to outdated decisions if mishandled. To mitigate this challenge, delay-aware MDP frameworks and artificial training delays have been proposed to align training conditions with operational realities [chen2021delay, xie2024computation]. Techniques such as state augmentation and temporal correlation exploitation—using models like Deep Deterministic Policy Gradient (DDPG) or Echo State Networks (ESN), help compensate for outdated CSI and allow ML agents to perform accurate resource allocation and power control in dynamic environments [omid2024reinforcement]. Additionally, decentralized MADRL architectures and hierarchical clustering strategies have been explored to accelerate beam hopping and spectrum management under uncertainty [zhao2023delay]. These innovations collectively underscore that robustness in ML-enabled SatCom is not merely a product of model accuracy but a consequence of architecture, training realism, hardware adaptability, and resilience to systemic delay and operational variability.

VI Performance Evaluation in Satellite Spectrum Management

In CogSat networks, where dynamic spectrum sharing and coexistence with terrestrial or other satellite systems are common, leveraging accurate performance evaluation methods is essential for adaptive resource management and overall system performance optimization. This section focuses on these performance metrics and discusses their evaluation criteria.

VI-A Spectrum utilization

A primary metric of overall system performance and a measure of the effectiveness of CogSat deployment, which also offers a normalized view of how well the spectrum is utilized in shared environments. ITU discussed metrics to measure spectrum in radio communication networks along three dimensions [ITU_spectrum_effi]. Namely, Spectrum Utilization Factor (SUF), Spectrum Utilization Efficiency (SUE), and Relative Spectrum Efficiency (RSE). The SUF ($U$) is defined as the product of frequency bandwidth ($B$), space ($S$), and time ($T$), which is given in Eq. 1. In which $S$ refers to the GEO metric space or area of interest, and in the context of SatCom networks, this can be a line representing a GEO or a LEO orbit. $T$ is the time denied to other potential users. ITU highlights the fact that time can be ignored in some scenarios, considering the continuity of the service. However, in cases such as broadcast and single-channel transmission, and in CR environments where frequency is shared, the time factor should be considered.

[TABLE]

The SUE is defined as a complex criterion $SUE=\{M,U\}$, where $M$ is defined as the useful effect obtained with the aid of a network of interest. The ITU further simplifies this as Eq. (2) to the ratio between the useful effect and SUF.

[TABLE]

In the above equation, $B^{\prime},S^{\prime}$, and $T^{\prime}$ represent actual measurements of occupational bandwidth, coverage area and operating time, respectively. The RSE, which is given in the equation below, is introduced as a ratio of SUE providing the same type of service or as a ratio relative to a theoretical system.

[TABLE]

In Eq. (3), $SUE_{a}$ and $SUE_{std}$ are the spectrum efficiencies of actual and standard communication systems, respectively. In addition, spectral efficiency can be measured as the total throughput achieved per unit bandwidth. This reflects the overall performance improvement of the system under constrained bandwidth utilization, which is what CR radio systems are primarily defined to achieve. Another approach to measure spectrum utilization is presented in Eq. (4), where it is presented as a function between the number of successfully allocated channels ($N_{s}$), number of channels in collision with PU system ($N_{c}$) and the total number of channels available for sharing ($N_{T}$). This is more suitable for CR systems where spectrum reuse is deployed.

[TABLE]

VI-B Radio Network Interference

Performance evaluation metrics, such as SINR, SNR and Interference to Noise Ratio (INR), are common physical layer radio network measurements that can also be leveraged to evaluate CogSat networks. Radio network interference is the primary factor in performance degradation in radio networks. The inter-network interference between PU and SU systems and intra-network interference within the SU system should be considered in evaluating interference in the context of CogSat systems [bohai]. SatCom links are designed with an interference tolerance level, also called interference temperature or interference margin. In [itu-r], ITU recommends fixed GEO service networks operating in frequencies below 30 GHz to design and operate their links to tolerate interference levels up to 25% of the total system noise power when the network does not practice frequency reuse. This interference margin is reduced to 20% when the networks leverage frequency reuse.

