SAGIN: Space-Air-Ground Integrated Networks

• Space-Air-Ground Integrated Networks (SAGIN) are heterogeneous architectures combining satellites, aerial platforms, and terrestrial systems to provide full, three-dimensional coverage.
• They employ unified interfaces, geometry-aware models, and dynamic resource orchestration to enhance global connectivity and resilient 6G services.
• Research in SAGIN focuses on quantum-secured communication, intelligent control via cybertwin, and cross-layer service optimization for diverse applications.

Space-Air-Ground-Integrated Networks (SAGIN) denote a heterogeneous integrated communication architecture that combines satellite systems, airborne networks, and ground networks into one coordinated system rather than a loose interconnection of heterogeneous networks. In the 6G literature, SAGIN is presented as a means to provide full network coverage, ubiquitous services, resilient communication, and seamless global coverage across terrestrial and non-terrestrial domains. The canonical three-layer description consists of satellites in the space layer, aerial vehicles or platforms in the air layer, and terrestrial devices or infrastructure in the ground layer; recent modeling work further formalizes six unique cross-spatial-layer transmission scenarios spanning three uplinks and three downlinks across these layers (Chen et al., 2023, Yin et al., 2022, Liu et al., 30 Apr 2025).

1. Architectural definition and segment roles

The architectural premise of SAGIN is the integration of complementary segments whose physical capabilities are markedly different. The space segment includes satellites in GEO, MEO, and LEO orbits and associated ground infrastructure such as control centers and measurement/control stations. The air segment includes high-altitude and low-altitude aerial platforms, including airships, balloons, aircraft, HAPs, and especially UAVs. The ground segment includes terrestrial cellular networks, MANETs, WLANs, base stations, servers, vehicles, and edge/cloud resources. The purpose of SAGIN is to make these layers interoperable through unified radio interfaces, unified network architecture, and unified intelligent control (Chen et al., 2023).

Several papers emphasize that the importance of SAGIN is partly a coverage argument. One 6G-oriented treatment states that terrestrial systems currently cover only about 20% of land area and less than 6% of the Earth’s surface, and therefore cannot alone realize “true global, three-dimensional, seamless coverage” across land, sea, air, and space. The same work organizes a SAGIN-assisted 6G architecture into a Basic Wireless Layer, an Integrated Network Layer, and an Exploration Layer, with functions ranging from waveform and synchronization support to unified interfaces, prediction-based handover, collaborative transmission, and satellite-ground frequency resource sharing (Lan, 25 Oct 2025).

At a finer granularity, the three-layer abstraction is often instantiated geometrically. In one unified model, the space layer contains satellites at altitude $H_s$, the aerial layer contains aerial vehicles or platforms at altitude $H_a$, and the ground layer contains terrestrial devices on the Earth surface approximated as a sphere of radius $R_e$. That model defines six cross-layer scenarios: Ground-to-Air, Air-to-Space, Ground-to-Space, Air-to-Ground, Space-to-Air, and Space-to-Ground. The explicit treatment of all six cases is significant because transmitter/receiver placement, antenna structure, and visibility rules differ across them, yet the paper shows that they can be represented within one spherical-geometry abstraction (Liu et al., 30 Apr 2025).

The literature also cautions against a common reduction of SAGIN to “coverage extension.” One survey states that SAGIN is intended not only for remote access but also for wide-area time-sensitive connectivity, wide-area large-scale connectivity, and wide-area high-precision positioning. Another work on cybertwin-enabled SAGIN similarly argues that SAGIN should be seen as a cyber-physical, intelligence-driven, service-oriented integrated network rather than merely a transport overlay (Chen et al., 2023, Yin et al., 2022).

2. Geometric, channel, and propagation foundations

A recurring methodological theme is that SAGIN cannot be modeled accurately with flat-Earth or purely planar abstractions. In the unified six-scenario formulation, the Earth is approximated as a sphere with center $O$ and radius $R_e$, using spherical coordinates

$(r,\vartheta,\varphi),$

where $r$ is radial distance from $O$, $\vartheta$ is azimuth, and $\varphi$ is polar angle. For a transmitter at altitude $H_a$0 and a receiver at altitude $H_a$1,

$H_a$2

All transmitters at the same altitude lie on a spherical surface $H_a$3, and the receiver’s observation region on that surface is modeled as a spherical dome. The area of the resulting spherical-cap region is

$H_a$4

which becomes the common coverage formula for all six cross-layer scenarios once the vertex angle $H_a$5 is derived from beamwidth or elevation constraints (Liu et al., 30 Apr 2025).

