- The paper introduces a unified S²C²I architecture that co-designs sensing, storage, communication, computing, and intelligence across HAP payloads, connectivity links, and cloud-edge networks.
- The paper’s emergency-response simulation shows integrated orchestration achieving 82.4% conjunctive service availability, compared with 34.7% for communication-only HAPs and 58.6% for function-separated designs, while reducing feeder traffic from 1.17 TB to 0.49 TB.
- The paper identifies deployment barriers including limited feeder capacity, payload energy and thermal constraints, missing HAP-specific standards and telemetry, inadequate end-to-end benchmarks, and the lack of flight-validated evidence for fully integrated platforms.
Motivation and positioning
This survey by Luo and Alouini (KAUST) argues that high-altitude platforms (HAPs) operating at 20–50 km altitude are evolving from communication relays into persistent infrastructures that jointly provide sensing, storage, communication, computing, and intelligence (S²C²I). The authors contend that existing surveys treat these functions in isolation—communication-centric surveys dominate the literature, while function-oriented works address only subsets (e.g., optical communication plus atmospheric sensing)—and that such fragmentation obscures the true system bottlenecks. Their central claim is that S²C²I integration on HAPs is a tightly coupled co-design problem under stringent payload mass, power, spectrum, thermal, and endurance constraints, not an aggregation of five independent capabilities.
The survey's comparative analysis of prior work is explicit: it classifies prior surveys into communication-centric foundations (e.g., Karapantazis and Pavlidou; Widiawan and Tafazolli), NTN-integration studies (Lou et al.; Abbasi et al.; Svistunov et al.), and function-oriented convergence works (Bakambekova et al.; Elkhazraji et al.). It claims to be the first to cover all five S²C²I dimensions with cross-functional coupling as the organizing principle—a claim supported by its own comparison table, in which no prior survey marks all six columns (five functions plus cross-function design).
HAP fundamentals and roles in SAGIN
The survey reviews the two HAP families: lighter-than-air systems (stratospheric airships such as Stratobus, targeting 12-month endurance at 20 km with 250 kg payload; balloons such as Loon) and heavier-than-air systems (solar UAVs such as Zephyr S, holding a 64-day endurance record with ~5 kg payload; hydrogen-powered Phantom Eye). Within the three-tier space-air-ground integrated network (SAGIN), HAPs occupy a middle layer whose quantitative advantages are stated concretely: a single platform covers a 50–200 km radius at 2–10 ms end-to-end latency; HAP-to-ground free-space path loss is roughly 20–30 dB lower than LEO links and 50–60 dB lower than GEO links at equal frequency. These figures position HAPs as filling a genuine capability gap between satellite coverage and short-range terrestrial/aerial infrastructure.
Six functional roles are enumerated: wide-area access node, wireless relay, integrated access-and-backhaul hub, aerial edge service node, low-altitude aerial coordinator, and cross-layer orchestrator. Field evidence cited includes NTT/AALTO trials achieving LTE downlink above 4 Mbps to commercial smartphones via Zephyr over a 20 km stratospheric link—an important reality check on direct-to-device feasibility, though the survey notes public releases omit raw channel traces and energy telemetry.
Unified architecture
The architectural contribution has three components. First, a five-link connectivity taxonomy: service links (sub-6 GHz/mmWave, 10 Mbps–1 Gbps), HAP-UAV links (5–30 km slant range), feeder links (Ka/Ku/E-band/FSO, 1–50 Gbps), inter-HAP links (mmWave/FSO), and HAP-satellite links. Feeder capacity is identified as the principal system bottleneck constraining raw-data offload, cache refresh, and model distribution—establishing why storage and computing must be onboard rather than ground-centric.
Second, a payload architecture progression from transparent (bent-pipe) through regenerative and CU/DU-split designs to a proposed onboard S²C²I converged architecture hosting DU/UPF functions alongside sensing pre-processing, hierarchical storage, edge AI inference, and a cross-function orchestrator. The authors acknowledge this fourth architecture carries the highest complexity, since cross-function coupling introduces joint resource-allocation challenges across all five dimensions; they also concede that full native integration of all five functions remains unexplored in existing primary works.
Third, a cloud-edge-HAP space continuum extending integration to network scale, with a global cloud layer (space cloud plus terrestrial data centers), a stratospheric HAP layer as regional hub, and a ground/low-altitude edge layer, connected by layered data-processing and control/model loops. The design intent is that only high-value features traverse feeder links while closed-loop latency-critical decisions remain onboard.
Enabling technologies
The technology review is organized by function but consistently emphasizes cross-functional coupling:
- Heterogeneous connectivity: RF serves as reliability/control plane; mmWave and sub-THz as high-capacity data planes (with the notable finding that PRACH, not link budget, limits large mmWave footprints); FSO for tens-of-Gbps trunks sensitive to turbulence and clouds; hybrid RF/FSO schemes for availability. The recurring lesson is that medium selection should be coordinated with computing placement and buffering—for example, data buffered during optical blockage and transmitted when conditions recover.
- Sensing and ISAC: radar/SAR payloads with joint trajectory-beamforming co-design, waveform-level SAR-communication integration via frequency-index modulation, MIMO max-min beampattern fairness, sensing-assisted beam alignment reducing inter-HAP search overhead, cooperative HAP-UAV sensing with sub-THz backhaul, and secure full-duplex ISAC with jammer-UAV coordination.
