SKYLINK: Integrated Sky–Ground Connectivity

• SKYLINK is a cross-domain research motif that defines multiple approaches to sky–ground connectivity, integrating terrestrial and non-terrestrial networks.
• It encompasses studies on distributed LEO link management, UAV route prediction, cross-view geo-localization, and programmable free-space optical communication.
• Researchers address dynamic topology and uncertainty using contextual bandits, AI orchestration, and measurement-driven methodologies across varied platforms.

In recent arXiv literature, “SKYLINK” denotes several technically distinct but conceptually related research efforts concerned with sky–ground connectivity, perception, and communication. The label appears in work on integrated terrestrial and non-terrestrial networking, fully distributed LEO link management, UAV connectivity prediction, cross-view geo-localization, and free-space optical transmission; adjacent empirical studies on Starlink emissions and mobile Starlink performance supply measurement context for the broader problem of sky-linked systems (Ozger et al., 2023, Sombre et al., 10 Sep 2025, Lu et al., 11 Feb 2026, Zhang et al., 29 Sep 2025, Cai et al., 11 May 2026, Grigg et al., 3 Jun 2025, Liu et al., 24 Jun 2025).

1. Terminological scope

No single standardized SKYLINK architecture is defined across the cited literature. Instead, the term is used as a research label for multiple system classes, each emphasizing a different technical bottleneck: 3D orchestration across TN/NTN domains, scalable routing in dynamic LEO constellations, route-level radio prediction for UAVs, cross-view retrieval between street and satellite imagery, or topological free-space optical signaling. This suggests that SKYLINK is presently better understood as a cross-domain research motif than as a unified protocol stack.

Usage	Technical focus	Representative source
6G-SKY / combined ASN	3D TN–NTN integration with DA2GC, A2AC, HAPS/HIBS, satellites	(Ozger et al., 2023)
SKYLINK in LEO networking	Fully distributed contextual link management over ISLs and GSLs	(Sombre et al., 10 Sep 2025)
SKYLINK for UAV radio maps	Transfer learning from ground crowdsourced data to aerial route prediction	(Lu et al., 11 Feb 2026)
SkyLink in CVGL	Street–satellite retrieval via UAV-mediated 3D scene alignment	(Zhang et al., 29 Sep 2025)
Optical SKYLINK	Programmable skyrmion states for free-space optical communication	(Cai et al., 11 May 2026)

2. Integrated sky–ground architectures and radio environments

The 6G-SKY project formulates a holistic, AI- and cloud-native architecture that integrates terrestrial networks and non-terrestrial networks into a combined airspace and NTN, or combined ASN. Its multi-layer design comprises a terrestrial layer with 4G/5G/5G-Advanced/6G TNs, an aerial layer with UAVs, flying taxis, commercial airplanes, and HAPS/HIBS, and a space layer with satellite systems. The architecture explicitly distinguishes service management, data exposure, application interaction/support functions, 3D end-to-end and domain-oriented orchestration, slicing, mobility handling, and multi-hop A2AC mesh routing through a limited number of multi-domain gateways. The associated link taxonomy includes space-to-ground, direct air-to-ground communication, air-to-air communication, space-to-air, HAPS/HIBS links, and terrestrial backhaul, with managed latency, interference coordination, beam tracking, and trustworthy AI treated as core design requirements (Ozger et al., 2023).

A more deployment-oriented formulation appears in partially integrated coordinated satellite–5G networking for the low-altitude economy. There, satellite beams and 5G carriers share time–frequency resources under coarse synchronization over a coordination interval $T$, while optimization relies only on large-scale CSI obtained from a Radio Map. The framework uses link-feature-aided clustering and divide-and-conquer scheduling rather than solving the mixed-integer problem directly, with deployment guidance that $T$ can be several or tens of seconds and a case study using $T=10$ s. Its representative rate model is based on

$$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$

with cross-system interference constrained through thresholding and clustering (Wang et al., 15 Mar 2026).

