Satellite-Terrestrial Edge Computing Networks

• Satellite-Terrestrial Edge Computing Networks are hybrid architectures integrating LEO satellites, terrestrial base stations, and cloud data centers for on-site computation, caching, and service provisioning.
• They employ layered, multi-tier execution models and innovative service abstractions such as serverless computing and VNF placement to optimize performance under strict delay, energy, and security constraints.
• Advanced methods including ILP, LP, deep reinforcement learning, and heuristics address complex control problems and resource trade-offs in dynamic three-dimensional networks.

Searching arXiv for recent and foundational papers on Satellite-Terrestrial Edge Computing Networks to ground the article. arXiv search query: "Satellite-Terrestrial Edge Computing Networks" Satellite-Terrestrial Edge Computing Networks (STECNs) are hybrid space-ground computing architectures in which low-Earth-orbit (LEO) satellites, terrestrial base stations, edge clouds, gateways, and cloud data centers cooperate as a distributed computing system rather than as a purely communications overlay. Across the literature, the defining shift is that satellites are treated not merely as relays or backhaul nodes, but as computation-, caching-, and service-capable edge elements that can execute tasks, host virtualized functions, support semantic processing, participate in multi-hop routing, and interoperate with terrestrial edge and cloud resources under delay, energy, bandwidth, and security constraints (Soret et al., 2024, Strinati et al., 2020, Huang et al., 2024).

1. Architectural scope and network layering

A recurring architectural pattern in STECN research is a layered system that spans ground, air, and space. In Earth Observation (EO), one representative design has three main blocks: EO satellites that acquire imagery, a LEO constellation that acts as a satellite edge layer, and a ground segment that performs final aggregation, semantic interpretation, and cloud-backed processing when needed. In geo-distributed edge-cloud processing, the corresponding three components are edge clouds, a LEO satellite network acting as backhaul/access transport, and a core cloud or data center. In 6G-oriented work, this expands into a hierarchical three-dimensional network with terrestrial base stations at the bottom, aerial layers such as UAVs and HAPS in the middle, and LEO/GEO satellites at the top (Soret et al., 2024, Zhao et al., 2024, Strinati et al., 2020).

Within these architectures, LEO satellites are repeatedly assigned a dual role: communication infrastructure and edge-computing substrate. The EO literature explicitly treats a LEO constellation as an edge layer that may be separate from EO satellites or overlap with them as multi-purpose spacecraft, with conventional RF, FSO, or hybrid FSO/RF links connecting the space segment and RF feeder links connecting to the ground. In one concrete simulation, the system uses 23 satellites, FSO inter-satellite links, and an RF feeder link. Other works instantiate satellites as cooperative edge clouds alongside one terrestrial BS, as cache nodes serving geographically distributed users, or as VNF hosts for remote IoT service chains (Soret et al., 2024, Huang et al., 2024, Zheng et al., 22 Aug 2025, Gao et al., 2021).

Aerial layers are not universal, but when included they alter STECN design substantially. The AA-MEC literature inserts aircraft as MEC nodes between satellites and terrestrial gateways, yielding a three-layer multi-layer network in which satellites can generate tasks, aircraft can host MEC servers, and gateways remain terrestrial compute endpoints. The 3D-network literature similarly argues that UAVs, HAPS, and satellites should be treated as full network elements providing communication, computation, and caching rather than as passive relays (Mankowski et al., 2022, Strinati et al., 2020).

This architectural diversity suggests that “satellite-terrestrial” in STECNs is best understood as a continuum of execution locations and transport paths rather than a fixed two-node interface. A plausible implication is that the term covers both classic space-ground integration and broader 3D systems whenever the control problem jointly spans non-terrestrial and terrestrial compute resources.

2. Execution models, service abstractions, and data handling

The execution model in STECNs is typically multi-tier. Tasks may be computed locally at the originating device, at a terrestrial BS or gateway MEC server, on a source satellite, on another satellite reached through inter-satellite links, or in a ground cloud. Several papers formalize this as binary or fractional offloading among local, edge, and cloud locations, while others represent it as VNF placement or route-based data-flow allocation (Huang et al., 2024, Wang et al., 2023, Zhu et al., 2021, Liao et al., 20 Feb 2025).

