SAGIN: Space-Air-Ground Integrated Networks

• SAGIN is a hierarchical integrated network combining space, air, and ground segments to enable global connectivity and heterogeneous service delivery.
• It leverages satellites for wide-area coverage, aerial platforms for flexible relaying, and terrestrial systems for high-throughput connectivity and advanced computation.
• Recent research in SAGIN explores joint optimization of communication, computation, caching, semantics, and AI-enabled orchestration to enhance overall network performance.

Searching arXiv for recent SAGIN papers to ground the article. Space-Air-Ground Integrated Networks (SAGINs) are integrated communication systems that combine space, air, and ground segments into a unified networking architecture for wide-area or global connectivity, heterogeneous service delivery, and cross-layer coordination. In the literature considered here, SAGINs are consistently described as multi-layer systems in which satellites provide wide-area or global coverage, aerial platforms provide flexible relaying and regional continuity, and terrestrial infrastructure provides dense access, rich computation, and service delivery. Across recent work, SAGINs are treated not only as communication fabrics but also as substrates for edge intelligence, semantic communication, federated learning, integrated sensing and control, and even quantum-enabled optical networking (Kato et al., 2018, Liu et al., 30 Apr 2025, Xu et al., 2024, Han et al., 2024, Trinh et al., 2 Mar 2025).

1. Definition, scope, and architectural role

SAGINs are defined as hierarchical integrated networks composed of three segments: the space segment, the air segment, and the ground segment. One formulation describes them as “hierarchical integrated networks” formed by multi-layer satellites, aerial platforms such as UAVs, balloons, and airships, and terrestrial communication and edge infrastructures such as 5G base stations and roadside units (Yu et al., 2021). Another explicitly models three spatial layers as satellites on the space layer, aerial vehicles on the aerial layer, and ground devices on the ground layer (Liu et al., 30 Apr 2025). A broader systems-oriented treatment characterizes SAGINs as a layered architecture in which packets may traverse multiple heterogeneous end-to-end paths involving base stations on the ground, UAVs in the air, and satellites in space (Kato et al., 2018).

The recurring motivation is that terrestrial infrastructure alone cannot provide uniform coverage, especially in deserts, isolated islands, mountains, oceans, rural regions, and disaster-stricken areas. This motivates the assignment of differentiated functions to each segment. The ground segment is repeatedly described as offering the richest computation resources and highest-throughput connectivity where infrastructure exists, while the air segment supplies flexible, rapidly deployable relaying and local augmentation, and the space segment provides ubiquitous or global coverage (Yu et al., 2021, Kato et al., 2018). This division of labor is central to the SAGIN concept: nearby terrestrial or aerial nodes provide locality and responsiveness, and satellites extend reach beyond terrestrial coverage.

Several papers extend this baseline architecture. One generalized model to “space-air-ground-sea integrated networks” adds maritime surface gateways, buoys, ships, and underwater optical/acoustic modems, explicitly treating sea-surface and underwater domains as first-class strata (Yang et al., 2 Sep 2025). Another reframes SAGIN as a “geometry-control hierarchy,” with LEO providing macro-scale coverage and coordination, HAPS providing regional continuity and backhaul support, and UAV and terrestrial layers performing micro-scale adaptation through movable antennas (Liu et al., 20 Apr 2026). A plausible implication is that SAGIN has evolved from a connectivity-centric abstraction toward a broader systems framework encompassing coordination, semantics, learning, and geometry-aware control.

2. Layered system models and cross-layer topology

A central technical issue in SAGIN research is that it is not a single link model but a collection of heterogeneous cross-layer communication scenarios. One unified geometric treatment formalizes six such scenarios: three uplinks—ground-to-air, air-to-space, ground-to-space—and three downlinks—air-to-ground, space-to-air, and space-to-ground (Liu et al., 30 Apr 2025). In that framework, the ground, air, and space layers are modeled as spheres of radii $R_e$, $R_e+H_{\mathrm a}$, and $R_e+H_{\mathrm s}$, respectively, and the receiver’s coverage region on the transmitter layer is represented as a spherical dome. The common coverage-area expression is

$$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$

where the scenario dependence is absorbed into $R_t$, $R_r$, and $\delta_{i,j}$ (Liu et al., 30 Apr 2025). This provides a unified coverage geometry for simulation, connectivity analysis, and node deployment generation across all six cross-layer cases.