This geometric treatment is complemented by more detailed propagation models. A practical satellite-ground channel model for SAGIN includes five explicit effects: Shadowed-Rician small-scale fading, path loss with bending rays due to atmospheric refraction, molecular absorption modeled by the Beer-Lambert law, Doppler including the Earth’s rotation, and weather attenuation according to ITU-R. The received signal is modeled as

$H_a$6

with the corresponding instantaneous SNR

$H_a$7

Within that model, the Shadowed-Rician parameters $H_a$8 and $H_a$9 separately capture the line-of-sight to scattering ratio and LoS shadowing severity. The numerical conclusions are that atmospheric refraction has a modest effect on path loss, Earth curvature must be considered particularly at small elevation angles, high-frequency carriers suffer from substantial path loss, and Goodput is especially suitable for characterizing coding, modulation, and Doppler-estimation effects (Zhang et al., 2024).

A further propagation line concerns THz joint communication and radar sensing in SAGIN. In that setting, non-terrestrial THz channels are frequently LoS-dominant and sparsely scattered; the paper states that for UAV channels the K-factor is greater than 50 at altitudes above 50 m. It also states that water vapor is by far the strongest absorber, with an absorption cross-section about six orders of magnitude larger than other air components, while rain is in the Mie regime and fog is much less harmful to THz than to FSO in many cases. For a 2000-km ground-to-satellite link, the reported free-space path loss values are about 184 dB for mmWave, 208 dB for THz, and 264 dB for FSO, yet the paper concludes that THz achieves the highest capacity among mmWave, THz, and FSO in the evaluated clear-sky, rain, and fog scenarios (Han et al., 25 Feb 2025).

These results collectively indicate that geometry is not a secondary detail in SAGIN. Coverage, path length, atmospheric traversal, visibility, Doppler, and spectral choice all couple directly to network-layer feasibility.

3. Virtualization, mission offloading, and resource orchestration

A major systems line treats SAGIN as a cooperative 3D resource pool whose communication, computation, storage, and sensing resources should be coordinated rather than statically partitioned. An influential formulation is bi-directional mission offloading, which rejects the assumption that offloading is only ground-to-space or ground-to-air. It argues simultaneously for Ground $R_e$0 Space/Air offloading to exploit coverage, mobility support, robustness, and backhauling capability, and Space/Air $R_e$1 Ground offloading to exploit richer spectrum, energy, computing, and storage on the ground. In that framework, network reconfiguration is the key enabler, and NFV with SFC provides the mechanism by which missions are decomposed into virtual network functions, chained, placed, and migrated across heterogeneous nodes (Zhou et al., 2019).

The SFC viewpoint is developed further through time-varying graph models. One work proposes a reconfigurable time expansion graph (RTEG),

$R_e$2

with nodes

$R_e$3

and links including G2U, U2U, U2S, S2S, S2G, and a storage link across time,

$R_e$4

This converts dynamic SAGIN resource allocation into time-indexed SFC deployment. The resulting optimization is a MILP that maximizes the number of successfully completed tasks, and the paper transforms it into a many-to-one two-sided matching game solved with a Gale-Shapley-based algorithm called MG-RTEG. In simulations with 6 UAVs and 1 satellite and 300 random tasks, MG-RTEG completed more tasks than AASO and FCFS in every time slot (Cao et al., 2023).

Related work addresses failure recovery. In an NFV-based SFC recovery model, failures of UAV or satellite nodes and their incident links trigger redeployment of affected VNFs in a time-expanded graph. The optimization objective is to minimize the total time consumption of all completed SFCs, and the proposed matching-game-based recovery algorithm FRMG-SAGIN reportedly optimizes total time consumption by about 25% compared with benchmark methods (Jia et al., 4 Feb 2025).

Resource orchestration also appears at the radio layer. An OFDMA-based UAV-assisted SAGIN resource-allocation study formulates a MINLP over subcarrier assignment $R_e$5 and power $R_e$6, with

$R_e$7

It compares alternating optimization (AO), damped iterative water filling (DIWF), and genetic algorithms (GA) over 10,000 Monte Carlo trials. The reported outcomes are that AO converged in about 2 to 11 iterations, DIWF converged smoothly with a final average sum rate around 34.56–34.575 Mbps, and GA achieved the best reported throughput with best fitness 36.55 Mbps and average 35.04 Mbps (Geddam et al., 16 Sep 2025).

Together, these studies depict orchestration in SAGIN as intrinsically cross-domain: VNFs, compute, storage, bandwidth, and backhaul are all jointly involved, and dynamic matching, reconfiguration, and service recovery are recurrent solution patterns.