However, these average interference thresholds are defined under general communication conditions, which might not be compatible with CogSat networks. In CSA CogSat environments, a strategically formulated Interference Power Constraints (IPCs) is required to govern the service quality. IPC regularizes the interference thresholds, and it can be deployed under two scenarios. Peak IPC enforces strict interference limits for all channel states, which is suitable for protecting PU QoS, and the average IPC allows for higher flexibility by averaging interference over time, which is beneficial for delay-tolerant PU applications [yixuan2022underlay, musavian2009fundamental]. When explicit IPC constraints are unavailable, CSA performance can also be evaluated by setting a maximum tolerable performance degradation (rate or outage) for the PU. It ensures that the SU’s concurrent access does not significantly degrade the PU’s communication quality. This methodology requires more extensive CSIs on PU transmit parameters, which are often challenging to obtain in practice [liang2020dynamic]. DARPA CR implementation is a good example of peak IPC, in which they defined and maintained 3 dB SNR degradation at a primary receiver [darpa].

VI-C Detection and False Alarm Probabilities

As discussed, SS plays a key role in realizing CogSat networks. OSA primarily relies on the effectiveness and reliability of SS to detect spectrum holes in PU transmission. The probability of detection, which measures the likelihood that the SU system correctly detects the presence of a PU in a channel, is a fundamental indicator of harmful interference to the PU s, as a higher probability of detection results in lower interference to the PU s. The probability of detection depends on several factors: the sensing time, SNR of the PU signal at the SU receiver, and the chosen detection technique (e.g., energy detection, matched filtering, or cyclostationary feature detection). In energy detection, for instance, detection probability is a function of the detection threshold, noise variance, and sample size. Achieving a higher PU detection probability typically requires more sensing time or sensitive receivers, which in turn can impact the agility and effectiveness of CogSat systems.

Conversely, the probability of a false alarm is the measure of the chance that an SU incorrectly detects a PU as active when it is idle. A high false alarm probability leads to missed transmission opportunities for SU s, reducing the spectrum utilization efficiency; therefore, a lower probability is preferred as it implies better access to spectrum holes for SU s [sharma2012satellite, cabric2004implementation]. Technically, false alarm probability is also influenced by the detection threshold, noise uncertainty, and environmental factors such as satellite movement and Doppler shifts. In energy detection schemes, a lower threshold value increases detection sensitivity but also raises false alarm probability. Therefore, careful design measures should be utilized, creating a trade-off between detection reliability and spectrum access opportunities, as the SU transmission in the PU s channel when it is active will lead to harmful interference.

VI-D Channel Availability

Under DSA paradigms, Channel Availability (CA) refers to the channel licensed to a PU being available for SU communications. Higher CA provides more freedom for the SU channel allocation, creating less interference to the PU system. In static satellite scenarios, such as fixed-beam GEO systems, CA primarily depends on the temporal activity pattern of the PU, which means that the channel is deemed available if the PU is inactive. However, in the context of LEO or MEO constellations, where satellites are inherently mobile, CA becomes a spatio-temporal variable. The relative motion between the primary and secondary satellite networks creates frequent changes in coverage overlap and interference regions, thereby affecting real-time CA. Hence, in CogSat environments, traditional CA estimations based solely on PU activity become insufficient for the reliable planning of DSA. Executing spectrum hand-off based on sensing decisions improves CA but introduces trade-offs concerning hand-off delays, false alarm probabilities, and data transmission durations as discussed in [lee2015channel].

VI-E Service Retainability

Service retainability in CogSat radio networks is a critical QoS metric that reflects the system’s ability to maintain uninterrupted communication sessions once established. In CR operations, especially under DSA schemes, this metric becomes increasingly important due to the opportunistic and often pre-emptive nature of spectrum access by SU s. A SU service may be interrupted mainly due to the arrival of a PU, channel failure, or lack of available backup spectrum in reserved bands.