Other work models SAGIN as a multi-tier association problem. A four-layer formulation with ground users, UAVs, HAPs, and satellites defines the end-to-end user rate as the bottleneck along the relay chain,

$$R^{\text{e2e}} = \min \{ R_{u\!-\!\text{UAV}},\; R_{\text{UAV}\!-\!\text{HAP}},\; R_{\text{HAP}\!-\!\text{SAT}} \},$$

and then casts cross-layer association as a multi-sided many-to-one transferable-utility matching game (Saliah et al., 2021). In that model, downstream nodes connect to at most one upstream node, while upstream nodes serve multiple downstream nodes up to quota. The significance is not only algorithmic; it formalizes a persistent SAGIN theme that local association choices are inadequate when end-to-end performance depends on a chain of layer-by-layer couplings.

The geometry and mobility of the non-terrestrial segments also enter explicitly into system constraints. For example, one LEO-relay model defines the maximal communication duration with a mobile user as

$T_0^S = \frac{L}{v^S},$

with $L = 2\theta^g(E+l)$, where $R_e+H_{\mathrm a}$0 is Earth radius and $R_e+H_{\mathrm a}$1 is orbital altitude (Xu et al., 2024). Another LEO-UAV model derives satellite access duration from

$R_e+H_{\mathrm a}$2

where $R_e+H_{\mathrm a}$3 depends on Earth radius, orbital altitude, and minimum elevation angle (Huang et al., 2024). These formulations make clear that intermittent contact time is not a peripheral implementation issue but a defining structural constraint of SAGIN.

3. Joint communication, computation, caching, and semantic resource design

A major direction in recent SAGIN work is the shift from pure communication modeling to integrated resource orchestration. In an edge-intelligence formulation for LLM-agent provisioning, the system contains one or more LEO satellites, multiple ground base stations with edge servers, and cloud data centers connected via backhaul. The operator set is

$R_e+H_{\mathrm a}$4

where operator $R_e+H_{\mathrm a}$5 is the LEO satellite and operators $R_e+H_{\mathrm a}$6 are ground BSs (Xu et al., 2024). Users request LLM agent services indexed by $R_e+H_{\mathrm a}$7, backed by models $R_e+H_{\mathrm a}$8, and requests are represented by

$R_e+H_{\mathrm a}$9

This model is significant because it elevates the running LLM instance itself to a managed resource.

That paper introduces “cached model-as-a-resource,” which treats an executable LLM instance and its useful inference state as part of the edge provisioning problem (Xu et al., 2024). Caching is represented by

$R_e+H_{\mathrm s}$0

and local execution proportion by

$R_e+H_{\mathrm s}$1

The framework imposes GPU memory, compute, and context-window constraints such as

$R_e+H_{\mathrm s}$2

and

$R_e+H_{\mathrm s}$3

Its distinctive freshness metric is Age of Thought (AoT), defined through

$R_e+H_{\mathrm s}$4

where $R_e+H_{\mathrm s}$5 is the batch of locally executed request tokens. AoT is used because standard cache metrics such as recency and frequency ignore the semantic statefulness of in-context LLM inference (Xu et al., 2024). A least-AoT replacement algorithm and a DQMSB auction are then used to optimize caching and operator allocation, with the auction reported to “enhance allocation efficiency by 23\% while guaranteeing strategy-proofness and free from adverse selection” (Xu et al., 2024).

A different line of work studies probabilistic semantic communication (PSC) in a satellite-UAV-ground relay SAGIN (Zhao et al., 2024). There the semantic compression ratio for user $R_e+H_{\mathrm s}$6 is

$R_e+H_{\mathrm s}$7

and stronger compression reduces communication load at the expense of increased semantic inference overhead. That overhead is modeled as a piecewise function,

$R_e+H_{\mathrm s}$8

with $R_e+H_{\mathrm s}$9 and $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$0 (Zhao et al., 2024). The corresponding network optimization minimizes total communication and computation energy under latency, power, computation, bandwidth, semantic compression, and UAV-location constraints. This suggests a broader principle: in SAGINs, semantics introduce an explicit communication-computation tradeoff that is especially valuable when the space segment is the bottleneck.

A related semantic design for image delivery proposes hybrid bit-level and generative semantic communication over a LEO-UAV-ground network (Huang et al., 2024). It defines semantic communication efficiency (SCC) as

$$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$1

where $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$2 is reconstruction quality and $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$3 is end-to-end delay. The optimization jointly selects satellite-UAV pairing, UAV trajectory, UAV-user pairing, inter-satellite forwarding, and transmission mode: $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$4 subject to access, motion, delay, and quality constraints (Huang et al., 2024). The use of SCC indicates that SAGIN resource allocation is increasingly formulated around task-level utility rather than channel-level throughput alone.