For low-altitude UAV operations, route-level radio maps are treated as a distinct SKYLINK function. “Transfer to Sky” defines the route prediction problem on $x(s)=(x(s),y(s),z(s))$ and learns $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$ from dense but non-uniform ground measurements $\mathcal{D}_g$ and extremely sparse aerial measurements $\mathcal{D}_a$, with $|\mathcal{D}_a|/|\mathcal{D}_g|<10^{-3}$. The proposed pipeline uses simulation pretraining with Sionna ray tracing, ADDA-based adversarial alignment, and decoder-only fine-tuning. On Meituan data, the reported route-level RSRP RMSE is 3.4 / 5.3 / 12.7 dB for best/mean/worst cases, compared with 9.5 / 12.1 / 18.3 dB for Kriging and 7.8 / 10.2 / 16.9 dB for the autoencoder baseline (Lu et al., 11 Feb 2026).

Urban visibility constraints are formalized analytically by the skyline process. In that framework, the first skyline process

$\omega_1(\psi)=\sup_{b_i\in V(\psi)} \theta_i$

gives the blockage elevation angle along azimuth $T$0, and a satellite at elevation $T$1 is line-of-sight visible only if $T$2. For dense urban settings, the analysis derives closed-form expressions for the blockage-angle CDF, ACF, and PSD, and reports that independence in dual-satellite blockage is approached around $T$3 for $T$4 m, $T$5 m, and $T$6. The same study reports that about 98% of satellites are blocked when $T$7 and $T$8 m (Lee et al., 27 Mar 2026).

3. Distributed LEO link management

In the networking literature proper, SKYLINK is a fully distributed learning strategy for link management in LEO satellite networks with inter-satellite links and ground-station-satellite links. The system models the constellation as a time-varying directed graph $T$9, where $T=10$0 includes satellites, ground stations, and the internet node $T=10$1, and where each satellite independently allocates incoming traffic across its outgoing links each slot. The global objective is to minimize the rate-weighted average cost $T=10$2, where path delay is capped at $T=10$3 and dropped traffic contributes $T=10$4, so minimizing cost jointly minimizes average delay and drop rate. Link selection is implemented as a contextual multi-armed bandit with tile coding over continuous distance contexts, using a lower-is-better UCB score

$T=10$5

followed by ranking and a capacity-aware water-filling allocator (Sombre et al., 10 Sep 2025).

The reported simulator uses a OneWeb near-polar Walker Star configuration with $T=10$6 satellites, $T=10$7 ground stations, $T=10$8 s, $T=10$9 ms, and 25.4 million users. Under that setup, SKYLINK reduces the weighted sum of average delay and drop rate by 29% relative to the bent-pipe approach and by 92% relative to Dijkstra, lowers drop rates by 95% relative to k-shortest paths, 99% relative to Dijkstra, and 74% relative to the bent-pipe baseline, and achieves up to 46% higher throughput, while maintaining constant computational complexity with respect to constellation size. In a failure scenario with 3% of satellites losing GSLs for two days, SKYLINK keeps average drop rate at or below 0.7% and limits cost increase to less than 10%, whereas the bent-pipe baseline exhibits a 67.6% cost increase (Sombre et al., 10 Sep 2025).

4. Measurement studies: interference, mobility, and visibility

A major empirical context for sky-linked systems is unintended Starlink emission in radio astronomy. The largest survey to date across the SKA-Low frequency range used the Engineering Development Array 2, a 256 dual-polarisation dipole SKA-Low prototype station, and analyzed approximately 76 million full-sky images over 694.2 hours. It reported 112,534 individual detections of 1,806 unique Starlink satellites, with 1,623 unique v2-mini Ku satellites corresponding to 76% of all v2-mini Ku satellites in orbit at the time, and 175 unique v2-mini Direct to Cell satellites corresponding to 71% of the DTC population. In the worst cases, nearly 30% of images contained at least one detectable Starlink satellite; excluding 136.7 MHz and $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$0 MHz datasets, the mean flux density due to Starlink emission was 93 Jy/beam, reaching 312 Jy/beam at 170.3 MHz. The study detected 13 satellites between 73.00–74.60 MHz and 703 between 150.05–153.00 MHz, observed a temporally shifting comb-like broadband structure, and found that XX and YY flux densities were strongly anti-correlated, implying significant linear polarisation and a time-varying polarisation vector (Grigg et al., 3 Jun 2025).