The service abstraction used on top of this execution model varies. In EO systems, the data path is increasingly semantic and goal-oriented: instead of forwarding raw pixels, the system may transmit compressed images for reconstruction, object detections, or bounding boxes for localization and tracking. The communication objective is therefore no longer bit accuracy alone, but task fulfillment. The EO framework uses a shared knowledge base between satellites and the ground station, learned offline at the ground and uploaded as pre-trained models to satellites; it also notes the need for KB alignment and continual learning because the KB can drift over time (Soret et al., 2024).

For general-purpose satellite edge platforms, one line of work evaluates virtual machines, containers, and serverless functions as organization paradigms for the LEO edge and concludes that serverless computing is the most promising abstraction. The argument is that satellites are moving servers, nearly identical hardware units, resource-poor and expensive to scale, non-serviceable after launch, and part of a fixed compute pool over a lifetime usually around 5 years. In that environment, virtual stationarity, service migration, replication, and fine-grained resource control become central, and serverless provides the platform with the most control over placement, handoff, failover, and multi-tenancy (Pfandzelter et al., 2021).

Other STECN service models are function- and content-centric rather than task-centric. VNF placement work represents user requests as service function chains deployed across satellites and cloud, subject to CPU, memory, bandwidth, and delay constraints. Caching work treats LEO satellites as edge cache nodes with local cache state, ISL forwarding, and latency-constrained content delivery. Computation reuse work goes further by storing reusable records of prior tasks in a satellite computation reuse table and combining locality-sensitive hashing with SSIM-based matching to reuse prior results rather than recomputing them (Gao et al., 2021, Zheng et al., 22 Aug 2025, Zhang et al., 15 Mar 2025).

A common misconception is that all STECN workloads are simple task-offloading problems. In fact, the literature spans semantic EO, service-function chaining, cache placement, privacy-preserving sequential offloading, secure association with artificial noise, integrated sensing-communication-computing, and computation reuse. This suggests that STECNs are increasingly defined by orchestration over heterogeneous service abstractions rather than by offloading alone.

3. Optimization and control methodologies

STECN control problems are usually formulated as mixed-integer, non-convex, or NP-hard programs because they couple discrete placement or association decisions with continuous communication and computing resource allocation. Representative objectives include minimizing total system energy under delay and satellite-energy constraints, minimizing access-network transmission duration, minimizing weighted energy subject to task delay deadlines, minimizing delay-energy cost under secrecy constraints, maximizing weighted overall computing capacity under hop and link constraints, and maximizing a joint sensing-communication-computing utility (Huang et al., 2024, Zhao et al., 2024, Martinez-Gost et al., 2021, Zheng et al., 22 Aug 2025, Liao et al., 20 Feb 2025, Zhu et al., 2023).

The solution techniques are correspondingly diverse. Exact or near-exact optimization appears in ILP, LP, SDR/SCA, KKT-based decompositions, Dinkelbach-style fractional programming, and column generation. One access-satellite selection problem is formulated as an ILP whose brute-force complexity grows roughly as $O(m^n)$, motivating the greedy DVA heuristic. Another route-and-compute problem is posed as a route-based LP and solved optimally with column generation, which activates only a subset of route variables and uses a hop-constrained Bellman-Ford pricing subproblem to identify improving columns (Zhao et al., 2024, Liao et al., 20 Feb 2025).

Distributed and decomposition-based methods are also prominent. Sat2C decouples routing from power allocation and edge-versus-cloud processing decisions. An ADMM-inspired method decomposes joint edge-cloud selection and bandwidth allocation into an association subproblem, a convex bandwidth subproblem, and a separable auxiliary-variable update. Alternating optimization is used to coordinate offloading selection, beamforming, power allocation, and CPU-resource allocation in 6G satellite-terrestrial computing (Martinez-Gost et al., 2021, Huang et al., 2024, Wang et al., 2023).

Learning-based control is especially common when the action space is large or the environment is dynamic. PPO is used for privacy-preserving offloading, security-sensitive offloading, and AIGC-assisted digital watermarking. Deep reinforcement learning with order-preserving quantization is used in DRTO so that offloading location and bandwidth allocation depend only on the current channel states. Soft actor-critic with GNN encoders is used for joint caching and routing over dynamic LEO graphs, and a hybrid quantum DDQN is proposed for discrete edge-cloud selection when classical and quantum neural networks process information in parallel (Lan et al., 2023, Lan et al., 2024, Chen et al., 2024, Zhu et al., 2021, Zheng et al., 22 Aug 2025, Huang et al., 2024).