[TABLE]

General form of calculating service retainability is given in Eq. 5, and the Forced Termination Probability ($FTP$) depends on multiple factors such as PU arrival rate and effective channel assignment rate [abdelgalel2024performance, tanveer2022reliability]. In CogSat scenarios, service retainability will also depend on the orbital dynamics of LEO and MEO satellites in addition to the static channel characteristics, which demands refined $FTP$.

VI-F Energy Consumption

Satellite communication systems operate under tight energy constraints due to a lack of power generation resources and their compact nature. Therefore, they are forced to embrace cross-layer design, onboard processing, and energy-aware scheduling to optimize power consumption [alagoz2011energy]. Depending on the channel capacity, the power consumption of a satellite can range from 500 to 2000 W [PEREZ199897]. Especially in the case of CogSat networks, additional computational tasks such as intelligent decision-making, beam-forming, and associated MIMO systems require a significant amount of system power. On the other hand, CR capabilities can be adapted to reduce the power consumption of the SatCom systems without sacrificing performance [hussein2021reducing]. Therefore, the amount of successfully transmitted data units per unit of energy consumed can be a Key Performance Indicator (KPI) for the overall CogSat system, reflecting on power utilization.

Energy-aware CR functionalities allow satellites to adaptively select modulation, coding, and power parameters based on environmental feedback, thereby minimizing redundant retransmissions and idle power dissipation. Metrics such as bits-per-Joule and energy-per-bit have been proposed to quantify these optimizations, ensuring that throughput remains efficient relative to the energy consumed per unit transmission [hussein2021reducing, alagoz2011energy]. Similar to mobile communication devices, power utilization is a primary metric in communication satellites and should be monitored regularly when operating under cognitive conditions. In addition, integration of ML and AI demands a significant amount of power from compact satellite systems, which primarily depend on solar power. This further highlights the importance of power consumption as a performance indicator in CogSat networks.

VI-G Latency, Delay and Jitter

Due to the signal traveling distance and associated SS and processing, CogSat networks are affected by additional delay components compared to general SatCom networks. Delay, the total time a data packet or signal takes to travel from the source to the destination, can be decomposed into several components such as transmission, propagation, process or decision and queuing delay. Latency mainly refers to the Rount Trip Time (RTT) or the reaction time. Jitter is the variation or inconsistency in the delay experienced by consecutive packets. Although propagation delay and jitter are primary parameters defining QoS, in CogSat systems, the delay associated with the process and queuing delay are of interest, as that is mostly affected by the CR functions.

Processing latency refers to the time a CogSat system takes to sense the spectrum environment, process the contextual information, and make an intelligent transmission decision—such as channel selection, power adaptation, or interference management. In CR environments, this translates into a critical KPI that directly impacts the system’s responsiveness and performance to real-time spectrum variations. Especially in OSA setups, performance latency reduces the SU transmission time, thus impacting the system QoS. Queuing in communication systems has been studied extensively in the literature, and advanced queuing algorithms have been proposed. However, CogSat networks require novel queuing algorithms due to the SS and intelligent decision-making processes incorporated in CogSat networks. The goal of these performance metrics is to push CogSat systems to bridge the gap between them and general SatCom networks, thus providing seamless transmission between the two networks.

VI-H Communication Overhead

In CogSat networks, signal overhead refers to the additional signaling required to facilitate DSM, coordination among satellites, and real-time SS. It quantifies the proportion of communication resources, such as bandwidth, power, and time consumed for control signaling rather than payload data [hui2025review]. The overflowing coordination and control messages, which add up to signal overhead, can lead to bandwidth and latency degradation in a communication system. In CogSat systems, especially those employing techniques like OSA, signaling overhead arises from SS reports, channel allocation decisions, hand-off signaling, and inter-satellite coordination. Additionally, due to the dynamic learning, adaptation, and optimization algorithms, an additional overhead/ signaling component is introduced in CR systems. These control signals are prioritized over the general traffic due to their importance to the system’s operation, thus adding a latency factor for the low-prioritized traffic. Therefore, the signaling overhead to payload ratio is a metric that reflects the efficiency of the communication protocols deployed in a CogSat system. Minimizing this ratio is crucial in bandwidth constrained and latency sensitive CogSat communication systems, thus highlighting its importance as a key performance metric in such systems.