4. AI-native orchestration, semantic intelligence, and learning over SAGINs

AI-native control has become a prominent organizing principle for SAGIN research. An early articulation argues that SAGIN optimization is harder than in terrestrial networks because of heterogeneity, self-organization, and time variability, and proposes AI as a way to map observed network state to routing and resource-control decisions (Kato et al., 2018). In the paper’s traffic-balancing example, a CNN uses traffic patterns and buffer states to decide path combinations across GEO and MEO satellites, improving throughput and postponing congestion relative to shortest-path routing (Kato et al., 2018). This is one of the earliest explicit claims that AI can serve as a control-plane mechanism for integrated space-air-ground operation.

More recent work moves from narrow prediction to multi-function LLM backbones. One SAGSIN study proposes a single LLM-based adaptation layer spanning radio, optical, and acoustic media (Yang et al., 2 Sep 2025). For long-range channel prediction, it uses a $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$5-parameter LLaMA-3 backbone fine-tuned with LoRA, after a two-stage compression front-end: reference-port selection and separable PCA. The resulting compressed representation uses a rank-$$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$6 basis, retains $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$7 of signal energy, reduces dimensionality to less than $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$8, and produces $$\mathrm{Area}(\mathcal{A}_{i,j})=2\pi R_t^2(1-\delta_{i,j}),$$9 tokens plus one query token (Yang et al., 2 Sep 2025). Trained on 50 past frames to predict 20 future frames, the predictor is reported to follow perfect-CSI capacity within $R_t$0 bit/s/Hz and incur less than $R_t$1 capacity penalty over the practical 5–14 dB SNR range (Yang et al., 2 Sep 2025). The same paper uses LLM-assisted semantic decoding for underwater image communication, claiming more than $R_t$2 dB SNR reduction for high-fidelity delivery (Yang et al., 2 Sep 2025). Although that work is cast as SAGSIN, its methods are explicitly positioned as transferable to standard SAGIN.

Learning is also being embedded directly into edge and distributed training workflows. A federated learning architecture for remote regions without terrestrial base stations treats space-layer satellites and air-layer platforms as both edge computing units and model aggregators (Han et al., 2024). Ground datasets are split into private and offloadable parts,

$R_t$3

with offloadable fraction

$R_t$4

The global objective is

$R_t$5

Its distinctive SAGIN contribution is seamless inter-satellite handover during FL rounds, where a current LEO transfers its model state and unprocessed data to the next incoming satellite over inter-satellite links (Han et al., 2024). The convergence bound explicitly incorporates round-dependent heterogeneity terms $R_t$6 and $R_t$7, reflecting the non-stationarity introduced by adaptive data movement.

Security is now being studied in a similarly AI-native manner. An LLM-based self-evolving defense framework for 6G SAGINs uses LLMs to interpret “massive unstructured multi-dimensional threat information” and generate security strategies (Qin et al., 6 May 2025). Its online component, LLM-6GNG, combines a Condensation Agent, attack correlation, keyword extraction, and multi-agent strategy generation. Its self-evolving component, 6G-INST, filters new instructions using a ROUGE-L threshold below $R_t$8, generates new strategy instructions with GPT-4 plus retrieval, and fine-tunes a twin agent to adapt to unknown attacks (Qin et al., 6 May 2025). The reported average improvement against unknown attacks is $R_t$9 across five attack types. This suggests that in SAGINs, adaptation to non-stationary cross-domain threats is becoming part of the network-control problem rather than an external management function.

5. Spectrum sharing, interference control, geometry, and physical-layer adaptation

SAGIN physical-layer research repeatedly emphasizes that shared-spectrum operation and changing geometry create constraints not well captured by terrestrial abstractions. In a hierarchical cognitive spectrum sharing architecture, the satellite network is treated as the primary network, the aerial network as a preferential secondary network, and the terrestrial network as an ordinary secondary network (Zhou et al., 2023). The joint beamforming problem maximizes terrestrial sum rate while satisfying a satellite interference threshold and an aerial minimum-rate guarantee: $R_r$0 subject to

$R_r$1

and

$R_r$2

The paper’s main architectural point is that spectrum sharing in SAGIN should be hierarchical rather than flat, because the aerial layer may require explicit QoS protection even while remaining secondary to the satellite layer (Zhou et al., 2023).