4. Intelligent control, learning, and geometry-aware adaptation

The move from static control to intelligent control is explicit in cybertwin-based proposals. A cybertwin-enabled 6G SAGIN architecture introduces a five-dimension digital twin model consisting of Sat-DT, UAV-DT, BS-DT, Agent-DT, and Inter-DT connections. The physical SAGIN collects data, while the cybertwin space at core and edge clouds stores, analyzes, learns, and returns control decisions. The paper organizes its technical agenda into cybertwin-based multi-source heterogeneous network integration, integrated cloud-edge-end, and integrated sensing-communication-computing, and treats the cybertwin as the “brain” of SAGIN (Yin et al., 2022).

A distinct but related line argues that SAGIN control is increasingly limited by the inability to adapt to rapidly changing spatial geometry. In a geometry-aware architecture for the low-altitude economy, LEO provides macro-scale coverage and global coordination, HAPS provides regional persistence and backhaul support, and UAV plus terrestrial layers carry the most dynamic local geometry. The proposed mechanism for fine-grained adaptation is the movable antenna (MA), defined as controlled displacement and/or orientation adjustment of antenna elements within a compact local region, typically on the order of several to tens of wavelengths. The paper explicitly distinguishes MA from beamforming, platform mobility, and RIS, and argues that MA enables real-time geometry control at the link level for blockage recovery, beam misalignment compensation, interference suppression, and improved sensing geometry (Liu et al., 20 Apr 2026).

Learning-based control over UAV-assisted SAGIN has been reviewed from a methodological perspective in terms of Q-learning, DQN, MAB, PSO, and satisfaction-based learning. A representative reward is

$R_e$8

combining fairness and load. Under a simulation setup with 22 LEO satellites, 4 SBSs, 4 UAVs in some experiments, satellite altitude 550 km, UAV max altitude 121.9 m, and 100 Monte Carlo runs, the paper reports that 3D satisfaction-CA outperforms the compared methods in most cases on outage users, load balancing, user rate, fairness, and reward, while DQN-CA and PSO require more simulation time (Arani et al., 2022).

Other intelligent-control studies are more task specific. A deep-learning-aided multi-objective routing method for maritime SAGIN uses real Iridium-NEXT-like satellite trajectories, real North Atlantic flight traces, and shipping data, and formulates routing through delay, throughput, and path-lifetime constraints. The paper reports that integrated AANET + LEO provides 100% coverage over the full day and yields up to 200% gain in maximum E2E throughput on average and up to 200% gain in maximum path-lifetime on average, while the learned routing policy achieves near Pareto-optimal performance using only local information (Liu et al., 2021). In task scheduling, CMADDPG combines dynamic UAV clustering with centralized-training/distributed-execution MADDPG, reducing the number of active decision-makers to cluster heads; simulations report at least a 25% improvement in system profit over MADDPG and 42.86% improvement over MAAC (Wang et al., 2024).

Generative AI has also entered the SAGIN literature. One survey-and-framework paper reviews VAEs, GANs, GDMs, and transformer-based models for channel modeling, CSI estimation, resource allocation, deployment, semantic communications, image processing, and security. Its concrete proposal is a GDM-based channel information map pipeline with image acquisition, image processing, elimination of redundant information, zero-shot classification, and channel information construction. In the reported case study, RMSE decreases from 109.32 to 0.58, the GDM classifier outperforms KNN, XGBoost, and CNN in accuracy, F1-score, and recall, and a downstream PPO strategy achieves higher cumulative reward using the constructed map (Zhang et al., 2023).

5. Privacy, security, and quantum-secured SAGIN

Security work on SAGIN begins from the observation that classical key-distribution assumptions become problematic in heterogeneous long-range networks and in the presence of quantum computers. One line therefore proposes quantum-secured SAGIN, integrating fiber-based QKD in the terrestrial layer with satellite-based QKD and UAV-based QKD as dynamic supplements. A universal QKD service provisioning framework is managed by a QKD global manager that reserves optical-fiber QKD services in advance and supplements them with satellite/UAV QKD during real-time transmission if the reserved keys are insufficient. The optimization is posed as a two-stage stochastic program minimizing reservation cost plus expected free-space supplementation cost under uncertain secure-communication demand (Xu et al., 2022).

A closely related resource-allocation treatment formulates QKD over SAGIN across optical fiber, UAV, and satellite layers with uncertainty in secret-key-rate requirements and weather conditions. It represents uncertainty by

$R_e$9

where $O$0 is the random secret-key-rate requirement and $O$1 is the random weather condition, and minimizes total deployment cost through reservation, utilization, and on-demand decisions. The paper reports that an optimal point is found around reserved QKD wavelength 80, at which the second-stage cost becomes 0, and states that on-demand costs are set to 2× reservation costs in experiments (Kaewpuang et al., 2022).