VII Challenges and Future Directions

VII-A Regulatory Challenges

Global Consensus

Implementation of spectrum sharing and trading in CogSat communications faces significant challenges due to inadequate regulatory frameworks. Without generalized and agreed standards on spectrum sharing, satellite operators have to make individual decisions, thus making the process complex and unmanageable, due to the number of satellite network operators, regional and national spectrum management, alongside security regulations. Therefore, coordination between national and international authorities is paramount in realizing CogSat networks, as terrestrial spectrum management is facilitated at the national level, while international cooperation is essential for satellite spectrum management. ITU, being the recognized body in managing global spectrum regulatory requirements, coming up with a consensus to realize a global CogSat framework mainly relies on them. Clear definitions of EIRP and out-of-band interference thresholds should be defined and agreed [sharma2013satellite]. Additionally, the secondary dynamic access to dedicated bands for the government and military requires careful regulatory attention, particularly because of the associated security risks and emergency availabilities [pedram2023role].

Compatibility and Spectrum Ownership

Similar to the involvement of policymakers in building a platform for CogSat communication networks, satellite operators and equipment manufacturers have a pivotal role to play in realizing successful CogSat networks. Equipment manufacturers have to explore methodologies to develop affordable communication equipment compatible with both terrestrial and satellite networks. Furthermore, business models have to be developed to share the spectrum ownership between the networks [zhang2022spectrum, hassan2021survey]. Below are several possible modes of spectrum ownership

• •

Temporarily transfer of usage rights to another entity on a short or medium-term basis, including the full transfer of associated rights and responsibilities.

• •

Temporarily lease on a short-term basis, to be used based on the traffic demand. The primary holders retain their rights and obligations to the shared spectrum.

• •

Spectrum trading, where the primary holders may also retain their rights for the traded spectrum.

• •

Spectrum pooling, which is implemented as either pure pooling or a hybrid approach (e.g., combining fixed bands with a shared pool).

Collaboration across all stakeholders—regulators, industries, and researchers is key to overcoming these regulatory challenges and ensuring the successful deployment of CogSat systems.

Enhanced Protocols

The unavailability of communication protocols for CogSat is another challenge that standardization bodies need to address, as their absence can lead to significant interoperability challenges between the communication nodes in a CogSat network. Further, these protocols should have seamless integration capabilities with the existing globally recognized communication protocols, thus mitigating the potential inefficiencies or conflicts in network deployments. The absence of standard protocols for CogSat systems has hindered the trust in CogSat networks as a whole, among the satellite network operators and equipment manufacturers, creating an obstacle in realizing CogSat networks. Data privacy and security of the CogSat network are another aspect lack in standardization. Particularly when PU and SU operate in shared spectrum environments, sensitive information can be exposed through unauthorized access [RSMA5]. Additionally, concerns related to national security and integrity can arise due to the potential use of CogSat s for unauthorized surveillance and eavesdropping, which also emphasizes the requirement for advanced, regularized security measures [pedram2023role].

VII-B Architectural Challenges

Cooperative Networks

CR deployments in Satellite Terrestrial Network (STN) s and multi-orbital satellite networks demand unified network architectures and seamless operation between the networks [janu2022machine, yao2023green]. The relative motion between non-GEO satellites and users generates complex dynamics. Therefore, sophisticated spectrum sharing CogSat techniques should be utilized to predict and adapt the radio conditions in such environments, considering the associated spatial, temporal, and spectral parameters. These CR methodologies often assume a cooperative architecture with geographical location and frequency parameters of the users shared between the network of interest [9628097, ruan2018performance, bohai]. In addition, this coordination should account for the propagation characteristics and data processing delays in making real-time decisions. Therefore, refined cooperative network paradigms should be developed to cater to these unique requirements.