The methodological picture is therefore not one of convergence on a single solver family. Rather, STECN research treats orchestration as a family of coupled combinatorial-control problems whose tractable solution depends on whether the dominant difficulty is route explosion, bandwidth bottlenecks, large discrete action spaces, or rapid topology variation.

4. Representative applications and workload classes

EO is one of the canonical STECN applications. The EO semantic framework supports image reconstruction, object detection and localization, and closed-loop object tracking. Its motivation is that feeder-link congestion, cloud-delayed delivery, and intermittent connectivity can make conventional store-and-forward architectures unsuitable for real-time or mission-critical tasks. The paper explicitly notes that, in sparse constellations, satellite processing can reduce wait times from “a few hours” to “a few tens of milliseconds,” while dense constellations with inter-satellite links can operate as a distributed edge-computing fabric (Soret et al., 2024).

Large-scale geo-distributed data processing is another major class. Here the edge sources are geographically distributed edge clouds rather than satellites or handsets, and the LEO constellation serves as the access network to a core cloud for time-sensitive big-data processing such as attack detection, scientific data aggregation, and dynamic content delivery. The key observation is that the access network is the bottleneck: approximate magnitudes cited are about 0.5 Gb/s for user access bandwidth, about 2.5 Gb/s for ground station access bandwidth, and about 20 Gb/s for inter-satellite links (Zhao et al., 2024).

6G-oriented integrated systems broaden the application scope further. The literature names interactive 3D video, remote sensing, pollution and weather monitoring, traffic control, agriculture, disaster recovery, remote broadband, maritime and aeronautical connectivity, transportation, and distributed intelligence for mobile IoT. AA-MEC adds two particularly concrete use cases: In-Flight Entertainment and Connectivity Services for aircraft and AI/ML task offloading from satellites when onboard compute is insufficient (Strinati et al., 2020, Mankowski et al., 2022).

Application-specific intelligent services also appear. One study models AIGC-assisted digital watermarking, in which visible LEO satellites act as mobile edge servers for DCT-, DWT-, or LSB-based watermarking services and inter-satellite migration is used when the processing satellite is not visible. Another paper studies integrated sensing, communication, and computing, where ground IoT devices sense and communicate in a time-division manner and can process the sensing results locally, at the edge, or in the cloud via the satellite communication link (Chen et al., 2024, Zhu et al., 2023).

Security- and privacy-sensitive workloads form a further branch. One paper balances completion time, energy, communication reliability, usage pattern privacy, and location privacy for satellite-assisted offloading; another minimizes delay, energy, and the number of malicious-satellite attacks under reliability constraints; a later secure scheduling framework adds secrecy-aware user association with artificial noise and delay-energy-aware task scheduling (Lan et al., 2023, Lan et al., 2024, Zheng et al., 22 Aug 2025).

5. Performance characteristics and empirically observed trade-offs

A central empirical theme is that STECN performance is governed by coupled trade-offs among accuracy, delay, energy consumption, bandwidth usage, and orchestration latency. In semantic EO, these trade-offs are quantified with task-achievement metrics such as MSE, SSIM, recall, precision, mAP, and IoU; time metrics from image capture to ground reception; and energy metrics determined by onboard processing architecture and transmission technology. The paper’s concrete ship-detection setup uses WorldView-3 parameters, 600 × 600 pixel frames, $GSD = 0.5$ m, YOLOv8 with $W_\text{YOLO}=79.1$ G floating-point operations per image, and a semantic output averaging 336 bits per image for ships and bounding boxes. It reports that low frame rates favor GS CPU processing for power efficiency, whereas higher frame rates require a combination of edge and cloud processing to satisfy timing constraints while minimizing power (Soret et al., 2024).

Access-layer orchestration can dominate end-to-end performance when many edge sources compete for LEO access. In the DVA study, compared with shortest-distance and longest-visible-time baselines, DVA reduces average transmission time by about 49.7% and 48.8%, improves access network throughput by about 2.28× and 2.30×, is only about 8% worse than the ILP optimum in transmission duration, and runs under 1 ms versus about 290 ms for OP/Gurobi (Zhao et al., 2024).