Privacy-preserving intelligence is treated separately through federated learning. The cybertwin-enabled SAGIN paper identifies AI-based networking decision and optimization and FL-based cross-layer privacy/security as open issues, arguing that FL alone is insufficient because wireless model transmission remains vulnerable to eavesdropping and therefore must be combined with secure physical-layer transmission (Yin et al., 2022). A later vision paper extends this trajectory from FL to QFL, positioning SAGIN as a distributed, heterogeneous, resource-constrained, and privacy-sensitive environment in which local training at satellites, UAVs, and ground devices is preferable to centralized data collection. In a UAV-network case study, QFL uses a 4-qubit circuit with RX, RY, RZ, and CRX gates, Adam, learning rate 0.001, 50 epochs, batch size 64, and NLL loss on MNIST, and the reported result is faster convergence than conventional FL (Quy et al., 2024).

Quantum communication work also extends beyond key distribution. An ORIS-based quantum SAGIN paper proposes optical reconfigurable intelligent surfaces on building rooftops to establish blocked optical links and control beam diameter. It studies four scenarios—Drone–ORIS–Drone, HAP–ORIS–Drone, Drone–ORIS–Airplane, and LEO Satellite–ORIS–Drone—and analyzes fidelity, secret-key rate, and entanglement rate. In the HAP-to-drone QKD scenario, the best SKR occurs around an ORIS-drone distance of about 339 m, with an optimal beam diameter of 73 cm in the examined setup (Trinh et al., 2 Mar 2025).

Taken together, these works frame secure SAGIN as more than encrypted transport. It is a multi-layer key-distribution, privacy-preserving learning, and in some cases quantum-networking problem.

6. Representative applications, performance patterns, and open problems

The application literature shows that SAGIN is not confined to a single sector. Maritime communications are a prominent case: integrating LEO satellites, passenger airplanes, ships, and on-shore base stations improves coverage quality and creates new routing tradeoffs between delay, throughput, and path lifetime. The in-flight connectivity literature provides another variant, in which satellites are treated not merely as transparent relays but as caching nodes connected by ISLs. In one IFC-oriented framework, satellites and GSs cooperatively deliver cached and non-cached files to aircraft, and an exact penalty method for cached files performs nearly as well as exhaustive search, while increasing the maximum number of ISLs significantly reduces average delay before saturation (Liu et al., 2021, Chen et al., 2024).

Emergency communications, overload relief, and low-altitude services recur across several papers. The user-association literature explicitly analyzes single-TBS failure, multiple-TBS outage, and public-event congestion, using proximity-based, SINR-based, priority-based, and distance-aware/capacity-aware hybrid strategies. The low-altitude economy literature, by contrast, emphasizes that communication, sensing, control, and navigation are geometry-dependent and that the main bottleneck is the mismatch between multi-scale network control and fast local geometry variation (Geddam et al., 16 Sep 2025, Liu et al., 20 Apr 2026).

The 6G architecture literature identifies additional system-level tensions. One paper argues that SAGIN congestion is exacerbated by conventional IP routing because mega-constellations contain many equal-cost or near-equal-cost paths that standard routing underutilizes. Its proposed multi-routing-plane based Flow Scheduling Strategy (MFSS) uses an auxiliary plane to exploit Equivalent Paths (EPs) after congestion prediction. In an evaluation over Starlink with 1584 satellites, 18 cities, and $O$2 MB as written in the paper, the reported result is that over 40% of end-to-end connections exceed 200 MB, competing methods stay below 20% for that threshold, and MFSS eliminates cases below 125 MB (Lan, 25 Oct 2025).

Open problems remain consistent across surveys and technical papers. The survey literature emphasizes dynamic nodes, high-speed topology changes, huge time and space span, heterogeneous interconnection, limited on-board resources, and security. The THz JCRS literature adds THz-specific CRB-rate tradeoffs, directional MAC/routing, hardware robustness, and space-debris detection. Generative-AI and cybertwin papers stress real-time responsiveness, integration overhead, scalability, and cross-layer privacy/security. A plausible implication is that future SAGIN research will continue to move away from single-layer optimization and toward joint treatment of geometry, communication, sensing, computing, and security as co-evolving design variables (Chen et al., 2023, Han et al., 25 Feb 2025, Zhang et al., 2023, Yin et al., 2022).

A final misconception addressed repeatedly in the literature is that SAGIN can be understood by simply appending satellites or UAVs to terrestrial systems. The surveyed work instead presents SAGIN as an integrated, hierarchical, and cross-layer system whose defining properties arise precisely from the interaction of orbital geometry, air mobility, terrestrial infrastructure, heterogeneous resources, and multi-domain control.