Latency and Delay

These are inherent challenges in SatCom, which can also affect real-time decision making and the responsiveness of CogSat networks [xie2024computation]. In CogSat systems, the radio transmission parameters adjustments, frequency reuse, and network routing decisions have to be swift and actionable, considering the dynamic conditions to minimize interference. However, the considerable propagation delays between the satellite and ground station, specifically in GEO communication, or between the different satellites in multi-orbital cooperative networks, can introduce complications in coordinating these decisions. For instance, LEOs demand rapid handovers and real-time adjustments for successful deployments of CR techniques. These operations are sensitive even to minor delays, which can result in QoS degradations in the CogSat deployments. Therefore, effective latency and delay management strategies, such as predictive algorithms, advanced caching, and data prioritization, are essential for an efficient CogSat network operation to maintain the QoS levels [omid2024reinforcement, zhao2023delay].

Scalability

Large-scale communication networks like satellite and mobile networks face rapid expansions, considering the growing global demand for connectivity. Therefore, CogSat networks must be designed to cater for rapid horizontal and vertical network expansions without compromising performance, reliability, or efficiency. Current mega constellations contain several thousand LEOs; thus the CogSat techniques should be capable of handling seamless communication, data processing, and spectrum management across a vast, distributed system. This demands advanced algorithms for dynamic satellite resource allocation under CR settings to maintain the QoS levels even under fluctuating conditions. Additionally, the CogSat methodologies must efficiently manage handovers, particularly in LEO and terrestrial coexisting CR networks, where both networks experience rapid handovers due to the user dynamics, further necessitating frequent communication to different ground stations or other satellites. Consequently, effective, scalable CogSat solutions must be developed to meet global demand and compatible with diverse applications, while maintaining the optimal performance levels in a growing network infrastructure.

Energy Efficiency

Power and computation are limited resources in satellite networks, as increasing them is associated with the manufacturing and launching costs, which can escalate the capital expenditure [hussein2021reducing]. Modern mega-satellite constellations have successfully reduced the weight of a LEO satellite to reduce the launching cost (the latest version of Starlink LEO only weighs 260 kg [starlink_weight]), thus limiting the onboard power and processing capabilities. However, this poses a significant challenge in realizing CogSat networks, as they require substantial computational resources to perform real-time SS, data analysis, decision-making, and dynamic network adjustments [8786872, ortiz2022supervised]. Apart from these functions, setups such as dual-CogSat networks demand frequent communication and coordination between neighboring satellites and ground network operation centers. This requirement is further elevated in hybrid CogSat networks, where rapid coordination between terrestrial networks is essential. To address these limitations, CogSat systems must employ energy-efficient algorithms and optimize resource allocations. Therefore, satellites can potentially offload some processing tasks to ground stations or more capable satellites, relaxing the strain on low-powered satellites.

Security and Privacy

These are two areas of paramount importance in modern networks, and they are further emphasized in SatCom, considering the global access and coverage spanning geographical boundaries [pedram2023role]. The spectrum sharing and sensing techniques used in the cooperative network architectures of CogSat networks are prone to vulnerabilities due to information sharing between the associated networks. Unauthorized access, eavesdropping, data interception, jamming, spoofing, and malicious use of spectrum resources are a few key vulnerabilities CogSat networks might experience. In Hybrid CogSat networks, terrestrial network integration adds another layer of complexity due to the secure communication requirements between heterogeneous network interfaces. Similarly, Dual CogSat networks demand secure cooperation between the satellite networks amid the dynamics of the environment. Facilitating these security requirements through advanced encryption and authentication processes can be challenging due to the limited computational resources in satellites. Therefore, CogSat networks should be equipped with a simple but robust, multi-layered security architecture. It should include strong encryption, authentication protocols, intrusion detection systems, and secure key management mechanisms compatible with the employed CR techniques [perera2024survey, chen2022blockchain]. Additional privacy-preserving techniques must be deployed in the CogSat setup to protect data integrity, considering the personal and confidential information communicated through modern broadband networks. Further, the CogSat network architecture should be resilient in detecting and mitigating in real-time, amid the dynamic network parameter changes and the decentralized nature of the CogSat operations.