Hybrid edge-cloud cooperation repeatedly outperforms fixed policies. Sat2C, which jointly selects routing, power allocation, and edge-versus-cloud processing for Earth surveillance, achieves energy savings of up to 18% relative to always-edge, always-cloud, or maximum-power baselines. DRTO achieves near-optimal offloading cost relative to exhaustive enumeration while reducing runtime by 42.6% versus DDLO, 87.3% versus coordinate descent, and 96.6% versus enumeration for $N=7$. In route-and-compute optimization, a three-layer protocol improves computed data volume by 40% over local-only processing while reducing optimization scale by up to 92–93% at high hop budgets through column generation (Martinez-Gost et al., 2021, Zhu et al., 2021, Liao et al., 20 Feb 2025).

Performance gains also appear in more specialized STECN functions. AA-MEC reports a 10.43% latency improvement over terrestrial-only MEC for latency-critical applications, at least 6.7% decrease in flow latency for IFECS, and at least 56.03% decrease for computation offloading under dynamic reconfiguration. GT-SAC improves delivery success rate and reduces communication traffic cost, with update traffic about 27.6% lower than SAC and about 62.3% lower than PCF when each LEO caches one content, alongside success-rate improvement about 59.2% over cloud. CCRSat reports task completion time reduction by up to 62.1% and computational resource consumption reduction by up to 28.8% (Mankowski et al., 2022, Zheng et al., 22 Aug 2025, Zhang et al., 15 Mar 2025).

These results do not imply a universal superiority of space-edge execution. Several studies explicitly show path dependence and workload dependence. Edge computing becomes more attractive when compression ratio is high, routes are long, or semantic extraction drastically shrinks payloads; cloud or terrestrial execution can remain preferable when route length is short, local processing energy is high, or terrestrial BS capacity is abundant. That pattern recurs across EO, Earth surveillance, and 6G satellite-terrestrial computing (Soret et al., 2024, Martinez-Gost et al., 2021, Wang et al., 2023).

6. Challenges, misconceptions, and research directions

The first persistent challenge is mobility and topology variation. Satellite visibility is intermittent, satellites move at orbital speed, and in 3D networks even access points may move. The literature therefore repeatedly raises virtual stationarity, service migration, mobility-aware handover, distributed control, reconfigurable optimization over snapshots, and route recomputation under time-varying visibility as core STECN requirements (Pfandzelter et al., 2021, Strinati et al., 2020, Mankowski et al., 2022).

The second challenge is resource scarcity in the space segment. Satellites are constrained in energy budgets, thermal dissipation capacity, mass budget, and launch economics; some models explicitly emphasize limited harvested solar energy. Queue stability is a recurring issue: each satellite must process and transmit data faster than it arrives, or data may need to be discarded or acquisition reduced. This makes hard resource arbitration unavoidable and weakens the common misconception that edge placement is always the latency-optimal and energy-optimal decision (Huang et al., 2024, Soret et al., 2024, Pfandzelter et al., 2021).

A third challenge is that communication, computation, caching, and security are inseparable in practice. Interference, mmWave sensitivity, antenna design for high-elevation users, secrecy-aware association, artificial-noise power splitting, malicious-satellite attacks, usage-pattern privacy, location privacy, and reliability constraints all appear as first-class variables rather than afterthoughts. A plausible implication is that future STECN controllers will need to internalize security and privacy as operational objectives rather than as external safeguards (Strinati et al., 2020, Lan et al., 2024, Lan et al., 2023, Zheng et al., 22 Aug 2025).

A fourth challenge is state and knowledge management. Semantic EO requires a shared knowledge base and continual KB alignment; serverless LEO platforms require externalized state and replicated state services; collaborative reuse and caching require consistency of reuse tables or cache state across moving nodes. The literature does not present a single unified solution to these state-management problems, which remain a major systems question (Soret et al., 2024, Pfandzelter et al., 2021, Zhang et al., 15 Mar 2025, Zheng et al., 22 Aug 2025).

Finally, the literature repeatedly corrects the misconception that satellites in integrated systems are only coverage extenders. In the surveyed works, satellites host MEC servers, execute AI/ML tasks, cache content, route offloaded tasks over ISLs, support VNF chains, provide semantic pre-processing, and participate in distributed optimization loops. The field’s open direction is therefore not merely better satellite access, but better joint orchestration of where to compute, what to transmit, how much to cache, when to cooperate with the cloud, and how to maintain service continuity and security across a dynamic multi-layer network (Soret et al., 2024, Strinati et al., 2020, Wang et al., 2023, Liao et al., 20 Feb 2025).