Another shared-spectrum study focuses on HAPS satellite uplink and HAPS ground downlink reuse in a RIS-aided SAGIN (Mamillapalli et al., 27 Jan 2026). It formulates a power-minimizing beamforming problem with nulling, SINR, and power-budget constraints: $R_r$3 and solves it using DDPG. The effective downlink channel includes a direct and RIS-reflected component,

$R_r$4

The DRL policy is reported to achieve up to $R_r$5 throughput improvement for a $R_r$6 RIS configuration over ZF beamforming (Mamillapalli et al., 27 Jan 2026). Here, the SAGIN-specific issue is cross-tier interference caused by HAPS uplink antenna back-lobes under shared-frequency operation.

A more conceptual but influential direction reframes SAGIN performance bottlenecks around spatial geometry rather than only bandwidth or power (Liu et al., 20 Apr 2026). That work defines spatial geometry as “the relative positions, orientations, and blockage relationships among communication nodes, obstacles, and sensing targets,” and argues that future dense low-altitude operations will be constrained by the ability to maintain favorable geometry in real time (Liu et al., 20 Apr 2026). Movable antennas are proposed as a local spatial degree of freedom, especially valuable at UAV and terrestrial layers. The paper contains no formal optimization problem, but its architectural principle of “scale-aware geometry adaptation” suggests that upper layers should provide macro-scale coordination while lower layers handle micro-scale geometry refinement (Liu et al., 20 Apr 2026).

Optical wireless and quantum extensions push this geometric dependence further. A perspective on optical reconfigurable intelligent surfaces proposes rooftop ORISs as a new ground-side infrastructure element for optical SAGINs, particularly when SAGIN evolves toward quantum networking (Trinh et al., 2 Mar 2025). ORISs are proposed to reflect and reshape free-space optical beams between drones, HAPs, airplanes, and LEO satellites, mitigating line-of-sight blockage, beam broadening, and geometrical and misalignment loss. The reported scenarios include HAP–ORIS–drone QKD and LEO–ORIS–drone quantum links (Trinh et al., 2 Mar 2025). This suggests that future SAGIN physical-layer design may increasingly depend on controllable propagation surfaces and beam-geometry adaptation across multiple media.

6. Twinning, open problems, and research directions

Recent work increasingly treats SAGIN not only as a communication network but as a closed-loop cyber-physical system. A SAGSIN twinning survey characterizes the system by a perception-communication-computing-actuation (PCCA) loop and argues that conventional digital twins are fundamentally mismatched to integrated space-air-ground(-sea) environments because of computational scarcity, synchronization delay, and cross-system semantic gaps (Qiu et al., 18 Dec 2025). Its proposed alternative, the Goal-Oriented Semantic Twin (GOST), is defined as a framework that “constructs task-specific digital models leveraging a unified knowledge base and, through the mining of semantic features and intelligent reasoning, achieves real-time state reflection and control command generation” (Qiu et al., 18 Dec 2025). The framework is organized into knowledge-based semantics, data-driven semantics, and goal-oriented principles, with evaluation dimensions including infrastructure foundation, data-to-model transformation, model quality, synchronization capability, and application utility.

Across the broader SAGIN literature considered here, several unresolved issues recur. One is that the “air” component is sometimes weakly modeled even when the label SAGIN is used; for example, one LLM-agent provisioning framework is described in the details as “primarily space-ground-cloud” despite the SAGIN label (Xu et al., 2024). Another is that several formulations contain typographical inconsistencies, especially in optimization and reward expressions (Xu et al., 2024, Huang et al., 2024, Mamillapalli et al., 27 Jan 2026). These are textual limitations, but they also indicate that some of the field’s abstractions remain in flux.

Substantive open directions are nevertheless clear. The SAGSIN LLM adaptation paper explicitly identifies on-device model compression, multimodal fidelity control, cross-layer resource orchestration, and trustworthy operation as major research directions (Yang et al., 2 Sep 2025). The geometry-aware vision article emphasizes platform-specific movable-antenna co-design, multi-timescale orchestration, protocol-visible MA abstractions, scalable channel awareness, access-backhaul coupling, and communication-sensing convergence (Liu et al., 20 Apr 2026). The federated learning work points to trajectory optimization of air nodes and richer upper-layer infrastructures such as GEO integration (Han et al., 2024). The semantic PSC study explicitly identifies multiple satellites, multiple UAVs, dynamic satellite motion, and UAV trajectory optimization as future directions (Zhao et al., 2024). The security framework implies a need for broader attack coverage, stronger systems metrics, and more realistic cross-domain evaluations (Qin et al., 6 May 2025).

Taken together, these works suggest that SAGIN research is moving toward a unifying view in which communication, computation, storage, semantics, geometry, and security are co-optimized rather than treated as separate subproblems. This suggests a shift from infrastructure integration alone to integrated service orchestration under severe heterogeneity, intermittent connectivity, and task-dependent utility.