A second measurement line characterizes mobile Starlink performance directly. The Starlink Robot mounts a Starlink Mini on a Unitree GO2 wheeled robot together with a fisheye camera, Livox Mid-360 LiDAR, IMU, GPS, wheel odometry, terminal gRPC metrics, packet capture, active probing, and satellite tracking, all aligned at sub-millisecond precision. Its preliminary results show that at pedestrian speeds of approximately 0.8–2.0 m/s, RTT distributions remain stable and are concentrated around 35–45 ms in speed-comparison runs, with open-sky RTT generally ranging from 20–40 ms and handovers occurring approximately every 15 s. Under tree canopy and reduced sky visibility, RTT instability increases markedly, with frequent spikes to 40–100 ms; the dominant explanatory variables are sky visibility, obstruction angles, satellite elevation/azimuth, and motion-induced orientation dynamics rather than speed itself (Liu et al., 24 Jun 2025).

5. SkyLink in cross-view geo-localization

In computer vision, SkyLink denotes a cross-view geo-localization framework that seeks to rank satellite images for a given street-level query image. Its central claim is that direct feature matching between street and satellite viewpoints is insufficient because of extreme viewpoint disparity, occlusion, limited field of view, and semantic degradation. The method therefore introduces three coupled modules: the Google Retrieval Enhancement Module, which adds the top 50% most similar street images and contributes 2,463 supplemental street-view images; Patch-Aware Feature Aggregation, which forms descriptors from multiple local patch features using residual MLP mixing; and a Multi-scale 3D Scene Bridge Module, which reconstructs point clouds from low-, medium-, and high-altitude UAV image sequences using VGGT and encodes them with PointCLIP (Zhang et al., 29 Sep 2025).

The learning objective combines self-supervised and cross-view contrastive terms. With street descriptors $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$1, satellite descriptors $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$2, UAV-scale descriptors $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$3, and temperature $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$4, the total loss is

$$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$5

with the best-performing configuration reported at $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$6. Using DINOv2-L, 448$$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$7448 inputs, SGD with momentum 0.9, and 40 epochs on two NVIDIA RTX 4090 GPUs, the method achieves Recall@1 $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$8, Recall@5 $$R_k=\sum_{n\in\mathcal{N}} x_{k,n} B_n \log_2\!\bigl(1+\mathrm{SINR}_{k,n}\bigr),$$9, Recall@10 $x(s)=(x(s),y(s),z(s))$0, and AP $x(s)=(x(s),y(s),z(s))$1 on University-1652 under Ground $x(s)=(x(s),y(s),z(s))$2 Satellite retrieval; the abstract additionally reports 25.75% Recall@1 on the UAVM2025 Challenge. Component ablations show that PAFA alone reaches 18.42% Recall@1, while adding SSL, the 3D bridge, and GREM progressively raises performance to the full 27.06% (Zhang et al., 29 Sep 2025).