- Onboard computing: HAP-centric edge computing for vehicular/IoT workloads (with a concrete requirement of >3000 GFLOPS for real-time processing at 10 frames/s in one studied setting), hierarchical HAP-UAV computing via DRL decompositions, SAGIN-wide orchestration with demand prediction, and green/emerging directions including data-center-enabled HAPs reporting ~12% electricity-cost savings with one DC-HAP and ~36% with four, and quantum-computing-enabled HAPs showing 21% energy reduction and 30% more qubits than terrestrial equivalents at stratospheric temperatures—the latter remaining speculative given unresolved radiation, link-loss, and error-correction issues.
- Storage and caching: a four-layer hierarchy spanning user-side caches fed by HAP broadcast, HAP fundamental-data libraries for computation-aware storage (tens of TB requirements cited for highway ITS scenarios), hierarchical HAP-UAV caching with federated popularity prediction, and multi-HAP cooperative caching with coded multicast. Cache state is shown to influence even payload operating mode selection in multimode HAP designs.
- AI orchestration: predictive DNNs for association and channel estimation, RL/MARL for distributed control (with the candid observation that classical RL underperformed swarm intelligence for multi-HAP coverage due to state discretization), and emerging LLM-agent hierarchies in which LLMs perform slow semantic planning while lightweight solvers and RL policies execute bounded actions.
Standardization, open resources, and evaluation
The standardization synthesis concludes that current frameworks can assemble a communication-centric HAP prototype but cannot guarantee interoperability of a fully integrated S²C²I platform. WRC-19 identified 31–31.3 GHz and 38–39.5 GHz worldwide; WRC-23 expanded HIBS operation at 2 GHz and 2.6 GHz. 3GPP Releases 17–19 provide transparent/regenerative payload models, onboard UPF, and store-and-forward, but lack HAP-specific profiles for quasi-stationary drift, HAP-UAV interfaces, and telemetry for sensing quality, cache freshness, or energy reserves. ETSI NFV-EVE 023 treating HAPs as constrained NFV infrastructure is highlighted as a step toward "HAP as mobile mini-data center," while O-RAN information models do not natively describe flight dynamics or propulsion constraints. IEEE 1920.1's inactive-reserved status illustrates the absence of an active end-to-end aerial-network framework.
For reproducibility, the survey assembles a modular evaluation chain from ERA5/NASA POWER environmental traces, QuaDRiGa/Sionna channel generation, 5G-LENA and OAI protocol evaluation, OpenSAND/Hypatia transport emulation, O-RAN SC/ETSI OSM orchestration, and AERPAW/Colosseum hardware validation—while plainly stating that no single open tool models the complete aircraft-atmosphere-radio-payload stack, and that field trials publish headline results rather than reusable benchmarks.
Evaluation methodology and case study
A four-level evaluation hierarchy is proposed: component/link validation, node-level payload feasibility under concurrent resource contention, network/service evaluation, and mission-level effectiveness. Its central methodological device is a conjunctive S²C²I service availability metric requiring simultaneous satisfaction of rate, closed-loop latency, sensing quality, data-freshness, task-completion, and protected-energy-reserve constraints—explicitly designed so that success in one function cannot mask failure in another.
The NS-3 emergency-response case study (100×100 km disaster area, three HAPs, twenty sensing UAVs, one gateway, 500 users, 200 IoT sensors) quantifies the thesis. The satellite-terrestrial baseline satisfies individual requirements for only 30.6–60.7% of the mission horizon. Communication-only HAPs raise rate satisfaction to 90.8%—the highest of any scheme—but leave latency and freshness satisfaction at 48.5% and 54.1%. Function-separated HAPs improve most dimensions yet drop energy-reserve satisfaction to 69.2% because independently optimized modules compete for onboard power. The S²C²I-integrated scheme achieves 88.7–94.3% across all six dimensions and raises conjunctive availability from 34.7% (communication-only) and 58.6% (function-separated) to 82.4%, while cutting feeder traffic from 1.17 TB to 0.49 TB. Notably, joint orchestration sacrifices 2.1 percentage points of rate satisfaction relative to the communication-only design—the paper's clearest demonstration that single-metric maximization and conjunctive service availability are competing objectives.
Limitations and open questions
The survey is candid about several dependencies. The case study is simulation-only (NS-3 on commodity hardware); no stratospheric flight validation of the conjunctive metric or the integrated orchestration scheme is provided. The proposed onboard S²C²I converged payload architecture is conceptual, and the authors acknowledge that full five-function native integration remains unexplored experimentally. Standardization gaps are substantive rather than incidental: no common HAP capability descriptor, no telemetry schema for energy/thermal/compute/storage/sensing state, and no cross-domain assurance procedure linking telecom orchestration to flight safety exist. On agentic AI, the paper identifies unsolved problems including memory synchronization across intermittently connected nodes, hallucination-driven unsafe allocations, and onboard model compression. Blockchain-based trust architectures are proposed with the caveat that consensus mechanisms must avoid proof-of-work and that ledgers verify provenance but not semantic correctness. Quantum-computing-enabled HAP results rest on favorable thermal assumptions without addressing radiation hardening or error correction. Finally, the evaluation chain depends on reanalysis datasets whose resolution may miss local turbulence and cloud blockage relevant to FSO availability.
Conclusion
This survey consolidates a fragmented HAP literature under a unified S²C²I framing, contributes a three-part architecture (connectivity taxonomy, converged payload, cloud-edge-HAP continuum), and proposes a conjunctive availability metric that operationalizes cross-functional co-design. Its strongest empirical support is the emergency-response case study, where joint orchestration more than doubles conjunctive service availability relative to a communication-only baseline while halving feeder traffic—at a modest cost in peak rate. The gap between this simulation evidence and certifiable, flight-validated deployment remains the survey's most significant open problem, and it frames the research agenda—trustworthy autonomy, semantic operation, digital twins, and standardizable HAP-native telemetry—that the paper leaves to subsequent work.