Adapting SDN and NFV

Empowering SatCom networks with these two technologies offers substantial benefits in realizing CogSat networks as they offer flexibility and programmability to network setups [li2016using]. CR networks require real-time control of network resources to manage the dynamic behaviors, which SDN can efficiently handle through its inherently flexible and centralized control mechanism. Programmability and network-wide optimization are additional features that CogSat networks can benefit from theSDN architecture, as it enables complex network policy implementations and offers agility to maintain the network’s QoS through seamless coordination [qi2022sdn, jiang2023software]. NFV enables the virtualization of CR functions, allowing these functions to run on general-purpose hardware rather than specialized, dedicated devices [LP9, maity2024traffic]. This increases the flexibility and scalability of deploying CR functionalities across the SatCom network. One of the key challenges satellite networks face is the lack of flexibility to deliver new services with the existing hardware, as hardware upgrades in orbiting satellites are practically impossible. NFV provides a solution to that problem, thus enabling the deployment of upgraded cognitive solutions with the same hardware. However, these technologies are yet to be fully realized in SatCom networks, posing a significant challenge in realizing CogSat networks.

VII-C Machine Learning Implementations

Heterogeneity

Integrating ML algorithms into CogSat networks presents substantial challenges, particularly when interfacing with existing systems. ML models can be leveraged for tasks such as DSM, resource optimization, and interference mitigation in CogSat systems (refer Table IV). However, these cognitive capabilities empowered through ML should be compatible with existing satellite and terrestrial networks that operate on fixed frequency allocations and standardized communication protocols. The heterogeneity of these systems creates significant barriers to interoperability, as almost all the existing networks lack the flexibility to accommodate adaptive decision-making processes demanded by CogSat networks. Furthermore, the seamless integration of ML-based cognitive functionalities requires extensive modifications to existing network architectures in addition to general CR methods, thus necessitating the deployment of middleware solutions or protocol converters specifically catered to ensure ML model compatibility between CogSat and general networks. The challenge is further complicated by the need to maintain backwards compatibility, which can present significant limitations to the extent to which ML-driven CR innovations can be fully realized. Ensuring smooth operation across CogSat and legacy systems demands robust and interoperable ML frameworks capable of managing the complex interactions between these networks while providing accurate network predictions amid the complex dynamics of the SatCom environment.

Communication Overhead and Security

CogSat networks demand continuous data exchange between satellites, ground stations, and other network elements to maintain synchronized CR functions. These data exchanges, essential for ML model updates, coordination, and real-time decision-making in CogSat s, introduce considerable communication overhead, particularly given the limited bandwidth and high latency characteristic of SatCom links. The requirement to balance data throughput amid energy efficiency and processing constraints exacerbates this challenge. Therefore, advanced ML algorithms capable of reducing the communication overhead need to be developed. In addition, data transmission approaches should be cautiously chosen, communicating only necessary updates after the initial handshake. Moreover, ML integration into CogSat networks heightens the importance of security and privacy. ML models in CogSat networks often operate in the control plane, therefore, to protect the satellite networks and ML model’s performance, advanced cryptographic techniques, such as homomorphic encryption and secure multi-party computation, should be implemented within the ML algorithms to safeguard data during transmission and processing. Additionally, privacy-preserving approaches such as differential privacy should be incorporated into CogSat systems to protect the integrity of sensitive information, even as data is shared and processed across the network. Addressing these challenges is vital in realizing ML-driven CogSat networks to ensure they operate in a secure and efficient manner.