6. Optical SKYLINK and programmable skyrmion communication

In photonics, SKYLINK refers to a free-space optical communication scheme based on dynamically reconfigurable optical skyrmions generated by a silicon microring-resonator optical phased array. The device operates at $x(s)=(x(s),y(s),z(s))$3 nm on SiO$x(s)=(x(s),y(s),z(s))$4, uses a cascaded $x(s)=(x(s),y(s),z(s))$5 Y-branch splitter tree with thermo-optic phase shifters, and interleaves inner-grating and outer-grating microring resonators to generate decoupled left-circularly polarized and right-circularly polarized radiation bases. Full-ring 3D-FDTD validation reports resonance at $x(s)=(x(s),y(s),z(s))$6 nm, loaded quality factors $x(s)=(x(s),y(s),z(s))$7 and $x(s)=(x(s),y(s),z(s))$8, DOCP values of $x(s)=(x(s),y(s),z(s))$9 and $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$0, polarization fractions of 90.27% and 91.40%, and peak upward efficiencies of approximately 1.43% and 3.04% for the LCP and RCP emitters, respectively (Cai et al., 11 May 2026).

The skyrmion state is described through the normalized Stokes vector $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$1, with skyrmion number

$R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$2

By programming the OPA phase gradients, the system tunes $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$3, yielding measured skyrmion numbers $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$4, $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$5, $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$6, and $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$7, while sweeping $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$8 switches between Néel-type and Bloch-type textures without changing $R_m(s)=f_m(x(s);\mathcal{D}_g,\theta)$9. A 4-symbol communication link based on the alphabet $\mathcal{D}_g$0 is evaluated under Kolmogorov phase-screen turbulence with 500 Monte Carlo trials per turbulence level. Compared with ideal LG-OAM encoding, the skyrmion-encoded link maintains a lower symbol error rate over a broader turbulence range, and at a representative distortion of approximately 1.09 rad the received $\mathcal{D}_g$1 values remain clustered near their calibrated centroids while OAM detection spreads across multiple $\mathcal{D}_g$2 values (Cai et al., 11 May 2026).

7. Recurring technical themes and open issues

Across these otherwise disparate uses of SKYLINK, several technical motifs recur. One is geometry-awareness: 6G-SKY relies on 3D orchestration and multi-layer mobility control; partially integrated satellite–5G sharing relies on position-aware large-scale CSI and a Radio Map; route-level UAV planning depends on 3D radio-map prediction; the skyline process reduces NTN access to stochastic blockage elevation; SkyLink geo-localization uses UAV-mediated 3D reconstructions as a shared embedding scaffold; and the Starlink Robot treats sky visibility and obstruction geometry as first-order predictors of connectivity (Ozger et al., 2023, Wang et al., 15 Mar 2026, Lu et al., 11 Feb 2026, Lee et al., 27 Mar 2026, Zhang et al., 29 Sep 2025, Liu et al., 24 Jun 2025).

A second motif is robustness under uncertainty, but the source of uncertainty changes by domain. In LEO routing, the uncertainty lies in time-varying topology, link capacity, and failures; in radio astronomy, it lies in unintended broadband and narrowband emissions entering protected RAS allocations; in mobile Starlink, it lies in occlusion and handover dynamics; in optical SKYLINK, it lies in atmospheric turbulence; and in cross-view retrieval, it lies in viewpoint disparity and semantic degradation. The mitigation mechanisms are correspondingly heterogeneous—contextual bandits and water-filling, continuous RFI monitoring and polarimetric discrimination, multi-modal occlusion sensing, topology-based decoding, or contrastive 3D alignment—but all are explicitly designed around non-stationarity rather than around static optimality (Sombre et al., 10 Sep 2025, Grigg et al., 3 Jun 2025, Liu et al., 24 Jun 2025, Cai et al., 11 May 2026, Zhang et al., 29 Sep 2025).

A plausible implication is that future uses of the SKYLINK label will continue to converge on systems that are position-aware, cross-layer, and measurement-driven. The cited literature also indicates persistent open issues: explicit UEMR regulation in protected RAS bands, robust handover and routing in 3D multi-connectivity, transferability of radio maps across cities and bands, dependency on UAV-derived 3D supervision in CVGL, and the speed and energy limits of thermo-optic optical phased arrays (Grigg et al., 3 Jun 2025, Ozger et al., 2023, Lu et al., 11 Feb 2026, Zhang et al., 29 Sep 2025, Cai et al., 11 May 2026).