Data Scarcity

The effectiveness of a ML model relies heavily on the quality and quantity of data available for training and validation. However, in the context of SatCom networks, obtaining such data representing the network’s diverse and dynamic nature has practical limitations. Satellites operate across different orbits, covering vast geographical areas with separate frequencies, leading to highly diverse conditions that are difficult to encapsulate comprehensively in a dataset. Moreover, labeled data specific to SatCom scenarios, such as spectrum usage patterns, interference levels, and environmental effects, is often scarce or unavailable. In addition, CogSat networks have not been realized, thus adding the requirement of converting/REM odeling data captured in current satellite deployments. This scarcity limits the effective learning abilities of ML models. Additionally, the quality of the data is often compromised by incomplete information due to unsynchronized, antagonistic data collection methods, which can degrade ML model performance. Therefore, to address these challenges, there is a need for advanced data collection methods specifically catered to satellite environments, guaranteeing both quality and quantity. As an alternative approach, synthetic data generation methods can be explored, considering practical limitations.

Generalization

For ML models to be effective in CogSat environments, they must generalize profoundly across various network and communication conditions. However, the highly dynamic and diverse nature of SatCom makes this difficult. CogSat networks are multi-orbital and must operate in varying spectral conditions and under fluctuating network conditions, often with limited prior data, as highlighted above. An ML model trained on data from a specific scenario may not perform accurately under new, unseen situations, which might lead to poor decision-making and reduced QoS. This lack of generalization can be particularly problematic when dealing with rare or extREM e events, such as unexpected interference or sudden changes in spectrum availability. To enhance ML model generalization, it is essential to develop robust training strategies that expose the model to a wide variety of conditions, further highlighting the challenge of data scarcity. Methodologies such as transfer learning and behavior cloning can improve the ML model’s adaptation to new environments, and implement continuous learning approaches where the model evolves as it encounters new data.

Scalability

In parallel to satellite constellation expansions, especially with the advent of mega LEO constellations consisting of thousands of satellites, the ML models must scale effectively to handle the growing volume of data and the increasing number of nodes. Thus the ML models in CogSat networks should have the capability to handle the increasing data inputs while ensuring the ability to make real-time decisions across a distributed network without compromising performance. The distributed network architecture of CogSat networks, where satellites collaborate across multiple orbits and coordinate with ground stations, further complicates scalability in such systems. centralized ML approaches may struggle to cope with the sheer scale, which further necessitates the decentralized learning methods for CogSat systems that distribute the learning process across multiple satellites or ground stations. As the categorization highlights in Table IV, there exists a clear gap in the literature in distributed learning for CogSat. Therefore, developing ML models that can scale efficiently while maintaining robustness, accuracy, and responsiveness is essential to realize CogSat networks on a large scale.

VIII Conclusions

The traditional exclusive satellite frequency allocation approach is leading to spectrum scarcity, which creates barriers for new Satellite Communication (SatCom)operators while limiting the capabilities of the existing service providers. Intelligent spectrum management approaches enabled by Cognitive Satellite (CogSat)provide feasible solutions to this upcoming issue, and Artificial Intelligence (AI)/ Machine Learning (ML)with superior decision-making and classification capabilities is identified as a key enabler of Cognitive Satellite (CogSat)systems. In this paper, we highlighted the unique characteristics of existing satellite systems and their roles in facilitating global communication through inter-satellite and satellite-terrestrial integrations. Furthermore, this paper extensively evaluated Cognitive Radio (CR)-enabled dynamic spectrum management approaches for SatCom in the context of Opportunistic Spectrum Access (OSA)and Concurrent Spectrum Access (CSA)and discusses the state-of-the-art ML approaches leveraged in Sectrum Sensing (SS), allocation, interference mitigation, and resource management for CogSat networks. However, the deployment of these CR methodologies in satellite networks enabling CogSat remains a complex challenge, due to the barriers caused by regulatory bodies and the limitations in network architecture and standardization. Moreover, leveraging ML in CogSat has challenges due to the computational and power constraints in modern satellite systems. This paper discusses these issues and details ML as a promising solution for scalable, efficient, and secure enabler of CogSat systems toward sustainable future satellite networks as a solution for global connectivity.

ACKNOWLEDGMENT

The authors acknowledge the guidance from Dr. Z. Krusevac for this work, who is with the Defence Science and Technology Group (DSTG), Australia.